text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1710.02210v1 [cs.AI] 5 Oct 2017
Exploration in Feature Space for
Reinforcement Learning
Suraj Narayanan Sasikumar
Under the Supervision of
Professor Marcus Hutter
A thesis submitted for the degree of
Master of Computing (Advanced)
The Australian National University
May 2017
c Suraj Narayanan Sasikumar 2017
Declaration
This thesis is an account of the research undertaken at the Research School of
Computer Science, The Australian National University, Canberra, Australia.
The work presented in this thesis represents work conducted between July 2016 and
May 2017. This work has been accepted for publication as "Count-Based Exploration
in Feature Space for Reinforcement Learning" at the 26th International Joint
Conference on Artificial Intelligence (IJCAI) Melbourne, Australia, August 19-25 2017
(Martin et al., 2017).
The implementation of the algorithm, and the design of the infrastructure required
to empirically evaluate the algorithm, are entirely my own original work. The
design of the algorithm itself was done in collaboration with my colleague Jarryd
Martin.
Except when acknowledged in the customary manner, the material presented in this
thesis is, to the best of my knowledge, original and has not been submitted in whole
or part for a degree in any university.
Suraj Narayanan Sasikumar
25 May 2017
Supervisors:
• Professor Marcus Hutter (Australian National University)
• Tom Everitt (Australian National University)
Convenor:
• Professor John Slaney (Australian National University)
iii
To my Friends and Family who literally made this possible.
Acknowledgments
At the culmination of two years of hard work, I would like to take this opportunity
to acknowledge the role they have played in my life.
• My supervisor Marcus Hutter, whose lectures inspired me to pursue Reinforcement Learning. Hearing him talk about Mathematics and AI with childlike
enthusiasm has always been an inspiration to me.
• Tom Everitt, his steadfast and professional approach to supervising my thesis
helped me a great deal in making this a success.
• Jarryd Martin, my collaborator in this Project. Words cannot express the happiness that I feel, to have a friend in you. The days we toiled over broken theories
and random agents were not for nothing. We did it. . . and it could never have
been achieved without you. Your determination, motivation and quite frankly
the sheer energy is something that I aspire to.
• John Aslanides, first a huge thank-you for talking your time to provide your
comments and feedback on this thesis. Your rationality and cut-to-the-chase
attitude. Your clarity of thought, work-ethic and passion for growth has been
inspirational for me.
• Lulu Huang, for talking care of me throughout this whole endeavour, and being
a wonderful roommate.
• Boris Repasky, the man with infinite rigour. All the arguments and debates
we’ve had throughout the year has only been for the better. Sina Eghbal, for
being there through tough times as an unassuming friend. Darren Lawton, for
forcing me to play Basketball. My AI Labmates, Sultan Javed, Owen Cameron,
Arie Slobbe, Elliot Catt, for the motivation and support.
• My father Sasikumar, sister Sumitha, for believing in me and being a vocal
supporter of my decisions. My in-laws Rasmi, M Vasudevan, and Geetha, for
supporting my decision to study. I miss you all very much.
• My wife Ramya Vasudevan, her sacrifices and compromises are the reason I am
able to do what I want to do, and I am forever indebted. None of this would
even make sense without you in my life.
• My mother Sudhamony whose guidance and sacrifices made the man I am.
Through out my life she has been a constant source of moral guidance. I am a
better person because of her.
vii
Abstract
The infamous exploration-exploitation dilemma is one of the oldest and most important problems in reinforcement learning (RL). Deliberate and effective exploration
is necessary for RL agents to succeed in most environments. However, until very
recently even very sophisticated RL algorithms employed simple, undirected exploration strategies in large-scale RL tasks.
We introduce a new optimistic count-based exploration algorithm for RL that is
feasible in high-dimensional MDPs. The success of RL algorithms in these domains
depends crucially on generalization from limited training experience. Function approximation techniques enable RL agents to generalize in order to estimate the value
of unvisited states, but at present few methods have achieved generalization about
the agent’s uncertainty regarding unvisited states. We present a new method for
computing a generalized state visit-count, which allows the agent to estimate the
uncertainty associated with any state.
In contrast to existing exploration techniques, our φ-pseudocount achieves generalization by exploiting the feature representation of the state space that is used
for value function approximation. States that have less frequently observed features
are deemed more uncertain. The resulting φ-Exploration-Bonus algorithm rewards
the agent for exploring in feature space rather than in the original state space. This
method is simpler and less computationally expensive than some previous proposals, and achieves near state-of-the-art results on high-dimensional RL benchmarks.
In particular, we report world-class results on several notoriously difficult Atari 2600
video games, including Montezuma’s Revenge.
ix
x
Contents
Declaration
iii
Acknowledgments
vii
Abstract
ix
Contents
xi
List of Figures
xiii
List of Tables
xv
List of Algorithms
xvii
1
Introduction
1.1 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 The Exploration/Exploitation Dilemma . . . . . . . . . . . . . . . . . . .
1.3 Summary of Contributions . . . . . . . . . . . . . . . . . . . . . . . . . .
2
Background and Related Work
2.1 Classical Reinforcement Learning . . . . . . . . . . . . . . . . . .
2.1.1 Reinforcement Learning Algorithms . . . . . . . . . . . .
2.1.2 Function Approximation . . . . . . . . . . . . . . . . . .
2.2 Exploration Strategies for Reinforcement Learning . . . . . . . .
2.2.1 Taxonomy of Exploration Strategies . . . . . . . . . . . .
2.3 The Optimism in the Face of Uncertainty Principle . . . . . . .
2.3.1 OFU using Count-Based Exploration Bonuses . . . . . .
2.3.2 Tabular Count-based Exploration Algorithms . . . . . .
2.3.3 Generalized Visit-counts for Exploration in Large MDPs
2.4 Bayes-Adaptive RL . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Intrinsic Motivation . . . . . . . . . . . . . . . . . . . . . . . . . .
3
1
1
2
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
7
10
11
12
12
13
14
15
15
16
Exploration in Feature Space
3.1 Drawbacks of Existing Exploration Methods for Large MDPs . . . .
3.1.1 Choosing a Novelty Measure . . . . . . . . . . . . . . . . . . .
3.1.2 Separate Generalization Methods for Value and Uncertainty .
3.2 Estimating Novelty in Feature Space . . . . . . . . . . . . . . . . . . .
3.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
19
19
19
21
22
22
xi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
xii
3.3
3.4
4
5
6
3.2.2 Design Decisions . . . . . . . . . . . . . . . .
The φ-EB Algorithm . . . . . . . . . . . . . . . . . .
3.3.1 Feature Visit-Density . . . . . . . . . . . . . .
3.3.2 The φ-pseudocount . . . . . . . . . . . . . .
3.3.3 The φ-Exploration Bonus algorithm (φ-EB)
3.3.4 LFA with φ-EB . . . . . . . . . . . . . . . . .
3.3.5 Complexity Analysis . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . .
Implementation
4.1 Software Architecture . . . . . . . . .
4.1.1 Modular Overview . . . . . . .
4.1.2 Agent-Environment Work-flow
4.2 Implementation Details . . . . . . . .
4.2.1 Feature Visit-Density . . . . . .
4.2.2 Updating Factor Densities . . .
4.2.3 Exploration Bonus . . . . . . .
4.2.4 Action Selection . . . . . . . . .
4.2.5 SARSA(λ)+φ-EB . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
23
24
24
26
26
28
29
30
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
31
32
34
35
35
37
40
40
43
Empirical Evaluation
5.1 Evaluation Framework . . . . . . . . . . . . . . .
5.1.1 Arcade Learning Algorithm (ALE) . . . .
5.1.2 Blob-PROST Feature Set . . . . . . . . . .
5.2 Empirical Evaluation . . . . . . . . . . . . . . . .
5.2.1 Evaluation Methodology . . . . . . . . .
5.2.2 Sparse Reward Games . . . . . . . . . . .
5.2.3 Dense Reward Games . . . . . . . . . . .
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Boltzmann vs. e-greedy Action Selection
5.3.2 Comparison with e-greedy . . . . . . . .
5.3.3 Comparison with Leading Algorithms .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
45
45
45
47
48
48
51
54
55
55
56
57
Conclusion and Future Work
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
59
List of Figures
1.1
The agent-environment interaction cycle (Sutton and Barto, 1998) . . . .
2
2.1
Generalized Policy Iteration (Sutton and Barto, 1998) . . . . . . . . . . .
8
3.1
3.2
3.3
3.4
3.5
Q*bert Level 1 . . . . . . . . . . . . . . . . . . . . . . . . . .
Q*bert Level 2 . . . . . . . . . . . . . . . . . . . . . . . . . .
Two levels of the Atari2600 game Q*bert . . . . . . . . . .
Novelty Measure in Feature Space . . . . . . . . . . . . . .
Flow Chart for computing the exploration bonus of φ-EB
4.1
Agent-Environment interaction framework for SARSA(λ)+φ-EB . . . . 32
5.1
5.2
5.3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
20
20
20
22
27
Game screens from 55 Atari 2600 games. . . . . . . . . . . . . . . . . . . 46
High level working of ALE for an RL algorithm. . . . . . . . . . . . . . . 46
Five games in which exploration is both difficult and crucial to performance. From top left: Montezuma’s Revenge, Venture, Freeway,
Frostbite, and Q*Bert. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4 Montezuma’s Revenge: Rooms visited by undirected exploration (DQN+egreedy, above) vs. directed exploration (DQN+CTS-EB, below) (Bellemare et al., 2016). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.5 Venture Outer Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.6 Venture Inner Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.7 Two visual states of venture (Atari 2600) . . . . . . . . . . . . . . . . . . 52
5.8 Freeway (Atari2600) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.9 Qbert (Atari2600) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.10 Frostbite (Atari2600) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.11 Average training score for Action Selection using Boltzmann vs. egreedy. Shaded regions describe one standard deviation. Dashed lines
represent min/max scores. . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.12 Average training score for SARSA-φ-EB vs. SARSA-e. Shaded regions
describe one standard deviation. Dashed lines represent min/max
scores. (Martin et al., 2017) . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
xiii
xiv
LIST OF FIGURES
List of Tables
5.1
Average evaluation score for leading algorithms. Sarsa-φ-EB and Sarsae were evaluated after 100M training frames on all games except Q*bert,
for which they trained for 80M frames. All other algorithms were evaluated after 200M frames. (Martin et al., 2017) . . . . . . . . . . . . . . . . 58
xv
xvi
LIST OF TABLES
List of Algorithms
1
2
φ-exploration bonus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
LFA with φ-EB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3
4
5
6
7
8
Implementation of Feature Visit Density . . . . . . . . . . . . .
Factor Distribution Update . . . . . . . . . . . . . . . . . . . . .
Exploration Bonus . . . . . . . . . . . . . . . . . . . . . . . . . .
Action Selection: e-Greedy . . . . . . . . . . . . . . . . . . . . .
Action Selection: Boltzmann Distributed . . . . . . . . . . . . .
Reinforcement Learning with SARSA(λ) and φ-EB exploration
xvii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
39
40
42
42
43
xviii
LIST OF ALGORITHMS
Chapter 1
Introduction
‘No great discovery was ever made
without a bold guess.’
Isaac Newton
1.1
Reinforcement Learning
Machine Learning is a field in computer science that allows computers to dynamically
generate novel algorithms that otherwise cannot be explicitly programmed. These
algorithms, called hypotheses, generalize patterns and regularities from observed realworld data using statistical techniques (Bishop, 2007). Reinforcement Learning (RL) is
a field of machine learning that deals with optimal sequential decision making in an
unknown environment with no explicitly labelled training data. The RL framework
is one of the fundamental models that best describes how intelligent beings interact
with their world to achieve a goal. An RL algorithm is given agency to interact
with its surroundings, and is aptly called an agent. The world with which the agent
interacts is called its environment. The unsupervised nature of RL algorithms means
that the agent has to develop a policy for acting in an unknown environment by trialand-error (Sutton and Barto, 1998). In every such interaction the agent performs an
action on the environment and receives a percept. The percept consists of the current
configuration of the environment, called state, and a scalar feedback signal, called
reward. The reward signal indicates how good the sequence of actions of the agent
was. The goal of an RL agent is based on the concept of the reward hypothesis:
Definition 1 (Reward Hypothesis). (Sutton, 1999) Any notion of a goal or purpose of an
intelligent agent can be described as the maximization of expected cumulative reward.
The existence of an extrinsic feedback signal makes RL algorithms also somewhat
supervised in nature - thus RL algorithms are in some sense both supervised and
unsupervised (Barto and Dietterich, 2004).
As an example, consider an agent playing a car racing game in which the goal is
to reach the finish line as soon as possible. To model the goal as a cumulative reward
maximization problem, we give the agent a negative reward every time step, thereby
1
2
Introduction
incentivizing the agent to reach the finish line as quickly as possible. This example
illustrates how an objective can be modelled as the maximization of expected cumulative reward, and the goal of an agent as a sequence of actions that achieves it. The
interaction between agent and environment continues until the agent converges to
an optimal sequence of actions for each state in the environment. This interaction
is called the agent-environment interaction cycle, as illustrated in Figure 1.1. Each
iteration of the interaction is called a time-step, often denoted by the subscript t to
distinguish states, actions, and percepts between time-steps.
Figure 1.1: The agent-environment interaction cycle (Sutton and Barto, 1998)
1.2 The Exploration/Exploitation Dilemma
In an online decision-making setting such as the reinforcement learning problem, an
agent is faced with two choices - explore or exploit. The term exploration in an active
learning system is defined as the process of deliberately taking a non-greedy action
with the sole aim of gathering more information about the environment. Exploration
plays a fundamental role in reinforcement learning algorithms. It is born out of the
notion that an optimal long-term policy might involve short-term sacrifices. Alternatively, exploitation is the act of taking the best possible action given the current
information about the environment. A central challenge in reinforcement learning is
to find the sweet spot between exploration and exploitation, i.e., to figure out when
to explore and when to exploit. This problem is known as the exploration-exploitation
dilemma.
At present there are a number of provably efficient exploration methods that
are effective in environments with low-dimensional state-action spaces. Most of the
exploration algorithms which enjoy strong theoretical guarantees implement the socalled "Optimism in the Face of Uncertainty" (OFU) principle. This heuristic encourages the agent to be optimistic about the reward it might attain in less explored parts
of the environment. The agent seeks out states with higher associated uncertainty,
and in doing so reduces its uncertainty in a very efficient way. Many algorithms that
implement this heuristic do so by adding an exploration bonus to the agent’s reward
signal. This bonus is usually a function of a state visit-count; the agent receives
higher exploration bonuses for exploring less frequently visited states (about which
§1.3 Summary of Contributions
3
it is less certain).
Unfortunately, these algorithms do not scale well to high-dimensional environments. In these domains, the agent can only visit a small portion of the state space
while it is training. The visit-count for most states is always zero, even after training
is finished. Nearly all states will be assigned the same exploration bonus throughout
training. This renders the bonus useless as a tool for efficient exploration. All unvisited states appear to the agent as equally uncertain. This problem arises because
these count-based OFU algorithms fail to generalise the agent’s uncertainty from one
context to another. Even if an unvisited state has very similar features to a frequently
visited one, the agent will treat the former as a complete unknown. Consequently
even the sophisticated algorithms that are suitable for the high-dimensional setting –
e.g. those that use deep neural networks for policy evaluation – tend to use simple,
inefficient exploration strategies.
Success in the high-dimensional setting demands that the agent represent the
state space in a way that allows generalisation about uncertainty. This sort of generalisation would allow that the agent’s uncertainty be lower for states with familiar
features, and higher for states with novel features, even if those exact states haven’t
been visited. What we require, then, is an efficient method for computing a suitable
similarity measure for states. That is the key challenge addressed in this thesis.
1.3
Summary of Contributions
This thesis presents a new count-based exploration algorithm that is feasible in environments with large state-action spaces. It can be combined with any value-based
RL algorithm that uses linear function approximation (LFA). The principal contribution is a new method for computing generalised visit-counts. Following Bellemare
et al. (2016), we construct a visit-density model in order to measure the similarity
between states. Our approach departs from theirs in that we do not construct our
density model over the raw state space. Instead, we exploit the feature map that is
used for value function approximation, and construct a density model over the transformed feature space. This model assigns higher probability to state feature vectors
that share features with visited states. Generalised visit-counts are then computed
from these probabilities; states with frequently observed features are assigned higher
counts. These counts serve as a measure of the uncertainty associated with a state.
Exploration bonuses are then computed from these counts in order to encourage the
agent to visit regions of the state-space with less familiar features.
Our density model can be trivially derived from any feature map used for LFA,
regardless of the application domain, and requires little or no additional design. In
contrast to existing algorithms, there is no need to perform a special dimensionality
reduction of the state space in order to compute our generalised visit-counts. Our
method uses the same lower-dimensional feature representation to estimate value
and to estimate uncertainty. This makes it simpler to implement and less computationally expensive than some existing proposals. Our evaluation demonstrates that
4
Introduction
this simple approach achieves near state-of-the-art performance on high-dimensional
RL benchmarks.
Chapter 2
Background and Related Work
‘Each night, when I go to sleep, I die.
And the next morning, when I wake up,
I am reborn.’
Mahatma Gandhi
In this chapter we give a formal treatment of the RL problem. We then provide
a taxonomy of RL algorithms and the challenges posed by classical RL algorithms.
Further, we talk about relevant research findings on how to solve these challenges.
2.1
Classical Reinforcement Learning
In Classical RL (CRL), the environment is assumed to be fully observable, ergodic,
and every state has the Markov property. The branch of reinforcement learning where
these assumptions are lifted is called General Reinforcement Learning (GRL) (Hutter,
2005).
Definition 2 (Markov property). Future states are only dependent on the current states
and action, and are independent of the history of percepts. Formally,
P ( s t +1 = s 0 , r t +1 = r | s t , a t , r t , s t −1 , a t −1 , . . . , r 1 , s 0 , a 0 ) = P ( s t +1 = s 0 , r t +1 = r | s t , a t )
for all s0 , r, and histories st , at , rt , st−1 , at−1 , . . . , r1 , s0 , a0
A Markov Decision Process (MDP) captures the above assumptions about the environment, and so in the CRL context the environment is modelled as an MDP (Puterman, 1994). Thus, the CRL problem now reduces to the problem of finding an
optimal policy for an unknown MDP.
Definition 3 (Markov Decision Process). A Markov Decision Process is a Tuple hS , A, P , R, γi
representative of a fully-observable environment in which all states are Markov.
• S is a finite set of states
• A is a finite set of actions
5
Background and Related Work
6
• Pssa 0 = P[st+1 = s0 | st = s, at = a] are the transition probabilities
a = E[r
0
• Rss
0
t+1 | st = s, at = a, st+1 = s ] is the expected value of the reward resulting
0
from the transition s, a, s
• γ is the discount factor which weights the relative importance of immediate rewards to
future rewards.
If the dynamics (transition and reward distributions) of the MDP are known, then
we can use dynamic programming methods to directly plan on the MDP to find an
optimal policy. In the RL context, in which the system dynamics are unknown, we
have to use iterative RL algorithms such as TD-learning (Sutton, 1988) to find a good
policy asymptotically.1
Definition 4 (Policy). A policy may be deterministic or stochastic. A deterministic policy
is a mapping from the states to actions.
π:S→A
A stochastic policy is a probability distribution over the set of actions given a state.
π ( a | st = s)
Value
The most common way to characterize the quality of a given policy is to define a
function that computes how valuable it is to follow the policy from a given state (or
state-action pair). This notion of value is expressed in terms of future rewards the
agent could expect, if it had chosen to follow the given policy.
Definition 5 (State-Value Function). The state-value function, V π (s) is a mapping from
states to R. The value of a state s ∈ S under policy π is the expected discounted cumulative
reward given that the agent starts in state s and follows policy π thereafter.
"
V ( s ) = Eπ
π
∞
∑ γ r t + k +1 | s t = s
#
k
k =0
Definition 6 (Action-Value Function). The action-value function, Qπ (s, a) is a mapping
from state-action pairs to R. The action-value of the state-action pair (s, a) under policy π is
the expected discounted cumulative reward given that the agent starts in state s, takes action
a, and follows policy π thereafter.
"
Qπ (s, a) = Eπ
∞
∑ γk rt+k+1 | st = s, at = a
#
k =0
1 Asymptotic
analysis is one of the few theoretical tools we have to analyse RL algorithms in a
domain-agnostic way.
§2.1 Classical Reinforcement Learning
7
Bellman Equations
Bellman equations form the basis for how to compute, approximate, and learn value
functions in the RL setup (Sutton and Barto, 1998). They arise naturally from the
structure of an MDP by capturing the recursive relationship between the value of a
state and the value of its successor states. The two Bellman equations for the statevalues and action-values can be defined as follows.
Definition 7 (Bellman Equation for state-value function of an MDP).
"
π
V (s) =
∑ π ( a | s) ∑
0
a∈A
Pssa 0
h
a
Rss
0
π
0
+ γV (s )
i
#
s ∈S
Definition 8 (Bellman Equation for action-value function of an MDP).
Qπ (s, a) =
∑
s0 ∈S
h
a
Pssa 0 Rss
0 +γ
∑
π ( a0 | s0 ) Qπ (s0 , a0 )
i
a0 ∈A
We can now use the value function to define a partial ordering over policies. A
policy is said to be better than another when the expected return of one policy is
greater than or equal to the other for all states. Formally, π % π 0 ⇐⇒ V π (s) ≥
0
V π (s) ∀s ∈ S . From the imposed partial ordering it has been shown that there
exists at least one policy, π ∗ , such that π ∗ % π for all policies π, although it might not
be unique (Bertsekas and Tsitsiklis, 1996). The Bellman Optimality Equations provide
a mathematical framework for talking about the optimal policy just by replacing the
sum over actions with a max operator. Intuitively, this represents a policy that is
greedy with respect to the value of its successor states.
Definition 9 (Bellman Optimality Equation for state-values).
∗
V π (s) ≡ V ∗ (s) = max
a∈A
∑
s0 ∈S
i
h
a
∗ 0
Pssa 0 Rss
0 + γV ( s )
Definition 10 (Bellman Optimality Equation for action-values).
π∗
∗
Q (s, a) ≡ Q (s, a) =
∑
s0 ∈S
Pssa 0
h
a
Rss
0
∗
0
0
+ γ max
Q (s , a )
0
i
a ∈A
For finite MDPs with known environment dynamics, the Bellman Optimality
Equations have a unique solution. Unfortunately in the RL setup we deal with an
unknown MDP. Thus, almost all of the RL algorithms approximate the Bellman Optimality Equations for an unknown MDP and try to iteratively find an optimal policy
asymptotically.
2.1.1
Reinforcement Learning Algorithms
The fundamental difference between an RL problem and a planning problem is the
knowledge of the environment dynamics. In a planning problem the model of the
8
Background and Related Work
environment is already known and the problem boils down to finding an optimal
policy in the environment. In an RL problem, the agent is dropped into an unknown
environment the dynamics of which is unknown. This makes reinforcement learning
a hard problem. This distinction gives rise to two categories of RL algorithms, namely
model-based and model-free.
The class of algorithms that learns the model of the environment, and then does
planning within the learned model are called model-based RL algorithms. These ala ) of the
gorithms learn the transition probabilities (Pssa 0 ) and reward functions (Rss
0
MDP by iteratively simulating the environment and updating the simulation to better represent the true environment. This approach to solve unknown MDP’s is computationally intensive, especially in large or continuous problems. Value iteration
and policy iteration are two dynamic programming algorithms that have a planningbased approach to the RL problem. On the other hand, model-free algorithms directly
learn the optimal policy using an intermediary quantity (usually the value-function).
Generalized Policy Iteration (GPI)
The overarching theme of almost all value-function based CRL algorithms is the backand-forth between two interacting processes, prediction and control, eventually resulting in convergence. Prediction refers to policy-evaluation where the value-function is
estimated for the current policy. Control on the other hand aims to find a policy that
is greedy with respect to the current value-function (state-value or action-value).
Figure 2.1: Generalized Policy Iteration (Sutton and Barto, 1998)
The ‘Prediction and Control’ process converges when it produces no significant
change, that is, the value-function is consistent with the current policy and the policy
is greedy with respect to the current value-function.
Temporal Difference Learning
Temporal Difference learning (TD learning) is a common RL algorithm; it is a modelfree algorithm that combines Monte Carlo methods with the ideas from dynamic
programming. TD learning allows the agent to directly learn from its experience
of the environment. Following the GPI theme, we need a strategy for prediction
§2.1 Classical Reinforcement Learning
9
and control. In TD prediction we use the sampling of Monte Carlo methods and
bootstrapping (updating from an existing estimate) of DP algorithms to estimate the
current value-function.
Definition 11 (Update formula for state-value function).
V (st ) ← V (st ) + α[rt+1 + γV (st+1 ) − V (st )]
TD(0) is a TD learning algorithm that updates state-values after each time-step,
so the learning process is fast and on-line. The target for the TD(0) update formula
uses the existing estimate of V (st+1 ), hence we say the algorithm bootstraps.
As the agent interacts with the environment more, TD learning is able to generate
a better estimate of the value-functions. In the limit, if each state (or state-action
pair) is visited infinitely often with some additional constraints on the learning rates,
convergence to the true value-function is guaranteed (Bertsekas and Tsitsiklis, 1996).
In TD control we want to optimize the value-function of an unknown environment. There are two classes of policy control methods, namely, on-policy and off-policy.
On-policy control uses the policy derived from the current value-function estimate
to update the future estimates. Alternatively, off-policy control uses a policy that is
greedy with respect to the current value-function to estimate future value-functions.
SARSA (on-policy) and Q-learning (off-policy) are two popular TD control algorithms that are known to learn an MDP asymptotically (Sutton and Barto, 1998;
Watkins and Dayan, 1992).
The important concept of why we are able to do model-free TD control lies in the
fact that we use (state,action)-value functions instead of state-value functions.
Definition 12 (Greedy policy control). Policy improvement is done by considering a new
policy, π 0 , which is greedy with respect to the current value-function.
π 0 (s) = arg max Rsa + Pssa 0 V (s0 )
a∈A
π 0 (s) = arg max Q(s, a)
a∈A
(Greedy w.r.t. state-value function)
(Greedy w.r.t. action-value function)
From the above policy improvement equations we can see that in order to be
greedy with respect to the state-value function, we require the model of the MDP. In
contrast, if the policy is greedy with respect to the action-value function, the model
dynamics of the MDP is not needed, and hence, it is model-free. Thus, optimizing
action-value functions to learn the optimal policy is at the heart of all model-free TD
control algorithms.
Challenges and Drawbacks
All the Classical Reinforcement Learning algorithms that we discussed above can be
categorized as tabular algorithms. That is, the algorithms use a table data structure to
associate each state (or state-action pair) with its current value estimate. As the agent
Background and Related Work
10
interacts with the environment and gains experience, the table values are updated
with better estimates of its value.
The main drawback of such a method is that it scales poorly. When the state-space
is very large or continuous, the fundamental requirement that the agent visits each
state (or state-action pair) multiple times (or infinitely often) is not satisfied; states
are at most visited once. An agent following a policy derived from these value estimates would do no better than a random policy. Moreover, the table size grows with
the number of states, making storage infeasible for problems with large/continuous
state-space.
A common approach to solving this problem is to find a way to generalize the
value-function from the limited experience of the agent (Sutton and Barto, 1998).
That is, we want to approximate the value-function for an unseen state (or stateaction pair) from the example values it has observed so far. Function approximation
is a generalization technique that does exactly this; it takes in observed values of a
desired function and attempts to generalize an approximation of the function.
2.1.2
Function Approximation
Function approximation (FA) is an instance of supervised learning (Sutton and Barto,
1998). It is viewed as a class of techniques used to approximate functions by using
example values of the desired function. In the RL context tabular methods become
infeasible in large or continuous state spaces. This challenge is mitigated by employing FA techniques to predict the value-function at unseen states. However, not all FA
methods are applicable to the RL setting. We require a training method which can
learn efficiently from on-line, non-i.i.d. data, and also handle non-stationary target
functions. The following are some of the function approximators that are used in the
RL context.
• Gradient-Descent Methods
– Artificial Neural Networks
– Linear Combination of Features
• State Aggregation
– k-Nearest Neighbors
– Soft Aggregation
State aggregation is a method of generalizing function approximation in which
states are grouped based on a criterion and then value is estimated as an attribute
of the group. When a state is re-visited the value corresponding to the state’s group
gets updated.
Linear combination of features, also known as Linear Function Approximation (LFA),
is essentially a linear mapping from the state space (of dimensionality D) to a feature space of dimension M, where often M < D. Each basis function of the feature
§2.2 Exploration Strategies for Reinforcement Learning
11
space is a mapping from the state space to a real-valued number that represents some
feature of the state-space.
Definition 13 (Linear-Approximate state(action)-value function). The approximate statevalue function of a state s ∈ S under a policy π is given by:
V̂ π (s) = θT φ(s) =
M
∑ θi φi (s)
i =1
M
Q̂π (s, a) = θT φ(s, a) =
∑ θi φi (s, a)
i =1
Where φ : S(×A) → R M , is a feature map, and θ ∈ R M is the parameter vector.
LFA has sound theoretical guarantees and also is very efficient in terms of both
data and computation (Sutton and Barto, 1998), making it a good candidate for the
implementation of our algorithm.
As mentioned previously, FA can be regarded as a technique to develop a generalization regarding value. In order to have a good capacity to generalize, a function
approximator must have relevant data about the state-space. Consider a pathological
case in which the agent does not explore at all: as a result the only data available for
FA would be concentrated in one region of the state space. This results in the estimation of values of unseen states being highly biased. In order to avoid this problem
we have to make sure that the agent visits most regions of the state-space, that is,
the agent has to explore the state-space efficiently. The main goal of this thesis is to
address the problem of how to explore efficiently in large state-spaces.
2.2
Exploration Strategies for Reinforcement Learning
In Section 1.2 we described the exploration/exploitation dilemma, which is a fundamental problem in RL. All exploration strategies attempt to manage the trade-off between
these two often opposed objectives. The simplest and most widely-used exploration
strategy is known as e-greedy. At each time-step t the agent chooses a greedy action
with probability 1 − e and with e probability the agent chooses a completely random
action. To ensure that the policy converges to the optimal policy it has to satisfy the
GLIE assumptions (Singh et al., 2000):
Definition 14 (Greedy in the Limit with Infinite Exploration). A policy is GLIE if it
satisfies the following two assumptions.
• Each action is taken infinitely often in every state that is visited infinitely often,
lim Nt (s, a) = ∞
t→∞
Where Nt (s, a) is the number of times action a has been chosen in state s up-to time-step
t.
Background and Related Work
12
• In the limit, the learning policy is greedy with respect to the Q-value function with
probability 1.
lim πt ( a | s) = 1, when, a = arg max
Qt (s, a0 )
0
t→∞
a ∈A
For example, e-greedy satisfies the GLIE assumptions when e is annealed to zero.
A common way to do this is by setting et ∝ 1/t.
In small, finite MDPs e-greedy satisfies the GLIE assumptions, but when the stateaction space is large/continuous the first GLIE assumption is violated and hence the
convergence guarantee is lost. e-greedy is a naïve approach to solving the exploration
problem, but we still use it in large MDPs because of its low resource requirements
when compared with alternatives (Bellemare et al., 2016). In this thesis we propose a
novel exploration strategy that improves upon e-greedy, and provides state-of-the-art
results in large problems with low computational overhead.
We now provide an exposition of various explorations strategies, their foundational principles, and an analysis of recent breakthroughs in the field of exploration.
2.2.1
Taxonomy of Exploration Strategies
The exploration-exploitation dilemma is still an open problem, but researchers have
made significant inroads into understanding the nature of the problem. Sebastian
Thrun classified exploration techniques into two families of exploration schemes, directed and undirected (Thrun, 1992). Undirected exploration strategies do not use any
information from the environment to make an informed exploratory action; they predominantly rely on randomness to do exploration. Softmax methods and e-greedy
are examples of undirected exploration techniques. The softmax action is sampled
from the Boltzman distribution
Boltzs ( a) =
exp( Q(s, a))
.
exp( Q(s, a0 ))
∑
a0 ∈A
On the other hand, directed exploration strategies use the knowledge about the
learning process to form an exploration-specific heuristic for action selection. This
heuristic directs the agent to take those actions that maximizes the information gain
about the environment. The exploration algorithm introduced in this thesis falls into
the category of directed exploration algorithms. In order to put it into context, we
first present an overview of the existing directed exploration strategies used in the
literature.
2.3 The Optimism in the Face of Uncertainty Principle
In the following chapter we present our directed exploration method, which implements the principle of "Optimism in the Face of Uncertainty" (OFU) as a heuristic
for exploration. In this section we review existing work on the OFU heuristic. The
principle is succinctly captured in Osband and Van Roy (2016):
§2.3 The Optimism in the Face of Uncertainty Principle
13
"When at a state, the agent assigns to each action an optimistically biased
while statistically plausible estimate of future value and selects the action
with the greatest estimate."
OFU is a heuristic to direct exploratory actions. OFU directs the agent to take
actions which have more uncertain value estimates. Instead of greedily taking the
action that has the highest estimated value, that agent is encouraged to take actions
which have a high probability of being optimal. To see that an apparently suboptimal
action may indeed have a high probability of being optimal, let us take an example.
Suppose that the agent has taken an action a ∈ A very often from a particular state
s ∈ S , and suppose that a also currently has the highest value-estimate Q̂π (s, a)
among the available actions. Now consider an alternative action a0 ∈ A that has only
been tried once from the state s, and suppose that the reward received was lower
than Q̂π (s, a). Action a has higher estimated value, but having tried it many times,
the agent’s uncertainty about its value is quite low. In contrast, the uncertainty about
the value of the alternative action Q̂π (s, a0 ) is very high, since it has been taken so
rarely. Thus, while the current estimate Q̂π (s, a0 ) may be lower than Q̂π (s, a), there is
a good chance that the agent was unlucky when taking a0 the first time, and that the
true action-value Qπ (s, a0 ) is much higher than both estimates. Thus it may be that
a0 has a higher probability of being the optimal action than does a, especially if their
estimated values are quite close. The OFU heuristic would bias the agent toward
taking action a0 instead of the greedy action a. An agent following this heuristic will
behave as if it is optimistic about action a0 , or more precisely, about its true actionvalue Qπ (s, a0 ). This optimism drives the agent to explore regions of the environment
about which it is more uncertain.
2.3.1 OFU using Count-Based Exploration Bonuses
Most of the exploration algorithms that enjoy strong theoretical efficiency guarantees,
implement the OFU heuristic. Many do so by augmenting the estimated value of a
state(-action pair) with an exploration bonus that quantifies the uncertainty in that
value estimate. An agent which acts greedily with respect to this augmented value
function will be biased to take actions with higher associated uncertainty. Most of
these algorithms are tabular and count-based in that they compute their exploration
bonuses using a table of state(-action) visit-counts. The visit-count serves as an approximate measure of the uncertainty associated with a state(-action), because more
novel state(-action) pairs will have lower visit-counts. State(-actions) with lower visit
counts are assigned higher exploration bonuses. This drives the agent to behave optimistically and explore less frequently visited regions of the environment, which may
yet prove to have higher value than familiar regions. Moreover, even if those regions
turn out to yield little reward when explored, the agent will have greatly reduced its
uncertainty about those regions. Indeed, the reduction in uncertainty would be much
smaller if the agent were to take an action that had already been tried many times.
The OFU heuristic is therefore a win-win approach for the agent. OFU algorithms are
more efficient than undirected exploration strategies like e-greedy because the agent
Background and Related Work
14
avoids actions that yield neither large rewards nor large reductions in uncertainty
(Osband et al., 2016).
2.3.2
Tabular Count-based Exploration Algorithms
One of the best known OFU methods is the UCB1 bandit algorithm,
which selects
q
2 log t
an action that maximises an upper confidence bound Q̂t ( a) +
, where Q̂t ( a)
N ( a)
is the estimated mean reward and N ( a) is the visit-count (Lai and Robbins, 1985).
The dependence of the bonus term on the inverse square-root of the visit-count is
justified using Chernoff bounds. In the MDP setting, the tabular OFU algorithm most
closely resembling our method is Model-Based Interval Estimation with Exploration
Bonuses (MBIE-EB) (Strehl and Littman, 2008).2 Empirical estimates P̂ and R̂ of
the transition and reward functions are maintained, and R̂(s, a) is augmented with
β
a bonus term √
, where N (s, a) is the state-action visit-count, and β ∈ R is a
N (s,a)
theoretically derived constant. The Bellman optimality equation for the augmented
action-value function is
Q̃π (s, a) = R̂(s, a) + p
β
N (s, a)
+ γ ∑ P̂ (s0 | s, a) max
Q̃π (s0 , a0 )
0
s0
a ∈A
Here the dependence of the bonus on the inverse square-root of the visit-count is
provably optimal (Kolter and Ng, 2009). This equation can be solved using any MDP
solution method.
While tabular OFU algorithms perform well in practice on small MDPs (Strehl
and Littman, 2004), their sample complexity becomes prohibitive for larger problems
(Bellemare et al., 2016). The sample complexity of an algorithm is a bound on the
number of timesteps at which the agent is not taking an e-optimal action with high
probability (Kakade, 2003). Loosely speaking, it measures the amount of experience
the agent must have before one can be confident it is basically performing opti2
|A|
mally. MBIE-EB, for example, has a sample complexity bound of Õ e|S|
3 (1− γ )6 . In the
high-dimensional setting – where the agent cannot hope to visit every state during
training – this bound offers no guarantee that the trained agent will perform well.
The prohibitive complexity of these tabular OFU algorithms is due in part to the fact
that a table of visit-counts is not useful if the state-action space is too large. Since
the agent will only visit a small fraction of that space, the visit-count for most states
will always be zero. These algorithms are therefore unable to usefully compare the
novelty of two unvisited states. All unvisited states have the same visit-count, and
hence the same exploration bonus. The optimistic agent will treat them all as equally
novel and equally appealing.
2 To
the best of our knowledge, the first work to use exploration bonuses in the MDP setting was the
Dyna-Q+ algorithm, in which the bonus is a function of the recency of visits to a state, rather than the
visit-count (Sutton, 1990)
§2.4 Bayes-Adaptive RL
15
2.3.3 Generalized Visit-counts for Exploration in Large MDPs
Tabular OFU algorithms fail on high-dimensional problems because they do not allow for generalization across the state space regarding uncertainty. Every unvisited
state is treated as entirely novel, regardless of any similarity between the unvisited
states and the visited states in the history. In order to explore efficiently in large domains, the agent must be able to make use of the fact that some unvisited states share
many features with visited states, while others share very few. If an unvisited state
has almost exactly the same features as a very frequently visited one, then it should
not be considered to be as uncertain as a state with unfamiliar features. An effective
OFU method for these problems would not just encourage the agent to visit unvisited states, but rather would drive the agent to visit states with novel or uncommon
features. We discuss this issue further in section Section 3.1.1.
Several very recent extensions of count-based exploration methods have achieved
this sort of generalisation regarding uncertainty, and have produced impressive results on high-dimensional RL benchmarks. These algorithms closely resemble MBIEEB, but they substitute the state-action visit-count for a generalised visit-count which
quantifies the similarity of a state to previously visited states. Bellemare et al. (2016)
construct a Context Tree Switching (CTS) density model over the state space such
that higher probability is assigned to states that are more similar to visited states
(Veness et al., 2012). A state pseudocount is then derived from this density. A subsequent extension of this work replaces the CTS density model with a neural network
(Ostrovski et al., 2017). Another recent proposal uses locality sensitive hashing (LSH)
to cluster similar states, and the number of visited states in a cluster serves as a generalised visit-count (Tang et al., 2016). As in the MBIE-EB algorithm, these counts
are used to compute exploration bonuses. These three algorithms outperform random strategies, and are currently the leading exploration methods in large discrete
domains where exploration is hard.
Before presenting our optimistic count-based exploration method in the following
chapter, we now briefly canvas two alternative frameworks for directed exploration,
and discuss their limitations.
2.4
Bayes-Adaptive RL
In the Bayesian approach to model-based reinforcement learning, we maintain a posterior distribution over the possible models of the environment given the experience
of the agent (Dearden et al., 1998). Bayesian inference is used to update the posterior with new information as the agent interacts with the environment, and also to
incorporate the agent’s prior distribution over the transition models.
Since the posterior is maintained over all possible models we can now talk about
the uncertainty pertaining to what is the best action to take. This uncertainty is
modelled as a Markov Decision Process defined over a set of hyper-states. A hyperstate acts as an information state which summarizes the information accumulated so
far. This augmented MDP, often referred to as the Bayes-Adaptive MDP (BAMDP),
16
Background and Related Work
can be solved with standard RL algorithms (Duff, 2002). In this framework an
agent acting greedily in the BAMDP whilst updating the posterior acts optimally
(according to its prior belief) in the original MDP. The Bayes-optimal policy for the
unknown environment is the optimal policy of the BAMDP, thereby providing an
elegant solution to the exploration-exploitation trade-off.
Unfortunately, the cardinality of the hyper-states grows exponentially with the
planning horizon thereby rendering exact solution to the BAMDP computationally
intractable for large problems (Duff, 2002).
2.5 Intrinsic Motivation
The final directed exploration heuristic that we discuss is born out of the so-called
intrinsic motivation framework. There appears to be a growing scientific consensus in
developmental psychology that human beings, from infants to adults, develop their
understanding of the world using certain cognitive systems such as intuitive theories,
social-structures, spatial systems, etc. (Spelke and Kinzler, 2007; Lake et al., 2016).
During curiosity-driven, creative, or risk-taking activities, rational agents use this
understanding to generate intrinsic goals. Accomplishing these intrinsic goals leads
to the accumulation of intrinsic rewards, thereby exhibiting an innate desire to explore,
manipulate, or probe their environment (Oudeyer, 2007).
Drawing parallels to reinforcement learning, the goal of a traditional RL agent
is to maximize its expected cumulative reward. This behaviour is extrinsically motivated since the reward signal is external to an agent. We say that an agent is intrinsically motivated if it has intrinsic goals and rewards. In the context of exploration
for RL, the aim of the intrinsic motivation approach is to use intrinsic reward as a
heuristic that assigns an exploratory value to the agent’s actions. For example, an agent
may receive intrinsic rewards for visiting novel parts of the environment that need
further exploration (Thrun, 1992).
Many formulations that quantify the exploratory value of an action has been put
forth, and most of them augment the environment’s reward function so as to motivate directed exploration. Schmidhuber (2010) proposed a measure for intrinsic
motivation by taking into account the improvement a learning algorithm effected on
its predictive world model. This measure tracks the progress of an agent’s ability to
better compress the history of states and actions (Steunebrink et al., 2013). Another
framework for intrinsically motivated learning is to maximize the mutual information.
An intrinsic reward measure called empowerment is formulated by searching for the
maximal mutual information (Mohamed and Rezende, 2015). The notion of maximizing information gain was demonstrated in a humanoid robot by the introduction
of artificial curiosity (Schmidhuber, 1991) as an intrinsic goal (Frank et al., 2014).
These formulations have some major drawbacks which hinder their suitability as
exploration heuristics. Firstly, they fail to provide any strong theoretical guarantees
of efficient exploration. Leike (2016) pointed out that since none of these heuristics
take into account the reward structure of the problem, they do not distinguish be-
§2.5 Intrinsic Motivation
17
tween regions of high and low expected reward. Secondly, these algorithms require
that we maintain the environment dynamics of the underlying MDP, which prevents
us from easily integrating them with model-free algorithms. Another major drawback is the computational overhead associated with calculating the heuristic. For
problems with large state/action spaces, computing the intrinsic reward becomes intractable for many heuristics (Bellemare et al., 2016). Most problems of interest have
extremely large state spaces, and hence the intrinsic motivation heuristic is currently
impractical as an exploration strategy in these domains.
18
Background and Related Work
Chapter 3
Exploration in Feature Space
‘To wander is to be alive.’
Roman Payne, Europa
In this chapter we introduce a simple, optimistic, count-based exploration strategy that achieves state-of-the-art results on high-dimensional RL benchmarks. In
Section 3.1 we begin by discussing the drawbacks of current exploration strategies.
In Section 3.2 we provide an exposition of the core ideas that underpin our algorithm.
Finally, in Section 3.3 we present our algorithm, as well as a number of related theoretical results.
3.1
Drawbacks of Existing Exploration Methods for Large MDPs
We introduced count-based exploration strategies for large MDPs in section Section 2.3.3. Even though they are the current state-of-the-art exploration algorithms in
these domains, we consider that there are some potential drawbacks to their common
approach to estimating novelty. The motivation for our algorithm arises from trying
to avoid these drawbacks.
3.1.1 Choosing a Novelty Measure
The aforementioned algorithms compute a generalized visit-count. This generalized
count is a novelty measure that quantifies the (dis)similarity of a state to those in
the history. These algorithms drive the agent towards regions of the state space with
high novelty. However, the effectiveness of these novelty measures depends on the
way in which they measure the similarity between states. If this similarity measure
is not chosen in a principled way, states may deemed similar in ways that are not
relevant to the given problem. Let us explore this issue by taking an example.
Example 1 (Confounded novelty). Alice is a foodie. She wants to explore the myriad
restaurants that are open in her city. Suppose that Alice’s novelty measure treats restaurants
as similar if they are geographically close. Alice consults her novelty measure to choose a
restaurant she has not tried yet, and it returns a Chinese restaurant in a distant suburb
19
20
Exploration in Feature Space
that she has not visited before. Alice scratches her head thinking: ‘I have been to a tonne of
Chinese restaurants; if only my novelty measure understood that and suggested a different
cuisine!’ Unfortunately, her novelty measure considers this restaurant very dissimilar from
the Chinese restaurants she has visited, simply because it is geographically distant from them.
The problem here is that Alice’s novelty measure does not know anything about
which features matter when evaluating the novelty of a restaurant. Let us now look at
an example from the recent exploration literature where this problem can be clearly
observed.
Inappropriate Novelty Measures in Practice
The problems that can arise from an unprincipled choice of novelty measure are well
illustrated in the experimental evaluation of Stadie et al. (2015). Their algorithm uses
an autoencoder to encode the state-space into a lower dimensional representation.
The encoding is then fed into a model dynamics prediction neural network which
estimates the novelty by providing an error-based bonus. This method, called Model
Prediction Exploration Bonuses (MP-EB), uses an error based estimator and is different from the visit-density model of Bellemare et al. (2016), but they both estimate
novelty. To generalize regarding value they use the DQN network, and so we refer
to their algorithm as DQN+MP-EB.
Figure 3.1: Q*bert Level 1
Figure 3.2: Q*bert Level 2
Figure 3.3: Two levels of the Atari2600 game Q*bert
During empirical evaluation of their algorithm an anomaly was detected in the
game Q*bert1 from the Arcade Learning Environment2 (ALE) benchmarking suit.
DQN+MP-EB algorithm scored lower than the baseline algorithm, DQN+e-greedy.
They attributed this anomaly to the fact that during each level change of Q*bert, the
color of the game changes dramatically, but neither the objective nor the structure
of the level changes (Figure 3.3). When their agent reached level 2 (Figure 3.2), it
1 In Q*bert, the goal of the agent is to jump on all the cubes without falling off the edge, or being
captured.
2 ALE is a performance evaluation platform consisting of Atari2600 games. It is considered as the
standard performance test bed for RL algorithms. We’ll discuss in depth about ALE in Chapter 4.
§3.1 Drawbacks of Existing Exploration Methods for Large MDPs
21
perceived the state to be completely novel because MP-EB is sensitive to color. This
tricked MP-EB into assigning high exploration bonus to all the states even though
the action-values of the states hadn’t changed. Hence the policy of the agent was
impacted adversely.
The pathology of DQN+MP-EB in the Q*bert game highlights a serious problem
with current novelty estimators they do not take into account the relevancy of a
state to the task an agent is trying to accomplish. We argue that a measure of novelty
should not just be an arbitrary generalized representation of how many times an
agent has visited a state, but should ideally be a measure of dissimilarity in facets
that are relevant to the agent’s goal. Two states can be different in many ways; the
challenge is to find out a similarity metric which is effective in achieving the agent’s
goal optimally. In Example 1, Alice’s novelty measure did not know that suggesting
a restaurant with a different cuisine would be more relevant to her task, thereby
naively suggesting a geographically distant unvisited restaurant.
3.1.2 Separate Generalization Methods for Value and Uncertainty
We contend that this deficiency is not peculiar to MP-EB, but rather that it may arise
whenever the novelty measure is not designed to be task-relevant. Indeed, all of
the aforementioned algorithms which compute a novelty measure share a common
structure which leaves them vulnerable to this problem. Each algorithm has two
quite unrelated components: a value estimator (an RL algorithm which performs
policy evaluation), and a novelty estimator. Each component involves an entirely
separate generalization method. The value estimator makes use of a feature representation of the state space in order to generalize about value. The novelty estimator separately utilizes a different, exploration-specific state space representation to
measure the similarity between states. For example, the #Exploration algorithm of
Tang et al. (2016) uses the DQN algorithm for value estimation. In order to estimate
novelty, however, #Exploration maps the state space into a lower-dimensional representation using locality sensitive hashing. The similarity measure induced by the
choice of hash codes is unlikely to resemble that which is induced by the features
learnt by DQN. The DQN-CTS-EB algorithm of Bellemare et al. (2016) has a similar
structure: DQN is used to estimate value, but the CTS density model is used to estimate novelty. Again, it is not obvious that there should be much in common between
the two similarity measures induced by these different state space representations.
One might think that this is natural; after all, each representation is used for a different purpose. However, there are two questions we can ask here. Firstly, is there
redundant computation due to performing a dimensionality reduction of the same
state-space twice? If so, can we reuse the same state space representation for both
value and novelty estimation? We address these questions in the following section.
Before moving on we should note that the concerns we express in this section
have already been raised in the literature. In their empirical evaluation Bellemare
et al. (2016) observed that their value estimator (DQN) was learning at a much slower
rate than their CTS density model (their novelty measure). The authors attribute this
Exploration in Feature Space
22
mismatch to the incompatibility between novelty and value estimators. They further
go on to suggest that designing density models to be compatible with value function
would be beneficial and a promising research direction.
The drawbacks we presented in this section suggest that there may be much room
for improvement in the design of novelty estimators for exploration. In the following
sections we describe our technique for estimating novelty by factoring in the insights
we gained from analyzing these drawbacks. We first provide a solid footing for some
of the assumptions that we made while designing the algorithm. We then go on to
present our core exploration algorithm, and then combine it with a model-free RL
algorithm (SARSA(λ)). In the coming chapters we present empirical evidence that
our RL algorithm achieves world-leading results on the ALE benchmarking suite.
3.2 Estimating Novelty in Feature Space
3.2.1
Motivation
Which representation of the state space is appropriate for novelty estimation? Intuitively, if we use some parameters to determine the value of a state, then naturally, two
such objects are considered dissimilar only if they differ in these parameters. Analogously, if the agent is using certain features to determine the value of a state, then
naturally, two such states should be considered dissimilar only if they differ in those
value-relevant features. This motivates us to construct a similarity measure that exploits the feature representation that is used for value function approximation. These
features are explicitly designed to be relevant for estimating value. If they were not,
they would not permit a good approximation to the true value function. This sets our
method apart from the approaches described in Section 2.3.3, which measure novelty with respect to a separate, exploration-specific representation of the state space,
one that bears no relation to the value function or the reward structure of the MDP.
We argue that measuring novelty in feature space is a simpler and more principled
approach, and hypothesise that more efficient exploration will result. Our proposal
ensures that generalization regarding novelty is done in the same space as generalization regarding value. Figure 3.4 illustrates the basic structure of our proposed
novelty estimator.
Figure 3.4: Novelty Measure in Feature Space
§3.2 Estimating Novelty in Feature Space
23
Let us make the idea more concrete with our running example.
Example 2 (Value-relevant exploration). After Alice’s disappointing restaurant visit last
time, she tweaked her novelty estimator such that it now generalizes based on value-relevant
features like the type of cuisine, the star rating, and the other features that truly determine
the quality of Alice’s dining experience. When Alice is ready to try something new, she can
rest assured that it’s going to be something novel in a way that is meaningful.
3.2.2 Design Decisions
Our exploration strategy, henceforth known as φ-exploration bonus (φ-EB), can be
thought of as exploration in the feature space. This makes the existence of a feature
map crucial to our strategy. Therefore we require that our algorithm be compatible
with Linear Function Approximation (LFA). Before the advent of neural networks
and subsequently DQN, large RL problems used linear function approximation to
estimate the value of a state. Our decision to use LFA as our value prediction module
has the following desirable benefits:
• Domain Independence: The visit-density models that we have seen so far
(MP-EB, CTS-EB, PixelCNN, etc.) are designed to work with RGB pixel values
from a video input. Though there are many domains that use video input to
train the agent, there are equally many other domains that have nothing to do
with a video input. For example, reinforcement learning is used in the financial
sector to optimize portfolios, asset allocations, and trading systems (Moody
and Saffell, 2001). Therefore developing a visit-density model that is domain
independent is a key challenge. Our φ-EB method estimates the novelty using
the same features that LFA uses to approximate the value function. This allows
our exploration strategy to be compatible with any value-based RL algorithm
that uses LFA.
• Indirect dependence on LFA: LFA is essentially a linear combination of features. The only requirement φ-EB has is the existence of a feature map, which
is implicitly satisfied with LFA. Because of this indirect dependence on LFA,
we hypothesis that it is possible to extend φ-EB to be compatible with valuenetworks that perform representation learning as well (e.g., DQN). Due to resource and time constraints we do not pursue empirical evidence for this claim,
but rather leave this as a possible future extension of our research.
• Single point of change: The best way to assess the performance impact of
changes to a system is to confine the change to a single module and then run
performance tests. Following this principle, we know that SARSA(λ) is a valuebased RL algorithm which uses LFA for value prediction and e-greedy for exploration (Sutton and Barto, 1998). SARSA(λ) has been studied, perfected and
validated through-out the ages. Therefore showcasing the performance gains
achieved by replacing e-greedy with our φ-EB exploration strategy allows for
a sound empirical proof for the efficacy of our algorithm.
Exploration in Feature Space
24
One drawback of using LFA for value prediction is that it requires a set of handcrafted features. This is easily mitigated by choosing the Arcade Learning Environment (ALE) as our evaluation platform (Bellemare et al., 2013), combined with the
Blob-PROST feature set (Liang et al., 2015). Using Blob-PROST as our feature set
has and added advantage. Blob-PROST is designed to mimic the features learned
by DQN, thus making our algorithm comparable with those using DQN for representation learning and value prediction. We’ll discuss in depth about the ALE and
Blob-PROST in Chapter 4.
3.3 The φ-EB Algorithm
The main original contribution of this work is a method for estimating novelty in
feature space. The challenge is to do so without explicitly computing the distance
between each new feature vector and all the feature vectors in the history. That approach quickly becomes infeasible because the cost of computing all these distances
grows with the size of the history. Our method instead constructs a density model
over feature space that assigns higher probability to states that share more features
with more frequently observed states. In order to formally describe our method we
first introduce some notation.
Notation
• φ : S → T ⊆ R M , The feature map used in LFA. Maps the state space into an
M-dimensional feature space, T .
• φt ≡ φ(st ), Feature vector observed at time t, whose ith component is denoted
by φt,i
• φ1:t ≡ (φ1 , . . . , φt ) ∈ T t , Sequence of feature vectors observed after t timesteps.
• φ1:t φ ≡ (φ1 , . . . , φt , φ) ∈ T t+1 , Sequence where φ1:t is followed by φ.
• T ∗ , Set of all finite sequences of feature vectors.
• ρ : T ∗ × T → [0, 1], The sequential density model (SDM) that maps a finite
sequence of feature vectors to a probability distribution.
We will now present the key component of our algorithm that allows us to estimate novelty in feature space.
3.3.1
Feature Visit-Density
Definition 15 (Feature visit-density). The feature visit-density ρt (φ) ≡ ρ(φ ; φ1:t ) at
time t is a probability distribution over the feature space T , representing the probability of
§3.3 The φ-EB Algorithm
25
observing the feature vector φ after observing the sequence φ1:t . It is modelled as a product
of independent factor distributions ρit (φi ) over individual features φi
M
ρt (φ) =
∏ ρit (φi )
i =1
This density model induces a similarity measure on the feature space. Loosely
speaking, feature vectors that share component features are deemed similar. This
enables us to use ρt (φ) as a novelty measure for states, because it represents the
frequency with which features are observed in the history. When confronted with
a new state, we are able to estimate how frequently its component features have
occurred in the history. If φ(s) has more novel component features, ρt (φ) will be
lower. By using a density model we are therefore able to measure novelty in a way
that usefully generalizes the agent’s uncertainty across the state space. To illustrate
this, let us consider an example.
Example 3. Suppose we use a 3-D binary feature map and that after 3 timesteps the history
of observed feature vectors is φ1:3 = (0, 1, 0), (0, 1, 0), (0, 1, 0). Let us estimate the feature
visit densities of two unobserved feature vectors φ0 = (1, 1, 0), and φ00 = (1, 0, 1). Using the
KT estimator for the factor models, we have ρ3 (φ0 ) = ρ13 (1) · ρ23 (1) · ρ33 (0) = 18 · 87 · 78 ≈ 0.1,
and ρ3 (φ00 ) = ρ13 (1) · ρ23 (0) · ρ33 (1) = ( 81 )3 ≈ 0.002. Note that ρ3 (φ0 ) > ρ3 (φ00 ) because
the component features of φ0 are more similar to those in the history. As desired, our novelty
measure generalizes across the state space.
Each factor distribution ρit (φi ) is modelled using a count-based estimator. A naive
option would be to use the empirical estimator which is the ratio of the number of
times a feature has occurred to the total number of time steps. Another class of
count-based estimators are the Dirichlet estimators which enjoy strong theoretical
guarantees (Hutter, 2013). We use the Krichevsky-Trofimov(KT) estimator which is a
Dirichlet-like estimator that is simple, easy to implement, scalable, and data efficient
(Krichevsky and Trofimov, 1981). If Nt (φi ) is the number of times the feature φi has
been observed, then the KT estimator is given by:
ρit (φi ) =
Nt (φi ) +
t+1
1
2
Using independent factor distributions for modelling the probability of each feature component inherently assumes that the features are independently distributed.
This is not always the case, especially in video-input based domains such as the ALE
we have many features that are strongly correlated. This doesn’t mean that we cannot
use fully factorized distributions. One of the early assumptions made by Bellemare
et al. (2016) about the density model is that the states are independently distributed.
This allowed them to factorize the states, and model each factor using a positiondependent CTS3 density model. Moreover, our empirical evaluations show that we
3A
Bayesian variable-order Markov model.
Exploration in Feature Space
26
achieve world leading results in hard exploration games suggesting that independent factored distributions produce good novelty measures. Thus by precedence and
by empirical data the independence assumption on the features is a well-justified
trade-off that makes the computation of novelty fast and data efficient.
3.3.2
The φ-pseudocount
Here we adopt a recently proposed method for computing generalised visit-counts
from density models (Bellemare et al., 2016). By analogy with the pseudocounts presented in that work, we derive two φ-pseudocounts from our feature visit-density.
Both variants presented generalize the same quantity, the state visitation count function Nt (s). The expression given in the following definition is derived in Bellemare
et al. (2016). We emphasize that our approach constitutes a departure from theirs,
because while they derive pseudocounts from a state visit-density model, we do so
using a feature visit-density model.
Definition 16 (φ-pseudocount). Let ρ0t (φ) ≡ ρt (φ ; φ1:t φ)4 be the probability that the
feature visit-density model would assign φ if it was observed one more time. Then the φpseudocount for a state s ∈ S is given by:
φ
N̂t (s) =
ρt (φ(s))(1 − ρ0t (φt (s)))
ρ0t (φ(s)) − ρt (φ(s))
3.3.3 The φ-Exploration Bonus algorithm (φ-EB)
Equipped with all the tools necessary for the construction of an exploration bonus
we now proceed to define the φ-EB algorithm. We provide a high level flow-chart
for the construction of the bonus in Figure 3.5, and the corresponding pseudo-code
in Algorithm 1. Having defined the φ-pseudocount (a generalised visit-count), we
follow traditional count-based exploration algorithms by computing an exploration
bonus that depends on this count. The functional form of the bonus is the same as in
MBIE-EB; we merely replace the empirical state-visit count with our φ-pseudocount.
Definition 17 (φ-exploration bonus). The exploration bonus for a state-action pair (s, a) ∈
S × A at time t is
β
φ
Rt (s, a) = q
φ
N̂t (s)
where β is a hyper-parameter that controls the agents level of optimism.
Loosely speaking, the hyper-parameter β can be viewed as a knob that tunes the
agent’s confidence in its estimate of the true action-value function. Higher values of
β makes the agent under-confident about value, and result in too much exploration.
Very low β values do not encourage enough exploration because the exploration
bonus is too small to dissuade the agent from acting greedily with respect to its
4 Also
called the recoding probability.
§3.3 The φ-EB Algorithm
27
current value estimates. In both scenarios the final policy of the agent is affected
adversely. The goal is to find a β value that gives good results across domains.
We performed a coarse parameter sweep among the games in the ALE evaluation
platform and concluded that β = 0.05 was the best value. Further details regarding
the selection of β value is discussed in Section 5.2.1.
Figure 3.5: Flow Chart for computing the exploration bonus of φ-EB
Exploration in Feature Space
28
Algorithm 1: φ-exploration bonus
Input: Density Model ρ
1 function FeatureVisitDensity( φ )
M
return ∏ ρi (φi )
2
i =1
3
1
2
3
4
5
1
2
3
1
2
3
4
5
6
7
end
Input: Feature Visit Count Nt ; Density Model ρ; Current Timestep t
function UpdateFeatureVisitDensity(φ)
for i=1 to M do
Nt (φi ) + 21
ρi (φi ) ←
t+1
end
end
function PseudoCount( p, p0 )
p (1 − p )
return
p0 − p
end
Input: LFA Feature Map φ; Exploration Coefficient β
function ExplorationBonus(s)
ρ(φ) ← FeatureVisitDensity(φ(s))
UpdateFeatureVisitDensity(φ(s))
ρ0 (φ) ← FeatureVisitDensity(φ(s))
N̂ φ (s) ← PseudoCount(ρ(φ), ρ0 (φ))
β
return q
N̂ φ (s)
end
3.3.4
LFA with φ-EB
One the advantages that we have in developing our algorithm for use with LFA is
that our exploration strategy is compatible with all value-based RL algorithms that
use LFA. As we will see, our empirical performance across a range of environments
suggests that one can plug our exploration strategy with little to no modification
into any of these algorithms and expect considerable gains in exploration efficiency.
In our empirical evaluation we use SARSA(λ) with replacing traces as our valuebased reinforcement learning algorithm. Algorithm 2 presents the pseudo-code for a
generic RL algorithm that uses the augmented reward r + for updating the function
§3.3 The φ-EB Algorithm
29
parameters θ of the approximate action-value function Q̂(s, a) = θT φ(s, a).
Algorithm 2: LFA with φ-EB
Input: LFA Feature Map φ; Training Horizon tend
1 t ←0
2 Initialize arbitrary θt
3 st , at ← initial state, action
4 while t < tend do
5
rt+1 , st+1 ← Act( at )
φ
6
Rt (st , at ) ← ExplorationBonus(st )
φ
7
rt++1 ← rt+1 + Rt (st , at )
8
at+1 ← NextAction(st+1 , θt )
9
θt+1 ←UpdateTheta(rt++1 , φ)
10
t ← t+1
11 end
12 return θtend
The functions NextAction and UpdateTheta are specific to the underlying value-based RL algorithm used, hence left unspecified. Act( at ) performs action at in the environment.
3.3.5 Complexity Analysis
Time Complexity
From Algorithm 1 it is trivial to see that a call to ExplorationBonus has a worstcase time complexity of O( M), where M is the dimension of the feature space. This
suggests that the time needed to compute the novelty of a state is independent of the
dimension of the state-space. Also, more often than not, the dimension of the feature
space is far smaller than that of the state space. Therefore, our algorithms generates
significant savings in computation over other density models whose time-complexity
scales with the number states. In practice, for a binary feature set like Blob-PROST
we process only those features that have been observed before. This is achieved by
maintaining a single prototypical factor density estimator for all previously unseen
features. We’ll discuss the implementation specific details in depth in Chapter 4.
Space Complexity
We look at Algorithm 2 to analyze what objects are needed to be persisted across
iterations so as to facilitate calculation of the exploration bonus. Clearly the factor
density estimators ρi (φi ), and the feature visit count Nt (φi ) are needed to evaluate
and update the feature visit density. Therefore it can be seen that our algorithm has a
worst case space complexity of O( M). Again, because the features in Blob-PROST are
binary valued, the KT estimator can be defined recursively. This allows for updating
the factor density online without the need to maintain a feature visit count Nt (φi ).
We’ll discuss more on this in Chapter 4.
30
Exploration in Feature Space
3.4 Summary
In this chapter we have presented the main contribution of our research. Motivated
by the drawbacks of current state-of-the-art exploration algorithms, we introduced
our novel exploration algorithm called φ-EB. Later, we provided an exposition on
the various components of the algorithm and also analysed its time and space complexity.
Now that we have presented our algorithm, we move on to implementation aspects. The next chapter provides a detailed overview of the evaluation test-bed, the
software architecture, and the implementation challenges faced during the Research
& Development of the algorithm.
Chapter 4
Implementation
’Any A.I. smart enough to pass a Turing
test is smart enough to know to fail it.’
Ian McDonald, River of Gods
This chapter is dedicated to developing a technically correct implementation of
our exploration strategy φ-EB, and its surrounding infrastructure. This allows us to
perform a sound empirical evaluation which is the focus of the next chapter.
In Section 4.1 we present the high-level architecture of the whole system, and how
the various components interact with each other. Later, in Section 4.2, we present the
implementation of our exploration strategy, φ-EB. Throughout the section we also
talk about the design aspects, and optimization’s that went into implementing φ-EB.
4.1 Software Architecture
Our implementation goal is to develop an RL software agent that uses φ-EB as its
exploration strategy. We present the high-level design of the algorithm in Figure 4.1.
The presented diagram is analogous to the Agent-Environment interaction cycle (Figure 1.1), but with more granularity. From an exploration-centric standpoint, we first
provide a concise overview of the components presented in the architecture, and then
an exposition on the implementation details for φ-EB.
31
Implementation
32
Figure 4.1: Agent-Environment interaction framework for SARSA(λ)+φ-EB.1
4.1.1
Modular Overview
Control
We use SARSA(λ) with replacing traces (Sutton and Barto, 1998) as our learning
algorithm. This decision was driven primarily by two factors. First, using the BlobPROST2 feature set meant that we are locked into the framework provided by Liang
et al. (2015). In our case this is in fact desirable. Replacing the exploration module of
1 Boxes with a tiny row and column, on top and left edges respectively, denote objects stored in
RAM. They can persists across cycles and episodes. Dotted arrows with instruction on them denote
operation on such objects.
2 Discussed in Section 5.1.2.
§4.1 Software Architecture
33
an open source, peer-reviewed and published implementation with our own exploration module enhances the credibility of any performance gains that result. Second,
we need a learning algorithm that works well with Linear Function Approximation
(LFA). When coupled with LFA, SARSA(λ) has better convergence guarantees than
Q-learning (Melo et al., 2008). Hence, SARSA(λ) is a suitable value estimation algorithm for our agent.
Exploration Strategy
This component is our φ-EB exploration strategy that was proposed in Chapter 3. We
implement it using the C++113 programming language. C++ offers a significant edge
over other languages in terms of efficiency and greater control over memory management. Due to the high dimensional nature of our problem, we need to extract
as much performance as possible from our code. Therefore implementing the exploration strategy in C++ is critical to the empirical success of our algorithm. Moreover,
a lock-in with the framework provided by Liang et al. (2015) meant that there was
no compelling reason to choose a different programming language. In the coming
sections we provide a detailed look at the design and implementation of φ-EB.
Prediction
We use LFA to generalize the action-value function for unknown state-action pairs.
For further discussion on LFA we refer the reader to Section 2.1.2. LFA uses the BlobPROST feature set from Liang et al. (2015) to approximate the action-value function.
Feature Map
We consider the feature map to be an integral part of the agent. The ability to discern
different features of a state is imperative to generalization regarding value, and by
extension, to exploration. Our agent explores in the feature space, and we want to
use LFA for value prediction. This necessitated the need for an efficient and effective
feature set. The Blob-PROST feature set from Liang et al. (2015) is the best feature set
available to date for the Arcade Learning Environment (ALE) evaluation platform.
More details on Blob-PROST in Section 5.1.2.
Environment
We chose the Arcade Learning Environment as our evaluation platform for the following reasons.
• ALE contains many games which vary in degree of exploration hardness. This
allows us to test the efficacy of our algorithm on a broad spectrum of games
(Bellemare et al., 2016).
3 C++11
is a major revision of C++. This particular version was chosen because it makes several useful
additions to core language libraries.
Implementation
34
• ALE is widely accepted as the standard for testing RL algorithms. The vast
majority of exploration specific research that is published post Bellemare et al.
(2013) has adopted the ALE platform to report empirical results (Mnih et al.,
2013, 2015; Stadie et al., 2015; Osband et al., 2016; Bellemare et al., 2016; Tang
et al., 2016; Ostrovski et al., 2017). Therefore, in order to compare and contrast our results with existing research, it is crucial that we choose ALE as our
evaluation platform.
More details on the Arcade Learning Environment in Section 5.1.1.
4.1.2
Agent-Environment Work-flow
We want to seamlessly integrate φ-EB into the control module. Therefore understanding the nuances of what happens in an agents cycle from an implementation perspective is critical. Figure 4.1 also doubles as a work-flow diagram for our
agent SARSA(λ)+φ-EB. Following the usual agent-environment interaction process
at timestep t, the agent performs an action at on the environment and receives an
extrinsic reward rt+1 . The agent also observes the new state of the environment, st+1 .
Inside the agent, the Blob-PROST feature map consumes the current state st+1 and
returns a feature vector φt+1 . The feature vector is then used by the LFA module
to do value prediction. The φ-EB module uses the stored feature φt to generate the
exploration bonus Rt (st , at )4 . SARSA(λ) updates the parameters of LFA with the TD
update and chooses the next action optimistically.
φ-EB
In the exploration strategy module the feature visit-density ρ is a hash map that
persists in memory across cycles and episodes. Each entry of ρ is a key-value pair
mapping individual features φi to its corresponding factor distribution ρi . In the
φ-EB module shown in Figure 4.1, the flow of control is as follows: compute ρ(φ)
as product of factors, update ρ with the observation φ, then compute ρ(φ) again.
Now calculate the pseudo-count and subsequently the exploration bonus Rt (st ). The
bonus Rt (st ) is considered as an intrinsic reward and is sent to the control module,
SARSA(λ).
LFA
The prediction module (LFA) approximates the next state action-value function using
the parameter vector θt as Q̂π (st+1 , at+1 ) = θTt φ(st+1 , at+1 ). LFA sends the next state
Q-value to the control module, SARSA(λ). The parameter vector θt is also an object
that is saved in memory and persisted across cycles and episodes.
4 Here
we can see that generalization regarding value and novelty are being done in the same space.
§4.2 Implementation Details
35
SARSA(λ)5
All the results from the various modules flow into the control module SARSA(λ).
The control module essentially has two tasks.
• Choose the next action at+1
The next state is chosen by being greedy with respect to next state action-value
obtained from LFA.
i
h
at+1 = arg max Q̂π (st+1 , a)
a∈A
• Update θ of LFA
First we augment the extrinsic reward rt+1 with the intrinsic reward Rt (st )
obtained from φ-EB module.
rt++1 = rt+1 + Rt (st )
The augmented reward rt++1 , the next state action-value Q̂π (st+1 , at+1 ), and the
current state action-value Q̂π (st , at ), both from LFA, is used to calculate the TD
error.
δt+1 = rt++1 + γ Q̂π (st+1 , at+1 ) − Q̂π (st , at )
Where γ is the discount factor. Next we update θ, and is updated using the
usual TD update formula.
θt+1 ← θt + αδt+1
Where α is the learning rate.
Now that we have a clear idea about the surrounding infrastructure, let’s move
on to the implementation details of φ-EB
4.2
Implementation Details
4.2.1 Feature Visit-Density
The central data structure that stores the factor distribution of each individual feature
is an unordered_map6 called fvd_maphφi , ρi i. Each entry is a key-value pair mapping
individual features to its corresponding factor distribution. This allows us to have
constant time look-up for the factor distribution of any feature. At first glance of
the theoretical formulation of feature visit-density (Section 3.3.1, Algorithm 1), the
implementation looks straight forward. Unfortunately that is not the case. We need
to take into account certain implementation specific aspects that are often subsumed
by mathematical formulation. Following are some of the important implementation
details that need to be considered for computing feature visit-density.
5 For
brevity we have left out the discussion on eligibility traces.
a hash map
6 Essentially
Implementation
36
• Sparse Feature Vector
In practice the feature vector φ is the list of features that are active in the current
timestep. Most of the time the set of observed features is in a vastly smaller
subspace of the feature space R M . Therefore, iterating till M to compute the
product of the factor distributions is quite wasteful. In order to overcome this
we maintain a prototype7 function that computes the KT-estimate of observing
the feature give that it has never been observed in t timesteps. Now whenever
a new feature is observed it is added to fvd_map with the current value of the
prototype. If Mt is the total number of features observed till timestep t, then
we can compute the feature visit-density in O( Mt ) time.
• Numerical Stability
Experience has taught us that when dealing with probabilities, innocent looking formulas such as ours can be deceiving. Since we are taking product of
probabilities, they are bound to numerically underflow. In our implementaM
M
tion, rather than computing ∏ ρi (φi ) we compute ∑ log ρi (φi ) . This allows
i =1
i =1
us to safely perform probability calculations without the worry of underflow.
• Inactive Features
During the evaluation of the feature visit-density we need to consider the factor
distributions for the features that are inactive but previously observed. Since
we have already observed φt , we can identify the features in fvd_map that are
not active. The probability density stored in fvd_map against some feature φi ,
is the probability of φi being
φi 6∈ φt , the probability of φi not
active. Assuming
being active is given by 1−fvd_map[φi ] . Therefore when evaluating feature
visit-density for φt we should also factor in the probability of inactive features
not occurring.
Algorithm 3 presents the implementation for computing the feature visit-density
with all the above mentioned optimization/requirements. One key observation is
that we return the log-probability. This is done to facilitate further log based proba-
7 In
this context, a prototype function creates an object of a specified type. Here, a KT estimator
which has seen t zeros.
§4.2 Implementation Details
37
bility computation that occur in other modules.
Algorithm 3: Implementation of Feature Visit Density
Input: Current Timestep t
1 function KT_Prototype()
0.5
2
return
t+1
3 end
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Input: Factor Distribution Map fvd_maphφi , ρi i
function LogFeatureVisitDensity(φ)
sum_log_rho ← 0
for i = 1 to |φ| do
if φi 6∈ fvd_map.keys then
fvd_map[φi ] = KT_Prototype()
end
sum_log_rho ← sum_log_rho + log fvd_map[φi ]
end
// O(1) look-up
/* Inactive features
for i = 1 to size(fvd_map.keys) do
*/
if φi 6∈ φ then
// O(1) look-up with flag trick
sum_log_rho ← sum_log_rho + log 1 − fvd_map[φi ]
end
end
return sum_log_rho
end
4.2.2 Updating Factor Densities
Recall that we use the Krichevsky-Trofimov (KT) estimator to compute the factor
densities. Given a sequence of symbols, the KT-estimator computes the probability
of the next symbol. For a binary symbol-set, the KT-estimator is given by.
Pr ( xt+1 = 1 | x1:t ) =
n1 + 12
n0 + n1 + 1
Where n1 is the number of 1’s seen so far in the sequence, and n0 is the number of
0’s seen so far.
The Blob-PROST feature set is binary valued, making the use of KT-estimators
ideal. Therefore, our factor density for a feature φi being active is given by.
i
ρi (φi ) ≡ Pr (φi = 1 | φ1:t
)=
Nt (φi ) +
t+1
1
2
Implementation
38
And the probability for the feature being inactive is.
ρi (φi = 0) = 1 − ρi (φi = 1)
i is the
Where Nt (φi ) is the number of times feature φi has been seen, and φ1:t
complete sequence of past observations for feature φi .
The factor density equation is neat and simple, but it requires that we maintain
a count for each feature. This is an unnecessary overhead and we can do better. We
now propose an update formula for ρi (φi ) and derive it.
Proposition 1 (Update formula for KT-estimate ρi (φi )). The factor distribution ρit at
timestep t for feature φi can be updated using the following update formula.
ρit+1 (φi )
=
ρit (φi )
t+1
t+2
!
+
φi
t+2
Where φi ∈ {0, 1}
Proof. From the equation for KT-estimates of ρit (φi ) we have,
Nt (φi ) +
ρit (φi ) =
t+1
!
Nt (φi ) +
t+1
=
ρit (φi )
t+2
t+2
1
2
1
2
(1)
In the next timestep t + 1, depending on the value of φi we have two cases.
• Case 1: Feature φi is active, i.e., φi = 1
The KT-estimate ρit+1 (φi ) can be written as,
ρit+1 (φi ) =
=
=
Nt+1 (φi ) + 21
( t + 1) + 1
Nt (φi ) + 1 + 12
t+2
Nt (φi ) + 12 + 1
t+2
Nt (φi ) + 12
1
ρit+1 (φi ) =
+
t+2
t+2
• Case 2: φi = 0
(Since φi = 1)
(2)
§4.2 Implementation Details
39
The KT-estimate ρit+1 (φi ) can be written as,
ρit+1 (φi ) =
=
=
Nt+1 (φi ) + 21
( t + 1) + 1
Nt (φi ) + 0 + 12
t+2
Nt (φi ) + 12 + 0
t+2
Nt (φi ) + 21
0
ρit+1 (φi ) =
+
t+2
t+2
(Since φi = 0)
(3)
In both cases, from Eq. (2) and (3) we can see that the value φi decides the
existence of an additional term. Therefore by observation we can combine the two
cases as follows.
Nt (φi ) + 12
φ
+ i
(4)
ρit+1 (φi ) =
t+2
t+2
Therefore, from Eq. (1) and (4) we get,
t+1
ρit+1 (φi ) = ρit (φi )
t+2
!
+
φi
t+2
Algorithm 4 presents the algorithm for updating the factor distributions. It uses
the update formula presented in Proposition 1 to efficiently update the factor distributions. In the implementation we can see that the update is performed in a two part
manner with linear time complexity, rather than a naive double-loop search.
Algorithm 4: Factor Distribution Update
1
2
3
4
5
6
7
8
Input: Factor Distribution Map fvd_maphφi , ρi i; Current Timestep t
function Update(φ)
for i = 1 to size(fvd_map.keys) do!
fvd_map[φi ] = fvd_map[φi ] ·
end
for i = 1 to |φ| do
fvd_map[φi ] = fvd_map[φi ] +
end
end
t +1
t +2
1
t +2
Implementation
40
4.2.3
Exploration Bonus
This is the entry point for our exploration strategy φ-EB. Due to the modular design
of our algorithm, this function mostly acts like a hub that calls other functions sequentially to get the data required to calculate the exploration bonus. Algorithm 5
presents the implementation to calculate exploration bonus. Note that the probabilities are in log space to avoid numerical stability issues.
Algorithm 5: Exploration Bonus
Input: Exploration Coefficient β
1 function ExplorationBonus( φ )
2
log(ρ) ←LogFeatureVisitDensity(φ)
3
Update(φ)
4
log(ρ0 ) ←LogFeatureVisitDensity(φ)
1
5
N̂ ←
// Pseudo-count
log(ρ0 )−log(ρ)
e
−1
β
// Exploration Bonus
6
return √
N̂
7 end
4.2.4
Action Selection
In the early stages of the project, our agent was facing some inexplicable issues. It
had really slow learning progress, and was getting stuck with a single action for long
periods of time. Fortunately, we had rich logs that helped us in identifying a pattern
to the problem.
We observed that during the initial training cycles, the value predictions from
LFA had very high variance due to lack of enough samples. In cases when there
was an abnormally high Q-value, our greedy optimistic agent always kept taking the
same action over and over again in a loop. We initially thought that, decay in the
corresponding exploration bonus, coupled with increase in optimistic estimates for
other states would lead to the agent breaking out of the loop. Even though eventually
the agent got out of the loop, it happened only after an exorbitantly large number
of episodes. From the logs we observed that each TD update only effected a small
change, and hence the reason why it took a large number of episodes to overcome
abnormally high Q-value.
If we were using e-greedy as the exploration strategy this would not be a problem.
With e-greedy, the agent takes more exploratory action in the initial training cycles.
Even if LFA produces highly varying Q-values initially, the agent doesn’t get stuck
for more that a few cycles. Thus, it can be noted that random exploration at the
beginning helps stabilize the action-values predicted by LFA.
Our goal is to replace e-greedy with our intrinsically motivated exploration strategy φ-EB. Unfortunately, the removal of e-greedy meant that the agent’s policy is
now deterministic and has the above debilitating side-effect. In order to solve this
§4.2 Implementation Details
41
crippling issue we experimented two approaches.
• Combine φ-EB with e-greedy
A similar problem was reported by Bellemare et al. (2016). Their solution was to
use e-greedy, not as an exploration strategy, but as a tool to introduce stochasticity in the agents policy. During the initial training cycles, when there is high
variance from the LFA estimates, taking a purely random action allows the
agent to get out of the greedy action loop. In this experiment e-greedy is implemented in the usual way - with probability e take a random action, and a
greedy optimistic action otherwise. Algorithm 6 presents the implementation.
• Combine φ-EB with Boltzmann distributed action selection
One motivation for our research is to make sure that the agent does not take
purely random actions. The approach from Bellemare et al. (2016) described
above introduces purely random actions. We present an alternate approach
which introduces stochasticity but in a directed manner.
We split our optimistic Q functions into two functions, QE and QI . QE is
trained using the extrinsic reward, whereas QI is trained on the exploration
bonus from φ-EB8 . The motivation here is that we now have a value-function
QI that directs the exploratory actions of the agent. For action selection we
construct the optimistic value function as the summed value function Q =
QE + QI . During action selection, with probability (1 − e) the agent takes the
action that is greedy with respect to Q, otherwise the agent takes a Boltzmann
distributed random action. The Boltzmann distribution is constructed from
the QI values using the discrete_distribution standard library. Hence the
selected random action is more likely to be an action that has higher exploratory
value. Algorithm 7 presents implementation for this approach.
Theoretically, the only difference between the above two approaches is the action
selection process during exploration. The first approach takes a uniformly random
action, whereas the second one takes a Boltzmann-distributed random action. Therefore during the implementation of the learning algorithm we implement the second
approach, and swap the action selection process with the first for experimentation.
8 When
using LFA, training is done on the LFA parameters. Therefore we essentially maintain two
sets of parameters, θE and θI
Implementation
42
Algorithm 6: Action Selection: e-Greedy
1
2
3
function NextAction( Q)
/* Q contains Q-values ∀ a ∈ A for some state.
a = arg max Q( a)
a∈A
/* rand(0,1) generates random number between 0,1
if rand(0,1) < e then
/* randInt(1,x) generates a uniformly random integer
between 1,x
4
5
6
7
8
*/
i = randInt(1,|A|)
return ai
end
return a
end
*/
*/
// random action
// Greedy Optimistic action
Algorithm 7: Action Selection: Boltzmann Distributed
2
function NextActionBoltz( QE , QI )
/* QE , QI contains Q-values ∀ a ∈ A for some state.
a = arg max QE ( a) + QI ( a)
3
/* rand(0,1) generates random number between 0,1
if rand(0,1) < e then
1
a∈A
/* boltzDistInt(W,1,x) generates an integer between 1,x
that is Boltzmann distributed according to W
4
5
6
7
8
*/
i = boltzDistInt(QI , 1, |A|)
return ai
// Boltzmann distributed random action
end
return a
// Greedy Optimistic action
end
*/
*/
§4.2 Implementation Details
43
4.2.5 SARSA(λ)+φ-EB
Now that we have all the modules necessary for learning, we present the implementation for our agent in Algorithm 8.
Algorithm 8: Reinforcement Learning with SARSA(λ) and φ-EB exploration
1
2
3
4
5
6
7
8
9
10
11
12
13
Input: Feature Map φ : S → R M ; Training Horizon tend
t←0
/* Each θ are an |A| × Mt matrix.
*/
Initialize arbitrary θEt , θIt st ← Initial state
QEt ← θEt φ(st )
// Vector containing QE -values ∀ a ∈ A
I
I
Qt ← θt φ(st )
// Vector containing QI -values ∀ a ∈ A
at ← NextActionBoltz( QEt , QIt , st )
while t < tend do
/* Re-estimate Q-values with updated θ values.
QEt ← θEt φ(st )
QIt ← θIt φ(st )
rt+1 , st+1 ← Act( at ) // Perform action in ALE.
φ
Rt ← ExplorationBonus(st )// Compute Intrinsic reward.
/* Predict next state Q-values.
*/
/* Boltzmann distributed action selection.
*/
QEt+1 ← θEt φ(st+1 )
QIt+1 ← θIt φ(st+1 )
at+1 ← NextActionBoltz( QEt+1 , QIt+1 )
/* Alternatively: e-greedy action selection.
// at+1 ← NextAction( QEt+1 + QIt+1 )
/* TD update
14
15
16
17
18
19
20
*/
δtE+1 = rt+1 + γQEt+1 ( at+1 ) − QEt ( at )
φ
δtI+1 = Rt + γQIt+1 ( at+1 ) − QIt ( at )
θEt+1 ← θEt + αδtE+1 I|A|× Mt
θIt+1 ← θIt + αδtI+1 I|A|× Mt
*/
*/
t ← t+1
end
return θEtend , θItend
Mt is the number feature observed til timestep t. We have removed the details regarding eligibility
traces for brevity and clarity.
Algorithm 8 presents the final version of the algorithm that we have implemented,
and for which empirical results are presented. In the algorithm shown we use the
Boltzmann distributed action selection approach. We can disable Line 13 and enable
the two comments below it to use the e-greedy action selection approach.
In the next chapter we discuss the experimental evaluation framework we used
to perform empirical evaluation. Then we showcase the state-of-the-art results that
44
our algorithms enjoys.
Implementation
Chapter 5
Empirical Evaluation
’What can be asserted without evidence
can also be dismissed without evidence.’
Christopher Hitchens
Empirical evidence is one of the fundamental requirements for validating any
scientific hypothesis. In order to validate the efficacy of our exploration algorithm,
this chapter showcases empirical results that represent a significant improvement
over existing algorithms.
In Section 5.1 we talk about the evaluation platform and the feature set that we
used to evaluate our exploration strategy. Section 5.1.1 introduces the Arcade Learning Environment (ALE) as our evaluation platform. We provides justification for
choosing ALE as an environment for our agent. Further, in Section 5.1.2 we introduce the Blob-PROST feature set, and the benefits our agent enjoys from using it.
In Section 5.2 we provide the necessary foundations needed to evaluate Algorithm 8. We discuss the aspects that need to be considered in choosing a particular
game for evaluation. Further, we talk about the parameters for empirical evaluation,
such are number of trial, training frames, etc.
Lastly, in Section 5.3 we discuss the results of our various experiments. We compare the two action selection process discussed in Section 4.2.4 and compare their
empirical performance. Then we compare the learning performance of our agent
with SARSA(λ)+e-greedy. Finally, we compare the evaluation scores for our agent
with other leading algorithms.
5.1
Evaluation Framework
5.1.1 Arcade Learning Algorithm (ALE)
The Arcade Learning Environment (ALE) (Bellemare et al., 2013) is a software framework that interfaces with the Stella emulator (Mott and Team, 1996) for the Atari
2600 games (Montfort and Bogost, 2009). The Atari 2600 platform contains hundreds
of games that vary in many aspects of game-playing such as sports, puzzle, action,
45
46
Empirical Evaluation
adventure, arcade, strategy etc. (Figure 5.1). Some of the games are quite challenging
for human players (Bellemare et al., 2013). Due to the diverse nature of the games,
a learning algorithm that can play the entire gamut of the Atari 2600 games can be
considered to be generally competent. The goal of the ALE framework is to provide
a platform for AI researchers to test their learning algorithm for general competence,
share empirical data with the research community, and further the goal of achieving
artificial general intelligence (Bellemare et al., 2013).
Figure 5.1: Game screens from 55 Atari 2600 games (Defazio and Graepel, 2014).
Figure 5.2: High level working of ALE for an RL algorithm.
§5.1 Evaluation Framework
47
The ALE contains a game-layer to facilitate reinforcement learning. The gamelayer takes in the action from the agent and validates if it is one of the predefined
18 discrete actions. The game-layer sends the validated action to the Stella emulator
which performs the action on the chosen Atari 2600 game. The resulting screen frame
and the RAM state is sent to the game-layer by the emulator. Each screen frame is
a 160 × 210 2D array, with each element representing a 7-bit pixel. The game-layer
analyses the frame and RAM state to identify the game score. It then transform the
game score into an appropriate reward signal expected by an RL agent. Based on the
configuration settings, the game-layer returns the frame array and/or RAM state as
the observation.
After the advent of the ALE, the majority of reinforcement learning publications
have used ALE to provide empirical evidence (Mnih et al., 2013, 2015; Stadie et al.,
2015; Osband et al., 2016; Bellemare et al., 2016; Tang et al., 2016; Ostrovski et al.,
2017). Now, ALE enjoys the status as the standard test bed for testing RL algorithms.
Therefore, in order to compare and contrast our results with existing research, it is
crucial that we choose ALE as our evaluation platform.
5.1.2 Blob-PROST Feature Set
We observed breakthrough performance by DQN to achieve human level performance on majority of ALE games (Mnih et al., 2015). Subsequently Liang et al. (2015)
did a systematic study to analyse the factors that resulted in such a dramatic increase
in performance. As part of this study, they created feature sets that incorporated key
representational biases encoded by DQN. One such data set is called Blob-PROST.
PROST in Blob-PROST stands for Pairwise Relative Offset in Space and Time.
They argue that in most games absolute position of objects are not as important as
their relative distance. Therefore, by taking into account the relative distance between
two objects on screen, they are able to encode information like “there is a green pixel
5 tiles above a blue pixel”. To encode the movements of objects, they take the relative
distance of an object in the current frame to the object five frames in the past.
In an Atari game screen we can assume that there are many blocks of contiguous
pixels with the same color. This is a common continuity assumption that is typically
made in the context of computer vision. Liang et al. (2015) exploits this assumption
and calls such blocks blobs. From a high level, Liang et al. (2015) first pre-process a
frame to find blobs, and then find features based on pairwise relative distances. We
refer the reader to Liang et al. (2015) for a full treatment on the construction of the
feature set.
The Blob-PROST feature set contains a total of 114,702,400 binary features. Even
though there are a large number of potential features, due to the sparsity of blobs,
most of the features are never generated. Also, given a specific game only a relatively
small number of features would be observed. Therefore, using Blob-PROST with
LFA is far more computationally efficient that DQN. Moreover, empirical results
from Liang et al. (2015) suggest that the fixed representations constructed in BlobPROST has the same quality as those learned by DQN.
Empirical Evaluation
48
5.2 Empirical Evaluation
The evaluation methodology was designed to investigate and answer the following
research questions:
• Does the novelty measure generalize state visit-counts when the generalization
is performed in the same space (feature space) as the generalization regarding
value?
• Is there performance improvement over different kinds of environments?
• How does the performance of our algorithm fare when compared to state-ofthe-art Deep RL algorithms?
5.2.1
Evaluation Methodology
We evaluate Algorithm 8 in the ALE using the Blob-PROST feature set. We evaluate
our algorithm on a subset of the games that are most relevant to the problem of
efficient directed exploration. An important factor that determines our selection is
the time and computational resources required to perform the evaluation. Finally,
we perform hyperparameter sweeps to appropriately tune our algorithm.
Figure 5.3: Five games in which exploration is both difficult and crucial to performance. From top left: Montezuma’s Revenge, Venture, Freeway, Frostbite, and
Q*Bert.
Choosing Evaluation Games
ALE contains many games which vary in the degree to which exploration is difficult (Bellemare et al., 2016). At the lower end of the difficulty spectrum are games
for which undirected exploration (e-greedy) is sufficient to learn a good policy (Mnih
et al., 2015). We evaluate our algorithm on hard games, i.e., games where e-greedy
§5.2 Empirical Evaluation
49
fails to improve substantially on a random policy. Hence we focus our evaluation
on games that are classified as hard in the taxonomy provided by Bellemare et al.
(2016). In their taxonomy they further split hard exploration games into games that
have sparse and dense rewards.
In dense-reward games, our RL agent can expect rewards, on average, every few
cycles. Dense-reward games are easier for an RL agent because policy iteration techniques such as SARSA generally require regular feedback on the agent’s policy to
learn; this is known as the issue of temporal credit assignment (Sutton and Barto,
1998).
On the other hand, sparse-reward games only dispense rewards infrequently. In
this setting, the agent must often perform long sequences of actions in the correct
order in order to receive a reward1 . Without a good exploration strategy, stumbling
upon a productive sequence of action is very challenging. Therefore, our main focus
will be on sparse-reward games in which exploration is difficult.
We compare our performance with that of state-of-the-art Deep RL algorithms
reported in the literature, most of which are variants of DQN (van Hasselt et al.,
2016). Recall that we use LFA for value prediction, while DQN uses neural networks. As discussed previously, the Blob-PROST feature set that we selected for LFA
has the property that it closely models the representations learned by DQN (Liang
et al., 2015). Since the Blob-PROST feature set has this property, it is appropriate
to compare our φ-EB algorithm to the other algorithms that use DQN. Now from
the empirical data presented in Liang et al. (2015), we look for games that perform
similar to DQN, and meet the exploration difficult criterion outlined above.
With these in mind, we choose the following five hard exploration games to evaluate our exploration strategy.
• Sparse Reward Games (Figure 5.3, Top Layer)
– Montezuma’s Revenge
– Venture
– Freeway
• Dense Reward Games (Figure 5.3, Bottom Layer)
– Frostbite
– Q*bert
Computational Roadblocks
The computational requirement for games in the ALE are very demanding, especially
in the absence of graphical processing units (GPUs), which excel at the dense linear
algebra computations common in vision-related tasks (Liang et al., 2015). Given the
high-dimensional nature of the problem, agents must be trained for several days on
1 Classic
examples of such games are the board games Chess and Go; rewards are only dispensed
once, at the very end of the game.
Empirical Evaluation
50
end to obtain satisfactory performance. Frostbite and Q*bert were especially computationally intensive, and we had to train for several weeks to obtain sufficient data.
Given our limited time and computational resources, we had to place the following
constraints on our evaluation.
• In Section 4.2.1 we remarked that due to sparse nature of the feature vector our
algorithm runs in O( Mt ) where Mt is the number of unique features observed
till time t. This means that different games run at different speeds. We trained
our algorithm on all games for 100 million frames except for Q*bert which was
trained only for 80 million frames.
• Our main focus is to showcase performance gains in sparse reward hard exploration games. Therefore we focused bulk of our computational resources into
running multiple trial for Montezuma’s Revenge and Venture. This is where
our algorithm leads other state-of-the-art exploration strategies.
Tuning the hyper-parameter β
We performed independent evaluation of all the chosen games for discrete values of
β in the range [0.0001, 0.5]. Recall that β is a parameter that controls the magnitude
of the exploration bonus. We observed that β = 0.05 is the best performing value
for all the games except for Freeway. Recall that because of the nature of the game,
there is a large number of unique Blob-PROST features active. If β is high enough,
the chicken just remains stationary and receives novelty rewards for observing all the
changes in the traffic. When we set β = 0.035, our agent performed much better and
delivered comparable results with the baseline algorithm.
Training Methodology
Training and evaluation of learning algorithms in high-dimensional spaces is computationally demanding. Due to the constantly evolving nature of the field, there
is no general consensus on how many cycles is required to train an agent. Some of
the major exploration algorithms published recently report training till 200 million
frames. Due to limited time and computational resources, this amount of training is
not feasible for us.
We perform empirical evaluation of Montezuma’s Revenge and Venture for five
independent trials, and two independent trials for Frostbite, Freeway, and Q*bert.
With the exception of Q*bert, all agents are trained for 100 Million frames, and then
evaluated for 500 episodes. The result for Q*bert is reported after training for 80
million frames with subsequent 500 episodes of training.
We use the average score per episode, which is a common metric used to report
scores (Bellemare et al., 2013; Mnih et al., 2015; Bellemare et al., 2016).
§5.2 Empirical Evaluation
51
5.2.2 Sparse Reward Games
Montezuma’s Revenge
Montezuma’s Revenge is widely regarded as one of the most difficult games in the
Atari 2600 suite. Learning algorithms typically suffer here due to the problem of
long term credit assignment and sparse rewards. For example, DQN with e-greedy
exploration achieves a score of 0 after training for 200 million frames (Mnih et al.,
2015). In order to get the very first reward of the game the agent must climb down
a ladder, jump onto the pole, jump onto a raised platform, climb down a ladder,
walk left and jump to avoid an enemy, climb another ladder, and finally obtain a
golden key. This long sequence of complex actions is required to simply achieve the
first reward in the first of 24 rooms, arranged in a pyramid structure (Figure 5.4).
It is evident from the complexity of the game that a random exploration strategy
will fail miserably. The challenges posed by this game make the game central to our
evaluation.
Figure 5.4: Montezuma’s Revenge: Rooms visited by undirected exploration
(DQN+e-greedy, above) vs. directed exploration (DQN+CTS-EB, below) (Bellemare
et al., 2016).
Our agent showed steady learning progress and was relatively quick to learn
the initial sequence of actions needed to get the golden key. Analysis from the logs
showed increasing novelty measure associated with novel events. For example, one
of the rooms had a dead-end for one side, but there was a reward there. When our
agent reached the dead-end and received the reward once, it kept getting stuck in
that region for a few episodes. During the episodes where the agent was stuck, we
could observe the novelty values of the other end slowly rising. After a few episodes
the agent started going in the opposite direction based on the novelty rewards it was
Empirical Evaluation
52
getting, and proceeded to explore further.
In Montezuma’s revenge, the total number of rooms visited is a good measure
of exploration efficiency (Bellemare et al., 2016). Prior to the work of Bellemare
et al. (2016) no agent had visited more that 3 rooms without domain specific tailoring. Bellemare et al. (2016)’s agent visited 15 room while our agent visited 14 rooms
in total (Figure 5.4). Both Bellemare et al. (2016)’s and our agent enjoy a peak score
of 6600, which is the highest reported score.
Venture
Venture is another hard, sparse reward exploration game with complex visual representations. Figure 5.7 represents the two different visual states of venture. In
Figure 5.5 the agent is depicted by the small tiny pink pixel towards the bottom
middle of the screen. When the agent is in the outer level it is powerless. The only
way to stay alive is to navigate the maze and/or entering one of the rooms to avoid
the green goblins. Once the agent enters the room, the screen changes to Figure 5.6.
Now our agent is transformed into a smiley face with the ability to fire a projectile
weapon. Here the agent is chased by evil blue robots which can be destroyed by the
agent’s weapon. The small tiny cup-like structure in the bottom left corner of the
room is the goal and the first reward. Destroying the blue monsters does not result
in any reward. The game presents some intricate dynamics, with a large state-space
to explore, and so is difficult for a naive learning algorithm. Therefore Venture is a
must-have game in our evaluation roster.
Figure 5.5: Venture Outer Level
Figure 5.6: Venture Inner Level
Figure 5.7: Two visual states of venture (Atari 2600)
The agent initially moves around the black region to the bottom of the screen
where it starts out (Figure 5.5). The novelty of the agent with empty black space
around it quickly reduces. We observed that the agent starts to hug the outer walls
of the rooms. The room walls together with the agents are novel features. Then the
agent moves along the room wall towards the entry of the room. When the agent
enters the room the screen now transforms as shown in Figure 5.6. We later observe
§5.2 Empirical Evaluation
53
that, with subsequent visits to the lower room our agent learned to shoot and kill
the blue robots and get the reward. Our agent achieves substantial improvement
over Bellemare et al. (2016). Even though they do not report their evaluation score
for Venture, Ostrovski et al. (2017) reports the score of DQN+CTS-EB as 82.2, whereas
our evaluation score is 1169.2.
Freeway
Figure 5.8: Freeway (Atari2600)
Here the agent is a chicken that must cross a busy highway (Figure 5.3 Top layer,
third game). The agent receives the reward only when it reaches the other end of the
road. Being hit by a car sets you back some positions down the road. Even though
this is not a complex game, it is still a hard exploration game. In freeway, cars are
always moving from left to right in every frame. Therefore, there is a large number
of active unique Blob-PROST features.
When tuning the hyper-parameter β we talked about the large number of unique
Blob-PROST features created due to the constant movement of traffic. These changing unique features constantly floods the agent with high novelty rewards. Therefore
the agent is willing to just stand and observe the traffic. When we reduced the exploration coefficient β, the agent’s extrinsic reward is no longer overwhelmed by the
novelty reward. Also, due to the initial optimism, the agent is encouraged to move
upwards. Once the agent manages to cross the road once, it reinforces the up action
and there the agent starts learning faster.
Empirical Evaluation
54
5.2.3
Dense Reward Games
Q*bert
Figure 5.9: Qbert (Atari2600)
In Q*Bert, the agent stands on a pyramid of cubes (Figure 5.9). The goal of the
agent is to jump on all the cubes without falling off the edge or being captured by
an adversary. When the agent has highlighted all the cubes by jumping on them at
least once, the level is cleared. The game has multiple levels. On higher levels, the
task still remain the same, but the enemies become smarter, making it increasingly
difficult to accomplish the task while avoiding capture. As discussed in Section 3.1.1
the only other difference between levels is the choice of color scheme.
Frostbite
Figure 5.10: Frostbite (Atari2600)
In Frostbite, the agent is an Eskimo jumping up and down on an ice platform which
magically results in the building of an igloo (Figure 5.10). Each time the agent jumps
on a pure white ice platform, the agent receives a reward. The agents has to consistently do this by avoiding obstacles/dangers, and picking up bonus points. Once the
igloo has been built the agent need to perform a level-end move by entering into the
igloo. Since there are multiple ways of maximizing the score, and the existence of a
§5.3 Results
55
level-end move makes it a hard exploration game. Our main aim with this game is
to confirm that our exploration strategy improves upon the baseline performance.
5.3
Results
From here on we denote our agent SARSA(λ)+φ-EB as SARSA-φ-EB and SARSA(λ)+egreedy as SARSA-e for notational brevity.
5.3.1 Boltzmann vs. e-greedy Action Selection
Figure 5.11: Average training score for Action Selection using Boltzmann vs. egreedy. Shaded regions describe one standard deviation. Dashed lines represent
min/max scores.
Recall that in Section 4.2.4 we discussed the problem of action selection and presented
two methods that introduce stochasticity into the agent’s policy. Here we compare
two agents differing only in the action selection functionality. SARSA-Boltz-φ-EB introduces stochasticity via a Boltzmann-distributed random action, whereas SARSAφ-EB does via e-greedy.
Figure 5.11 shows the learning curve for the two agent evaluated for the games
Montezuma’s Revenge and Venture. From the plots we can observe that during the
initial training period Boltzmann-distributed action selection has an advantage over
epsilon-greedy. In the long run both have essentially the same average score. This
behavior is expected since e is annealed with training cycles, and hence with time
the agent takes fewer purely exploratory actions.
What we hoped to see in this experiment is for the Boltzmann distributed action
selection to send the agent into a steeper learning trajectory. Unfortunately, this experiment concludes that there is no long term gain to having a Boltzmann-distributed
action selection process.
Empirical Evaluation
56
5.3.2
Comparison with e-greedy
Overview
Figure 5.12 shows the comparison between learning curves for our agent SARSAφ-EB, and the benchmark implementation SARSA-e. The plots clearly illustrates
that SARSA-φ-EB significantly outperforms SARSA-e on both Montezuma’s Revenge
and Venture; two of hardest exploration games in the ALE. In Q*bert and Frostbite
SARSA-φ-EB consistently outperforms SARSA-e, but not by a huge margin. Frequent
rewards from these games give a constant feedback to SARSA-e, helping it to chart
a positive learning path. With β = 0.05 freeway fails to obtain any score, but with
β = 0.035 SARSA-φ-EB achieved a marginally better performance that the baseline
SARSA-e.
Figure 5.12: Average training score for SARSA-φ-EB vs. SARSA-e. Shaded regions
describe one standard deviation. Dashed lines represent min/max scores. (Martin
et al., 2017)
§5.3 Results
57
Montezuma’s Revenge
After an average of 20 Million frames the policy of SARSA-e converges. It learns the
policy to consistently get the golden key, buy rarely leaves the room. Even when
SARSA-e leaves the room it is always to the left room and gets immediately killed by
the laser beams. SARSA-e never learns how to get through the laser beams. SARSAφ-EB performs remarkably well, visiting 14 rooms in total and observing a peak
score of 6600 around the 40 Million frame mark. SARSA-φ-EB learns to go through
laser doors by timing the movements perfectly, avoid dead-ends, duck and jump over
skulls, etc.
Venture
SARSA-e fails to score any points while SARSA-φ-EB performs exceptionally well in
Venture. It is the most impressive performance gain in this evaluation. SARSA-φ-EB
can quickly identify that spending time in the black region is not novel. It is attracted
(because of novelty) to the walls of the room quickly. The agent then moves along
the wall until it sees a nearby opposite wall or the room entrance. Once the agent
starts entering a room consistently, it quickly learns how to attack blue robots and
obtains the reward. The positive learning trend suggest that even higher scores are
possible with more training.
Freeway
With β = 0.05 our agent SARSA-φ-EB fails to cross the road for reasons previously
mentioned in Section 5.2 and Section 5.2.2. But with β = 0.035, SARSA-φ-EB explores
quickly initially and received the reward by crossing the road for the first time. Due
to the undirected nature of exploration, SARSA-e has to first random walk across the
road to receive a reward. Once both the algorithms receive a reward they perform
similarly, with SARSA-φ-EB doing marginally better.
Recall that the components of SARSA-φ-EB were designed such that there is only
a single point of change from SARSA-e. The only difference between SARSA-φ-EB
and SARSA-e is our exploration module, φ-EB. Hence we can conclude that the
empirical performance gains is solely due to the introduction of our exploration algorithm. Further we can conclude that φ − EB is significantly better that e-greedy in
sparse reward hard exploration domains, and consistently outperforms e-greedy in
all domains.
5.3.3
Comparison with Leading Algorithms
Table 5.1 shows the evaluation score for the final policy learnt by our agent. In order
to compare and contrast the efficacy of our algorithm, Table 5.1 also presents the
evaluation scores reported by leading RL algorithms.
Empirical Evaluation
58
Venture
Montezuma
Freeway
Frostbite
Q*bert
Sarsa-φ-EB
1169.2
2745.4
0.0
2770.1
4111.8
Sarsa-e
0.0
399.5
29.9
1394.3
3895.3
DQN+CTS-EB
N/A
3459
N/A
N/A
N/A
A3C+
0
142
27
507
15805
MP-EB
N/A
0
12
380
N/A
DDQN
98
0
33
1683
15088
Dueling
497
0
0
4672
19220
DQN-PA
1172
0
33
3469
5237
Gorila
1245
4
12
605
10816
TRPO
121
0
16
2869
7733
TRPO-Hash
445
75
34
5214
N/A
Table 5.1: Average evaluation score for leading algorithms. Sarsa-φ-EB and Sarsa-e
were evaluated after 100M training frames on all games except Q*bert, for which
they trained for 80M frames. All other algorithms were evaluated after 200M
frames. (Martin et al., 2017)
Our agent reports an average evaluation score of 2745.4 on Montezuma’s Revenge. This score is the second-highest reported score for Montezuma, the leading
score reported for DQN+CTS-EB by Bellemare et al. (2016). Algorithms such as MPEB (Stadie et al., 2015), TRPO-Hash (Tang et al., 2016), A3C+ (Bellemare et al., 2016)
does not score more than 200 points in Montezuma’s Revenge, despite the presence
of an exploration strategy module.
SARSA-φ-EB evaluated on Venture reported a score of 1169.2, the third-highest
reported score. Again it outperforms all the other exploration algorithms. Bellemare
et al. (2016) does not report their score for Venture. A recent evaluation of DQN+CTSEB done by Ostrovski et al. (2017) reported that DQN+CTS-EB obtained a score of
82.2 on venture.
It is known that non-linear algorithms perform much better on Q*bert therefore
our results is not surprising (Wang et al., 2015). Frostbite is notoriously slow to train
because of the large number of features. Given the dense reward nature of the game
and the small training time, we achieve competitive results.
Our results for Montezuma and Venture are extremely competitive and can be considered state-of-the-art. The key observation from Table 5.1 is this, if we look at the
score of Montezuma and Venture together, ours is the only algorithm that achieve
state-of-the-art results in both the games.
Chapter 6
Conclusion and Future Work
In this thesis we have presented the φ-Exploration Bonus (φ-EB) method, a novel
approach to perform directed exploration in large problem domains. The algorithm
is simple to implement, and is compatible with any value-based RL algorithm that
uses Linear Function Approximation (LFA) to predict value. Our method also enjoys lower computational requirements when compared to other leading exploration
strategies. Our empirical evaluation demonstrates that measuring novelty in feature
space is a simple and effective way to drive efficient exploration on MDPs of practical
interest. It also lends support to our hypothesis that defining a novelty measure in
feature space is a principled way to generalize state-visit counts to large problems.
In contrast to other approaches, measuring novelty in feature space avoids building
an exploration-specific state-representation. Instead, our method exploits the taskrelevant features that are already being used for value estimation.
There are myriad ways in which this work could be extended, and the problem
of efficient exploration in large MDPs is still wide open. A promising direction for
future research would be a rigorous empirical comparison of the various generalized
count-based algorithms which we have discussed. At present, many of the reported
results are not helpful in deciding which is the better approach to measuring novelty or generalizing visit-counts. Different exploration algorithms are presented in
conjunction with totally different value-estimation methods, and it can be difficult
to discern whether or not the exploration method used is responsible for the quality of the results. More theoretical understanding of the problem of exploration in
the high-dimensional setting is also sorely needed, and we hope to build upon the
results presented here in future work.
59
60
Conclusion and Future Work
Bibliography
Barto, A. and Dietterich, T., 2004. Reinforcement learning and its relationship to
supervised learning. Handbook of Learning and Approximate Dynamic Programming,
(2004), 47–64. (cited on page 1)
Bellemare, M. G.; Naddaf, Y.; Veness, J.; and Bowling, M., 2013. The Arcade
Learning Environment: An Evaluation Platform for General Agents. Journal of
Artificial Intelligence Research, 47 (7 2013), 253–279. doi:10.1613/jair.3912. (cited on
pages 24, 34, 45, 46, and 50)
Bellemare, M. G.; Srinivasan, S.; Ostrovski, G.; Schaul, T.; Saxton, D.; and
Munos, R., 2016. Unifying Count-Based Exploration and Intrinsic Motivation.
CoRR, abs/1606.0 (6 2016), 1–26. (cited on pages xiii, 3, 12, 14, 15, 17, 20, 21, 25,
26, 33, 34, 41, 47, 48, 49, 50, 51, 52, 53, and 58)
Bertsekas, D. P. and Tsitsiklis, J. N., 1996. Neuro-Dynamic Programming, vol. 5.
Athena Scientific. ISBN 1886529108. doi:10.1109/MCSE.1998.683749. (cited on
pages 7 and 9)
Bishop, C. M., 2007. Pattern Recognition and Machine Learning. Journal of Electronic
Imaging, 16, 4 (1 2007), 049901. doi:10.1117/1.2819119. (cited on page 1)
Dearden, R.; Friedman, N.; and Russell, S., 1998. Bayesian Q-learning. Proceedings
of the fifteenth national/tenth conference on Artificial intelligence/Innovative applications
of artificial intelligence, (1998), 761–768. (cited on page 15)
Defazio, A. and Graepel, T., 2014. A Comparison of learning algorithms on the
Arcade Learning Environment. CoRR, abs/1410.8 (10 2014). (cited on page 46)
Duff, M., 2002. Optimal Learning: Computational procedures for Bayes-adaptive Markov
decision processes. Ph.D. thesis, University of Massachusetts at Amherst. (cited on
page 16)
Frank, M.; Leitner, J.; Stollenga, M.; Förster, A.; and Schmidhuber, J., 2014.
Curiosity driven reinforcement learning for motion planning on humanoids. Frontiers in Neurorobotics, 7, JAN (2014), 1–15. doi:10.3389/fnbot.2013.00025. (cited on
page 16)
Hutter, M., 2005. Universal Artificial Intellegence. Texts in Theoretical Computer
Science An EATCS Series. Springer Berlin Heidelberg, Berlin, Heidelberg. ISBN
978-3-540-22139-5. doi:10.1007/b138233. (cited on page 5)
61
62
BIBLIOGRAPHY
Hutter, M., 2013. Sparse Adaptive Dirichlet-Multinomial-like Processes. Journal of
Machine Learning Research, 30 (5 2013). (cited on page 25)
Kakade, S. M., 2003. On the Sample Complexity of Reinforcement Learning. In
International Conference on Machine Learning, 133. doi:10.1.1.164.7844. (cited on
page 14)
Kolter, J. Z. and Ng, A. Y., 2009. Near-Bayesian exploration in polynomial time.
In Proceedings of the 26th Annual International Conference on Machine Learning - ICML
’09, 1–8. ACM Press, New York, New York, USA. doi:10.1145/1553374.1553441.
(cited on page 14)
Krichevsky, R. E. and Trofimov, V. K., 1981. The Performance of Universal Encoding. IEEE Transactions on Information Theory, 27, 2 (1981), 199–207. doi:
10.1109/TIT.1981.1056331. (cited on page 25)
Lai, T. and Robbins, H., 1985. Asymptotically efficient adaptive allocation rules. Advances in Applied Mathematics, 6, 1 (3 1985), 4–22. doi:10.1016/0196-8858(85)90002-8.
(cited on page 14)
Lake, B. M.; Ullman, T. D.; Tenenbaum, J. B.; and Gershman, S. J., 2016. Building
Machines That Learn and Think Like People. Behavioral and Brain Sciences, (11
2016), 1–101. doi:10.1017/S0140525X16001837. (cited on page 16)
Leike, J., 2016. Exploration Potential. CoRR, abs/1609.0 (9 2016), 1–9. (cited on page
16)
Liang, Y.; Machado, M. C.; Talvitie, E.; and Bowling, M., 2015. State of the Art
Control of Atari Games Using Shallow Reinforcement Learning. Proceedings of the
2016 International Conference on Autonomous Agents & Multiagent Systems, (12 2015),
485–493. (cited on pages 24, 32, 33, 47, and 49)
Martin, J.; Narayanan S, S.; Everitt, T.; and Hutter, M., 2017. Count-Based Exploration in Feature Space for Reinforcement Learning. In Proceedings of the TwentySixth International Joint Conference on Artificial Intelligence. AAAI Press. (cited on
pages iii, xiii, xv, 56, and 58)
Melo, F. S.; Meyn, S. P.; and Ribeiro, M. I., 2008. An analysis of reinforcement
learning with function approximation. Proceedings of the 25th international conference
on Machine learning, (2008), 664–671. doi:10.1145/1390156.1390240. (cited on page
33)
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra,
D.; and Riedmiller, M., 2013. Playing Atari with Deep Reinforcement Learning.
arXiv preprint arXiv:1312.5602, (12 2013). (cited on pages 34 and 47)
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.; Bellemare, M. G.;
Graves, A.; Riedmiller, M.; Fidjeland, A. K.; Ostrovski, G.; Petersen, S.; Beattie, C.; Sadik, A.; Antonoglou, I.; King, H.; Kumaran, D.; Wierstra, D.; Legg,
BIBLIOGRAPHY
63
S.; and Hassabis, D., 2015. Human-level control through deep reinforcement
learning. Nature, 518, 7540 (2015), 529–533. doi:10.1038/nature14236. (cited on
pages 34, 47, 48, 50, and 51)
Mohamed, S. and Rezende, D. J., 2015. Variational Information Maximisation for
Intrinsically Motivated Reinforcement Learning. ArXiv e-prints, (9 2015). (cited on
page 16)
Montfort, N. and Bogost, I., 2009. Racing the beam: the Atari Video computer
system. Platform studies, (2009), xii, 180 p. doi:10.1162/leon.2010.43.2.188. (cited on
page 45)
Moody, J. and Saffell, M., 2001. Learning to trade via direct reinforcement. IEEE
Transactions on Neural Networks, 12, 4 (2001), 875–889. doi:10.1109/72.935097. (cited
on page 23)
Mott, B. W. and Team, S., 1996. Stella: a multiplatform Atari 2600 VCS emulator.
(cited on page 45)
Osband, I.; Blundell, C.; Pritzel, A.; and Van Roy, B., 2016. Deep Exploration
via Bootstrapped DQN. Advances In Neural Information Processing Systems, (2 2016).
(cited on pages 14, 34, and 47)
Osband, I. and Van Roy, B., 2016. Why is Posterior Sampling Better than Optimism
for Reinforcement Learning. arXiv preprint, (7 2016). (cited on page 12)
Ostrovski, G.; Bellemare, M. G.; Oord, A. v. d.; and Munos, R., 2017. Count-Based
Exploration with Neural Density Models. arXiv preprint, (3 2017). (cited on pages
15, 34, 47, 53, and 58)
Oudeyer, P.-Y., 2007. What is intrinsic motivation? A typology of computational
approaches. Frontiers in Neurorobotics, 1, NOV (2007), 1–14. doi:10.3389/neuro.12.
006.2007. (cited on page 16)
Puterman, M., 1994. Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley Series in Probability and Statistics. John Wiley & Sons, Inc., Hoboken,
NJ, USA, 1 edn. ISBN 9780470316887. doi:10.1002/9780470316887. (cited on page
5)
Schmidhuber, J., 1991. A Possibility for Implementing Curiosity and Boredom in
Model-Building Neural Controllers. In From Animals to Animats:Proceedings of the
First International Conference on Simulation of Adaptive Behavior, 222–227. MIT Press,
Paris, France. ISBN 0-262-63138-5. (cited on page 16)
Schmidhuber, J., 2010. Formal Theory of Creativity, Fun, and Intrinsic Motivation.
IEEE Transactions on Autonomous Mental Development, 2, 3 (9 2010), 230–247. doi:
10.1109/TAMD.2010.2056368. (cited on page 16)
64
BIBLIOGRAPHY
Singh, S.; Jaakkola, T.; Littman, M. L.; Szepes, C.; and Hu, A. S., 2000. Convergence Results for Single-Step On-Policy Reinforcement-Learning Algorithms.
Machine Learning, 39, 1998 (2000), 287–308. (cited on page 11)
Spelke, E. S. and Kinzler, K. D., 2007. Core knowledge. Developmental Science, 10, 1
(1 2007), 89–96. doi:10.1111/j.1467-7687.2007.00569.x. (cited on page 16)
Stadie, B. C.; Levine, S.; and Abbeel, P., 2015. Incentivizing Exploration In Reinforcement Learning With Deep Predictive Models. arXiv, (2015), 1–11. (cited on
pages 20, 34, 47, and 58)
Steunebrink, B. R.; Koutník, J.; Thórisson, K. R.; Nivel, E.; and Schmidhuber,
J., 2013. Resource-Bounded Machines are Motivated to be Effective, Efficient, and
Curious. In Artificial General Intelligence: 6th International Conference, AGI 2013, Beijing, China, July 31 – August 3, 2013 Proceedings, vol. 7999 LNAI, 119–129. Springer
Berlin Heidelberg. ISBN 978-3-642-39521-5. doi:10.1007/978-3-642-39521-5{_}13.
(cited on page 16)
Strehl, A. L. and Littman, M. L., 2004. An empirical evaluation of interval estimation for Markov decision processes. In Proceedings - International Conference on
Tools with Artificial Intelligence, ICTAI, 128–135. doi:10.1109/ICTAI.2004.28. (cited
on page 14)
Strehl, A. L. and Littman, M. L., 2008. An analysis of model-based Interval Estimation for Markov Decision Processes. Journal of Computer and System Sciences, 74,
8 (12 2008), 1309–1331. doi:10.1016/j.jcss.2007.08.009. (cited on page 14)
Sutton, R. and Barto, A., 1998. Reinforcement Learning: An Introduction, vol. 1. MIT
press Cambridge. ISBN 0262193981. (cited on pages xiii, 1, 2, 7, 8, 9, 10, 11, 23, 32,
and 49)
Sutton, R. S., 1988. Learning to Predict by the Methods of Temporal Differences.
Machine Learning, 3, 1 (1988), 9–44. doi:10.1023/A:1022633531479. (cited on page
6)
Sutton, R. S., 1990. Integrated Architectures for Learning, Planning, and Reacting
Based on Approximating Dynamic Programming. ML, (1990), 216–224. doi:10.1.1.
51.7362. (cited on page 14)
Sutton, R. S., 1999. Reinforcement Learning: Past, Present and Future. In AsiaPacific Conference on Simulated Evolution and Learning, 195–197. Springer. doi:10.
1007/3-540-48873-1{_}26. (cited on page 1)
Tang, H.; Houthooft, R.; Foote, D.; Stooke, A.; Chen, X.; Duan, Y.; Schulman, J.;
De Turck, F.; and Abbeel, P., 2016. #Exploration: A Study of Count-Based Exploration for Deep Reinforcement Learning. arXiv preprint arXiv:1611.04717, (2016),
1–11. (cited on pages 15, 21, 34, 47, and 58)
BIBLIOGRAPHY
65
Thrun, S. B., 1992. Efficient Exploration In Reinforcement Learning. In Science,
January, 1–44. (cited on pages 12 and 16)
van Hasselt, H.; Guez, A.; and Silver, D., 2016. Deep Reinforcement Learning with
Double Q-learning. In AAAI, 2094–2100. (cited on page 49)
Veness, J.; Ng, K. S.; Hutter, M.; and Bowling, M., 2012. Context tree switching. In
Data Compression Conference Proceedings, 327–336. doi:10.1109/DCC.2012.39. (cited
on page 15)
Wang, Z.; Schaul, T.; Hessel, M.; van Hasselt, H.; Lanctot, M.; and de Freitas,
N., 2015. Dueling Network Architectures for Deep Reinforcement Learning. CoRR,
abs/1511.0, 9 (11 2015), 1–16. doi:10.1109/MCOM.2016.7378425. (cited on page 58)
Watkins, C. J. C. H. and Dayan, P., 1992. Q-learning. Machine Learning, 8, 3-4 (1992),
279–292. doi:10.1007/BF00992698. (cited on page 9)
| 2cs.AI
|
Processor Allocation for Optimistic Parallelization of
Irregular Programs∗
Francesco Versaci†
Keshav Pingali‡
TU Wien & University of Padova
University of Texas at Austin
arXiv:1201.3778v2 [cs.PL] 27 Jun 2012
Abstract
Optimistic parallelization is a promising approach for the parallelization of irregular
algorithms: potentially interfering tasks are launched dynamically, and the runtime system detects conflicts between concurrent activities, aborting and rolling back conflicting
tasks. However, parallelism in irregular algorithms is very complex. In a regular algorithm
like dense matrix multiplication, the amount of parallelism can usually be expressed as a
function of the problem size, so it is reasonably straightforward to determine how many
processors should be allocated to execute a regular algorithm of a certain size (this is
called the processor allocation problem). In contrast, parallelism in irregular algorithms
can be a function of input parameters, and the amount of parallelism can vary dramatically during the execution of the irregular algorithm. Therefore, the processor allocation
problem for irregular algorithms is very difficult.
In this paper, we describe the first systematic strategy for addressing this problem.
Our approach is based on a construct called the conflict graph, which (i) provides insight
into the amount of parallelism that can be extracted from an irregular algorithm, and
(ii) can be used to address the processor allocation problem for irregular algorithms. We
show that this problem is related to a generalization of the unfriendly seating problem and,
by extending Turán’s theorem, we obtain a worst-case class of problems for optimistic
parallelization, which we use to derive a lower bound on the exploitable parallelism. Finally, using some theoretically derived properties and some experimental facts, we design
a quick and stable control strategy for solving the processor allocation problem heuristically.
Keywords: Irregular algorithms, Optimistic parallelization, Automatic parallelization, Amorphous data-parallelism, Processor allocation, Unfriendly seating, Turán’s theorem.
∗ The
original publication is available at www.springerlink.com [24]
Email: [email protected]. This works was partially supported by PAT-INFN Project
AuroraScience, by MIUR-PRIN Project AlgoDEEP, and by the University of Padova Projects STPD08JA32 and
CPDA099949
‡ [email protected]
† Contact author.
1
1
Introduction
The advent of on-chip multiprocessors has made parallel programming a mainstream concern. Unfortunately writing correct and efficient parallel programs is a challenging task for
the average programmer. Hence, in recent years, many projects [14, 10, 3, 20] have tried
to automate parallel programming for some classes of algorithms. Most of them focus on
regular algorithms such as Fourier transforms [9, 19] and dense linear algebra routines [4].
Automation is more difficult when the algorithms are irregular and use pointer-based data
structures such as graphs and sets. One promising approach is based on the concept of amorphous data parallelism [17]. Algorithms are formulated as iterative computations on work-sets,
and each iteration is identified as a quantum of work (task) that can potentially be executed
in parallel with other iterations. The Galois project [18] has shown that algorithms formulated in this way can be parallelized automatically using optimistic parallelization): iterations
are executed speculatively in parallel and, when an iteration conflicts with concurrently executing iterations, it is rolled-back. Algorithms that have been successfully parallelized in this
manner include Survey propagation [5], Boruvka’s algorithm [6], Delauney triangulation and
refinement [12], and Agglomerative clustering [21].
In a regular algorithm like dense matrix multiplication, the amount of parallelism can
usually be expressed as a function of the problem size, so it is reasonably straightforward
to determine how many processors should be allocated to execute a regular algorithm of a
certain size (this is called the processor allocation problem). In contrast, parallelism in irregular algorithms can be a function of input parameters, and the amount of parallelism can
vary dramatically during the execution of the irregular algorithm [16]. Therefore, the processor allocation problem for irregular algorithms is very difficult. Optimistic parallelization
complicates this problem even more: if there are too many processors and too little parallel
work, not only might some processors be idle but speculative conflicts may actually retard
the progress of even those processors that have useful work to do, increasing both program
execution time and power consumption. This paper1 presents the first systematic approach to addressing the processor allocation problem for irregular algorithms under optimistic parallelization, and
it makes the following contributions.
• We develop a simple graph-theoretic model for optimistic parallelization and use it
to formulate processor allocation as an optimization problem that balances parallelism
exploitation with minimizing speculative conflicts (Section 2).
• We identify a worst-case class of problems for optimistic parallelization; to this purpose,
we develop an extension of Turán’s theorem [2] (Section 3).
• Using these ideas, we develop an adaptive controller that dynamically solves the processor allocation problem for amorphous data-parallel programs, providing rapid response
to changes in the amount of amorphous data-parallelism (Section 4).
2
Modeling Optimistic Parallelization
A typical example of an algorithm that exhibits amorphous data-parallelism is Dalauney
mesh refinement, summarized as follows. A triangulation of some planar region is given,
containing some “bad” triangles (according to some quality criterion). To remove them, each
bad triangle is selected (in any arbitrary order), and this triangle, together with triangles
that lie in its cavity, are replaced with new triangles. The retriangulation can produce new
bad triangles, but this process can be proved to halt after a finite number of steps. Two bad
triangles can be processed in parallel, given that their cavities do not overlap.
1A
brief announcement of this work has been presented at SPAA’11 [23]
2
(i)
(ii)
(iii)
Figure 1: Optimistic parallelization. (i) Nodes represent possible computations, edges conflicts
between them. (ii) m nodes are chosen at random and run concurrently. (iii) At runtime the
conflicts are detected, some nodes abort and their execution is rolled back, leaving a maximal
independent set in the subgraph induced by the initial nodes choice.
There are also algorithms, which exhibit amorphous data-parallelism, for which the order
of execution of the parallel tasks cannot be arbitrary, but must satisfy some constraints (e.g.,
in discrete event simulations the events must commit chronologically). We will not treat this
class of problems in this work, but we will focus only on unordered algorithms [16]. A different
context in which there is no roll-back and tasks do not conflict, but obey some precedence
relations, is treated in [1].
Optimistic parallelization deals with amorphous data-parallelism by maintaining a workset of the tasks to be executed. At each temporal step some tasks are selected and speculatively launched in parallel. If, at runtime, two processes modify the same data a conflict is
detected and one of the two has to abort and roll-back its execution. Neglecting the details of
the various amorphous data-parallel algorithms, we can model their common behavior at a
higher level with a simple graph-theoretic model: we can think a scheduler as working on a
dynamic graph Gt = (Vt , Et ), where the nodes represent computations we want to do, but we
have no initial knowledge of the edges, which represent conflicts between computations (see
Fig. 1). At time step t the system picks uniformly at random mt nodes (the active nodes) and
tries to process them concurrently. When it processes a node it figures out if it has some connections with other executed nodes and, if a neighbor node happens to have been processed
before it, aborts, otherwise the node is considered processed, is removed from the graph and
some operations may be performed in the neighborhood, such as adding new nodes with
edges or altering the neighbors. The time taken to process conflicting and non-conflicting
nodes is assumed to be the same, as it happens, e.g., for Dalauney mesh refinement.
2.1
Control Optimization Goal
When we run an optimistic parallelization we have two contrasting goals: we both want
to maximize the work done, achieving high parallelism, but at the same time we want to
minimize the conflicts, hence obtaining a good use of the processors time. (Furthermore,
for some algorithms the roll-back work can be quite resource-consuming.) These two goals
are not compatible, in fact if we naïvely try to minimize the total execution time the system
is forced to use always all the available processors, whereas if we try to minimize the time
wasted from aborted processes the system uses only one processor. Therefore in the following
we choose a trade-off goal and cast it in our graph-theoretic model.
Let G = (V, E) be a computations/conflicts (CC) graph with n = |V | nodes. When
a scheduler chooses, uniformly at random, m nodes to be run, the ordered set πm (·) by
which they commit can be modeled as a random permutation: if i < j then πm (i ) commits
before πm ( j) (if there is a conflict between πm (i ) and πm ( j) then πm (i ) commits and πm ( j)
3
aborts, if πm (i ) aborted due to conflicts with previous processes πm ( j) can commit, if not
conflicting with other committed processes). Let k t (πm ) be the number of aborted processes
due to conflicts and rt (πm ) ∈ [0, 1) the ratio of conflicting processors observed at time t (i.e.
rt (πm ) , k t (πm )/m). We define the conflict ratio r̄t (m) to be the expected r that we obtain
when the system is run with m processors:
r̄t (m) , Eπm [rt (πm )] ,
(1)
where the expectation is computed uniformly over the possible prefixes of length m of the n
nodes permutations. The control problem we want to solve is the following: given r (τ ) and
mτ for τ < t, choose mt = µt such that r̄t (µt ) ' ρ, where ρ is a suitable parameter.
Remark 1. If we want to dynamically control the number of processors, ρ must be chosen
different from zero, otherwise the system converges to use only one processor, thus not being
able to identify available parallelism. A value of ρ ∈ [20%, 30%] is often reasonable, together
with the constraint mt ≥ 2.
3
Exploiting Parallelism
In this section we study how much parallelism can be extracted from a given CC graph and
how its sparsity can affect the conflict ratio. To this purpose we obtain a worst case class of
graphs and use it to analytically derive a lower bound for the exploitable parallelism (i.e., an
upper bound for the conflict ratio). We make extensive use of finite differences (i.e., discrete
derivatives), which are defined recursively as follows. Let f : Z → R be a real function
defined on the integers, then the i-th (forward) finite difference of f is
∆if (k ) = ∆if−1 (k + 1) − ∆if−1 (k ) ,
with ∆0f (k ) = f (k ) .
(2)
(In the following we will omit ∆’s superscript when equal to one, i.e., ∆ , ∆1 .)
First, we obtain two basic properties of r̄, which are given by the following propositions.
Proposition 1. The conflict ratio function r̄ (m) is non-decreasing in m.
To prove Prop. 1 we first need a lemma:
Lemma 1. Let k̄ (m) , Eπm [k (πm )]. Then k̄ is a non-decreasing convex function, i.e. ∆k̄ (m) ≥ 0
and ∆2k̄ (m) ≥ 0.
Proof. Let k̃ (πm , i ) be the expected number of conflicting nodes running r = m + i nodes
concurrently, the first m of which are πm and the last i are chosen uniformly at random
among the remaining ones. By definition, we have
Eπm k̃ (πm , i ) = k̄ (m + i ) .
(3)
In particular,
k̃ (πm , 1) = k (πm ) + Pr [(m + 1)-th conflicts] ,
(4)
k̄ (m + 1) = Eπm k̃(πm , 1) = k̄ (m) + η ,
(5)
which brings
with η = k̄ (m + 1) − k̄ (m) = ∆k̄ (m) ≥ 0, hence proving the monotonicity of k̄. Consider now
k̃ (πm , 2) = k (πm ) + Pr [(m + 1)-th conflicts] + Pr [(m + 2)-th conflicts] .
(6)
If the (m + 1)-th node does not add any edge, then we have
Pr [(m + 1)-th conflicts] = Pr [(m + 2)-th conflicts] ,
(7)
but since it may add some edges the probability of conflicting the second time is in general
larger and thus ∆2k̄ (m) ≥ 0.
4
Prop. 1. Since r̄ (m) = k̄ (m)/m, its finite difference can be written as
∆r̄ (m) =
m∆k̄ (m) − k̄ (m)
.
m ( m + 1)
(8)
Because of Lemma 1 and being k̄ (1) = 0 we have
k̄(m + 1) ≤ m∆k̄ (m) ,
(9)
which finally brings
∆r̄ (m) =
m∆k̄ (m) − k̄ (m)
k̄ (m + 1) − k̄ (m)
∆k̄ (m)
≥
=
≥0 .
m ( m + 1)
m ( m + 1)
m ( m + 1)
(10)
Proposition 2. Let G be a CC graph, with n nodes and average degree d, then the initial derivative of
r̄ depends only on n and d as
d
∆r̄ (1) =
.
(11)
2( n − 1)
Proof. Since
∆r̄ (1) =
k̄ (2)
∆k̄ (1) − k̄ (1)
=
,
2
2
(12)
we just need to obtain k̄ (2). Let k̃ be defined as in the proof on Lemma 1 and π1 = v a node
chosen uniformly at random. Then
dv
Ev [ d v ]
d
k̄ (2) = Ev k̃ (v, 1) = Ev
=
=
.
(13)
n−1
n−1
n−1
A measure of the available parallelism for a given CC graph has been identified in [15]
considering, at each temporal step, a maximal independent set of the CC graph. The expected
size of a maximal independent set gives a reasonable and computable estimate of the available
parallelism. However, this is not enough to predict the actual amount of parallelism that a
scheduler can exploit while keeping a low conflict ratio, as shown in the following example.
Example 1. Let G = Kn2 ∪ Dn where Kn2 is the complete graph of size n2 and Dn a disconnected
graph of size n (i.e. G is made up of a clique of size n2 and n disconnected nodes). For this graph every
maximal independent set is maximum too and has size n + 1, but if we choose n + 1 nodes uniformly
at random and then compute the conflicts we obtain that, on average, there are only 2 independent
nodes.
A more realistic estimate of the performance of a scheduler can be obtained by analyzing
the CC graph sparsity. The average degree of the CC graph is linked to the expected size of
a maximal independent set of the graph by the following well known theorem (in the variant
shown in [2] or [22]):
Theorem 1. (Turán, strong formulation). Let G = (V, E) be a graph, n = |V | and let d be the
average degree of G. Then the expected size of a maximal independent set, obtained choosing greedily
the nodes from a random permutation, is at least s = n/(d + 1).
Remark 2. The previous bound is existentially tight: let Kdn be the graph made up of s =
n/(d + 1) cliques of size d + 1, then the average degree is d and the size of every maximal
(and maximum) independent set is exactly s. Furthermore, every other graph with the same
number of nodes and edges has a bigger average maximal independent set.
5
The study of the expected size of a maximal independent set in a given graph is also
known as the unfriendly seating problem [7, 8] and is particularly relevant in statistical physics,
where it is usually studied on mesh-like graphs [11]. The properties of the graph Kdn has
suggested us the formulation of an extension of the Turán’s theorem. We prove that the
graphs Kdn provide a worst case (for a given degree d) for the generalization of this problem
obtained by focusing on maximal independent set of induced subgraphs. This allows, when
given a target conflict ratio ρ, the computation of a lower bound for the parallelism a scheduler
can exploit.
Theorem 2. Let G be a graph with same nodes number and degree of Kdn and let EMm ( G ) be the
expected size of a maximal independent set of the subgraph induced by a uniformly random choice of
m nodes in G, then
EMm ( G ) ≥ EMm (Kdn ) .
(14)
To prove it we first need the following lemma.
j
Lemma 2. The function η j ( x ) , ∏i=1 (n − i − x ) is convex for x ∈ [0, n − j].
Proof. We prove by induction on j that, for x ∈ [0, n − j],
η j0 ( x ) ≤ 0 ,
ηj (x) ≥ 0 ,
η j00 ( x ) ≥ 0 .
(15)
Base case Let η0 ( x ) = 1. The properties above are easily verified.
Induction Since η j ( x ) = η j−1 ( x )(n − j − x ), we obtain
η j0 ( x ) = −η j−1 ( x ) + (n − j − x )η j0−1 ( x ) ,
(16)
which is non-positive by inductive hypotheses. Similarly,
η j00 ( x ) = −2η j0−1 ( x ) + (n − j − x )η j00−1 ( x )
(17)
is non-negative.
Thm. 2. Consider a random permutation π of the nodes of a generic graph G that has the same
number of nodes and edges of Kdn . We assume the prefix of length m of π (i.e. π (1), . . . , π (m))
forms the active nodes and focus on the following independent set ISm in the subgraph induced: a node v is in ISm ( G, π ) if and only if it is in the first m positions of π and it has no
neighbors preceding it. Let bm ( G ) be the expected size of ISm ( G, π ) averaged over all possible
π’s (chosen uniformly):
bm ( G ) , Eπ [# ISm ( G, π )] .
(18)
Since for construction bm ( G ) ≤ EMm ( G ) whereas bm (Kdn ) = EMm (Kdn ), we just need to prove
that bm (Kdn ) ≤ bm ( G ). Given a generic node v of degree dv and a random permutation π, its
probability to be in ISm ( G, π ) is
1
Pr [v ∈ ISm ( G, π )] =
n
m j −1
∑∏
j =1 i =1
n − i − dv
.
n−i
(19)
By the linearity of the expectation we can write b as
j −1
1 vn m
n − i − dv
bm ( G ) =
∑ ∏ n − i = Ev
n v∑
= v1 j =1 i =1
bm (Kdn ) =
m j −1
∑∏
j =1 i =1
n−i−d
=
n−i
m j −1
∑∏
j =1 i =1
6
"
m j −1
∑∏
j =1 i =1
n − i − dv
n−i
n − i − Ev [ d v ]
.
n−i
#
,
(20)
(21)
To prove that EMm ( G ) ≥ EMm (Kdn ) is thus enough showing that
"
∀j
Ev
j
j
#
∏(n − i − dv ) ≥ ∏ (n − i − Ev [dv ]) ,
i =1
(22)
i =1
which can be done applying Jensen’s inequality [13], since in Lemma 2 we have proved the
j
convexity of η j ( x ) , ∏i=1 (n − i − x ).
Corollary 1. The worst case for a scheduler among the graphs with the same number of nodes and
edges is obtained for the graph Kdn (for which we can analytically approximate the performance, as
shown in §3.1).
Proof. Since
r̄ (m) =
m − EMm ( G )
1
= 1 − EMm ( G ) ,
m
m
(23)
the thesis follows.
3.1
Analysis of the Worst-Case Performance
Theorem 3. Let d be the average degree of G = (V, E) with n = |V | (for simplicity we assume
n/(d + 1) ∈ N). The conflict ratio is bounded from above as
!
m
n
n−d−i
r̄ (m) ≤ 1 −
1−∏
.
(24)
m ( d + 1)
n+1−i
i =1
Proof. Let s = n/(d + 1) be the number of connected components in Kdn . Because of Thm. 2
and Cor. 1 it suffices to show that
!
m
n
−
d
−
i
EMm (Kdn ) = s 1 − ∏
.
(25)
n+1−i
i =1
The probability for a connected component k of Kdn not to be accessed when m nodes are
chosen is given by the following hypergeometric
n−d−1
d+1
m
m
0
n−d−i
Pr[k not hit] =
=∏
.
(26)
n+1−i
n
i =1
m
Let Xk be a random variable that is 1 when component k is hit and 0 otherwise. We have
d −i
that E[ Xk ] = 1 − ∏im=1 nn−
+1−i and, by the linearity of the expectation, the average number of
components accessed is
"
#
!
s
s
m
n−d−i
E ∑ Xk = ∑ E[ Xk ] = s 1 − ∏
.
(27)
n+1−i
i =1
k =1
k =1
Corollary 2. When n and m increase the bound is well approximated by
n
m d +1
r̄ (m) ≤ 1 −
1− 1−
.
m ( d + 1)
n
7
(28)
Proof. Stirling approximation for the binomial, followed by low order terms deletion in the
resulting formula.
Corollary 3. If we set m = αs =
αn
d +1
we obtain
"
d +1 #
α
1
1
1− 1−
≤ 1−
1 − e−α .
r̄ (m) ≤ 1 −
α
d+1
α
4
(29)
Controlling Processors Allocation
In this section we will design an efficient control heuristic that dynamically chooses the number of processes to be run by a scheduler, in order to obtain high parallelism while keeping
the conflict ratio low. In the following we suppose that the properties of Gt are varying slowly
compared to the convergence of mt toward µt under the algorithm we will develop (see §4.1),
so we can consider Gt = G and µt = µ and thus our goal is making mt converge to µ.
Since the conflict ratio is a non-decreasing function of the number of launched tasks m
(Prop. 1) we could find m ' µ by bisection simply noticing that
r̄ (m0 ) ≤ ρ ≤ r̄ (m00 )
m0 ≤ µ ≤ m00 .
⇒
(30)
The control we propose is slightly more complex and is based on recurrence relations, i.e.,
we compute mt+1 as a function F of the target conflict ratio ρ and of the parameters which
characterize the system at the previous timestep:
mtF+1 = F (ρ, rt , mt ) .
(31)
The initial value m0 for a recurrence can be chosen to be 2 but, if we have an estimate of the
CC graph average degree d, we can choose a smarter value: in fact applying Cor. 3 we are
sure that using, e.g., m = 2(dn+1) processors we will have at most a conflict ratio of 21.3%.
Our control heuristic (Algorithm 1) is a hybridization of two simple recurrences. The first
recurrence is quite natural and increases m based on the distance between r and ρ:
Recurrence A:
mtA+1 = (1 − rt + ρ)mt .
(32)
The second recurrence exploits some experimental facts. In Fig. 2 we have plotted the conflict
ratio functions for three CC graphs with the same size and average degree (note that initial
derivative is the same for all the graphs, in accordance with Prop. 2). We see that conflict
ratios which reach a high value (r̄ (n) > 21 ) are initially well approximated by a straight line
(for m such that r̄ (m) ≤ ρ = 20 ÷ 30%), whereas functions that deviates from this behavior do
not raise too much. This suggests us to assume an initial linearity in controlling mt , as done
by the following recurrence:
mtB+1 =
Recurrence B:
ρ
mt .
rt
(33)
The two recurrences can be roughly compared as follows (see Fig. 3): Recurrence A has
a slower convergence than Recurrence B, but it is less susceptible to noise (the variance that
makes rt realizations different from r̄t ). This is the reason for which we chose to merge
them in an hybrid algorithm: initially, when the difference between r and ρ is big, we use
Recurrence B to exploit its quick convergence and then Recurrence A is adopted, for a finer
tuning of the control.
8
Algorithm 1: Pseudo-code of the proposed hybrid control algorithm
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
// Tunable parameters
m0 = 2;
mmax = 1024;
mmin = 2;
T = 4;
rmin = 3%;
α0 = 25%;
α1 = 6%;
// Variables
m ← m0 ;
r ← 0;
t ← 0;
// Main loop
while nodes to elaborate 6= 0 do
t ← t + 1;
if m > mmax then m ← mmax ;
else if m < mmin then m ← mmin ;
Launch the scheduler with m nodes;
r ← r + new conflict ratio;
if (t mod T ) = T − 1 then
r ← r/T;
r
α ← 1− ;
ρ
if α > α0 then
if r < rmin then r ← rmin ;
lρ m
m←
m ;
r
else if α > α1 then
m ← d(1 − r + ρ) me;
r ← 0;
1
0.8
r̄(m)
Upper bound
0.6
Random graph
Cliques + discon. nodes
Common tangent
0.4
0.2
0
1
200 400 600 800 1000 1200 1400 1600 1800 2000
m
Figure 2: A plot of r̄ (m) for some graphs with n = 2000 and d = 16: (i) the worst case upper
bound of Cor. 2 (ii) a random graph (edges chosen uniformly at random until desired degree is
reached; data obtained by computer simulation) (iii) a graph unions of cliques and disconnected
nodes.
9
300
d = 4, Rec. A
d = 4, Hybrid
200
mt
d = 16, Rec. A
d = 16, Hybrid
100
0
0
20
40
t
60
80
100
Figure 3: Comparison between two realizations of the hybrid algorithm and one that only uses
Recurrence A, for two different random graphs (n = 2000 in both cases). The hybrid version has
different parameters for m greater or smaller than 20. ρ was chosen to be 20%. The proposed
algorithm proves to be both quick in convergence and stable.
4.1
Experimental Evaluation
In the practical implementation of the control algorithm we have made the following optimizations:
• Since rt can have a big variance, especially when m is small, we decided to apply the
changes to m every T steps, using the averaged values obtained in these intervals, to
smooth the oscillations.
• To further reduce the oscillations we apply a change only if the observed rt is sufficiently
different from ρ (e.g. more than 6%), thus avoiding small variations in the steady state,
which interfere with locality exploitation because of the data moving from one processor
to another.
• Another problem that must be considered is that for small values of m the variance is
much bigger, so it is better to tune separately this case using different parameters (this
optimization is not shown in the pseudo-code).
To validate our controller we have run the following simulation: a random CC graph of
fixed average degree d is taken and the controller runs on it, starting with m0 = 2. We are
interested in seeing how many temporal steps it takes to converge to mt ' µ. As can be
seen in [15] the parallelism profile of many practical applications can vary quite abruptly,
e.g., Delauney mesh refinement can go from no parallelism to one thousand possible parallel
tasks in just 30 temporal steps. Therefore, an algorithm that wants to efficiently control the
processors allocations for these problems must adapt very quickly to changes in the available
parallelism. Our controller, that uses the very fast Recurrence B in the initial phase, proves
to do a fast enough job: as shown in Fig. 3 in about 15 steps the controller converges close to
the desired µ value.
5
Conclusions and Future Work
Automatic parallelization of irregular algorithms is a rich and complex subject and will offer
many difficult challenges to researchers in the next future. In this paper we have focused
on the processor allocation problem for unordered data-amorphous algorithms; it would
be extremely valuable to obtain similar results for the more general and difficult case of
ordered algorithms (e.g., discrete event simulation), in particular it is very hard to obtain good
estimates of the available parallelism for such algorithms, given the complex dependencies
arising between the concurrent tasks. Another aspect which needs investigation, especially in
10
the ordered context, is whether some statical properties of the behavior of irregular algorithms
can be modeled, extracted and exploited to build better controllers, able to dynamically adapt
to the different execution phases.
As for a real-world implementation, the proposed control heuristic is now being integrated
in the Galois system and it will be evaluated on more realistic workloads.
Acknowledgments
We express our gratitude to Gianfranco Bilardi for the valuable feedback on recurrence-based
controllers and to all the Galois project members for the useful discussions on optimistic
parallelization modeling.
References
[1] Agrawal, K., Leiserson, C.E., He, Y., Hsu, W.J.: Adaptive work-stealing with parallelism
feedback. ACM Trans. Comput. Syst. 26(3), 7:1–7:32 (Sep 2008), http://doi.acm.org/
10.1145/1394441.1394443
[2] Alon, N., Spencer, J.: The probabilistic method. Wiley-Interscience (2000)
[3] An, P., Jula, A., Rus, S., Saunders, S., Smith, T.G., Tanase, G., Thomas, N., Amato, N.M.,
Rauchwerger, L.: Stapl: An adaptive, generic parallel C++ library. In: Dietz, H.G. (ed.)
LCPC. Lecture Notes in Computer Science, vol. 2624, pp. 193–208. Springer (2001)
[4] Blackford, L.S., Choi, J., Cleary, A., D’Azevedo, E., Demmel, J., Dhillon, I., Dongarra, J.,
Hammarling, S., Henry, G., Petitet, A., Stanley, K., Walker, D., Whaley, R.C.: ScaLAPACK
Users’ Guide. Society for Industrial and Applied Mathematics, Philadelphia, PA (1997)
[5] Braunstein, A., Mézard, M., Zecchina, R.: Survey propagation: An algorithm for satisfiability. Random Struct. Algorithms 27(2), 201–226 (2005)
[6] Eppstein, D.: Spanning trees and spanners. In: Sack, J., Urrutia, J. (eds.) Handbook of
Computational Geometry, pp. 425–461. Elsevier (2000)
[7] Freedman, D., Shepp, L.: Problem 62-3, an unfriendly seating arrangement. SIAM Review 4(2), p. 150 (1962), http://www.jstor.org/stable/2028372
[8] Friedman, H.D., Rothman, D., MacKenzie, J.K.: Problem 62-3. SIAM Review 6(2), pp.
180–182 (1964), http://www.jstor.org/stable/2028090
[9] Frigo, M., Johnson, S.G.: The design and implementation of FFTW3. Proceedings of the
IEEE 93(2), 216–231 (2005), special issue on “Program Generation, Optimization, and
Platform Adaptation”
[10] Frigo, M., Leiserson, C.E., Randall, K.H.: The implementation of the Cilk-5 multithreaded language. In: PLDI. pp. 212–223 (1998)
[11] Georgiou, K., Kranakis, E., Krizanc, D.: Random maximal independent sets and the
unfriendly theater seating arrangement problem. Discrete Mathematics 309(16), 5120
– 5129 (2009), http://www.sciencedirect.com/science/article/B6V00-4W55T4X-2/2/
72d38a668c737e68edf497512e606e12
[12] Guibas, L.J., Knuth, D.E., Sharir, M.: Randomized incremental construction of delaunay
and voronoi diagrams. Algorithmica 7(4), 381–413 (1992)
11
[13] Jensen, J.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta
Mathematica 30(1), 175–193 (1906)
[14] Kalé, L.V., Krishnan, S.: Charm++: A portable concurrent object oriented system based
on C++. In: OOPSLA. pp. 91–108 (1993)
[15] Kulkarni, M., Burtscher, M., Cascaval, C., Pingali, K.: Lonestar: A suite of parallel irregular programs. In: ISPASS. pp. 65–76. IEEE (2009)
[16] Kulkarni, M., Burtscher, M., Inkulu, R., Pingali, K., Cascaval, C.: How much parallelism
is there in irregular applications? In: Reed, D.A., Sarkar, V. (eds.) PPOPP. pp. 3–14. ACM
(2009)
[17] Méndez-Lojo, M., Nguyen, D., Prountzos, D., Sui, X., Hassaan, M.A., Kulkarni, M.,
Burtscher, M., Pingali, K.: Structure-driven optimizations for amorphous data-parallel
programs. In: Govindarajan, R., Padua, D.A., Hall, M.W. (eds.) PPOPP. pp. 3–14. ACM
(2010)
[18] Pingali, K., Nguyen, D., Kulkarni, M., Burtscher, M., Hassaan, M.A., Kaleem, R., Lee,
T.H., Lenharth, A., Manevich, R., Méndez-Lojo, M., Prountzos, D., Sui, X.: The tao of
parallelism in algorithms. In: Proceedings of the 32nd ACM SIGPLAN conference on
Programming language design and implementation. pp. 12–25. PLDI ’11, ACM, New
York, NY, USA (2011), http://doi.acm.org/10.1145/1993498.1993501
[19] Püschel, M., Moura, J., Johnson, J., Padua, D., Veloso, M., Singer, B., Xiong, J., Franchetti,
F., Gacic, A., Voronenko, Y., Chen, K., Johnson, R., Rizzolo, N.: Spiral: Code generation
for dsp transforms. Proceedings of the IEEE 93(2), 232–275 (Feb 2005)
[20] Reinders, J.: Intel threading building blocks. O’Reilly & Associates, Inc., Sebastopol, CA,
USA (2007)
[21] Tan, P.N., Steinbach, M., Kumar, V.: Introduction to Data Mining. Addison-Wesley (2005)
[22] Tao, T.: Additive combinatorics. Cambridge University Press (2006)
[23] Versaci, F., Pingali, K.: Brief announcement: processor allocation for optimistic parallelization of irregular programs. In: Proceedings of the 23rd ACM symposium on Parallelism in algorithms and architectures. pp. 261–262. SPAA ’11, ACM, New York, NY,
USA (2011), http://doi.acm.org/10.1145/1989493.1989533
[24] Versaci, F., Pingali, K.: Processor allocation for optimistic parallelization of irregular
programs. In: Murgante, B., Gervasi, O., Misra, S., Nedjah, N., Rocha, A., Taniar, D.,
Apduhan, B. (eds.) Computational Science and Its Applications – ICCSA 2012, Lecture
Notes in Computer Science, vol. 7333, pp. 1–14. Springer Berlin / Heidelberg (2012),
http://dx.doi.org/10.1007/978-3-642-31125-3_1
12
| 6cs.PL
|
Texture Segmentation Based Video Compression
Using Convolutional Neural Networks
Chichen Fu⋆ , Di Chen⋆ , Edward Delp⋆ , Zoe Liu† , Fengqing Zhu⋆
University
†
Google, Inc. (United States)
arXiv:1802.02992v1 [cs.CV] 8 Feb 2018
⋆ Purdue
Abstract
There has been a growing interest in using different approaches to improve the coding efficiency of modern video codec
in recent years as demand for web-based video consumption increases. In this paper, we propose a model-based approach that
uses texture analysis/synthesis to reconstruct blocks in texture regions of a video to achieve potential coding gains using the AV1
codec developed by the Alliance for Open Media (AOM). The
proposed method uses convolutional neural networks to extract
texture regions in a frame, which are then reconstructed using a
global motion model. Our preliminary results show an increase
in coding efficiency while maintaining satisfactory visual quality.
Introduction
With the increasing amount of videos being created and consumed, better video compression tools are needed to provide
fast transmission and high visual quality. Modern video coding
standards utilize spatial and temporal redundancy in the videos
to achieve high coding efficiency and high visual quality with
motion compensation techniques and 2-D orthogonal transforms.
However, efficient exploitation of statistical dependencies measured by a mean squared error (MSE) does not always produce
the best psychovisual result, and may require higher data rate to
preserve detail information in the video.
Recent advancement in GPU computing has enabled the
analysis of large scale data using deep learning method. Deep
learning techniques have shown promising performance in many
applications such as object detection, natural language process,
and synthetic images generation [1, 2, 3, 4]. Several methods
have been developed for video applications to improve coding efficiency using deep learning. In [5], sample adaptive offset (SAO)
is replaced by a CNN-based in-loop filter (IFCNN) to improve
the coding efficiency in HEVC. By learning the predicted residue
between the quantized reconstructed frames obtained after deblocking filter (DF) and the original frames, IFCNN is able to reconstruct video frames with higher quality without requiring any
bit transmission during coding process. Similar to [5], [6] proposes a Variable-filter-size Residue-learning CNN (VRCNN) to
improving coding efficiency by replacing DF and SAO in HEVC.
VRCNN is based on the concept of ARCNN [7] which is originally designed for JPEG applications. Instead of only using spatial information to train a CNN to reduce the coding artifacts in
HEVC, [8] proposed a spatial temporal residue network (STResNet) as an additional in-loop filter after SAO. A rate-distortion
optimization strategy is used to control the on/off switch of the
proposed in-loop filter. There are also some works that have been
done in the decoder of HEVC to improve the coding efficiency.
In [9], a deep CNN-based auto decoder (DCAD) is implemented
in the decoder of HEVC to improve the video quality of decoded
video. DCAD is trained to learn the predict residue between decoded video frames and original video frames. By adding the predicted residual generated from DCAD to the compressed video
frames, this method enhances the compressed video frames to
higher quality.
In summary, the above methods improve the coding efficiency by enhancing the quality of reconstructed video frames.
However, they require different trained models for video reconstruction at different quantization levels.
We are interested in developing deep learning approaches to
only encode visually relevant information and use a different coding method for “perceptually insignificant” regions in a frame,
which can lead to substantial data rate reductions while maintaining visual quality. In particular, we have developed a model based
approach that can be used to improve the coding efficiency by
identifying texture areas in a video frame that contain detail irrelevant information, which the viewer does not perceive specific
details and can be skipped or encoded at a much lower data rate.
The task is then to divide a frame into “perceptually insignificant”
texture region and then use a texture model for the pixels in that
region.
In 1959, Schreiber and colleagues proposed a coding method
that divides an image into textures and edges and used it in image
coding [10]. This work was later extended by using the human visual system and statistical model to determine the texture region
[11, 12, 13]. More recently, several groups have focused on adapting perceptual based approaches to the video coding framework
[14]. In our previous work [15], we introduced a texture analyzer before encoding the input sequences to identify detail irrelevant regions in the frame which are classified into different texture
classes. At the encoder, no inter-frame prediction is performed for
these regions. Instead, displacement of the entire texture region is
modeled by just one set of motion parameters. Therefore, only the
model parameters are transmitted to the decoder for reconstructing the texture regions using a texture synthesizer. Non-texture
regions in the frame are coded conventionally. Since this method
uses feature extraction based on texture segmentation technique,
a proper set of parameters are required to achieve accurate texture segmentation for different videos. Deep learning methods
usually do not require such parameter tuning for inference. As a
result, deep learning techniques can be developed to perform texture segmentation and classification for the proposed model-based
video coding. A Fisher vector convolution neural networks (FVCNN) that can produce segmentation labels for different texture
classes was proposed in [16]. One of the advantage of FV-CNN
is that the image input size is flexible and is not limited by the
network architecture. Instead of doing pixel-wise classification
on texture regions, a texture classification CNN network was described in [17]. To reduce computational expenses, [17] uses a
small classification network to classify image patches with size of
227 × 227. A smaller network is needed to classify smaller image patches in our case. In this paper, we propose a block-based
texture segmentation method to extract texture region in a video
frame using convolutional neural networks. The block-based segmentation network classifies each 16 × 16 block in a frame as
texture or non-texture. The identified texture region is then synthesized using the temporal correlations among the frames. Our
method was implemented using the AOM/AV1 codec. Preliminary results show significant bitrate savings while maintaining a
satisfactory visual quality.
The CNN architecture of our method is shown in Figure 2.
Our CNN network is inspired by the VGG network architecture
[18]. The input of our architecture is a 16 × 16 color image block.
The architecture consists of convolutional layers followed by a
batch normalization rectified linear unit (ReLU) and a max pooling operation. Three fully connected layers with dropout operations and a softmax layer produces class probabilities. The output
of our network is the probability that a 16 × 16 block is labeled as
texture or non-texture, which indicates reliably of the texture/nontexture block label produced by the network. The kernel size of
the convolutional layer is 3 × 3 and is padded by 1. The max pooling layer down samples the image by 2 and doubles the number
of feature maps.
Method
Figure 1. Block Diagram of Proposed Method
Figure 1 shows a block diagram of the proposed system. The
original frames are first analyzed by texture detector to generate
texture segmentation masks. Then, AOM/AV1 global motion tool
is used to warping the identified texture region in a reference
frame to synthesize the identified texture region in the current
frame. The texture synthesis for each frame uses provided segmentation masks without sending residues for identified texture
region.
Texture Analysis
To provide texture information for the AOM/AV1 encoder,
we use a block-based deep learning texture detector to analyze
video frames and produce segmentation masks. Our deep learning
detector obtains texture segmentation masks by classifying each
16 × 16 block in a frame.
CNN Architecture
Figure 2. CNN Architecture for Texture Block Classification
Figure 3. CNN Training Data Preparation Procedure
Training
The Salzburg Texture Image Database (STex) and the
Places365 [19] were used for training the CNN. Images in the
STex dataset are ”pure” texture images. Images in the Place365
dataset are of general scenes. The texture class samples were created by cropping 512 × 512 STex images into 256 × 256 and
128 × 128 image patches and resizing them to 16 × 16 patches.
ThePlace365 images were resized into 16 × 16 image patches.
Since the texture content in the Places365 images were lost during resizing operation, we can use resized Places365 images as
non-texture class samples. In total, 1740 texture class images and
36148 non-texture class images were generated. The CNN training data preparation procedure is shown in Figure 2.
The proposed CNN architecture was implemented in Torch
[20]. Mini-batch gradient descent is used with a fixed learning
rate of 0.01, a momentum of 0.9 and a weight decay of 0.0005.
The batch size of 512 image patches was trained in each iteration.
For each epoch, 74 iterations were performed to cover the entire training set. The training set was shuffled before each epoch.
Since our training set is unbalanced for the texture class images
and non-texture class images, the class weight of each class was
set proportion to the inverse of class frequency. The error of training was calculated using cross entropy criterion and was con-
verted to probability score using softmax regression. A total of
100 epochs were trained using one NVIDIA GTX Titan GPU.
video. The texture motion parameters are sent to the decoder in
the uncompressed header of inter predicted frames.
Inference
Texture Block Decision
After training the CNN, texture segmentation is performed
using test video frames. Each frame is divided into 16 × 16 adjacent non-overlapping blocks. Each block in the video frames is
classified as either texture or non-texture block. The segmentation
mask for each frame is formed by grouping the classified blocks
in the frame.
If a block is entirely inside the texture region of the current
frame and the warped area in the reference frame is also inside the
texture region of the reference frame, then this block is considered
a texture block. If the block is a texture block, we do not further
split it into smaller sub-blocks. If the block contains no texture
region, RD optimization is performed for block partitioning and
mode decision.
Texture-Based Video Coding
We use AOM/AV1 codec [21] to implement and test the proposed method. In our implementation, a video sequence is first divided into group of frames (GF group). In each GF group, texture
synthesis method is disabled for the first frame and enabled for the
rest of the frames. When a frame contains identified synthesizable
texture regions, the texture region is synthesized by warping the
corresponding region from the reference frames. This is implemented by using the global motion mode [22] in AOM/AV1. The
texture analyzer is integrated into the encoder as follows and is
illustrated in Figure 4:
Figure 4. Encoding Operation Flow
Texture Synthesis
The global motion tool of AOM/AV1 codec was used to synthesize texture blocks by warping the reference frame. On the
encoder side, the texture blocks select global motion mode and
only use the reference frames that have texture motion parameters. To avoid artifacts on the block edges of the texture blocks,
all texture blocks within one frame use the same reference frame.
The residual of the texture blocks is set to zero.
Experimental Result
To evaluate our method, four pairs of data rates and PSNRs
are calculated for each test sequences using the baseline codec.
The baseline codec and our proposed method both use a singlelayer coding structure. There are 4-16 frames in each GF group.
Our proposed method enables texture mode for all the frames except the GOLDEN and ALTREF frames in AV1, which means
that there is at least one frame in each GF group that does not use
texture mode. Bjontegaard metric [23] is used to evaluate the performance of our method. Bjontegaard metric are BD-RATE and
BD-PSNR which measure average data rate change and PSNRs
change respectively. The data rate is computed by dividing the
WebM file size by the number of frames. The WebM file is the
output of the encoder. PSNR is calculated based only on nontexture regions of the frame for the decoded video. We obtain four
different pairs of data rates and PSNRs for each test sequence.
The BD-RATE and BD-PSNR were calculated using these four
pairs of data rate and PSNR points.
Table 1. BD-RATE and BD-PSNR
1. Load and store texture mask for the current frame and all its
reference frames;
2. Estimate texture motion parameters for all the reference
frames and replace at least one set of the codec’s global motion parameters with the texture motion parameters;
3. Check if a block is a texture block;
4. Stop further block splitting for texture blocks;
5. Choose global motion mode for texture blocks;
6. Set the residual of texture block to be zero.
Texture Motion Parameters
In order to get more accurate motion estimation of texture
region, instead of using global motion parameters from the codec
which are estimated based on the reference frame and the entire
current frame, we obtained the texture motion parameters for each
frame based on the reference frame and the texture region in the
current frame. This indeed reduces the artifacts on the block edges
of the texture blocks and non-texture blocks in the reconstructed
Sequence
BD-RATE (%) BD-PSNR (dB)
Flower garden
-16.49
1.25
Football
-5.68
0.42
Figure 5 shows the reconstructed sample video frames. The
quantitative evaluation results are shown in Table 1, 2, 3. The
evaluation results show large data rate saving for our sample decoded videos while the reconstructed video frames showing no
significant visual artifacts. For sequence with fast motion, motion artifacts are observed between the edge of texture regions and
non-texture regions illustrated in Figure 6. In our current implementation, texture synthesis is performed using the codec’s builtin global motion function. Inaccurate motion parameters can lead
to motion artifacts in some sequences. A potential approach to
fix the artifacts is to use different motion models when estimating
the motion parameters. By addressing what types of motion may
cause such visual artifacts, the most suitable motion model can
Figure 5. Sample Reconstructed Video Frames
Figure 6. Sample Reconstructed Video Frame with Motion Artifacts Between the Edge of Texture Regions and Non-Texture Regions
be selected to fit the scheme in different videos. The complexity of the motion models used for each identified texture regions
can be adjusted depending on the texture content. In addition, we
would like to categorize different types of motion present in the
video and assess how well our texture based method performs for
different types of motion. This can then be used as feedback to
automatically determine which mode is best suited for encoding
the frame/block using machine learning methods.
Figure 7 shows the texture segmentation results. Our CNN
detector is more likely to find texture blocks that only contain one
single type of texture within a 16 × 16 block. If a block contains
different types of texture content, it may not be detected as a texture block. Using more training samples that contains different
types of textures could improve the performance of our CNN detector. Additional post-processing techniques such as connected
components can be used to remove small texture regions.
Conclusion
In this paper, we proposed a texture segmentation based
video coding method for AOM/AV1 codec. A deep learning based
texture segmentation method was developed to detect texture re-
gions in a frame that is “perceptually insignificant.” The proposed
method is implemented in the AOM/AV1 codec by enabling the
global motion mode to ensure temporal consistency. Preliminary
results showed a large data rate saving without noticeable visual
artifacts.
References
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Proceedings of the
Neural Information Processing Systems, pp. 1097–1105, December
2012, Lake Tahoe, NV.
[2] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 3431–3440, June
2015, Boston, MA.
[3] J. Schmidhuber, “Deep learning in neural networks: An overview,”
Neural networks, vol. 61, pp. 85–117, 2015.
[4] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
learning applied to document recognition,” Proceedings of the IEEE,
vol. 86, no. 11, pp. 2278–2324, November 1998.
[5] W. Park and M. Kim, “Cnn-based in-loop filtering for coding effi-
Table 2. Data Rate Savings Obtained for Flower Garden Sequence
Quantization Level Data Rate - AOM/AV1 (bits/frames) Data Rate - Our Method (bits/frames) Data Rate Savings (%)
16
115330
136080
15.24
24
81558
94695
13.87
28
63621
73768
13.76
32
51326
59326
13.48
Table 3. Data Rate Savings Obtained for Football Sequence
Quantization Level Data Rate - AOM/AV1 (bits/frames) Data Rate - Our Method (bits/frames) Data Rate Savings (%)
16
112720
122110
7.69
24
79621
83874
5.71
28
62905
65811
4.42
32
49600
51705
4.71
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
Figure 7. Sample Texture Segmenatation Results
[18]
[6]
[7]
[8]
[9]
ciency improvement,” Image, Video, and Multidimensional Signal
Processing Workshop, pp. 1–5, 2016.
Y. Dai, D. Liu, and F. Wu, “A convolutional neural network approach
for post-processing in hevc intra coding,” International Conference
on Multimedia Modeling, pp. 28–39, 2017.
C. Dong, Y. Deng, C. C. Loy, and X. Tang, “Compression artifacts reduction by a deep convolutional network,” Proceedings of
the IEEE International Conference on Computer Vision, pp. 576–
584, 2015.
C. Jia, S. Wang, X. Zhang, S. Wang, and S. Ma, “Spatialtemporal residue network based in-loop filter for video coding,”
arXiv preprint, p. arXiv:1709.08462, 2017.
T. Wang, M. Chen, and H. Chao, “A novel deep learning-based
[19]
[20]
[21]
[22]
method of improving coding efficiency from the decoder-end for
hevc,” Data Compression Conference, pp. 410–419, 2017.
W. F. Schreiber, C. F. Knapp, and N. D. Kay, “Synthetic highs, an
experimental tv bandwidth reduction system,” J. Soc. Motion Picture Televis. Eng., vol. 68, pp. 525–537, 1959.
H. Peterson, “Image segmentation using human visual system properties with applications in image compression,” Ph.D. dissertation,
Purdue Univ, May 1990, West Lafayette, IN.
M. Kunt, A. Ikonomopoulos, and M. Kocher, “Second-generation
image-coding techniques,” Proc. IEEE, vol. 73, no. 4, pp. 549–574,
April 1985.
E. J. Delp, R. L. Kashyap, and O. Mitchell, “Image data compression using autoregressive time series models,” Pattern Recognit., pp.
313–323, June 1979.
P. Ndjiki-Nya, D. Doshkov, H. Kaprykowsky, F. Zhang, D. Bull,
and T. Wiegand, “Perception-oriented video coding based on image analysis and completion: A review,” Signal Processing: Image
Communication, vol. 27, no. 6, pp. 579–594, July 2012.
M. Bosch, F. Zhu, and E. J. Delp, “Segmentation-based video compression using texture and motion models,” IEEE Journal of Selected Topics in Signal Processing, vol. 5, no. 7, pp. 1366–1377,
2011.
M. Cimpoi, S. Maji, and A. Vedaldi, “Deep filter banks for texture
recognition and segmentation,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3828 – 3836,
2015.
V. Andrearczyk and P. F. Whelan, “Using filter banks in convolutional neural networks for texture classification,” Pattern Recognition Letters, pp. 63–69, 2016.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint, p.
arXiv:1409.1556, 2014.
B. Zhou, A. Khosla, A. Lapedriza, A. Torralba, and A. Oliva,
“Places: An image database for deep scene understanding,” arXiv
preprint, p. arXiv:1610.02055, 2016.
R. Collobert, K. Kavukcuoglu, and C. Farabet, “Torch7: A
matlab-like environment for machine learning,” Proceedings of the
BigLearn workshop at the Neural Information Processing Systems,
pp. 1–6, December 2011, Granada, Spain.
“Alliance
for
open
media,
press
release
online,”
http://aomedia.org/press-release.
S. Parker, Y. Chen, D. Barker, P. D. Rivaz, and D. Mukherjee, “Global and locally adaptive warped motion compensation in
video compression,” Proc. IEEE Int. Conf. Image Process. (ICIP),
September 2017.
[23] G. Bjntegaard, “Calculation of average psnr differences between
rdcurves,” VCEGM33, 13th VCEG meeting, March 2001, Austin,
Texas.
Author Biography
Chichen Fu received his BS in Electrical Engineering from
Purdue University (2014) and he is pursuing his PhD in Electrical and Computer Engineering from Purdue University. His
research interest include machine learning, image processing and
computer vision.
Di Chen is a PhD candidate in Video and Image Processing
Laboratory (VIPER) at Purdue University, West Lafayette. Her
research focuses on video analysis and compression. Currently,
she is working on texture segmentation based video compression
using convolutional neural networks.
Edward J. Delp was born in Cincinnati, Ohio. He is currently The Charles William Harrison Distinguished Professor of
Electrical and Computer Engineering and Professor of Biomedical Engineering at Purdue University. His research interests include image and video processing, image analysis, computer vision, image and video compression, multimedia security, medical imaging, multimedia systems, communication and information theory. Dr. Delp is a Life Fellow of the IEEE, a Fellow of the
SPIE, a Fellow of IS&T, and a Fellow of the American Institute of
Medical and Biological Engineering.
Zoe Liu is a software engineer with Google’s WebM team
and has been a key contributor to the open source video codec
standard AOMedia Codec One (AV1). Zoe received her PhD from
Purdue University and her ME and BE from Tsinghua University,
Beijing. Either as a principal contributor or as a Technical Lead,
Zoe had previously devoted her effort to the design and development of several video call products, including FaceTime, Tango
Video Call, and Google Glass Video Call. Her main research
interests include video compression, processing, and communications.
Fengqing Zhu is an Assistant Professor of Electrical and
Computer Engineering at Purdue University, West Lafayette, IN.
Dr. Zhu received her Ph.D. in Electrical and Computer Engineering from Purdue University in 2011. Prior to joining Purdue in 2015, she was a Staff Researcher at Huawei Technologies
(USA), where she received a Huawei Certification of Recognition
for Core Technology Contribution in 2012. Her research interests include image processing and analysis, video compression,
computer vision and computational photography.
| 1cs.CV
|
arXiv:1401.0976v1 [cs.CC] 6 Jan 2014
Progress on Polynomial Identity Testing - II
Nitin Saxena
To my grand-advisor Professor Somenath Biswas
Abstract. We survey the area of algebraic complexity theory; with the
focus being on the problem of polynomial identity testing (PIT). We
discuss the key ideas that have gone into the results of the last few
years.
Mathematics Subject Classification (2010). Primary 68Q25, 68W30; Secondary 12Y05, 13P25.
Keywords. arithmetic circuit, identity testing, hitting-set, rank, lower
bound, Jacobian, concentration, shift, morphism.
Contents
1. Introduction
2. Shallow circuits, deep interconnections
3. Faithful morphisms, hitting-sets
4. Rank concentration, shift, hitting-sets
5. Open ends
References
1
4
6
10
13
14
1. Introduction
Algebraic complexity theory is the study of computation via algebraic models,
hence, algebraic techniques. In this article we work with only one model –
arithmetic circuit (in short, circuit). A circuit C(x1 , . . . , xn ), over a ring R,
computes a polynomial f in R[x1 , . . . , xn ]. Its description is in the form of a
rooted tree; with the leaves inputting the variables or constants, the internal
nodes computing addition or multiplication, and the root outputting the f .
The edges in C, called wires, carry the intermediate polynomials and could
also be used to multiply by a constant (from R). By the size, respectively the
depth, of C we mean the natural thing (sometimes to avoid “trivialities” we
2
Saxena
might want to take into account the bit-size needed to represent an element
in R).
A moment’s thought would suggest that a circuit is a rather compact
way of representing polynomials. Eg. a circuit of size s could produce a polynomial of degree 2s (hint: repeated squaring). In fact, a single product gate
could multiply s linear polynomials and produce nΩ(s) many monomials.
Thus, a circuit is an ‘exponentially’ compact representation of some polynomial families. Conversely, are there ‘explicit’ polynomial families (say nvariate n-degree) that require exponential (i.e. 2n ) sized circuits? We “expect” almost every polynomial to be this hard, but, the question of finding
an explicit family is open and is the main goal motivating the development
of algebraic complexity.
One can try to directly give a good lower bound against circuits by
designing an explicit polynomial family {fn } and prove that it requires a
‘large’ sized circuit family {Cn }. The other, indirect, way is to design an
efficient hitting-set H for the circuit family, i.e. if Cn 6= 0 then ∃a ∈ H,
Cn (a) 6= 0. This ‘flip’ from lower bounds to algorithms was first remarked by
[HS80] and now it has several improved versions [KI04, Agr05, Agr06]. This is
a remarkable phenomena and is one of the primary motivations to study the
question of PIT: Given a circuit C test it for zeroness, in time polynomial in
size(C). The hitting-set version of PIT is also called blackbox PIT (contrasted
with whitebox PIT).
The last 10 years have seen a decent growth of algebraic tools and techniques to understand the properties of polynomials that a circuit computes.
The feeling is that these polynomials are special, different from general polynomials, but a strong enough algebraic ‘invariant’ or a combinatorial ‘concept’ is still lacking. There have been several articles surveying the known
techniques and the history of PIT [Sax09, AS09, SY10, CKW11, Sap13]. In
this survey we will attempt not to repeat what those surveys have already
covered. So, we will focus only on the new ideas and assume that the reader
has given at least a cursory glance at the older ones. We directly move on to
the Leitfaden.
1.1. Survey overview
This article deals mainly with three broad topics – the ‘universality’ of depth3 circuits, the design of hitting-sets via ‘faithful’ morphisms and that via rank
‘concentration’. A major emerging area that we skip in this article is that of
PIT vis à vis GCT (geometric complexity theory) program [Mul11, Mul12a,
Mul12b]; the algebraic-geometry interpretations there are interesting though
any concrete PIT algorithm, or application, is yet to emerge.
Shallow circuits. A depth-2 circuit (top + gate) of size s, over a field, essentially computes a sum of s monomials. Such polynomials are called sparse
polynomials; blackbox PIT for them was solved few decades ago. So, our next
Progress on PIT
3
stop is depth-3: Polynomials of the form
C=
k Y
d
X
Li,j ,
i=1 j=1
where Li,j are linear polynomials in F[x1 , . . . , xn ]. Significant research has
been done with this model, but both sub-exponential PIT and exponential
lower bounds are open here. Recently, a remarkable universality result was
shown for depth-3 [GKKS13]: If an n-variate poly(n)-degree polynomial can
be nontrivially computed by a circuit, then it can be nontrivially computed in
depth-3. This ‘squashing’ of depth means that it suffices to focus on depth-3
for PIT purposes.
If we consider a depth-2 circuit (top × gate), over a ring R, then again
we get some remarkable connections. Fix R to be the 2 × 2 matrix algebra
M2 (F), and consider the circuit
D=
d
Y
Li ,
i=1
where Li are linear polynomials in R[x1 , . . . , xn ]. Traditionally, D is called
a width-2 algebraic branching program (ABP). It was shown by [SSS09] that
depth-3 PIT efficiently reduces to width-2 ABP PIT.
Faithful morphisms. It was observed in the last few years that in all the known
hitting-sets, the key idea in the proof is to work with a homomorphism ϕ and
an algebraic property that the image of ϕ should preserve. [SS12] used a
(Vandermonde-based) map ϕ : F[x1 , . . . , xn ] → F[y1 , . . . , yk ] that preserves
the ‘linear’ rank of any k linear polynomials. This gave the first blackbox
PIT for bounded top fanin depth-3, over any field.
[BMS13, ASSS12] used a (Vandermonde & Kronecker-based) map ϕ :
F[x1 , . . . , xn ] → F[y1 , . . . , yk ] that preserves the ‘algebraic’ rank (formally,
transcendence degree) of certain k polynomials. This gave the first blackbox
PIT (and lower bounds) for several well-studied classes of constant-depth
circuits. One drawback of the technique is that it requires zero/large characteristic fields.
Rank concentration. Inspired from the tensors, a restricted circuit model
called multilinear read-once ABP (ROABP) has been intensively studied.
Let R be the w × w matrix algebra Mw (F) and let {Si } be a partition of
Qd
[n]. Consider the circuit D = i=1 Li , where Li are linear polynomials in
R[xSi ] (i.e. the linear factors have disjoint variables). For D [FSS13] gave a
hitting-set in time poly(wn)log w·log n , i.e. quasi-poly-time. The proof is based
on the idea, following [ASS13], that after applying a small (Kronecker-based)
‘shift’, D gets the property: The rank of its coefficients (viewed as F-vectors)
is concentrated in the ‘low’ support monomials. Thus, checking the zeroness
of these low monomials is enough!
We conjecture that rank-concentration, after a ‘small’ shift, should be
attainable in any ABP D. But, currently the proof techniques are not that
4
Saxena
strong. Recently, [AGKS13] have achieved rank-concentration in multilinear
depth-3 circuits where the partitions (corresponding to each product gate)
are ‘close’ to each other in the sense of ‘refinement’.
2. Shallow circuits, deep interconnections
In this section we exhibit the key ideas behind the universality of two shallow
circuits.
2.1. The depth-3 chasm
In the study of circuits one feels that low-depth should already hold the key.
This feeling was confirmed in a series of work [VSBR83, AV08, Koi12, Tav13]:
Any poly(n)-degree n-variate polynomial
computed by a poly(n)-sized circuit
√
C can also be computed by a nO( n) sized depth-4 circuit!
The idea for this is, in retrospect, simple – since the degree is only
poly(n), first, squash the depth of C to O(log n) by only a polynomial blowup
in the size (the product gates we get are quite balanced). Next, identify a
subcircuit
√ C2 by picking those gates whose output polynomial has degree at
least n, and call the remaining subcircuit C1 . We view C2 as our circuit of
interest that takes gates√
of C1 as input. It can be shown that C2 computes a
polynomial of degree ≈ n of its input variables (which are poly(n) many).
√
Obviously, each gate of C1 also computes a polynomial of degree ≈ n of
its input variables
(which are x1 , . . . , xn ). Thus, C2 finally computes a sum
√
√
poly(n)+ n
√
products, each product has n factors, and each factor
of ≈
n
√
√
√ n degree- n monomials. To put it simply, C can
is itself a sum of ≈ n+
n
√
P Q√n P Q√n
be expressed as a
circuit of size nO( n) . The details of this
proof can be seen in [Tav13].
The strength of depth-4 is surprising. Recently, an even more surpris√
ing reduction has been shown [GKKS13] – that to depth-3 (again, nO( n)
sized). We will now sketch the proof; it ties together the known results in an
unexpected way.
P Qa P Qa
Essentially, √
the idea is to modify a
circuit C of size s :=
na (where a := n) by using two polynomial identities that are in a way
“inverse” of each other, and are to do with powers-of-linear-forms. First,
replace the product gates using Fischer’s identity:
Lemma 2.1 ([Fis94]). Any degree a monomial can be expressed as a linear
combination of 2a−1 a-th powers of linear polynomials, as:
!a
a
X
X
a−1
−1
· (−1)#{i|ri =−1} .
ri yi
y1 +
y1 · · · ya = (2
· a!) ·
r2 ,...,ra ∈{±1}
i=2
P Va P
We denote this type of a circuit by the notation
, where
P Qathe
P Qa
wedge signifiesP
theVpowering
by
a.
The
above
identity
transforms
the
P
V
P
a
a
circuit C to a
circuit, of size ≈ s.
Progress on PIT
5
Next, the two power gates are ‘opened’ up using an identity introduced
by the author:
Lemma 2.2 ([Sax08]). For any a, m, there exist degree-a univariate polynomials fi,j such that
(y1 + · · · + ym )a =
ma+1
m
X Y
fi,j (yj ).
i=1 j=1
P Va P Va P
Let us carefully see
circuit C
P
P the jugglery on C. The
e
has the expression C = i Ti , where each Ti has the form ( sj=1 ℓi,ji,j )a with
linear ℓi,j ’s. We want to open up the top power gate of C. By Lemma 2.2 we
get
sa+1
s
X Y
e
fu,j (ℓi,ji,j ).
Ti =
u=1 j=1
Since fu,j is a univariate, it splits into linear polynomials when the base
field F is algebraically closed.
ℓi,j is already a linear polynomial, we deduce
PAs
QP
that Ti , and hence C, is a
circuit of size poly(s).
Finally, note that for the above arguments to work we require F to be
algebraically closed and char(F) > a. Lemma 2.2 has been generalized to all
fields by [FS13b], so it is likely that this depth-3 reduction can be extended
to all fields.
√
The optimality of n n -size, in this reduction, is open. However, [KSS13]
showed that any decent improvement would lead to a proof of V N P 6= V P .
2.2. The width-2 chasm
QP
Here we look at
circuits over a matrix algebra. Though the model D =
Q
L
,
with
linear
L
∈
R[x1 , . . . , xn ], seems innocuous at first sight, a closer
i
i
i
look proves the opposite! It can be shown fairly easily that: A polynomial
computed by a constant-depth circuit (over a field) can as well be computed
by a D over a 3 × 3 matrix algebra [BC88]. On the other extreme, by taking
R = Mn (F) we can compute the determinant of a matrix in Fn×n [MV97],
hence, arithmetic formulas (not general circuits!) can be simulated in this
model [Val79].
Perhaps surprisingly, [SSS09] showed that: A polynomial computed by
a depth-3 circuit (over a field) can as well be computed by a D over a 2Q×P
2
matrix algebra. This, togetherwith the previous subsection, makes the
circuits over M2 (F) quite strong.
Pk
Say, we want to express the depth-3 circuit C =
i=1 Ti in a 2 × 2
Qd
matrix product. Firstly, we express a product Ti = j=1 ℓi,j as:
′
ℓi,1 0
ℓi,d−1 0
Ti Ti
1 ℓi,d
, where Ti′ := Ti /ℓi,d .
···
=
·
0 1
0 1
0 1
0
1
Once we have such k 2 × 2 matrices, each containing Ti in the (1, 2)th place, we would like to sum the Ti ’s in a ‘doubling’ fashion (instead of
one-by-one).
6
Saxena
M1 M2 g
L1 L2 f
&
We describe one step of the iteration. Let
0
M3
0
L3
be encapsulating two intermediate summands f and g. With the goal of
getting (a multiple of) f + g we consider the following, carefully designed,
product:
M1 M2 g
L2 M3
0
L1 L2 f
·
·
0
M3
0
L1 M2
0
L3
=
L1 M1 L2 M3
0
L2 M3 L1 M2 (f + g)
L3 M3 L1 M2
After log k such iterations, we get a multiple of C in the (1, 2)-th entry of
the final 2×2 matrix product. Note that the middle matrix, introduced in the
LHS above, potentially doubles (in the degree of the entry polynomials) in
each iteration. Thus, finally, D is a product of poly(d2log k ) linear polynomials
over M2 (F). Thus, the size blowup is only polynomial in going from depth-3
to width-2.
3. Faithful morphisms, hitting-sets
In algebraic complexity the study of certain maps has been fruitful – homomorphisms ϕ : R := F[x1 , . . . , xn ] → F[y1 , . . . , yk ] =: R′ such that the
algebraic ‘relationship’ of certain polynomials {f1 , . . . , fk } does not change
in the image of ϕ. When fi ’s are linear this boils down to a linear algebra
question and we can easily design ϕ in time poly(n) (hint: employ Vandermonde matrix). This business becomes complicated when fi ’s are non-linear.
Then we have to ask how are fi ’s represented. If they are given via monomials
then we invoke the Jacobian criterion to design ϕ, but the time complexity
becomes exponential in k. Several variants of such faithful maps are discussed
in the PhD thesis [Mit13]. We sketch the ideas behind two basic maps here.
3.1. Bounded fanin depth-3 blackbox PIT
P
Let C =
i∈[k] Ti be a depth-3 circuit. When k is constant, C is naturally called bounded fanin depth-3. This case of PIT has, by now, a rich
history [DS07, KS07, KS11, SS11, KS09, SS13, SS12]. Several new techniques have sprung up from this model – a locally decodable code structure, a
rank-preserving map via extractors, Sylvester-Gallai configurations (higherdimensions and all fields) and rank bounds. We will sketch here the main idea
behind the poly-time blackbox PIT of bounded fanin depth-3. The details are
quite technical and could be seen in [SS13, SS12].
Vandermonde map. We define a homomorphism Ψβ , for a β ∈ F, as:
∀i ∈ [n], Ψβ : xi 7→
k
X
j=1
β ij yj ,
Progress on PIT
7
x2
x2 + 2x1
x1
x1
T1
x3 + 10x1
x3 − x1
x3 + 3x1
T2
x4 + x2
x4 + 4x2 − x1
x4 + x2 + x1
x4 + x2 − 2x1
T3
Figure 1. Nodes and paths in C = T1 + T2 + T3 + . . .
and Ψβ (α) = α for all α ∈ F. This (naturally) defines the action of Ψβ , on all
the elements of R, that preserves the ring operations. We have the following
nice property, as a consequence of [GR08, Lemma 6.1]:
Lemma 3.1 (Ψβ preserves k-rank). Let S be a subset of linear forms in R
with rk(S) ≤ k, and |F| > nk 2 . Then ∃β ∈ F, rk(ψβ (S)) = rk(S).
Intuitively, Ψβ is faithful to any algebraic object involving the elements
in span(S). The proof of this lemma is by studying the coefficient-matrix of
the linear polynomials in S, and its change under Ψβ . This map has a role
to play in bounded fanin depth-3 owing to a certain structural theorem from
[SS13] – certificate for a non-identity.
To discuss this certificate we need a definition, that of ‘paths’ of ‘nodes’
in C (assumed to be nonzero). A path p with respect to an ideal I is a
sequence of terms {p1 , p2 , . . . , pb } (these are products of linear forms) with
the following property. Each pi divides Ti , and each pi is a ‘node’ of Ti with
respect to the ideal hI, p1 , p2 , . . . , pi−1 i.1 So p1 is a node of T1 wrt I, p2 is a
node of T2 wrt hI, p1 i, etc.
Let us see an example of a path (h0i, p1 , p2 , p3 ) in Figure 1. The oval
bubbles represent the list of forms in a product gate, and the rectangles
enclose forms in a node. The arrows show a path. Starting with the zero
ideal, nodes p1 := x21 , p2 := x2 (x2 + 2x1 ), and p3 := (x4 + x2 )(x4 + 4x2 −
x1 )(x4 + x2 + x1 )(x4 + x2 − 2x1 ) form a path. Initially the path is just the
zero ideal, so x21 is a node. Note how p2 is a power of x2 modulo radsphp1 i,
and p3 is a power of x4 modulo radsphp1 , p2 i.
The non-identity certificate theorem [SS13, Theorem 25] states that for
any non-identity C, there exists a path p such that modulo hpi, C reduces to
a single nonzero multiplication term.
1 By
a node pi we mean that some nonzero constant multiple of pi is identical to a powerof-a-linear-form modulo radsphI, p1 , p2 , . . . , pi−1 i, where radsp is the ideal generated by
the set of all the linear polynomials that divide pj , j ∈ [i − 1] and the generators of I.
8
Saxena
Theorem 3.2 (Certificate for a non-identity).
Let I be an ideal generated by
P
some multiplication terms. Let C =
T
i∈[k] i be a depth-3 circuit that is
nonzero modulo I. Then ∃i ∈ {0, . . . , k − 1} such that C[i] 2 mod I has a path
p satisfying: C ≡ α · Ti+1 6≡ 0 (mod I + hpi) for some α ∈ F∗ .
The proof of this theorem involves an extension of Chinese remaindering to ideals that are generated by multiplication terms. Once we have this
structural result about depth-3, observe that we would be done if we could
somehow ensure Ti+1 ∈
/ hpi (in our application I is zero). How do we preserve
this ideal non-membership under a cheap map?
Notice that the rank of the set S0 of linear polynomials that divide the
nodes in the path p is < k (since path length is below k). Moreover, Ti+1
factors into at most d linear polynomials, denote the set by S1 . So if we apply
a map that preserves the rank of each of the d sets S0 ∪ {ℓ}, ℓ ∈ S1 , then,
intuitively, the ideal non-membership should be preserved. As rk(S0 ∪ {ℓ}) ≤
k we can employ the previously discussed map Ψβ (over a field satisfying
|F| > dnk 2 ). This idea could be easily turned into a proof; details are in
[SS12].
Finally, what we have achieved is the construction of a map Ψβ , in
time poly(dnk), that reduces the variables of C from n to k and preserves
nonzeroness. Once this is done, the poly(ndk ) blackbox PIT follows from the
brute-force hitting-set.
3.2. Depth≥ 3 results
Looking at the success of bounded fanin depth-3 one wonders about the
analogous depth-4 model:
X Y
fi,j , where fi,j are sparse polynomials.
(3.1)
C =
i∈[k] j∈[d]
Here we are thinking of a bounded k. But now even k = 2 seems nontrivial!
In fact, a simpler PIT case than this is an old open question in a related area
[vzG83].
This bounded top fanin depth-4 PIT is an important open question currently. What is doable are other restricted models of depth-4. Inspired from
the last subsection we ask: Is there a notion of ‘rank’ for general polynomials,
are there easy ‘faithful’ maps, and finally is all this useful in PIT?
There are several notions of rank in commutative algebra. The one we
[BMS13] found useful is – transcendence degree (trdeg). We say that a set
S of polynomials {f1 , . . . , fm } ⊂ F[x1 , . . . , xn ] is algebraically dependent if
there exists a nonzero annihilating polynomial A(y1 , . . . , ym ), over F, such
that A(f1 , . . . , fm ) = 0. The largest number of algebraically independent
polynomials in S is called trdeg(S). With this notion we call a homomorphism
ϕ faithful if trdeg(S) = trdeg(ϕ(S)). The usefulness of ϕ (assuming that one
can come up with it efficiently) was first proved in [BMS13]:
2 We
mean C[i] :=
P
j∈[i]
Tj .
Progress on PIT
9
Lemma 3.3 (Faithful is useful). Let ϕ be a homomorphism faithful to f =
{f1 , . . . , fm } ⊂ F[x]. Then for any C ∈ F[y], C(f ) = 0 ⇔ C(ϕ(f )) = 0.
This implies that we can use a faithful map to ‘reduce’ the number of
variables n without changing the nonzeroness of C. The strategy can be used
in cases where trdeg(f ) is small, say, smaller than a constant r.
The only missing piece is the efficiency of ϕ3 . To do this we need three
fundamental ingredients – an efficient criterion for algebraic independence
(Jacobian), its behavior under ϕ (chain rule), and standard maps (Vandermonde & Kronecker based).
Lemma 3.4 (Jacobian criterion). Let f ⊂ F[x] be a finite set of polynomials
of degree at most d, and trdeg(f ) ≤ r. If char(F) = 0 or char(F) > dr ,
then trdeg(f ) = rkF(x) Jx (f ), where Jx (f ) := (∂fi /∂xj )m×n is the Jacobian
matrix.
There are several proofs of this, see [Jac41, For91, BMS13, MSS12].
This gives us an efficient way to capture trdeg, when the characteristic is
zero/large. Let us now see how the Jacobian matrix changes under ϕ.
Lemma 3.5 (Chain rule). Jy (ϕ(f )) = ϕ (Jx (f ))·Jy (ϕ(x)), where ϕ applied to
a matrix/set refers to the matrix/set obtained by applying ϕ to every entry.
This is a simple consequence of the chain rule of ‘derivatives’. It suggests
that for ϕ to preserve the trdeg of the polynomials, we need to control – (1)
the image of the original Jacobian under ϕ, and (2) the Jacobian of the image
of x. In our applications, the former is achieved by a Kronecker-based map
(i.e. sparse PIT tricks, eg. [BHLV09]) and the latter by Vandermonde map
(as seen in the previous subsection).
This general ‘recipe’ has been successfully implemented to various circuit models. The case of the circuit C ′ (x) := C(f ), where trdeg(f ) ≤ r and
fi ’s are polynomials of sparsity at most s, was worked out in [BMS13]. The
proof follows exactly the above strategy. The time complexity is polynomial
in size(C ′ ) and (s · deg(C ′ ))r , where the exponential dependence comes from
the sparsity estimate of Jx (f ) (and of course the final brute-force hitting-set
for the r-variate ϕ(C ′ )).
[ASSS12] extended the recipe to depth-4 circuits (3.1) where the number
of fi,j ’s where any variable appears is bounded by r4 . This model is called
occur-r depth-4; it generalizes the well-studied multilinear read-r depth-4.
Interestingly, slightly modified techniques also provided exponential lower
bounds against these special models. This required proving some combinatorial properties of the derivatives of immanant (eg. permanent, determinant).
The faithful maps recipe has been able to unify all the assorted poly-time
hitting-sets known. However, one drawback is that it needs the characteristic to be zero/large. Baby steps in resolving that issue have been taken by
[MSS12].
3 It
can be shown, from first principles, that a faithful r-variate map always exists [BMS13].
that this does not mean that trdeg(fi,j |i, j) is bounded.
4 Note
10
Saxena
4. Rank concentration, shift, hitting-sets
The hitting-sets that we saw till now were for models where some parameter
was kept bounded. But we could also study models with a ‘structural’ restriction, eg. multilinearity. This route
P has also been successful and enlightening.
We call a depth-3 circuit C = i Ti multilinear if the linear factors in Ti
involve disjoint variables. Hence, each product gate Ti induces a partition Pi
on the variables (or indices) [n]. Moreover, we call C set-multilinear if these
partitions are all equal!
There is a large body of work on the set-multilinear model [RS05,
AMS10, FS12, FS13b, ASS13, FS13a, FSS13, AGKS13]. The motivation
for this model is, on the one hand, the algebraic concept of tensors, and,
on the other hand, the interest in read-once boolean branching programs
[Nis92, IMZ12, Vad12]. Interestingly, [FSS13] has shown (extending the ideas
of [ASS13]) that the situation in the arithmetic world is exponentially better
than that in the boolean one!
Here we will exhibit the key ideas of [ASS13] and [AGKS13] on two toy
cases that are already quite instructive; this saves us from the gory technical
machinery that drives the more general cases.
4.1. Multilinear ROABP
[ASS13] gave the first quasi-poly-time hitting-set for set-multilinear depth-3
(and extensions to constant-depth, non-multilinear versions). This was generalized by [FSS13]
Q to any depth; in fact, they dealt directly with the multilinear
ROABP D = i Li over Mw (F), where Li ’s are linear polynomials in disjoint
variables. Both the papers proved ‘low-support rank concentration’ in their
models.
For the following discussion we fix a base commutative ring R = Hw (F)
called the Hadamard algebra (instead of the w × w matrix algebra). This is
basically (Fk , +, ⋆), where + is the vector addition and ⋆ is the coordinatewise vector product (called the Hadamard product).
ℓ-concentration. We say that a polynomial f ∈ R[x1 , . . . , xn ] is ℓ-concentrated
if
rkF {coef f (xS ) | S ⊆ [n], |S| < ℓ} = rkF {coef f (xS ) | S ⊆ [n]},
where coef f extracts a coefficient in f .
I.e. the coefficient-vectors of ‘lower’ monomials already span every possible coefficient-vector in f . We are interested in studying whether circuits
compute an ℓ-concentrated polynomial for small ℓ (say, log n instead of n).
By itself this is not true, eg. the trivial circuit D = x1 · · · xn is not even nconcentrated. But, maybe we can transform f a bit and then attain (log n)concentration? In this case, D′ := D(x1 + 1, . . . , xn + 1) is suddenly 1concentrated!
It was shown by [ASS13] that any D, above R, becomes (log k)-concentrated after applying a ‘small’ shift; the price of which is nlog k time. Once
we have this it directly
applies to the set-multilinear depth-3 model. Since,
P
a depth-3 C =
i∈[k] Ti can be rewritten as C = [1, . . . , 1] · D, where
Progress on PIT
11
T1
D = ... is of the promised sort over R = Hk (F) (since D completely
Tk
factorizes into disjoint-variate linear polynomials). So, ℓ-concentration in D
implies an easy way to check C for zeroness – test the coefficients of the
monomials below ℓ-support in C.
Glimpse of a proof. We now show how to achieve ℓ-concentration, ℓ =
O(log k), in the following toy model:
Y
(1 + zi xi ), where zi ∈ Hk (F).
(4.1)
D =
i∈[n]
Because of the disjointness of the factors
Q it can be seen, as a simple
exercise, that: D is ℓ-concentrated iff DS := i∈S (1 + zi xi ) is ℓ-concentrated,
for all S ∈ [n]
ℓ . Thus, from now on we assume, wlog, n = ℓ.
Shift D by formal variables t, and normalize, to get a new circuit:
Y
(1 + zi′ xi ), where zi′ ∈ Hk (F(t)).
D′ =
i∈[ℓ]
We can express the new coefficients as:
zi′ = zi /(1 + zi ti ), ∀i ∈ [ℓ].
Conversely, we write:
Q
zi = zi′ /(1 − zi′ ti ), ∀i ∈ [ℓ].
(4.2)
We write zS for i∈S zi . Now the goal is to ‘lift’ an F-dependence of
zS ’s to the zS′ ; which ultimately shows the condition on the shift that shall
yield concentration.
Consider the 2ℓ vectors {zS | S ⊆ [ℓ]}. If ℓ > log k then there is a
nontrivial linear dependence amongst these vectors, say,
X
αS zS = 0, where αS ∈ F.
S⊆[ℓ]
Rewriting this in terms of zS′ we get:
Y
X
αS ·
zi′ /(1 − zi′ ti ) = 0.
S⊆[ℓ]
Or,
X
S⊆[ℓ]
i∈S
αS · zS′ ·
Y
i∈[ℓ]\S
(1 − zi′ ti ) = 0.
(4.3)
′
Let us collect the ‘coefficient’ of z[ℓ]
in the above expression. It comes
out to,
X
αS · (−1)|[ℓ]\S| · t[ℓ]\S .
(4.4)
S⊆[ℓ]
If we can ensure this expression to be nonzero then Equation (4.3) tells
′
us that z[ℓ]
is in the F(t)-span of the ‘lower’ zS′ . But, ensuring the nonzeroness
of Equation (4.4) is easy – use ti ’s such that all the (≤ ℓ)-support monomials
12
Saxena
tS are distinct. We can use standard sparse PIT tricks [BHLV09] for this, in
time poly(nℓ ).
What we have shown is that, after applying a Kronecker-based shift, the
circuit D becomes ℓ-concentrated; all this in time nO(log k) . This ‘recipe’ of
studying the generic shift, via some combinatorial properties of the ‘transfer’
equations (4.2), is generalized in [ASS13] to other D; and further improved in
[FSS13] to multilinear ROABP. It is not known how to design such hittingsets, even for the toy case, in poly-time.
4.2. Towards multilinear depth-3
It is tantalizing to achieve ℓ-concentration in multilinear depth-3 (before embarking on the general depth-3!). A partial result in that direction was obtained in [AGKS13]. We will sketch their ideas in a toy model.
Consider a multilinear depth-3 circuit C with only two partitions being induced by the product gates – P1 = {{1}, · · · , {n}} and an arbitrary
partition P2 . Say, the number of the corresponding product gates is k1 respectively k2 (summing to k). We can say, naturally, that P1 is a refinement
of P2 (denoted P1 ≤ P2 ) because: For every color (or part) S ∈ P2 there exist
colors in P1 whose union is exactly S. In this
refinement situation [AGKS13]
QP
showed that, again, a suitable shift in the
circuit D (corresponding to
C) achieves ℓ-concentration in time poly(nlog k ).
Glimpse of a proof. We can assume P2 different from P1 , otherwise this case
is no differentPfrom the last subsection. We assume that the first k1 product
gates in C = i∈[k] Ti respect P1 and the rest k2 respect P2 . The correspond
T1
ing circuit D where we desire to achieve concentration is D = ... over
Tk
R = Hk (F). But
now
the
linear
factors
of
D
are
not
necessarily
in
disjoint
0
0
1
x1 x2
= x1 +
+
· x2 ·
· x2 over
variables. Eg.
1
1
0
x1 + x2
H2 (F).
To get some kind of a reduction to the set-multilinear case, we prove rank
concentration in parts. First, we consider those monomials (called P1 -type)
that could only be produced by the ‘upper’ part of D (i.e. the first k1 product
gates of C). Such a monomial, say indexed by S ⊆ [n], is characterized by
the presence of i, j ∈ S that are in the same color of P2 . For a fixed such i, j
we can “access” all such monomials by the derivative ∂ 2 D/∂xi ∂xj =: ∂i,j D.
Notice that this differentiation kills the ‘lower’ part of D and only the P1 part remains. So, we can prove (2 + log k1 )-concentration in the monomials
containing i, j as in Section 4.1. This proves O(log k1 )-concentration in the
monomials of P1 -type.
Next, we want to understand the remaining monomials (called P2 -type);
those that could be produced by the ‘lower’ part of D (i.e. the last k2 product
gates of C). These, obviously, could also be produced by the upper part of
D. Let us fix such a monomial, say x1 · · · xℓ . Assume that S1 , . . . , Sℓ ∈ P2
Progress on PIT
13
are the colors that contain one of the indices 1, . . . , ℓ. Consider the subcircuit
Dℓ that in its i-th coordinate, ∀i ∈ [k], simply drops those factors of Ti that
are free of the variables S1 ∪ · · · ∪ Sℓ . The problem here is that Dℓ may be a
‘high’ degree circuit (≈ n instead of ℓ) and so we cannot use a proof like in
Section 4.1.
But, notice that all the degree-(≥ ℓ) monomials in Dℓ are P1 -type; where
we know how to achieve ℓ-concentration. So, we only have to care about
degree-(≤ ℓ) P2 -type monomials in Dℓ . There, again, (log k)-concentration
can be shown using Section 4.1 and the well-behaved transfer equations.
This sketch, handling two refined partitions, can be made to work for
significantly generalized models [AGKS13]. But, multilinear depth-3 PIT is
still open (nothing better than exponential time known).
Remark 4.1. Using a different technique [AGKS13] also proves constantconcentration, hence designs poly-time hitting-sets, for certain constant-width
ROABP. These models are arithmetic analogues of the boolean ones – width2 read-once branching programs [AGHP92, NN93] and constant-width readonce permutation branching programs [KNP11].
5. Open ends
The search for a strong enough technique to study arithmetic circuits continues. We collect here some easy-to-state questions that interest us.
Top fanin-2 depth-4. Find a faithful map ϕ that preserves
the
Q algebraic indeQ
pendence of two products-of-sparse polynomials i fi and j gj . If we look
at the relevant 2 × 2 Jacobian determinant, say wrt variables X := {x1 , x2 },
then the question boils down to finding a hitting-set for the special rational
P det JX (fi ,gj )
function i,j
. Can this version of rational sparse PIT be done in
fi gj
sub-exponential time?
Independence over Fp . Currently, there is no sub-exponential time algorithm/heuristic known to test two given circuits for algebraic independence
over a ‘small’ finite field Fp . The reason is that something as efficient as the
Jacobian criterion is not readily available, see [MSS12].
Model in Eqn.(4.1). Find a poly-time hitting-set for this simple model. Note
that a poly-time whitebox PIT is already known [RS05].
Multilinear depth-3. Achieve o(n)-concentration in multilinear depth-3 circuits, in no(n) time. Here, the presence of an exponential lower bound against
the model [RY09] is quite encouraging.
14
Saxena
References
[AGHP92] Noga Alon, Oded Goldreich, Johan Håstad, and René Peralta, Simple construction of almost k-wise independent random variables, Random Struct. Algorithms 3 (1992), no. 3, 289–304, (Conference version
in FOCS 1990).
[AGKS13] Manindra Agrawal, Rohit Gurjar, Arpita Korwar, and Nitin Saxena,
Hitting-sets for low-distance multilinear depth-3, Electronic Colloquium
on Computational Complexity (ECCC) 20 (2013), 174.
[Agr05]
Manindra Agrawal, Proving lower bounds via pseudo-random generators,
Proceedings of the 25th Annual Foundations of Software Technology and
Theoretical Computer Science (FSTTCS), 2005, pp. 92–105.
[Agr06]
, Determinant versus permanent, Proceedings of the 25th International Congress of Mathematicians (ICM), vol. 3, 2006, pp. 985–997.
[AMS10]
Vikraman Arvind, Partha Mukhopadhyay, and Srikanth Srinivasan, New
Results on Noncommutative and Commutative Polynomial Identity Testing, Computational Complexity 19 (2010), no. 4, 521–558, (Conference
version in CCC 2008).
[AS09]
Manindra Agrawal and Ramprasad Saptharishi, Classifying polynomials
and identity testing, Indian Academy of Sciences, Platinum Jubilee P1
(2009), 1–14.
[ASS13]
Manindra Agrawal, Chandan Saha, and Nitin Saxena, Quasi-polynomial
hitting-set for set-depth-∆ formulas, STOC, 2013, pp. 321–330.
[ASSS12] Manindra Agrawal, Chandan Saha, Ramprasad Saptharishi, and Nitin
Saxena, Jacobian hits circuits: hitting-sets, lower bounds for depthD occur-k formulas & depth-3 transcendence degree-k circuits, STOC,
2012, pp. 599–614.
[AV08]
Manindra Agrawal and V. Vinay, Arithmetic circuits: A chasm at depth
four, FOCS, 2008, pp. 67–75.
[BC88]
Michael Ben-Or and Richard Cleve, Computing Algebraic Formulas Using a Constant Number of Registers, STOC, 1988, pp. 254–257.
[BHLV09] Markus Bläser, Moritz Hardt, Richard J. Lipton, and Nisheeth K. Vishnoi, Deterministically testing sparse polynomial identities of unbounded
degree, Inf. Process. Lett. 109 (2009), no. 3, 187–192.
[BMS13]
Malte Beecken, Johannes Mittmann, and Nitin Saxena, Algebraic independence and blackbox identity testing, Inf. Comput. 222 (2013), 2–19,
(Conference version in ICALP 2011).
[CKW11] Xi Chen, Neeraj Kayal, and Avi Wigderson, Partial Derivatives in
Arithmetic Complexity (and beyond), Foundation and Trends in Theoretical Computer Science 6 (2011), no. 1-2, 1–138.
[DS07]
Zeev Dvir and Amir Shpilka, Locally Decodable Codes with Two Queries
and Polynomial Identity Testing for Depth 3 Circuits, SIAM J. Comput.
36 (2007), no. 5, 1404–1434, (Conference version in STOC 2005).
[Fis94]
Ismor Fischer, Sums of like powers of multivariate linear forms, Mathematics Magazine 67 (1994), no. 1, 59–61.
Progress on PIT
15
[For91]
Krister Forsman, Constructive commutative algebra in nonlinear control
theory, Ph.D. thesis, Dept. of Electrical Engg., Linköping University,
Sweden, 1991.
[FS12]
Michael A. Forbes and Amir Shpilka, On identity testing of tensors,
low-rank recovery and compressed sensing, STOC, 2012, pp. 163–172.
[FS13a]
, Explicit Noether Normalization for Simultaneous Conjugation
via Polynomial Identity Testing, APPROX-RANDOM, 2013, pp. 527–
542.
[FS13b]
, Quasipolynomial-time Identity Testing of Non-Commutative
and Read-Once Oblivious Algebraic Branching Programs, FOCS, 2013.
[FSS13]
Michael A. Forbes, Ramprasad Saptharishi, and Amir Shpilka, Pseudorandomness for multilinear read-once algebraic branching programs,
in any order, Electronic Colloquium on Computational Complexity
(ECCC) 20 (2013), 132.
[GKKS13] Ankit Gupta, Pritish Kamath, Neeraj Kayal, and Ramprasad Saptharishi, Arithmetic circuits: A chasm at depth three, FOCS, 2013.
[GR08]
Ariel Gabizon and Ran Raz, Deterministic extractors for affine sources
over large fields, Combinatorica 28 (2008), no. 4, 415–440, (Conference
version in FOCS 2005).
[HS80]
Joos Heintz and Claus-Peter Schnorr, Testing Polynomials which Are
Easy to Compute (Extended Abstract), STOC, 1980, pp. 262–272.
[IMZ12]
Russell Impagliazzo, Raghu Meka, and David Zuckerman, Pseudorandomness from shrinkage, FOCS, 2012, pp. 111–119.
[Jac41]
Carl Gustav Jacob Jacobi, De determinantibus functionalibus, J. Reine
Angew. Math. 22 (1841), no. 4, 319–359.
[KI04]
Valentine Kabanets and Russell Impagliazzo, Derandomizing Polynomial Identity Tests Means Proving Circuit Lower Bounds, Computational Complexity 13 (2004), no. 1-2, 1–46, (Conference version in STOC
2003).
[KNP11]
Michal Koucký, Prajakta Nimbhorkar, and Pavel Pudlák, Pseudorandom generators for group products: extended abstract, STOC, 2011,
pp. 263–272.
[Koi12]
Pascal Koiran, Arithmetic circuits: The chasm at depth four gets wider,
Theor. Comput. Sci. 448 (2012), 56–65.
[KS07]
Neeraj Kayal and Nitin Saxena, Polynomial Identity Testing for Depth
3 Circuits, Computational Complexity 16 (2007), no. 2, 115–138, (Conference version in CCC 2006).
[KS09]
Neeraj Kayal and Shubhangi Saraf, Blackbox polynomial identity testing
for depth-3 circuits, FOCS, 2009, pp. 198–207.
[KS11]
Zohar Shay Karnin and Amir Shpilka, Black box polynomial identity
testing of generalized depth-3 arithmetic circuits with bounded top fanin, Combinatorica 31 (2011), no. 3, 333–364, (Conference version in CCC
2008).
[KSS13]
Neeraj Kayal, Chandan Saha, and Ramprasad Saptharishi, A superpolynomial lower bound for regular arithmetic formulas, Electronic Colloquium on Computational Complexity (ECCC) 20 (2013), 91.
16
Saxena
[Mit13]
Johannes Mittmann, Independence in Algebraic Complexity Theory, Ph.D. thesis, Mathematisch-Naturwissenschaftlichen Fakultät der
Rheinischen Friedrich-Wilhelms-Universität Bonn, Germany, December
2013.
[MSS12]
Johannes Mittmann, Nitin Saxena, and Peter Scheiblechner, Algebraic
Independence in Positive Characteristic – A p-adic Calculus, Electronic
Colloquium on Computational Complexity TR12-014 (2012), (accepted
in Trans. Amer. Math. Soc., 2013).
[Mul11]
Ketan Mulmuley, On P vs. NP and geometric complexity theory: Dedicated to Sri Ramakrishna, J. ACM 58 (2011), no. 2, 5.
[Mul12a]
, Geometric Complexity Theory V: Equivalence between Blackbox
Derandomization of Polynomial Identity Testing and Derandomization
of Noether’s Normalization Lemma, FOCS, 2012, pp. 629–638.
[Mul12b]
, The GCT program toward the P vs. NP problem, Commun. ACM 55 (2012), no. 6, 98–107.
[MV97]
Meena Mahajan and V. Vinay, Determinant: Combinatorics, Algorithms, and Complexity, Chicago J. Theor. Comput. Sci. (1997), (Conference version in SODA 1997).
[Nis92]
Noam Nisan, Pseudorandom generators for space-bounded computation,
Combinatorica 12 (1992), no. 4, 449–461, (Conference version in STOC
1990).
[NN93]
Joseph Naor and Moni Naor, Small-Bias Probability Spaces: Efficient
Constructions and Applications, SIAM J. Comput. 22 (1993), no. 4,
838–856, (Conference version in STOC 1990).
[RS05]
Ran Raz and Amir Shpilka, Deterministic polynomial identity testing in
non-commutative models, Computational Complexity 14 (2005), no. 1,
1–19, (Conference version in CCC 2004).
[RY09]
Ran Raz and Amir Yehudayoff, Lower bounds and separations for constant depth multilinear circuits, Computational Complexity 18 (2009),
no. 2, 171–207, (Conference version in CCC 2008).
[Sap13]
Ramprasad Saptharishi, Unified Approaches to Polynomial Identity
Testing and Lower Bounds, Ph.D. thesis, Department of CSE, IIT Kanpur, India, April 2013.
[Sax08]
Nitin Saxena, Diagonal Circuit Identity Testing and Lower Bounds,
ICALP (1), 2008, pp. 60–71.
[Sax09]
, Progress on Polynomial Identity Testing, Bulletin of the EATCS
(2009), no. 90, 49–79.
[SS11]
Nitin Saxena and C. Seshadhri, An Almost Optimal Rank Bound for
Depth-3 Identities, SIAM J. Comput. 40 (2011), no. 1, 200–224, (Conference version in CCC 2009).
[SS12]
, Blackbox identity testing for bounded top-fanin depth-3 circuits:
The field doesn’t matter, SIAM J. Comput. 41 (2012), no. 5, 1285–1298,
(Conference version in STOC 2011).
[SS13]
, From Sylvester-Gallai configurations to rank bounds: Improved
blackbox identity test for depth-3 circuits, J. ACM 60 (2013), no. 5, 33,
(Conference version in STOC 2010).
Progress on PIT
[SSS09]
17
Chandan Saha, Ramprasad Saptharishi, and Nitin Saxena, The Power
of Depth 2 Circuits over Algebras, FSTTCS, 2009, pp. 371–382.
[SY10]
Amir Shpilka and Amir Yehudayoff, Arithmetic Circuits: A survey of
recent results and open questions, Foundations and Trends in Theoretical
Computer Science 5 (2010), no. 3-4, 207–388.
[Tav13]
Sébastien Tavenas, Improved Bounds for Reduction to Depth 4 and
Depth 3, MFCS, 2013, pp. 813–824.
[Vad12]
Salil P. Vadhan, Pseudorandomness, Foundations and Trends in Theoretical Computer Science 7 (2012), no. 1-3, 1–336.
[Val79]
Leslie G. Valiant, Completeness classes in algebra, STOC, 1979, pp. 249–
261.
[VSBR83] Leslie G. Valiant, Sven Skyum, Stuart J. Berkowitz, and Charles Rackoff, Fast Parallel Computation of Polynomials Using Few Processors,
SIAM J. Comput. 12 (1983), no. 4, 641–644.
[vzG83]
Joachim von zur Gathen, Factoring Sparse Multivariate Polynomials,
FOCS, 1983, pp. 172–179.
Nitin Saxena
Department of CSE
IIT Kanpur
Kanpur 208016
India
e-mail: [email protected]
| 0math.AC
|
Finding Minimal Spanning Forests in a Graph
arXiv:1705.00774v1 [math.CO] 2 May 2017
Abdel-Rahman Madkour, Phillip Nadolny, Matthew Wright
St. Olaf College, Northfield, MN USA
May 3, 2017
Abstract
We propose two algorithms for solving a graph partitioning problem motivated
by computational topology. Specifically, given a weighted, undirected graph G and a
positive integer k, we desire to find k disjoint trees within G such that each vertex of
G is contained in one of the trees and the weight of largest tree is as small as possible.
We are unable to find this problem in the graph partitioning literature, but we show
that the problem is NP-complete.
We propose two approximation algorithms, one that uses a spectral clustering approach and another that employs a dynamic programming strategy, that produce nearoptimal partitions on a family of test graphs. We describe these algorithms and analyze
their empirical performance.
1
Introduction
We consider the following graph partitioning problem, motivated by work in computational
topology: given an edge-weighted, undirected graph G and a positive integer k, we desire to
find k subtrees T1 , . . . , Tk within G such that (1) each vertex of G is contained in Ti for some
i and (2) the maximum weight of any Ti is as small as possible. Although graph partitioning problems have been well-studied, we have been unable to find this particular problem
in the literature [3, 8]. Furthermore, standard variations of graph partitioning problems,
such as Euclidean sum-of-squares clustering, are known to be NP-hard [5]. Our particular
graph partitioning problem is also NP-hard, but a solution to this problem is important for
efficiently parallelizing an expensive computation in multi-dimensional persistent homology
[4]. Thus, we desire an efficient algorithm that gives a near-optimal solution to the graph
partitioning problem described above.
Motivation
The computational topology software RIVET performs an expensive computation that is
organized by a graph structure [4]. Specifically, the program must compute a “barcode
template” at each vertex of an undirected, edge-weighted graph G, which arises as the
dual graph of a certain line arrangement. In this paper we will avoid the details of the
barcode template, but simply regard it as “data” that must be computed at each vertex of
G. The computation of barcode templates is typically expensive. However, computations
at adjacent vertices are similar, and it is sometimes possible to use the result computed at
one vertex to quickly obtain the barcode template at an adjacent vertex. Edge weights in
the graph give computational cost estimates of this process. Specifically, if vertices v1 is
connected to vertex v2 by an edge of weight w, then the time required to obtain the data
at v2 , given the computed data at v1 , is roughly proportional to w.
Thus, RIVET is designed to perform an expensive computation at some starting vertex,
and then follows edges of G to each other vertex, obtaining the data at each vertex by
1
updating the data from the previous vertex. Since finding the shortest path that covers
all vertices in a graph is NP-hard, RIVET uses an approximation based on a minimum
spanning tree (MST) of G [4].
However, it is highly desirable to parallelize the RIVET computation, dividing the computational work among multiple processors. That is, if k processors are available, we would
like to partition the vertices of G into k sets of “nearby” vertices so that a single processor
can compute the data required at all vertices in a single set. To ensure that the entire
computation finishes as quickly as possible, we desire that vertices within a single set can
be connected by a tree with edges of low weight, and that the maximal weight of any such
tree is as low as possible. Thus, we arrive at the problem discussed in this paper.
Outline of Paper
We begin by discussing some mathematical background for this problem, including a proof
that the problem is NP-complete, in Section 2. We then discuss two approximation algorithms that we have implemented: a spectral algorithm in Section 3 and a dynamic
programming (DP) algorithm in Section 4. Section 5 presents some empirical data about
the performance of these algorithms on a family of test graphs, and Section 6 offers some
conclusions.
2
Mathematical Background
Terminology
We begin by reviewing some standard terminology from graph theory, and then formally
introduce the concept of a minimum spanning k-forest.
A graph G is an ordered pair (V, E), where V is a set of vertices and E is a set of edges,
each of which connects two vertices. Formally, an edge is an unordered pair of distinct
vertices in V , and each such pair occurs at most once in E. We sometimes say that the
graph is undirected (since the edges are unordered pairs) and that the graph is simple (since
each edge occurs at most once).
A subgraph of G is a pair (V ′ , E ′ ), with V ′ ⊆ V and E ′ ⊆ E, that is itself a graph. The
degree of a vertex v is the number of edges that contain v.
A graph is weighted if each edge is assigned a number, which is referred to as the “weight”
of the edge. For our purposes, we restrict edge weights to nonnegative integers. Let w(G)
denote the weight of graph G, which we define to be the sum of the weights of all edges in
the graph.
A cycle in a graph is an ordered set of vertices {v1 , v2 , . . . , vj } such that the graph
contains an edge (vi , vi+1 ) for all i ∈ {1, . . . , j − 1} and also an edge (vj , v1 ). A tree is a
connected graph with no cycles.
A spanning tree of a graph G is a subgraph of G that has no cycles (and is thus a tree),
but contains all vertices of G. A minimum spanning tree (MST) of G is a spanning tree that
has minimal weight among all spanning trees of G. Any weighted graph G has one or more
minimum spanning trees. Minimum spanning trees can be found by various algorithms,
such as Kruskal’s algorithm, which runs in O(|E| log |E|) time [2].
We now define a minimal spanning k-forest of G, which is the central object of study in
this paper.
Definition 1. A spanning k-forest of G is a collection of k trees, T1 , . . . , Tk , each a
subgraph of G, such that each vertex of G is contained in Ti for some i ∈ {1, . . . , k}.
A minimum spanning k-forest of G is a spanning k-forest such that the quantity max{w(T1 ), . . . , w(Tk )} is minimal among all spanning k-forests of G. The quantity
max{w(T1 ), . . . , w(Tk )} is the weight of the minimum spanning k-forest.
2
Note that a particular graph may have many minimum spanning k-forests, since there
may be many spanning k-forests that contain “heaviest” trees of the same weight.
Problem Statement
Given Definition 1, our problem can be stated as follows:
Main Problem. Given a graph G and an integer k > 1, find a minimal spanning k-forest
of G.
In the computational setting described in the Introduction, each tree of the minimal
spanning k-forest gives a set of vertices at which a single processor will compute the required
data, along with edges connecting the vertices used in the computational process. The
definition of minimal spanning k-forest minimizes the weight of the “heaviest” tree, which
ensures that the longest-running processor finishes as quickly as possible.
Connection to Minimum Spanning Trees
It may seem that a minimum spanning k-forest of G should be related to one or more minimum spanning trees of G. While minimum spanning k-forests and minimum spanning trees
are similar concepts, the minimality criterion of a spanning k-forest makes its construction
much more difficult than that of a MST. Specifically, the minimality criterion of a spanning
k-forest involves only the maximum-weight tree of the forest, not the total weight of all
trees.
Given any spanning tree T of a graph G, one can simply remove any k − 1 edges from
T to produce a spanning k-forest of G. However, if T is a minimal spanning tree of G, it
might not be possible to obtain a minimal spanning k-forest of G by removing edges from
T , as Figure 1 shows.
1
1
G
1
1
1
1
1
1
T1
2
1
1
2
T2
Figure 1: The only MST of graph G (left) contains all edges of G except for the
edge of weight 2. The minimum spanning 2-forest of G consists of two trees T1 and
T2 (right), and is not contained in the MST.
NP-Completeness
The Main Problem is NP-complete. We give a proof for the decision version of the Main
Problem: given a graph G and a positive number W , decide whether G has a minimal
spanning 2-forest of weight W . NP-completeness of the Main Problem then follows.
Theorem 2. The problem of deciding whether a graph G has a minimal spanning 2-forest
of weight W is NP-complete.
Proof. We show that any algorithm that can solve the problem of deciding whether a graph
G has a minimal spanning 2-forest of weight W (for any W ) can also decide whether a
multiset of positive integers can be partitioned into two subsets with the same sum, which
is an NP-hard problem.
Let x1 , x2 , . . . , xm be positive integers whose sum is 2W . Consider the graph in Figure 2.
The graph contains 3m edges. Of these, m are drawn vertically, with weights x1 , . . . , xm .
The remaining edges have weight 0.
3
xm
0
A
..
.
0
x3
0
0
0
B
0
x2
0
0
x1
Figure 2: Graph G used in the proof of Theorem 2.
We show that a minimal spanning 2-forest of G gives a partition of {x1 , . . . , xm }. First,
observe that since G contains 2m + 2 vertices, a minimal spanning 2-forest F of G must
contain exactly 2m edges that form no cycles.
All of the edges with weights x1 , . . . , xm must be in F . Otherwise, F would contain an
isolated vertex, but a spanning 2-forest that contains an isolated vertex cannot be minimal
for m > 1.
Thus, F must also contain m of the edges of weight 0. Together with the edges of weight
xi , these edges form two trees. Except in the trivial case where some xi = W , it must be
that vertices A and B are in separate trees.
Let TA be the tree containing A, and let TB be the tree containing B. Then w(TA ) =
xi1 + · · · + xip and w(TB ) = xj1 + · · · + xjp , where {i1 , . . . , ip } and {j1 , . . . , jq } form a
partition of {1, . . . , m}. Moreover, F is minimal if and only if the difference between w(TA )
and w(TB ) is as small as possible. This difference is zero if and only if G has a minimal
spanning 2 forest of weight W , which is the case if and only if it is possible to partition the
set {x1 , x2 , . . . , xm } into two subsets, both of which sum to W .
Therefore, an algorithm that can decide whether there exists a minimal spanning 2-forest
of G of weight W also solves the partition problem, which implies that the decision version of
the Main Problem is NP-hard. Since this problem is clearly also NP, it is NP-complete.
Given that the Main Problem is NP-complete, we desire to find an efficient algorithm
that produces an approximate solution to this problem. Roughly speaking, an “approximate
solution” to Main Problem is a spanning k-forest of G such that the max-weight tree of the
forest is not much larger than w
k , where w is the weight of a minimum spanning tree of G.
We now present two such algorithms. The first involves a spectral clustering technique
that is similar to the “normalized cuts” method of image segmentation. The second algorithm uses a dynamic programming (DP) approach to partition a minimum spanning tree of
G into k parts. These algorithms perform well in practice, returning approximate solutions
to Main Problem in our tests (see Section 5 for empirical data).
3
Spectral Algorithm
It seems natural to recast the Main Problem as a clustering problem: we wish to find k
clusters of the vertices of G such that the vertices within each cluster are connected by
relatively low-weight edges, and the edges that connect different clusters have relatively
higher weights. Thus, we can find an approximate solution by using a spectral clustering
algorithm to obtain such clusters of the vertices in G, and then finding a minimum spanning
tree within each cluster. Specifically, our algorithm is based on the normalized spectral
clustering of Shi and Malik, commonly known as the normalized cuts algorithm [7].
4
Step 1: Calculating Similarity
Spectral clustering requires a similarity metric: a real-valued function on pairs of vertices
that takes large values on pairs of similar vertices and values close to zero on pairs of
dissimilar vertices. In our case, two vertices are “similar” if they are connected by a lowweight path in G. Specifically, suppose V = {v1 , v2 , . . . , vm }. Let P (i, j) be the weight of
the minimal-weight (“shortest”) path between vertices vi and vj in G, and let L be the
maximum weight of any shortest path between any two vertices in G. We then define the
function ω(i, j) as
ω(i, j) = L − P (i, j)
The function ω gives the notion of similarity between vertices which we use for our spectral
clustering. Let W be the m × m matrix of similarity values; that is, W (i, j) = ω(i, j). Note
that W is a symmetric matrix.
Step 2: Finding an Eigenvector
We now introduce some terminology to explain the normalized cuts algorithm. Suppose we
desire to partition the vertex set V of G into two nonempty disjoint subsets A and B (with
A ∪ B = V ). We define a cut in graph G as follows:
X
cut(A, B) =
ω(u, v).
u∈A,v∈B
Choosing A and B to minimize cut(A, B) may result in a lopsided partition: for example,
A might consist of a single vertex. To avoid this, we use a normalized cut. We first
define a measure
P of association between a vertex subset A and the entire vertex set V :
assoc(A, V ) =
u∈A,t∈V ω(u, t). This allows us to normalize a graph cut, which takes
into account the association between each subset and the entire vertex set. Concretely, a
normalized cut is defined:
Ncut(A, B) =
cut(A, B)
cut(A, B)
+
.
assoc(A, V ) assoc(B, V )
(1)
Similarly, we define normalized association:
Nassoc(A, B) =
assoc(A, A) assoc(B, B)
+
.
assoc(A, V ) assoc(B, V )
(2)
Shi and Malik prove in that minimizing Ncut is equivalent to maximizing Nassoc. While
this problem is NP-hard, an approximate solution can be obtained by solving a generalized
eigenvalue problem [7].
With matrix W as defined above, let D be a m × m diagonal matrix where Dii is the
sum of elements in row Wi . Consider the equation
(D − W )y = λDy.
(3)
Solving equation (3) for λ and y is a symmetric generalized eigenvalue problem. A vector y
corresponding to the smallest positive eigenvalue λ of (3) can be used to obtain subsets A
and B that approximately minimize Ncut(A, B).
Step 3: Constructing the subgraphs
The entries in eigenvector y indicate the partition of V as follows: if yi > 0, then vi ∈ A;
otherwise vi ∈ B. We then construct a subgraph GA with vertex set A by including all edges
in G that connect two vertices within A; we similarly construct a subgraph GB with vertex
set B. If k = 2, then we find a MST within each subgraph; this provides an approximate
5
solution to Main Problem. An issue arises when a value yi is very close to 0, as vi could be
adjacent to a cut edge and may be placed in a vertex set A such that no edge in G connects
vi to any vertex in A. If this occurs then we run Dijkstra’s Algorithm to connect vi with
the nearest vertex vj ∈ A adding all the vertices in the path to A.
More than two blocks
If k, the desired number of trees, is a power of 2, then we simply apply the process described
in Steps 1 through 3 recursively on the subgraphs GA and GB , until we have obtained the
desired number of subgraphs. We then find a MST within each subgraph. This strategy is
performs very well in practice when k is a power of 2, as seen in Section 5.
Unfortunately, the spectral algorithm does not produce desirable solutions to the Main
Problem if k is not a power of 2. In other applications, spectral clustering can be used
to find more blocks than desired in a partition, then blocks are combined as necessary
(e.g., via k-means clustering) to obtain the desired number of blocks [7]. However, this
approach is problematic in our case since the algorithm yields subtrees of roughly equal
weight; combining two or more subtrees is likely to produce a new subtree much larger than
the others. We are not aware of a method for adapting the spectral algorithm to obtain an
approximately minimal spanning k-forest when k is not a power of 2.
Summary
In short, the spectral approach converts the graph into two matrices based on a defined
similarity metric. These matrices are used to solve a generalized eigenvalue problem to obtain an eigenvector that determines a partition of the vertex set into two blocks. Subgraphs
are then constructed by connecting every vertex in a given block to vertices in the same
block with the edges in the original graph, which partitions the graph into two blocks. This
process is repeated to obtain the number of desired blocks.
Computational Complexity
Computation of ω(u, v) for all pairs of vertices u and v requires computing the shortest
paths between all pairs of vertices. Our implementation computes all shortest paths using
the Floyd-Warshall algorithm, which runs in O(|V |3 ) time [2]. More efficient algorithms,
such as the Pettie and Ramachandran algorithm, are available [6].
The eigenvector can also be found via various algorithms. Since these methods involve
convergence of matrices under iterations of the algorithm, stating precise runtime bounds
is difficult. We use the software LAPACK (Linear Algebra PACKage) to find the eigenvector. Specifically, we call the LAPACK routine DSPGVX, which uses generalized Schur
decomposition (also called QZ decomposition) [1].
4
Dynamic Programming Algorithm
The dynamic programming approach partitions a minimum spanning tree of G into k pieces,
obtaining an approximate minimum spanning k-forest of G. Thus, the algorithm involves
two steps. First, we find a MST of G. For this step, we developed a modified Kruskal’s
algorithm to reduce the branching of the MST, which makes our second step more effective.
In the second step, we apply a dynamic programming strategy to the MST, removing k − 1
edges to obtain a spanning k-forest of G.
Step 1: Modified Kruskal’s Algorithm
Kruskal’s algorithm is a standard algorithm for finding a MST: the algorithm begins by
initializing a forest F that contains each vertex in G as a separate tree, and a set S containing
6
all edges of G. The algorithm considers all edges in S in order of increasing weight. If an
edge e connects two different trees in F , then the algorithm replaces these two trees with a
single tree formed by their union together with e. Any edge e in S that does not connect
two different trees would create a cycle, and is removed from S. If G is connected, then the
algorithm stops when F contains a single (spanning) tree.
We use a modified version of Kruskal’s algorithm to find a MST of G with relatively low
degree. Specifically, we modify Kruskal’s algorithm by refining the ordering of edges in S.
When S contains two or more edges of equal weight, each of which connects two trees in
F , we select the one that minimizes the degree of the nodes that would be connected in the
resulting tree. Our algorithm appears in Algorithm 1.
This strategy of giving priority to edges that connect low-degree vertices at each step of
Kruskal’s algorithm is not guaranteed to always produce the minimum-degree MST, but we
have found that it produces relatively low-degree MSTs in practice.
Algorithm 1 Modified Kruskal’s Algorithm
Input: Connected edge-weighted graph G = (V, E)
Output: Minimum spanning tree T of G of low degree
1: Initialize F to contain each v ∈ V as a tree
2: S ← E
3: while F contains more than 1 tree do
4:
Remove from S all edges that do not connect two trees in F
5:
S ′ ← set of lowest-weight edges in S
6:
e ← null; d ← ∞; T1 ← null; T2 ← null
7:
for each e′ = (v1 , v2 ) ∈ S ′ do
8:
T1′ ← tree in F containing v1
9:
T2′ ← tree in F containing v2
10:
d′ ← max degree of v1 and v2 in T1′ ∪ e′ ∪ T2′
11:
if d′ < d then
12:
e ← e′ ; d ← d′ ; T1 ← T1′ ; T2 ← T2′
13:
Replace T1 and T2 with T1 ∪ e ∪ T2 in F
14:
Remove e from S
Step 2: Dynamic Programming Strategy
Let T be the MST of G found by Algorithm 1 above. We now use a dynamic programming
process to find k − 1 edges that we can “cut” from T to produce a spanning k-forest of G.
Our algorithm appears in Algorithm 2.
We first find a path P in T of maximal weight; let s and t be vertices at either end of
this path. Let e1 , . . . , en denote the edges along path P from s to t. We will regard these
edges as arranged horizontally with s on the left and t on the right, as in Figure 3. For
1 ≤ i ≤ n, let ti be the weight of the portion of T to the left of ei , including ei . For the
tree depicted in Figure 3, t1 = 2, t2 = 6, and so on. For notational convenience below, let
t0 = 0 and tn+1 = tn , which is the weight of T .
s
e1
e2
e3
e4
e5
e6
t
Figure 3: Sample tree T to illustrate the dynamic programming algorithm. Edges
e1 , . . . , e6 comprise path P and each have weight 2; all other edges have weight 1.
We iterate over each edge in path P (from e1 to en ) and calculate two quantities associated with each edge: a “penalty” and a “discount.” For edge ei , the penalty is denoted pi
7
Algorithm 2 Dynamic Programming Partitioning
Input: Tree T , k = number of parts desired
Output: Cut set S of edges to remove from T
1: Find the longest path P in T ; let P = {e1 , e2 , . . . , en }
2: for i from 1 to n do
3:
ti ← weight of T to the left of and including edge ei
4: t0 ← 0; tn+1 ← tn
5: r0 ← 0
6: p0 ← 0
7: for i from 1 to
n + 1 do
w(T ) − rj − w(ei )
(ti − w(ei ) − tj ) −
8:
pi ← min
+ pj
0≤j<i
k
9:
qi ← j that minimizes the quantity above
10:
ri ← rqi + ei
11: Initialize the cut set S to the empty set
12: c ← qn+1
13: while c 6= 0 do
14:
Add ec to the set S
15:
c ← qc
⊲ Index of first edge to cut
⊲ Index of next edge to cut
and the discount is denoted ri . Intuitively, the penalty pi is a positive number that is close
to zero if it is possible to cut edge ei , along with some edges from the set {e1 , . . . , ei−1 },
and obtain a set of trees whose weights are near w(Tk)−ri . The discount ri is the sum of the
weight of edge ei and weights of edges to the left of ei that are cut to obtain the lowest
penalty.
For notational convenience, let p0 = 0. Then, for i = 1, . . . , n + 1, the penalty pi is
defined:
w(T ) − rj − w(ei )
(4)
+ pj .
pi = min
(ti − w(ei ) − tj ) −
0≤j<i
k
In equation (4), the quantity (ti −w(ei )−tj ) is the weight of the subtree between edges ej and
w(P )−rj −w(ei )
ei , and the quantity
is the weight of each subtree if it were possible to obtain
k
k subtrees by removing edge ei and the edges counted in the discount rj . Furthermore, to
make sense of equation (4), we define w(en+1 ) = 0; intuitively, we may regard en+1 as an
edge of weight 0 to the right of en , which is always the last edge that our algorithm cuts
from T . Thus, penalty pn+1 is the total penalty associated with the best possible cut set
for the entire tree T .
For each i from 1 to n + 1, the algorithm stores not only pi and ri , but also the index
j that minimizes the quantity in equation (4) above; let qi be this index. Thus we can
compute the discount ri = rqi + w(ei ), which is used in the calculation of penalties pℓ for
ℓ > i. Moreover, at the conclusion of the dynamic programming process, the qi allow us to
recover the edges that we will cut from T to minimize the penalty pn+1 .
Specifically, to obtain the set of edges that we must cut, we backtrack from along P from
en to e1 (right to left in Figure 3). The index rightmost edge to cut is given by qn+1 ; the
index of the next edge to cut is given by qqn+1 , and so on. We have found the entire cut set
when we encounter some qj = 0. Removing the edges in the cut set from T , we obtain a
spanning forest F .
Example. Running Algorithm 2 on the tree T depicted in Figure 3 with k = 3, we obtain
the values displayed in the following table.
8
i
ti
pi
qi
ri
0
1
2
3
4
5
6
7
0
2
6
12
20
26
29
29
0
9
5
1
10/3
14/3
4
4
0
0
0
3
3
4
4
0
2
2
2
4
4
6
4
The algorithm proceeds along path P from left to right. Cutting edge e1 would result in
an isolated vertex (weight 0), while the rest of the tree has weight 27, so penalty p1 =
0 − 27
3 = 9 and discount r1 = 2. If edge e2 is cut, then it is best to not cut e1 , but to
obtain a subtree of weight 4, so penalty p2 = 4 − 27
3 = 5 and discount r1 = 2. Similarly, if
edge e3 is cut, then the penalty p3 = 1 is obtained by cutting neither e2 nor e1 . However, if
edge e4 is cut, then the lowest penalty is obtained by also cutting e3 ; the subtree between
10
e3 and e4 has weight 6, and with two edges cut, penalty p4 = 6 − 25
3 + 1 = 3 and discount
r + 4 = 4. The other penalties and discounts are obtained similarly.
The algorithm then backtracks to find the cut set. Since q7 = 4, edge e4 will be cut.
Next, q4 = 3 implies that edge e3 will be cut. However, q3 = 0, so no other edges will be
cut. Therefore, edges e3 and e4 are cut, resulting in a spanning forest F containing trees of
weight 10, 6, and 9.
While the penalty calculation in Algorithm 2 gives preference to cuts that result in
)
subtrees of weight close to w(T
k , the algorithm is not guaranteed to return a cut set of size
k − 1. Thus, F is not guaranteed to contain k trees. In practice, the algorithm often does
return a cut set of size k − 1. However, if the algorithm returns a cut set of size less than
k − 1, we find additional edges to cut by by repeatedly applying Algorithm 2 to the largest
remaining subtree. If the algorithm returns a cut set of size greater than k − 1, then we
join the pair of smallest subtrees in F that are adjacent in T , continuing to do so until F
contains exactly k trees.
Computational Complexity
It is well known that the standard Kruskal’s Algorithm runs in O(|E| log |V |) time, using
a union-find data structure to store the forest F [2]. In our modified algorithm, at each
iteration of the while loop in Algorithm 1, we must identify the edge, among edges of lowest
weight that do not create cycles, that connects minimum-degree vertices. Maintaining a list
of vertex degrees in F allows for constant-time vertex degree lookup, but we must still find
the minimum among O(|E|) edges at each iteration of the while loop. Since this loop runs
O(|E|) times, the complexity of our modified Kruskal’s algorithm is O(|E|2 ).
Our dynamic programming algorithm begins by finding the longest path P in T . This
requires finding the distance between all pairs of nodes in T , which can be accomplished in
O(|V |2 ) time by a depth-first search of T . We then iterate over every edge in P and perform
a constant-time calculation for each previously-considered edge in P . Since the number of
edges in P is not more than |E|, this requires O(|E|2 ) time. Since E = V − 1, the overall
complexity of the dynamic programming algorithm is O(|E|2 ).
5
Empirical Data
We tested our algorithms, with various values of k, on a family of 90 graphs. Since the
motivation for our work arises from the computational topology software RIVET, we chose
graphs representative of those that are encountered in RIVET. These graphs are the dual
graphs of planar line arrangements [4].
9
k
Algorithm
2
2
3
4
4
5
6
7
8
8
DP
Spectral
DP
DP
Spectral
DP
DP
DP
DP
Spectral
Heaviest Subtree
Ratio
178,945
166,010
109,338
93,334
90,061
69,136
67,524
53,709
53,076
50,755
0.550
0.510
0.336
0.287
0.277
0.212
0.208
0.165
0.163
0.156
Table 1: Empirical data for Graph G1 . The MST of G1 has weight 325,427.
Methodology
Specifically, we first generated a collection of m random lines in the plane. These lines form
the boundaries of a collection of two-dimensional cells (2-cells) in the plane; each 2-cell has
positive, possibly infinite, area. The 2-cells determine a graph G (the dual graph of the
line arrangement) as follows: each 2-cell is a vertex of G, and two vertices are connected
by an edge if and only if the corresponding 2-cells have a common boundary (this common
boundary is a segment contained in one of the lines). For each line ℓ, we chose a random
integer between 1 and 1000, which is used to weight all edges in G that connect cells whose
common boundary is line ℓ.
We repeated the above process 90 times, for choices of m between 5 and 100, to generate
90 graphs. We then used our algorithms described in Section 3 and Section 4 to produce
spanning k-forests from each graph, for values of k between 2 and 32.
Data from Two Graphs
We present here data of the minimal spanning forests obtained from two of our graphs,
which are representative of what we observed from our experiments.
Graph G1 is the dual graph of an arrangement of 50 lines; this graph contains 1154
vertices and 2213 edges. A minimum-spanning tree T1 of G1 has weight 325,427. We tested
our algorithms on this graph for each integer k from 2 to 8. However, we only ran the
spectral algorithm when k is a power of 2 because the algorithm does not give desirable
results for other choices of k, as explained in Section 3.
Table 1 summarizes the spanning k-forests found by our algorithms when run on Graph
G1 , for integers k from 2 to 8. Values in the Heaviest Subtree column give the weight of
the heaviest tree in the spanning k-forest found by the algorithm. The ratio of the weight
of the heaviest subtree to the MST weight is given in the Ratio column. A lower ratio is
better; a ratio of k1 would be ideal in some sense. Indeed, since some edges of T1 are cut
and do not appear in any tree of the spanning k-forest, the ratio could be less than k1 , but
this does not occur in Table 1.
Graph G2 is the dual graph of an arrangement of 100 lines, and contains 4442 vertices
and 8543 edges. A minimum-spanning tree T2 of G2 has weight 1,627,441. We tested our
algorithms on this graph, letting k take on powers of 2 up to 32. This allowed us to directly
compare our algorithms for each value of k. Table 2 summarizes the spanning k-forest found
by our algorithms for Graph G2 ; the meaning of each column is the same as in the previous
table.
10
k
Algorithm
2
2
4
4
8
8
16
16
32
32
DP
Spectral
DP
Spectral
DP
Spectral
DP
Spectral
DP
Spectral
Heaviest Subtree
Ratio
813748
837742
522254
424436
274441
229936
169941
128014
72667
68909
0.500
0.515
0.321
0.261
0.169
0.141
0.104
0.079
0.045
0.042
Table 2: Empirical data for Graph G2 . The MST of this graph has weight 1,627,441.
k
Spectral average
Spectral st. dev.
DP average
DP st. dev.
t-statistic
p-value
2
4
8
16
32
0.517
0.011
0.530
0.049
1.398
0.173
0.274
0.011
0.288
0.024
3.063
0.005
0.149
0.008
0.158
0.019
2.681
0.012
0.082
0.008
0.084
0.008
0.913
0.369
0.048
0.014
0.044
0.003
−1.674
0.105
Table 3: Summary of tests from 30 large graphs. For each algorithm and each value
of k, we give the mean and standard deviation of 30 ratios computed. We used a
paired t-test to compare the ratios; the t-statistics and 2-sided p-values appear here.
Summary
The results reported in Table 1 and Table 2 are typical of our tests. As a summary of our
other tests, Table 3 gives statistics from our tests on our 30 largest graphs, with k taking
on powers of 2, up to 32. For each graph, algorithm, and value of k, we computed the ratio
of the weight of the largest subtree found by the algorithm to the weight of the MST of
the graph. The table gives the average and standard deviation of 30 such ratios for each
algorithm and each k. We used a paired t-test to compare the ratios achieved by the two
algorithms; the table gives the t-statistics and 2-sided p-values for each k.
Our spectral algorithm performed very well in our tests with k a small power of 2,
consistently producing ratios close to k1 . Our dynamic programming algorithm performed
reasonably well in all cases, but generally not quite as well as the spectral algorithm when k ∈
{2, 4, 8, 16}. Notably, for small k the ratios achieved by the spectral algorithm exhibit less
variation than those achieved by the DP algorithm, as indicated by the standard deviations
in Table 3. Furthermore, the difference between the ratios produced by the two tests were
most statistically significant for k = 4 and k = 8, as shown by the p-values in Table 3.
Interestingly, we noticed that for some graphs, the DP algorithm performs better for
certain smaller values of k than for certain larger values of k. For example, when run on
a particular graph, the algorithm returns a smaller largest subtree when k = 7 than when
k = 8. This is due to the fact that the initial cutting process does not always produce exactly
k subtrees, in which case we cut the largest subtree again. Of course, we do not expect this
algorithm to return the optimal partition, but it does seem to return approximately optimal
partitions in every case.
11
Runtime
The DP algorithm ran much faster than the spectral algorithm in our tests, but the runtime
of the spectral algorithm was dominated by the runtime of the Floyd-Warshall algorithm for
finding all shortest paths. We report here a few sample runtimes for the graphs discussed
above, recorded when testing our algorithms on a typical laptop computer.
For graph G1 with k = 8, the DP algorithm ran in 0.2 seconds. The spectral algorithm
ran in 26.0 seconds; of this time, 24.0 seconds was spent running the Floyd-Warshall algorithm. For the graph G2 with k = 32, the DP algorithm ran in 8.3 seconds. The spectral
algorithm ran in 1,546.2 seconds, of which 1,431.1 seconds was spent running the FloydWarshall algorithm. Thus, it seems that more than 90% of the spectral algorithm runtime is
spent in the Floyd-Warshall algorithm. As mentioned in Section 3, more efficient algorithms
for finding all shortest paths are available.
Code
The code used to generate the synthetic graphs and for the both the spectral and DP
algorithms can be found at: https://github.com/a3madkour/Minimal-Spanning-Trees
6
Conclusion
Though the Main Problem is NP-Hard, we have two algorithms that produce approximate
solutions. Our spectral algorithm performed very well in our tests with k a power of 2.
However, as stated in Section 3, we are unable to obtain desirable results with the spectral
algorithm when k is not a power of 2. Our dynamic programming algorithm performed well
in all cases, but less so than the spectral algorithm when k is a small power of 2.
Future Work
NP-Hardness for Trees We suspect that the Main Problem remains NP-hard when the
input graph is a tree, rather than a general graph, but we were not able to prove this. If
there does exist a polynomial-time algorithm to find a minimum spanning k-forest in a tree,
then such an algorithm can be used in an approximation method for finding a minimum
spanning k-forest in a general graph.
APX-Hardness for Graphs We attempted to find an algorithm that can produce a
spanning k-forest whose weight is within a constant factor of the minimum spanning kforest. However, it is an unclear if such an algorithm can run in polynomial time; that is,
we are not sure if the Main Problem is APX-hard.
Spectral Clustering for non powers of 2 In Section 3 we mentioned that there was
no obvious way of using the spectral algorithm when k is not a power of 2. However, there
may yet exist a method of using the spectral approach effectively for values of k that are
not powers of 2.
DP algorithm questions We suspect that it is possible to improve the performance
of our dynamic programming algorithm. For example, it may be possible to modify our
“penalty” calculation to obtain a better set of subtrees in the first iteration of the algorithm.
Furthermore, when the first iteration of the algorithm does not return exactly k trees, what
is the best strategy for making additional cuts? These questions require further research.
12
Acknowledgements
This work was supported by nsf-dms 1606967, the Collaborative Undergraduate Research
and Inquiry (CURI) program at St. Olaf College, and the Kay Winger Blair endowment.
References
[1] E. Anderson, et al., LAPACK Users’ Guide,
http://dx.doi.org/10.1137/1.9780898719604.
3rd Edition,
SIAM, 1999,
[2] S. Dasgupta, C. Papadimitriou, and U. Vazirani, Algorithms, McGraw Hill, 2008.
[3] P.-O. Fjällström, “Algorithms for Graph Partitioning: A Survey”, Linköping
Electronic Articles in Computer and Information Science, Vol. 3. (1998),
http://www.ep.liu.se/ea/cis/1998/010/cis98010.pdf.
[4] M. Lesnick, M. Wright, “Interactive Visualization of 2-D Persistence Modules”,
preprint, http://arxiv.org/abs/1512.00180.
[5] M.
I.
Malinen,
P.
Fränti,
“Balanced
K-Means
for
Clustering”,
Lecture Notes in Computer Science,
8621(2014),
pp. 32–41,
http://link.springer.com/chapter/10.1007/978-3-662-44415-3_4.
[6] S. Pettie, V. Ramachandran, “Computing Shortest Paths with Comparisons and Additions”, Proceedings of the Thirteenth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2002, pp. 267–276,
http://dl.acm.org/citation.cfm?id=545381.545417.
[7] J.
Shi,
J.
Malik,
“Normalized
Cuts
and
Image
IEEE Trans. Pattern Anal. Mach. Intell.,
22(2000),
http://ieeexplore.ieee.org/document/868688/.
Segmentation”,
pp. 888-905,
[8] D. Xu, Y. Tian, “A Comprehensive Survey of Clustering Algorithms”, Ann. Data Sci.
(2015) 2: 165, http://link.springer.com/article/10.1007/s40745-015-0040-1.
13
| 8cs.DS
|
arXiv:1602.08025v1 [math.AC] 25 Feb 2016
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
BOGDAN ICHIM AND ANDREI ZAROJANU
A BSTRACT. In this paper we study the behavior of the size of a monomial ideal under
polarization and under generic deformations. As an application, we extend a result relating the size and the Stanley depth of a squarefree monomial ideal obtained by Herzog,
Popescu and Vladoiu, together with a parallel result obtained by Tang.
1. I NTRODUCTION
Let S = K[X1 , ..., Xn], with K a field, and let I ⊂ S be a monomial ideal. The notion of
size of a monomial ideal was introduced by Lyubeznik in [12]. In time, it has been used
by several authors, see for example [8], [9], [15], [16] and [17].
Several algebraic or combinatorial invariants associated to a monomial ideal are known
to have a nice behavior under polarization. For example, see [3], [5] or [10]. In the first
part of this paper we study the behavior of the size I under polarization. In Section 3 we
establish that
size I p ≤ size I + c,
where I p ⊂ S′ = K[X1 , ..., Xn′ ] is the polarization of I and c = n′ − n (see Theorem 3.7).
The equality does not hold in general, as shown in Example 3.9.
In the main result of this paper, that is Theorem 3.10, we provide a complete description
of the (particular) situation when the equality
size I p = sizeI + c
does hold.
A counterexample by H. Shen shows that the second statement of [8, Lemma 3.2] is
false when I is not squarefree. It follow that the proof [8, Theorem 3.1] is correct only
when I is squarefree, and that the statement of [8, Theorem 3.1] is in fact a conjecture
in general. As an application of out main result we deduce in Corollary 3.15 that this
conjecture is true under the conditions described in Theorem 3.10. In the same Corollary
and under the same conditions we also obtain an extension of [17, Theorem 3.2].
The notion of deformation of a monomial ideal was introduced by Bayer et al. [2] and
further developed in Miller et al. [13]. The most important deformations are the generic
deformations, which attracted the attention of several researchers, see for example [1] or
[11]. In the last part of this paper (Section 4) we briefly study the behavior of the size I
under generic deformations. We find that
size Iε ≤ size I,
where Iε is a generic deformation of I (see Proposition 4.4).
2010 Mathematics Subject Classification. Primary: 05E40; Secondary: 16W50.
Key words and phrases. monomial ideal; size; polarization; deformation.
1
2
BOGDAN ICHIM AND ANDREI ZAROJANU
2. P REREQUISITES
Let S = K[X1 , ..., Xn], with K a field. For n ∈ N we use the notation [n] := {1, . . ., n}.
2.1. The size of an ideal. In this Subsection we recall the definition of size and we make
some easy remarks; these will be needed in the sequel.
Let I ⊂ S be a monomial ideal and I =
s
T
Ui an irredundant primary decomposition of
i=1
I, with Ui monomial ideals. Let Ui be Vi -primary. Then each Vi is a monomial prime ideal
and Ass(S/I) = {V1 , . . . ,Vs }.
Definition 2.1. Following Lyubeznik [12, Proposition 2] the size of I, denoted in the
following by size I, is the number v + (n − h) − 1, where v is the minimum number t ≤ s
such that there exist i1 < · · · < it with
v
s
u t
s
u
t ∑ Ui = ∑ Ui ,
j
j=1
i=1
and where h = height ∑si=1 Ui .
A monomial ideal is irreducible if it is generated by powers of some variables. An
irreducible decomposition of a monomial ideal I is an expression I = I1 ∩. . .∩Ir with the I j
irreducible. Every irreducible monomial ideal is primary, so an irreducible decomposition
is a primary decomposition [7, Theorem 1.3.1].
q
p
Remark 2.2. Observe that ∑tj=1 Ui j = ∑tj=1 Vi j and ∑si=1 Ui = ∑si=1 Vi , so that the
size of I only depends on the set of associated prime ideals of S/I. Indeed, consider an
irredundant irreducible decomposition
I=
r
\
l=1
Ql , r ≤ s,
where Ql are monomial ideals and where Pl =
Ui =
ki
T
j=1
√
Ql . Then, for each i ∈ [s] we have that
Ql j where Pl j = Vi . Using the definition of size I we get the same result as above.
From now on we will only consider irredundant irreducible decompositions.
Remark 2.3. Working with size we can assume that ∑ri=1 Pi = m, where m is the maximal
ideal. Otherwise, set X = {X1, . . . , Xn }, Z = {Xk | Xk ∈
/ ∑si=1 Pi }, T = K[X \ Z] and let
J = I ∩ T . Then the sum of the associated prime ideals of J is the maximal ideal of T and
size I = sizeJ + |Z|.
Remark 2.4. Let I ⊂ S be a monomial ideal and let I =
r
T
i=1
Qi be an irredundant decom-
r
√
√
T
position of I as an intersection of irreducible ideals, where Qi = Pi . Then I =
Pi
i=1
√
and so we have that size I = size I. In general it easy to see that, if Ass S/I ⊆ Ass S/J
for two monomial ideals, then size I ≥ size J.
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
3
2.2. The polarization of a monomial ideal. In the following we study the behavior of
the size of a monomial ideal under polarization. We recall the definition of polarization
following Herzog and Hibi [7]. Let I ⊂ S be a monomial ideal with generators u1 , . . . , um ,
a
where ui = ∏nj=1 X j i j for i = 1, . . ., m. For each j let a j = max{ai j : i = 1, . . ., m}. Set
a = (a1 , . . ., an ) and S′ to be the polynomial ring
S′ := K[Xk,l : 1 ≤ k ≤ n, 1 ≤ l ≤ a j ].
Then the polarization of I is the squarefree monomial ideal I p ⊂ S′ generated by
v1 , . . . , vm , where
n
ai j
vi = ∏ ∏ Xk,l for i = 1, . . ., m.
k=1 l=1
The Stanley depth of an S-module M is a combinatorial invariant denoted in the following by sdepth M. We skip the details since this invariant will only appear briefly in
Corollary 3.15. For an excellent account on the subject, the reader is referred to Herzog’s
survey [6]. The following Theorem follows immediately from the main result of [10].
Theorem 2.5. Let I ⊂ S = K[X1 , ..., Xn] be a monomial ideal and I p ⊂ S′ = K[X1 , ..., Xn′ ]
be the polarization of I. Set c = n′ − n. Then
(1) sdepth I p = sdepth I + c;
(2) sdepth S p /I p = sdepth S/I + c.
Finally, we recall the most important known results relating sdepth and size.
Theorem 2.6. Let I be as squarefree monomial ideal of S.Then
(1) sdepth I ≥ size I + 1 (see [8, Theorem 3.1]);
(2) sdepth S/I ≥ size I (see [17, Theorem 3.2]).
For an extension of Theorem 2.6 see [4].
3. T HE
BEHAVIOR OF SIZE UNDER POLARIZATION
Let S = K[X1 , ..., Xn], with K a field, and I ⊂ S be a monomial ideal. Throughout the
section, we fix I =
r
T
Qi to be the unique irredundant irreducible decomposition of I. For
i=1
each i ∈ [r] we have
ai
ai
Qi = (X1 1 , . . ., Xn n ),
where aik ∈ N for all k ∈ [n]. In this writing we use the following convention: Assume Qi
ai
to be Pi -primary. Then, if Xk 6∈ Pi , we set aik = 0 and Xk k = 0.
n
Let ak = max{a1k , . . . , ark }, for all k ∈ [n]. Denote by n′ = ∑ ak and set c = n′ − n. Set
k=1
′
Let
Ip
⊂ S′
S := K[Xk,l : 1 ≤ k ≤ n, 1 ≤ l ≤ a j ].
p
be the polarization of I and Qi ⊂ S′ be the polarization of Qi . Then
p
I =
r
\
i=1
p
Qi
4
BOGDAN ICHIM AND ANDREI ZAROJANU
by [3, Proposition 2.3]. Moreover, by [3, Proposition 2.5], it holds
p
Qi =
\
(X1,b1 , . . ., Xn,bn )
1≤bk ≤aik
1≤k≤n
with the convention that, if aik = 0, then bk = Xk,0 = 0.
Definition 3.1. Remark that for all i ∈ [r] there exists at least one index k ∈ [n] such that
j
aik ≥ ak for all j ∈ [r]. We say that aik is a top power for Qi . For a subset N ⊆ [n] we set
BiN = {aik : aik is a top power for Qi and k ∈ N}. Note that Bi[n] 6= 0/ for all i ∈ [r]. For a
subset R ⊆ [r] we introduce the notion top base (denoted by C by the following recursive
algorithm). Let M ∈ Mr,n (Z) be the matrix with Mi j = aij , for all i ∈ [r], j ∈ [n].
Algorithm 1: Function which computes a top base
Data: r, n ∈ N, a Matrix M ∈ Mr,n (Z) and the Sets R ⊆ [r], N ⊆ [n]
Result: A List C containing a top base
1 Vector C = NewVector(r,0), R = [r], N = [n];
2 List BuildTopBase(M,R,N);
begin
3
if R = 0/ then
4
return C;
5
6
7
8
9
10
11
12
13
14
i = min(R);
List BiN = ReadTopPowers (i,N,M);
if BiN = 0/ then
Set R̃ = R \ {i};
return BuildTopBase(M,R̃,N);
else
for j=begin(N) to j=end(N) do
if aij = max(BiN ) then
C[i] ←− (aij , i, j));
Set R̃ = R \ {i};
Set Ñ = N \ { j};
return BuildTopBase(M,R̃, Ñ);
Below we describe the key steps.
• line 1. We initialize the Vector C to have 0 on all components and lenght r.
• line 6. We read the top powers from M on the line i. If we find any of these on the
columns j ∈ N then we include them in the list BiN .
• line 7. If there aren’t any top powers on line i and the columns j ∈ N from M then
we skip to the next line in M.
Remark that a top base is not unique and it depends on the choice of the maximal top
powers from each line, as the following example shows.
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
5
Example 3.2. Let I = (x10 , y10 , z) ∩ (x10 , y2 ) ∩ (x, z4 ). Then we have that
10 10 1
M = 10 2 0
1 0 4
where the top powers are boldfaced.
We see that {a1 , a2 , a3 } = {10, 10, 4} and we start the
top base. At the first step we can select c1 = a11 = 10.
✚ ✚
✚
10 10 1
10
10
✚
10 2 0 −→ ✚
✚
10 2
1 0 4
1✁ 0
algorithm above to compute a
1✁
0
4
Now in the second line we don’t have any top powers, thus c2 = 0 so we go to the third
line from where we get c3 = 4. So we obtained the top base {10, 0, 4}.
✚ ✚
✚ 1
✚ ✚
✚ 1
10
10
10
10
✁
✁
✚
✚
✚
✚ 2 0 −→ ✚
✚ 2
10
10
✁ 0
✁
1✁ 0 4
1✁ 0 4
Now, if we choose c1 = a12 we obtain a different top base {10, 10, 4}.
✚ ✚
✚ 1
✚ ✚
✚ 1
10 10 1
10
10
10
10
✁
✁
✚
✚
10 2 0 −→ 10 2✁ 0 −→ ✚
✚ 2
10
✁ 0
✁
1 0 4
1 0✁ 4
1✁ 0✁ 4
Definition 3.3. Consider the top base C from the above algorithm and set ci = C[i][1], i ∈
[r]. In the following we denote
Qi,p j = (X1,min{ j,ai } , . . . , Xn,min{ j,ain } ) for j ∈ [ci ],
1
where Xi,0 = 0 and if ci = 0, then considerQi, j = (1) in intersections of ideals andQi, j = (0)
in sums of ideals,
p
Qi =
ci
\
j=1
p
Qi, j = (X1,1, . . . , Xn,1 ) ∩ (X1,2, . . . , Xn,2) ∩ . . . ∩ (X1,min{ci ,ai } , . . . , Xn,min{ci ,ain } )
1
and
Ip =
r
\
p
Qi .
i=1
Remark 3.4. Following Algorithm 1 we have that {c1 , . . . , cr } \ {0} ⊂ {a1 , . . . , an }. Thus
r
n
r
i=1
k=1
i=1
∑ ci ≤ ∑ ak , which implies that ∑ ci ≤ c + r.
q
q
w+1
Q jk = ∑rj=1 Q j }, w = size I .
Definition 3.5. Set A = { j1 , . . . , jw+1 } such that ∑k=1
We also use the notations |A| = m and A = ∪m
h=1 Ah .
6
BOGDAN ICHIM AND ANDREI ZAROJANU
Remark 3.6. We may suppose (eventually after renumbering the ideals Qi ) that we always
have A1 = {1, 2, . . ., w + 1} ∈ A, where w = size I.
Theorem 3.7. size I p ≤ size I + c.
Proof. From the definition of size we see that size Qip = sizeQip and that size I p = sizeI p,
where
Ip=
r
\
p
Qi =
i=1
Qip j,
\
1≤i≤r
1≤ j≤ci
r
therefore I p has an irredundant decomposition composed of D = ∑ ci irreducible monoi=1
mial ideals.
s
w+1
p
According to Remark 3.6 we can assume that Q j ⊂
∑ Qi , for all t + 2 ≤ j ≤ r
i=1
s
w+1
p
and set a = r − (w + 1). We see that Q j ⊂
∑ Qi , for all w + 2 ≤ j ≤ r implies that
i=1
p
w+1
p
Q j ⊂ ∑ Qi 1, for all w + 2 ≤ j ≤ r.
1
i=1
Then
(3.1)
size I p = sizeI p = size
"
\
1≤i≤w+1
1≤ j≤ci
p
Qi
!
j
Notice that the last term has D − a intervals. Then
1 − 1 ≤ c + w = size I + c.
∩
p
\
Qi
!#
.
j
w+2≤i≤r
2≤ j≤ci
p
size I ≤ D − a − 1
= D−r +w+
r
Remark 3.8. We see from equation 3.1 that size I p ≤ size I + ∑ (ci − 1) ≤ sizeI + c.
i=1
ci >0
Thus, if {ai : ai > 1, i ∈ [n]} * {c1 , . . ., cr }, we have that
size I p ≤ size I +
r
n
i=1
ci >0
i=1
∑ (ci − 1) < size I + ∑ (ai − 1) = size I + c.
Example 3.9. Let I = (X12, X22) ∩ (X32, X42 ) = Q1 ∩ Q2 ⊂ K[X1 , X2 , X3, X4 ] a monomial
ideal. Then size I p = 3 < size I + c = 5. Indeed we see that size I = 1 and that c = 4.
Also note that
p
p
p
p
size I = sizeQ1 1 ∩Q1 2 ∩Q2 1 ∩Q2 2 = 3.
Let I ⊂ S = K[X1 , . . . , Xn] be a monomial ideal and let I =
r
T
Qi an irredundant decom-
i=1
ai
ai
position of I as an intersection of irreducible ideals, where the Qi = (X1 1 , . . . , Xn n ), aij ∈
N, j ∈ [n] are monomial ideals. Let Qi be Pi -primary. For every k ∈ [n], we set:
Tk = {t : Xks ∈ G(Qt ), s ≥ 2};
Lk = {l : Xk ∈ Pl };
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
7
Uk = {u : Xk ∈ Qu , max(Bu[n] ) > 1}.
We may now formulate the main result of this paper, which fully describes the (particular) cases when the equality holds.
Theorem 3.10. Let I ⊂ S be a monomial ideal. Then size I p = size I + c if and only if
aii
Qi =< {Xi1 1 , Xi2 , . . . Xis }| s ∈ [n], ai1 ≥ 1 > for all i ∈ [r] and, if Xk ∈ Pi ∩ Pj , 1 ≤ i < j ≤
n, k ∈ [n], then one of the following is true:
(1) Tk = 0/ or
(2) Tk 6= 0/ and
(A) ∀t ∈ Tk , {t, l} * Ah , ∀ l ∈ Lk \ {t}, ∀ h ∈ [m] and
(B) if there exists t ∈ Tk ∩ Ah for some h ∈ [m], then Uk = 0.
/
Proof. ”⇐=”
ai
(1) Let Qi = (Xi i , Xi1 , . . . , Xit ), i ∈ [r], ai > 1, t ∈ [n]. If such a Qi does not exist, then
c = 0 and I = I p . Thus we have that Xi ∈
/ Pj for all j ∈ [r] \ {i}, thus i ∈ Ah , for all h ∈ [m].
We may assume that Qi , i ∈ [s] have a top power ≥ 2 and that Qi , i > s are generated by
variables, that is max(Bi[n] ) = 1, for all i > s. Then the ideals in the second intersection
from the last term in the equation 3.1 are generated by variables. Then we get that
s
n
i=1
i=1
size I p = ∑ ci + w − s = ∑ (ai − 1) + w = size I + c,
because for all j ∈ [aii ] and i ∈ [s] we only have the ideal (Xi, j , Xi1,1 , . . . , Xit ,1 ) to cover Xi, j .
(2) Let size I = w. Consider that Qw+2 , . . . , Qs have top powers ≥ 2 and that Qs+1 , . . ., Qr
have top powers = 1. Then according to algorithm 1 we have that ci ≤ 1, for all i > s and
using equation 3.1 we see that for computing size I p we don’t need the ideals Qi , i > s.
Moreover, we shall consider only the ideals Qi , w + 1 < i ≤ s with ci > 1. Thus we may
suppose that ci > 1, w + 1 < i ≤ s.
j
If, for example, we have that C[i] = (ai , i, i), i ≥ w + 2, we show that in the sum of the
other intervals we can cover only one variable from {Xi,1, . . ., Xi,a j } and that variable is
i
j
Xi,1 from our choice in definition 3.3. Indeed, if we have xti ∈ G(Q1 ), t < ai , then c1 can
be at most 1, so we cover Xi,1. Condition (A) tells us that xi ∈
/ Pj , 1 < j ≤ w+1. Condition
j
j
d
(B) tells us that if we have, for example, C[1] = (a1 , 1, 1) and C[i] = (ai , i, i), i 6= 1, ai > 1,
then x1 ∈
/ G(Qi ), so that we can not cover X1,1 in the sum of the other intervals, that is
p
j
Qi l, 2 ≤ l ≤ ai .
It follows that
size I p =
w+1
∑
i=1
s
ci +
∑
w+1
(ci − 1) − 1 = size I +
i=w+2
∑ (ci − 1) +
i=1
s
∑
(ci − 1) = size I + c.
i=w+2
”=⇒”
Assume that size I p = size I + c. Then Remark 3.8 gives us the following inclusion
{ai : ai > 1, i ∈ [n]} ⊂ {c1 , . . ., cr }. Let size I = w.
8
BOGDAN ICHIM AND ANDREI ZAROJANU
a1
a1
j
First suppose that X1 1 , X2 2 ∈ G(Q1 ), a1 > 1, j ∈ [2]. If c1 > a11 , then using Remark
j
3.8 we get that there exists 1 < j ≤ r such that c j = a1 = a1 . Then from equation 3.1
p
p
we can skip the ideal Q j 2 because we have X1,2 ∈ Q1 2 and we find all the other variables
p
X p,2 , 2 ≤ p ≤ n in Qw 2, where cw = a p . We will call this procedure the absence of variable
X1,2 . Thus we get that size I p < size I + c.
Now, if c1 = a11 = a1 let 2 ≤ j ≤ r such that c j = a2 . Then again, using the absence of
j
variable X2,2 , we get that size I p < size I + c. Thus we see that ai = 1 except for at most
one i ∈ [n], for all j ∈ [r].
a1
a2
Suppose that {1, 2} ⊂ A1 , X1 1 ∈ G(Q1 ), X1 1 ∈ G(Q2 ), a11 > 1, a21 ≥ 1, a11 ≥ a21 . If
c1 = a11 , then
!
!#
"
size I p ≤ size
≤ size
∑
i=2
Qi
\
Qip j
j
1≤i≤w+1
1≤ j≤ci
"
2≤i≤w+1
1≤ j≤ci
w+1
≤
p
\
!
∩
∩
p
\
Qi
\
Qip j ∩Q1p 2 . . . ∩Q1p a1
1
j
w+2≤i≤r
2≤ j≤ci
w+2≤i≤r
2≤ j≤ci
!#
r
ci +
∑
i=w+2
(ci − 1) + c1 − 1
= size I − 1 + c.
p
p
p
p
We skipped the ideal Q1 1 because Q1 1 ⊂ Q1 2 + ∑w+1
j=2 Q j 1.
j
If c1 > a11 , then there exists 2 ≤ j ≤ r such that c j = a1 = a1 > 1 and in equation 3.1
w+1
we can skip over the ideal Q pj because Q pj ⊂ Q pj + ∑ Qip 1, thus size I p ≤ size I + c − 1.
1
1
2
i=1
i6= j
a1
au
Now suppose that 1 ∈ A1 , X1 1 ∈ G(Q1 ), a11 > 1, X1 , X2 2 ∈ G(Qu ) and cu = au2 > 1. As
p
p
p
we have seen above, we may assume that c1 = a11 . Then we have Q1 1 ⊂ ∑w+1
j=2 Q j 1 +Qu 2,
thus we get size I p ≤ size I + c − 1.
Example 3.11. Consider the monomial ideal I = (X12, X2 )∩(X2 , X3)∩(X3 , X4)∩(X2 , X4 ) =
4
T
i=1
Qi ⊂ K[X1 , X2 , X3 , X4]. Then A = {{1, 2, 3}, {1, 2, 4}} and c = 1, thus size I = 2. Us-
ing Theorem 3.10 we get that size I p = size I + c. Indeed, size I p = size(X1,1, X2,1 ) ∩
(X1,2 , X2,1) ∩ (X2,1 , X3,1) ∩ (X3,1, X4,1 ) ∩ (X2,1, X4,1 ).
Example 3.12. Let I = (X12, X2 ) ∩ (X1 , X3 ) ∩ (X2 , X3) =
3
T
i=1
Qi ⊂ K[X1 , X2, X3 ]. Then A =
{{1, 2}, {1, 3}, {2, 3}} and c = 1, thus size I = 1. We have that size I p = size(X1,2 , X2,1) ∩
(X1,1 , X3,1) = sizeQ1p 2 ∩Q2p = 1, thus size I p < sizeI + c. We see that I does not respect
condition (2)(A) from Theorem 3.10.
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
Example 3.13. Let I = (X12, X2 ) ∩ (X3, X4) ∩ (X1, X42) =
9
4
T
Qi ⊂ K[X1 , X2, X3 , X4]. Then
i=1
A = {{1, 2}} and c = 2 thus size I = 1. We have that size I p = size(X1,2 , X2,1)∩(X3,1 , X4,1 )∩
(X1,1 , X4,2) = size Q¯1p 2 ∩ Q¯2p ∩ Q¯3p 2 = 2, thus size I p < size I + c. We see here that X4 re-
spects all the conditions from Theorem 3.10, but X1 does not respect the condition (2)(B).
Example 3.14. Let I = (X1k+1 , X2k ) ∩ (X1, X2k+1 ) ⊂ K[X1 , X2 ]. Then size I p = size I + c − k.
Indeed, we have size I = 0, c = 2k and size I p = size(X1,2, X2,1 ) ∩ . . . ∩ (X1,k+1 , X2,k ) ∩
(X1,1 , X2,k+1 ). We see that each variable appears only once, thus the size can not be
smaller.
As an application of our main result, by using Theorem 3.10 and Theorem 2.5, we
easily deduce the following extension of Theorem 2.6.
Corollary 3.15. Let I be a monomial ideal of S such that, either I is squarefree, or I is as
described in Theorem 3.10. Then
(1) sdepth I ≥ size I + 1;
(2) sdepth S/I ≥ size I.
4. T HE
BEHAVIOR OF SIZE UNDER GENERIC DEFORMATIONS
The notion of deformation of a monomial ideal was introduced by Bayer et al. [2] and
further developed in Miller et al. [13].
Definition 4.1. (1) Let M = {m1 , . . . , mr } ⊂ S be a set of monomials. For 1 ≤ i ≤ r let
ai = (ai1 , . . . , ain ) ∈ Nn denote the exponent vector of mi . A deformation of M is a set of
vectors εi = (ε1i , . . . , εni ) ∈ Nn for 1 ≤ i ≤ r subject to the following conditions:
(4.1)
aij > akj =⇒ aij + ε ij > akj + ε kj
and
aik = 0 =⇒ εki = 0.
(2) Let I ⊂ S be a monomial ideal with generating set GI . A deformation of I is a
deformation of GI . We set Iε := (g · xεg : g ∈ GI ) to be the ideal generated by the deformed
generators.
The most important deformations are the generic deformations. Let us recall the definition from [13].
Definition 4.2. (1) A monomial m ∈ S is said to strictly divide another monomial m′ ∈ S
′
if m | mXi for each variable Xi dividing m′ .
(2) A monomial ideal I ⊂ S is called generic if for any two minimal generators m, m′ of
I having the same degree in some variable, there exists a third minimal generator m′′ that
strictly divides lcm(m, m′ ).
(3) A deformation of a monomial ideal I is called generic if the deformed ideal Iε is
generic.
Definition 4.3. [14] A monomial ideals is considered generic if for any m1 , m2 ∈ Min (I)
and any variable Xi if degXi (m1 ) = degXi (m2 ), then degXi (m1 ) = 0. That is, no minimal
generator of I has the same exponent for any variable.
10
BOGDAN ICHIM AND ANDREI ZAROJANU
r
T
Proposition 4.4. Let I ⊂ S = K[X1 , . . ., Xn ] be a monomial ideal and let I =
Qi an
i=1
√
irredundant decomposition of I as an intersection of irreducible ideals, where Qi = Pi .
Let Iε be the generic deformation of I. Then size I ≥ size Iε .
Proof. We follow the proof of [14] Theorem 6, except we consider any generic transformation, not just the generic ones. By construction, we have an unique map ϕ from the
individual powers of variables amongst the generators of I to individual powers of variables amongst the generators of Iε . Let Iε =
r
T
Qi be the unique irredundant irreducible
i=1
primary decomposition of Iε , then by [14] Theorem 6 we have that I =
r
T
ϕ (Qi ) is an
i=1
irreducible primary decomposition ofq
I. The decomposition may not be irredundant.
p
If we consider sizeIε = t < r with ∑tj=1 Qi j = ∑ri=1 Qi . then we also have
v
s
s
u t
r
r
u
t ∑ ϕ (Qi ) = ∑ ϕ (Qi ) = ∑ Qi .
j
j=1
i=1
i=1
Thus we get the inequality size I ≥ size Iε , because the decomposition for I may be redundant.
Example 4.5. Let I = (xyt, xyw, xtw, yzt, yzw, ztw) ⊂ K[z, y, z,t, w] and let ε1 be a deformation and ε2 a generic deformation such that Iε1 = (xyt, xyw, xt 3 w, yzt 2, yzw, ztw) and
Iε2 = (x3 y4t, x2 y2 w, xt 3 w3 , y3 zt 2, yz2 w2 , z3tw4 ). Then we have the following unique irreducible primary decompositions:
I =(x, z) ∩ (y,t) ∩ (y, w) ∩ (t, w);
Iε1 =(x, z) ∩ (y,t) ∩ (y, w) ∩ (t, w) ∩ (x,t 2, w) ∩ (y, z,t 3);
Iε2 =(x, z) ∩ (y,t) ∩ (y3, y2 w, yw2 , w3 ) ∩ (t, w) ∩ (y4, y3t 2, y2 w, yw2 ,t 3, w4 )∩
(x3 , x2 w,t 2, w2 ) ∩ (y4 , y3 z, y2 w, z2 , w3 ) ∩ (x3 , x2 w, z, w3 ) ∩ (x2 , xw3 , y3 , yw2 , w4 )∩
(y2 , yz2 , z3 ,t 3) ∩ (x2 , z2, zt 2,t 3 ) ∩ (x, y3 , yz2 , z3 ) ∩ (x3 , x2 y2 w, x2 z2 , y3 , yz2 w2 , z3 , w3 )∩
(x3 , x2 y2 , y3 , yz2 , z3 ,t 3).
We have that 2 = size I > size Iε1 = size Iε2 = 1, thus the inequality from the above proposition may be strict.
ACKNOWLEDGEMENTS
They were partially supported by project PN-II-RU-TE-2012-3-0161, granted by the
Romanian National Authority for Scientific Research, CNCS - UEFISCDI, during the
preparation of this work.
R EFERENCES
[1] J. Apel, On a conjecture of R. P. Stanley; Part I–Monomial Ideals. J. Algebr. Comb. 17 (2003), 39–56.
ON THE BEHAVIOR OF THE SIZE OF A MONOMIAL IDEAL
11
[2] D. Bayer, I. Peeva and B. Sturmfels, Monomial resolutions. Mathematical Research Letters 5 (1998),
31–46.
[3] S. Faridi, Monomial ideals via square-free monomial ideals. Lecture Notes in Pure and Applied Mathematics 244 (2005), 85–114.
[4] S. Fakhari, On the Stanley depth and size of monomial ideals. Preprint, arXiv:1503.03736.
[5] V. Gasharov, I. Peeva and V. Welker, The LCM-lattice in monomial resolutions. Mathematical Research Letters 6 (1999), 521–532.
[6] J. Herzog, A survey on Stanley depth. In Monomial Ideals, Computations and Applications, Springer
LMS 2083 (2013), 3–45.
[7] J. Herzog and T. Hibi, Monomial ideals. Springer GTM 260 (2011).
[8] J. Herzog, D. Popescu and M. Vladoiu, Stanley depth and size of a monomial ideal. Proc. Am. Math.
Soc. 140 (2012), 493–504.
[9] J. Herzog, Y. Takayama and N. Terai, On the radical of a monomial ideal. Archiv der Mathematik 85
(2005), 397–408.
[10] B. Ichim, L. Katthän and J.J. Moyano-Fernández, The behavior of Stanley depth under polarization.
Journal of Combinatorial Theory, Series A, 135 (2015), 332–347.
[11] B. Ichim, L. Katthän and J.J. Moyano-Fernández, Stanley depth and the lcm-lattice. Preprint,
arXiv:1405.3602.
[12] G. Lyubeznik, On the Arithmetical Rank of Monomial ideals. J. Algebra 112 (1988), 86–89.
[13] E. Miller, B. Sturmfels and K. Yanagawa, Generic and cogeneric monomial ideals. J. of Symbolic
Computation 29 (2000), 691–708.
[14] A. Milowski, Computing irredundant irreducible decompositions and the scarf complex of large scale
monomial ideals. Thesis, San Francisco State University (2004).
[15] D. Popescu, The Stanley Conjecture on Intersections of Four Monomial Prime Ideals. Communications in Algebra 41 (2013), 4351–4362.
[16] Y. H. Shen, Stanley depth and Stanley support-regularity of monomial ideals. Collectanea Mathematica, in press (2015).
[17] Z. Tang, Stanley depths of certain Stanley-Reisner rings. J. Algebra 409 (2014), 430–443.
S IMION S TOILOW I NSTITUTE OF M ATHEMATICS OF
5, P.O. B OX 1-764, 014700 B UCHAREST, ROMANIA
E-mail address: [email protected]
THE
ROMANIAN ACADEMY, R ESEARCH U NIT
S IMION S TOILOW I NSTITUTE OF M ATHEMATICS OF
5, P.O. B OX 1-764, 014700 B UCHAREST, ROMANIA
E-mail address: andrei [email protected]
THE
ROMANIAN ACADEMY, R ESEARCH U NIT
| 0math.AC
|
Fundamental Limits on Latency in Transceiver
Cache-Aided HetNets
Jaber Kakar, Soheil Gherekhloo and Aydin Sezgin
Institute of Digital Communication Systems
Ruhr-Universität Bochum, 44780 Bochum, Germany
Email: {jaber.kakar, soheyl.gherekhloo, aydin.sezgin}@rub.de
hr /
nr
UE
RN
µN L
arXiv:1707.09029v1 [cs.IT] 27 Jul 2017
In recent years, mobile usage characteristics in wireless networks have changed profoundly from conventional connectioncentric (e.g., phone calls) to content-centric (e.g, HD video)
behaviors. As a result, content caching and heterogenous network (HetNet) technology are two major solutions for next
generation (5G) mobile networks. Advantages of these solutions
are two-fold: Firstly, caching the most popular contents in
network edges, e.g., base stations and relays, alleviates backhaul
traffic and reduces latency. Secondly, seamless uniform quality
of service through improved performance, particularly, at cell
edges is guaranteed. It is therefore to be expected that future
networks will be heterogenous in nature, vastly deploying relay
nodes (RN) (e.g., in LTE-A [1]) endowed with (adaptive) cache
capabilties. A simplistic HetNet modeling this aspect is shown
in Fig. 1. In this model, the RN acts as a cache-aided transceiver.
Thus, aspects of both transmitter and receiver caching in a
single RN is captured through this network model enabling
low-latency transmission and reception of requested files by
RN and the user equipment (UE). In this work, we focus
on characterizing the fundamental trade-off on latency of this
particular network.
Various works showed that both receiver (Rx) and transmitter
(Tx) caching offer great potential for reducing latency. Rx
caching, on the one hand, was first studied in [2] for a shared
link with one server and multiple cache-enabled receivers. The
authors show that caching can exploit multicast opportunities
and as such significantly reduces the delivery latency over the
shared link. On the other hand, the impact of Tx caching
on the latency has mainly been investigated by analyzing the
latency-centric metric; the inverse degrees-of-freedom (DoF),
for Gaussian networks. To this end, the authors of [3] developed
a novel achievability scheme characterizing the metric as a
hd /nd
DeNB
s
I. I NTRODUCTION
Cloud
h s/n
Abstract—Stringent mobile usage characteristics force wireless networks to undergo a paradigm shift from conventional
connection-centric to content-centric deployment. With respect
to 5G, caching and heterogenous networks (HetNet) are key
technologies that will facilitate the evolution of highly contentcentric networks by facilitating unified quality of service in terms
of low-latency communication. In this paper, we study the impact
of transceiver caching on the latency for a HetNet consisting of
a single user, a receiver and one cache-assisted transceiver. We
define an information-theoretic metric, the delivery time per bit
(DTB), that captures the delivery latency. We establish coinciding
lower and upper bounds on the DTB as a function of cache
size and wireless channel parameters; thus, enabling a complete
characterization of the DTB optimality of the network under study.
As a result, we identify cache beneficial and non-beneficial channel
regimes.
Index Terms—Transceiver caching, latency, 5G, information
theory, relay
Fig. 1: System model of cloud-aided HetNet for (a) Gaussian and (b)
Linear Deterministic model
function of the cache storage capability for a 3-user Gaussian
interference network. The cache placement was designed to
facilitate transmitter cooperation such that interference coordination techniques can be applied. A converse on this metric was
developed in [4] for a network with arbitrary number of edge
nodes and users showing the optimality of schemes presented in
[3] for certain regimes of cache sizes. Extensions of this work
include the characterization of the latency-memory tradeoff
to cloud and cache-assisted networks [5]. Two new lines of
research are to determine the fundamental limits of joint TxRx caching at distinct nodes [6] and transceiver caching. This
paper focuses on the latter.
In this paper, we are studying the fundamental limits on the
latency for a transceiver cache-aided HetNet consisting of a
donor eNB (DeNB), a transceiver and a user. We measure the
performance through the latency-centric metric delivery time
per bit (DTB). To this end, we establish coinciding lower
(converse) and upper bounds (achievability) on the DTB for the
linear deterministic model (LDM) [7] of our proposed network
for various channel regimes. Hereby, the LDM serves as an
approximation of the additive Gaussian noise model.
Notation: For any two integers a and b with a ≤ b, we
define [a : b] , {a, a + 1, . . . , b}. When a = 1, we simply
write [b] for {1, . . . , b}. The superscript (·)† represents the
transpose of a matrix. Furthermore, we define the function
(x)+ , max{0, x}.
II. S YSTEM M ODEL
We study the downlink of a cache-aided HetNet as shown
in Fig. 1. The HetNet consists of a causal full-duplex RN and
a macro donor eNB (DeNB) which serve a single user – a
small-cell user (UE) – over a wireless channel. Simultaneosuly,
the RN also acts as a user requesting information from the
DeNB. At every transmission interval, we assume that RN and
UE request files from the set W of N popular files, whose
elements are all of L bits in size. The transmission interval
terminates when the requested files have been delivered. The
system model, notation and main assumptions for a single
transmission interval are summarized as follows:
• Let W = {W1 , . . . , WN } denote the library of popular files,
where each file Wi is of size L bits. Each file Wi is chosen
uniformly at random from [2L ]. RN and UE request files Wdr
and Wdu from the library W, respectively. The demand vector
d = (dr , du )† ∈ [N ]2 denotes the request pattern of RN and
UE.
• The RN is endowed with a cache capable of storing µN L
bits, where µ ∈ [0, 1] corresponds to the fractional cache size.
It denotes how much content can be stored at the RN relatively
to the entire library W.
• The cloud server has access to all N files. The DeNB is
connected to the cloud via a fronthaul link of infinite capacity
CF = ∞ bits per channel use.
• Global channel state information (CSI) for a single transmission interval is summarized by the channel vector h =
(hd , hr , hs )† ∈ C3 , where on the one hand hd and hs represent
the complex channel coefficients from DeNB to the UE and RN,
respectively. On the other hand, hr is the channel from RN to
UE. We assume that all channel coefficients are assumed to be
drawn i.i.d. from a continuous distribution and are kept fixed
over a transmission interval.
Communication over the wireless channel occurs in two consecutive phases, placement phase followed by the delivery phase.
In the following, we will describe the modeling of placement
and delivery phase and the key performance metric termed as
delivery time per bit (DTB) formally.
1) Placement phase: During this phase, the RN is given full
access to the database of N files. The cached content at the
RN is generated through its caching function.
Definition 1. (Caching function) The RN maps each file Wi ∈
W to its local file cache content
Si = φi (Wi ),
∀i = 1, . . . , N.
All Si are concatenated to form the total cache content
S = (S1 , S2 , . . . , SN )
at the RN. Hereby, due to the assumption of symmetry in
caching, the entropy H(Si ) of each component Si , i =
1, . . . , N , is upper bounded by µN L/N = µL. The definition of
the caching function presumes that every file Wi is subjected to
individual caching functions. Thus, permissible caching policies
allow for intra-file coding but avoid inter-file coding. Moreover,
the caching policy is typically kept fixed over multiple transmission intervals. Thus, it is indifferent to the user’s request
pattern and of channel realizations.
2) Delivery phase: In this phase, a transmission policy at DeNB
and RN is applied to satisfy the given user’s requests d under
the current channel realizations h.
Definition 2. (Encoding functions) The DeNB encoding function
ψS : [2N L ] × [N ]2 × C3 → CT
determines the transmission signal xTs subjected to an average
power constraint of P . Hereby, the codeword xTs is a function
of W, d and h conveyed to both RN and UE over T channel
uses. The encoding function of the causal full-duplex RN at the
t–th time instant is defined by
ψr[t] : [2µN L ] × Ct−1 × [N ]2 × C3 → C,
t ∈ [1 : T ].
[t]
ψr
For any time instant t,
accounts for the simultaneous
reception and transmission through incoming and outgoing
wireless links at the RN. To be specific, at the t–th channel
[t]
use the encoding function ψr maps the cached content S, the
received signal yrt−1 (see Eq. (2)), the demand vector d and
global CSI given by h to the codeword xr [t] while satisfying
the average power constraint given by the parameter P .
Definition 3. (Decoding functions) The decoding operation at
the UE follows the mapping
ηu : CT × [N ]2 × C3 → [2L ].
The decoding function ηu takes as its arguments h, the available
demand pattern d and the channel outputs yuT given by
yuT = hd xTs + hr xTr + zTu
(1)
to provide an estimate Ŵdu = ηu yuT , d, h of the requested
file Wdu . The term zTu denotes complex i.i.d. Gaussian noise of
zero mean and unit power. In contrast to decoding at the UE,
the RN explicitly leverages its cached content according to
ηr : CT × [2µN L ] × [N ]2 × C3 → [2L ]
to generate an estimate Ŵdr = ηr yrT , S, d, h on the requested
file Wdr . Hereby, yrT is the received signal
yrT = hs xTs + zTr
(2)
corrupted through additive zero mean, unit-power i.i.d. Gaussian noise zTr at the RN in T channel uses.
A proper choice of a caching, encoding and decoding function
that satisfies the reliability condition; that is, the worst-case
error probability
Pe = max2 max P(Ŵdj 6= Wdj )
d∈[N ] j∈{r,u}
(3)
approaches 0 as L → ∞, is called a feasible policy.
Definition 4. (Delivery time per bit) [4] The DTB for given
request pattern d and channel realization h is defined as
∆(µ, h, P ) = max2 lim sup
d∈[N ]
L→∞
T (d, h)
.
L
(4)
The minimum DTB ∆∗ (µ, h, P ) is the infimum of the DTB of
all achievable schemes.
Remark 1. The DTB measures the per-bit latency, i.e., the
latency incurred when transmitting the requested files through
the wireless channel, within a single transmission interval for
the worst-case request pattern of RN and UE.
To gain initial insight into the DTB for the Gaussian system
model, we suggest to approximate (1) and (2) by the lineardeterministic model (LDM) [7]. In the LDM, input symbols at
the DeNB and RN are given by the binary input vectors xs
and xr ∈ Fq2 where q = max{nd , nr , ns }. Hereby, the integers
nk ∈ N>0 , k ∈ {d, r, s}, given by
nk = dlog P |hk |2 e
(5)
approximate the number of bits per channel use which can
be communicated over each link reliably. The channel output
symbols yrT and yuT received in T channel uses at the UE and
RN are given by deterministic functions of the inputs; that is,
yrT = Sq−ns xTs ,
yuT
where S ∈
Fq×q
2
q−nd
=S
xTs
(6a)
q−nr
⊕S
xTr ,
is a down-shift q × q matrix defined by
†
0
0q−1
.
S=
Iq−1 0q−1
(6b)
(7)
The input-output equation (6) approximates the input-output
equation of the Gaussian channel given in (1) and (2) in the high
SNR regime. A graphical representation of the transmitted and
received binary vectors xl [i] and yj [i], l ∈ {r, s}, j ∈ {r, u},
in the i-th channel use is shown in Fig. 3a. Each sub-block of
length (rate) R in the figure represents a sub-signal being able
to hold R bits for transmission. For any wireless link of the
network under study, only the most n = (nd , nr , ns )† significant bits are received at the destinations while less significant
bits are not. We denote the DTB for the LDM by ∆det (µ, n).
The remainder of this paper focuses on characterizing the DTB
on the basis of the LDM.
III. MAIN RESULT
In this section, we state our main results on the optimal
DTB ∆∗det (µ, n) for various channel regimes. This results in
Theorems 1 and 2. To state these theorems we define the
following disjoint channel regimes {Ci }4i=1 :
C1 = nd ≥ max{nr , ns }
C = n ≥ n ≥ n
2
d
s
r
.
(8)
C
=
min{n
,
n
}
≥ nd
3
r
s
C 4 = ns ≥ nd ≥ nr
First, we provide the following lemma that explains sub-channel
regimes derived from Ci that are utilized in the theorems in
greater detail.
Lemma 1. For the LDM-based cache-aided HetNet in Fig. 1
with µ = 0, the optimal DTB is given by
2
1
1
∗
, ,
.
∆det (µ = 0, n) = max
max{nd , ns } ns max{nd , nr }
(9)
Proof: For µ = 0, the RN has no relevant information on
its own and on the UE’s requested file Wdr and Wdu . Thus, the
DeNB is involved in broadcasting files Wdr and Wdu to RN
and the UE while the RN acts as a transceiver. The resulting
channel constitutes a broadcast channel with one-sided noisy
receiver cooperation. For a feasible scheme, either user can
reliably decode Wdj (Wdr and Wdu ), j = {r, u}, if it is
aware of yjT (Sq−max{nd ,ns } xTs ). These observations can be
used to generate lower bounds on the DTB ∆∗det (µ = 0, n)
that correspond to elements in the outer max-expression of (9).
Since the requested files are of the same size, an optimal scheme
that minimizes the latency would try to split the transmission
load equally to the UE and RN. Depending on the channel
conditions, the load balancing is done as follows. For instance,
when nd ≥ ns , nd/2 bits can be send in one channel use
from the DeNB to both RN and UE, if the weaker channel
ns is stronger than nd/2, i.e., nd/2 ≤ ns . On the other hand
if nd ≥ ns and nd/2 ≥ ns , the latency is governed by the
weaker channel and ns bits are transmitted in one channel use
to each user. In summary, for nd ≥ ns we can reliably convey
min{nd/2, ns } bits per channel use to each user, or in other
words, ∆∗det (µ = 0, n) = max{2/nd , 1/ns } channel uses are
needed to provide each user with one bit. For nd ≤ ns , a
similar observation holds. Details for this case are omitted to
conserve space. This concludes the proof.
Remark 2. It is easy to verify from Lemma 1 that the
optimal DTB corresponds
to 2/nd (2/ns ) for I0 = 2n
s ≥
nd ≥ ns
I1 = 2 max{nd , nr } ≥ ns ≥ nd
and
C
1/ns (1/max{nd ,nr }) for I C =
to
2n
≤
n
}
I
=
s
d
0
1
2 max{nd , nr } ≤ ns .
Now we are able to state the theorems. We make the
distinction between the cases when the direct channel is either
weaker or stronger than the DeNB-RN channel (nd ≥ ns and
nd ≤ ns ).
Theorem 1. (nd ≥ ns ) For the LDM-based cache-aided HetNet
in Fig. 1 with µ ∈ [0, 1], the optimal DTB in channel regimes
R1 = C1 ∩ I0
R0 = C ∩ I C
1
1
0
,
(10)
R
=
C
∩
I
2
2
0
0
R2 = C2 ∩ I0C
is given by
2−µ
nd n
o
2−2µ 2−µ
∗
∆det (µ, n) = max nd , nr
n
2−µ
max 1−µ ,
ns
max{nd ,nr }
for n ∈ R1
for n ∈ R2
o
for n ∈
.
R01 , R02
(11)
Theorem 2. (nd ≤ ns ) For the LDM-based cache-aided HetNet
in Fig. 1 with µ ∈ [0, 1], the optimal DTB in the channel
regimes
R31 = C3 ∩ {nd ≤ ns ≤ 2nd }
R32 = C3 ∩ {2nd ≤ ns ≤ 2nr }
,
(12)
R03 = C3 ∩ I1C
R4 = C4 ∩ I1
R0 = C ∩ I C
4
4
1
corresponds to
∆∗det (µ, n) =
n
o
2−µ
2−2µ
max
,
,
∆
(n)
+
LB
n ns ns +(nr −n
o d)
max 2−2µ , ∆ (n)
LB
n ns
o
2−µ
max
,
∆
(n)
+
LB
ns +(nr −nd )
∆LB (n)
for n ∈ R31
for n ∈ R32
for n ∈ R4
for n ∈ R03 , R04
(13)
where ∆LB (n) = 1/max{nd ,nr }.
Proof: (Theorems 1 and 2) The lower bound (converse)
on the DTB is provided in section IV. The upper bound
(achievability) can be found in section V.
Remark 3. In all channel regimes of (10) and (12) (except
of R03 and R04 ), caching at the RN decreases the latency
with increasing fractional cache size (cf. Fig. 2) until the best
possible latency ∆∗det (1, n) = ∆LB (n) is achieved at some
µ00 (n). This latency is identical to the optimal DTB when the
,
the information subsets, when using B1 –B5 we arrive at the
following five inequalities bounding T :
2
ns
A
DTB – ∆∗det (µ, n)
µ0 (n) =
00
µ (n) =
nr −nd
ns/2+nr −nd
T ≥
nd +nr −ns
nr
T ≥
C
1
nr
µ0 (n)
0
µ00 (n)
T ≥
1
ns/2+nr −nd
B
1
µ
Fig. 2: DTB as a function of µ for n ∈ R31
RN acts as an additional transmitter with access to the entire
library of files (µ = 1). Interestingly, our results reveal that
for nd ≤ ns only, a fractional cache size µ00 (n) ≤ 1 suffices
to attain the exact same performance. For nd ≥ ns , however
µ00 (n) = 1 to achieve a DTB of ∆LB (n).
Remark 4. In all channel regimes R03 and R04 (R03 ∪ R04 =
I1C ), transceiver caching is not beneficial. The optimal latency
corresponds to the optimal DTB given in Lemma 1 for n ∈ I1C
(see Remark 2).
IV. L OWER B OUNDS ON DTB (C ONVERSE )
In this section, we develop lower bounds on the DTB
∆det (µ, n) to settle the optimality of our proposed achievability
scheme for various regimes of channel parameters nd , nr and
ns . For the sake of notational simplicity, we outline the proof
as if there were only N = 2 files. Nonetheless, the lower
bound remains valid for N > 2. We consider one option of
a worst-case demand pattern d = (dr , du )† = (1, 2)† ; that is,
RN and UE request distinct files Wdr = W1 and Wdu = W2 .
Throughout this section, we will use L to denote the Fano term
with its property L → 0 as L → ∞ [8]. For a given channel
realization n = (nd , nr , ns )† , we establish lower bounds on the
delivery time T as the converse on ∆det (µ, n). To this end, we
deploy distinct subsets of information resources B1 , . . . , B5 that
allow for reliable decoding of its corresponding message set as
L → ∞. This is defined as follows:
B1 = {yrT , S1 , S2 |W2 } → W1
B2 = {yuT } → W2
B3 = {Sq−nd xTs , yrT , S1 , S2 } → W1 , W2
B4 = {yuT , S1 } → W1 , W2 for nd ≥ ns
B5 = {xTs,[nd +1:ns ] , yuT , S1 } → W1 , W2 for ns ≥ nd .
For instance, we can write H(W1 , W2 |B5 ) ≤ LL . As
opposed to B2 , subsets B1 , B3 , B4 and B5 include cached
content(s) S1 (and S2 ) of requested file(s) W1 (and W2 ).
Thus, applying standard information theoretic bounding techniques with subsets B as side information will generate cachedependent and independent bounds through B1 , B3 , B4 , B5 and
B2 , respectively. The reasoning behind the choice of subsets
B1 –B4 are rather intuitive and we omit further details on
them for the sake of brevity. We will explain the choice of
B5 in the next paragraph. In accordance with the ordering of
T ≥
T ≥
(1 − µ)L
LL
−
,
ns
ns
L
LL
−
,
max{nd , nr }
max{nd , nr }
(2 − 2µ)L
LL
−
,
max{nd , ns }
max{nd , ns }
(2 − µ)L
LL
−
,
max{nd , nr }
max{nd , nr }
(2 − µ)L
LL
−
.
ns + (nr − nd )+
ns + (nr − nd )+
(14)
(15)
(16)
(17)
(18)
Dividing inequalities (14)–(18) by L and taking the limit for
L → ∞ will yield the desired lower bounds.
In the remainder of this section, we will explain the intuition
behind the choice of B5 = {xTs,[nd +1:ns ] , yuT , S1 } applicable for
ns ≥ nd only. We note that knowing yrT on top of B5 suffices
for reliable decodability. In fact, we will show that yrT can be
directly recovered from B5 . To understand this fact, we make
two key observations:
(A) First, we note that through the channel output yuT ⊆ B5
spanning T channel uses, any hypothetical decoder obtains
file W2 and ultimately S2 since S2 = φ2 (W2 ) (yuT →
W2 → S2 ).
(B) Second, any decoder can retrieve yrT
from
{xTs,[nd +1:ns ] , yuT , S1 , S2 } in a recursive manner.
Note that (B) presumes observation (A) due to the awareness of
S2 from yuT . Key observation (B) can be explained thoroughly
through T recursive steps t = 1, . . . , T . In this context, the
terms steps and channel use are equipollent. We now describe
the main operations of step t = 1 in detail. First, all arguments
of RN’s encoding function for channel use t = 1 are readily
available allowing for the recovery of xr [1] according to
xr [1] = ψr[1] (S1 , S2 , yr [0] = ∅, d, h).
(19)
Next, one can obtain the top most nd samples of xs [1] specified
by xs,[1:nd ] [1] from the known vectors xr [1] and yu [1] in
Equation (6b). Concatenating xs,[1:nd ] [1] and xs,[nd +1:ns ] [1] to
a single vector generates xs [1]; thus, also yr [1] (cf. (6a)). With
[2]
yr [1] in place, the RN encoding function ψr at channel use
t = 2 can be invoked, and, therefore step t = 2 initiates. The
decoder proceeds this way in the remaining T −1 steps until all
instances (up to the T –th instance) of yrT have been generated.
Finally, now having yrT ∪B5 , decoding of W1 and W2 becomes
possible. On the example of B5 , the main bounding techniques
applied to establish the lower bounds look as follows:
2L = H W1 , W2
= I W1 , W2 ; xTs,[nd +1:ns ] , yuT , S1
+ H W1 , W2 |xTs,[nd +1:ns ] , yuT , S1
(a)
≤ H xTs,[nd +1:ns ] + H yuT + H S1 + LL
(b)
≤ T ns + (nr − nd )+ + µL + LL ,
(20)
where (a) follows from Fano’s inequality and the fact that
conditioning does not increase entropy and (b) is because the
Bern(1/2) distribution maximizes the binary entropy
of each
element of xTs,[nd +1:ns ] and yuT as well as H S1 ≤ µL.
Cached Bits
W1
W2
DeNB
W2c̄
ns/2
W1c̄
c̄
W1
c
W1
c̄
W2
c
W2
Cached Bits
UE
W2c
nd ≥ ns/2
0nr −ns
0nr −ns
nr − nd
0nd
nr
W2c̄
W1c̄
received
(a) B = µ0 (n),
W2c
1
ns/2+nr −nd
transmitted
with µ0 (n) =
c
W1
c̄
W2
c
W2
UE
02nd −ns
nd − ns/2
RN
ns
DeNB
W2c̄
ns/2
W1
W2
c̄
W1
nr −nd
ns/2+nr −nd
W2c̄
ns − nd
W1c̄
ns − nd
nd
W2c
RN
W2c̄
0nr −ns
0nr −ns
02nd −ns
ns
W2c
W2c̄
nr
W1c̄
0ns −nd
received
transmitted
(b) C = µ00 (n),
1
nr
nd + nr
−ns
with µ00 (n) =
nd +nr −ns
nr
Fig. 3: Achievability scheme for corner points (a) B and (b) C in channel regime R31 when nr ≥ ns . The length of each sub-block
represents the entropy of the corresponding messages. For instance in (a) H(W2c̄ ) = ns/2.
V. ACHIEVABILITY OF M INIMUM DTB
In this section, we provide details on our scheme that
achieves the minimum DTB. Due to page limitations, the
scheme is described for the single channel regime R31 . Details
on the achievability for the remaining channel regimes follow
similar lines of argument. We first observe that the minimum
DTB is a convex function of the fractional cache size µ.
Lemma 2. The minimum DTB ∆∗det (µ, n) is a convex function
of µ ∈ [0, 1] for any given channel n.
Proof: The proof is omitted due to lack of space.
Due to the convexity of the DTB, we infer that it suffices
to establish the achievability of the corner points A , B and
C (cf. Fig. 2). In fact, the optimal DTB at fractional cache
size µ which lies between two neighboring corner points (say
A and B for instance) cache sizes’ is achieved through file
splitting and time sharing between the policies at those two
corner points. This strategy is only operational at a channel
regime under which the transmission policy at A , RA , and
the transmission policy at B , RB , are both feasible. This is
given by the non-empty set RA ∩RB . The next two sub-sections
establish the achievability for A , B and C , respectively.
In either sub-section, it is assumed that the RN and the UE
request different files denoted by W1 and W2 .
A. Achievability at µ = 0
For µ = 0, we can directly apply the results of Lemma
1. For RA = I1 ⊇ {nd ≤ ns ≤ 2nd }, this establishes the
achievability of a 2/ns -DTB.
B. Achievability at µ0 (n) and µ00 (n)
The optimal schemes at both µ0 (n) and µ00 (n) operate
over one single channel use under the same channel regime
RB = RC = R31 . Both transmission schemes are very much
alike as they are based upon the same premise. That is, the
DeNB transmits parts of the requested files W1 and W2 that
are not cached (W1c̄ and W2c̄ ) at the RN. The RN, on the other
hand, applies receiver caching for its own sake and transmitter
caching for the sake of the UE to make the remaining, cached
parts of W1 and W2 (W1c and W2c ) available. As µ increases,
more information on the requested files are available locally
at the RN; thus, reducing, the required rates on W1c̄ and W2c̄ .
The proper rate allocation of Wic̄ and Wic , i = 1, 2, at µ0 (n)
and µ00 (n) (µ0 (n) ≤ µ00 (n)) is specified in Figs. 3a and 3b,
respectively. The required fractional cache size at these corner
points corresponds to
H(Wic )
µ=
,
i = 1, 2.
(21)
H(Wi )
As the transmission schemes need one channel use to convey
transceiver RN and the UE with files Wi of equal rate, the
achievable DTB becomes the inverse of the file rate, i.e., 1/L,
where L = H(Wic )+H(Wic̄ ). This establishes the achievability
for corner points B and C .
VI. C ONCLUSION
In this paper, we have studied the fundamental latency of the
transceiver cache-aided wireless HetNet in the downlink given
in Fig. 1. We utilize the DTB as the performance metric that
captures the worst-case delivery latency of requested files. The
LDM is used as an approximation method for Gaussian channels, to completely characterize the optimal tradeoff between
storage and latency for the given HetNet in various channel
regimes.
R EFERENCES
[1] Network, Evolved Universal Terrestrial Radio Access, “M2 Application
Protocol (M2ap) (Release 10),” 2011.
[2] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” Trans.
on Info. Theory, vol. 60, no. 5, pp. 2856–2867, May 2014.
[3] ——, “Cache-aided interference channels,” in ISIT, June 2015, pp. 809–
813.
[4] A. Sengupta, R. Tandon, and O. Simeone, “Cache aided wireless networks:
Tradeoffs between storage and latency,” in CISS, March 2016, pp. 320–325.
[5] R. Tandon and O. Simeone, “Cloud-aided wireless networks with edge
caching: Fundamental latency trade-offs in fog radio access networks,” in
ISIT, July 2016, pp. 2029–2033.
[6] N. Naderializadeh, M. A. Maddah-Ali, and A. S. Avestimehr, “Fundamental
limits of cache-aided interference management,” in ISIT, July 2016, pp.
2044–2048.
[7] A. S. Avestimehr, S. N. Diggavi, and D. N. C. Tse, “Wireless network
information flow: A deterministic approach,” Trans. on Info. Theory,
vol. 57, no. 4, pp. 1872–1905, April 2011.
[8] T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley,
2006.
| 7cs.IT
|
Gröbner Bases of Generic Ideals
Juliane Capaverde and Shuhong Gao
arXiv:1711.05309v2 [math.AC] 8 Dec 2017
December 11, 2017
Abstract
Let I = (f1 , . . . , fn ) be a homogeneous ideal in the polynomial ring K[x1 , . . . , xn ]
over a field K generated by generic polynomials. Using an incremental approach
based on a method by Gao, Guan and Volny, and properties of the standard monomials of generic ideals, we show how a Gröbner basis for the ideal (f1 , . . . , fi ) can
be obtained from that of (f1 , . . . , fi−1 ). If deg fi = di , we are able
P to give
a comi−1
plete description of the initial ideal of I in the case where di ≥
j=1 dj − i − 1.
It was conjectured by Moreno-Socı́as that the initial ideal of I is almost reverse
lexicographic, which implies a conjecture by Fröberg on Hilbert series of generic
algebras. As a result, we obtain a partial answer to Moreno-Socı́as Conjecture: the
initial ideal of I is almost reverse lexicographic if the degrees of generators satisfy
the condition above. This result improves a result by Cho and Park. We hope this
approach can be strengthened to prove the conjecture in full.
1
Introduction
Let R = K[x1 , . . . , xn ] be the polynomial ring in n variables over an infinite field K.
A homogeneous ideal I = (f1 , . . . , fr ) is said to be of type (n, d1 , . . . , dr ) if, for each
1 ≤ i ≤ r, fi is a homogeneous polynomial of degree di . We are interested in answering
the following question: fixing the graded reverse lexicographic order, what is the shape
of the Gröbner bases of most ideals of type (n, d1 , . . . , dr )?
An ideal I = (f1 , . . . , fr ) of type (n, d1 , . . . , dr ) can be identified, through the coefficients of f1 , . . . , fr , with a point of the affine space V = RN1 × RN2 × · · · × RNr , where
Ni = di +n−1
. That is, V is viewed as the set of all ideals of type (n, d1 , . . . , dr ). A
di
property is said to be generic if it holds on a Zariski-open subset of V . We also say such
a property holds for most ideals of type (n, d1 , . . . , dr ), or that it holds for generic ideals.
It is known that Hilbert functions and initial ideals are constant in a nonempty open
subset of V (see [12]). A longstanding conjecture by Fröberg gives a formula for the
generic Hilbert series.
Conjecture 1.1 (Fröberg [6]). If I is a generic ideal of type (n, d1 , . . . , dr ), then the
Hilbert series of R/I, SR/I (z), is given by
Qr
di
i=1 (1 − z )
.
SR/I (z) =
(1 − z)n
P
d
The notation above
thePfollowing: if ∞
d=0 ad z is a power series with integer
P∞ means
∞
d
d
= d=0 bd z , where bd = ad if ai > 0 for 0 ≤ i ≤ d, and
coefficients, then
d=0 ad z
bd = 0 otherwise.
1
Moreno-Socı́as gave a conjecture describing the initial ideal of generic ideals with
respect to the graded reverse lexicographic order, related to the weak reverse lexicographic
property.
Definition 1.2. Let J = (xα1 , . . . , xαr ) be a monomial ideal, and suppose xα1 , . . . , xαr
are minimal generators, that is, these monomials are not divisible by one another. J is
said to be almost reverse lexicographic, or weakly reverse lexicographic, if, for every i, J
contains every monomial xα such that deg xα = deg xαi and xα > xαi .
Conjecture 1.3 (Moreno-Socı́as [11]). If I is a generic homogeneous ideal in R, then
the initial ideal of I is almost reverse lexicographic.
The following is a weaker version of the Moreno-Socı́as conjecture, restricted to generic
ideals in R = K[x1 , . . . , xn ] generated by n generic forms.
Conjecture 1.4. Let I = (f1 , . . . , fn ) be a generic ideal in R. Then in(I) is almost
reverse lexicographic.
It turns out that Conjecture 1.4 implies the case where the number r of polynomials
is different from the number of variables (see [12] for a proof of this fact). In this paper,
only the case where r = n is treated.
Partial answers have been given to both conjectures. Moreno-Socı́as Conjecture was
proven to hold in the case n = 2 by Aguirre et al. [1] and Moreno-Socı́as [11], n = 3
by Cimpoeaş
[4], n = 4 by Harima and Wachi [10], and the case for d1 , . . . , dn satisfying
Pi
di > j=1 dj − i + 1 by Cho and Park [3]. The Fröberg Conjecture is known to be true for
n = 2 [6], n = 3 [2], r = n + 1 [6], and d1 = · · · = dr = 2 and n ≤ 11, d1 = · · · = dr = 3
and n ≤ 8 [7].
It was pointed out in [3] that the Moreno-Socı́as Conjecture and the Fröberg Conjecture are closely related to the Lefschetz and Stanley properties.
Definition 1.5. Let I be a zero-dimensional homogeneous ideal in R, and A = R/I. Let
δ = max{i | Ai 6= 0}.
(i) A has the strong (resp. weak) Lefschetz property if there is a linear form L such
that multiplication map
Ld
Ai −→ Ai+d
is of maximal rank (that is, it is injective if dimK Ai ≤ dimK Ai+d , and surjective if
dimK Ai ≥ dimK Ai+d ) for each i and each d (resp. d=1).
(ii) A has the strong Stanley property if there is a linear form L such that the multiplication map
Lδ−2i
Ai −→ Aδ−i
is bijective for each 0 ≤ i ≤ ⌊δ/2⌋.
In [12] Pardue shows that the Moreno-Socı́as conjecture implies a series of other
conjectures. In particular, it implies the Fröberg conjecture, a fact that was also proven
by Cho and Park [3]. The connection between Moreno-Socı́as Conjecture and Fröberg
Conjecture and the Lefschetz property is illustrated by the following conjecture, which is
equivalent to the Fröberg Conjecture.
2
Conjecture 1.6 (Pardue [12]). If I is a generic ideal of type (n, d1 , . . . , dn ), then xn−1
is a weak Lefschetz element on R/ in(I) + (xn , . . . , xn−i+1 ), for each 0 ≤ i ≤ n − 1.
In [1], Aguirre et al gave the Gröbner basis of generic ideals in the case n = 2.
They used Buchberger’s Criterion to prove a given set, which was suggested by computer
calculations, was indeed a Gröbner bases. Our approach is similar to theirs, but instead
of using Buchberger’s Algorithm, we apply an incremental method from [9]. In this
paper, we show how a Gröbner basis for the generic ideal (I, g) can be obtained from
the Gröbner basis of I when a generic polynomial g is added, employing properties of
the standard basis of I. More precisely, we give a proof that the quotient R/I has the
strong Lefschetz and strong Stanley properties, and use this fact to establish the form
of the initial ideal of (I, g). As a result, we obtain a description
Pof theinitial ideal of
i−1
I = (f1 , . . . , fn ) in the case the degrees d1 , . . . , dn satisfy di ≥
j=1 dj − i − 1. Our
construction shows that Moreno-Socı́as Conjecture is true for these ideals, thus we give
a partial answer to the conjecture. Our result is somewhat more general then the one
given by Cho and
PPark in
[3], where they showed Moreno-Socı́as to be true for degrees
i−1
satisfying di >
j=1 dj − i + 1. We expect that our method can be strengthened to
fully prove the conjecture.
2
Incremental Gröbner bases
Let I be any ideal in R and suppose G is a Gröbner basis for I with respect to some
monomial order. The initial term of a polynomial f will be denoted by in(f ), and the
initial ideal of I will be denoted by in(I). Let g be any polynomial in R. We now describe
the method given in [8] to obtain a Gröbner basis of the ideal (I, g).
Let B = {xα1 , xα2 , . . . , xαN } be the set of monomials that are not in in(I), which we
call standard monomials with respect to I. Note that when I is not zero-dimensional, we
have N = ∞.
Suppose xαi g ≡ hi (mod G), where hi ∈ R is a K-linear combination of monomials in
B, for 1 ≤ i ≤ N. We can write this as
α
h1
x 1
h2
xα2
(1)
.. · g ≡ .. (mod G).
.
.
xαN
hN
We apply row operations to both sides of Equation (1) as follows: for 1 ≤ i < j ≤ N
and a ∈ K, subtract from the j-th row the i-th row multiplied by a. Our goal is to
eliminate equal leading terms. So if in(hi ) = in(hj ), with i < j, we use a row operation to
eliminate the leading term of hj . This means we only perform row operations downward.
We start with h1 , using the first row to eliminate the leading term of all hj bellow that
have the same leading monomial as h1 . Then we pass to the leading monomial of the new
second row, and eliminate the leading terms of all hj ’s with the same leading monomial.
Then we go to the new third row, and so on. Since the monomial order is a well ordering,
any decreasing sequence of monomials must be finite. Hence we perform only a finite
number of row operations on row j, using rows above it. By induction, we may assume
3
that Equation (1) can be transformed into the form
u1
v1
u2
v2
.. · g ≡ .. (mod G)
.
.
uN
(2)
vN
where ui , vi ∈ R are K-linear combinations of monomials in B, and for 1 ≤ i < j ≤ N
with vi , vj 6= 0, we have in(vi ) 6= in(vj ), that is, the nonzero rows in the right-hand side
of (2) have distinct leading monomials.
e = G ∪ {vi |1 ≤ i ≤ N}. Then G
e is a
Theorem 2.1 (Gao, Guan and Volny [8]). Let G
Gröbner basis of (I, g).
3
Gröbner bases of generic ideals
Now, let us return to the generic setting. We use here the definition of generic ideals
from [7], which is more suitable for our approach. Suppose K is an extension of a base
field F .
Definition 3.1.
(i) A polynomial f ∈ R of degree d is called generic over F if
X
f=
cα xα ,
α
where the sum runs over all monomials of degree d in R, and the coefficients cα are
algebraically independent over F .
(ii) An ideal I ⊂ R is generic if it is generated by generic polynomials f1 , . . . , fr with
all the coefficients algebraically independent over F . I is a generic ideal of type
(n, d1 , . . . , dr ) if I is a generic ideal in R generated by generic polynomials of degree
d1 , . . . , dr .
We apply the method above to generic ideals. Let S = K[x1 , . . . , xn , z] denote the
polynomial ring in n + 1 variables. Let f1 , . . . , fn and g be generic polynomials in S such
that I = (f1 , . . . , fn ) is a generic ideal of type (n + 1, d1 , . . . , dn ), and (f1 , . . . , fn , g) is a
generic ideal of type (n + 1, d1 , . . . , dn , d). Let G denote the reduced Gröbner basis of I
with respect to the grevlex order. Let E be the set of standard monomials with respect
to I. We denote by E the column vector whose entries are the monomials in E listed in
decreasing order. The set of elements of degree i in E is denoted by Ei .
Reducing g modulo G we obtain a polynomial that is a linear combination of monomials of degree d in E with coefficients that are still algebraically independent over F ,
and also algebraically independent over the extension of F generated by the coefficients
of elements of G. Thus, from now on we assume that g is reduced modulo G, that is,
we take g to be a linear combination of monomials in Ed with coefficients algebraically
independent over the extension of F generated by coefficients of elements of G.
Let M be the matrix satisfying
Eg ≡ ME (mod G).
4
(3)
Note that all polynomials involved are homogeneous. So, for a monomial m ∈ Ei , the
product mg is homogeneous and its reduced form is a homogenous polynomial of degree
i + d, that is, a K-linear combination of monomials in Ei+d only. Also, the row operations
can only be performed using two rows containing polynomials of the same degree. Thus,
we consider rows of different degrees separately. Let Mi denote the matrix such that
Ei g ≡ Mi Ei+d (mod G),
(4)
where Ei denotes the column vector whose entries are the monomials in Ei listed in
decreasing order.
Lemma 3.2. The rows of Mi are linearly independent.
Proof. Denote the rows of Mi by v1 , . . . , vℓ , and suppose Ei = (m1 , . . . , mℓ )T . Assume
c1 v1 + · · · + cℓ vℓ = 0, with c1 , . . . , cℓ ∈ K. Then
(c1 m1 + · · · + cℓ mℓ )g ≡ c1 v1 + · · · + cℓ vℓ ≡ 0 (mod G).
Since g is regular, g is not a zero divisor in R/I, and it follows that c1 m1 + · · · + cℓ mℓ = 0
in R/I. Since the monomials in Ei are K-linear independent, it follows that cj = 0 for
all 1 ≤ j ≤ ℓ. Hence, v1 , . . . , vℓ are linearly independent.
Thus, each matrix Mi has rank |Ei |. To be able to describe in(I, g), we need to see
which columns are linearly independent. The monomials in Ei+d corresponding to the
linearly independent columns are the ones that will be added to in(I) to form in(I, g).
Note that some of the monomials might be redundant, that is, multiples of monomials
previously added to the basis. For the Moreno-Socı́as Conjecture to be true, the columns
corresponding to the greatest non-redundant monomials in Ei+d should be linearly independent.
Define π : S −→ R to be the ring homomorphism that takes z to zero, fixing the
elements in K and the variables x1 , . . . , xn . Let J = π(I) ⊂ R be the image of I. Then
J is a generic ideal of type (n, d1 , . . . , dn ) in R, and using a property of the reverse
lexicographic order we have that π(in(I)) = in(J) (see [5, Proposition 15.12]. Moreover,
since z is regular in S/I, by Theorem 15.13 from [5], z is regular in S/ in(I), and by
Proposition 15.14, also from [5], we have that in(I) is generated by monomials that are
not divisible by z. Thus, in(I) and in(J) have the same minimal generators. Let B ⊂ R
denote the set of standard monomials with respect to J. It follows that
E = {mz ℓ |m ∈ B, ℓ ≥ 0}
= B ∪ zB ∪ z 2 B ∪ z 3 B ∪ · · · .
Now, J is a generic ideal in R generated by n generic forms, and for this type of ideals
some properties are known. For instance, Bi = 0, for all i > δ, where δ = d1 +· · ·+dn −n.
So, for each 0 ≤ i ≤ δ,
Ei = Bi ∪ zBi−1 ∪ z 2 Bi−1 ∪ · · · ∪ z i−1 B1 ∪ z i B0 ,
and for i ≥ δ,
Ei = z i−δ Bδ ∪ z i−δ+1 Bδ−1 ∪ · · · ∪ z i−1 B1 ∪ z i B0 .
This implies that for i > δ, Ei = z i−δ Eδ and
Ei g ≡ z i−δ Mδ Eδ+d (mod G).
5
Thus, the Gröbner basis elements obtained at this point are redundant, and we only need
to consider Ei g for 0 ≤ i ≤ δ.
Also note that, since Ei = zEi−1 ∪ Bi , Mi−1 is a submatrix of Mi for all 1 ≤ i ≤ δ,
and all matrices Mi are submatrices of Mδ . The rows and columns of Mδ can be indexed
by the elements of B. For 0 ≤ i ≤ δ, Mi is formed by blocks Γj,k , for 0 ≤ j ≤ i and
0 ≤ k ≤ d + i, where the entries of Γj,k are the coefficients of the monomials in z d+i−k Bk
in the reduced form of the polynomials in z i−j gBj . So Mi can be written as
Bd+i−1
Bd+i
Γi,d+i
Γi,d+i−1
Γ
i−1,d+i Γi−1,d+i−1
..
Mi = ...
.
Γ
Γ
1,d+i
1,d+i−1
Γ0,d+i
Γ0,d+i−1
···
···
···
..
.
B1
Γi,1
Γi−1,1
..
.
···
···
Γ1,1
Γ0,1
B0
Γi,0
Bi
Γi−1,0 Bi−1
..
.. ,
.
.
Γ B
1,0
1
Γ0,0
B0
and Equation (4) takes the form
Bd+i
Γi,d+i
Γi,d+i−1 · · · Γi,0
Bi
Γi−1,d+i Γi−1,d+i−1 · · · Γi−1,0 zBd+i−1
zBi−1
(mod G).
.. g ≡
..
.
.
.
.
.
d+i
i
z B0
z B0
Γ0,d+i
Γ0,d+i−1 · · · Γ0,0
3.1
Standard bases of generic ideals and the structure of the
blocks Γj,k
Our goal is to study the structure of the blocks Γj,k . First we establish some properties
of the sets Bi .
As noted before, the ideal J = π(I) is a generic ideal in R, generated by the generic
polynomials π(f1 ), . . . , π(fn ), with deg(π(fi )) = di , for 1 ≤ i ≤ n. Let A = R/J, and
define
δ
δ∗
σ
µ
=
=
=
=
d1 + · · · + dn − n,
d1 + · · · + dn−1 − (n − 1),
min{δ ∗ , ⌊δ/2⌋},
δ − 2σ.
The Hilbert series of A is known to be a symmetrical polynomial of degree δ, given by
S(z) =
Qn
δ
− z dj ) X
=
aν z ν ,
(1 − z)n
ν=0
j=1 (1
with 0 < a0 < · · · < aσ = · · · = aσ+µ > · · · > aδ > 0 (see [11, Proposition 2.2]). Here
aν = |Bν |.
To prove the properties of B we need in our proofs, we use a result from [11]. For
e ≥ 0, define
αn−1 e
n−1
e e = {xα1 1 · · · xαn−1
B
|xα1 1 · · · xn−1
xn ∈ B} ⊂ K[x1 , . . . , xn−1 ].
6
Proposition 3.3 (Moreno-Socı́as [11]). With the notation above,
e0 = B
e1 = · · · = B
eµ,
B
and
for 0 ≤ λ < σ.
e µ+1 = B
e µ+2 , . . . , B
e δ−1 = B
eδ ,
B
e δ−2λ = {m ∈ B
e 0 | deg(m) ≤ λ},
B
The following lemma implies that A has the strong Stanley property, and xn is a
strong Stanley element for A.
Lemma 3.4. Let 0 ≤ i ≤ 2δ . Then Bδ−i = xnδ−2i Bi .
eie = B
e e+δ−2i , for all e ≥ 0. For e > i, we
Proof. Note that Bδ−i = xnδ−2i Bi if and only if B
δ−i
e+δ−2i
e
e
e
have Bi = Bδ−i
= ∅. So we need to show equality for 0 ≤ e ≤ i.
ee = B
e e+δ−2i = B
e0 .
If both e ≤ µ and e + δ − 2i ≤ µ, then B
i
i−e
δ−i
ee = B
e 0 , and B
e e+δ−2i = {m ∈ B
e 0 | deg(m) ≤ λ},
If e ≤ µ and e + δ − 2i > µ, then B
i
i−e
i−e
δ−i
where e + δ − 2i = δ − 2λ or e + δ − 2i = δ − 2λ − 1. We need to see that i − e ≤ λ. In
the first case we have
2i − 2e
2i − e
≥
= i − e,
λ=
2
2
and in the second case,
λ=
as e ≥ 1.
Now, if e > µ, then
2i − e − 1
2i − e − e
≥
= i − e,
2
2
e e = {m ∈ B
e 0 | deg(m) ≤ λ},
B
where e = δ − 2λ or e = δ − 2λ − 1, and
e e+δ−2i = {m ∈ B
e 0 | deg(m) ≤ λ′ },
B
where e + δ − 2i = δ − 2λ′ or e + δ − 2i = δ − 2λ′ − 1. We want to see that i − e ≤ λ if
and only if i − e ≤ λ′ .
Case 1: Suppose e = δ − 2λ and e + δ − 2i = δ − 2λ′ . This happens when δ is even,
giving λ′ = λ + i − 2δ . If i − e ≤ λ, then λ′ = λ + i − 2δ = i − 2e ≥ i − e. If i − e ≤ λ′ , then
λ ≥ λ + i − 2δ = λ′ ≥ i − e, as i − 2δ ≤ 0.
Case 2: Suppose e = δ − 2λ and e + δ − 2i = δ − 2λ′ − 1, with δ odd and λ′ =
λ + i − δ−1
. If i − e ≤ λ, then λ′ = λ + i − δ−1
= i − 2e ≥ i − e. And if i − e ≤ λ′ , then
2
2
δ−1
δ−1
λ ≥ λ + i − 2 = λ′ ≥ i − e, as i − 2 ≤ 0.
.
Case 3: Suppose e = δ −2λ−1 and e+δ −2i = δ −2λ′ , with δ odd and λ′ = λ+i− δ+1
2
δ+1
e
′
If i − e ≤ λ, then λ = λ + i − 2 = i − 2 + 1 ≥ i − e, as e > µ ≥ 1 (µ = 0 would
contradict the fact that δ is odd). If i − e ≤ λ′ , then λ = λ + i − δ+1
= i − 2e + 1 ≥ i − e.
2
Case 4: Finally, Suppose e = δ − 2λ − 1 and e + δ − 2i = δ − 2λ′ − 1, with δ an even
integer and λ′ = λ + i − 2δ . If i − e ≤ λ, then λ′ = λ + i − 2δ = i − e+1
≥ i − e, as e > µ ≥ 0.
2
If i − e ≤ λ′ , then λ ≥ λ + i − 2δ = i − e+1
≥
i
−
e.
2
7
The next lemma establishes that A has the strong Lefschetz property, and xn is a
strong Lefschetz element for A.
Lemma 3.5. Let 0 ≤ j ≤ δ and r ≥ 0. Then multiplication by xrn from Aj to Aj+r is
either injective or surjective. More precisely:
(i) Suppose |Bj | ≤ |Bj+r |. Let S denote the subset of Bj+r consisting of |Bj | smallest
monomials in Bj+r . Then
S = xrn Bj .
(ii) Suppose |Bj | ≥ |Bj+r |. Let S denote the subset of Bj consisting of |Bj+r | smallest
monomials in Bj . Then
Bj+r = xrn S.
Proof. First, suppose 0 ≤ j ≤ δ/2 and j + r ≤ δ − j. Then |Bj | ≤ |Bj+r |. By Lemma
3.4, Bδ−j = xnδ−2j Bj , so multiplication by xnδ−2j from Aj to Aδ−j is bijective. This multiplication can be seen as the composition
xδ−2j−r
xr
n
Aj −→
Aj+r n−→ Aδ−j
so that multiplication by xrn from Aj to Aj+r must be injective. Moreover, if m is a
monomial in Bj , xnδ−2j m is in Bδ−j , which implies xrn m ∈ Bj+r . So, xrn Bj ⊆ Bj+r .
Suppose Bj = {xα1 , . . . , xαN }, with xα1 < · · · < xαN , and suppose m is a monomial in
Bj+r such that m < xαi xrn . Then xrn divides m, and m′ = m/xrn ∈ Bj , with m′ < xαi .
This proves (i).
Now suppose 0 ≤ j ≤ δ/2 and j + r ≥ δ − i. Then |Bj | ≥ |Bj+r |. Let m be a
2(j+r)−δ
2(j+r)−δ ′
monomial in Bj+r . Since Bj+r = xn
Bδ−j−r , we can write m = xn
m , for some
′
′′
2j+r−δ ′
m ∈ Bj , so m = xrn m′′ .
monomial m ∈ Bδ−j−r . By the previous paragraph, m = xn
So multiplication by xrn is surjective. Moreover, the monomials m′′ ∈ Bj that are taken
to m ∈ Bj+r are in the image of Bδ−j−r under multiplication by x2j+r−δ
, and, by part (i),
n
correspond to the smallest monomials in Bj .
If δ/2 ≥ j ≥ δ, then |Bj | ≥ |Bj+r |, and the same argument from the previous
paragraph works.
Since g is a combination of all monomials in Ed = Bd ∪ zBd−1 ∪ · · · ∪ z d B0 , we can
write
g = vd · Bd + vd−1 · Bd−1 z + · · · + v1 · B1 z d−1 + v0 · B0 z d ,
where vi is a row vector of coefficients. We denote the last entry of vi by ci . The following
Lemma gives some of the structure of the blocks Γj,k .
Lemma 3.6. Let 0 ≤ i ≤ δ, 0 ≤ j ≤ i and j ≤ k ≤ δ.
(i) Suppose |Bj | ≤ |Bk |. Then the entries on the diagonal of the square submatrix of
Γj,k formed by the last |Bj | columns have the form ck−j + L, where L is linear on
other coefficients in vk−j , vk−j+1, . . . , vδ and does not involve ck−j . Also, ck−j does
not appear in the other entries of Γj,k .
∗ · · · ∗ ck−j + L
∗
···
∗
∗ · · · ∗
∗
ck−j + L · · ·
∗
Γj,k =
(5)
.
..
∗ ··· ∗
∗
∗
8
...
ck−j + L
(ii) Suppose |Bj | ≥ |Bk |. Then the entries on the diagonal of the square submatrix of
Γj,k formed by the last (bottom) |Bk | rows have the form ck−j + L, where L is linear
on other coefficients in vk−j , vk−j+1, . . . , vδ and does not involve ck−j .
∗
∗
···
∗
..
.
∗
∗
···
∗
c
+
L
∗
·
·
·
∗
(6)
Γj,k =
k−j
∗
c
+
L
·
·
·
∗
k−j
.
..
∗
∗
· · · ck−j + L
Proof. (i) Let xα ∈ Bj , and consider the term ck−j xnk−j z d+j−k of g. By Lemma 3.5, the
monomial xα xnk−j is in Bk , that is, it is reduced modulo G. So in the reduced form
of the product xα z i−j · g, ck−j will certainly appear in the coefficient of the monomial
xα xnk−j z d+i−k . Larger monomials that appear in the product might not be reduced, and
the reduction would result in a coefficient of the form ck−j + L, as claimed. Since the
coefficient ck−j comes from a unique term in g, it cannot appear in any other entries.
(ii) Again, we let xα be a monomial in Bj . Suppose that xα is among the |Bk |
smallest monomials in Bj . By Lemma 3.5, the monomial xα xnk−j is in Bk , so ck−j appears
in the coefficient of the monomial xα xnk−j z d+i−k in the reduced form of xα z i−j · g. In the
reduction process, possibly larger terms will be reduced resulting in a coefficient of the
form ck−j + L. Note that ck−j might appear in the top rows of Γk,j , that is, ck−j appears
only once in each of |Bk | the bottom rows, but we cannot guarantee it does not appear
in other entries in the top rows.
3.2
Main results
In the next lemmas we handle the case with δ − d ≤ i ≤ δ. Let Θi denote the square
submatrix of Mi formed by columns corresponding to the |Ei | largest monomials in Ei+d .
We want to show that Θi is nonsingular, for all 0 ≤ i ≤ δ. The determinant of Θi is
a polynomial in the coefficients of g. We need to see that this polynomial is nonzero.
Our goal is to show there is a term that can be obtained as a product of entries in a
unique way, and hence cannot be cancelled. In this case the |Ei | largest monomials in
Ed+i are the monomials in z i+d−δ Bδ , z i+d−δ+1 Bδ−1 , . . . , z 2i+d−δ Bδ−i , and Θi is formed by
the following blocks
Bδ−1
Bδ
Γi,δ
Γi,δ−1
Γ
i−1,δ Γi−1,δ−1
..
Θi = ...
.
Γ
Γ1,δ−1
1,δ
Γ0,δ
Γ0,δ−1
···
···
···
..
.
Bδ−i+1
Γi,δ−i+1
Γi−1,δ−i+1
..
.
···
···
Γ1,δ−i+1
Γ0,δ−i+1
Bδ−i
Γi,δ−i
Bi
Γi−1,δ−i Bi−1
..
.. .
.
.
B
Γ
1,δ−i
1
Γ0,δ−i
B0
Lemma 3.7. Suppose δ − d ≤ i ≤ δ, and i ≥ δ/2. Then the term
(i+1)a0 i(a1 −a0 ) (i−1)(a2 −a1 )
cδ−i−1 cδ−i−2
cδ−i
(2i−δ+1)(aδ−i −aδ−i−1 )
· · · c0
(7)
can be obtained from the product of entries of Θi , with exactly one entry from each column
and row.
9
Proof. We will show how to select entries from Θi in steps. In each step, we pick entries
from a certain set of blocks Γj,k . We start at step 0, selecting entries from the blocks on
the diagonal of Θi , and then blocks above the diagonal in the next step, and so on.
Let 0 ≤ ℓ ≤ δ − i. At step ℓ we select aℓ − aℓ−1 entries from blocks
Γi,δ−ℓ , Γi−1,δ−ℓ+1 , . . . , Γℓ,δ−i .
(8)
The entries selected are the ones in the bottom aℓ rows, skipping the bottom aℓ−1 , and
aℓ right-most columns, skipping the last aℓ−1 columns. These entries have the form
cδ−ℓ−i + L.
..
.
cδ−ℓ−i+L
∗
∗
∗
∗
∗
∗
cδ−ℓ−i+L
∗
∗
∗
∗
∗
∗
aℓ − aℓ−1
cδ−ℓ−i+L
∗
cδ−ℓ−i+L
aℓ−1
∗
∗
cδ−ℓ−i+L
} |
{z
}
aℓ−1
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
∗
cδ−ℓ−i+L
∗
∗
∗
{z
|
aℓ − aℓ−1
Note that for any of the blocks Γj,k in (8), since ℓ ≤ δ − i ≤ i, and ℓ ≤ j ≤ i, it follows
that aℓ ≤ aj . Also, since δ − i ≤ k ≤ δ − ℓ, we have aℓ = aδ−ℓ ≤ ak . Thus, we indeed
have enough entries to pick in all blocks.
Furthermore, for a group of rows corresponding to Bj , we picked entries from the
bottom a0 rows of the block Γj,δ+j−i , then entries from the next a1 − a0 rows from the
block Γj,δ+j−i−1, and so on, so that we never select entries from the same rows. The same
reasoning applies to columns. Fixing a group of columns corresponding to Bk , we pick a0
entries from right column of Γk+i−δ,k , then a1 − a0 entries from the next columns, and so
on, never repeating columns. So we select a single entry from each row and each column.
At each step ℓ, for 0 ≤ ℓ ≤ δ − i, we picked aℓ − aℓ−1 entries of the form cδ−ℓ−i from
i − ℓ + 1 blocks. Taking the product of all entries selected, we have a polynomial in the
coefficients of g of the form
(i+1)a0 i(a1 −a0 ) (i−1)(a2 −a1 )
cδ−i−1 cδ−i−2
cδ−i
(2i−δ+1)(aδ−i −aδ−i−1 )
· · · c0
+ other terms.
Lemma 3.8. There is only one way of selecting entries from Θi and obtaining the term
in Equation (7).
Proof. We use induction to show that for ℓ = δ − i, δ − i − 1, . . . , 0, there is only one way
of obtaining the power of cℓ in Equation (7) from the product of entries of Θi .
We first consider cδ−i . We now use induction to show that for all 0 ≤ j ≤ i, the
only entry available to select in the last row of the set of rows corresponding to Bj is the
one on the last column of the block Γj,δ+j−i , of the form cδ−i + L. Note that the only
coefficient in Equation (7) that appears in the bottom row of Θi , corresponding to B0 ,
is cδ−i , which is in the last column of the block Γ0,δ−i . So we pick this entry. Suppose
now that the only way of selecting an entry containing a coefficient in Equation (7) from
the last row of the block corresponding to Bj is picking the one containing cδ−i , from the
10
last column of the block Γj,δ+j−i . This means that the other entries in this row involving
coefficients in Equation (7) cannot be selected at this point, and so entries in the last
columns of the blocks Γj,δ+j−i−1, Γj,δ+j−i−2, · · · have been selected in previous steps, from
blocks below. Passing to the set of rows corresponding to Bj+1 , it follows that entries
have been selected on the last columns of blocks Γj+1,δ+j−i , Γj+1,δ+j−i−1,..., and hence we
are left with no choice other than selecting the entry from the last column of the block
Γj+1,δ+j−i+1, which has the form cδ−i . This proves our claim.
Let 1 ≤ ℓ ≤ δ−i, and suppose we have selected entries involving cδ−i , cδ−i−1 , . . . , cδ−i−ℓ+1
as in Lemma 3.7, and that this selection was the only possible choice. This means that
we have already picked entries from the bottom aℓ−1 rows of all blocks B0 , . . . , Bi . So let
us consider the next aℓ − aℓ−1 rows. Starting with the block Bℓ , note that the coefficients
from Equation (7) that appear in this block are cδ−i , cδ−i−1 , . . . , cδ−i−ℓ . But with the selections we have already made, the exponents of cδ−i , cδ−i−1 , . . . , cδ−i−ℓ+1 in Equation (7)
were reached, so that at this point we cannot select the entries involving these coefficients.
Hence, the only choice left is selecting the entries of the form cδ−i−ℓ + L from Γℓ,δ−i .
Let ℓ+1 ≤ j ≤ i. Suppose we already picked entries of the form cδ−i−ℓ +L as in Lemma
3.7 from blocks Bℓ , Bℓ+1 , . . . , Bj−1. Consider block Bj . Still assuming that entries have
been selected from the bottom aℓ−1 rows, we pass to the next aℓ − aℓ−1 . The selections
made in blocks bellow prevent us from picking the entries involving cδ−i−ℓ−1 , . . . , c0 . Also,
we cannot select entries where coefficients cδ−i , . . . , cδ−i−ℓ+1 appear. Thus, we are left with
entries containing cδ−i−ℓ .
Lemma 3.9. Suppose δ − d ≤ i ≤ δ, and i ≤ δ/2. Then the term
(i+1)a0 i(a1 −a0 ) (i−1)(a2 −a1 )
cδ−i−1 cδ−i−2
cδ−i
(ai −ai−1 )
· · · ci
(9)
can be obtained from the product of entries of Θi , with exactly one entry from each column
and row.
Proof. The proof is the same as Lemma 3.7, except that in this case we select entries in
steps ℓ, for 0 ≤ ℓ ≤ δ − 2i.
The same proof of Lemma 3.8 works to show the following.
Lemma 3.10. There is only one way of selecting entries from Θi and obtaining the term
in Equation (9).
Corollary 3.11. det Θi 6= 0 for δ − d ≤ i ≤ δ.
When d ≥ δ, Corollary 3.11 holds for all matrices Θi , and we have the following.
Proposition 3.12. If d ≥ δ, then
in(I, g) = (in(I), z d−δ Bδ , z d−δ+2 Bδ−1 , . . . , z δ+d−3 B1 , z δ+d−1 B0 ).
Proof. Fix 1 ≤ i ≤ δ. Since Θi is the submatrix of Mi formed by columns corresponding
to the monomials in Bδ z d+i−δ , . . . , Bδ−i z d+2i−δ , and by Corollary 3.11 is nonsingular, we
can perform row operations on Mi and change Equation (4) into
ui
wi
ui−1
wi−1
.. · g ≡ .. (mod G),
.
.
u0
w0
11
where the entries of each wj are polynomials with distinct initial terms, so that each
monomial in Bδ z d+i−δ , . . . , Bδ−i z d+2i−δ occurs as leading monomial of some polynomial
in w0 , . . . , wi . But the monomials in Bδ z d+i−δ , . . . , Bδ−i+1 z d+2(i−1)−δ are redundant
as they are multiples of monomials that occur as leading terms when we perform row
operations on Ei−1 g ≡ Mi−1 Ei+d−1 (mod G). Thus, only the monomials in Bδ−i z d+2i−δ
are minimal generators of in(I, g).
e ⊂ K[x1 , . . . , xn , z] denote the set of standard
Corollary 3.13. Suppose d ≥ δ, and let B
monomials of (I, g). Then
e0 = B0
B
e1 = B1 ∪ zB0
B
e2 = B2 ∪ zB1 ∪ z 2 B0
B
..
.
eδ = Bδ ∪ zBδ−1 ∪ · · · ∪ z δ B0
B
eδ+1 = z B
eδ
B
..
.
e
eδ
Bd−1 = z d−δ−1 B
ed = z d−δ+1 B
eδ−1
B
ed+1 = z d−δ+3 B
eδ−2
B
..
.
e
Bd+δ−1 = z d+δ−1 B0 .
We now turn to the case d < δ and 0 ≤ i < δ − d. In this case, it is clear that
the monomials of degree d + i that enter the basis of in(I, g) are not necessarily the
largest monomials in Ei+d , and hence the square submatrix of Mi formed by columns
corresponding to those monomials is not necessarily nonsingular. In fact, the blocks Γj,k
for 0 ≤ j < i and d < k ≤ d + i have all entries equal to zero, thus the matrix formed
by columns corresponding to the greatest monomials might have rows of zeroes (it is
certainly the case for d < δ − 2).
However, for the Moreno-Socı́as Conjecture to be true, what we need is the greatest
non-redundant monomials to enter the basis at each step. Let i⋆ = ⌊ δ−d
⌋. We conjecture
2
the following.
Conjecture 3.14. Suppose d1 ≤ d2 ≤ · · · ≤ dn ≤ d, and 0 ≤ i ≤ i⋆ . Let Θi denote the
square submatrix of Mi formed by the columns corresponding to the ai largest monomials
of Bi+d , the ai−1 largest monomials of zBi+d−1 , and so on, up to the a0 largest monomials
of z i Bd . Then Θi is nonsingular.
Conjecture 3.15. Suppose d1 ≤ d2 ≤ · · · ≤ dn ≤ d, and i⋆ < i < δ − d. Let Θi denote
the square submatrix of Mi formed by columns corresponding to
(i) all monomials in Bd+j , for δ − d − i ≤ j ≤ i, and
(ii) the aj largest monomials in Bd+j , for 0 ≤ j < δ − d − i.
Then Θi is nonsingular.
12
In fact, matrices Θj for 0 ≤ j ≤ i⋆ are submatrices of Θi for i⋆ < i < δ − d, and if we
can prove the smaller matrices are nonsingular, we are actually able to prove all Θi are
nonsingular.
Proposition 3.16. Conjecture 3.14 implies Conjecture 3.15.
Proof. Let i⋆ < i < δ − d. Let Λi denote the submatrix of Θi formed by the following
blocks
Γi,d+i
Γi,d+i−1
···
Γi,δ−i
Γi−1,d+i
Γi−1,d+i−1 · · · Γi−1,δ−i
Λi =
..
..
..
.
.
.
.
.
.
Γδ−d−i,d+i Γδ−d−i,d+i−1 · · · Γδ−d−i,δ−i
Then, Θi can be written as
Θi =
Ω
Λi
0 Θδ−d−i−1
that is, the columns formed by Λ0i are the ones in Conjecture 3.15(i), and the columns
Ω
formed by Θδ−d−i−1
are the columns in (ii).
Now, det Θi = det(Λi ) · det(Θδ−d−i−1 ). If Conjecture 3.14 is true, then det Θδ−d−i−1 6=
0. So we need to see that det Λi 6= 0. In fact, an argument similar to that applied in
Lemmas 3.7-3.10 can be used. We claim the term
(2i+d−δ+1)ad+i (2i+d−δ)(ad+i−1 −ad+i )
cd−1
cd
(a
δ−i
· · · cδ−2i
−aδ−i+1 )
(10)
appears in the determinant of Λi . Again we start by selecting entries from the blocks on
the diagonal at step 0, and then from the blocks above the diagonal at step 1, and so on.
In general, at step ℓ, for 0 ≤ ℓ ≤ 2i + d − δ, we select entries from the blocks
Γi,di −ℓ , Γi−1,d+i−ℓ−1, . . . , Γδ−d−i+ℓ,δ−i .
We select the entries in the diagonal of the bottom ad+i−ℓ rows and right-most ad+i−ℓ
columns, skipping the bottom ad+i−ℓ+1 . The proof that these selections can be made,
and that this is the only way of obtaining the term (10) is identical to Lemma 3.7 and
Lemma 3.8.
The condition d1 ≤ d2 ≤ · · · ≤ dn ≤ d is necessary. When the degrees are not in this
order, our description fails, as we can see from the next example.
Example 3.17. Let I be the ideal in K[x1 , x2 , x3 , z] generated by generic forms of degree
4, 4 and 5, and let g be a generic polynomial of degree 3. According to Conjecture 3.14,
we consider the matrix Θ3 formed by the a3 = 10 greatest monomials in B6 , a2 = 6
greatest monomials in zB5 , a1 = 3 greatest monomials in z 2 B4 , and a0 = 1 greatest
monomial in z 3 B3 . This matrix, however, is singular, as the row corresponding to x3 z 2 · g
is zero. In fact, the monomial x42 z 2 , which is the third greatest monomial in z 2 B4 , is not
in the basis of in(I, g).
♦
When d = δ − 1, the only matrix treated in Conjecture 3.14 is Θ0 , which is a one by
one matrix whose single entry is the leading coefficient of g, and thus is nonzero. For
13
d = δ − 2, Θ0 is once again a one by one matrix whose entry is lc(g), and Θ1 is given by
cδ−2 + L
∗
···
∗
∗
∗
cδ−2 + L · · ·
∗
∗
Ω
Γ1,δ−1
..
..
..
..
..
=
Θ1 =
.
.
.
.
.
0
lc(g)
∗
∗
· · · cδ−2 + L
∗
0
0
···
0
lc(g)
so det Θ1 = lc(g) · det Γ1,δ−1 , and the determinant of Γ1,δ−1 is nonzero because the term
1
caδ−2
appears in it. This, together with the results from the previous section, proves the
following.
Proposition 3.18. Suppose d ≥ δ − 2. If in(I) is almost reverse lexicographic, then
in(I, g) is almost reverse lexicographic.
Proof. For all t ≥ d, the minimal generators of degree t introduced to the basis of in(I, g)
are the largest monomials in Et .
Using induction we have a partial answer to Moreno-Socı́as Conjecture.
Theorem 3.19. Let
I = (f1 , . . . , fn ) ⊂ K[x1 , . . . , xn ] be a generic ideal, with deg(fi ) = di
Pi−1
and di ≥
j=1 dj − i − 1. Then in(I) is almost reverse lexicographic.
Proof. The result clearly holds for n = 1. Assuming it holds for n − 1, the initial ideal
of (f1 , . . . , fn−1 ) ⊂ K[x1 , . . . , xn−1 ] is almost reverse lexicographic. By Proposition 3.18,
in(I) is almost reverse lexicographic.
The theorem above is somewhat
than the result given in [3], where Cho
Pmore general
i−1
and Park proved the case di >
j=1 dj − i + 1. We believe that our approach is
promising, and that by investigating further the properties of B(I) and the structure of
the matrices from Conjecture 3.14, we could be able to give an answer to Moreno-Socı́as
Conjecture.
If Conjecture 3.14 is true, then we can give a description of in(I, g) as follows. We
will use the following notation: for a set S = {s1 , . . . , sℓ } and 1 ≤ a ≤ b ≤ ℓ, let
S [a,b] = {sa , . . . , sb },
S (a,b] = {sa+1 , . . . , sb }.
If δ − d ≡ 0 (mod 2), the initial ideal of (I, g) can be described as
[1,a0 ]
in(I, g) = ( in(I), Bd
(a
,a
⋆ −1 ]
d+i
z 2 Bd+ii ⋆−1
−1
⋆
[1,a
[1,a ]
⋆
]
−1
, Bd+11 , . . . , Bd+i⋆i −1
, Bd+i⋆ ,
(a
,a
⋆ −2 ]
d+i
, z 4 Bd+ii ⋆−2
−2
⋆
z δ−d+2 Bd−1 , . . . , z δ+d−2 B1 , z δ+d B0
e is
The corresponding set of standard monomials B
e0 = B0
B
e1 = B1 ∪ zB0
B
e2 = B2 ∪ zB1 ∪ z 2 B0
B
14
(a0 ,ad ]
, . . . , z δ−d Bd
).
,
..
.
ed−2
= Bd−1 ∪ z B
(a0 ,ad ]
ed−1
= B
∪ zB
ed−1
B
ed
B
ed+1 =
B
..
.
ed+i⋆ =
B
ed+i⋆ +1 =
B
..
.
eδ =
B
eδ+1 =
B
..
.
e
Bδ+d−1 =
d
(a1 ,ad+1 ]
Bd+1
ed−2
∪ zB
ed+i⋆ −1
zB
ed+i⋆ −2
z3B
ed−1
z δ−d+1 B
ed−2
z δ−d+3 B
e0 .
z δ+d−1 B
If δ − d ≡ 1 (mod 2), the initial ideal of (I, g) can be described as
[1,a0 ]
in(I, g) = ( in(I), Bd
(a ⋆ ,ad+i⋆ ]
zBd+ii ⋆
z
δ−d+2
[1,a ⋆ ]
[1,a ]
, Bd+11 , . . . , Bd+i⋆i ,
(a
,a
⋆ −1 ]
d+i
, z 3 Bd+ii ⋆−1
−1
Bd−1 , . . . , z
⋆
δ+d−2
B1 , z
(a0 ,ad ]
, . . . , z δ−d Bd
δ+d
B0
,
).
e is
The corresponding set of standard monomials B
e0 = B0
B
e1 = B1 ∪ zB0
B
e2 = B2 ∪ zB1 ∪ z 2 B0
B
..
.
e
ed−2
Bd−1 = Bd−1 ∪ z B
ed = B (a0 ,ad ] ∪ z B
ed−1
B
d
(a1 ,ad+1 ]
Bd+1
ed−2
ed+1 =
∪ zB
B
..
.
ed+i⋆ −1
ed+i⋆ = B (ai⋆⋆,ad+i⋆ ] ∪ z B
B
d+i
ed+i⋆ +1 = z 2 B
ed+i⋆ −1
B
4
ed+i⋆ −2
ed+i⋆ +2 = z B
B
..
.
e
ed−1
Bδ = z δ−d+1 B
eδ+1 = z δ−d+3 B
ed−2
B
..
.
eδ+d−1 = z δ+d−1 B
e0 .
B
From the description above, we have that Conjecture 3.14 implies that in(I, g) is
almost reverse lexicographic, and hence also implies the Moreno-Socı́as conjecture.
15
Example 3.20. Let f1 , f2 be generic polynomials of degrees d1 = d2 = 4, and let I =
(f1 , f2 ). The initial ideal of I is given by
in(I) = (x41 , x31 x2 , x21 x32 , x1 x52 , x72 ).
Then δ = 6, and we consider g of degree d = 4 = δ − 2.
g = b1 x21 x22 + b2 x1 x32 + b3 x42 + b4 x31 z + b5 x21 x2 z + b6 x1 x22 z + b7 x32 z + b8 x21 z 2 + b9 x1 x2 z 2
+b10 x22 z 2 + b11 x1 z 3 + b12 x2 z 3 + b13 z 4 .
We give the matrices Θi below. We write entries as functions of the coefficients bi ’s.
All entries have the form bi + L(b1 , . . . , bi−1 ) or L(b1 , . . . , bi ). We show only the entries of
the first form, ignoring the L portion. The entries selected to form the terms in Lemma
3.7 and Lemma 3.9 are shown in boldface. We start with Θ6 = M6 :
x62
x1 x42
x52
x21 x22
x1 x32
x42
x31
x21 x2
x1 x22
x32
2
x1
x1 x2
x22
x1
x2
1
x62 x1 x42 x52 x21 x22 x1 x32 x42
x31 x21 x2 x1 x22 x32
x21 x1 x2 x22
x1
x2
1
b13
b13
b12
b13
b10
b12
b11 b12
b7
b13
b10
b9 b10
b7
b3 b6 b7
b3
b2 b3
b13
b13
b13
b12
b11 b12
b11
b10
b9 b10
b8 b9
b6 b7
b5 b6
b1 b2
b13
b13
b12
b10
b7
b3
b13
b11 b12
b13
b11 b12
b13
b11 b12
b13
b9 b10
b11 b12
b13
b8 b9 b10
b11 b12
b13
b4 b5 b6 b7 b8 b9 b10 b11 b12 b13
.
The entries in boldface give a nonzero term in det Θ6 . Since the determinant is nonzero,
performing row operations on
E6 · g ≡ M6 E10 (mod G),
all monomials in E10 will appear as leading monomials on the right-hand side. Thus, the
monomials
x62 z 4 , x1 x42 z 5 , x52 z 5 , x21 x22 z 6 , x1 x32 z 6 , x42 z 6 , x31 z 7 , x21 x2 z 7 , x1 x22 z 7 , x32 z 7 , x21 z 8 , x1 x2 z 8 , x22 z 8 , x1 z 9 , x2 z 9 , z 10
are in the basis of in(I, g). The matrix Θ5 is obtained from Θ6 by removing the top row
and right-most column. Again we show in boldface the entries that are used to guarantee
16
that the determinant of this matrix is nonzero. This is the form of Θ5
x62 x1 x42 x52 x21 x22 x1 x32 x42
x1 x42
x52 b12
x21 x22
x1 x32
4
x2 b10
x31
x21 x2
2
x1 x2
x32 b7
x21
x1 x2
x22 b3
x1
x2
1
x31 x21 x2 x1 x22 x32
x21 x1 x2 x22
x1
x2
b13
b13
b13
b12
b11 b12
b10
b9 b10
b7
b6 b7
b3
b2 b3
b13
b13
b12
b11 b12
b11
b10
b9 b10
b8 b9
b6 b7
b5 b6
b1 b2
b13
b13
b13
b12
b10
b7
b3
b13
b11 b12
b13
b11 b12
b13
b11 b12
b13
b9 b10
b11 b12
b13
b8 b9 b10
b11 b12
b13
b4 b5 b6 b7 b8 b9 b10 b11 b12
.
So, performing row operations on
E5 · g ≡ M5 E9 (mod G)
leads to the 15 greatest monomials in E9 being leading monomials on the right-hand side,
which means that
x62 z 3 , x1 x42 z 4 , x52 z 4 , x21 x22 z 5 , x1 x32 z 5 , x42 z 5 , x31 z 6 , x21 x2 z 6 , x1 x22 z 6 , x32 z 6 , x21 z 7 , x1 x2 z 7 , x22 z 7 , x1 z 8 , x2 z 8
are in in(I, g). Next, we consider
E4 · g ≡ M4 E8 (mod G).
The matrix M4 is 13 × 16, and Θ4 is the 13 × 13 submatrix given by
x21 x22
x1 x32
x42
x31
x21 x2
x1 x22
x32
x21
x62 x1 x42 x52 x21 x22 x1 x32 x42
b12
b10 b11 b12
b10
b7 b9 b10
x1 x2
b7
x22 b3 b6 b7
x1
b3
x2
b2 b3
1
x31 x21 x2 x1 x22 x32
x21 x1 x2 x22
b13
b13
b13
b13
b12
b13
b11 b12
b13
b11 b12
b13
b10
b11 b12
b13
b9 b10
b11 b12
b13
b8 b9 b10
b11 b12
b13
b6 b7
b9 b10
b11 b12
b5 b6 b7
b8 b9 b10
b11 b12
b1 b2 b3 b4 b5 b6 b7 b8 b9 b10
,
which is also a submatrix of Θ5 , obtained by removing the rows corresponding to B5 and
the columns corresponding to B1 . After row operations,
x62 z 2 , x1 x42 z 3 , x52 z 3 , x21 x22 z 4 , x1 x32 z 4 , x42 z 4 , x31 z 5 , x21 x2 z 5 , x1 x22 z 5 , x32 z 5 , x21 z 6 , x1 x2 z 6 , x22 z 6
are leading monomials. Similarly, removing from Θ4 the rows corresponding to B4 and
17
the columns corresponding to B2 , we get Θ3 given by
x62 x1 x42 x52 x21 x22 x1 x32 x42
x31 x21 x2 x1 x22 x32
b13
x31
x21 x2
x1 x22
x32
x21
b13
b12
b10
b11 b12
b
b
b
b11 b12
7 9 10
b
b11
10
x1 x2
b7
b9 b10
x22 b3 b6 b7 b8 b9 b10
x1
b3
b6 b7
x2
b2 b3 b5 b6 b7
1
b1 b2 b3 b4
b13
b12
.
b11 b12
b11 b12
b9 b10
b8 b9 b10
b5 b6 b7
The leading monomials obtained are
b13
x62 z, x1 x42 z 2 , x52 z 2 , x21 x22 z 3 , x1 x32 z 3 , x42 z 3 , x31 z 4 , x21 x2 z 4 , x1 x22 z 4 , x32 z 4 .
Next, Θ2 is given by
6
x21
x2
x1 x2
x22
b3
x1
x2
1
x1 x42 x52 x21 x22 x1 x32 x42
b7
b6
b3
b2
b10
b9
b7 b8
b6
b3 b5
b1
b10
b9 b10
,
b7
b6 b7
b2 b3
and the elements in E6 that are leading monomials are
x62 , x1 x42 z, x52 z, x21 x22 z 2 , x1 x32 z 2 , x42 z 2 .
The matrix Θ1 is given by
x1 x42 x52 x21 x22
x1
x2
1
b3
b6
b2 b3 b5
b1
!
,
and the monomials of degree 5 that enter the basis of in(I, g) are
x1 x42 , x52 , x21 x22 z.
Finally, Θ0 is the 1 × 1 matrix
x21 x22
1
( b1 ),
and the monomial x21 x22 is in in(I, g). Putting all the leading monomials we found together,
and discarding the redundant ones, we have
in(I, g) = (x41 , x31 x2 , x21 x22 , x1 x42 , x52 , x1 x32 z 2 , x42 z 2 , x31 z 4 , x21 x2 z 4 ,
x1 x22 z 4 , x32 z 4 , x21 z 6 , x1 x2 z 6 , x22 z 6 , x1 z 8 , x2 z 8 , z 10 ),
which is an almost reverse lexicographical monomial ideal.
18
♦
References
[1] Aguirre, E., Jarrah, A. S., and Laubenbacher, R. Generic ideals and
Moreno-Socı́as conjecture. In Proceedings of the 2001 International Symposium on
Symbolic and Algebraic Computation (New York, 2001), ACM, pp. 21–23.
[2] Anick, D. J. Thin algebras of embedding dimension three. J. Algebra 100, 1 (1986),
235–259.
[3] Cho, Y. H., and Park, J. P. Conditions for generic initial ideals to be almost
reverse lexicographic. J. Algebra 319, 7 (2008), 2761–2771.
[4] Cimpoeaş, M. Generic initial ideal for complete intersections of embedding dimension three with strong Lefschetz property. Bull. Math. Soc. Sci. Math. Roumanie
(N.S.) 50(98), 1 (2007), 33–66.
[5] Eisenbud, D. Commutative algebra, vol. 150 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1995. With a view toward algebraic geometry.
[6] Fröberg, R. An inequality for Hilbert series of graded algebras. Math. Scand. 56,
2 (1985), 117–144.
[7] Fröberg, R., and Hollman, J. Hilbert series for ideals generated by generic
forms. J. Symbolic Comput. 17, 2 (1994), 149–157.
[8] Gao, S., Guan, Y., and Volny IV, F. A new incremental algorithm for computing Gröbner bases. In ISSAC’10: Proceedings of the 2010 International Symposium
on Symbolic and Algebraic Computation (Munich, Germany, 2010), ACM, pp. 13–19.
[9] Gao, S., Volny, IV, F., and Wang, M. A new framework for computing
Gröbner bases. Math. Comp. 85, 297 (2016), 449–465.
[10] Harima, T., and Wachi, A. Generic initial ideals, graded Betti numbers, and
k-Lefschetz properties. Comm. Algebra 37, 11 (2009), 4012–4025.
[11] Moreno-Socı́as, G. Degrevlex Gröbner bases of generic complete intersections.
J. Pure Appl. Algebra 180, 3 (2003), 263–283.
[12] Pardue, K. Generic sequences of polynomials. J. Algebra 324, 4 (2010), 579–590.
19
| 0math.AC
|
arXiv:1602.05784v3 [math.CO] 17 Jun 2016
ON SUBTILINGS OF POLYOMINO TILINGS
JACOB TURNER:
Abstract. We consider a problem concerning tilings of rectangular regions
by a finite library of polyominoes. We specifically look at rectangular regions
of dimension n ˆ m and ask whether or not a tiling of this region can be
rearranged so that tiling of the n ˆ m rectangle can be realized as a tiling of
an n ˆ m1 rectangle and an n ˆ m2 rectangle, m “ m1 ` m2 . We call this a
subtiling. We show that the associated decision problem is NP-complete when
restricted to rectangular polyominoes. We also show that for certain finite
libraries of polyominoes, if m is sufficiently large, a subtiling always exists and
give bounds.
Keywords: Polyomino; tilings; algebraic combinatorics
1. Introduction
We begin by considering the following combinatorial optimization problem. Suppose one has a set of objects that one can place in some rectangular container without having any gaps. However, the container may be considered suboptimal as one
of its dimensions is very large. Instead, one would rather place the objects into two
containers, both of which have more reasonable dimensions. This can be done of
course, but at the expense of packing efficiency; one may be forced to accept that
there will be wasted space in at least one of the containers.
We model this problem in terms of polyomino tilings. Suppose you are given
a finite set of shapes L, which we will take to be polyominoes. Recall that a
polyomino is a region of R2 , defined up to translation, consisting of squares of
equal size arranged by coincident edges. In this paper, we take the squares making
up a polyomino to all be the unit square.
We say that a set of polyominoes L can tile a compact region R Ď R2 if there
is a multiset with elements drawn from L such that they can be arranged in R so
the every point in R is covered by an element of the multiset and no overlap of two
different elements is two dimensional.
We are interested in tiling rectangular regions of R2 by a set of polyominoes L.
We denote an n ˆ m rectangular region by Rnˆm . Suppose we having a tiling of
Rnˆm by a multiset T of polyominoes with elements drawn from L. In general there
may be several ways that the elements of T tile Rnˆm and it might be possible do
so by tiling the regions Rnˆm1 and Rnˆm2 where m “ m1 ` m2 , with m1 , m2 ‰ 0,
using the elements from the multiset T situtating these two tiled regions side by
side. If this happens, we say that tiling of Rnˆm admits a subtiling. If a subtiling
exists once one or more of the polyominoes in T has been replaced with a rotated
: Korteweg-de Vries Institute for Mathematics, University of Amsterdam, 1098 XG Amsterdam, Netherlands.
1
ON SUBTILINGS OF POLYOMINO TILINGS
2
copy of itself, we say the tiling admits a subtiling allowing rotations. We make the
following definition:
Definition 1.1. Suppose one is given a tiling T of Rnˆm by a multiset of polyominoes T . Any tiling of Rnˆm by the same multiset T is called a rearrangement of
T . We consider T to be a rearrangement of itself. A tiling of Rnˆm by a multiset
T 1 , reached by replacing any of the elements in T by a rotated copy, is called a
rearrangement of T with rotations. We emphasize that while T may contain multiple copies of a particular polyomino, each copy can be rotated different degrees
independent of each other.
An equivalent way of defining a subtiling is that a tiling T of Rnˆm can be
rearranged to give a tiling of Rnˆm that is formed by juxtaposing tilings of Rnˆm1
and Rnˆm2 with m “ m1 ` m2 and m1 , m2 ‰ 0, side by side in one of the two
obvious ways. The analagous statement holds for subtilings allowing rotations.
This motivates the following definition.
Definition 1.2. Given a finite set of polynominoes, define βn pLq to be the smallest
natural number such that for any tiling of Rnˆm by elements of L with m ą βn pLq,
the tiling admits a subtiling. If arbitrarily large tilings exist with no subtilings,
then βn pLq :“ 8. We define βnR pLq analagously for subtilings allowing rotations.
Note that these numbers depend on n.
The main problem of interest for us is finding upper bounds on βn pLq. This
forces us to address a couple of other questions that we will attempt to give at least
partial answers to.
‚ How hard is it to tell if a given tiling admits a subtiling, with or without
allowing rotations?
‚ For which L and which n is βn pLq and βnR pLq finite?
We can answer the first question for finite sets of rectangular polyominoes and
investigation into the latter question will occupy much of this paper. In this paper,
we find several sets of polyominoes L such that βn pLq is finite and give bounds
on it. We will not devote much space to βnR pLq in this paper, but we note that
βn pLq ě βnR pLq.
Since the popularization of polyominoes in the mid 20th century, there has been
a vast amount of literature on the subject and subsequent generalizations to polyforms, especially the question of determining when polyforms tile given regions
[17, 9, 26, 1, 2, 28, 8].
Many polyomino problems have been recreational: for example puzzles of Martin
Gardner [14, 15, 13] or more recently, the Eternity puzzle [22, 35]; a survey of the
recreational aspects of the subject can be found in [29]. Others are problems well
known as favorites to give young mathematics students to solve on homeworks, e.g.
the problems of tiling mutilated chess boards by dominos. Some of these problems
can be quite difficult.
In addition, polyomino tiling problems have played an important part in complexity theory where they have proven to be a source of NP-hard or even undecidable
problems [4, 34, 3, 30]. Furthermore, as polyomino tilings are ubiquitous in combinatorial optimization and operations research, they have even found their way into
other scientific and industrial settings with applications ranging from genetics to
antenna design [36, 6, 23, 7].
ON SUBTILINGS OF POLYOMINO TILINGS
3
The question of determining if a subtiling must exist is similar to the question
of existence of fault-free tilings. A fault in a tiling of a rectangular region Rnˆm
(typically with dominos) is a horizontal or vertical line that can be drawn through
the interior of Rnˆm and does not pass through the interior of any tile. A fault is
called vertical or horizontal if the defining line is vertical or horizontal, respectively.
A fault-free tiling is a tiling with no fault. The problem of finding fault-free tilings
is credited to Robert I. Jewett in [16] where it was also shown that a fault-free
rectangle must have length exceeding four, with a 5 ˆ 6 rectangle yielding the
smallest construction. An elegant proof is given that no such tiling exists for the
6 ˆ 6 rectangle.
Necessary and sufficient conditions for a rectangular region to admit a fault-free
tiling of dominos has been worked out [18]. Later, the rectangular polyominoes
that admit fault-free tilings was completely classified [33]. The condition of having
a subtiling is different as we ignore horizontal faults, consider sets of many different
polyominoes, and we allow a tiling to be rearranged.
This paper is organized as follows: In Section 2, we formalize our problem as
well as proving that determining if a tiling admits a subtiling is NP-complete when
L only consists of rectangular polyominoes. We also prove that some conditions
on L implies βn pLq is finite using direct arguments. In Section 3, we establish a
connection between the finiteness of βn pLq and certain rings being Noetherian. We
then prove that if a finite set of polyominoes satisfies a technical condition, then we
can show βn pLq will be finite. In Section 4, we prove that certain families of finite
sets of polyominoes actually satisfy the technical condition in Section 3. In Section
5, we find upper bounds for βn pLq for special sets L.
2. Preliminaries and first results
Throughout this paper, we will have a finite set of polyominoes L which will
be used to tile a rectangular region. While it is natural for such a set L to be
closed under reflections and rotations, our proofs only require that the set be closed
under reflections across a vertical line (we call this a vertical reflection). Many
of the examples we will consider, however, will only involve sets of rectangular
polyominoes and is thus closed under reflections.
Definition 2.1. If βnR pLq ă 8, we say that L is n-Noetherian. If βn pLq ă 8,
we say that L is strongly n-Noetherian. Being strongly n-Noetherian implies being
n-Noetherian as βnR pLq ď βn pLq.
The justification of terminology of n-Noetherian will be given later in Section 3.
It is clear that for every n, there exists a set L that is (strongly) n-Noetherian: let
L consist solely of the 1 ˆ 1 domino. We now give an example of a set L that is not
2-Noetherian.
Example 2.2. Let L be the set consisting of the following three polyominoes and
their vertical reflections.:
,
,
We can construct tilings of R2ˆN , for N arbitrarily large, such that no subtiling
exists. Consider the following general tiling:
¨¨¨
ON SUBTILINGS OF POLYOMINO TILINGS
4
It is clear that this tiling has no subtiling, even if we allow all reflections and
rotations, and can be made arbitrarily wide.
2.1. The computational complexity of deciding if a subtiling exists. In
Example 2.2, we could visually tell that those particular tilings had no subtiling.
However, this is not an easy problem in general. This is in contrast to checking to
see if a tiling is fault-free, which is very easy. We will show that even if we restrict
ourselves to rectangular polyominoes, this problem is hard.
We are given as input a tiling of a rectangle Rnˆm for some n and m by rectangular polyominoes. The polyominoes in the tiling are given as ordered pairs of
binary numbers ph, wq where h is the height and w the width of the corresponding
polyominoes. The unit square in Rnˆm that the bottom left corner of each polyomino occupies in the tiling is also specified as an ordered pair of binary numbers
with the convention that p0, 0q denotes the bottom left unit square in Rnˆm .
Problem 2.3. Define the problem TRANS ´ ST as the problem of deciding if a
tiling of Rnˆm by rectangular polynominoes admits a subtiling where rearrangement
is restricted to translations. The problem GEN ´ ST asks if there exists a subtiling
allowing rotations.
We show that TRANS ´ ST and GEN ´ ST are NP-complete. Many polyomino
packing problems are NP-complete and often can be reduced to the subset sum
problem. Here we will reduce to the following related NP-complete problem.
Problem 2.4. The partition problem is the problem of deciding if a multiset
M
ř
of positive integers can be written as the disjoint union S1 \ S2 where sPS1 s “
ř
sPS2 s.
Theorem 2.5 ([24]). The partition problem is NP-complete.
We note that for rectangular polyominoes, reflections are trivial. As such, the
difference in TRANS ´ ST and GEN ´ ST is only that in GEN ´ ST we can attempt
to rotate pieces in our rearrangements.
Theorem 2.6. The problems TRANS ´ ST and GEN ´ ST are both NP-complete.
Proof. Clearly both problems are in NP as a subtiling can be easily verified. We wish
to reduce both problems to the partition problem.řSuppose weřare given a multiset
of positive integers M “ tm1 , . . . , mℓ u and 2N “ mPM m (if mPM m is not even,
we know no partition is possible). Then let L be the set of polyominoes consisting
of the p2N ` 1q ˆ mi polyomino for every 1 ď i ď ℓ and also the 1 ˆ N polyomino.
Then we tile the region Rp2N `2qˆ2N by placing the two 1 ˆ N polyominoes in the
first row and one of each p2N ` 1q ˆ mi for 1 ď i ď ℓ polyomino side by side the
the bottom 2N ` 1 rows. This is clearly a tiling of Rp2N `2qˆ2N .
With this tiling, rearrangements allowing rotations are impossible (apart from
rotations of nπ radians). First of all, we cannot rotate any of the p2N ` 1q ˆ mi
polyominoes by π{2 radians as their height exceeds the width of the rectangular
region being tiled. We can rotate the 1 ˆ N polyominoes π{2 radians, however.
Suppose we do this for one of the 1 ˆ N polyominoes.
It now lies within a single column whose remaining N ` 2 rows must be occupied
by other polynominoes. However, none of the p2N ` 1q ˆ mi polyominoes can share
a column with the N ˆ 1 polyomino as the height of the rectangular region being
tiled is 2N ` 2. So it can only share a column with the other 1 ˆ N polyomino, or
ON SUBTILINGS OF POLYOMINO TILINGS
5
if we rotate it, an N ˆ 1 polynomio. But in either scenario, we see that some of the
rows of the column in question will be left unoccupied.
So this reduction is valid for both TRANS ´ ST and GEN ´ ST since we are
using rectangular polyominoes and reflections are trivial. Lastly, we note that a
subtiling exists for this tiling if and only if the polyominoes of the form 1 ˆ mi
can be partitioned into two disjoint sets S1 and S2 such that the lengths of the
polyominoes in S1 sum to N and likewise for the lengths of the polyominoes in S2 .
As such, a subtiling exists if and only if M has a partition of the desired kind.
2.2. The case of a single rectangular polyomino. Returning our attention to
the question of determining L such that βnR pLq ă 8, we first look at the most basic
case of a single rectangular polyomino. The result we prove allows all translations,
reflections, and rotations of the aˆb polyomino. Necessary and sufficient conditions
for an a ˆ b rectangular polyomino to tile Rnˆm have been worked out previously
in the literature and this can be used to decide when a rectangular polyomino is
n-Noetherian for a particular n. In particular, we have the following theorem.
Theorem 2.7 ([9, 26]). An a ˆ b polyomino tiles Rnˆm if and only if the following
conditions are satisfied.
(a) Either m or n is divisible by a, and either m or n is divisible by b.
(b) Both m and n can be written as linear combinations of a and b with coefficients
in N.
In the following theorem, the property of being Noetherian actually comes from
the fact that if we are given two natural numbers a, b with greatest common divisor
d, then for m sufficiently large and divisible by d, it is always possible to write
m “ ax ` by for x, y P N using the Euclidean algorithm. Let us denote the a ˆ b
polyomino by paˆb . The following result follows directly from Theorem 2.7.
Theorem 2.8. The set L “ tpaˆb u is n-Noetherian for every n P N.
Proof. There are three cases we consider. The first is that n is divisible by neither
a nor b. Then if any tiling is to exist, n “ ax ` by, with x, y P N, otherwise
no tilings exist and paˆb is trivially n-Noetherian. If n is of such a form, then if
paˆb tiles Rnˆm , m must be divisible by ab and can trivially be written as a linear
combination of a and b with coefficients in N. Then if m ą ab and paˆb tiles Rnˆm ,
we have that m “ dab for d ą 1. Let d “ pd1 ` d2 q, d1 , d2 ě 1 and m1 “ d1 ab and
m2 “ d2 ab. Then by Theorem 2.7, Rnˆm1 and Rnˆm2 are both tilable by paˆb
and so for any tiling of Rnˆm with m ą ab, the tiling admits a subtiling allowing
rotations.
Now suppose n is divisible by exactly one of a and b, without loss of generality
let it be a. Then trivially n can be written as linear combination of a and b with
coefficients in N. Then if paˆb tiles Rnˆm , we have that b divides m. Then we
repeat the argument above for when m ą b.
The last case is when ab divides n. Once again, n can be written as a linear
combination of a and b with coefficients in N. Let gcdpa, bq “ d . Then if paˆb tiles
Rnˆm , we have that m is divisible by d by condition (b) of Theorem 2.7. There
exists an N P N (which we may assume is divisible by d) such that for every m ě N
divisible by d, m “ ax ` by for some x, y P N. Thus for m ě 2N divisible by
d, we know that Rnˆm admits a subtiling allowing rotations since we can choose
ON SUBTILINGS OF POLYOMINO TILINGS
6
m1 , m2 ě N , both divisible by d, and Rnˆm1 and Rnˆm2 are tilable by paˆb by
Theorem 2.7.
It is not too difficult to come up with bounds for βn ppaˆb q. If n is not divisible
by a nor b, then βn ppaˆb q “ 2ab. Similarly, if n is divisible by precisely one of a
or b, say a, then βn ppaˆb q “ 2b. However, if ab divides n, it is a bit trickier to
determine βn ppaˆb q.
2.3. Sets of Tall Rectangular Polynominoes. There is another family of rectangular polyominoes we can show are strongly n-Noetherian by elementary arguments. Let L be a set of rectangular polyominoes such that every element has
height ą n{2 and at most one polyomino of unit height. Let the polyomino of unit
height have width dividing the greatest common divisor of the widths of all of the
other polyominoes in L. Then we have the following theorem.
Theorem 2.9. If L is as above, then it is strongly n-Noetherian. Furthermore,
βn pLq equals the largest width of any polyomino in L.
Proof. Let us first focus on those polyominoes of height ą n{2. It is clear that no
two such polyominoes can occupy space in the same column of Rnˆm . As such, we
shall arrange these polyominoes in Rnˆm as follows.
n n´1
¨ ¨ ¨ t n2 u ` 1
In the picture, the rectangle labeled n denotes a contiguous block of height n
rectangular polynominoes. The same for n ´ 1, all the way down to t n2 u ` 1. The
grey area is then filled with the copies of the polyomino of unit height. This can
always be done as the width of this polyomino divides the width of any of the
polyominoes of height ą n{2 and thus divides the width of any of these contiguous
blocks.
This rearrangement admits a subtiling. The subtiling Rnˆm1 can be formed by
taking any polyomino p of height ą n{2 and tiling the space above it with the
polyomino of unit height; the remaining rectangle Rnˆm2 may be tiled in any way.
The tiling of Rnˆm1 described admits no subtiling and m1 must in fact be the
width of p. This shows that βn pLq is in fact the maximum width of any rectangular
polyomino in L of height ą n{2.
3. Rephrasing the problem in terms of rings generated by monomials
We now recast this problem in terms of rings, justifying the use of the term
“Noetherian”. Let us first restrict ourselve to rearrangements only allowing translations. Let L be a finite collection of polyominoes. We view each polyomino as an
algebraically independent variable. We can form the formal commutative ring over
C generated by elements of L, denote it CrLs.
However, let us consider the subring Rn rLs Ď CrLs, whose homogeneous elements are equivalence classes of tilings of Rnˆm for a fixed n. Two tilings are
equivalent if one can be rearranged to give the other using translations. This is
equivalent to the commutativity of CrLs. Equivalence classes of tilings are monomials in the ring CrLs. In the following discussion we will say “tiling” instead of
“equivalence class of tilings” when it is clear from context.
ON SUBTILINGS OF POLYOMINO TILINGS
7
Given tilings of Rnˆm1 and Rnˆm2 , denoted T1 and T2 , respectively, we define
T1 ˚ T2 to be the tiling of Rnˆpm1 `m2 q given by adjacent juxtaposition of T1 and
T2 . There are two ways to naturally do this, but they are in the same equivalence
class and so this product is well defined.
Now suppose we wanted to allow reflections and rotations in our rearrangements.
Then we would make sure that L is closed under rotations and reflections. Then
we consider a quotient ring of Rn rLs by imposing the relation that any polyomino
p (viewed as a variable in CrLs) is equal to the polyominoes that can be reached
by rotating and reflecting p. We denote this ring by R̃n rLs.
It is clear that if L is strongly n-Noetherian then Rn rLs is Noetherian. This is
because every tiling of Rnˆm for m ą βn pLq can be written as a product of tilings
of Rnˆm1 and Rnˆm2 with m “ m1 ` m2 . So Rn is generated by finitely many
monomials. Likewise, R̃n rLs is Noetherian if L is n-Noetherian. We now show that
the converse holds as well.
Proposition 3.1. The ring Rn rLs is Noetherian if and only if L is strongly nNoetherian. Likewise, R̃n rLs is Noetherian if and only if L is n-Noetherian.
Proof. We have already discussed one direction for the two rings. Now let us
suppose that Rn rLs is Noetherian. Since Rn rLs is ring generated by monomials
in algebraically independent variables, it is well know that the ideal of relations
are generated by those of the form X1 ˚ X2 “ Y1 ˚ Y2 where X1 , X2 , Y1 , and
Y2 are monomials. These are called binomial relations. This corresponds to the
observation that a monomial may be factored into smaller monomials non-uniquely.
Since the ideal of relations is a homogeneous ideal, if Rn rLs is Noetherian, then it
must be the case that for sufficiently large N , if T is monomial of degree at least
N then T “ T1 ˚ T2 . This implies that L is strongly n-Noetherian. For the ring
R̃n rLs, the argument is the same where we note that the kernel of the map taking
Rn rLs to R̃n rLs was also generated by binomial relations.
Proving the ring CrLs is Noetherian for various L is too hard, so we instead use
another ring whose Noetherianess will imply CrLs is Noetherian. This ring will not
be isomorphic to CrLs in general and will restrict the translations we can make in
rearranging tilings to only horizontal translations. To construct this new ring, we
assume that L only contains rectangular polyominoes.
Given n P N, let Ipnq :“ tpℓ, ℓ ` 1, . . . , ℓ ` kq| ℓ, ℓ ` k P rnsu, where rns “
p1, . . . , nq, be the set of all contiguous intervals contained in rns. For an interval
ι “ pℓ, . . . , ℓ`kq P Ipnq, let |ι| “ k be the length of that interval. Given a polyomino
of height k, we will create several different variables associated to it, one for each
element of Ipnq of length k. Let ω be the maximum width of any polyomino in L.
Let us define the following vector space:
à
EndpVι q,
V pnq :“
ιPIpnq
EndpVι q :“
|ι|
â
EndpCω q.
j“1
Â
The coordinate ring of V pnq is isomorphic to
ιPIpnq CrEndpVι qs, where the
ring CrEndpVι qs is the coordinate ring of EndpVι q. Let us fix ι P Ipnq. Then for
every polyomino p P L of height |ι|, we associate to it the polynomial TrpMιw q
ON SUBTILINGS OF POLYOMINO TILINGS
8
where w is the width of p and Mι is the matrix of variables in EndpVι q. Then let
CrLsι Ď CrEndpVι qsÂbe the subring generated by 1 and these polynomials. Then
we define Pn rLs :“ ιPIpnq CrLsι which is a subring of the coordinate ring of V pnq.
We can think of Pn rLs as a ring generated by tuples pp, ιq where p P L and ι
is a tuple containing the rows of Rnˆm p must be placed in. For this reason, the
rearrangements will only consist of horizontal translations since row data is implicit
in each variable. We would now like to show that these variables are algebraically
independent. For that we need the following result.
Theorem 3.2 ([32, 27]). For the generic matrix M P EndpCn q, the polynomials
TrpM q, . . . , TrpM n q are algebraically independent.
Proposition 3.3. The ring Pn rLs is a Noetherian polynomial ring.
Proof. That
 Pn rLs is Noetherian is clear as it has finitely many generators. Since
Pn rLs “ ιPIpnq CrLsι , it suffices to show that CrLsι is a polynomial ring for any
ι P Ipnq. We know that CrLsι is generated by polynomials of the form TrpMιw q
where w ď ω. The matrix Mι is the generic p|ι| ¨ ωq ˆ p|ι| ¨ ωq matrix and so by
Theorem 3.2, these polynomials are algebraically independent. Thus CrLsι is a
polynomial ring.
Â
Â
n
n
Consider the vector space EndpCωn q – i“1 EndpCω q :“ i“1 EndpWi q. Let
pλ1 , . . . , λn´1 q P pCˆ qn´1 act on EndpWi q by multiplication by λi if i ‰ n and
otherwise by multiplication by pλ1 ¨ ¨ ¨ λn´1 q´1 . This induces an action on Pn pLq as
follows. Given the generator TrpMιw q, pλ1 , . . . , λn´1 q multiplies this polynomial by
ź
λw
i if n R ι and
iPι
pλ1 ¨ ¨ ¨ λn´1 q´w
ź
λw
i otherwise.
iPιztnu
The group pCˆ qn´1 is a semi-simple algebraic group and this action is clearly
rational. This means we can use Hilbert’s famous result on invariant rings.
Theorem 3.4 ([19, 20]). For a commutative Noetherian C-algebra R acted on
rationally by semi-simple algebraic group G, the invariant ring RG is Noetherian.
We now look at the closely related action of pλ1 , . . . , λn q P pCˆ qn on EndpWi q
by multiplication by λi , even when i “ n. This action again extends to an action on Pn rLs. This action induces a multigrading on the ring. Given a monomial X P Pn rLs, we say that is is multihomogeneous of degree pw1 , . . . , w ` nq if
wn
1
pλ1 , . . . , λn q.X “ pλw
1 ¨ ¨ ¨ λn qX.
Definition 3.5. We say that a set of rectangular polyominoes L is n-representable
if p1 ¨ ¨ ¨ pℓ is a monomial ř
in Pn rLs of homogeneous degree pw1 , . . . , wm q, then there
n
is a tiling of Rnˆm , m “ i“1 wi by the polyominoes associated to p1 , . . . , pℓ which
respects their row assignments.
In the above definition, we remind the reader that polynomials p1 , . . . , pℓ correspond to polyominoes and a designation of which rows the polyomino must be
placed in. We need that multihomogeneous polynomials always correspond to
tilings of rectangular regions. We can now present our main result.
Theorem 3.6. If L is n-representable, then L is strongly n-Noetherian.
ON SUBTILINGS OF POLYOMINO TILINGS
9
Proof. The ring Pn rLs is Noetherian and by Theorem 3.4, so is Pn rLsGn , where
Gn “ pCˆ qn´1 acts on Pn rLs as described above. It is clear that the homogeneous
elements of Pn rLsG are the multihomogeneous elements of degree pm, . . . , mq for
m P N. Then if L is n-representable, the homogeneous elements of Pn rLsGn correspond to equivalence classes of tilings of rectangular regions. Since this ring is
Noetherian and generated by monomials of algebraically independent variables (by
Proposition 3.3), we can use the same argument as in the proof of Proposition 3.1
to say that every monomial of sufficiently large degree must factor into two monomials of smaller degree. This implies that L is n-Noetherian and furthermore, the
arrangements only allow horizontal translations. So it is strongly n-Noetherian.
4. Examples of representable sets of polyominoes
In the previous section we showed that n-representable sets of polyominoes are
strongly n-Noetherian. However, the condition of being n-representable is rather
opaque. In this section, we discuss some concrete sets of polyominoes that are
n-representable for specific n.
Let us first expound a bit further on the definition of n-representable. Suppose
we are given a multiset P “ tppaˆb , ιq| paˆb P L, ι P Ipnqu where L is finite set
of rectangular polyominoes. As before, we use paˆb to mean the a ˆ b rectangular
polyomino. The tuple ppaˆb , ιq indicates the rows in Rnˆm that paˆb should span.
Then suppose the following conditions hold:
ÿ
ab “ nm and
ppaˆb ,ιqPP
ÿ
b “ m for all 1 ď i ď n.
(1)
ppaˆb ,ιqPP
iPι
Then the set L is n-representable if whenever the above two conditions are met,
P tiles Rnˆm , respecting the row assignments of the polyominoes,
Observation 4.1. Note that if L consists only of polyominoes of the form 1 ˆ k,
it is n-representable for any n.
Example 4.2. We now give an example of a set of polyominoes that are not
5-representable. Let L “ tp1ˆ3 , p1ˆ2 , p1ˆ1 , p3ˆ1 u. Now consider the following set
P “ tpp1ˆ3 , t1uq, pp1ˆ3 , t5uq, pp3ˆ1 , t1, 2, 3uq, pp3ˆ1, t2, 3, 4uq,
pp3ˆ1 , t3, 4, 5uq, pp1ˆ1 , t3uq, pp1ˆ2 , t2uq, pp1ˆ2 , t4uqu.
We see that sum of the areas of the polyominoes in P is 3 ˚ 5 ` 2 ˚ 2 ` 1 “ 4 ˚ 5.
We also see that
ÿ
b “ 4 for all 1 ď i ď 5.
ppaˆb ,ιqPP
iPι
If L were 5-representable, we should be able to tile R5ˆ4 with polyominoes in P
satisfying the row assignments. We show that this is not the case. First note that
there is a 1 ˆ 3 polyomino in both the top and bottom row. Also, there is a 3 ˆ 1
polyomino that intersects the first row, and another 3 ˆ 1 polyomino that intersects
the last row. So if we could tile R5ˆ4 , there are only two ways that these four tiles
could be placed. One way is shown below. The other way is the vertical reflection
of the following picture.
ON SUBTILINGS OF POLYOMINO TILINGS
10
We now see that there are two ways to place the last 3 ˆ 1 polyomino, which must
span rows 2,3, and 4. But either way we do this, we make it impossible to place
one of the two 1 ˆ 2 polyominoes. So we cannot tile R5ˆ4 using these polyominoes
while respecting their row assignments. So L is not 5-representable.
In Example 4.2, we showed that set was not 5-representable by using polyominoes
to construct a region inside of R5ˆ4 that could not be tiled with polyominoes of
unit height. If we want to prove that sets of polyominoes are n-representable, we
need to understand which regions of R2 can always be tiled with polyominoes of
unit height.
Definition 4.3. A subregion R Ď Rnˆm is called row-convex if for any two points
x, y P R that can be connected by a horizontal line, the convex hull of those two
points lies entirely within R. R inherits rows and columns from Rnˆm . Let |Ri |
denote the width of Ri , the ith row of R.
Lemma 4.4. Let R be a row-convex region and consider any multiset of the form
P “ tppaˆb , ιq| p1ˆb P L, ι P Ipnqu. Then if
ÿ
b “ |Ri | for all 1 ď i ď n,
pp1ˆb ,ιqPP
iPι
there is a tiling of R by P respecting row assignments.
Proof. Since
ÿ
b “ |Ri | for all 1 ď i ď n,
pp1ˆb ,ιqPP
iPι
we can simply place those polyominoes assigned to row i in any order we want
since R is row-convex. The tiling of each row does not affect any other row, so R
is tilable by P respecting row assignments.
If L is a set of rectangular polyominoes, L is n-representable if when tiling Rnˆm
and Equations 4 hold, those polyominoes which are not unit height can be always
be arranged such that the remaining untiled region is row-convex. Then appealing
to Lemma 4.4, Rnˆm can be tiled.
Theorem 4.5. For n P N, if L consists of rectangular polyominoes whose height
is either 1 or at least n ´ 1, then L is n-representable. Furthermore, if n “ 4, then
any set of rectangular polyominoes where all of the polynominoes of height two have
the same width is n-representable.
Proof. Let P be a multiset satisfying Equations 4 for n, m P N. To prove the
theorem, we simply need to show that those polyominoes of height ě n ´ 1 can
be translated horizontally in their assigned rows in Rnˆm such that the remaining
untiled region is row-convex. Then we appeal to Lemma 4.4.
Let us first handle the case that all polyominoes have at at least n ´ 1 or 1. We
then arrange those polyominoes in P as in the following picture.
n n´1
n´1
ON SUBTILINGS OF POLYOMINO TILINGS
11
Here the numbers in the boxes indicate that the box is a contiguous block of polyominoes of that height. Every possible row assignment of every polyomino of height
n ´ 1, n is shown in this picture. The grey region is the remaining untiled area. As
we can see, it is row convex. Therefore, any set of rectangular polyominoes meeting
the specified height conditions is n-representable.
Now let n “ 4. We once again need to rearrange the polyominoes of height ě 2
in such a way that a row-convex region is left untiled. There are two scenarios,
both pictured below.
2
2 3 2
2 3
2
2
2 3
2 3
Once again, the boxes represent contiguous blocks of polyominoes of the designated height and every row assignment is accounted for. The grey region is the
part left untiled and in both pictures is row-convex. We note that we need the
assumption that all polyominoes of height two have the same width to ensure that
the two height two regions on the left hand side of the picture can be assumed to
have equal width.
Corollary 4.6. For n P r3s, every set of rectangular polyominoes is strongly nNoetherian. Furthermore, if L consists only of rectangular polyominoes of height 1
or height at least n ´ 1, then L is strongly n-Noetherian. Lastly if n “ 4 and all
polyominoes of height two in L have the same width, it is 4-Noetherian.
Proof. This follows immediately from Theorem 3.6 and Theorem 4.5.
It is interesting to note that as long as n ď 3, every set of rectangular polyominoes
is n-representable. However, this does not continue to hold beyond n “ 5 as
was shown in Example 4.2. Examples can also be found of sets of rectangular
polyominoes that are not 4-representable. However, we do not know if Example 4.2
is strongly 5-Noetherian or not.
5. Finding upper bounds on subtilings
For an n-representable set L, finding a bound on βn pLq is equivalent to bounding
the degree of the generators of the corresponding ring Pn rLsGn , where as before
Gn “ pCˆ qn´1 . There are results from invariant theory about bounding the degrees
of generators. We look at the most general theorem first, before specializing to
a particular case. For the sake of brevity, we assume some basic knowledge of
algebraic geometry in the subsequent exposition.
We first need to realize Gn as an affine variety. The set Cn´1 is clearly an affine
variety with coordinate ring Crx1 , . . . , xn´1 s. Then the ring Crx1 , . . . , xn´1 , ts{I,
where I “ xxi t ´ 1, i P rn ´ 1sy, is the coordinate ring of pCˆ qn´1 , as each element
must now be invertible. So Gn is an affine subvariety of Cn defined by quadratic
polynomials.
Given a rational representation ρ : G Ñ GLpV q, we have a map g ÞÑ taij pgqu
where taij pgqu is a matrix in GLpV q and aij pgq are rational functions in g. Let
Aρ “ maxpdegpaij qq.
ON SUBTILINGS OF POLYOMINO TILINGS
12
Theorem 5.1 ([10]). Let G be an affine variety of Cn defined by polynomials
h1 , . . . , hℓ . Let ρ : G Ñ GLpV q be a rational representation of G with finite kernel.
Define H “ maxpdegphi qq and let m “ dimpGq. Then CrV sG is generated by
polynomials of degree at most
maxp2,
3
dimpCrV sG qH n´m Am
ρ q.
8
In the above theorem, the Krull dimension of CrV sG appears in the formula.
In general, we do not know this dimension for Pn rLsGn , but is upper bounded by
the Krull dimension of Pn rLs. Since Pn rLs is polynomial ring by Proposition 3.3,
SpecpPn rLsq is isomorphic to affine space and the Krull dimension of Pn rLs is equal
to the dimension of this space.
Given a set of rectangular polyominoes L, let us define hn pLq “ ta ď n| paˆb P
Lu and Lpaq “ tpaˆb | paˆb P Lu. Note that the polyomino paˆb can be placed
in n ´ a ` 1 different contiguous rows in RÀ
nˆm , provided a ď n. Since Pn rLs “
Â
CrLs
,
SpecpP
rLsq
is
isomorphic
to
ι
n
ιPIpnq SpecpCrLsι q. Since SpecpCrLsι q
ιPIpnq
is also isomorphic to affine space, by Proposition 3.3,
ř its dimension is the number
of polyominoes in L of height |ι|. So dimpPn rLsq “ iPhn pLq pn ´ i ` 1q|Lpiq|.
Theorem 5.2. Given an n-representable set L, let A be the maximum area of a
polyomino in L with height not exceeding n, then
ˆ
˙
ÿ
3
βn pLq ď max 2, An´1 ¨
pn ´ i ` 1q|Lpiq|
4
iPhn pLq
Proof. This will follow from Theorem 3.6 and Theorem 5.1 after we substitute in
the appropriate quantities. We note that H “ 2 as Gn is defined by quadratics.
Furthermore, m “ n ´ 1 and t “ n. By the way Gn acts on Pn rLs, we see that the
maximum degree of a monomial is equal to the maximum area of polyomino in L
with height not exceeding n. Thus Aρ “ A. Lastly
ÿ
Pn rLsGn ď
pn ´ i ` 1q|Lpiq|.
iPhn pLq
5.1. The case where all polyominoes have unit height. Let us look at the
special case where L consists only of polyominoes of the form 1 ˆ k. As already
observed, such a set is n-representable for every n. For this special case, we can
actually realize Pn rLsGn in a different way. Let n be fixed and ω the maximum
width of a polyominoÂin L. Now we look at the subvariety V Ă EndppCω qbn q of
n
matrices of the form i“1 Mi where each Mi P EndpCω q.
n
Now let GLn,ω :“ ˆi“1 GLpCω q act on V by
pg1 , . . . , gn q.
n
â
i“1
Mi :“
n
â
gi Mi gi´1 .
i“1
If CrVs denotes the coordinate ring of V, then CrVsGLn,ω is the corresponding
invariant ring. We claim that Pn rLsGn is isomorphic to a quotient ring of CrVsGLn,ω .
n
Definition 5.3. For σ “ pσ1 , . . . , σn q P Sm
, let σi “ pr1 ¨ ¨ ¨ rk qps1 ¨ ¨ ¨ sl q ¨ ¨ ¨ be a
n
disjoint cycle decomposition. For such a σ P Sm
, define Trσ “ Tσ1 ¨ ¨ ¨ Tσn on V,
ON SUBTILINGS OF POLYOMINO TILINGS
where
Tσi p
n
â
13
Mj q “ TrpMik qTrpMil q ¨ ¨ ¨ .
j“1
We see that the functions in the above definition are precisely the polynomial
invariants of Pn rLsGn when L is the set of rectangular polyominoes tp1ˆk | k P rωsu.
Theorem 5.4 ([37]). The invariant ring CrVsGLn,ω is generated by the polynomials
n
Trσ for σ P Sm
such that the disjoint cycle decomposition of σ contains only cycles
of length at most ω.
The above theorem tells us that if L “ tp1ˆk | k P rωsu, then Pn rLsGn “
CrVsGLn,ω . In general, let us define wpLq “ tk| 1 ˆ k P Lu and
Ântake the subvariety
of V such that for any M P V and any decomposition M= i“1 Mi , TrpMiℓ q “ 0
for ℓ P rωs, ℓ R wpLq, for all 1 ď i ď n. This defines another GLn,ω -stable
subvariety of V, denote it VL . This means that CrVL sGLn,ω is a quotient ring of
CrVsGLn,ω [31]. As such, a degree bound on CrVsGLn,ω gives a degree bound on
CrVL sGLn,ω – Pn rLsGn .
We can use some classical tools from invariant theory to find bounds for this
ring. We introduce some definitions.
Definition 5.5. The null cone of an action G ñ V is the set vectors v such that
0 P G.v (the Zariski closure of the orbit of v). We denote it by NV . Equivalently,
NV are those v P V such that f pvq “ f p0q for all invariant polynomials f .
If one wants to compute the generators of an invariant ring, this can be accomplished by determining the null cone. If one knows precisely the generators of the
ideal of the null cone, then these are precisely the generators of the invariant ring
[19, 12]. However, it usually easier to find sets of invariant polynomials whose vanishing locus is NV . For that, one can use the following result (similar to Theorem
5.1).
Theorem 5.6 ([11]). Let G ñ V , be a reductive group acting rationally on a vector
space. Let f1 , . . . , fℓ be homogeneous invariants, with maximum degree p, such that
their vanishing locus is NV . Then CrV sG is generated by polynomials of degree at
most
3
maxp2, dimpCrV sG qp2 q.
8
In the proof of the above Theorem, the same arguments can be used to prove
the following slightly stronger statement that we shall need.
Corollary 5.7. Let G ñ V be a reductive group acting rationally on a vector
space and X Ď V be a G-stable subvariety. If CrXsG is Cohen-Macaulay, then let
f1 , . . . , fℓ be homogeneous invariants of maximum degree p whose vanishining locus
is the null cone of the action of G on X. Then CrXsG is generated by polynomials
of degree at most
3
maxp2, dimpCrV sG qp2 q.
8
We now need to prove that CrVL sGLn,ω is Cohen-Macaulay. However, we know
that this ring is isomorphic to Pn rLsGn which is Cohen-Macualay by the following
theorem.
ON SUBTILINGS OF POLYOMINO TILINGS
14
Theorem 5.8 ([21]). For a reductive group G acting rationally on a vector space
V , CrV sG is Cohen-Macaulay.
When studying orbit closures, the following theorem is a powerful tool when
dealing with reductive groups.
Theorem 5.9 (The Hilbert-Mumford Criterion [25]). For a linearly reductive group
G acting on a variety V , if G.wzG.w ‰ H, then there exists a v P G.wzG.w and
a 1-parameter subgroup (or cocharacter) λ : k ˆ Ñ G (where λ is a homomorphism
of algebraic groups), such that limtÑ0 λptq.w “ v.
A cocharachterÂof GLn,ω is a product of cocharachters λptq “ ˆni“1 λi ptq of
n
GLpCω q. If λptq.p i“1 Mi q Ñ 0, it is easy to see that for some i, λi ptq.Mi Ñ 0.
This means that Mi is in the null cone of GLpCω q ñ EndpCω q by conjugation. It
is well known that Â
this implies that Mi is nilpotent (cf. [5]). So the null cone of V
n
are those matrices i“1 Mi where at least one Mi is nilpotent.
Definition 5.10. Given a tuple pd1 , . . . , dn q, let d “ lcmtd1 , . . . , dn u. Then define
n
n
ź
â
Trd1 ,...,dn p Mi q :“
TrpM di qd{di .
i“1
i“1
GLn,ω
Theorem 5.11. The ring CrVL s
is generated by polynomials of degree at most
3 n
maxp2, pω qlcmtw| w P wpLqu2 q.
8
Proof. We first prove that the vanishing locus of Trd1 ,...,dn is precisely that null
cone. It is clear that every element ofÂ
the null cone is in the vanishing locus of
these polynomials. Now suppose M “ ni“1 Mi is not in the null cone. Then for
each i, there is a ki P rωs such that TrpMiki q ‰ 0 as Mi is not nilpotent. Then
Trk1 ,...,kn pM q ‰ 0. So the vanishing locus of these polynomials is precisely the
null cone. But we note that by restricting these functions to the variety VL , the
Trd1 ,...,dn is identically zero unless all di P wpLq and so can be excluded from the
set of polynomials whose vanishing locus defines the null cone. Thus the degree of
these polynomials is bounded by lcmtw| w P wpLqu. The dimension of CrVL sGLn,ω is
upper bounded by the dimension of CrEndpCw qbn s “ ω n , completing the proof.
Corollary 5.12. If L is a set of rectangular polynomials all of unit height then
3
βn pLq ď maxp2, pω n qlcmtw| w P wpLqu2 q
8
where ω is the maximum width of any polyomino in L.
Proof. This follows from Theorem 5.11 and the observation that Pn rLsGn is iso
morphic to a quotient of CrVL sGLn,ω .
6. Conclusion
We end this paper with several questions. We introduced the notion of nrepresentable sets L, and found some families of examples. However, we would
be interested in knowing if other examples exist, possibly by choosing rectangles in
L more carefully than simply restricting their heights.
Of course, n-representability seems a much stronger condition than being strongly
n-Noetherian. In our proofs, we only allowed horizontal translations. If all translations and rotations are allowed, it is tempting to believe that every set of rectangular
ON SUBTILINGS OF POLYOMINO TILINGS
15
polyominoes is n-Noetherian. However, it doesn’t seem terribly unlikely that this
is not the case as well. So a question of interest is if there exists a set of rectangular
polyominoes L that are not (strongly or otherwise) n-Noetherian.
We restricted ourselves to rectangular polyominoes, but what if other relatively
nice polyominoes are allowed. Does there exists a set of polyominoes L that includes
a hook, for example, that are (strongly) n-Noetherian? We also looked only at
vertical faults, but another natural variant is to ask whether a tiling of Rnˆm can
be rearranged to give a vertical or horizontal fault.
Lastly, while we were able to give upper bounds, they may be far from optimal.
And we did not give any lower bounds. The least common multiple of the widths
of polyominoes on L is an easy lower bound to see. However, it it s not clear how
sharp this lower bound is in general.
Acknowledgments. The author would like to thank Viresh Patel, Guus Regts,
and Lex Schrijver for their helpful discussions. The research leading to these results
has received funding from the European Research Council under the European
Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement
No 339109.
References
[1] Frank W Barnes. Algebraic theory of brick packing i. Discrete Mathematics, 42(1):7–26, 1982.
[2] Frank W Barnes. Algebraic theory of brick packing ii. Discrete Mathematics, 42(2-3):129–144,
1982.
[3] Danièle Beauquier, Maurice Nivat, Eric Rémila, and Mike Robson. Tiling figures of the plane
with two bars. Computational Geometry, 5(1):1–25, 1995.
[4] Robert Berger. The undecidability of the domino problem, volume 66. American Mathematical
Soc., 1966.
[5] Michel Brion. Representations of quivers. 2008.
[6] Lance D Chambers. Practical handbook of genetic algorithms: complex coding systems, volume 3. CRC press, 1998.
[7] Ho-Lin Chen and Ashish Goel. Error free self-assembly using error prone tiles. In DNA
computing, pages 62–75. Springer, 2004.
[8] John H Conway and Jeffrey C Lagarias. Tiling with polyominoes and combinatorial group
theory. Journal of combinatorial theory, Series A, 53(2):183–208, 1990.
[9] Nicholas G de Bruijn. Filling boxes with bricks. The American Mathematical Monthly,
76(1):37–40, 1969.
[10] Harm Derksen. Polynomial bounds for rings of invariants. Proceedings of the American Mathematical Society, 129(4):955–963, 2001.
[11] Harm Derksen and Gregor Kemper. Computational invariant theory, volume 130. Springer
Science & Business Media, 2013.
[12] Harm Derksen and Hanspeter Kraft. Constructive invariant theory, 1997.
[13] Martin Gardner. The Unexpected Hanging. JSTOR, 1969.
[14] Martin Gardner. Mathematical games: The fantastic combinations of John Conway’s new
solitaire game “life”. Scientific American, 223(4):120–123, 1970.
[15] Martin Gardner. More about tiling the plane: the possibilities of polyominoes, polyiamonds,
and polyhexes. Scientific American, 233(2):112–115, 1975.
[16] Solomon W Golomb. Polyominoes. 1965.
[17] Solomon W Golomb. Polyominoes which tile rectangles. Journal of Combinatorial Theory,
Series A, 51(1):117–124, 1989.
[18] Ron L Graham. Fault-free tilings of rectangles. In The Mathematical Gardner, pages 120–126.
Springer, 1981.
[19] D. Hilbert. Über die vollen Invariantensysteme. Math. Ann., 42:313–373, 1893.
[20] David Hilbert. Über die theorie der algebraischen formen. Mathematische Annalen,
36(4):473–534, 1890.
ON SUBTILINGS OF POLYOMINO TILINGS
16
[21] Melvin Hochster and Joel L Roberts. Rings of invariants of reductive groups acting on regular
rings are cohen-macaulay. Advances in mathematics, 13(2):115–175, 1974.
[22] Ed Pegg Jr. The Eternity Puzzle. http://www.mathpuzzle.com/eternity.html . Accessed
February, 2016.
[23] Serdar Karademir, Oleg A Prokopyev, and Robert J Mailloux. Irregular polyomino tiling
via integer programming with application in phased array antenna design. Journal of Global
Optimization, pages 1–37, 2015.
[24] Richard M Karp. Reducibility among combinatorial problems. Springer, 1972.
[25] George R Kempf. Instability in invariant theory. Annals of Mathematics, pages 299–316,
1978.
[26] David A Klarner. Packing a rectangle with congruent n-ominoes. Journal of Combinatorial
Theory, 7(2):107–115, 1969.
[27] Hanspeter Kraft and Claudio Procesi. Classical invariant theory, a primer. Lecture Notes,
Version, 2000.
[28] William Rex Marshall. Packing rectangles with congruent polyominoes. Journal of Combinatorial Theory, Series A, 77(2):181–192, 1997.
[29] George Martin. Polyominoes: A guide to puzzles and problems in tiling. Cambridge University
Press, 1991.
[30] Cristopher Moore and John Michael Robson. Hard tiling problems with simple tiles. Discrete
& Computational Geometry, 26(4):573–590, 2001.
[31] Masayoshi Nagata. Invariants of group in an affine ring. Kyoto Journal of Mathematics,
3(3):369–378, 1963.
[32] Claudio Procesi. The invariant theory of nˆ n matrices. Advances in Mathematics, 19(3):306–
381, 1976.
[33] Peter J Robinson. Fault-free rectangles tiled with rectangular polyominoes. In Combinatorial
Mathematics IX, pages 372–377. Springer, 1982.
[34] Raphael M Robinson. Undecidability and nonperiodicity for tilings of the plane. Inventiones
mathematicae, 12(3):177–209, 1971.
[35] Alex Selby. Eternity page. http://www.archduke.org/eternity. Accessed February, 2016.
[36] Yoshiyasu Takefuji. Neural network parallel computing, volume 164. Springer Science & Business Media, 2012.
[37] Jacob Turner and Jason Morton. A complete set of invariants for lu-equivalence of density
operators. arXiv preprint arXiv:1507.03350, 2015.
| 0math.AC
|
Electronic Journal of Statistics
Vol. 10 (2016) 2584–2640
ISSN: 1935-7524
DOI: 10.1214/16-EJS1186
arXiv:1511.06268v2 [math.ST] 7 Nov 2016
The relative effects of dimensionality
and multiplicity of hypotheses on the
F -test in linear regression
Lukas Steinberger
Department of Statistics and Operations Research
University of Vienna
Oskar-Morgenstern-Platz 1
1090 Vienna, Austria
e-mail: [email protected]
Abstract: Recently, several authors have re-examined the power of the
classical F -test in a non-Gaussian linear regression under a “large-p, largen” framework [e.g. 27, 29]. They highlight the loss of power as the number
of regressors p increases relative to sample size n. These papers essentially
focus only on the overall test of the null hypothesis that all p slope coefficients are equal to zero. Here, we consider the general case of testing
q linear hypotheses on the p + 1-dimensional regression parameter vector
that includes p slope coefficients and an intercept parameter. In the case of
Gaussian design, we describe the dependence of the local asymptotic power
function on both the relative number of parameters p/n and the relative
number of hypotheses q/n being tested, showing that the negative effect of
dimensionality is less severe if the number of hypotheses is small. Using the
recent work of Srivastava and Vershynin [23] on high-dimensional sample
covariance matrices we are also able to substantially generalize previous
results for non-Gaussian regressors.
MSC 2010 subject classifications: Primary 62F03, 62F05; secondary
62J05, 60F05, 60B20.
Keywords and phrases: high-dimensional linear regression, F-Test, multiple hypothesis testing, large-p asymptotics.
Contents
1
2
3
4
5
6
Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
Model formulation and main results . . . . . . . . . .
2.1 Technical conditions . . . . . . . . . . . . . . . . .
Testing only a small number of hypotheses . . . . . .
Discussion of the technical assumptions . . . . . . . .
Numerical results . . . . . . . . . . . . . . . . . . . . . .
Proofs of main results . . . . . . . . . . . . . . . . . . .
6.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Asymptotic normality of the noise term . . . . .
6.3 Asymptotic χ2 -distribution of the noise term . .
6.4 Asymptotic behavior of the non-centrality term
6.4.1 The case of Theorem 2.1(i) . . . . . . . .
2584
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2585
2587
2592
2593
2595
2597
2601
2601
2605
2609
2611
2612
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
6.4.2 The case of Theorem 2.1(ii) .
6.4.3 The cases of Theorem 2.1(iii)
Acknowledgements . . . . . . . . . . . . . . . .
A Auxiliary results of Section 4 . . . . . . .
B Proofs of auxiliary results of Section 6.2 .
C Other technical results . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . .
and Theorem 3.1
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
2585
2612
2615
2618
2618
2621
2628
2640
1. Introduction
Following a suggestion of Bai and Saranadasa [1] to investigate classical statistical procedures in high-dimensional settings, Wang and Cui [27] re-examine the
usual F -test in the linear regression model under a “large p, large n” asymptotic
framework. They derive the asymptotic power in a fairly general, non-Gaussian
setting, highlighting the dependence of the local power function on the dimensionality of the problem, i.e., on the limit ρ = lim p/n ∈ (0, 1), where n is sample
size, and p is the number of regressors in the model. In particular, they find that
the rejection probability of the F -test for H0 : Rβ = r0 , where R = [0, Iq ] and
p/n → ρ, q/n → ρ, satisfies
r
√
1−ρ
(1−ν)
→
− 0.
(1.1)
P(Fn > fq,n−p−1 ) − Φ −ζ1−ν + n∆β
2ρ
(1−ν)
Here Fn is the usual F -statistic, fq,n−p−1 is the appropriate F -quantile, Φ is the
cdf of the standard normal distribution, ζ1−ν = Φ−1 (1 − ν) and ∆β = (Rβ −
r0 )0 (RΣ−1 R0 )−1 (Rβ − r0 )/σ 2 is the scaled distance from the null hypothesis.
From this approximation we see that the local asymptotic power of the F -test
depends monotonically on the value of ρ and inflates to the nominal significance
level ν as ρ increases to one. The result of Wang and Cui [27] is consistent with
the derivations of the local asymptotic power in the case of Gaussian errors,
as obtained by Zhong and Chen [29]. Both of these studies consider only the
overall F -test for the null hypothesis that all, or almost all (cf. Condition (C3)
in Wang and Cui [27]) of the p slope coefficients are equal to zero. Also, they do
not consider hypotheses involving the intercept parameter. Here, we extend this
analysis and study the problem of testing q general linear hypothesis (including
also hypotheses on the intercept term), without the restriction that (p − q)/n →
0. In this sense, we examine the effect of the dimension of the null hypothesis
(i.e., the number of linear restrictions being tested) on the local asymptotic
rejection probability of the F -test. We find that when testing the null hypothesis
H0 : R0 γ = r0 , for some q × (p + 1) matrix R0 of rank q ≤ p + 1, such that
p/n → ρ1 and q/n → ρ2 ≤ ρ1 , the rejection probability of the F -test satisfies
s
!
√
(1 − ρ1 )(1 − ρ1 + ρ2 )
(1−ν)
P(Fn > fq,n−p−1 ) − Φ −ζ1−ν + n∆γ
→ 0. (1.2)
2ρ2
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2586
Now the asymptotic rejection probability depends also on the mean µ ∈ Rp
of the random design through ∆γ = (R0 γ − r0 )0 (R0 S −1 R00 )−1 (R0 γ − r0 )/σ 2 ,
where γ = (α, β 0 )0 is the vector of regression coefficients including an intercept
parameter α ∈ R and
1
µ0
S=
.
µ Σ + µµ0
This limiting expression coincides with that in (1.1) if ρ1 = ρ2 and R0 = [0, R].
But (1.2) refines the statement in (1.1) and shows the impact of both the relative
number of regressors ρ1 and the relative number of hypotheses ρ2 . These quantities affect the asymptotic rejection probability monotonically, which is consistent with small sample analyses in the Gaussian error case [cf. 12]. However, in
contrast to the complicated nature of the cdf of the non-central F -distribution
as a function in p, q and the non-centrality parameter, our asymptotic approximation to the rejection probability depends on the quantities ρ1 , ρ2 and ∆γ
only through elementary operations and the Gaussian cdf, and it is valid for a
large class of error distributions. In particular, we see that even if ρ1 is close
to 1, the F -test still has power if ρ2 is sufficiently small. In a second step, and
under slightly more restrictive assumptions on the data generating process, we
also investigate the case where only a very small relative number of hypotheses
q/n is tested, i.e., q is bounded as n → ∞ and ρ2 = 0, and the result in (1.2)
no longer holds.
Our work heavily builds on the ideas of Wang and Cui [27] (hereafter abbreviated as WC). The first part of the present work is concerned with reproducing
their results under substantially more general assumptions. First of all, here we
do not require independence between the random design and the error terms,
but we assume only the usual first and second order specification of conditional
moments of the errors given the design. This extension requires a slight modification of the result of Bhansali, Giraitis and Kokoszka [5] on the asymptotic
normality of certain quadratic forms as applied by WC (cf. Lemma 6.1). Furthermore, we do not assume that the n × p design matrix X, after standardization,
consists of i.i.d. components, as is needed for the application of the famous BaiYin Theorem [2] used by WC in order to control extreme eigenvalues of large
sample covariance matrices. Instead, we apply a recent result of Srivastava and
Vershynin [23] which essentially requires only certain moment restrictions on
the i.i.d. rows of X. For our extensions, we also develop a novel result on the
diagonal entries of a fairly general random projection matrix that might be of
interest on its own (see Lemma 6.3). It has the statistical interpretation that
in a moderately high-dimensional regression the leverage values hi , i.e., the diagonal entries of the projection matrix U (U 0 U )−1 U 0 , where U = [ι, X] is the
design matrix including an intercept column, typically behave like p/n. Finally,
we point out that since we also consider tests on the intercept parameter, the
distribution of the F -statistic, in general, also depends on the mean µ of the
random design vectors x1 , . . . , xn . This causes certain technical complications
due to non-centrality issues which are often avoided in the literature on random
design regression by restricting to the case µ = 0. Here, we present a detailed
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2587
treatment of the general case.
The paper is organized as follows. Section 2 introduces the setup and notation
and presents our main results in Theorem 2.1 and Corollary 2.2, which provide
a precise formulation of the statement in (1.2). In Section 3, we specifically
consider the situation where q is fixed and also provide a unifying result that does
not distinguish between large or small q. Next, in Section 4 we provide a detailed
discussion of our technical assumptions and explain the main differences to those
imposed by WC. The results of an extensive numerical study are reported in
Section 5. Finally, Section 6 provides the basic steps in the proof of our main
results. Some of the more technical arguments are deferred to the appendices.
2. Model formulation and main results
We consider a random array {(yi,n , xi,n ) : 1 ≤ i ≤ n, n ≥ 1} where, for each n ∈
N, the pairs (yi,n , xi,n )ni=1 are i.i.d. observations of a real valued response variable
y1,n and pn -dimensional random regressors x1,n with pn < n − 1, satisfying
E[y1,n |x1,n ] = αn + βn0 x1,n and Var[y1,n |x1,n ] = σn2 ∈ (0, ∞). Equivalently,
writing εi,n = yi,n − E[yi,n |xi,n ], the observations can be represented as
yi,n = αn + βn0 xi,n + εi,n ,
i = 1, . . . , n,
(2.1)
where the (εi,n )ni=1 are i.i.d., satisfying E[εi,n |xi,n ] = 0 and Var[εi,n |xi,n ] =
σn2 . Note that ε1,n does not need to be independent of x1,n . For identifiability,
we also assume that Σn := Var[x1,n ] is positive definite and we define µn :=
E[x1,n ]. Furthermore, we adopt the matrix notation Yn = (y1,n , . . . , yn,n )0 , Xn =
[x1,n , . . . , xn,n ]0 , εn = (ε1,n , . . . , εn,n )0 , γn = (αn , βn0 )0 and Un = [ιn , Xn ], where
ιn = (1, 1, . . . , 1)0 ∈ Rn . For notational convenience we will drop the subscript n
whenever there is no risk of confusion, i.e., we write Y = Yn , X = Xn , α = αn ,
β = βn , etc., keeping in mind that, unless noted otherwise, all quantities to
follow depend on sample size n. With this, the model equations in (2.1) become
Y = U γ + ε.
(2.2)
We want to test a general linear hypothesis on the coefficients γ, i.e.,
H0 : R0 γ = r0
vs.
H1 : R0 γ 6= r0 ,
(2.3)
where R0 is a q × (p + 1) matrix with rank R0 = q ≤ p + 1 and r0 ∈ Rq . Without
restriction we may assume that R0 has orthonormal rows (premultiply (2.3) by
(R0 R00 )−1/2 ). We test H0 by use of the F -statistic Fn defined as
Fn =
(R0 γ̂n − r0 )0 (R0 (U 0 U )−1 R00 )−1 (R0 γ̂n − r0 )/q
,
(Y − U γ̂n )0 (Y − U γ̂n )/(n − p − 1)
(2.4)
provided that all the appearing quantities are well defined, and Fn = 0, otherwise. The F -statistic is then compared to the 1 − ν quantile of an F -distribution
(1−ν)
with q and n − p − 1 degrees of freedom, which we denote by fq,n−p−1 . Here,
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2588
γ̂n = (α̂n , β̂n0 )0 is the OLS estimate in the unrestricted model. We also define
the usual estimator of the error variance σ̂n2 = kY − U γ̂n k2 /(n − p − 1), that
appears in the denominator of the F -statistic.
In Section 6 we prove the following results, involving the scaled distance from
the null hypothesis
∆γ := (R0 γ − r0 )0 (R0 S −1 R00 )−1 (R0 γ − r0 )/σ 2 ,
where
1
µ0
1
1
=
E
µ Σ + µµ0
x1
S=
x01
= E [U 0 U/n] .
A list of further technical conditions is given below in Section 2.1.
Theorem
2.1. Consider the
q linear, homoskedastic model (2.1) and set sn =
1
1
n +1)/n+qn /n)
. Moreover, suppose
2 qn + n−pn −1 and bn = (1−(pn +1)/n)(1−(p
2qn /n
that lim supn pn /n < 1, qn → ∞ and ∆γ = o(qn /n) as n → ∞. If either one of
the following three cases applies then the F -statistic satisfies
√
w
s−1/2
(Fn − 1) − n∆γ bn −−−−→ N (0, 1).
(2.5)
n
n→∞
(i) The assumptions (A1).(a,b,c,d) and (A2) on the random design and on
the error distribution, are satisfied, and either R0 = Ipn +1 for all n ∈ N,
or R0 = [0, Ipn ] for all n ∈ N. (In this case, either qn = pn or qn = pn + 1,
and thus (pn − qn )/n → 0 holds.)
(ii) The assumptions (A1).(a,b,c,d,e) and (A2) on the random design and
on the error distribution, are satisfied, (pn − qn )/n → 0, and (R0 γ −
√
r0 )0 R0 SR00 (R0 γ − r0 )/σ 2 = O( qn /n) holds.1
(iii) Assumption (A2) on the error distribution is satisfied and the design vectors x1,n , . . . , xn,n are i.i.d. Gaussian with mean µn ∈ Rpn and positive
definite covariance matrix Σn .2
By a simple argument involving Polya’s theorem this translates into the following corollary on the rejection probability of the F -test.
Corollary 2.2. If lim supn pn /n < 1 and qn → ∞, as n → ∞, and the conclusion of Theorem 2.1 holds, then the rejection probability of the F -test satisfies
s
√
(1 − pnn+1 )(1 − pnn+1 + qnn )
(1−ν)
→ 0.
P(Fn > fqn ,n−pn −1 )−Φ −ζ1−ν + n∆γ
2qn /n
Here, ζ1−ν = Φ−1 (1−ν) is the 1−ν quantile of the standard normal distribution
and ν ∈ (0, 1) does not depend on n.
1 Notice that this additional requirement implies and strengthens the assumption that
∆γ = o(qn /n). Simply observe that, by block matrix inversion, R0 S −1 R00 = (R0 SR00 −
R0 SR10 (R1 SR10 )−1 R1 SR00 )−1 , where R1 is a (pn + 1 − qn ) × (pn + 1) matrix with orthonormal
rows which are also orthogonal to the rows of R0 . Therefore, ∆γ ≤ (R0 γ − r0 )0 R0 SR00 (R0 γ −
√
r0 )/σ 2 = O( qn /n).
2 It is easy to see that the normality assumption implies Assumption (A1).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2589
Proof. It is easy to see, using Polay’s theorem and Lemma C.8, that f˜n :=
−1/2 (1−ν)
sn (fqn ,n−pn −1 − 1) satisfies f˜n → ζ1−ν . Now use the conclusion of Theorem 2.1, Polya’s theorem and the Lipschitz continuity of Φ to obtain
(1−ν)
|P(Fn > fq,n−p−1 ) − Φ(−ζ1−ν + ηn2 )|
= |P(s−1/2
(Fn − 1) > f˜n ) − Φ(−ζ1−ν + ηn2 )|
n
= |P(s−1/2
(Fn − 1) − ηn2 ≤ f˜n − ηn2 ) − Φ(ζ1−ν − ηn2 )|
n
≤ |P(s−1/2 (Fn − 1) − η 2 ≤ f˜n − η 2 ) − Φ(f˜n − η 2 )|
n
n
n
n
+ |Φ(f˜n − ηn2 ) − Φ(ζ1−ν − ηn2 )|
≤ sup |P(s−1/2
(Fn − 1) − ηn2 ≤ t) − Φ(t)| + o(1) −−−−→ 0,
n
t∈R
where ηn2 :=
√
n→∞
n∆γ bn .
If (pn − qn )/n → 0 and 0 < lim inf n qn /n ≤ lim supn pn /n < 1, then Corollary 2.2 recovers the result of WC under weaker assumptions on the joint distribution of the design and the errors, and for a null hypothesis that possibly
restricts also the intercept parameter α (cf. the assumptions of Theorem 2.1(i)
and (ii)). In this case the factor bn above asymptotically reduces to
s
s
(1 − pnn+1 )(1 − pnn+1 + qnn )
1 − pnn+1
bn =
∼
,
2qn /n
2 pnn+1
as in (1.1). It highlights the dependence of the power function on the relative
number of regressors pn /n. However, since (pn − qn )/n → 0, the individual
roles of pn and qn can not be discerned. This shortcoming is removed here,
but it comes at the price of a stronger design condition (cf. Theorem 2.1(iii)).
It is tempting, however, to conjecture that Assumptions (A1) and (A2) are
actually sufficient also for the general case. Corollary 2.2 nicely shows the effect
of both the dimension of the parameter space as well as the dimension of the
null hypothesis, on the asymptotic power function. In particular, we see that
even in a case where the relative number of regressors pn /n is large, the classical
F -test still has power, as long as we are interested in testing only a relatively
small number of hypotheses. However, we should make a cautionary remark at
this point. In Theorem 2.1 and Corollary 2.2 we have assumed that qn → ∞.
If the number of hypotheses q being tested is too small, then the asymptotic
approximation presented above will not be very accurate, in the same way the
χ2q distribution is not very accurately approximated by the normal if q is small.
Therefore, in the next section we specifically study the case where qn is bounded.
In Theorem 2.1, we treat the special cases of R0 = Ip+1 and R0 = [0, Ip ]
separately, because here it is considerably much easier to deal with the noncentrality term in the decomposition of the F -statistic (see Section 6.4.1). In
particular, in this case we do not need to impose further restrictions on the
distance from the null ∆γ other than that it is of order o(qn /n) as n → ∞ and
we can also work with weaker design conditions (cf. Theorem 2.1(i)).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2590
Remark 2.3 (On the detection boundary of the F -test). In the classical setting,
where q and p are fixed, while n goes to infinity, it is well known that the
detection boundary of the F -test is n−1/2 . This means that a violation of the
null hypothesis H0 : R0 γ = r0 that is of the order kR0 γ − r0 k2 n−1/2 leads
to non-trivial asymptotic power, while a slower order yields asymptotic power
equal to the size of the test and a larger order yields asymptotic power equal
to one. However, in general, when q = qn and p = pn are allowed to grow with
sample size n, the detection boundary of the F -test is no longer n−1/2 but rather
1/4
qn /n1/2 . To see this, first we ignore the influence of the nuisance parameters
and set µ = 0, Σ = Ip and σ 2 = 1, so that S = Ip+1 and ∆γ = kR0 γ − r0 k22 .
From Corollary 2.2, we see that in order to obtain non-trivial asymptotic power,
i.e., asymptotic power in the open interval (ν, 1), the non-centrality term
!2 s
√
n1/2
1
pn + 1
pn + 1 qn
n∆γ bn =
kR
γ
−
r
k
1
−
1
−
+
,
0
0 2
1/4
2
n
n
n
qn
has to stay away from 0 and ∞. If we exclude the pathological case lim supn pn /n =
1, then this requirement is met if the violation of the null hypothesis is of the
1/4
order kR0 γ − r0 k2 qn /n1/2 .
Remark 2.4 (Gaussian errors and fixed design). In the classical case where the
error ε follows a spherical normal distribution which is independent of the design,
the F -statistic (2.4) follows a non-central F -distribution with qn and n − pn − 1
degrees of freedom and non-centrality parameter ςn2 = n∇n , conditional on X,
where ∇n = (R0 γ − r0 )0 (R0 (U 0 U/n)−1 R00 )−1 (R0 γ − r0 )/σ 2 [cf. 22, page 41].
Nevertheless, even in this traditional case, only basic monotonicity results are
(1−ν)
available for the power function P(Fn > fqn ,n−pn −1 |X) as a function of qn ,
2
n − pn − 1 and ςn [e.g., 12, 26]. In Section 6.4 we investigate ∇n as n → ∞, such
that lim supn pn /n < 1. Our results provide approximations for the average (or
(1−ν)
unconditional) rejection probability, i.e., for E[P(Fn > fqn ,n−pn −1 |X)], which
are given by the Gaussian cdf applied to an elementary function in pn /n, qn /n
and ∆γ = (R0 γ − r0 )0 (R0 S −1 R00 )−1 (R0 γ − r0 )/σ 2 and which are therefore easy
to interpret (cf. Corollary 2.2).
Remark 2.5 (On omitted variable misspecification). One major motivation
for us to extend the result of WC to scenarios where there is some dependence between the design and the errors, and also among the components of
the standardized design vectors themselves (see Section 4 for details), was to
treat simple sub-models of high-dimensional linear models. These sub-models
typically exhibit misspecification due to omitted regressor variables. Consider
an i.i.d. sample (yi , zi )ni=1 from the high-dimensional linear model
yi = θ0 zi + ui ,
where the zi are random d-vectors with d n that are independent of the ui ,
which satisfy E[ui ] = 0 and E[u2i ] ∈ (0, ∞). Moreover, assume that the marginal
distribution of the regressors zi can be represented as zi = Ω1/2 z̃i , where Ω1/2 is
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2591
symmetric and positive definite and z̃i has a Lebesgue density fz̃ which is such
that the components of z̃i are independent with zero means, unit variances and
bounded 8-th moments, so that in particular Cov[zi ] = Ω.
Now suppose we want to use only a small number p of the d available regressors, with p < n, so that classical regression methods are feasible within this
subset regression problem. These working regressors can be described as
xi = M 0 zi ,
where M is a d×p matrix of full rank p < d. For instance, M could be a selection
matrix so that xi consists of a certain choice of p components of zi . In such a
situation, the sample (yi , xi )ni=1 need not follow a linear homoskedastic model
as in (2.1), because, in general, the conditional expectation E[yi |xi ] is not linear
in xi and the conditional variance Var[yi |xi ] is not constant if the pair (yi , xi )
is not jointly Gaussian. However, one can always write
yi = β̃ 0 xi + ξi ,
where β̃ = argminb E[(yi − b0 xi )2 ] and ξi = yi − β̃ 0 xi . Here, the parameter β̃
corresponds to the best linear predictor of yi given xi , and it may be of interest
to test whether H0 : β̃ = 0, i.e., whether the selected regressors xi = M 0 zi have
any value for linearly predicting the response variables yi .
Clearly, in the present setting the errors ξi are not independent of the design
vectors xi and, in general, E[ξi |xi ] 6= 0 and Var[ξi |xi ] 6= Var[ξi ]. Now consider
the unobservable corrected sample (yi∗ , xi )ni=1 , where
s
Var[ξi ]
∗
0
(ξi − E[ξi |xi ]).
yi = β̃ xi +
Var[ξi |xi ]
The corrected sample clearly follows a linear homoskedastic model as in (2.1),
because E[yi∗ |xi ] = β̃ 0 xi and Var[yi∗ |xi ] = Var[ξi ]. Moreover, the corrected sample satisfies Assumption (A1) in view of Lemma A.2 applied with Γ = M 0 Ω1/2 ,
µ = 0 and mn = d, and because the design matrix X = [x1 , . . . , xn ]0 has a
Lebesgue density on Rn×p . Of course, in general, the actual sample (yi , xi )ni=1
may be very different from the corrected sample (yi∗ , xi )ni=1 , but the results
of Steinberger and Leeb [24] suggest that if d p, then E[ξi |xi ] ≈ 0 and
Var[ξi |xi ] ≈ Var[ξi ], at least for a large collection of selection matrices M . Therefore, the observed sample and the corrected sample should be very similar if d
is large relative to p, and one might expect that also for a sufficiently regular
statistic T , we have T ((yi , xi )ni=1 ) ≈ T ((yi∗ , xi )ni=1 ). We suspect that the results
of Steinberger and Leeb [24] can also be used to establish
p the validity of Assumption (A2) for the corrected sample, with ei,n := Var[ξi ]/ Var[ξi |xi ] and
ε̃i,n := ξi − E[ξi |xi ]. Thus, we expect the F -test for H0 : β̃ = 0 based on the
sample (yi , xi )ni=1 to be asymptotically valid and even possess similar power as
the F -test in a correctly specified model, for most choices of M , provided that
d = dn and p = pn tend to infinity along with n, such that dn pn . The details
of this line of reasoning will be further developed elsewhere.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2592
2.1. Technical conditions
Throughout this paper, the readerP
will encounter several different norms. For
k
vectors v ∈ Rk we write kvk = ( i=1 vi2 )1/2 for the usual Euclidean norm,
k×`
whereas for matrices M ∈ R
we distinguish between the spectral norm
kM kS = (λmax (M 0 M ))1/2 and the Frobenius norm kM kF = (trace(M 0 M ))1/2 .
We write PM for the matrix of orthogonal projection onto the column span of
M . If M satisfies rank M = ` ≤ k, then PM = M (M 0 M )−1 M 0 . We also make
use of the stochastic Landau notation. For a sequence of real random variables
zn , we say that zn = OP (1) if the sequence is bounded in probability, i.e., if
supn∈N P(|zn | > δ) → 0 as δ → ∞, and we say that zn = oP (1) if zn → 0 in
probability. For a non-stochastic real sequence an 6= 0 we write zn = OP (an ) if
zn /an = OP (1) and zn = oP (an ) if zn /an = oP (1).
The following is a list of technical assumptions needed in the proof of Theorem 2.1.
(A1) (a) The design vectors xi,n are linearly generated as follows:
xi,n = µn + Γn zi,n ,
where Γn is a pn × mn matrix with mn ≥ pn , such that Γn Γ0n = Σn .
The random mn -vectors z1,n , . . . , zn,n are i.i.d. and satisfy E[z1,n ] =
0
0, E[z1,n z1,n
] = Imn .
(b) The (n − 1) × (pn + 1) matrix Un,−1 = [0, In−1 ]Un has rank pn + 1
with probability one. (In particular, P(det(Un0 Un ) = 0) = 0.)
(c) For every n ∈ N, the random mn -vector z1,n from Assumption (A1).(a)
also has the following property. There exist universal positive constants c and C, not depending on n, such that for every orthogonal projection P in Rmn and for every t > C rank P , we have
P(kP z1,n k2 > t) ≤ Ct−1−c .
(d) Let z1,n be the random mn -vector from Assumption (A1).(a). For
Lk,n := supkvk=1 (E|v 0 z1,n |k )1/k we have L4,n = O(1) as n → ∞. For
0
every symmetric matrix Mn ∈ Rmn ×mn we have Var[z1,n
Mn z1,n ] =
2
2
O(trace Mn ) + (trace(Mn )) o(1), as n → ∞.
(e) In addition to (A1).(d), we also have L8,n = O(1) as n → ∞ and for
0
any projection matrix Pn in Rmn , (E[(z1,n
Pn z1,n )4 ])1/4 = O(kPn k2F )
as n → ∞.
(A2) The error terms εi,n can be written as εi,n = ei,n ε̃i,n , where ei,n is xi,n measurable and such that maxi=1,...,n ei,n = OP (1), and the ε̃i,n have
the following properties. There exists a universal constant κ > 0, not
depending on n, such that E[(E[|ε̃1,n /σn |4 |x1,n ])1+κ ] = O(1) as n → ∞,
√
and maxi E[(ε̃i,n /σn )4 |xi,n ] = oP ( qn ).
This set of assumptions is weaker than the analogous conditions imposed by
WC to treat the case (pn − qn )/n → 0. A detailed discussion and comparison of
the differences between our treatment and the one of WC is deferred to Section 4.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2593
3. Testing only a small number of hypotheses
In Theorem 2.1 above, we needed the assumption that qn → ∞ as n → ∞ in
order to achieve asymptotic normality of the F -statistic. Since the asymptotic
distribution of the F -statistic is well understood in the Gaussian error and fixed
design case (cf. Remark 2.4 and Lemma C.8), we expect the asymptotic distribution of the F -statistic to be χ2q rather than Gaussian if qn = q is fixed. The
following result establishes this asymptotic χ2 distribution for the F -statistic
when the error distribution of the model 2.1 is arbitrary up to bounded fourth
moments. For this result we need somewhat stronger assumptions than those of
Theorem 2.1. The proof is deferred to Section 6.
Theorem 3.1. Consider the linear, homoskedastic model (2.1) and assume that
the design vectors x1,n , . . . , xn,n are i.i.d. Gaussian with mean µn ∈ Rpn and
positive definite covariance matrix Σn , and that the design Xn = [x1 , . . . , xn ]0
is independent of the errors εn = (ε1,n , . . . , εn,n )0 which satisfy E[(ε1,n /σn )4 ] =
O(1). Moreover, suppose that lim supn pn /n < 1, ∆γ = o(qn /n) and that the
null hypothesis does not involve a restriction on the intercept parameter (i.e.,
the first column of R0 is equal to zero). If qn = q ∈ N does not depend on n,
then
p
√
w
(3.1)
s−1/2
(Fn − 1) − n∆γ bn −−−−→ (2q)−1/2 χ2q − q/2,
n
n→∞
where sn and bn are as in Theorem 2.1.
Together with the asymptotic normality of the F -statistic in the case qn → ∞
(cf. Theorem 2.1), we can establish a non-central F approximation for the F statistic for any number of tested hypotheses qn .
Corollary 3.2. Consider the linear, homoskedastic model (2.1) and assume that
the design vectors x1,n , . . . , xn,n are i.i.d. Gaussian with mean µn ∈ Rpn and
positive definite covariance matrix Σn , and that the design Xn = [x1 , . . . , xn ]0
is independent of the errors εn = (ε1,n , . . . , εn,n )0 which satisfy E[(ε1,n /σn )4 ] =
O(1). Moreover, suppose that lim supn pn /n < 1, ∆γ = o(qn /n) and that the
null hypothesis does not involve a restriction on the intercept parameter (i.e.,
the first column of R0 is equal to zero). Then,
sup |P(Fn ≤ t) − P(Fqn ,n−pn −1 (λn ) ≤ t)|
t∈R
−−−−→
n→∞
0,
where Fqn ,n−pn −1 (λn ) denotes a random variable following the non-central F
distribution with qn and n − pn − 1 degrees of freedom and non-centrality parameter λn = ∆γ (n − pn − 1 + qn ).
Proof. Suppose the claim does not hold. Then there exists a subsequence n0 ,
such that the supremum converges to a positive number along n0 . Then, by compactness of the closed unit interval and the extended real line, we can extract a
further subsequence n00 , such that pn00 /n00 → ρ1 ∈ [0, 1), qn00 /n00 → ρ2 ∈ [0, ρ1 ],
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2594
and qn00 → q ∈ [1, ∞], as n00 → ∞. If q = ∞, then we are in the setting of The√
−1/2
orem 2.1(iii) and we obtain asymptotic normality of sn (Fn − 1) − n∆γ bn .
Since the limiting distribution function is continuous, we get uniform convergence of the corresponding distribution functions, in view of Polya’s theorem.
√
−1/2
Since by Lemma C.8, sn (Fqn ,n−pn −1 (λn ) − 1) − n∆γ bn is also asymptotically standard normally distributed, we get a contradiction in that case. If
q < ∞, then qn00 = q for all large n00 , because qn00 is integer valued. Thus The√
−1/2
orem 3.1 applies
and shows that sn (Fn − 1) − n∆γ bn converges weakly to
p
(2q)−1/2 χ2q − q/2. Since the limiting distribution function is again continuous,
√
−1/2
and since Lemma C.8 shows that sn (Fqn ,n−pn −1 (λn ) − 1) − n∆γ bn has the
same asymptotic distribution, we also get a contradiction in this case, upon
using the same argument as before, involving Polya’s theorem.
Corollary 3.2 provides a unified treatment of the asymptotic behavior of
the F -statistic without distinguishing between the cases qn = q and qn →
∞. However, this does not immediately lead to a neat formula for the local
asymptotic power function of the F -test because of the complicated nature of
the cdf of the non-central F -distribution.
Remark 3.3 (Fixed vs. random design). It is instructive to compare the noncentral F approximation of Corollary 3.2 to the distribution of the F -statistic in
the Gaussian error and fixed design case (cf. Remark 2.4). In the former case the
non-centrality parameter is given by n∆γ (n − pn − 1 + qn )/n, while in the latter
case it is n∇n . Now if the design X is random with i.i.d. rows and S = E[U 0 U/n],
U = [ι, X], it turns out that ∆γ := (R0 γ − r0 )0 (R0 S −1 R00 )−1 (R0 γ − r0 )/σ 2 is
not a good approximation for ∇n := (R0 γ − r0 )0 (R0 (U 0 U/n)−1 R00 )−1 (R0 γ −
r0 )/σ 2 if pn /n is not close to qn /n, even if n is very large. In fact, we need a
correction factor of E[χ2n−pn −1+qn /n] = (n − pn − 1 + qn )/n in order to account
for the additional randomness coming from the design X (cf. Lemma C.5). This
correction factor, however, is close to one if qn /n ≈ pn /n. If the number qn
of hypotheses to be tested is much smaller than the number of parameters pn ,
this correction is quite significant. Thus, the limiting distribution of the F statistic under random design with E[U 0 U/n] = S is, in general, not equal to
the distribution of the F -statistic under Gaussian errors and fixed design X
satisfying U 0 U/n = S. This distinction only occurs for qn pn . In particular,
this issue disappears completely if pn /n ≈ 0.
Remark 3.4 (On confidence sets for R0 γ). Corollary 3.2 can immediately be
used to construct asymptotically valid confidence sets for R0 γ. Simply define
o
n
(1−ν)
CIν = r ∈ Rq : (R0 γ̂n − r)0 (R0 (U 0 U )−1 R00 )−1 (R0 γ̂n − r) ≤ fq,n−p−1 σ̂n2 q ,
and note that
(1−ν)
(1−ν)
P(R0 γ ∈
/ CIν ) = P(Fn > fq,n−p−1 ) → P(Fq,n−p−1 (0) > fq,n−p−1 ) = ν,
where Fn is the F -statistic under the null hypothesis r0 = R0 γ.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2595
4. Discussion of the technical assumptions
We pause for a moment to discuss the meaning of our Assumptions (A1) and
(A2), as well as the other conditions used in Theorem 2.1, and we comment on
the main differences to the conditions imposed in WC.
First of all, Assumption (A1).(a) of linear generation of the design from possibly much higher dimensional random vectors also appears in WC who take
it as a modification from Bai and Saranadasa [1]. We point out that this is a
straight forward generalization of the case m = p, where moment restrictions
have to be imposed directly on the design vectors xi (note that the components of x1 may not be independent, even after standardization, whereas x1 can
still be linearly generated from a vector z1 whose components are independent).
Moreover, this assumption also allows for the interpretation that there is actually a much higher dimensional set of explanatory variables zi available whose
dimensionality m (possibly m n) has already been reduced to p < n. See also
Remark 2.5.
Together with (A1).(a), our Conditions (A1).(d,e) replace and considerably
relax Assumption (C1) in WC, which reads as follows.
(C1) xi is linearly generated by a m-variate random vector zi = (zi1 , ..., zim )0 so
that xi = Γzi +µ, where Γ is a p×m matrix for some m ≥ p such that ΓΓ0 =
4
Σ, each zil has finite 8-th moment, E[zi ] = 0, Var[zi ] = Im , E[zik
] = 3+∆
Pd
`d
`d
`2
`1
`1 `2
]
·
·
· E[z1i
·
·
·
z
]E[z
],
and for any
z
]
=
E[z
`
≤
8,
E[z
1i2
1id
1i1
1i1 1i2
j=1 j
d
3
where ∆ is some finite constant.
In fact, in addition to (C1), WC also need the 8-th moments of zik to be uniformly bounded so that none of them goes off to infinity as n (and m = mn )
increases. The factorization requirement of the 8-th mixed moments in (C1)
is a straight forward relaxation of an independence assumption. However, just
like the independence assumption, it rules out many spherical distributions (cf.
Lemma A.1(i)). Therefore, moment conditions like (A1).(d,e) are much more
natural to accommodate both product and spherical distributions. In fact, Con8
dition (C1), together with uniform boundedness of E[zik
], is strictly stronger
than our Assumptions (A1).(a,d,e) (cf. Lemma A.1(iii) and Lemma A.2(ii)).
Our Assumption (A1).(b) is important to guarantee that the F -statistic is
equal to the expression on the right-hand-side of (2.4), at least with asymptotic
probability one, which is used in WC implicitly. The reason that we not only
require almost sure invertibility of U 0 U but also of the design matrix based on
n − 1 observations is only of a technical nature and plays an important role
in the proof of Lemma 6.3 (cf. the end of Subsection 6.2), which is based on
leave-one-out ideas. This lemma replaces the strong assumption of WC that
there exists a global constant c1 > 0, such that the smallest eigenvalue of the
sample covariance matrix satisfies λmin (X̃ 0 X̃/n) ≥ c1 , almost surely, where
3 Note that this formulation, as it stands, is self-contradictory. Clearly, one has to assume
2 ] = E[z 1 z 1 ] =
that the indices i1 , . . . , id are distinct, or otherwise (C1) implies 1 = E[z11
11 11
E[z11 ]E[z11 ] = 0.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2596
X̃ = (X − ιµ0 )Σ−1/2 is the design matrix based on the standardized regressors
(cf. page 147 in WC).4
Finally, Assumption (A1).(c) is taken directly from Srivastava and Vershynin
[23] to control the extreme eigenvalues of large sample covariance matrices, and
different sets of sufficient conditions for (A1).(c) can be found in that reference.
In WC, control of extreme eigenvalues is accomplished by use of the celebrated
Bai-Yin Theorem of Bai and Yin [2] (cf. Lemma 2 in WC). However, this comes
at the price of the implicit assumption that the standardized design vector
Σ−1/2 (x1 − µ) has independent components.5
Altogether, our design condition (A1) includes linear functions of both, product distributions with uniformly bounded 8-th marginal moments and a large
class of spherically symmetric distributions (cf. Lemma A.1 and Lemma A.2 in
Appendix A).
The Assumption (A2) on the error distribution extends the fourth moment
condition (C2) in WC, which simply states that E[(ε1,n /σn )4 ] = O(1), as n →
∞.6 If the errors are independent of the design and qn → ∞ (as is the case in
WC), then (C2) and (A2) are equivalent. However, Condition (A2) is suited to
also allow for some amount of dependence between the errors and the design.
This dependence is ruled out in WC, because they use results of asymptotic
normality of quadratic forms from Bhansali, Giraitis and Kokoszka [5] that
apply only in the independence case (see Lemma 6.1 and the discussion at the
beginning of Subsection 6.2). We note that an (8 + κ)-th moment condition like,
e.g., supn E[(ε̃1 /σn )8+κ ] ≤ K, together with maxi ei = OP (1), is sufficient for
(A2), provided that lim inf n qn /n > 0.
The additional requirement of Theorem 2.1, that lim supn pn /n < 1 and
qn → ∞, simply describes the regime of the relative number of parameters and
hypotheses we are interested in. The corresponding assumption (C3) in WC and
also parts (i) and (ii) of Theorem 2.1 additionally require that (pn − qn )/n → 0.
This is a more serious restriction which is convenient in the present strategy
of proof to show that the non-centrality term in the F -statistic under the local
alternative degenerates to the correct value (cf. Section 6.4.2). This, however,
means that asymptotically we are only dealing with hypotheses where almost
all of the p parameters are restricted, since qn /pn → 1 in this regime. It is
therefore important to extend the analysis of the rejection probability of the F test also to the regime where qn pn in order to asses the different contributions
of the overall dimensionality and the multiplicity of hypothesis testing to the
asymptotic rejection probability. This is what we do in Theorem 2.1(iii) and
4 Notice
that this assumption rules out, for example, Gaussian design [see, e.g., 11, Theorem
2.1].
5 This is particularly inconvenient if one is interested in the case where the random
vectors z1 , . . . , zn in (C1) (and (A1).(a)) have independent components. If both z1 and
Σ−1/2 (x1 − µ) = (ΓΓ0 )−1/2 Γz1 have independent components and at least two rows of
(ΓΓ0 )−1/2 Γ have only non-zero entries (this can be relaxed even further), then, by the DarmoisSkitovich Theorem [cf. 7, Theorem 5.3.1.], z1 must already be Gaussian.
6 In WC it is implicitly assumed that lim inf σ 2 > 0.
n n
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2597
in Section 3 in the Gaussian design setting. The requirement that qn → ∞ as
n → ∞ is essential in Theorem 2.1 in order to obtain a Gaussian limit. This
assumption is dropped in Theorem 3.1 and Corollary 3.2.
Finally, the assumption in Theorem 2.1 that ∆γ = o(qn /n), is rather natural
in the case where lim inf n qn /n > 0, where it simply reduces to ∆γ = o(1). In
this case, it says that we study the asymptotic rejection probability only in a
shrinking neighborhood of the null hypothesis. If lim inf n qn /n > 0, we also do
not need to specify a rate at which ∆γ approaches zero. Note, however, that
part (ii) of Theorem 2.1 actually does require a specific rate of contraction which
is, again, only needed for technical reasons in establishing the asymptotic behavior of the non-centrality term (cf. Section 6.4). The corresponding Assumption (C4) in WC is rather dubious and seems to arise from a miscalculation when
dealing with said non-centrality term. In fact, they also need the O(n−1/2 ) rate
of (R0 γ − r0 )0 R0 SR00 (R0 γ − r0 )/σn2 imposed by our Theorem 2.1(ii) and nothing
more.7
In the case where lim inf n qn /n = 0, we need the rate ∆γ = o(qn /n) in order
to ensure that the mixed term in the expansion of the F -statistic is asymptotically negligible (cf. the discussion following display (6.7)). Note that in the
extreme case where qn = q is fixed, the assumption ∆γ = o(qn /n) appears to be
somewhat restrictive because it rules out n∆γ 1, as required for non-trivial
local asymptotic power (cf. Remark 2.3). In the classical case where qn = q (and
pn = p) is fixed, the classical approach based on the asymptotic normality of
the OLS estimator γ̂n allows for a much wider range of local alternatives than
our present strategy. Of course, the asymptotic normality of the whole vector
γ̂n breaks down if the dimensions pn and qn are too large relative to sample
size n (see, e.g., Portnoy [20, 21]), which is why we here use a different strategy involving the assumption ∆γ = o(qn /n). Judging from the simulations of
Section 5 below, it seems as if some bound on ∆γ that is proportional to qn
is essential for the validity of the normal approximation to the power function,
because this approximation turns out to be accurate for a larger range of ∆γ if
q gets larger.
5. Numerical results
In order to better understand the quality of the theoretical approximations
to the power function of the F -test derived above, we conducted an extensive
simulation study. We roughly follow the simulation setup of WC but our focus
is more on the role of the number of tested hypotheses q. The linear model we
considered was
yi = β 0 xi + εi ,
7 See the first display on page 146 in WC, where also the matrix X X T needs to be
2 2
√
standardized. Also, there is a scaling factor of n missing in that argument, which is necessary
to bring the non-centrality term to the same scale as the noise term.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2598
where the εi were i.i.d. with mean zero and variance one, independent of the
design. We tried different error distributions but found little effect on the power
p
function so we report only the results for N (0, 1) errors and for t(5)/ 5/3
errors.8 The p-dimensional design vectors xi were generated as i.i.d. realizations
of a moving average process
xij =
T
X
αt zi(j+t−1) ,
t=1
where zi = (zi1 , . . . , zi(p+T −1) )0 was either a (p + T − 1)-dimensional standard
normal vector or generated with i.i.d. Γ(1, 1) − 1 entries9 and T = 10. The
coefficients αt were generated as i.i.d. uniform from (0, 1) only once and then
held fixed across all the simulations to follow.
We tested only null hypotheses of the form
H0 : β1 = · · · = βq = 0,
at level ν = 0.05, so that the distribution of the F -statistic does neither depend
on the mean of the design vectors nor on the intercept parameter, and hence it is
no loss of generality that we have omitted both. The signal β was generated such
that half of the tested coefficients where equal to one and all the other coefficients
were equal to zero. The signal was then scaled appropriately to produce a range
of alternatives at which the power function was evaluated numerically.
Since WC have already extensively studied the effect of the dimensionality
p on the power of the F -test, we here focus more on the effect of the number of tested restrictions q. For our first set of Montecarlo experiments we fixed
n = 100 and p = 60 and looked at the cases q = 4 and q = 50. Figure 1 shows the
simulated power of the F -test (solid lines) as a function of the scaled distance
from the null hypothesis ∆γ , where for each value of ∆γ , 10.000 Montecarlo
p
samples were generated. The left column shows the results for t(5)/ 5/3 distributed errors and a design that was generated from a moving average process
with Γ(1, 1)−1 innovations. The right column was generated with i.i.d. standard
Gaussian errors and a Gaussian moving average design. As a first observation
we note that the influence of non-Gaussianity on the simulated power (solid
lines) is almost negligible at the present sample size of n = 100. Now Figure 1
should be inspected from top to bottom. In the first row we clearly see the gain
of power as the number of hypotheses q decreases from q = 50 to q = 4. We also
compare the simulated power to the Gaussian approximation from our asymptotic result of Theorem 2.1 (dotted lines). The picture is qualitatively the same
as in WC, who considered q = p − 2 and who concluded “that there is a good
conformity between the empirical power and the theoretical power of the [. . . ]
8 Of course, one can always completely change the picture by, e.g., taking an error distribution that does not have a finite variance. In that case none of our theoretical approximations
is of much use.
9 The Γ(1, 1)-distribution is just the standard exponential distribution, but we keep the Γ
notation in the plots to facilitate comparability with WC.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
N(0, 1) errors, N(0, 1) design
0.8
power
0.6
q=4
q = 50
0.0
0.0
0.2
q=4
q = 50
Gauss
Montecarlo
0.4
0.6
0.4
Gauss
Montecarlo
0.2
power
0.8
1.0
5 3 errors, Γ(1, 1) − 1 design
1.0
t(5)
0.0
0.5
1.0
1.5
0.0
0.5
Δγ
1.5
N(0, 1) errors, N(0, 1) design
0.25
Gauss
Montecarlo
q=4
q = 50
0.05
0.05
0.10
q=4
q = 50
0.20
power
0.15
Gauss
Montecarlo
0.15
0.20
0.25
0.30
5 3 errors, Γ(1, 1) − 1 design
0.10
power
1.0
Δγ
0.30
t(5)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.00
0.05
0.10
Δγ
0.15
0.20
0.25
0.30
Δγ
N(0, 1) errors, N(0, 1) design
0.8
Gauss
Montecarlo
NCF
0.2
q=4
q = 50
0.0
0.0
0.2
q=4
q = 50
0.6
power
0.4
Gauss
Montecarlo
NCF
0.4
0.6
0.8
1.0
5 3 errors, Γ(1, 1) − 1 design
1.0
t(5)
power
2599
0.0
0.5
1.0
Δγ
1.5
0.0
0.5
1.0
1.5
Δγ
Fig 1. Simulated power functions of the F -test (Montecarlo) at level ν = 0.05, compared to
the Gaussian approximation of Corollary 2.2 (Gauss) and the non-central F approximation
of Corollary 3.2 (NCF).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2600
F -test [. . . ]” [27, p. 142]. It seems hard to evaluate the quality of approximation
directly from this picture in absolute terms. Moreover, the Gaussian approximation does not seem to become much more accurate when q increases, contrary
to what was suggested by Theorem 2.1. In fact, however, Theorem 2.1 says that
we should look at the Gaussian approximation only locally, for values of ∆γ
that are of a smaller order than q/n. Indeed, if we look at the power function in
a smaller neighborhood around the null (cf. the second row of Figure 1) we see
a much better agreement between the simulated true power and the Gaussian
approximation. We also see that when q/n is larger, the approximation is accurate on a larger interval around the null, as predicted by the theory. However,
a global Gaussian approximation to the power of the F -test seems to be too
much to ask for. Finally, the last row of Figure 1 is identical to the first row,
except that we have added the theoretical approximation based on the noncentral F -distribution as in Corollary 3.2 (NCF). Compared to the Gaussian
approximation, the non-central F approximation is remarkably accurate over
the entire range of alternatives considered, not just locally, and for both large
and small values of q.
To investigate also the quality of our approximations in small samples, we
have repeated the simulations above with n = 30 and p = 20. We present
only one instance of this second round of simulations in Figure 2 to discuss the
main differences to the case where n = 100. We still find that the non-central F
approximation is much better than the Gaussian approximation, but clearly also
the quality of the former deteriorates considerably compared to the case n = 100.
However, the local behavior of the power function is still picked up quite well
even in the small sample scenario. Moreover, it is remarkable how similar the
picture with t(5)-errors is to the picture with normal errors already for n = 30.
This suggests that the dominant reason for the imperfect approximation by
the non-central F -distribution is the randomness of the design rather than the
non-Gaussianity of the errors (cf. Remark 3.3).
Finally, as in WC, we also investigate the distribution of the F -statistic under
the null hypothesis H0 : β1 = · · · = βq = 0. For different choices of n, p and q, we
−1/2
have generated 10.000 Montecarlo samplespof sn (Fn − 1) as before, but with
vanishing true signal β = 0, and for t(5)/ 5/3-errors and regressors generated
from Γ(1, 1)−1. To investigate also the impact of a non-symmetric error distribution we have repeated all the simulations also with the distribution of the errors
and the design interchanged.10 The plots in Figure 3 were generated by applying the R-function ‘density’ [25] to the samples of standardized F -statistics with
default settings. As before, the dashed lines correspond to the (appropriately
scaled and centered) asymptotic non-central F -approximation of Corollary 3.2
and the black dotted line is the standard normal density. The overall picture is
the same as for the power function. The non-central F approximation is remarkably accurate even for moderate sample sizes like n = 50. As predicted by the
10 We do not report any results for Gaussian errors here because in that case the F -statistic
follows exactly a central Fq,n−p−1 -distribution under the null hypothesis, irrespective of the
design (cf. Remark 2.4).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
N(0, 1) errors, N(0, 1) design
0.8
0.2
0.6
0.4
Gauss
Montecarlo
NCF
q=4
q = 16
0.0
q=4
q = 16
power
0.2
0.6
0.4
Gauss
Montecarlo
NCF
0.0
power
0.8
1.0
5 3 errors, Γ(1, 1) − 1 design
1.0
t(5)
2601
0
1
2
3
4
Δγ
5
6
0
1
2
3
4
5
6
Δγ
Fig 2. Simulated power functions of the F -test (Montecarlo) at level ν = 0.05,Gaussian
approximation (Gauss) and the non-central F approximation (NCF) for sample size n = 30
and p = 20 regressor variables.
theory, for large n and large q the null distribution is well approximated by the
normal, whereas for small q the normal approximation fails. We also note that,
again, the approximation accuracy appears to be rather insensitive to changes
of the error and design distributions.
6. Proofs of main results
In this section we give a high-level description of the proofs of both Theorem 2.1
and Theorem 3.1. The more technical parts of the argument are collected in
the appendices. The following outline section pertains to both proofs and uses
only assumptions that are invoked by both theorems. Note that because of
compactness and, in particular, lim supn pn /n < 1, it is no restriction to assume
that pn /n and qn /n are convergent sequences with limits ρ1 ∈ [0, 1) and ρ2 ∈
[0, ρ1 ], respectively.
6.1. Outline
In addition to the general model assumptions of Section 2, the present outline
section
6.1 only uses the conditions ∆γ = o(qn /n), P(det(U 0 U ) = 0) = 0 and
√
2
n(σ̂n /σn2 − 1) = OP (1). All of these are satisfied under the assumptions of any
part of Theorem 2.1 as well as under the assumptions of Theorem 3.1, in view
of Lemma C.1.
The first part of Section 6.1 closely follows the classical approach for the
decomposition of the F -statistic as described, e.g., in Rao and Toutenburg [22,
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
Γ(1, 1) − 1 errors, t(5)
0.5
0.4
Density
q=4
q = 30
0.1
0.0
0.1
0.0
-2
-1
0
−1 2
sn
1
2
3
-3
-2
-1
0
−1 2
(Fn − 1)
sn
1
2
3
(Fn − 1)
Γ(1, 1) − 1 errors, t(5)
5 3 design
0.6
5 3 errors, Γ(1, 1) − 1 design
0.6
t(5)
0.4
p = 80
Density
q=4
q = 70
0.1
0.0
0.0
0.1
0.2
0.3
q=4
q = 70
0.3
0.5
0.4
p = 80
Montecarlo
NCF
n = 100
0.5
Montecarlo
NCF
n = 100
0.2
-3
-2
-1
0
−1 2
sn
2
3
-3
-2
-1
0
−1 2
sn
1
2
3
(Fn − 1)
Γ(1, 1) − 1 errors, t(5)
5 3 design
0.6
5 3 errors, Γ(1, 1) − 1 design
0.6
t(5)
1
(Fn − 1)
0.5
q=4
q = 100
q = 350
0.2
0.1
0.0
0.0
0.1
0.2
Density
q=4
q = 100
q = 350
0.3
p = 400
0.4
0.4
p = 400
Montecarlo
NCF
n = 500
0.5
Montecarlo
NCF
n = 500
0.3
Density
p = 40
0.2
0.3
q=4
q = 30
Montecarlo
NCF
n = 50
0.3
0.5
0.4
p = 40
0.2
Density
Montecarlo
NCF
n = 50
-3
Density
5 3 design
0.6
5 3 errors, Γ(1, 1) − 1 design
0.6
t(5)
2602
-3
-2
-1
0
−1 2
sn
(Fn − 1)
1
2
3
-3
-2
-1
0
−1 2
sn
1
2
3
(Fn − 1)
Fig 3. Kernel density estimates of simulated standardized F -statistics under the null hypothesis (solid lines), appropriately standardized non-central F -approximation of Corrolary 3.2
(dashed lines) and the standard normal density (black dotted lines) for different choices of n,
p and q, and different design and error distributions.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2603
Chapter 3.7]. These arguments are kept to a minimum but are included nonetheless to make the notation more intelligible.
Recall the F -statistc Fn as defined in (2.4). For the following preliminary
consideration, we work only on the event Cn = {ω : σ̂n2 (ω) > 0, det(U 0 U (ω)) 6=
0}, where σ̂n2 = kY − U γ̂n k2 /(n − p − 1). On this event, Fn is given by
Fn =
(R0 γ̂n − r0 )0 (R0 (U 0 U )−1 R00 )−1 (R0 γ̂n − r0 )/q σn2
.
σn2
σ̂n2
(6.1)
Setting δγ = (R0 γ − r0 )/σn , we have
(R0 γ̂n − r0 )/σn = (R0 (U 0 U )−1 U 0 Y − r0 )/σn = R0 (U 0 U )−1 U 0 (ε/σn ) + δγ ,
and thus, the first fraction in (6.1) reads
(ε/σn )0 U (U 0 U )−1 R00 (R0 (U 0 U )−1 R00 )−1 R0 (U 0 U )−1 U 0 (ε/σn )/q
+ 2(ε/σn )0 U (U 0 U )−1 R00 (R0 (U 0 U )−1 R00 )−1 δγ /q
+ δγ0 (R0 (U 0 U )−1 R00 )−1 δγ /q.
Next, if q < p + 1, choose a (p + 1 − q) × p matrix R1 , whose rows form an
orthonormal basis for the orthogonal complement of the rows of R0 . Recall that
R0 was chosen such that R0 R00 = Iq . Hence, T := [R00 , R10 ]0 is a (p + 1) × (p + 1)
orthogonal matrix. Partitioning U T 0 = [U0 , U1 ], where U0 = U R00 and U1 =
U R10 , and using block matrix inversion, we see that
R0 (U 0 U )−1 R00 = [Iq , 0]T (U 0 U )−1 T 0 [Iq , 0]0
= [Iq , 0](T U 0 U T 0 )−1 [Iq , 0]0
= (U00 (In − PU1 )U0 )−1 .
Similarly, we get
R0 (U 0 U )−1 U 0 = [Iq , 0]T (U 0 U )−1 T 0 T U 0
= (U00 (In − PU1 )U0 )−1 U00 (In − PU1 ).
Now, by writing W = (In − PU1 )U0 , on Cn , we can simplify the F -statistic to
read
(σ̂n2 /σn2 )Fn = (ε/σn )0 PW (ε/σn )/q + 2(ε/σn )0 W δγ /q + δγ0 W 0 W δγ /q.
The above representation remains correct also in the case where q = p + 1
provided that the matrix U1 is removed wherever it appears, i.e., W = U0 in
−1/2
this case. The correct centering and scaling of Fn is sn (Fn − 1), for sn =
2(1/q + 1/(n − p − 1)) (cf. Lemma C.8). After noting that PW = PU − PU1 and
abbreviating Mn = (PU − PU1 )/q − (In − PU )/(n − p − 1), we obtain
s−1/2
(Fn − 1) = s−1/2
(ε/σn )0 Mn (ε/σn )
n
n
σn2
σ̂n2
+ s−1/2
2(ε/σn )0 W δγ /q + δγ0 W 0 W δγ /q
n
(6.2)
σn2
,
σ̂n2
(6.3)
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2604
on the event Cn . Now, to get rid of the restriction to Cn , define Gn by
Gn = s−1/2
(ε/σn )0 Mn (ε/σn ) + 2s−1/2
(ε/σn )0 W δγ /q + s−1/2
δγ0 W 0 W δγ /q,
n
n
n
(6.4)
and note that this is well defined everywhere. It is now elementary to verify that
−1/2
we can study the asymptotic behavior of Gn instead of sn (Fn − 1). Simply
2
note that if Gn − ηn converges weakly to some√limiting distribution L, for an
appropriate centering sequence ηn2 with ηn2 = o( n), then, on Cn ,
w
s−1/2
(Fn − 1) − ηn2 = (Gn − ηn2 )(σn2 /σ̂n2 ) + ηn2 (σn2 /σ̂n2 − 1) −−−−→ L,
n
n→∞
(6.5)
√
and P(Cn ) → 1 as n → ∞, because P(det(U 0 U ) = 0) = 0, n(σ̂n2 /σn2 − 1) =
OP (1) and P(σ̂n2 = 0) ≤ P(|σ̂n2 /σn2 − 1| > 1/2) → 0, as claimed at the beginning
of this section.
In what follows, we will establish that the first term on the right of the equal
−1/2
sign in (6.4), which we denote by Qn := sn ε0 Mn ε/σn2 , satisfies Qn → L,
weakly, for an appropriate limit distribution L. The last summand in (6.4) can be
−1/2
abbreviated to sn n∇n /q, where ∇n = δγ0 W 0 W δγ /n = δγ0 (R0 (U 0 U/n)−1 R00 )−1 δγ
(as in Remark 2.4). It will play the role of a non-centrality term and it will
be shown to be asymptotically non-random. Note that if we can also show
√
−1/2
−1/2
sn n∇n /q = oP ( q), then the mixed term in (6.4) satisfies sn (ε/σn )0 W δγ /q =
oP (1). Indeed, the conditional mean of the latter expression given X is equal to
√ √
−1/2
2
n∇n /q)/( q sn q) =
zero, and its conditional variance is equal to s−1
n n∇n /q = (sn
√
−1/2
oP (1), provided that sn n∇n /q = oP ( q).
Suppose, for now, that we have already established both, the weak convergence
w
Qn −−−−→ L,
(6.6)
n→∞
and also the fact that
s−1/2
n∇n /q −
n
√
n∆γ bn
i.p.
−−−−→
n→∞
0,
(6.7)
where bn is as in Theorem
2.1. Then, because ∆γ = o(q/n), we have ηn2 :=
√
√
√
n∆γ bn = o( q) = o( n), as required for the argument in (6.5). It also follows
√
√
−1/2
that sn n∇n /q = o( q) + oP (1) = oP ( q), so we have asymptotic negligibility of the mixed term in (6.4) by the argument in the previous paragraph.
Altogether, we arrive at
w
Gn − ηn2 = Qn + oP (1) −−−−→ L,
n→∞
which establishes the conclusion of Theorem 2.1 and Theorem 3.1, for an appropriate choice of L, provided that (6.6) and (6.7) hold.
For Theorem 2.1, we will prove the weak convergence (6.6) with L = N (0, 1),
under the general Assumptions (A1).(a,b,c,d) and (A2) in Section 6.2, and the
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2605
convergence in (6.7) under each of the sets of assumptions of Theorem 2.1(i),
2.1(ii) and 2.1(iii), respectively (cf. Section 6.4).
For Theorem 3.1, we will prove the convergence (6.6) with L = (2q)−1/2 χ2q −
p
q/2 in Section 6.3, and the convergence in (6.7) is established by the same
argument as in the case of Theorem 2.1(iii) in Section 6.4, which does not require
qn → ∞ nor Assumption (A2), so that it goes through also in the setting of
Theorem 3.1.
6.2. Asymptotic normality of the noise term
This section establishes the weak convergence in (6.6) with L = N (0, 1). For this
claim we only use the Assumptions (A1).(a,b,c,d), (A2), as well as pn /n → ρ1 ∈
[0, 1), qn /n → ρ2 ∈ [0, ρ1 ] and qn → ∞. The following lemma is a variation of
Theorem 2.1 in Bhansali, Giraitis and Kokoszka [5] on the asymptotic normality
of quadratic forms for the case where the matrix and the enclosing vectors
may exhibit a certain dependence between each other. Its proof is deferred to
Appendix B.
Lemma 6.1. Let (Ω, F, P) be the common probability space on which all the
random quantities below are defined. For every n ∈ N, let Gn ⊆ F be a subsigma algebra, let An = (aij,n )ni,j=1 be a real random symmetric n × n matrix
that is Gn measurable and such that An (ω) 6= 0, ∀ω ∈ Ω. Let Z1,n , . . . , Zn,n be
real random variables that are conditionally independent, given Gn , and such that
2
for i ≤ n, almost surely, E[Zi,n |Gn ] = 0, E[Zi,n
|Gn ] = 1 and E[|Zi,n |4 |Gn ] < ∞.
Moreover, assume that, as n → ∞,
kAn k2S
4
max E[Zj,n
|Gn ]
kAn k2F j=1,...,n
2
maxj (A2n )jj
4
max
E[Z
|G
]
j,n n
j=1,...,n
kAn k2F
n
4
X
a2jj,n E[Zj,n
|Gn ]
and
2
kA
k
n F
j=1
i.p.
−−−−→
n→∞
i.p.
−−−−→
n→∞
i.p.
−−−−→
n→∞
0,
0,
0.
Then, for Zn = (Z1,n , . . . , Zn,n )0 , we have
Zn0 An Zn − E[Zn0 An Zn |Gn ] w
√
−−−−→ N (0, 1).
n→∞
2kAn kF
Remark 6.2. The proof of Lemma 6.1 essentially follows the rationale of
Bhansali, Giraitis and Kokoszka [5] with the obvious modification that all the
moments of Zi,n have to be replaced by conditional moments. Note that if the
Z1,n , . . . , Zn,n are the first n elements of a sequence of i.i.d. random variables
and An is non-random, as in Bhansali, Giraitis and Kokoszka [5], then the assumptions of Lemma 6.1 reduce to those imposed by Theorem 2.1(iii) in that
reference, except for the additional requirement that E[Z14 ] < ∞, as needed here.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2606
By the method of Bhansali, Giraitis and Kokoszka [5] we can not get rid of this
additional requirement because their truncation argument does not apply in the
case of dependence between An and Zn .
With Lemma 6.1 at hand, we can proof the asymptotic normality of
Qn
=
s−1/2
(ε/σn )0 Mn (ε/σn ).
n
Under the linear model (2.1), the components ε1 , . . . , εn of ε are conditionally
independent given the design X, with E[εi /σn |X] = 0, E[(εi /σn )2 |X] = 1 and
E[(εi /σn )4 |X] < ∞, almost surely, in view of Assumption (A2). Moreover, the
random matrix Mn = (PU − PU1 )/q − (In − PU )/(n − p − 1) is σ(X)-measurable
and satisfies trace(Mn ) = 0 and
kMn k2F = trace Mn2 = trace [(PU − PU1 )/q 2 + (In − PU )/(n − p − 1)2 ]
= 1/q + 1/(n − p − 1),
with probability one, in view of (A1).(b). Also, kMn k2F ≥ 1/(n−p−1) ≥ 1/n, everywhere, because trace(PU ) ≤ p+1. With this and in view of E[ε0 Mn ε|X]/σn2 =
trace(Mn ) = 0, almost surely, we see that
Qn = s−1/2
n
(ε/σn )0 Mn (ε/σn ) − E[(ε/σn )0 Mn (ε/σn )|X]
ε 0 Mn ε
√
=
,
σn2
2kMn kF
at least on a set of probability one, and it remains to verify the convergence
conditions of Lemma 6.1. For the first one, note that kMn k2S ≤ (1/q + 1/(n −
p − 1))2 and hence, almost surely,
kMn k2S
q
maxj=1,...,n E[(εj /σn )4 |X]
4
max
E[(ε
/σ
)
|X]
≤
1
+
.
j
n
2
kMn kF j=1,...,n
n−p−1
q
But clearly maxj E[(εj /σn )4 |X]/q ≤ OP (1) maxj E[(ε̃j /σn )4 |xj ]/q = oP (q −1/2 )
under Assumption (A2). Therefore, the upper bound in the previous display
converges to zero in probability. For the second condition, since the diagonal
entries of a projection matrix are between 0 and 1, we see that (Mn2 )jj ≤ 1/q 2 +
1/(n − p − 1)2 , and thus
2
maxj (Mn2 )jj
4
max E[(εj /σn ) |X]
j=1,...,n
kMn k2F
≤
1 + q 2 /(n − p − 1)2
(maxj E[(ε̃j /σn )4 |xj ])2
(max ei )8
,
i
1 + q/(n − p − 1)
q
which converges to zero in probability under Assumption (A2) and qn → ∞.
Establishing the validity of the last condition is slightly more involved. Since
kMn k2F is of order 1/q, we have to show that
q
n
X
j=1
m2jj E[(εj /σn )4 |X] ≤ (max ei )4 q
i
n
X
j=1
m2jj E[(ε̃j /σn )4 |X] = oP (1),
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2607
where Mn = (mij )ni,j=1 . By Assumption (A2), (maxi ei )4 = OP (1). Now, take
expectation and use Hölder’s inequality with a, b > 1 such that 1/a + 1/b = 1
to obtain
n
X
E q
m2jj E[(ε̃j /σn )4 |X]
j=1
n
1/b X
1/a
≤ E[(E[(ε̃1 /σn )4 |x1 ])b ]
q
(E[m2a
.
jj ])
(6.8)
j=1
Now choose b = 1+κ and invoke Assumption (A2) to show that the b-th moment
of the conditional expectation in (6.8) is O(1). To establish that
q
n
X
j=1
1/a
(E[m2a
jj ])
i.p.
−−−−→
n→∞
0,
(6.9)
we distinguish between the cases ρ2 > 0 and ρ2 = 0.
If ρ2 > 0, we write the diagonal elements of Mn = (PU − PU1 )/q − (In −
PU )/(n − p − 1) as
1 − (PU )jj
(PU )jj − (PU1 )jj
−
q
n−p−1
p+1
p+1−q
(PU )jj − p+1
(PU )jj − n + n − (PU1 )jj
n
=
+
,
q
n−p−1
mjj =
and note that (PU )jj − (p + 1)/n ∈ [−1, 1] and (PU1 )jj − (p + 1 − q)/n ∈ [−1, 1],
in order to get the bound
1/a
E[m2a
jj ]
p+1
n
n
1
(PU )jj −
+
+
= 2 E
n
n
q
n−p−1
2a #!1/a
p+1−q
n
− (PU1 )jj
n
q
(
2a "
2a #
1
n
n
p+1
2a−1
≤ 2 2
+
E (PU )jj −
+
n
q
n−p−1
n
2a "
2a #)!1/a
n
p+1−q
E (PU1 )jj −
,
q
n
where we have used the inequality (c + d)2a ≤ 22a−1 (c2a + d2a ) for c, d ∈ R.
Hence, if we can show that the (PU )jj , for j = 1, . . . , n, and also the (PU1 )jj ,
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2608
for j = 1, . . . , n, are identically distributed, then
n
n
X
1/a
(E[m2a
jj ])
≤
2(2a−1)/a O(1)×
j=1
1/a
× E |(PU )11 − (p + 1)/n|2a + E |(PU1 )11 − (p + 1 − q)/n|2a
,
because n/q → 1/ρ2 . Since a = (1 + κ)/κ is fixed, it then remains to show that
|(PU )11 − (p + 1)/n| → 0 and |(PU1 )11 − (p + 1 − q)/n| → 0, in probability. The
desired properties of the diagonal entries of PU and PU1 are now established by
the following lemma, which applies under the Assumptions (A1).(a,b,d), and
whose proof is deferred to Appendix B.
Lemma 6.3. For every n ∈ N, let x1,n , . . . , xn,n be i.i.d. random pn -vectors
that satisfy Assumptions (A1).(a,b) with µn ∈ Rpn and positive definite covariance matrix Σn . Moreover, suppose that the random vector
z1,n from Assump0
tion (A1).(a) also satisfies Var[z1,n
Mn z1,n ] = O(trace Mn2 )+(trace(Mn ))2 o(1),
as n → ∞, for every symmetric mn × mn matrix Mn . Let Rn be a nonrandom (pn + 1) × kn matrix such that rank Rn = kn ≤ pn + 1 and define
Xn = [x1,n , . . . , xn,n ]0 and Wn = [ι, Xn ]Rn , where ι = (1, . . . , 1)0 ∈ Rn . Furthermore, let h1,n , . . . , hn,n denote the diagonal entries of the projection matrix
PWn . Then, the (hj,n )nj=1 are exchangeable random variables and
|h1,n − kn /n|
−−−−→
n→∞
0,
in probability.
Remark 6.4. Note that for Rn = Ipn +1 , the h1 , . . . , hn in Lemma 6.3 are the
leverage values of the regression with design matrix U = [ι, X].
Altogether, we see that in the case ρ2 > 0, (6.9) holds and the weak convergence
ε0 Mn ε w
−−−−→ N (0, 1),
Qn = s−1/2
n
n→∞
σn2
follows, as required in (6.6).
To treat the case ρ2 = 0, we recall from Section 6.1 that PU − PU1 =
PW = U (U 0 U )−1 R00 (R0 (U 0 U )−1 R00 )−1 R0 (U 0 U )−1 U 0 , almost surely, by Assumption (A1).(b), and write the diagonal elements of Mn = PW /q − (In − PU )/(n −
p − 1) as before as
mjj =
(PW )jj −
q
q
n
+
(PU )jj − p+1
n
.
n−p−1
Note that the (PW )jj = (1, x0j )(U 0 U )−1 R00 (R0 (U 0 U )−1 R00 )−1 R0 (U 0 U )−1 (1, x0j )0 ,
for j = 1, . . P
. , n, are exchangeable random variables, because x1 , . . . , xn are i.i.d.
n
and U 0 U = j=1 (1, x0j )0 (1, x0j ) is a function in x1 , . . . , xn that is invariant under
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2609
permutations of its arguments. Therefore,
q
n
X
1/a
(E[m2a
≤ q
jj ])
j=1
n
X
"
E
j=1
"r
=
E
|(PW )jj − nq |
1
+
q
n−p−1
2a #!1/a
2a #!1/a
√
nq
q
n
(PW )11 −
+
.
q
n
n−p−1
By boundedness ofpthe diagonal entries of a projection matrix, it remains to
show that Hn := n/q((PW )11 − q/n) → 0, in probability, as n → ∞. By
exchangeability, P
and Assumption (A1).(b), it follows that q = trace(PW ) =
n
E[trace(PW )] = j=1 E[(PW )jj ] = nE[(PW )11 ], almost surely, and thus E[Hn ] =
p
p
0. Moreover, Hn ≥ − q/n, and for ε > 0, 0 = E[Hn ] ≥ E[Hn 1{Hn >ε} ] − q/n,
p
which implies that for n large (such thatp q/n < ε), εP(|Hn | > ε) ≤ εP(Hn >
ε)+εP(Hn < −ε) ≤ E[Hn 1{Hn >ε} ]+0 ≤ q/n. This establishes the asymptotic
normality required in (6.6) also in the case ρ2 = 0.
6.3. Asymptotic χ2 -distribution of the noise term
In this section we show p
that under the assumptions of Theorem 3.1, (6.6) holds
with L = (2q)−1/2 χ2q − q/2. First note that
0
2
ε0 Mn ε
−1/2 ε PW ε
−1/2 σ̂n
=
s
−
s
.
n
n
σn2
σn2 q
σn2
p
−1/2
By Lemma C.1, we have sn σ̂n2 /σn2 → q/2, in probability. Since we do not
test the intercept parameter α, we can write R0 = [0, T0 ], for some q × p matrix
T0 , and
1 0···0
0
R1 = .
T1
..
Qn = s−1/2
n
0
for some (p − q) × p matrix T1 , such that T̄ := [T00 , T10 ]0 is p × p orthogonal.
If q = p, then T0 = Ip and R1 = (1, 0, . . . , 0). With this notation, we get
U = [ι, X], U0 = U R00 = XT00 and U1 = U R10 = [ι, XT10 ]. From this we see
that the distribution of W = (In − PU1 )U0 = (In − P[ι,(In −Pι )XT10 ] )XT00 does
not depend on µ, and without loss of generality we may assume that µ = 0.
Moreover, by standard properties of orthogonal projections,
PW = PU − PU1 = PX + P(In −PX )ι − PXT10 + P(In −PXT 0 )ι
1
= PX T̄ 0 − PXT10 + P(In −PX )ι − P(In −PXT 0 )ι
1
= P(In −PXT 0 )XT00 + P(In −PX )ι − P(In −PXT 0 )ι .
1
1
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2610
Now, abbreviate A = (In − PXT10 )XT00 and B = P(In −PX )ι − P(In −PXT 0 )ι , and
1
note that PA is uniformly distributed on the Grassmann manifold of n × n
projection matrices of rank q, because for an n × n orthogonal matrix O we
have OX = (OXΣ−1/2 )Σ1/2 ∼ X, OPA O0 = OA(A0 O0 OA)−1 A0 O0 , and
OA = (In − OXT10 (T1 X 0 O0 OXT10 )−1 T1 X 0 O0 )OXT00
∼ (In − XT10 (T1 X 0 XT10 )−1 T1 X 0 )XT00 = A.
Moreover, trace(B) = 0, almost surely, and thus E[ε0 Bε/σn2 |X] = 0 and by
standard calculations using independence (cf. the proof of Lemma C.1) and the
fact that PX = PX T̄ 0 = P[XT00 ,XT10 ] ,
n
X
2
Var[ε0 Bε/σn2 |X] = 2trace B 2 +
(E[(εi /σn )4 ] − 3)Bii
i=1
≤ (2 + E[(ε1 /σn )4 ])trace B 2
ι0 (In − PX )(In − PXT10 )ιι0 (In − PXT10 )(In − PX )ι
= O(1) 2 − 2
ι0 (In − PX )ιι0 (In − PXT10 )ι
ι0 (In − PX T̄ 0 )ι
= O(1) 1 − 0
.
ι (In − PXT10 )ι
In view of Lemma C.7 with µ = 0 and using the first two moments of the χ2
distribution, we see that the expressions ι0 (In − PX T̄ 0 )ι/n and ι0 (In − PXT10 )ι/n
both converge to 1 − ρ1 , in probability, and thus the whole expression on the
last line of the previous display converges to zero in probability. Altogether, we
see that
ε 0 PA ε p
Qn = (sn q 2 )−1/2 2 − q/2 + oP (1).
σn
Because PA is uniformly distributed on the Grassmann manifold, it can be
stochastically represented as PA ∼ CC 0 , where C is a random n × q matrix that
is uniformly distributed on the Stiefel manifold of order n × q, i.e., V has orthonormal columns and its distribution is both left and right invariant under the
action of the appropriate orthogonal group. Since C and ε are independent, the
so called Diaconis-Freedman effect as described in Dümbgen and Conte-Zerial
[9, Theorem 2.1] entails that the conditional distribution of C 0 ε/σn given C,
converges weakly in probability to a q-dimensional standard normal distribution because kε/σn k2 /n → 1 in probability and ε0 ε̄/(nσn2 ) → 0 in probability,
where ε̄ is an independent copy of ε. Weak convergence in probability implies
convergence of the conditional characteristic functions of C 0 ε/σn given C, in
probability, which, by boundedness, implies convergence of the unconditional
characteristic functions. Consequently, we obtain the weak convergence
p
p
w
Qn ∼ (sn q 2 )−1/2 kC 0 ε/σn k2 − q/2 + oP (1) −−−−→ (2q)−1/2 χ2q − q/2.
n→∞
This establishes (6.6) with L as claimed.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2611
6.4. Asymptotic behavior of the non-centrality term
Finally, we have to establish the convergence in (6.7) in the three cases of
Theorem 2.1(i), 2.1(ii) and 2.1(iii) as well as under the assumptions of The−1/2
orem 3.1. We begin by a representation of sn n∇n /q that pertains to all of
these cases. In this preliminary consideration we only use Assumpitons (A1).(a)
and (b) which are assumed to hold in each of the cases under investigation.
Recall the conventions and definitions of Section 6.1, in particular, U = [ι, X],
T = [R00 , R10 ]0 , U0 = U R00 , U1 = U R10 , W = (In − PU1 )U0 , δγ = (R0 γ − r0 )/σn ,
∆γ = δγ0 (R0 S −1 R00 )−1 δγ , and ∇n = δγ0 W 0 W δγ /n = δγ0 (R0 (U 0 U/n)−1 R00 )−1 δγ .
Pn
Write µ̂n = i=1 xi /n = X 0 ι/n and Σ̂n = X 0 X/n − µ̂n µ̂0n = X 0 (In − Pι )X/n
and partition the (p + 1) × (p + 1) orthogonal matrix T as
R0
t T0
T =
= 0
,
R1
t1 T1
where t0 ∈ Rq and T0 ∈ Rq×p . Since
0
−1
(U U/n)
=
1
µ̂n
µ̂0n
Σ̂n + µ̂n µ̂0n
−1
=
1 + µ̂0n Σ̂−1
n µ̂n
−1
−Σ̂n µ̂n
almost surely, by (A1).(b), we have
q
√ 0 0
n
−1/2
nδγ (U0 U0 /n)δγ ,
q (sn q)
q
√
n
−1/2
nδγ0 U00 (In − PU1 )U0 δγ /n,
s−1/2
n∇n /q =
n
q (sn q)
q
√
0 −1
n (sn q)−1/2 nδγ0 (T0 Σ̂−1
δγ ,
n T0 )
q
−µ̂0n Σ̂−1
n
,
Σ̂−1
n
if q = p + 1,
if q ≤ p,
if t0 = 0.
(6.10)
Notice that the last two cases are not mutually exclusive, but the case t0 = 0 is
a sub-case of the case q ≤ p. The representation of ∇n in the case t0 = 0 will
come in handy. With the notation
Ω00 Ω01
R0 SR00 R0 SR10
Ω=
=
= T ST 0 , where
Ω10 Ω11
R1 SR00 R1 SR10
1
µ0
1
0
1 x1 = E[U 0 U/n],
S=
=E
µ Σ + µµ0
x1
under (A1).(a), and by the simple block matrix inversion argument R0 S −1 R00 =
−1
[Iq , 0](T ST 0 )−1 [Iq , 0]0 = (Ω00 − Ω01 Ω−1
, involving the orthogonality of
11 Ω10 )
T , we analogously get
0
if q = p + 1,
δγ Ω00 δγ ,
(6.11)
∆γ = δγ0 (Ω00 − Ω01 Ω−1
Ω
)δ
,
if q ≤ p,
10
γ
11
0
−1 0 −1
δγ (T0 Σ T0 ) δγ ,
if t0 = 0.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2612
6.4.1. The case of Theorem 2.1(i)
Under the assumptions
which can
p of Theorem 2.1(i), choose bn as in the theorem,p
be written as bn = (1 − (p + 1)/n)(1 − (p + 1)/n + q/n)/(2q/n) = n/q(sn q)−1/2 (n−
p − 1 + q)/n. To establish
the convergence in (6.7) in the case of q = p + 1, first
p
note that now bn = n/q(sn q)−1/2 , and consider
√
√
bn nδγ0 (U00 U0 /n − Ω00 ) δγ ,
(6.12)
√
which
has mean zero. For the variance, we observe that Var[ nδγ0 (U00 U0 /n)δγ ] =
P
n
n−1 i=1 Var[δγ0 R0 (1, x0i )0 (1, x0i )R00 δγ ] ≤ E[|δγ0 R0 (1, x01 )0 |4 ] = O(|δγ0 Ω00 δγ |2 ) =
p
O(∆2γ ), in view of Lemma C.3(i) and Assumption (A1).(d), and bn ∆γ = O(1) n/q∆γ =
o(1), by assumption. This clearly covers also the case R0 = Ip+1 . If R0 = [0, Ip ],
then t0 = 0, T0 = Ip , q = p and the difference in (6.7) reads
r
√ 0
√
n
n−1
−1/2
(s
q)
δγ .
s−1/2
n∇
/q
−
n∆
b
=
nδ
Σ̂
−
Σ
n
n
γ
n
n
n
γ
q
n
s−1/2
n∇n /q −
n
n∆γ bn
=
The √
mean of this expression is, again, equal to zero. For its variance we find that
Var[ nδγ0 Σ̂n δγ ] = O(|δγ0 Σδγ |2 ), in view of Lemma C.3(ii) together with Assumpp
p
tion (A1).(d), and n/q(sn q)−1/2 δγ0 Σδγ = O(1) n/q∆γ → 0, by assumption.
This finishes the proof of Theorem 2.1(i).
6.4.2. The case of Theorem 2.1(ii)
For part (ii) we only need to consider the case where q ≤ p, as the case q = p + 1
has already been treated above (simply restrict to the subsequence
n0 such that
p
qn0 ≤ pn0 ). We establish the convergence in (6.7) for bn = n/q(sn q)−1/2 rather
than bn as in the Theorem. It should
p be clear, however, that this is no restriction.
[Indeed, if (6.7) holds with bn = n/q(sn q)−1/2 and b̃n = bn (n − p − 1 + q)/n
√
√
√
is as in√the theorem, then n∆γ bn − n∆γ b̃n = (bn − b̃n ) n∆γ = ((p + 1)/n −
√
q/n)bn nO( q/n), by the additional assumption of Theorem 2.1(ii). Since in
the present case (p − q)/n → 0, the previous expression converges to zero as
n → ∞.] Now, since q ≤ p, the quantity of interest reads
√
√
s−1/2
n∇n /q − n∆γ bn = bn nδγ0 (U00 U0 /n − Ω00 ) δγ
n
−1 0 !
√ 0
U00 U1 U10 U1
U1 U0
−1
δγ .
(6.13)
+ bn nδγ Ω01 Ω11 Ω10 −
n
n
n
The first term on the right-hand-side has already been studied in (6.12), and
the same argument applies, except that now δγ0 Ω00 δγ 6= ∆γ , in general. But
p
bn δγ0 Ω00 δγ = O(1) n/qδγ0 R0 SR00 δγ = O(n−1/2 ) → 0 by the additional assumption of Theorem 2.1(ii). For the remaining term in (6.13), as in WC, we begin
by approximating U10 U1 /n by Ω11 . This can only be successful because here we
are dealing with a sample covariance matrix of dimension p + 1 − q, based on n
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2613
independent observations and we assume that (p + 1 − q)/n → 0. We abbreviate
−1/2
−1/2
Ũ0 = U0 Ω00 and Ũ1 = U1 Ω11 and consider the absolute difference
"
#
√ 0 U00 U1 U10 U1 −1 U10 U0
U00 U1 −1 U10 U0
nδγ
−
Ω11
δγ
n
n
n
n
n
!−1
0
√ 0 1/2 Ũ00 Ũ1
Ũ 0 Ũ0 1/2
Ũ
Ũ
1 1
− Ip+1−q 1 Ω00 δγ
= nδγ Ω00
n
n
n
≤
Ũ10 Ũ1
n
!−1
Ũ1 Ũ10
n
− Ip+1−q
S
S
Ũ00 Ũ0
n
√
nδγ0 Ω00 δγ .
(6.14)
S
Now, Lemma C.3(iii) with kn = p + 1 − q shows that kŨ10 Ũ1 /n − Ip+1−q kS →
0 in probability, since kn /n → 0, and it also establishes the boundedness
in probability of kŨ1 Ũ10 /nkS kŨ00 Ũ0 /nkS . The assumptions of this lemma are
clearly satisfied under (A1).(a,c,d). Now the convergence in spectral norm implies the convergence of the extreme eigenvalues of Ũ10 Ũ1 /n to 1, and thus,
also the extreme eigenvalues of the inverse converge to
√ 1, which means that
k(Ũ10 Ũ1 /n)−1 − Ip+1−q kS → 0, in probability. Since bn nδγ0 Ω00 δγ = O(1), by
the additional assumption of Theorem 2.1(ii), the upper bound in (6.14) converges to zero in probability. Thus, we have shown that we can replace U10 U1 /n
in (6.13) by Ω11 , without changing the limit.
To finish part (ii) it remains to show that bn Bn converges to zero in probability, where
√ 0 U00 U1 −1 U10 U0
−1
Bn := nδγ
Ω11
− Ω01 Ω11 Ω10 δγ .
n
n
To evaluate its expectation, write
Bn =
n
1 X√ 0
−1
0 0
0
0
nδγ R0 (1, x0i )0 (1, x0i )R10 Ω−1
11 R1 (1, xj ) (1, xj )R0 − Ω01 Ω11 Ω10 δγ .
n2 i,j=1
Since E[R0 (1, x0i )0 (1, x0i )R10 ] = R0 SR10 = Ω01 , all summands in E[Bn ] with distinct indices i 6= j disappear. Using parts (i) and (iv) of Lemma C.3, which
apply in view of Assumptions (A1).(a,d), we arrive at
1√ 0
0 0
0
0
|E[Bn ]| =
nδγ E[R0 (1, x01 )0 (1, x01 )R10 Ω−1
11 R1 (1, x1 ) (1, x1 )R0 ]δγ
n
1√ 0
−
nδγ Ω01 Ω−1
11 Ω10 δγ
n
q
≤ n−1/2
0 0 2
E[|δγ0 R0 (1, x01 )0 |4 ]E[|(1, x01 )R10 Ω−1
11 R1 (1, x1 ) | ]
+ n−1/2 δγ0 Ω01 Ω−1
11 Ω10 δγ
q
= n−1/2 O(|δγ0 R0 SR00 δγ |2 )O((p + 1 − q)2 ) + n−1/2 δγ0 Ω01 Ω−1
11 Ω10 δγ
√ 0
= O( nδγ Ω00 δγ (p + 1 − q)/n) + n−1/2 δγ0 Ω01 Ω−1
11 Ω10 δγ .
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2614
After
√ multiplying by bn , the upper bound still converges to zero, because, first,
bn nδγ0 Ω00 δγ = O(1) and (p + 1 − q)/n → 0, by the additional assumptions
−1/2 0
of Theorem 2.1(ii), and because bn n−1/2 δγ0 Ω01 Ω−1
δγ Ω00 δγ =
11 Ω10 δγ ≤ bn n
−1
O(n ), where the inequality holds in view of the second case in (6.11).
In order to show that the distribution of Bn also concentrates around its
mean, we make use of the Efron-Stein inequality.11 We use the abbreviations
0
0 0
0
0 −1
0 0
D = δγ0 Ω01 Ω−1
11 Ω10 δγ , Li = δγ R0 (1, xi ) , Qij = (1, xi )R1 Ω11 R1 (1, xj ) and dep×n
p×(n−1)
fine the functions g : R
→ R, for k = 1, . . . , n, by
Pn → R and gk√: R
g(x1 , . . . , xn ) = n−3/2 i,j=1 Li Qij Lj − nD = Bn and
n
X
gk (x1 , . . . , xk−1 , xk+1 , . . . , xn ) = n−3/2
Li Qij Lj −
√
nD.
i,j=1
i6=k,j6=k
By the Efron-Stein inequality [18, Theorem 9],
Var[Bn ] ≤
n
X
E[(g(x1 , . . . , xn ) − gk (x1 , . . . , xk−1 , xk+1 , . . . , xn ))2 ].
(6.15)
k=1
Now, for k ∈ {1, . . . , n}, g(x1 , . . . , xn ) can be expressed as
n−3/2
n
X
i,j=1
i6=k,j6=k
Li Qij Lj +
n
X
Li Qik Lk +
i=1
i6=k
n
X
i=1
i6=k
√
Lk Qki Li + Lk Qkk Lk
− nD.
Using the fact that Qij = Q0ij = Qji , the differences g − gk in (6.15), are equal
to
n
X
Li Qik Lk + Lk Qkk Lk .
(6.16)
n−3/2 2
i=1
i6=k
We need to bound the expectation of the squared expression. To this end, we
calculate the expectation of Li Qik Lk Lj Qjk Lk for arbitrary indices i, j, k, as
well as in the special case where i 6= k, j 6= k and i 6= j. Observe that
−1/2
−1/2
Qik is the inner product of Ω11 R1 (1, x0i )0 and Ω11 R1 (1, x0k )0 and therefore, by Cauchy-Schwarz inequalities in both Euclidean and Lp space, satisfies
−1/2
−1/2
−1/2
E[|Qik |4 ] ≤ E[kΩ11 R1 (1, x0i )0 k4 kΩ11 R1 (1, x0k )0 k4 ] ≤ E[kΩ11 R1 (1, x01 )0 k8 ] =
4
E[Q11 ]. Moreover, parts (i) and (v) of Lemma C.3, whose assumptions are implied by the conditions (A1).(a,d,e), establish the facts E[L41 ] = O(|δγ0 Ω00 δγ |2 ),
E[L81 ] = O(|δγ0 Ω00 δγ |4 ) and E[Q411 ] = O(|p + 1 − q|4 ), provided that at least
one of the assumptions (a) or (b) of Lemma C.3(v) holds. But this follows from
Lemma C.4, because if t0 = 0, then from the representations of ∇n and ∆γ in
11 An
explicit argument for the concentration aspect of Bn is missing in WC.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2615
(6.10) and (6.11), we see that the distribution of the quantity of interest does
not depend on µ and we may restrict to µ = 0, whereas, if t0 6= 0, Lemma C.4
shows that T1 has full rank.12 With this, in general, we obtain
q
|E[Li Qik Lk Lj Qjk Lk ]| = |E[Li Lj L2k Qik Qjk ]| ≤ E[L2i L2j L4k ]E[Q2ik Q2jk ]
≤ (E[L81 ])1/2 (E[Q4ik ])1/4 (E[Q4jk ])1/4 ≤ (E[L81 ])1/2 (E[Q411 ])1/2
= O(|δγ0 Ω00 δγ |2 ) O(|p + 1 − q|2 ) = O(q|p + 1 − q|2 /n2 ).
If i 6= k, j 6= k and i 6= j, using the abbreviations v = R00 δγ and M = R10 Ω−1
11 R1 ,
we get the smaller bound
|E[Li Qik Lk Lj Qjk Lk ]| = |E[E[Li Qik |xk , xj ]Lj Qjk L2k ]|
= |E[v 0 SM (1, x0k )0 E[Lj Qjk |xk ]L2k ]| = |E[(v 0 SM (1, x0k )0 )2 L2k ]|
q
q
≤ E[(v 0 SM (1, x0k )0 )4 ]E[L4k ] = O(|v 0 SM SM Sv|2 )O(|δγ0 Ω00 δγ |2 )
0
0
2
2
= O(δγ0 Ω01 Ω−1
11 Ω10 δγ )O(δγ Ω00 δγ ) ≤ O(|δγ Ω00 δγ | ) = O(q/n ),
0
where we have used Lemma C.3 and δγ0 Ω01 Ω−1
11 Ω10 δγ ≤ δγ Ω00 δγ again. It is
now easy to bound the expectations in (6.15). When squaring the expression in
(6.16) we first note the leading factor n−3 . Next, we expand the square of the
bracket term in (6.16) and take expectation. From the previous considerations
we see that those summands in the resulting sum involving Lk Qkk Lk are of order
O(q|p + 1 − q|2 /n2 ), and there are O(n) of them. Together with the leading
factor n−3 and the summation in (6.15) we arrive at a total contribution of
O(q|p+1−q|2 /n3 ) from all those summands involving Lk Qkk Lk . This expression
has to be multiplied by b2n = O(n/q) to yield O(|p + 1 − q|2 /n2 ) = o(1). The
remaining terms are of the form |E[Li Qik Lk Lj Qjk Lk ]| with i 6= k and j 6= k.
Of those, there are a number of O(n) summands where i = j, but they are
again of order O(q|p + 1 − q|2 /n2 ) and therefore, as in the case before, their
total contribution to (6.15) is asymptotically negligible, even after multiplying
by b2n . Finally, there is a number of O(n2 ) remaining summands as above, but
with i 6= k, j 6= k and i 6= j. Therefore, by the refined bound above, they are of
order O(q/n2 ), so that their total contribution to the variance bound in (6.15)
is O(q/n2 ). Together with the factor b2n we arrive at an additional term of order
O(1/n) = o(1). Hence, we see that the variance of bn Bn goes to zero as n → ∞
and the proof of Theorem 2.1(ii) is finished.
6.4.3. The cases of Theorem 2.1(iii) and Theorem 3.1
In this section we establish (6.7) under the assumptions that the design vectors
x1 , . . . , xn are Gaussian, pn /n → ρ1 ∈ [0, 1), qn /n → ρ2 ∈ [0, ρ1 ] (cf. the
beginning of Section 6) and ∆γ = o(qn /n).
12 It should be noted that the case t = 0 corresponds to a null-hypothesis that does not
0
involve a restriction on the intercept parameter α, i.e., H0 : R0 γ = r0 can be expressed as
H0 : T0 β = r0 , in this case.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2616
We restrict to q ≤ p, because this is assumed in Theorem 3.1 and the case
q = p + 1 of Theorem 2.1(iii) has already been treated above in Section 6.4.1,
in much higher generality. We use different arguments for the two cases t0 = 0
and t0 6= 0. If t0 = 0, then, by (6.10) and (6.11), we see that the quantity of
interest is given by
√
s−1/2
n∇n /q − n∆γ bn
n
r
√
n
n − (p + 1) + q
0 −1
0
−1 0 −1
(sn q)−1/2 n δγ0 (T0 Σ̂−1
T
)
δ
−
δ
(T
Σ
T
)
δ
.
=
γ
0
γ
n
0
γ
0
q
n
By Lemma C.5, this expression has mean zero and variance equal to
2(n − (p + 1) + q)
n
n
(δγ0 (T0 Σ−1 T00 )−1 δγ )2
= ∆2γ O(1) = o(1).
2
sn q
n
q
For the case t0 6= 0 we recall T = (R00 , R10 )0 and introduce the matrix ΣT by
0 0
T0 ΣT00 T0 ΣT10
Σ00 Σ01
0 0
0
ΣT = Var[T (1, x1 ) ] = T
T =
=
,
0 Σ
T1 ΣT00 T1 ΣT10
Σ10 Σ11
and note that by Lemma C.4 the sub matrix Σ11 = Var[R1 (1, x01 )0 ] of order
(p + 1 − q) is regular. In the present case the difference in (6.7) is given by
√
s−1/2
n∇n /q − n∆γ bn
n
r
n
−1/2
= (sn q)
×
(6.17)
q
0
√ 0 U0 (In − PU1 )U0
n − (p + 1) + q
nδγ
− (Ω00 − Ω01 Ω−1
δγ .
11 Ω10 )
n
n
The distribution of this quantity is slightly more complicated than that of
the corresponding object in the case t0 = 0, because now we have to deal
with non-centrality issues due to µ 6= 0. We take a closer look at the random part. The joint distribution of the first row of U0 and the first row of
U1 is T (1, x01 )0 ∼ N (T (1, µ0 )0 , ΣT ). Therefore, the conditional distribution of
0 0
R0 (1, x01 )0 given R1 (1, x01 )0 is given by N (µ̃ + Σ01 Σ−1
11 R1 (1, x1 ) , Σ00·1 ), where
−1
−1
0 0
0 0
µ̃ = R0 (1, µ ) − Σ01 Σ11 R1 (1, µ ) and Σ00·1 = Σ00 − Σ01 Σ11 Σ10 . Hence, conditional on U1 , the rows of U0·1 := U0 − U1 Σ−1
11 Σ10 are i.i.d. N (µ̃, Σ00·1 ) and since
this distribution is free of U1 , this also means that U1 and U0·1 are indepen0
(In − PU1 )U0·1 . So if V is a
dent. Also, we clearly have U00 (In − PU1 )U0 = U0·1
random n × q matrix independent of U1 , whose rows are i.i.d. N (0, Σ00·1 ), then
U00 (In − PU1 )U0 has the same distribution as (V + ιµ̃0 )0 (In − PU1 )(V + ιµ̃0 ) =
V 0 (In − PU1 )V + µ̃ι0 (In − PU1 )V + V 0 (In − PU1 )ιµ̃0 + µ̃µ̃0 ι0 (In − PU1 )ι. For the
Schur complement of Ω11 in Ω we use the representation given by Lemma C.6.
Plugging this back into (6.17) and removing the leading (sn q)−1/2 term that
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2617
converges to a positive constant, it remains to study the limiting behavior of
√ 0 V 0 (In − PU1 )V
n − (p + 1) + q
nδγ
− Σ00·1
δγ
(6.18)
n
n
√
ι0 (In − PU1 )V δγ
+ 2 nδγ0 µ̃
(6.19)
n
0
0
2
(δγ µ̃) √
ι (In − PU1 )ι(1 + ν)
n
n − (p + 1) + q
n
+
−1
,
1+ν
n
n − (p + 1) + q
n
(6.20)
p
0 0
multiplied by n/q, where ν = (1, µ0 )R10 Σ−1
11 R1 (1, µ ) is defined as in Lemma C.6.
0
0
From that lemma we also see that δγ Σ00·1 δγ + (δγ µ̃)2 /(1 + ν) = δγ0 Ω00·1 δγ =
∆γ = o(q/n). Since Σ00·1 is the Schur complement of the positive definite
matrix Σ11 within the positive semidefinite matrix ΣT , it follows that Σ00·1
is itself positive semidefinite (consider the minimizer of the quadratic form
u 7→ (v 0 , u0 )0 ΣT (v 0 , u0 )0 in those variables u ∈ Rp+1−q corresponding to the
block Σ11 ). Consequently, δγ0 Σ00·1 δγ ≥ 0 and both δγ0 Σ00·1 δγ and (δγ0 µ̃)2 /(1 + ν)
are bounded by δγ0 Ω00·1 δγ = ∆γ = o(q/n). Now, we first show that the quantity
in (6.20) converges to zero in probability. By Lemma C.7 with λn = nν and
k = p + 1 − q, we have
0
√
ι (In − PU1 )ι(1 + ν)
n
−1
n
n
n − (p + 1) + q
√
ξ/(n − (p + 1) + q)
= n
−1
(ξ + ζ)/(n(1 + ν))
√
√
n (ξ/(n − (p + 1) + q) − 1) + n (1 − (ξ + ζ)/(n(1 + ν)))
=
,
(ξ + ζ)/(n(1 + ν))
where ξ ∼ χ2n−(p+1)+q independent of ζ ∼ χ2p+1−q (nν). The term in the denominator has mean (n + nν)/(n(1 + ν)) = 1 and variance 2(n + 2nν)/(n2 (1 + ν)2 ) =
O(1/n), by independence, and thus, converges to 1 in√probability. From the form
of these moments we also conclude that the term
√ n (1 − (ξ + ζ)/(n(1 + ν)))
is OP (1). Moreover, it is easy to see that also n (ξ/(n − (p + 1) + q) − 1) is
OP (1), which entails that the entire expression in the previous display is of order OP (1), and hence, thepexpression in (6.20) converges to zero in probability
even after multiplying by n/q. Next, the mixed term in (6.19) is easily seen to
have conditional distribution N (0, 4δγ0 Σ00·1 δγ ι0 (In − PU1 )ι(δγ0 µ̃)2 /n) given U1 .
By the previous considerations, the conditional variance is equal to
4δγ0 Σ00·1 δγ ι0 (In − PU1 )ι(δγ0 µ̃)2 /n = O(∆2γ )
ι0 (In − PU1 )ι(1 + ν)
= OP (∆2γ ),
n
which converges to zero in probability, even after multiplying by n/q, and this
implies convergence to zero of the scaled mixed term itself. Finally, for the expression in (6.18), we note that conditional on U1 , V 0 (In − PU1 )V has a Wishart
distribution with scale matrix Σ00·1 and n−(p+1)+q degrees of freedom. Therefore, (6.18) has mean zero and δγ0 V 0 (In −PU1 )V δγ ∼ δγ0 Σ00·1 δγ χ2n−(p+1)+q [cf. 19,
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2618
Theorem 3.4.2], which entails that the variance of (6.18), multiplied by n/q, is
(n/q)(δγ0 Σ00·1 δγ )2 2(n − (p + 1) + q)/n = o(q/n), since δγ0 Σ00·1 δγ ≤ ∆γ = o(q/n).
This finishes the proofs of Theorem 2.1 and Theorem 3.1.
Acknowledgements
I want to thank Hannes Leeb and the three referees for many valuable suggestions that helped to considerably improve the paper. This research was supported by the Austrian Science Fund (FWF): [P 28233-N32].
Appendix A: Auxiliary results of Section 4
Lemma A.1. For m ∈ N, let Z = (Z1 , . . . , Zm )0 be a spherically symmetric
random vector in Rm such that E[ZZ 0 ] = Im , let V ∼ N (0, Im ) and let z1 , . . . , zn
be i.i.d. copies of Z. For p ≤ m, let Γ be a p×m matrix of full rank p, let µ ∈ Rp
and define xi = Γzi + µ.
(i) Fix r ∈ N. If E[kZk2r ] < ∞ and for every choice of non-negative integers
Pm
Qm
Qm
`
`
`j with j=1 `j ≤ 2r, we have E[ j=1 Zj j ] = j=1 E[Zj j ], then E[Z12l ] =
E[V12l ] and E[kZk2l ] = E[kV k2l ] for every l = 1, . . . , r.13
(ii) If 2 ≤ p ≤ n − 2 and Z also satisfies P(kZk = 0) = 0, then the random vectors x1 , . . . , xn satisfy Assumptions (A1).(a,b). Moreover, if also
E[kZk8 ] < ∞, E[kZk4 ]/E[kV k4 ] → 1 and E[kZk8 ]/E[kV k8 ] = O(1), as
m → ∞, then also Assumptions (A1).(d,e) hold.
(iii) √
If Z follows the uniform distribution on the ball (of appropriate radius
m + 2, to ensure E[ZZ 0 ] = Im ) and 2 ≤ p ≤ n − 2, then the xi , for
i = 1, . . . , n, satisfy the full Assumption (A1), but not Assumption (C1).
Proof. We make use of the well known fact that any spherical distribution can be
represented as Z = bkZk, where b and kZk are independent, and b is uniformly
distributed on the unit P
m-sphere S m−1 [cf. 8].
m
For part (i), set ` = j=1 `j and let ei ∈ Rm denote the i-th element of the
m
standard basis in R and note that
m
m
m
Y
Y
Y
`
E
Zj j = E (e0j Z)`j = E kZk`
(e0j b)`j
j=1
j=1
j=1
m
Y
= E kZk` E (e0j b)`j .
j=1
Of course, the same argument can be carried through for the spherical vector
13 Due to symmetry, we always have E[Z l ] = 0 = E[V l ] if l is odd and the former moment
1
1
exists.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
V ∼ N (0, Im ), so that we have
hQ
i
`j
m
E
E kZk`
j=1 Zj
i=
hQ
,
`j
m
E [kV k` ]
E
V
j=1 j
2619
(A.1)
Qm
`
provided that all the `j are even, so that E[ j=1 Vj j ] 6= 0. Now choose the
`j to be either equal to 2 or 0, such that ` is any even number from 2 to 2r.
Therefore, since E[Z12 ] = 1 = E[V12 ] and by our factorization assumption, the
left-hand-side of (A.1) is equal to one, so that we have established the equality
of even moments of kZk and kV k. To see that also the even moments of Z1 and
V1 coincide, simply choose `1 = ` = 2l, for some l ∈ {1, . . . , k} and `j = 0, if
j 6= 1.
For part (ii), to establish Assumption (A1).(b), first note that the column
span of U−1 = [ι, [z2 , . . . , zn ]0 Γ0 + ιµ0 ] does not depend on µ ∈ Rp , where ι =
(1, . . . , 1)0 ∈ Rn−1 . So we may assume without restriction that µ = 0. Moreover,
[ι, [z2 , . . . , zn ]0 Γ0 ] and
0
0 0 1
[ι, [z2 , . . . , zn ] Γ ]
0 Σ−1/2
also have the same column span, such that it suffices to determine the rank of
[ι, [x̃2 , . . . , x̃n ]0 ], where x̃i = (ΓΓ0 )−1/2 Γzi is spherically symmetric with P(kx̃i k =
0) = 0. Next, we claim that the matrix Mk = [x̃2 , . . . , x̃k+1 ]0 has full rank p,
almost surely, provided that k ≥ p. To see this, simply write Mk = D1 D2−1 Λ,
almost surely, where D1 is k × k diagonal with entries kx̃2 k, . . . , kx̃k+1 k, D2
is k × k diagonal with i.i.d. χp entries and Λ is k × p and has i.i.d. N (0, 1)
entries. Now it is easy to see that D1 D2−1 is almost surely of full rank, and
thus, P(rank(Mk ) = p) = P(det(Λ0 Λ) 6= 0) = 1, where the last equality follows
from the well known fact that the zero-set of a non-constant polynomial is a
Lebesgue null-set. It remains to show that the event A = {ι ∈ span(Mn−1 )}
has probability zero. Define B = {rank(Mp ) = p}, ιk = (1, . . . , 1)0 ∈ Rk and the
function v : Rp×p → Rp by v(M ) = M −1 ιp , if det(M ) 6= 0, and v(M ) = 0, else.
Note that P(B) = 1 and v is Borel measurable. Since p + 2 ≤ n, we see that
A ∩ B is a subset of the event where both ιp = Mp v(Mp ) and 1 = x̃0p+2 v(Mp ),
the probability of which is clearly bounded by
P(1 = x̃0p+2 v(Mp )) = P kx̃p+2 k−1 = (x̃0p+2 /kx̃p+2 k)v(Mp ), kx̃p+2 k =
6 0
= E P kx̃p+2 k−1 = (x̃0p+2 /kx̃p+2 k)v(Mp ), kx̃p+2 k =
6 0 Mp , kx̃p+2 k .
But the conditional probability in the previous display is equal to zero, almost surely, because v(Mp ), x̃p+2 /kx̃p+2 k and kx̃p+2 k are independent and
x̃p+2 /kx̃p+2 k is uniformly distributed on the unit sphere in Rp , and therefore its
inner product with any fixed vector has a Lebesgue density on R provided that
p ≥ 2. For Assumptions (A1).(d,e), recall the moments of the χ2 -distribution
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
with m-degrees of freedom E[kV k2k ] =
ing as in part (i) yields
Qk−1
j=0 (m + 2j)
E[|v 0 Z|8 ] = E[|v 0 b|8 ]E[kZk8 ] =
2620
[cf. 15]. The same reason-
E[kZk8 ]
E[|v 0 V |8 ] = O(1),
E[kV k8 ]
uniformly in v ∈ S m−1 . Similarly, for a symmetric matrix M ∈ Rm×m ,
E[kZk4 ]
E[(V 0 M V )2 ],
E[kV k4 ]
and one easily calculates E[(V 0 M V )2 ] = (trace(M ))2 + 2trace M 2 . Therefore,
E[(Z 0 M Z)2 ] = E[(b0 M b)2 ]E[kZk4 ] =
Var[Z 0 M Z] = (trace(M ))2
E[kZk4 ]
E[kZk4 ]
−
1
+
2
trace M 2 .
4
4
E[kV k ]
E[kV k ]
Finally, for a projection matrix P ∈ Rm×m , V 0 P V follows a χ2 -distribution
with rank P = kP k2F degrees of freedom, and thus
1/4
E[(Z 0 P Z)4 ]
=
=
E[kZk8 ]
E[kV k8 ]
1/4
E[kZk8 ]
E[kV k8 ]
1/4
E[(V 0 P V )4 ]
3
Y
1/4
1/4
(kP k2F + 2j)
= O(kP k2F ).
j=0
To establish part (iii), we first verify the conditions of part (ii). The finiteness
of the 8-th moment of the radial component and P(kZk = 0) = 0 are immediate.
It is also elementary to calculate the higher non-central moments E[kZk2k ] =
(m+2)k m/(m+2k). [Use, for example, the formula
for the volume of the
√
√ m-ball
of radius r > 0 to obtain P(kZk ≤ x) = (x/ m + 2)m , for x ∈ [0, m + 2].]
Qk−1
Comparing this to the moments of the χ2m distribution E[kV k2k ] = j=0 (m+2j)
for k = 2, 4, we see that for m → ∞ the moment ratios behave as desired.
Therefore, Assumptions (A1).(a,b,d,e) hold in this case. Finally, the validity of
Assumption (A1).(c) follows from Srivastava and Vershynin [23, Section 1.4].
But Condition (C1) can not be satisfied in view of part (i) (with r = 4) and the
fact that E[kZk4 ] 6= E[kV k4 ].
Lemma A.2. Let Z = (Z1 , . . . , Zm )0 be a random m-vector with E[Z] = 0 and
E[ZZ 0 ] = Im and let z1 , . . . , zn be i.i.d. copies of Z. For p ≤ m, let Γ be a p × m
matrix of full rank p, let µ ∈ Rp and define xi = Γzi + µ.
(i) If Z has independent components, whose 8-th moments are uniformly
bounded, then the xi satisfy Assumptions (A1).(a,c,d,e).
(ii) If z1 , . . . , zn are as in Condition (C1) and, in addition, the components
of Z have 8-th moments that are uniformly bounded, then the xi satisfy
Assumptions (A1).(a,d,e).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2621
Proof. To establish part (i), we use the results of Whittle [28]. Theorem 2 in
that reference shows that for a unit vector v = (v1 , . . . , vm )0 ∈ Rm ,
4
X
E[|v 0 Z|8 ] ≤ C
vj2 (E[|Zj |8 ])1/4 ≤ C max E[|Zj |8 ],
j
j
for some numerical constant C > 0, and thus, L8 = O(1) as m → ∞, in view
of uniform boundedness of E[|Zj |8 ]. Next, for a symmetric matrix M ∈ Rm×m ,
the same theorem yields
q
X
2
E[|Zj |4 ]E[|Zk |4 ] ≤ C max E[|Zj |4 ]trace M 2 ,
Var[Z 0 M Z] ≤ C
Mjk
j
j,k
and, for a projection matrix P ∈ Rm×m ,
(E[(Z 0 P Z)4 ])1/4 − E[Z 0 P Z] ≤ E[(Z 0 P Z − E[Z 0 P Z])4 ]
1/2
X
1/4
2
≤ C 1/4
Pjk
E[|Zj |8 ]E[|Zk |8 ]
1/4
j,k
≤
C max E[|Zj |8 ]
j
1/4
kP kF ,
where the first inequality is the reverse triangle inequality for the L4 -norm. Now
the previous chain of inequalities implies that
(E[(Z 0 P Z)4 ])1/4 ≤ E[Z 0 P Z] + DkP kF ≤ kP k2F + DkP k2F = O(kP k2F ),
since E[Z 0 P Z] = trace(P ) = trace P 2 = kP k2F = rank P is integer, and where
D > 0 is an appropriate constant, not depending on m. The validity of Assumption (A1).(c) follows from the arguments in Section 1.4 in Srivastava and
Vershynin [23].
For part (ii), simply note that under the factorization assumption in (C1) all
the moments occurring in Conditions (A1).(d,e) are identical to those calculated
under independence of the components of Z. Therefore, the result follows from
part (i).
Appendix B: Proofs of auxiliary results of Section 6.2
Proof of Lemma 6.1. For ease of notation we drop the subscript n that indexes
the position of the matrix An in the array, i.e., we write A = An and denote by
aij the ij-th entry of that matrix. Similarly, we write Z = (Z1 , . . . , Zn )0 , where
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2622
Zi = Zi,n . Now, expand
Z 0 AZ − E[Z 0 AZ|Gn ]
=
n
X
Zi Zj aij +
=
2Zj
j=1
=
T̄n
ajj (Zj2 − 1)
j=1
i6=j
n
X
n
X
j−1
X
Zi aij
+
i=1
n
X
ajj (Zj2 − 1)
j=1
Tn∗ ,
+
where we adopt the convention that empty sums are equal to zero. We show
√
√
i.p.
w
→ N (0, 1) and Tn∗ /( 2kAkF ) −−→ 0, as n → ∞.
that T̄n /( 2kAkF ) −
The desired convergence of Tn∗ follows from the straight forward calculation
#
"
2
n
X
a2jj E[(Zj2 − 1)2 |Gn ]
Tn∗
√
Gn =
E
2kAk2F
2kAkF
j=1
=
≤
n
X
a2jj (E[Zj4 |Gn ] − 1)
j=1
n
X
1
2
j=1
2kAk2F
a2jj E[Zj4 |Gn ]
,
kAk2F
and by assumption.
To see the weak convergence of T̄n , for j = 1, . . . , n, define
Vn,j
=
√
2Zj
j−1
X
Zi aij /kAkF ,
i=1
Fn,0 = Gn and Fn,j = σ (Gn , Zi : i ≤ j), by which we mean the smallest sigma
algebra for which Z1 , . . . , Zj are measurable and which alsopcontains Gn . Note
that for each n, j ∈ N, Fn,j−1 ⊆ Fn,j ⊆ F, kAn kF =
trace(A2 ) is Fn,0
measurable and Vn,j is Fn,j measurable. Moreover, we have
n
√
X
T̄n
=
Vn,j .
2kAkF
j=1
Now, by the central limit theorem for dependent random variables [see 10, 13,
and notice the discussion in Helland [13] following eq. (2.7)], it remains to verify
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2623
that
n
X
j=1
n
X
j=1
n
X
i.p.
E[Vn,j |Fn,j−1 ] −−−−→ 0,
(B.1)
n→∞
i.p.
Var[Vn,j |Fn,j−1 ] −−−−→ 1,
n→∞
and
i.p.
2
E[Vn,j
1|Vn,j |>δ |Fn,j−1 ] −−−−→ 0
n→∞
j=1
(B.2)
for all δ > 0,
(B.3)
as in equations (2.5)-(2.7) in Helland [13]. The convergence in (B.1) is trivial,
since E[Vn,j |Fn,j−1 ] = 0, in view of the conditional independence of the Zi given
Gn .
Pn
Pn
2
For (B.2), abbreviate Tn =
j=1 Var[Vn,j |Fn,j−1 ] =
j=1 E[Vn,j |Fn,j−1 ]
and use conditional independence again to obtain
2
E[Vn,j
|Fn,j−1 ]
=
2kAk−2
F
j−1
X
!2
Zi aij
.
i=1
Expanding the squared sum gives
j−1
X
!2
Zi aij
=
j−1
X
i=1
Zi Zk aij akj =
j−1
X
Zi2 a2ij + 2
i=1
i,k
j−1
X
Zi Zk aij akj ,
i<k
and therefore, the absolute difference |Tn − 1| can be bounded as
|Tn − 1|
=
kAk−2
F
n
X
2
j=1
≤
kAk−2
F
2
n
X
i<j
j−1
X
Zi2 a2ij
+2
i=1
(Zi2
j−1
X
!
Zi Zk aij akj
− kAk2F
i<k
−
1)a2ij
+4
j−1
n X
X
j=1 i<k
Zi Zk aij akj +
n
X
!
a2jj
.
j=1
To establish the convergence in (B.2), it remains to show convergence to zero
in probability of the terms in absolute values on the last line of the preceding
−2
display multiplied by kAkP
F .
n
−2
2
First, note that kAkF
by assumpj=1 ajj converges to zero in probability
Pn
4
2
2
tion and because of E[Zj |Gn ] ≥ 1. Now, write Tn,1 =
i<j (Zi − 1)aij and
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
Tn,2 =
2624
√ Pn Pj−1
2 j=1 i<k Zi Zk aij akj and observe that
2
E[Tn,1
|Gn ] =
n
n
X
X
E[(Zi21 − 1)(Zi22 − 1)|Gn ]a2i1 j1 a2i2 j2
i1 <j1 i2 <j2
n j1 ∧j
2 −1
X
X
=
(E[Zi4 |Gn ] − 1)a2ij1 a2ij2
j1 ,j2
i=1
X
n j1 ∧j
2 −1
X
≤
max
j=1,...,n
E[Zj4 |Gn ]
j1 ,j2
a2ij1 a2ij2 ,
i=1
and
n jX
1 −1 jX
2 −1
X
2
E[Tn,2
|Gn ] = 2
E[Zi1 Zk1 Zi2 Zk2 |Gn ]ai1 j1 ak1 j1 ai2 j2 ak2 j2
j1 ,j2 i1 <k1 i2 <k2
=
n
X
2
j1 ∧j
2 −1
X
j1 ,j2
≤
aij1 akj1 aij2 akj2
i<k
max E[Zj4 |Gn ]
X
n j1 ∧j
2 −1
X
j=1,...,n
j1 ,j2
aij1 akj1 aij2 akj2 .
i6=k
Therefore, if we define
the triangular truncation operator à of the symmetric
P
matrix A by à = s>t es ast e0t , where es ∈ Rn is the s-th element of the standard
basis in Rn , we see that
!2
n
n
2
X
X
trace Ã0 Ã
= trace
et1 as1 t1 e0s es2 as2 t2 e0t
1
2
s1 >t1 s2 >t2
= trace
s−1
n
X
X
!2
et1 ast1 ast2 e0t2
s=1 t1 ,t2 =1
n
X
=
trace
s1 ,s2 =1
n
X
=
sX
1 −1
et1 as1 t1 as1 t2 e0t2
t1 ,t2 =1
s1 ∧s
2 −1
X
sX
2 −1
!
eu1 as2 u1 as2 u2 e0u2
u1 ,u2 =1
as1 t1 as1 t2 as2 t2 as2 t1 ,
(B.4)
s1 ,s2 =1 t1 ,t2 =1
and, in turn, that
2
2
E[Tn,1
|Gn ] + E[Tn,2
|Gn ]
X
n j1 ∧j
2 −1
X
≤
max E[Zj4 |Gn ]
aij1 akj1 aij2 akj2
j=1,...,n
=
j1 ,j2
i,k
4
max E[Zj |Gn ] trace (Ã0 Ã)2 .
j=1,...,n
(B.5)
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2625
Now, convergence to zero of kAk−2
F (|Tn,1 | + |Tn,2 |) in probability follows from
the above considerations and Lemma 2.1 in Bhansali, Giraitis and Kokoszka [5],
which yields the inequality
h
i
4
2
0
2
E kAk−4
G
≤
2
max
E[Z
|G
]
kAk−4
(|T
|
+
|T
|)
n
n
n,1
n,2
j
F
F kà ÃkF
j=1,...,n
kAk2S
,
≤ 2C 2
max E[Zj4 |Gn ]
j=1,...,n
kAk2F
where C > 0 is a global constant, not depending on n. Thus, by assumption,
the bound on the far right-hand-side of the preceding display converges to zero,
in probability, which establishes the convergence in (B.2).
Finally, for (B.3) we abbreviate mn = maxj E[Zj4 |Gn ] and use the upper
2
4
bound Vn,j
1|Vn,j |>δ ≤ δ −2 Vn,j
. Now,
!4
j−1
n
n
X
X
X
4
E[Zj4 |Gn ]E
Zi aij
E
E[Vn,j
|Fn,j−1 ] Gn = 4kAk−4
Gn
F
j=1
j=1
= 4kAk−4
F
n
X
E[Zj4 |Gn ] 3
j−1
X
i=1
a2i1 j a2i2 j +
j−1
X
i1 6=i2
j=1
≤ 4mn (mn + 3)kAk−4
F
E[Zi4 |Gn ]a4ij
i=1
j−1
n
X
X
a2i1 j a2i2 j ,
j=1 i1 ,i2 =1
and furthermore
kAk−4
F
j−1
n
X
X
a2i1 j a2i2 j = kAk−4
F
j=1 i1 ,i2 =1
j−1
n
X
X
j=1
≤
=
kAk−4
F
!2
a2ij
i=1
max
j
n
X
i=1
!
a2ij
n
X
a2ij
i,j=1
maxj (A2 )jj
.
kAk2F
Together with mn ≥ 1 and our assumption, this implies that the upper bound
on the second-to-last display converges to zero in probability.
Proof of Lemma 6.3. For convenience, we drop the subscript n that indexes the
position in the array whenever there is no risk of confusion. Let wi0 = (1, x0i )R
−1/2
−1/2
denote the i-th row of the matrix W and define w̃i = ΩW wi , W̃ = W ΩW
P
−1/2
−1/2
n
0
and S1 = W̃ 0 W̃ − w̃1 w̃10 = i=2 w̃i w̃i0 = ΩW R0 U−1
U−1 RΩW , where
1
µ0
R
ΩW = E[w1 w10 ] = R0
µ Σ + µµ0
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2626
is positive definite and U−1 is defined as in Assumption (A1).(b). This assumption also entails that W 0 W , W̃ 0 W̃ and S1 are invertible with probability one,
where we denote the corresponding null set by N . For convenience, we redefine
these quantities in an arbitrary invertible and measurable way on N . Moreover,
we must also have pn + 2 ≤ n under (A1).(b).
Since hj = hj,n = wj0 (W 0 W )−1 wj on N c , permuting the h1 , . . . , hn is equivalent to a permutation of w1 , . . . , wn , which are i.i.d., and therefore their joint
distribution is invariant under permutation. Hence, the hj are exchangeable
random variables.
Pn In particular, the hj are identically distributed and therefore
the fact that j=1 hj = trace(PW ) = kn , on N c , entails that E[h1 ] = kn /n.
We also note for later use that Var[h1 ] = E[h21 ] − E[h1 ]2 ≤ E[h1 ] − E[h1 ]2 =
(1 − kn /n)kn /n, since 0 ≤ h1 ≤ 1. It only remains to show that the variance
actually converges to zero.
As a preliminary consideration, we study h1 in the case where tn := kn /n →
t ∈ [0, 1]. The general case of possibly non-converging tn then follows from a
standard subsequence argument (see the end of the proof). The case t ∈ {0, 1} is
immediate, because here Var[h1 ] → 0 as n → ∞, and thus, h1 → t in probability,
by the arguments in the previous paragraph. Assume now that t ∈ (0, 1). Note
that PW = PW̃ and use the Sherman-Morrison formula to obtain
h1 = w̃10 (S1 + w̃1 w̃10 )
−1
w̃1 =
w̃10 S1−1 w̃1
,
1 + w̃10 S1−1 w̃1
at least on N c . For α ≥ 0, set Jα = (S1 + (n − 1)αIkn )−1 and define the random
function
w̃10 Jα w̃1
Ψn (α) =
,
1 + w̃10 Jα w̃1
which satisfies Ψn (0) = h1 , almost surely. Since y 7→ y/(1 + y) is non-decreasing
on [0, ∞), and M 7→ M −1 is non-increasing on invertible hermitian matrices
[cf. 6, p. 114], the function Ψn is non-increasing on [0, ∞). We establish the
convergence in probability of Ψn (0) by first analyzing the limiting behavior of
Ψn (α) as n → ∞, for every α > 0. To this end, we consider the conditional
mean and variance of w̃10 Jα w̃1 given S1 .
Since w̃1 and S1 are independent and E[w̃1 w̃10 ] = Ikn , one easily calculates
E[w̃10 Jα w̃1 |S1 ] = trace(Jα ). The conditional variance is slightly more involved.
−1/2
Abbreviate ΣW = Var[w1 ], µ̄ = E[w̃1 ] = ΩW R0 (1, µ0 )0 and use Assumption (A1).(a) to write
0 0
0
−1/2
w̃1 = µ̄ + ΩW R0
.
0 Γ
z1
−1/2
−1/2
Also notice that ΣW = ΩW − R0 (1, µ0 )0 (1, µ0 )R, and hence ΩW ΣW ΩW
=
−1/2
−1/2
0
Ikn − µ̄µ̄ . Since ΩW ΣW ΩW is positive semidefinite, this also implies that
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
kµ̄k ≤ 1. Now, decompose the quantity of interest
0 0
0
−1/2
w̃10 Jα w̃1 = µ̄0 Jα µ̄ + 2µ̄0 Jα ΩW R0
0 Γ
z1
0 0
0
0
−1/2
−1/2 0 0
+ (0, z10 )
RΩ
J
Ω
R
.
α W
W
0 Γ0
0 Γ
z1
2627
(B.6)
Conditional on S1 , the variance of µ̄0 Jα µ̄ is zero, the variance of half of the
mixed term is
−1/2
µ̄0 Jα ΩW
−1/2
ΣW ΩW
Jα µ̄ = µ̄0 Jα2 µ̄ − (µ̄0 Jα µ̄)2
and the variance of the last term in (B.6) is
2
Var[(0, z10 )M (0, z10 )0 |S1 ] = Var[z10 M22 z1 |S1 ] = O(trace M22
) + (trace(M22 ))2 o(1)
≤ O(trace M 2 ) + (trace(M ))2 o(1),
by assumption, and where we have abbreviated the symmetric positive semidefinite matrix in between the vectors (0, z10 ) and (0, z10 )0 in (B.6) by M and used
the notation M22 to denote its bottom
right sub matrix of order m × m. Now,
−1/2
−1/2
trace(M ) = trace Jα ΩW ΣW ΩW
= trace(Jα ) − µ̄0 Jα µ̄ and trace M 2 =
−1/2
−1/2
−1/2
−1/2
trace Jα ΩW ΣW ΩW Jα ΩW ΣW ΩW
= trace Jα2 − 2µ̄0 Jα2 µ̄ + (µ̄0 Jα µ̄)2 .
For α > 0, kJα kS ≤ [α(n − 1)]−1 and kµ̄k2 ≤ 1. Therefore, µ̄0 Jα µ̄ and µ̄0 Jα2 µ̄
converge to zero, almost surely, for every α > 0. Thus, in order to show that
Var[w̃10 Jα w̃1 |S1 ] converges
to zero, almost surely, for every α > 0, it suffices to
show that trace Jα2 → 0 as n → ∞, almost surely, and that trace(Jα ) is almost
surely convergent. Moreover, if we can even show that trace(Jα ) → ψα ∈ [0, ∞),
almost surely, for every α > 0, then we also have w̃10 Jα w̃1 → ψα , in probability
(since E[w̃10 Jα w̃1 |S1 ] = trace(Jα )), and thus Ψn (α) → Ψ(α) := ψα /(1 + ψα ), in
probability, for every α > 0.
Therefore, we need to study the limiting behavior of
kn
1 X
1
kn
(n − 1)` kn j=1 (λj + α)`
Z ∞
kn
=
(y + α)−` dF S1 /(n−1) (y),
(n − 1)` 0
trace (S1 + (n − 1)αIkn )−` =
(B.7)
for
· · ≤ λkn are the ordered eigenvalues of S1 /(n−1) =
Pn` = 1, 2,0 where 0 ≤ λ1 ≤S1·/(n−1)
denotes the corresponding empirical spectral
j=2 w̃j w̃j /(n − 1) and F
distribution function. Now one easily verifies the assumptions of Theorem 1.1 in
Bai and Zhou [3]. First note that the w̃j are i.i.d. and E[w̃1 w̃10 ] = Ikn . Second,
by the same argument following (B.6) and for an arbitrary non-random kn × kn
matrix B with bounded spectral norm, we have
E[|w̃10 B w̃1 − trace(B)|2 ] = Var[w̃10 (B/2 + B 0 /2)w̃1 ]
= O(trace (B/2 + B 0 /2)2 ) + (trace(B) + O(1))2 o(1) + O(1)
= O(kn kBk2S ) + kn2 kBk2S o(1) + O(1) = o(n2 ).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2628
Recall also that for now tn = kn /n → t ∈ (0, 1). Therefore, F S1 /(n−1) converges weakly, almost
Lebesgue
p surely to the Marčenko-Pastur distribution with √
P
(y
−
a)(b
−
y)/(2πty)
on
[a,
b],
where
a
=
(1
−
t)2 and
density f M
(y)
=
√
b = (1 + t)2 . Now we see that we can not use the same strategy to establish the convergence of (B.7) in the case where α = 0, because the function
hα (y) = (y + α)−1 is not bounded on (0, ∞) in that case. However, for α > 0,
hα is bounded and continuous on [0, ∞) and therefore the integral in (B.7)
converges almost surely,
Z b
Z ∞
a.s.
`
S1 /(n−1)
h`α (y)f M P (y) dy ∈ (0, ∞).
hα (y) dF
(y) −−→
0
a
Since kn /(n − 1) → t ∈ (0, 1) and kn /(n − 1)2 → 0, this means that trace(Jα ) →
Rb
ψα := t a hα (y)f M P (y) dy and trace Jα2 → 0, almost surely, for every α > 0.
As discussed at the end of the previous paragraph, this entails that Ψn (α) →
Ψ(α), in probability, for every α > 0, where the function Ψ is given by
Rb
t a hα (y)f M P (y) dy
.
Ψ(α) =
Rb
1 + t a hα (y)f M P (y) dy
Now it is easy to see (e.g., by the dominated convergence theorem) that the
function Ψ is continuous and non-increasing on [0, ∞) (recall that here t < 1
and a > 0). Moreover, the limiting integral for α = 0 can be evaluated as
Rb
h (y)f M P (y) dy = 1/(1 − t) [cf. 14, Lemma B.1] resulting in Ψ(0) = t.
a 0
Let us briefly recapitulate what we have found so far. First of all, we have
seen that h1 = h1,n → t in probability, as n → ∞, if tn → t ∈ {0, 1}. For
tn → t ∈ (0, 1), we know that E[h1,n ] = kn /n → t and 0 ≤ Ψn (α) ≤ Ψn (0) =
h1 ≤ 1 almost surely. Moreover, Ψn (α) → Ψ(α) in probability, for every α > 0,
and Ψ(α) → Ψ(0) = t as α → 0. Thus, Lemma C.2 applies and we obtain that
h1,n → t in probability, also in the case t ∈ (0, 1).
Finally, consider ∆n := |h1 −kn /n| with arbitrary tn = kn /n ∈ [0, 1]. Suppose
that c := lim sup E[∆n ] > 0. Then there exists a subsequence n0 , such that
E[∆n0 ] → c, as n0 → ∞. By compactness, there exists a further subsequence n00 ,
such that tn00 → t ∈ [0, 1], as n00 → ∞. But in this case, our previous arguments
have shown that ∆n00 → 0 in probability, which also entails that E[∆n00 ] → 0,
by boundedness, contradicting the assertion that E[∆n0 ] → c > 0.
Appendix C: Other technical results
Pn
Lemma C.1. Under the model (2.1), suppose that n1 i=1 E[(εi /σn )4 |xi ] =
OP (1) and
√ rank(U ) = pn +1, almost surely, for all n ∈ N. If lim supn→∞ pn /n <
1, then n|σ̂n2 /σn2 −1| = OP (1). In particular, we have P(σ̂n2 = 0) → 0 as n → ∞.
Pn
Remark. The assumption that n1 i=1 E[(εi /σn )4 |xi ] = OP (1) is clearly weaker
than a uniform bound on E[(ε1 /σn )4 ] or a uniform bound P
on E[ε41 ] together with
n
1
2
lim inf n σn > 0. Clearly, also Assumption (A2) implies n i=1 E[(εi /σn )4 |xi ] =
OP (1).
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2629
Proof of Lemma C.1. Recall that σ̂n2 = ε0 M ε, almost surely, where M := M (X) :=
(In − PU )/(n − p − 1) is a function of the design matrix X. Note that y1 , . . . , yn
are conditionally independent given X, and hence, also εi = yi − E[yi |xi ], for
i = 1, . . . , n, are conditionally independent given X. Therefore, one easily obtains the almost sure identities
E[σ̂n2 /σn2 |X] = trace(M )
Var[σ̂n2 /σn2 |X] = 2trace M 2 +
and
n
X
(E[(εi /σn )4 |xi ] − 3)Mii2 .
i=1
By our assumption on U , with probability one, trace(M ) = 1 and trace M 2 =
1/(n − p − 1), whereas Mii2 ≤ 1/(n − p − 1)2 holds everywhere, since the diagonal
entries of the projection matrix In − PU are always between 0 and 1. Taken
together, we see that E[σ̂n2 /σn2 |X] = 1 and Var[σ̂n2 /σn2 |X] ≤ 2/(n − p − 1) +
P
n
4
2
i=1 E[(εi /σn ) |xi ]/(n − p − 1) . Now, the conditional Markov inequality yields
√
√
P( n|σ̂n2 /σn2 − 1| > δ) = E[P( n|σ̂n2 /σn2 − 1| > δ|X) ∧ 1]
i
h n
2
2
Var[σ̂
/σ
|X]
∧
1
≤E
n
n
δ2
1
≤ P(n Var[σ̂n2 /σn2 |X] > δ) +
∧1 .
δ
Since n Var[σ̂n2 /σn2 |X] = OP (1), in view of the
Pnprevious considerations and the
assumptions lim supn→∞ pn /n < 1 and n1 i=1 E[(εi /σn )4 |xi ] = OP (1), this
finishes the proof of the first claim. The second assertion follows immediately,
because of P(σ̂n2 = 0) ≤ P(|σ̂n2 /σn2 − 1| > 1/2).
Lemma C.2. For n ∈ N and α > 0, let hn and Ψn (α) be real random variables
such that 0 ≤ Ψn (α) ≤ hn ≤ 1 almost surely, and, for t ∈ [0, 1], let Ψ : [0, ∞) →
[0, 1] be such that Ψ(α) → t as α → 0. If for every α > 0, Ψn (α) →
− Ψ(α) in
probability, and E[hn ] →
− t as n → ∞, then hn →
− t in probability, as n → ∞.
Remark. Lemma C.2 is an asymptotic version of the well known fact that a
random variable h that satisfies h ≥ t ∈ R and E[h] = t must be equal to t,
almost surely.
Proof of Lemma C.2. Fix δ > 0, choose α = α(δ) > 0 such that |Ψ(α)−t| < δ/2
and do the following standard bound,
P(hn < E[hn ] − δ)
≤
P(|Ψn (α) − E[hn ]| > δ)
+ P(hn < E[hn ] − δ, |Ψn (α) − E[hn ]| ≤ δ).
But |Ψn (α) − E[hn ]| ≤ |Ψn (α) − Ψ(α)| + |Ψ(α) − t| + |t − E[hn ]| ≤ δ/2 + oP (1),
whereas |Ψn (α) − E[hn ]| ≤ δ and hn < E[hn ] − δ together imply that Ψn (α) ≥
E[hn ]−δ > hn , which, by assumption, happens only on a set of probability zero.
Therefore, the upper bound in the previous display converges to zero. Now, by
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2630
boundedness of hn we have
E[|hn − E[hn ]|]
≤
≤
E[|hn − E[hn ] + δ|] + δ
E[(hn − (E[hn ] − δ))1{hn ≥E[hn ]−δ} ]
+ P(hn < E[hn ] − δ) + δ
≤
E[hn 1{hn ≥E[hn ]−δ} ] − E[hn ]P(hn ≥ E[hn ] − δ)
+ 2δ + o(1),
and we also see that both E[hn 1{hn ≥E[hn ]−δ} ] and E[hn ]P(hn ≥ E[hn ] − δ) converge to t. Since δ > 0 was arbitrary, we must have lim sup E[|hn − E[hn ]|] = 0
and thus, convergence in probability of hn to t.
Lemma C.3. For every n ∈ N, let x1,n , . . . , xn,n be i.i.d. random pn -vectors
satisfying xi,n = µn + Γn zi,n as in Assumption (A1).(a) with positive semidefinite covariance matrix Σn = Γn Γ0n . Set Xn = [x1,n , . . . , xn,n ]0 , Σ̂n = Xn0 (In −
Pι )Xn /n and
1
1
µ0n
0 0
1
1 x01 =
1 µ0n .
Sn = E
=
+
0
x1
µn Σn + µn µn
0 Σn
µn
Moreover, let Rn be a kn × (pn + 1) matrix such that Rn Rn0 = Ikn (i.e., kn ≤
pn + 1) and set Ωn = Rn Sn Rn0 .
(i) Let un ∈ Rpn +1 . If supkwk=1 E[|w0 z1,n |` ] = O(1) as n → ∞, for some fixed
` ∈ N, not depending on n, then E[|u0n (1, x01,n )0 |` ] = O(|u0n Sn un |`/2 ) as
n → ∞.
(ii) Let vn,1 , vn,2 ∈ Rpn . If supkwk=1 E[|w0 z1,n |4 ] = O(1) as n → ∞, then
√ 0
0
0
Var[ nvn,1
Σ̂n vn,2 ] = O(vn,1
Σvn,1 vn,2
Σvn,2 ).
(iii) If Σn is positive definite, z1,n satisfies (A1).(c) and supkwk=1 E[|w0 z1,n |4 ] =
O(1), then the design matrix of the transformed data Wn = [ι, X]Rn0 satisfies
(
oP (1), if kn /n → 0,
−1/2
0
−1/2
Ωn (Wn Wn /n)Ωn
− Ikn
=
S
OP (1), if kn = O(n).
0
(iv) If Σn is positive definite and Var[z1,n
M z1,n ] = O(trace M 2 )+(trace(M ))2 o(1),
as n → ∞, for every symmetric matrix M ∈ Rmn ×mn , then we have
0
0 2
2
E[|(1, x01,n )Rn0 Ω−1
n Rn (1, x1,n ) | ] = O(kn ).
(v) Suppose that Σn is positive definite and that supkwk=1 E[|w0 z1,n |8 ] = O(1)
0
and (E[|z1,n
P z1,n |4 ])1/4 = O(kP k2F ), as n → ∞, for every projection matrix P in Rmn , and partition Rn = [t1 , T1 ] with t1 ∈ Rkn . If for every n ∈ N either one of (a) µn = 0, or (b) rank T1 = kn holds, then
0
0 4
4
E[|(1, x01,n )Rn0 Ω−1
n Rn (1, x1,n ) | ] = O(kn ).
Proof. For ease of notation we will drop the subscript n whenever there is no risk
of confusion. A simple calculation involving the elementary inequality |a + b|` ≤
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2631
2`−1 (|a|` + |b|` ) and the notation un = u = (u0 , u0−1 )0 , with u−1 ∈ Rpn , yields
E[|u0 (1, x01 )0 |` ] = E[|u0 (1, µ0 )0 + u0 (0, z10 Γ0 )0 |` ]
≤ 2`−1 E |u0 (1, µ0 )0 |` + |u0−1 Γz1 |`
= 2`−1 |u0 (1, µ0 )0 (1, µ0 )u|`/2 + |u0−1 Σu−1 |`/2 E |w0 z1 |` ,
where w = Γ0 u−1 /kΓ0 u−1 k, if kΓ0 u−1 k > 0 and w = 0, else. In the sum
u0 (1, µ0 )0 (1, µ0 )u + u0−1 Σu−1 = u0 Sn u both summands are non-negative and
thus both summands are bounded by u0 Sn u. Therefore, the upper bound in the
previous display is itself bounded by a constant multiple of |u0 Sn u|`/2 . This was
the claim of part (i).
For part (ii), first note that because the distribution of Σ̂n does not depend on
µ, we may assume that µ = 0, without loss of generality. By the same argument
as above but with µ = 0, u0 = 0 and u−1 is either vn,1 or vn,2 , we see that
0
0
E[|vn,s
x1 |4 ] = O(|vn,s
Σvn,s |2 ), s = 1, 2. Now
√ 0
0
0
Σ̂n vn,2 ] = n Var[vn,1
X 0 Xvn,2 /n − vn,1
X 0 ιι0 Xvn,2 /n2 ]
Var[ nvn,1
n
n
X
X
1
1
0
0
0
0
≤ 2n 2
Var[vn,1
xi vn,2
xi ] + 4 Var
vn,1
xi vn,2
xj
n i=1
n
i,j=1
q
0 x |4 ]E[|v 0 x |4 ] + 2
≤ 2 E[|vn,1
1
n,2 1
n3
n
X
0
0
0
0
E[vn,1
xi vn,2
xj vn,1
xk vn,2
xl ]
i,j,k,l=1
2
0
0
0
0
x1 |2 |vn,2
x1 |2 ]
= O(vn,1
Σvn,1 vn,2
Σvn,2 ) + 2 E[|vn,1
n
4 X
0
0
0
0
+ 3
E[vn,1
xi vn,2
xi ]E[vn,1
xj vn,2
xj ]
n
i6=j
2 X
0
0
E[|vn,1
xi |2 ]E[|vn,2
xj |2 ],
+ 3
n
i6=j
0
0
0
0
Σvn,s and
x1 |2 ] = vn,s
which is of order O(vn,1
Σvn,1 vn,2
Σvn,2 ) because E[|vn,s
q
0
0
0 x |2 ]E[|v 0 x |2 ].
E[vn,1
x1 vn,2
x1 ] ≤ E[|vn,1
1
n,2 1
For parts (iii), (iv) and (v) we make the following preliminary considerations.
First, note that in all three of these statements Σ is assumed to be positive
definite and thus Ω is regular. Abbreviate µ̄0 := (1, µ0 )R0 Ω−1/2 and
0 0
ΣW = R
R0 .
0 Σ
Since ΣW = Ω − R(1, µ0 )0 (1, µ0 )R0 we have
Ω−1/2 ΣW Ω−1/2 = Ikn − µ̄µ̄0 = A
1 − kµ̄k2
0
0
Ikn −1
A0 ,
(C.1)
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2632
for some orthogonal matrix A whose first column is µ̄/kµ̄k if kµ̄k > 0, and
A = Ikn if µ̄ = 0. Here, quantities of dimension kn − 1 have to be removed
in case kn = 1. The matrix Ω−1/2 ΣW Ω−1/2 in the previous display is positive
semidefinite, which means that 0 ≤ kµ̄k ≤ 1. For later use, we partition the
matrix B := Ω−1/2 A as B = [b1 , B1 ] where b1 ∈ Rkn and note that B satisfies
1 − kµ̄k2
0
B 0 ΣW B =
,
(C.2)
0
Ikn −1
and Ikn = B 0 ΩB = B 0 ΣW B + B 0 R(1, µ0 )0 (1, µ0 )R0 B, which entails that
kµ̄k2 0
0
0 0
0
0
B R(1, µ ) (1, µ )R B =
.
0
0
(C.3)
This finishes the preliminary considerations.
Now, for the proof of part (iii), write the quantity of interest as
Ω−1/2 (W 0 W/n)Ω−1/2 − Ikn
= A0 Ω−1/2 (W 0 W/n)Ω−1/2 A − A0 A
1 − kµ̄k2
0
= kB 0 (W 0 W/n)B − Ikn kS ≤ B 0 Σ̂W B −
0
Ikn −1 S
2
kµ̄k 0
+ B 0 R(1, µ̂0 )0 (1, µ̂0 )R0 B −
0
0 S
0
2
0
b1 Σ̂W b1 − (1 − kµ̄k )
b1 Σ̂W B1
=
B10 Σ̂W b1
B10 Σ̂W B1 − Ikn −1 S
0
b1 R(1, µ̂0 )0 (1, µ̂0 )R0 b1 − kµ̄k2 b01 R(1, µ̂0 )0 (1, µ̂0 )R0 B1
+
.
B10 R(1, µ̂0 )0 (1, µ̂0 )R0 b1
B10 R(1, µ̂0 )0 (1, µ̂0 )R0 B1 S
where
Σ̂W
0
=R
0
S
0
R0
Σ̂n
S
(C.4)
(C.5)
and µ̂ = X 0 ι/n.
For a partitioned matrix as above we have
2
c11 c012
c11 w1 + c12 w−1
= sup
c21 C22 S kwk=1 c21 w1 + C22 w−1
2
≤ (|c11 | + kc12 k)2 + (kc21 k + kC22 kS )2 .
Therefore, it suffices to show that the norms of the respective blocks are OP (1),
if kn = O(n), and converge to zero in probability, if kn /n → 0.
We begin with the terms involving µ̂ in (C.5). First,
E [b01 R(1, µ̂0 )0 (1, µ̂0 )R0 b1 ] − kµ̄k2
0
0
1
0
1
= b1 R
+
0 Σ/n
µ
= (1 − kµ̄k2 )/n −−−−→ 0,
n→∞
0
µ
R0 b1 − kµ̄k2
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2633
in view of E[µ̂µ̂0 ] = Σ/n + µµ0 , (C.2) and (C.3). Moreover, the variance satisfies
n
h
i
X
1
2
Var (b01 R(1, µ̂0 )0 ) = 4 Var
b01 R(1, x0i )0 b01 R(1, x0j )0
n
i,j=1
n
X
1
= 4 Var
b01 R[(1, x0i )0 (1, x0j ) − (1, µ0 )0 (1, µ0 )]R0 b1
n
i6=j
#
n
X
0
0 0
0
0
+
b1 R[(1, xi ) (1, xi ) − S]R b1
i=1
n
n
2 XX
E[b01 R[(1, x0i )0 (1, x0j ) − (1, µ0 )0 (1, µ0 )]R0 b1 ×
≤ 4
n
i6=j r6=s
+
n
X
b01 R[(1, x0r )0 (1, x0s ) − (1, µ0 )0 (1, µ0 )]R0 b1 ]
!
E[(b01 R(1, x0i )0 (1, x0i )R0 b1 )2 ] .
i=1
To work out the combinatorics of the quadruple sum above, abbreviate Fij =
b01 R[(1, x0i )0 (1, x0j ) − (1, µ0 )0 (1, µ0 )]R0 b1 and note that E[Fij ] = 0 if i 6= j and
E[Fij Frs ] = 0 if all four indices are distinct. Moreover, there are only O(n3 )
summands in which not all four indices are distinct, i.e., there are only O(n3 )
non-zero summands. Moreover, the non-zero summands can always be bounded
by
q
2 ] = E[F 2 ] = Var[F ] = Var[b0 R(1, x0 )0 (1, x0 )R0 b ]
|E[Fij Frs ]| ≤ E[Fij2 ]E[Frs
ij
1
ij
1
i
j
≤ E[((1, x0i )R0 b1 )2 ((1, x0j )R0 b1 )2 ] = (E[((1, x01 )R0 b1 )2 ])2 ,
if i 6= j and r 6= s. Since E[(b01 R(1, x01 )0 )2 ] = b01 RSR0 b1 = b01 Ωb1 = 1, by
definition of B, we see that the quadruple sum in the second-to-last display is
of order O(n3 ). The remaining sum in the same display is of order O(n), since
E[(b01 R(1, x01 )0 )4 ] = O(1), by part (i) and the assumption supkwk=1 E[|w0 z1 |4 ] =
O(1). Thus, we have shown that b01 R(1, µ̂0 )0 (1, µ̂0 )R0 b1 −kµ̄k2 → 0, in probability.
Next, consider kb01 R(1, µ̂0 )0 (1, µ̂0 )R0 B1 k2 ≤ |b01 R(1, µ̂0 )0 |2 kB10 R(1, µ̂0 )0 k2 . The
first factor in the upper bound was just shown to be OP (1). For the second factor
note that E[kB10 R(1, µ̂0 )0 k2 ] = trace((B10 ΣW B1 /n + B10 R(1, µ0 )0 (1, µ0 )R0 B1 )) =
(kn −1)/n, by (C.2) and (C.3). Since kB10 R(1, µ̂0 )0 (1, µ̂0 )R0 B1 kS = kB10 R(1, µ̂0 )0 k2 ,
we see that the spectral norm in (C.5) is OP (1) if kn = O(n), and converges to
zero in probability, if kn /n → 0.
For the spectral norm in (C.4), we may restrict to µ = 0. First, write
R = [t1 , T1 ] with t1 ∈ Rkn and use (C.2) to see that E[b01 Σ̂W b1 ] − (1 − kµ̄k2 ) =
b01 ΣW b1 (n − 1)/n − (1 − kµ̄k2 ) = (1 − kµ̄k2 )/n → 0, whereas Var[b01 Σ̂W b1 ] =
part (ii) with vn =√
vn,1 = vn,2 =
Var[b01 T1 Σ̂n T10 b1 ] → 0, in view of the result in √
T10 b1 /n1/4 , which satisfies vn0 Σvn = b01 ΣW b1 / n = (1 − kµ̄k2 )/ n → 0. For
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2634
the off-diagonal block B10 Σ̂W b1 , note that it has mean zero in view of (C.2).
Pkn −1
Pkn −1
Therefore, E[kB10 Σ̂W b1 k2 ] = j=1
E[(e0j B10 Σ̂W b1 )2 ] = j=1
Var[e0j B10 Σ̂W b1 ],
where e1 , . . . , ekn −1 is the standard basis in Rkn −1 . Now, Var[e0j B10 Σ̂W b1 ] =
Var[e0j B10 T1 Σ̂n T10 b1 ], and part (ii) applies with vn,1 = T10 B1 ej /n1/4 and vn,2 =
√
√
0
0
Σvn,2 =
T10 b1 /n1/4 , which satisfy vn,1
Σvn,1 = e0j B10 ΣW B1 ej / n = 1/ n and vn,2
√
√
0
2
0
b1 ΣW b1 / n = (1 − kµ̄k )/ n, in view of (C.2). Therefore, E[kB1 Σ̂W b1 k2 ] =
Pkn −1
= O(kn /n). Hence, the only remaining term is kB10 Σ̂W B1 −
j=1 O(1/n)P
n
1
Ikn −1 kS ≤ k n i=1 B10 T1 xi x0i T10 B1 − Ikn −1 kS + kB10 T1 µ̂µ̂0 T10 B1 kS . For the second term in the upper bound, one easily finds its expected value to be (kn −1)/n,
as in the previous paragraph. For the spectral norm of the remaining covariance
term we verify the strong regularity (SR) condition of Srivastava and Vershynin
[23, Theorem 1.1] for the random (kn − 1)-vectors x̄i = B10 T1 xi = B10 T1 Γzi .
First, note that the x̄i are independent and isotropic, since µ = 0 and E[x̄i x̄0i ] =
B10 T1 ΣT10 B1 = B10 ΣW B1 = Ikn −1 . Fix a projection matrix P in Rkn −1 and
note that Γ0 T10 B1 P B10 T1 Γ is a projection matrix in Rmn of the same rank as
P . Since the zi satisfy Assumption (A1).(c) and kP x̄1 k2 = kP B10 T1 Γz1 k2 =
z10 Γ0 T10 B1 P B10 T1 Γz1 = kΓ0 T10 B1 P B10 T1 Γz1 k2 , we see that the (SR) condition
holds for x̄1 and with the same constants c, C as in (A1).(c).
Pn Therefore, Corollary 1.4 of Srivastava and Vershynin [23] shows that k n1 i=1 B10 T1 xi x0i T10 B1 −
Ikn −1 kS is OP (1) if kn = O(n), and converges to zero, in probability, if kn /n → 0.
This finishes part (iii).
For the proof of parts (iv) and (v), take ` ∈ N and consider the elementary
bound
E[|(1, x01 )R0 Ω−1 R(1, x01 )0 |` ] = E (1, µ0 )R0 Ω−1 R(1, µ0 )0
i
`
+2(1, µ0 )R0 Ω−1 R(0, z10 Γ0 )0 + (0, z10 Γ0 )R0 Ω−1 R(0, z10 Γ0 )0
`
(C.6)
≤ 2`−1 kµ̄k2` + 2`−1 2` E µ̄0 Ω−1/2 R(0, z10 Γ0 )0
io
h
`
.
+E (0, z10 Γ0 )R0 Ω−1 R(0, z10 Γ0 )0
Partition R = [t1 , T1 ] as above and abbreviate M = Γ0 T10 Ω−1 T1 Γ, so that
the expectation on the last line of the previous display can be written as
E[|z10 M z1 |` ]. Now, if ` = 2, this
as E[|z10 M z1 |2 ] = Var[z10 M z1 ]+
can be evaluated
2
2
2
0
(E[z1 M z1 ]) = O(trace M ) + (trace(M )) o(1) + (trace(M ))2 under the as
sumption of part (iv). Since trace(M ) = trace Ω−1/2 ΣW Ω−1/2 = kn −kµ̄k2 and
trace M 2 = kn −1+(1−kµ̄k2 ), by (C.1), we see that E[|z10 M z1 |2 ] = O(kn2 ). Furthermore, E[|µ̄0 Ω−1/2 R(0, z10 Γ0 )|2 ] = µ̄0 Ω−1/2 ΣW Ω−1/2 µ̄ = kµ̄k2 − kµ̄k4 ≤ 1/4,
which finishes part (iv).
For part (v) we begin with the expectation in (C.6) with ` = 4, which can be
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2635
written as
E[|µ̄0 Ω−1/2 T1 Γz1 |4 ] ≤ kµ̄0 Ω−1/2 T1 Γk4 sup E[|w0 z1 |4 ]
kwk=1
0
= O(|µ̄ Ω
−1/2
ΣW Ω−1/2 µ̄|2 ) = O(1).
For E[|z10 M z1 |4 ], we begin with case (a) µ = 0. Then
−1
1 0
1 0
1
0
E[|z10 M z1 |4 ] = E (0, z10 )
R
R
R0
R
0 Γ0
0 Σ
0
4
0
(0, z10 )0 ,
Γ
and we denote the matrix corresponding to the quadratic form in the vector
(0, z10 )0 on the right-hand-side of this display by P . Clearly, P is a projection
matrix which we partition as
p11 p021
P =
,
p21 P22
with p11 ∈ [0, 1]. Exploiting the idempotency and symmetry of P , one can
show that the generalized Schur complement of p11 in P , i.e, the matrix P22 −
p21 p†11 p021 , is again a projection matrix [cf. 4, Corollary 2.1], where p†11 = p−1
11 ,
†
† 0
14
if p11 6= 0, and p11 = 0, else. Moreover, since |kP22 kS − kp21 p11 p21 kS | ≤
kP22 − p21 p†11 p021 kS ≤ 1 and kP22 kS ≤ kP kS = 1, we see that kp21 (p†11 )1/2 k2 =
kp21 p†11 p021 kS ≤ 2 and that the Frobenius norm
of the generalized
Schur complement satisfies kP22 − p21 p†11 p021 k2F = trace (P22 − p21 p†11 p021 ) ≤ trace(P22 ) ≤
trace(P ) = kn . Therefore, using our assumptions, we calculate
E[|z10 M z1 |4 ] = E[|z10 P22 z1 |4 ]
≤ 23 E[|z10 (P22 − p21 p†11 p021 )z1 |4 ] + E[|z10 p21 p†11 p021 z1 |4 ]
≤ 23 O(kn4 ) + E[|(p†11 )1/2 p021 z1 |8 ] = O(kn4 ).
Finally, in the case (b), where rank T1 = kn , the matrix T1 ΣT10 in the representation Ω = RSR0 = T1 ΣT10 + R(1, µ0 )0 (1, µ0 )R0 , is regular and thus we can invert
Ω by the Sherman-Morrison formula to get
M = Γ0 T10 Ω−1 T1 Γ
= Γ0 T10 (T1 ΣT10 )−1 T1 Γ −
Γ0 T10 (T1 ΣT10 )−1 R(1, µ0 )0 (1, µ0 )0 R0 (T1 ΣT10 )−1 T1 Γ
.
1 + (1, µ0 )R0 (T1 ΣT10 )−1 R(1, µ0 )0
Therefore, we make use of the abbreviations P = Γ0 T10 (T1 ΣT10 )−1 T1 Γ and v =
Γ0 T10 (T1 ΣT10 )−1 R(1, µ0 )0 to bound the fourth moment of the quadratic form
14 Baksalary, Baksalary and Szulc [4] actually prove a more general result. The special case
we are interested in here can also be easily derived by direct calculation.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2636
z10 M z1 by
E[|z10 M z1 |4 ]
≤2
3
≤ 23
4 #!
|v 0 z1 |2
+E
1 + kvk2
!
4
kvk2
8
0
8
O(kP kF ) +
sup E[|w z1 | ] .
1 + kvk2
kwk=1
E[|z10 P z1 |4 ]
"
Since kP k8F = (trace(P ))4 = kn4 , the upper bound is of order O(kn4 ), which
finishes the proof of part (v).
Lemma C.4. Let 1 ≤ q ≤ p be positive integers. If T is a (p + 1) × (p + 1)
orthogonal matrix that is partitioned as
t T0
T = 0
,
t1 T1
where t0 ∈ Rq , t1 ∈ Rp+1−q , T0 ∈ Rq×p and T1 ∈ R(p+1−q)×p , then kt0 k > 0 if
and only if, rank T1 = p + 1 − q.
Proof. By orthogonality,
Ip+1
0
t t + T0 T00
= T T = 0 00
t1 t0 + T1 T00
0
t0 t01 + T0 T10
,
t1 t01 + T1 T10
and kt0 k2 + kt1 k2 = 1. Hence,
T1 T10 = Ip+1−q − t1 t01
has eigenvalues 1, with multiplicity p − q, and a single eigenvalue 1 − kt1 k2 =
kt0 k2 , which is strictly positive if and only if, rank T1 = p + 1 − q.
Lemma C.5. If the n × p random matrix X has i.i.d. rows following the
N (0, Σ)-distribution with positive definite Σ, v ∈ Rp , v 6= 0 and T ∈ Rq×p
0 −1
v ∼ v 0 (T Σ−1 T 0 )−1 v χ2n−1−(p−q) /n,
has orthonormal rows, then v 0 (T Σ̂−1
n T )
where Σ̂n = X 0 (In − Pι )X/n is the sample covariance matrix.
Proof. It is well known that nΣ̂n ∼ Wp (Σ, n−1) has a Wishart distribution with
scale matrix Σ and n−1 degrees of freedom [e.g., 19, Theorem 3.4.4.(c)]. If q = p,
0 −1
then T is orthogonal and nv 0 (T Σ̂−1
v = v 0 T nΣ̂n T 0 v ∼ v 0 T ΣT 0 v χ2n−1 =
n T )
0
−1 0 −1
2
v (T Σ T ) v χn−1 [cf. 19, Theorem 3.4.2]. So assume that q < p. Let S ∈
R(p−q)×p be such that R = [S 0 , T 0 ]0 is an orthogonal matrix. Then, by block
matrix inversion of
1
S Σ̂n S 0 S Σ̂n T 0
RΣ̂n R0 =
∼ Wp (RΣR0 , n − 1),
0
0
n
T Σ̂n S T Σ̂n T
0 −1
we see that the matrix (T Σ̂−1
= ([0, Iq ](RΣ̂n R0 )−1 [0, Iq ]0 )−1 = T Σ̂n T 0 −
n T )
0
0 −1
0
T Σ̂n S (S Σ̂n S ) S Σ̂n T is the Schur complement of S Σ̂n S 0 in RΣ̂n R0 , which
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2637
follows the Wq (Ω22·1 , n − 1 − (p − q))-distribution divided by n, where Ω22·1 =
0 −1
T ΣT 0 −T ΣS 0 (SΣS 0 )−1 SΣT 0 [cf. 19, Theorem 3.4.6.(a)]. Therefore, nv 0 (T Σ̂−1
v∼
n T )
0
2
−1 0 −1
v Ω22·1 v χn−1−(p−q) , and Ω22·1 = (T Σ T ) .
Lemma C.6. Let µ ∈ Rp and Σ be a symmetric, positive definite p × p matrix.
Let T = [R00 , R10 ]0 be a (p+1)×(p+1) orthogonal matrix such that R0 ∈ Rq×(p+1)
and set
1
µ0
Ω00 Ω01
R0 SR00 R0 SR10
0
S=
=
, Ω = T ST =
,
µ Σ + µµ0
R1 SR00 R1 SR10
Ω10 Ω11
0 0
R0
Σ00 Σ01
ΣT = T
T0 =
[0, Ip ]0 Σ[0, Ip ] R00 R10 =
.
0 Σ
R1
Σ10 Σ11
If Σ11 := R1 [0, Ip ]0 Σ[0, Ip ]R10 is regular, then the Schur complement of Ω11 in
Ω is related to the Schur complement of Σ11 in ΣT by Ω00 − Ω01 Ω−1
11 Ω10 =
−1
0
0 0
Σ00 − Σ01 Σ−1
11 Σ10 + µ̃µ̃ /(1 + ν), where µ̃ = (R0 − Σ01 Σ11 R1 )(1, µ ) and ν =
0 0
(1, µ0 )R10 Σ−1
11 R1 (1, µ ) .
Proof. First note that Ωij = Σij + Ri (1, µ0 )0 (1, µ0 )Rj , for i, j ∈ {0, 1}. Abbreviate µ̃i = Ri (1, µ0 )0 , for i = 0, 1 and ν = µ̃01 Σ−1
11 µ̃1 and use the Sherman-Morrison
formula to write
0
0 −1
Ω00 − Ω01 Ω−1
(Σ10 + µ̃1 µ̃00 )
11 Ω10 = Σ00 + µ̃0 µ̃0 − (Σ01 + µ̃0 µ̃1 )(Σ11 + µ̃1 µ̃1 )
0 −1
Σ−1
11 µ̃1 µ̃1 Σ11
(Σ10 + µ̃1 µ̃00 )
= Σ00 + µ̃0 µ̃00 − (Σ01 + µ̃0 µ̃01 ) Σ−1
−
11
1+ν
= Σ00 − Σ01 Σ−1
11 Σ10
−1
0
0 −1
0
+ µ̃0 µ̃00 − µ̃0 µ̃01 Σ−1
11 Σ10 − Σ01 Σ11 µ̃1 µ̃0 − µ̃0 µ̃1 Σ11 µ̃1 µ̃0
0 −1
Σ−1 µ̃1 µ̃01 Σ−1
Σ−1 µ̃1 µ̃01 Σ−1
Σ−1
11
11
11 µ̃1 µ̃1 Σ11
Σ10 + µ̃0 µ̃01 11
Σ10 + Σ01 11
µ̃1 µ̃00
1+ν
1+ν
1+ν
Σ−1 µ̃1 µ̃01 Σ−1
11
µ̃1 µ̃00
+ µ̃0 µ̃01 11
1+ν
ν2 0
0
µ̃
= Σ00 − Σ01 Σ−1
Σ
+
µ̃
(1
−
ν)µ̃
+
µ̃
10
0
0
0
11
1+ν 0
ν
ν
−1
0 −1
+ µ̃0
− 1 µ̃1 Σ11 Σ10 + Σ01 Σ11 µ̃1
− 1 µ̃00
1+ν
1+ν
+ Σ01
0 −1
Σ−1
11 µ̃1 µ̃1 Σ11
Σ10
1+ν
= Σ00 − Σ01 Σ−1
11 Σ10
+ Σ01
−1
−1
0
0 −1
+ µ̃0 µ̃00 − µ̃0 µ̃01 Σ−1
11 Σ10 − Σ01 Σ11 µ̃1 µ̃0 + Σ01 Σ11 µ̃1 µ̃1 Σ11 Σ10 /(1 + ν)
= Σ00 − Σ01 Σ−1
11 Σ10 +
where µ̃ = µ̃0 − Σ01 Σ−1
11 µ̃1 .
µ̃µ̃0
,
1+ν
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2638
Lemma C.7. Let k, n be positive integers such that k < n − 1. If X is a
random n × k matrix whose rows are i.i.d. distributed according to N (µ, Σ),
where µ ∈ Rk and Σ is positive definite, then
1 0
ι (In − PX )ι
n
∼
ξ
,
ξ+ζ
where ξ and ζ are independent and distributed according to ξ ∼ χ2n−k and ζ ∼
χ2k (λn ), with non-centrality parameter λn = nµ0 Σ−1 µ.
Remark. Lemma C.7 is a slight variation of Lemma A.2 in Leeb [17].
Proof. Note that PX = PX̄ , where X̄ = XΣ−1/2 has i.i.d. rows following the
ˆ = X̄ 0 (I − P )X̄/n =
ˆn = X̄ 0 ι/n and Σ̄
N (Σ−1/2 µ, In ) distribution. Writing µ̄
n
n
ι
0
0
ˆ
ˆ
X̄ X̄/n − µ̄n µ̄n , for the sample mean and sample covariance matrix of the transformed data, we have, at least on an event of probability one,
1 0
ι (In − PX )ι
n
=
=
=
ˆ + µ̄
ˆ0n (Σ̄
ˆn µ̄
ˆ0n )−1 µ̄
ˆn
1 − µ̄
n
"
#
ˆ −1 µ̄
ˆ n )2
ˆ0n Σ̄
(µ̄
n
0
−1
ˆ
ˆn Σ̄n µ̄
ˆ−
1 − µ̄
ˆ −1 µ̄
ˆ0n Σ̄
1 + µ̄
n ˆn
1
.
ˆ −1 µ̄
0
ˆ Σ̄
ˆ
1 + µ̄
n
n
n
ˆ has a standard Wishart distribution with n−1 degrees of freedom and
Since nΣ̄
n
ˆn , we get from Mardia, Kent and Bibby [19, Theorem 3.4.7]
is independent of µ̄
ˆ −1 µ̄
ˆ )−1 (µ̄
ˆn = nkµ̄
ˆn k2 (µ̄
ˆn /kµ̄
ˆn k)0 (nΣ̄
ˆn /kµ̄
ˆn k)
ˆn , the quantity µ̄
ˆ0n Σ̄
that, conditional on µ̄
n
n
2
2
ˆ
has the same distribution as nkµ̄n k /ξ, where ξ ∼ χn−k is independent
of
√
ˆn . The proof is finished upon noting that ζ := nkµ̄
ˆn k2 = kX̄ 0 ι/ nk2 ∼
µ̄
χ2k (nµ0 Σ−1 µ) and that 1/(1 + ζ/ξ) = ξ/(ξ + ζ).
Lemma C.8. Let qn ≤ pn + 1 < n be positive integer sequences such that
lim supn pn /n < 1, and let ∆n be a non-negative real sequence of order o(qn /n).
Moreover, let sn and bn be as in Theorem 2.1. Then a non-central F -distributed
random variable Fqn ,n−pn −1 (λn ) with qn and n − pn − 1 degrees of freedom and
non-centrality parameter λn = ∆n (n − pn − 1 + qn ) satisfies
√
s−1/2
(Fqn ,n−pn −1 (λn ) − 1) − n∆n bn
n
(
N (0, 1),
if qn → ∞,
w
p
−−−−→
−1/2 2
n→∞
(2q)
χq − q/2, if qn = q, ∀n ∈ N.
Remark. This result is elementary and follows from basic properties of the
non-central χ2 distribution [cf. 16, Chapter 29.5]. For the convenience of the
reader we include a proof nonetheless. The proof also nicely resembles the main
steps of the much more involved argument needed to treat the non-Gaussian
cases of Theorem 2.1 and Theorem 3.1.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2639
Proof. Let Y1 , . . . , Yqn , X1 , . . . , Xn−pn −1 be i.i.d. standard normal and let µ =
(µ1 , . . . , µqn )0 ∈ Rqn such that µ0 µ = λn . Then Fqn ,n−pn −1 (λn ) can be represented as
Pqn
(Yi + µi )2 /qn
.
Fqn ,n−pn −1 (λn ) ∼ Pn−pni=1
−1
Xj2 /(n − pn − 1)
j=1
Therefore,
√
s−1/2
(Fqn ,n−pn −1 (λn ) − 1) − n∆n bn
n
Pqn
Pn−pn −1 2
Pqn
0
1
1
2
Yi µi + µqnµ √
Xj + q2n i=1
i=1 Yi − n−pn −1
j=1
−1/2 qn
∼ sn
− n∆n bn
Pn−pn −1 2
1
Xj
j=1
n−pn −1
n−p
qn
n −1
X
X
1
1
= s−1/2
Xj2 1 + OP ((n − pn )−1/2 )
Yi2 −
n
qn i=1
n − pn − 1 j=1
+
+
−1/2
1 + OP ((n − pn )
λn
√
qn sn
)
2
√
qn
X
Yi µi
qn sn i=1
√
1 + OP ((n − pn )−1/2 ) − n∆n bn .
Pqn
2
For the mixed term qn √
i=1 Yi µi we note that it has mean zero and variance
sn
equal to 4(qn2 sn )−1 λn = 4(qn sn )−1 ∆n (n/qn )(n − pn − 1 + qn )/n = o(1), since
∆n = o(qn /n), by assumption, and qn sn = 2(1 + qn /(n − pn − 1)) ≥ 2. So
the mixed term converges to zero
√ in probability. For the non-centrality term
λ√n
−1/2
(1
+
O
((n
−
p
)
))
−
n∆n bn we first note that
P
n
qn sn
√
n∆n bn =
λn
n∆n n − pn − 1 + qn
=
√
√ ,
qn sn
n
qn sn
so that
√
λn
λn
1 + OP ((n − pn )−1/2 ) − n∆n bn =
√
√ OP ((n − pn )−1/2 )
qn sn
qn sn
p
n
n − pn − 1 + qn
=
∆n
(qn sn )−1/2 OP ( qn /(n − pn )) = oP (1),
qn
n
because qn /(n − pn ) ≤ n/(n − pn ) = (1 − pn /n)−1 is bounded in view of our
assumption that lim supn pn /n < 1. The same assumption also guarantees that
n − pn → ∞ as n → ∞, so it remains to establish the appropriate convergence
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2640
of
qn
n−p
n −1
X
X
1
1
s−1/2
Y2−
Xj2
n
qn i=1 i
n − pn − 1 j=1
qn
n−p
n −1
X
X
1
1
= s−1/2
(Y 2 − 1) −
(Xj2 − 1)
n
qn i=1 i
n − pn − 1 j=1
r
qn
n − pn − 1
1 X 2
√
=
(Y − 1)
n − pn − 1 + qn 2qn i=1 i
r
n−p
n −1
X
qn
1
p
−
(Xj2 − 1).
n − pn − 1 + qn 2(n − pn − 1) j=1
(C.7)
(C.8)
First, we consider the case where qn → ∞. Define the vectors
q
Pqn
!
2
n−pn −1
√1
(Y
−
1)
i=1 i
2qn
n−p −1+q
P
n−pn −1
Zn := √ 1
and vn := q n q n .
(Xj2 − 1)
n
j=1
−
2(n−pn −1)
n−pn −1+qn
By independence and the CLT, Zn converges weakly to a bivariate standard
normal distribution as n → ∞ and kvn k2 = 1. Thus, vn0 Zn → N (0, 1), weakly,
as n → ∞, by compactness of the unit circle. If qn = q ∈ N does not depend
on n, then the expression in (C.8) converges to zero in probability
and the
p
expression in (C.7) is distributed as (1 + o(1))[(2q)−1/2 χ2q − q/2].
References
[1] Bai, Z. D. and Saranadasa, H. (1996). Effect of high dimension: By an
example of a two sample problem. Statist. Sinica 6 311–329.
[2] Bai, Z. D. and Yin, Y. Q. (1993). Limit of the smallest eigenvalue of a
large dimensional sample covariance matrix. Ann. Probab. 21 1275–1294.
[3] Bai, Z. and Zhou, W. (2008). Large sample covariance matrices without
independence structures in columns. Statist. Sinica 18 425–442.
[4] Baksalary, J. K., Baksalary, O. M. and Szulc, T. (2004). Properties
of Schur complements in partitioned idempotent matrices. Linear Algebra
Appl. 379 303–318.
[5] Bhansali, R. J., Giraitis, L. and Kokoszka, P. S. (2007). Convergence
of quadratic forms with nonvanishing diagonal. Statist. Probab. Lett. 77
726–734.
[6] Bhatia, R. (1997). Matrix Analysis. Graduate Texts in Mathematics 169.
Springer, New York.
[7] Bryc, W. (1995). The Normal Distribution: Characterizations with Applications. Springer, New York.
[8] Cambanis, S., Huang, S. and Simons, G. (1981). On the theory of elliptically contoured distributions. J. Multivariate Anal. 11 368–385.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2641
[9] Dümbgen, L. and Conte-Zerial, P. (2013). On low-dimensional projections of high-dimensional distributions. In From Probability to Statistics
and Back: High-Dimensional Models and Processes. A Festschrift in Honor
of Jon Wellner, (M. Banerjee, F. Bunea, J. Huang, V. Koltchinskii and
M. H. Maathius, eds.). IMS Collections 9 91–104. IMS, Beachwood, Ohio.
[10] Dvoretzky, A. (1972). Asymptotic Normality for Sums of Dependent
Random Variables. In Proc. 6th Berkeley Symp. Math. Statist. Probab. 2
513–535. University of California Press.
[11] Edelman, A. (1991). The distribution and moments of the smallest eigenvalue of a random matrix of Wishart type. Linear Algebra Appl. 159 55–80.
[12] Ghosh, B. K. (1973). Some monotonicity theorems for χ2 , F and t distributions with applications. J. Roy. Statist. Soc. Ser. B 35 480–492.
[13] Helland, I. S. (1982). Central Limit Theorems for Martingales with Discrete or Continuous Time. Scand. J. Statist. 9 79–94.
[14] Huber, N. and Leeb, H. (2013). Shrinkage estimators for prediction outof-sample: Conditional performance. Commun. Statist. - Theory Methods
42 1246–1264.
[15] Johnson, L. N., Kotz, S. and Balakrishnan, N. (1995a). Continuous
Univariate Distributions 1. Wiley, New York.
[16] Johnson, L. N., Kotz, S. and Balakrishnan, N. (1995b). Continuous
Univariate Distributions 2. Wiley, New York.
[17] Leeb, H. (2009). Conditional predictive inference post model selection.
Ann. Statist. 37 2838–2876.
[18] Lugosi, G. (2009). Concentration of measure inequalities. Lecture notes
available online.
[19] Mardia, K. V., Kent, J. T. and Bibby, J. M. (1995). Multivariate
Analysis. Academic Press, London.
[20] Portnoy, S. (1985). On the Central Limit Theorem in Rp when p → ∞.
Probab. Theory Relat. Fields 73 571–583.
[21] Portnoy, S. (1988). Asymptotic Behavior of Likelihood Methods for Exponential Families when the Number of Parameters Tends to Infinity. Ann.
Statist. 16 356–366.
[22] Rao, C. R. and Toutenburg, H. (1995). Linear Models: Least Squares
and Alternatives. Springer Series in Statistics. Springer, New York.
[23] Srivastava, N. and Vershynin, R. (2013). Covariance estimation for
distributions with 2 + ε moments. Ann. Probab. 41 3081–3111.
[24] Steinberger, L. and Leeb, H. (2014). On conditional moments of highdimensional random vectors given lower-dimensional projections. arXiv
preprint arXiv:1405.2183.
[25] R Development Core Team (2011). R: A Language and Environment
for Statistical Computing R Foundation for Statistical Computing, Vienna,
Austria ISBN 3-900051-07-0.
[26] van Aubel, A. and Gawronski, W. (2003). Analytic Properties of Noncentral Distributions. Appl. Math. Comput. 141 3–12.
[27] Wang, S. and Cui, H. (2013). Generalized F test for high dimensional
linear regression coefficients. J. Multivariate Anal. 117 134–149.
L. Steinberger/Dimensionality and multiplicity of hypotheses in the F -test
2642
[28] Whittle, P. (1960). Bounds for the Moments of Linear and Quadratic
Forms in Independent Variables. Theory Probab. Appl. 5 302–305.
[29] Zhong, P.-S. and Chen, S. X. (2011). Tests for high-dimensional regression coefficients with factorial designs. J. Amer. Statist. Assoc. 106
260–274.
| 10math.ST
|
arXiv:1504.04853v3 [math.AC] 25 May 2016
ON THE ASYMPTOTIC BEHAVIOR OF THE LINEARITY DEFECT
HOP D. NGUYEN AND THANH VU
Abstract. This work concerns the linearity defect of a module M over a noetherian local
ring R, introduced by Herzog and Iyengar in 2005, and denoted by ldR M . Roughly speaking,
ldR M is the homological degree beyond which the minimal free resolution of M is linear. In the
paper, it is proved that for any ideal I in a regular local ring R and for any finitely generated Rmodule M , each of the sequences (ldR (I n M ))n and (ldR (M/I n M ))n is eventually constant. The
first statement follows from a more general result about the eventual constancy of the sequence
(ldR Cn )n where C is a finitely generated graded module over a standard graded algebra over
R. The second statement follows from the first together with a result of Avramov on small
homomorphisms.
1. Introduction
Let (R, m) be a commutative noetherian local ring with maximal ideal m and residue field
k = R/m. Let I ⊆ m be an ideal of R. The study of the asymptotic behavior of, for instances,
graded components of a graded module, or powers of an ideal, is a classical topic in commutative
algebra. A primal example of good asymptotic behavior is that if I is m-primary, then the
Hilbert-Samuel function of I, given by n 7→ length of (R/I n ), is eventually a polynomial function
when n tends to infinity. To give another example, Brodmann [3] proved that for any I, the
sequence (depth(R/I n ))n is eventually constant for n large enough. Aside from qualitative
statements, for each invariant, it is also of interest to study the point when nice asymptotic
behaviors of powers occur and the nature of the limiting values at high enough powers. See, for
example, [5], [13] and the references therein for more details.
In this paper, we study the linearity defect introduced by Herzog and Iyengar [15] (see Section
2). One of the motivations for studying the linearity defect is the research on the linear part of
minimal free resolutions over the exterior algebras in [9]. Similarly to the projective dimension,
the finiteness of the linearity defect has strong consequences on the structure of a module: If
ldR M is finite, then the Poincaré series of M is rational with denominator depending only on
R (see [15, Proposition 1.8]). However, there remain many open questions on the finiteness of
the linearity defect; see [1], [6], [22] for more details.
The linearity defect was studied by many authors, see for example [1], [6], [15], [19], [21], [22],
[23]. Nevertheless, it is still an elusive invariant. The problem is highly non-trivial as to bound
efficiently the linearity defect even for familiar classes of ideals like monomial ideals. Beyond
componentwise linear ideals [12] (which have linearity defect zero), there are few interesting and
large enough classes of ideals whose linearity defect is known. In this paper, the above remarks
notwithstanding, we show that the linearity defect behaves in a pleasant way asymptotically.
2010 Mathematics Subject Classification. 13D02, 13D05, 13H99.
Key words and phrases. Powers of ideals; linearity defect; asymptotic behavior.
This work is partially supported by the NSF grant DMS-1103176.
1
2
HOP D. NGUYEN AND THANH VU
Denote by gl ld R = sup{ldR M : M is a finitely generated R-module} the so-called global linearity defect of R. For instances, regular rings (or more generally local rings which are both
Koszul and Golod, see [15, Corollary 6.2]) have finite global linearity defect. The main result of
this paper is
Theorem 1.1. Let (R, m) be a local ring such that gl ld R < ∞, e.g., R is regular. Let I ⊆ m
be an ideal of R and M a finitely generated R-module. Then for all n large enough, each of the
numbers ldR (I n M), ldR (I n M/I n+1 M), and ldR (M/I n M) is a constant depending only on I and
M.
The proof essentially depends on a characterization of linearity defect in terms of Tor due
to Şega [22], and the theory of Rees algebras. Theorem 1.1 can be divided into two parts:
the first concerns with the sequences (ldR (I n M))n , (ldR (I n M/I n+1 M))n and the second with
(ldR (M/I n M))n . The proof of the second part is obtained from the proof of the first and a
result of Avramov [2] on small homomorphisms. The first part is obtained from the following
general result. Below, recall that S is called a standard graded algebra over R if it is an N-graded
ring with S0 = R and S is generated over R by elements of S1 .
Theorem 1.2. Let (R, m) be a local ring such that gl ld R < ∞. Let S be a standard graded
algebra over R and C a finitely generated graded S-module. Then the sequence (ldR Cn )n is
eventually constant.
Theorem 1.2 is motivated by previous work of Herzog and Hibi [13, Theorem 1.1] on depth. In
Theorem 1.2, if we allow S to be generated by finitely many elements of positive degrees, then
(ldR Cn )n is quasiperiodic (namely, there exist an integer p ≥ 1 and numbers ℓ0 , . . . , ℓp−1 such
that for all n large enough, ldR Cn = ℓi if n is congruent to i modulo p, where 0 ≤ i ≤ p − 1). In
the end of Section 3, we turn to the asymptotic linearity defect of the integral closure I n and the
saturation Ien of I n . Theorem 1.2 is also suitable for studying the sequence (ldR I n )n , at least
over regular local rings. We give an example in which the sequence (ldR Ien )n is quasiperiodic but
not eventually constant. We do not know any example for which (ldR Ien )n is non-quasiperiodic.
Theorem 1.1 implies that the sequence (ldR I n )n is eventually constant. It remains mysterious
to us how to bound effectively the asymptotic value of the sequence (ldR I n )n . A rare result
in this direction is [14, Theorem 2.4]. There the authors establish a necessary and sufficient
condition for all the powers of a polynomial ideal to have linearity defec zero, using the theory
of d-sequences [17]. It would be interesting to study possible generalizations and analogues of
this result.
We assume that the reader is versed with the standard concepts and facts of commutative
algebra, which may be found in [4], [8].
2. Linearity defect
Let (R, m) be a commutative noetherian local ring with residue field k. Let I be a proper
M IjM
ideal of R. Let M be a finitely generated R-module. We call grI M =
the associated
I j+1 M
j≥0
graded module of M with respect to I.
Let F be the minimal free resolution of M:
F : ···
/
Fi
/
Fi−1
/
···
/
F1
/
F0
/
0.
ON THE ASYMPTOTIC BEHAVIOR OF THE LINEARITY DEFECT
3
Since ∂(F ) ⊆ mF the (by homological degree) graded submodule
F iF : · · ·
/
Fi+1
/
Fi
/
mFi−1
/
···
/
mi−j Fj
/
···
of F is stable under the differential; said otherwise, F i F is a subcomplex of F .
We define the so-called linear part of F by the formula
∞
M
F iF
R
.
lin F :=
F i+1 F
i=0
Note that linR F is a complex of graded modules over grm R, and (linR F )i = (grm Fi )(−i) for
every i ≥ 0. Following [15], the linearity defect of M is defined by
ldR M := sup{i : Hi (linR F ) 6= 0}.
If M ∼
= 0, we set ldR M = 0. This convention guarantees that the maximal ideal (0) of the field
k has linearity defect zero.
Remark 2.1. (1) The definition of the linear part works also for standard graded algebras. Let
(R, m) be a standard graded algebra over a field k, with the graded maximal ideal m, and M a
finitely generated graded R-module. Let F be the minimal graded free resolution of M over R.
Then the above construction of the linear part goes through for F .
In this case, the linear part linR F has a simple interpretation. For each i ≥ 1, apply the
following rule to all entries in the matrix representing the map Fi → Fi−1 : keep it if it is a linear
form, and replace it by 0 otherwise. Then the resulting complex is linR F .
(2) Although our results deal only with local rings, their counterparts for standard graded
algebras are also valid and can be established by the same methods.
3. Asymptotic behavior of the linearity defect
Let (R, m) be a noetherian local ring such that gl ld R < ∞. We divide Theorem 1.1 into two
parts: the first concerns with the stability of ldR (I n M) and ldR (I n M/I n+1 M) for large n, and
the second with the stability of ldR (M/I n M). The proof of the second part follows from the
first part and a result due to Avramov on small homomorphisms. The first part follows from
Theorem 1.2.
3.1. The first part of Theorem 1.1. Before deducing the first part of Theorem 1.1 from
Theorem 1.2, recall that Rees(I) = R ⊕ I ⊕ I 2 ⊕ · · · denote the Rees algebra of I, whose grading
is given by deg I n = n. Since R is noetherian, Rees(I) is a standard graded R-algebra.
L
L
n
n
n+1
Proof of the first part of Theorem 1.1. Clearly
M are finitely
n≥0 I M and
n≥0 I M/I
generated graded modules over Rees(I). By Theorem 1.2, we see that each of the sequences
(ldR I n M)n and (ldR I n M/I n+1 M)n is eventually constant.
To prove Theorem 1.2, it is harmless to replace S by a standard graded polynomial ring over
R which surjects onto it. Hence below, we will adopt the hypothesis that S = R[y1 , . . . , ym ] is a
polynomial extension of R with deg yi = 1. With this hypothesis, we make Theorem 1.2 more
precise by giving an upper bound for the smallest integer from which the sequence (ldR Cn )n
becomes constant. This requires certain information about the minimal graded free resolution
of C as an S-module.
4
HOP D. NGUYEN AND THANH VU
Definition 3.1. For each finitely generated graded S-module C, let pdegS (C) be the minimal
number such that Ci = 0 for all i ≥ pdegS (C) or Ci 6= 0 for all i ≥ pdegS (C). If C = 0, we set
pdegS (0) = −∞. Note that pdegS (C) is well-defined since S is standard graded.
We can compute the number pdegS (C) effectively, using two simple facts:
(i) pdegS (C) = pdegS/mS (C/mC). This is by Nakayama’s lemma.
(ii) pdegS/mS (C/mC) is bounded above by the point where the Hilbert function and Hilbert
polynomial of C/mC as a module over S/mS = k[y1 , . . . , ym ], start to agree. The latter
number is given, for example, in [4, Proposition 4.12].
Given a finitely generated graded S-module C, define the constant N(C) as follows. For i = 0,
denote n(0) = pdegS (C).
For 1 ≤ i ≤ min{gl ld R, pdS C}, denote by c(i, q) the number
c(i, q) := pdegS Im µi,q ,
where µi,q denotes the map TorSi (S/mq+1 S, C) → TorSi (S/mq S, C). Let
···
/
Fi
/
Fi−1
/
···
/
F0
/
0
be the minimal graded free resolution of C over S; a reference for the existence of such a
resolution is [4, Section 1.5].
Denote by Mi = Im(Fi → Fi−1 ) the i-syzygy of C. Using the Artin-Rees lemma, choose T (i)
to be the minimal number h ≥ 1 such that
mq Fi−1 ∩ Mi = mq−h (mh Fi−1 ∩ Mi )
for all q ≥ h. Denote
n(i) = max{c(i, 1), . . . , c(i, T (i))}.
Finally, let N(C) = max {n(0), n(1), . . . , n(min{gl ld R, pdS C})}. Since gl ld R is a finite number, N(C) is also finite.
Remark 3.2. In principal, the numbers T (i) in the definition of N(C) should not be difficult to
compute. Indeed, let n denote the graded maximal ideal of the associated graded ring
grmS (S) = (grm R) ⊗R S.
In concrete terms, n = (m/m2 ) ⊗R S. Furthermore, denote by Ki the kernel of the natural
surjective map grmS (Fi−1 ) → grmS (Fi−1 /Mi ). Then there is an equality
T (i) = sup{q : (Ki /nKi )q 6= 0}.
The proof is straightforward; see [16, Proposition 2.1] for an analogous statement.
As explained above, we can deduce Theorem 1.2 from the next result
Theorem 3.3. Let (R, m) be a local ring with gl ld R < ∞. Let S be a standard graded polynomial
ring over R and C a finitely generated graded S-module. Then for all n ≥ N(C), ldR Cn is a
constant independent of n.
In the proof, we will use the following characterization of linearity defect due to Şega.
ON THE ASYMPTOTIC BEHAVIOR OF THE LINEARITY DEFECT
5
Theorem 3.4 (Şega, [22, Theorem 2.2]). Let (R, m) be a noetherian local ring, M a finitely
generated R-module, and d ≥ 0 an integer. The following statements are equivalent:
(i) ldR M ≤ d;
q+1
q
(ii) The natural morphism TorR
, M) −→ TorR
i (R/m
i (R/m , M) is the zero map for every
i > d and every q ≥ 0.
Proof of Theorem 3.3. Since S is a flat R-algebra, there is an isomorphism of R-modules
S
q
q
∼
TorR
i (R/m , Cn ) = Tori (S/m S, C)n
(3.1)
for all i, q ≥ 0, n ∈ Z.
Let N = N(C) and e = sup{ldR Cn : n ≥ N} ≤ gl ld R < ∞. We prove that ldR Cn = e
for any n ≥ N. There is nothing to do if e = 0, so we assume that e ≥ 1. Note that, since
TorSi (S/mq S, C) = 0 for i > pdS C, Isomorphism (3.1) yields e ≤ min{gl ld R, pdS C}.
Denote by µe,q
n the following map
TorSe (S/mq+1 S, C)n −→ TorSe (S/mq S, C)n .
Choose m ≥ N such that ldR Cm = e. Since ldR Cm = e > e − 1, Theorem 3.4 implies that
e,q
µm
6= 0 for some q ≥ 0. To prove the inequality ldR Cn ≥ e, also by Theorem 3.4, it suffices to
show that µe,q
n 6= 0 for some q ≥ 0.
Firstly, consider the case q < T (e). Since n, m ≥ N ≥ c(e, q), the definition of c(e, q) implies
e,q
that µne,q and µe,q
m are both zero or both non-zero. This implies that µn 6= 0, as desired.
Secondly, consider the case q ≥ T (e). Denote T = T (e), we claim that µe,T
6= 0. As
n
e,T
m, n ≥ c(e, T ), µe,T
and
µ
are
both
zero
or
both
non-zero,
so
it
suffices
to
prove
that
µe,T
m
n
m 6= 0.
e,T
Assume the contrary, so µm = 0. Let F be the minimal graded free resolution of C over S.
Denote Mi = Im(Fi → Fi−1 ), the i-th syzygy of C as an S-module if i ≥ 1 and M0 = C. Denote
M = Me and P = Fe−1 . Clearly
TorSe (S/mq S, C) ∼
= TorS1 (S/mq S, Me−1 ) ∼
=
mq P ∩ M
,
mq M
mq+1 P ∩ M + mq M
Im µe,q ∼
.
=
mq M
The equality µe,T
m = 0 then yields
(mT +1 P ∩ M + mT M)m = (mT M)m .
e,q
= 0. Indeed,
We will show that µm
mq+1 P ∩ M + mq M
m
(3.2)
= mq−T (mT +1 P ∩ M + mT M) m
= mq−T mT +1 P ∩ M + mT M m
= mq−T (mT M)m
= (mq M)m .
In the above string, the first equality holds because of the inequality q ≥ T = T (e) and the
definition of T (e), the second and fourth because m ⊆ S0 , the third because of (3.2).
e,q
Therefore, µm
= 0. But this is a contradiction, so the proof of the theorem is finished.
6
HOP D. NGUYEN AND THANH VU
3.2. The second part of Theorem 1.1. We state a simple result which is useful for the
proof of the second part of Theorem 1.1, which concerns with the stability of the sequence
(ldR (M/I n M))n .
Lemma 3.5. Let 0 −→ M −→ P −→ N −→ 0 be an exact sequence of non-trivial, finitely
generated R-modules. Let F, G be the minimal free resolution of M, P , respectively. Assume that
there is a lifting ϕ : F −→ G of M → P such that ϕ(F ) ⊆ m2 G. Then there is an equality
ldR N = max{ldR P, ldR M + 1}.
Proof. Let W be the mapping cone of ϕ. Since ϕ(F ) ⊆ m2 G, W is a minimal free resolution of
N. By simple computations, we get a direct sum decomposition
linR W ∼
= linR G ⊕ (linR F )[−1]
Hence by accounting, ldR N = max{ldR P, ldR M + 1}, as desired.
Combining the first part of Theorem 1.1 with a result due to Avramov [2], we will provide the
Proof of the second part of Theorem 1.1. We have to prove the eventual constancy of the sequence (ldR (M/I n M))n . If there exists an n0 ≥ 1 such that I n0 M = 0 then I n M = 0 for all
n ≥ n0 , and we are done. Hence below, we assume that I n M 6= 0 for all n ≥ 1.
By [2, Corollary A.4], there exists d ≥ 1 such that for any P ⊆ md M, the map TorR
i (k, P ) →
TorR
(k,
M)
is
zero
for
all
i
≥
0.
Applying
the
same
result,
there
exists
e
≥
1
such
that
for any
i
R
R
d+e
d
P ⊆ m M, the map Tori (k, P ) → Tori (k, m M) is zero for all i ≥ 0.
R
R
n
d
d
Take n ≥ d + e. Then the maps TorR
i (k, I M) → Tori (k, I M) and Tori (k, I M) →
R
n
Tori (k, M) are zero for all i ≥ 0. Let F, G, H be the minimal free resolution of I M, I d M, M,
respectively. Take any lifting λ : F → G of the map I n M → I d M, then λ(F ) ⊆ mG. Similarly,
for any lifting ψ : G → H of the map I d M → M, we have ψ(G) ⊆ mH. Therefore we obtain a
lifting ϕ = ψ ◦ λ : F → H on the level of minimal free resolutions of the map I n M → M which
satisfies ϕ(F ) ⊆ m2 G.
By Lemma 3.5, we have for any n ≥ d + e the equality
ldR (M/I n M) = max{ldR M, ldR (I n M) + 1}.
By the first part of Theorem 1.1, for n large enough, ldR (I n M) is a constant independent of
n. Hence the same is true for ldR (M/I n M). This concludes the proof of the second part of
Theorem 1.1.
3.3. Final remarks. The following example illustrates the value of the constant N = N(C) in
Theorem 3.3, with the help of Macaulay2 [10].
Example 3.6. Let R = Q[x, y, z] be a polynomial ring of dimension 3 and I = (x2 , xy, z 2 ). Denote
S = R[w0 , w1 , w2 ] a standard graded polynomial extension of R which surjects onto the Rees
algebra E = Rees(I) by mapping w0 7→ x2 , w1 7→ xy, w2 7→ z 2 . The ring E has the following
presentation
S
.
E∼
=
2
(w0 y − w1 x, w0 z − w2 x2 , w1 z 2 − w2 xy)
The minimal graded free resolution of E over S is as follows
ON THE ASYMPTOTIC BEHAVIOR OF THE LINEARITY DEFECT
F :0→
S(−2)
L
w2 x −z 2
−w1
y
7
2
2
2
w0 y − w1 x w0 z − w2 x w1 z − w2 xy
w0 −x
−−−−−−−−−−−→ S(−1)3 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ S → 0.
S(−1)
Using the notation of the proof of Theorem 3.3, we will show that N = 1, namely all the powers
of I have the same linearity defect, which turns out to be 1. Since pdS E = 2 < gl ld R = 3,
N = max{n(0), n(1), n(2)}. The graded structure of E tells us that n(0) = pdegS (E) = 0.
Let J ⊆ S, M2 ⊆ G be the first and second syzygies of E, where G denotes the module
F1 = S(−1)3 . We claim that T (1) = 2 and T (2) = 1, namely,
mq S ∩ J = mq−2 (m2 S ∩ J), for all q ≥ 2,
(3.3)
mq G ∩ M2 = mq−1 (mG ∩ M2 ), for all q ≥ 1.
(3.4)
For (3.3): one sees immediately that both sides are equal to mq−1 (w0 y − w1 x)S + mq−2 (w0 z 2 −
w2 x2 , w1 z 2 − w2 xy)S.
For (3.4): we have M2 = (w2 xe1 − w1 e2 + w0 e3 , −z 2 e1 + ye2 − xe3 ), where e1 , e2 , e3 is the
standard basis of G sitting in degree 1. It is not hard to check that both sides of (3.4) are equal
to
mq (w2 xe1 − w1 e2 + w0 e3 ) + mq−1 (−z 2 e1 + ye2 − xe3 ).
The above arguments yield n(1) = max{c(1, 1), c(1, 2)} and n(2) = c(2, 1). We prove that
n(1) = 1 and n(2) = −∞.
For each q ≥ 1, TorS1 (S/mq S, E) = TorS1 (S/mq S, S/J) = (J ∩ mq S)/(Jmq S). Therefore
the image of TorS1 (S/mq+1 S, E) → TorS1 (S/mq S, E) is Im µ1,q = (J ∩ mq+1 S + Jmq S)/(Jmq S).
Computations show that
S2
1,1
Im µ =
,
mS 2 + (w0 e11 − w1 e12 )
where e11 , e12 is a basis for S 2 , both of degree 1, and
Im µ1,2 =
S5
,
mS(e23 , e24 , e25 ) + m2 S(e21 , e22 ) + (−xe21 + ye22 , w0 e21 − w1 e22 + w2 e25 )
where e21 , . . . , e25 are a basis for S 5 , all of them of degree 1. Thanks to routine Gröbner basis
arguments, the residue classes w0i e12 ∈ Im µ1,1 and w0i e22 ∈ Im µ1,2 are always non-zero for every
i ≥ 0. Hence c(1, 1) = c(1, 2) = 1, and thus n(1) = 1.
Denote by f1 , f2 the standard basis of F2 where deg f1 = 2, deg f2 = 1. Since TorS2 (S/mS , E) =
H2 (F ⊗S S/m2 S), computations show that
(i) TorS2 (S/mS , E) is generated by xf2 , yf2 , zf2 ∈ F2 ⊗ (S/m2 S),
(ii) TorS2 (S/mS, E) is generated by f2 ∈ F2 ⊗ (S/mS).
As TorS2 (S/mS, E) is killed by mS, the map TorS2 (S/mS , E) → TorS2 (S/mS, E) is zero; this yields
n(2) = c(2, 1) = −∞.
Putting everything together, N = max{n(0), n(1), n(2)} = max{0, 1, −∞} = 1.
8
HOP D. NGUYEN AND THANH VU
Theorem 3.3 has the following consequence on the linearity defect of the integral closure of
powers. Recall that the integral closure I of I is the ideal consisting of elements x ∈ R which
satisfies a relation
xn + a1 xn−1 + · · · + an−1 x + an = 0
where n ≥ 1 and ai ∈ I i for any 1 ≤ i ≤ n.
Corollary 3.7. Let (R, m) be a regular local ring and I ⊆ m an ideal. Then the sequence
(ldR I n )n is eventually constant.
Proof. Denote C = R ⊕ I ⊕ I 2 ⊕ · · · , then C is a graded module over Rees(I) with deg I n = n.
By [18, Proposition 5.3.4], C is a finitely generated Rees(I)-module. An application of Theorem
1.2 yields the desired conclusion.
Recall that the saturation of I is Ie = {x ∈ R : xmd ⊆ I for some d ≥ 1}. The next example
shows that the (graded) analog of Corollary 3.7 for saturation of powers does not hold.
Example 3.8. Consider the ideal I = (x(y 3 − z 3 ), y(x3 − z 3 ), z(x3 − y 3)) ⊆ R = C[x, y, z]. The
ideal I defines a reduced set of 12 points in P2 , the so-called Fermat configuration (see the proof
of [11, Proposition 2.1]). We show that the saturation ideals Ies do not have eventually constant
linearity defect.
T
For s ≥ 1, denote by I (s) = R ∩ P ∈Ass(I) I s RP the s-th symbolic power of I. Since I is the
defining ideal of a reduced set of points, we get that Ies = I (s) . From [11, Proposition 1.1], we
3s = (Ie3 )s . By [14, Theorem 2.4], ld If
3s = 0 for all s ≥ 1. Indeed, computations
deduce that If
R
with Macaulay2 [10] show that x, y + z, z is a d-sequence with respect to Rees(Ie3 ).
3s+1 = 1 for all s ≥ 1. First, since depth R/I]
3s+1 ≥ 1, by [6, Proposition
Now we show that ldR I]
6.3],
3s+1 ≤ dim R − 1 = 2.
ldR R/I]
3s+1 ≤ 1.
Hence ldR I]
Let H = (x3 − y 3 )(y 3 − z 3 )(z 3 − x3 ). We will show that the minimal non-zero component of
3s+1 = I (3s+1) is of degree 9s + 4 and
I]
(3s+1)
I
= (H s )Ih4i ∼
= I(−9s).
h9s+4i
3s+1
If this is the case, then I]
h9s+4i has linearity defect at least 1, as I does. (For the inequality
ldR I ≥ 1, use Römer’s theorem [19, Theorem 5.6] and the fact that I is generated in degree 4
3s+1 ≥ 1 for every s ≥ 1. All in all, we obtain
but does not have 4-linear resolution). Hence ldR I]
3s+1 = 1 for every s ≥ 1.
ldR I]
Now for our purpose, it suffices to prove the following claim:
(3s+1)
Ihdi
= (H s )Ihd−9si
(3.5)
holds for all d ≤ 9s + 4. We are grateful to Alexandra Seceleanu for providing us the following
nice argument.
We will proceed by induction on s; the starting case s = 0 is trivial. Assume that s > 0.
Let G be a homogeneous element of I (3s+1) of degree d. Here
Q9 the geometry of the Fermat
configuration comes into play. We have a decomposition H = i=1 hi , where each hi is a linear
ON THE ASYMPTOTIC BEHAVIOR OF THE LINEARITY DEFECT
9
form and no two of them are proportional. According to [11, Section 1.1], for each i, hi passes
through exactly 4 points (among the 12 points of the configuration). Moreover, each point of
the configuration lies on 3 of the 9 lines defined by the hi s.
Now as G lies in I (3s+1) , G passes through each point of the configuration with multiplicity at
least 3s + 1. Thus the curves (G) and (hi ) intersect with multiplicity at least 4(3s + 1), which is
strictly larger than d = (deg G) · (deg hi ). From that, Bezout’s theorem forces G to be divisible
by hi for all 1 ≤ i ≤ 9. In particular G is divisible by H. Writing G = HG′ , then as H vanishes
(3s+1−3)
(3(s−1)+1)
exactly 3 times at each of the points, we must have G′ ∈ Ihd−9i
= Ihd−9i
. Finally, the
induction hypothesis gives us the claim.
So we conclude that the sequence ldR Ies is not eventually constant when s goes to infinity.
Remark 3.9. More generally than Theorem 1.2, one can prove the following: If S is a noetherian
R-algebra which is generated by elements of positive degrees, and C is a finitely generated graded
S-module, then the sequence (ldR Cn )n is quasiperiodic, i.e. there exist a number p ≥ 1 and
fn = ℓi if n is congruent to
integral constants ℓ0 , . . . , ℓp−1 such that for all n ≫ 0, we have ldR C
i ∈ {0, . . . , p − 1} modulo p.
The proof uses the fact that any high enough Veronese subring of S is standard graded (after
normalizing the grading), and Theorem 1.2. We leave the details to the interested reader (see
an analogous statement in [7, Theorem 4.3]).
By [20, Theorem 4.3], for the ideal I in Example 3.8, the graded R-algebra R ⊕ Ie ⊕ Ie2 ⊕ · · ·
is finitely generated. This fact and the above general version of Theorem 1.2 guarantee the
quasiperiodic behavior of the sequence (ldR Ien )n in the example.
We do not know any example where the sequence (ldR Ien )n is not quasiperiodic. In view of
[7, Example 4.4] on bad behavior of regularity for saturations of powers, it is desirable to seek
for one.
Acknowledgments
The content of this paper has been worked out when the first author was visiting the Department of Mathematics, University of Nebraska – Lincoln (UNL) in March 2015. We are grateful
to Luchezar Avramov, Dale Cutkosky, Hailong Dao, Thomas Marley and Alexandra Seceleanu
for inspiring comments and useful suggestions. The first author would like to thank the colleagues at Department of Mathematics of the UNL, among them Luchezar Avramov, Roger and
Sylvia Wiegand for their hospitality and warm support. Finally, both authors are grateful to the
anonymous referee for her/his careful reading of the manuscript and many thoughtful comments
which greatly improve the presentation.
References
[1] R. Ahangari Maleki and M.E. Rossi, Regularity and linearity defect of modules over local rings. J. Commut.
Algebra 6, no. 4 (2014), 485–504.
[2] L. L. Avramov, Small homomorphisms of local rings. J. Algebra 50 (1978), 400–453.
[3] M. Brodmann, The asymptotic nature of the analytic spread. Math. Proc. Cambridge Philos. Soc. 86 (1979),
no. 1, 35–39.
[4] W. Bruns and J. Herzog, Cohen-Macaulay rings. Rev. ed.. Cambridge Studies in Advanced Mathematics 39,
Cambridge University Press (1998).
10
HOP D. NGUYEN AND THANH VU
[5] M. Chardin, Powers of ideals and the cohomology of stalks and fibers of morphisms. Algebra Number Theory
7 (2013), no. 1, 1–18.
[6] A. Conca, S.B. Iyengar, H.D. Nguyen and T. Römer, Absolutely Koszul algebras and the Backelin-Roos
property. Acta Math. Vietnam. 40 (2015), 353–374.
[7] S.D. Cutkosky, J. Herzog and N.V. Trung, Asymptotic behaviour of the Castelnuovo-Mumford regularity.
Compositio Math. 118 (1999), no. 3, 243–261.
[8] D. Eisenbud, Commutative algebra. With a view toward algebraic geometry. Graduate Texts in Mathematics
150. Springer-Verlag, New York (1995).
[9] D. Eisenbud, G. Fløystad and F. -O. Schreyer, Sheaf cohomology and free resolutions over exterior algebras.
Trans. Amer. Math. Soc. 355 (2003), 4397–4426.
[10] D. Grayson and M. Stillman, Macaulay2, a software system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2.
[11] B. Harbourne and A. Seceleanu, Containment counterexamples for ideals of various configurations of points
in PN . J. Pure Appl. Algebra 219 (2015), 1062–1072.
[12] J. Herzog and T. Hibi, Componentwise linear ideals. Nagoya Math. J. 153 (1999), 141–153.
[13] J. Herzog and T. Hibi, The depth of powers of an ideal. J. Algebra 291 (2005), no. 2, 534–550.
[14] J. Herzog, T. Hibi and H. Ohsugi, Powers of componentwise linear ideals. Abel Symp. 6, Springer, Berlin
(2011).
[15] J. Herzog and S.B. Iyengar, Koszul modules. J. Pure Appl. Algebra 201 (2005), 154–188.
[16] J. Herzog, V. Welker and S. Yassemi, Homology of powers of ideals: Artin-Rees numbers of syzygies and the
Golod property. Preprint (2011), http://arxiv.org/abs/1108.5862.
[17] C. Huneke, The theory of d-sequence and powers of ideals. Adv. in Math. 46 (1982), no. 3, 249–279.
[18] C. Huneke and I. Swanson, Integral closure of ideals, rings and modules. London Math. Soc., Lecture Note
Series, vol. 336. Cambridge University Press, Cambridge (2006).
[19] S.B. Iyengar and T. Römer, Linearity defects of modules over commutative rings. J. Algebra 322 (2009),
3212–3237.
[20] U. Nagel and A. Seceleanu, Ordinary and symbolic Rees algebras for ideals of Fermat point configurations.
Preprint (2015), available online at http://arxiv.org/abs/1509.04977.
[21] R. Okazaki and K. Yanagawa, Linearity defect of face rings. J. Algebra 314 (2007), 362–382.
[22] L.M. Şega, On the linearity defect of the residue field. J. Algebra 384 (2013), 276–290.
[23] K. Yanagawa, Linearity defect and regularity over a Koszul algebra. Math. Scand. 104 (2009), no. 2, 205–220.
Fachbereich Mathematik/Informatik, Institut für Mathematik, Universität Osnabrück, Albrechtstr. 28a, 49069 Osnabrück, Germany
Dipartimento di Matematica, Università di Genova, Via Dodecaneso 35, 16146 Genoa, Italy
E-mail address: [email protected]
Department of Mathematics, University of Nebraska-Lincoln, Lincoln, NE 68588, USA
E-mail address: [email protected]
| 0math.AC
|
1
Enriched K -Tier HetNet Model to Enable the
Analysis of User-Centric Small Cell Deployments
arXiv:1606.06223v2 [cs.IT] 18 May 2017
Chiranjib Saha, Mehrnaz Afshang, and Harpreet S. Dhillon
Abstract—One of the principal underlying assumptions of
current approaches to the analysis of heterogeneous cellular
networks (HetNets) with random spatial models is the uniform
distribution of users independent of the base station (BS) locations. This assumption is not quite accurate, especially for
user-centric capacity-driven small cell deployments where lowpower BSs are deployed in the areas of high user density, thus
inducing a natural correlation in the BS and user locations. In
order to capture this correlation, we enrich the existing K-tier
Poisson Point Process (PPP) HetNet model by considering user
locations as Poisson Cluster Process (PCP) with the BSs at the
cluster centers. In particular, we provide the formal analysis of
the downlink coverage probability in terms of a general density
functions describing the locations of users around the BSs. The
derived results are specialized for two cases of interest: (i) Thomas
cluster process, where the locations of the users around BSs are
Gaussian distributed, and (ii) Matérn cluster process, where the
users are uniformly distributed inside a disc of a given radius.
Tight closed-form bounds for the coverage probability in these
two cases are also derived. Our results demonstrate that the
coverage probability decreases as the size of user clusters around
BSs increases, ultimately collapsing to the result obtained under
the assumption of PPP distribution of users independent of the
BS locations when the cluster size goes to infinity. Using these
results, we also handle mixed user distributions consisting of two
types of users: (i) uniformly distributed, and (ii) clustered around
certain tiers.
Index Terms—Stochastic geometry, heterogeneous cellular network, Poisson cluster process, Poisson point process, user-centric
deployments.
I. I NTRODUCTION
Increasing popularity of Internet-enabled mobile devices,
such as smartphones and tablets, has led to an unprecendented
increase in the global mobile data traffic, which has in turn
necessitated the need to dramatically increase the capacity
of cellular networks. Not surprisingly, a key enabler towards
increasing network capacity at such a rate is to reuse spectral
resources over space and time more aggressively. This is
already underway in the form of capacity-driven deployment
of several types of low-power BSs in the areas of high
user density, such as coffee shops, airport terminals, and
downtowns of large cities [2], [3]. Due to the coexistence
of the various types of low-power BSs, collectively called
small cells, with the conventional high-power macrocells, the
resulting network is often termed as a heterogeneous cellular
network (HetNet). Because of the increasing irregularity of
The authors are with Wireless@VT, Department of ECE, Virginia Tech,
Blacksburg, VA, USA. Email: {csaha, mehrnaz, hdhillon}@vt.edu. The support of the US NSF (Grants CCF-1464293 and CNS-1617896) is gratefully
acknowledged. This work was presented in part at the IEEE ICC in Kuala
Lumpur, Malaysia, in May 2016 [1].
BS locations in HetNets, random spatial models have become
preferred choice for the accurate modeling and tractable analysis of these networks. The most popular approach is to model
the locations of different classes of BSs by independent PPPs
and perform the downlink analysis at a typical user chosen
independent of the BS locations; see [4]–[6] and the references
therein. However, none of the prior works has focused on
developing tools for the more realistic case of user-centric
deployments in which the user and BS locations are correlated.
Developing new tools to fill this gap is the main goal of this
paper.
A. Related Works
Stochastic geometry has recently emerged as a useful tool
for the analysis of cellular networks. Building on the single-tier
cellular model developed in [7], a multi-tier HetNet model was
first developed in[8], [9], which was then extended in [10]–
[12]. While the initial works were mainly focused on the
downlink coverage and rate analyses, the models have since
been extended in multiple ways, such as for load aware
modeling of HetNets in [13], traffic offloading in [14], and
throughput optimization in [15]. Please refer to [4]–[6], [16],
[17] for more pedagogical treatment of the topic as well as
extensive surveys of the prior art. While PPP remains a popular
abstraction of spatial distribution of cellular BSs randomly and
independently coexisting over a finite but large area, a common
assumption of the aforementioned analysis, as noted above,
is that the users are uniformly distributed independent of the
BS locations. However, in reality, the users form hotspots,
which are where some types of small cells, such as picocells
are deployed to enhance coverage and capacity [18]. As a
result, the user-centric deployment of small cells is one of the
dominant themes in future wireless architectures [19]. In such
architectures, one can envision small cells being deployed to
serve clusters of users. Such models are also being used by
the standardization bodies, such as 3GPP [2], [3]. While there
have been attempts to model such clusters of small cells by
using PCP, e.g., see [20]–[24], the user distribution is usually
still assumed to be independent of the BS locations.
As noted above, modeling and performance analysis of
user-centric capacity-driven deployment of small cells require
accurate characterization of not only the spatial distribution of
users but also correlation between the BS and user locations.
Existing works, however sparse, on the analysis of correlated
non-uniform user distributions can be classified into two main
directions. The first is to characterize the performance through
detailed system-level simulations [25]–[28]. As expected, the
2
general philosophy is to capture the capacity-centric deployments by assuming higher user densities in the vicinity of
small cell BSs, e.g., see [25]. In [26], the authors proposed
non-uniform correlated traffic pattern generation over space
and time based on log-normal or Weibull distribution. On
similar lines, [27] has introduced a low complexity PPP
simulation approach for HetNets with correlated user and BS
locations. System level simulation shows that network performance significantly deteriorates with increased heterogeneity
of users if there exists no correlation among the users and the
small cell BS locations. But the HetNet performance improves
if the small cell BSs are placed at the cluster centers which
are determined by means of clustering algorithms from a given
user distribution [28].
The second direction, in which the contributions are even
sparser, is to use analytic tools from stochastic geometry to
characterize the performance of HetNets with non-uniform
user distributions. One notable contribution in this direction
is the generative model proposed in [29], where non-uniform
user distribution is generated from the homogeneous PPP by
thinning the BS field independently, conditional on the active
link from a typical user to its serving BS. While the resulting
model is tractable, it suffers from two shortcomings: (i) it is
restricted to single-tier networks and extension to HetNet is not
straightforward, and (ii) even for single-tier networks, it does
not allow the inclusion of any general non-uniform distribution
of users in the model. In [30], the authors proposed a mixture
of correlated and uncorrelated user distribution with respect
to small cell BS deployment and evaluated the enhancement
in coverage probability as a function of correlation coefficient.
Correlation has been introduced by generating users initially as
an independent PPP and later shifting them towards the BSs
with some probability. In [15], the authors have considered
clustered users around femto-BSs as uniformly distributed on
the circumference of a circle with fixed radius. Besides, some
other attempts have been made at including non-uniform user
distributions using simple models, especially in the context of
indoor communications, e.g., see [31]. In [32], both the user
and small scale BS locations are modeled as correlated Matérn
cluster processes having the same “parent” point process. But
the analysis is simplified by assuming the distance between
a user and its serving small cell BS either being fixed or a
uniformly distributed random variable. Overall, we are still
somewhat short-handed when it comes to handling the analysis
of user-centric deployments, which is the main focus of this
paper. With this brief overview of the prior art, we now discuss
our contributions next.
This model is flexible enough to include any kind of user
distribution around any arbitrary number of BS tiers as well
as user distribution that is homogeneous and independent of
the BS locations. This approach builds on our recent work on
modeling device-to-device networks using PCPs [33], [34].
2) Downlink Analysis: We derive exact expression for the
coverage probability of a typical user chosen randomly from
one of the clusters in this setup. The key step of our approach
is the treatment of the cluster center as an individual singleton
tier. This enables the characterization of key distance distributions, which ultimately lead to easy-to-use expressions for the
Laplace transform of interference distribution in all cases of
interest. Using these components, we derive the coverage probability of a randomly chosen user from one of the user clusters.
After characterizing the coverage probability under a general
distribution of users, we specialize our results for two popular
PCPs, viz. Thomas and Matérn cluster processes. Next, we
provide upper and lower bounds on coverage probability which
are computationally more efficient than the exact expressions
and reduce to closed form expressions for no shadowing when
the user distribution is modelled as Thomas or Matérn cluster
process. Although our analysis primarily focuses on users
clustered around BS locations, we also consider users that are
independently and homogeneously located over the network
modeled as a PPP and use previously derived results for
coverage [10] in conjunction to evaluate the overall coverage
probability for any randomly chosen user in our HetNet setup
with mixed user distribution.
3) System Design Insights: Our analysis leads to several
system-level design insights. First, it can be observed that
the coverage probability under the assumption of BS-user
correlation is significantly greater than that derived under the
assumption of independence. While the assumption of independence of BS and user locations does simplify analyses, the
resulting coverage probability predictions may be significantly
pessimistic. That being said, our results concretely demonstrate that the difference between the coverage probabilities
corresponding to user-centric and independent BS deployment
becomes less significant as the cluster sizes (of user cluster)
increase. In the limit of cluster size going to infinity, the
new coverage results are shown to mathematically converge
to the results obtained under independent user distribution
assumption. Second, as opposed to the previous works, the
coverage probability of users clustered around BSs under
interference-limited open access network is a function of
BS transmission power. Our analysis shows that coverage
probability can be improved by increasing transmission power
of small cell BSs located at centers of the user clusters.
B. Contributions and Outcomes
1) New HetNet Model: In this paper, we develop a new and
more practical HetNet model for accurately capturing the nonuniform user distribution as well as correlation between the
locations of the users and BSs. In particular, the user locations
have been modeled as superposition of PCPs. Correlation
between the users and BSs under user-centric capacity-driven
deployment has been captured by assuming the BS locations
as the parent point processes of the cluster processes of users.
II. S YSTEM M ODEL
A. BS Deployment
Consider a K-tier HetNet, where BSs across tiers (or
classes) differ in terms of their transmit powers and deployment densities. For mathematical convenience and notational
simplicity, define K = {1, 2, . . . , K} as the indices of the K
tiers. The locations of the k th -tier BSs are modelled by an
(BS)
(BS)
independent homogeneous PPP Φk
of density λk
> 0
3
(a) Uniformly distributed users independent of
the BS locations (prior art).
(b) Users clustered around small cell BSs (this
paper).
(c) Mixed (clustered and uniformly distributed)
user distribution (this paper).
Fig. 1: Macro (green squares) and small cell BSs (black dots) are distributed as independent PPPs λ2 = λ02 = λ1 /10. The uniformly
distributed users are represented by small blue dots and clustered users by small red dots. The average number of users per cluster (wherever
applicable) is 10.
(k ∈ K). The k th -tier BSs are assumed to transmit at the
same power Pk . As is usually the case, we assume that a
fraction of k th -tier BSs are in open access for the user of
interest while the rest are in closed access. The k th -tier open
and closed access BSs are modelled by two independent PPPs
(BS,o)
(BS,c)
Φk
and Φk
with densities λk and λ0k , respectively,
(BS)
(BS,o)
(BS,c)
(BS)
where Φk
= Φk
∪ Φk
and λk
= λk + λ0k .
B. User Distribution
Unlike prior art that focused almost entirely on the performance analysis of users that are uniformly distributed in
the network independent of the BS locations, we focus on a
correlated setup where users are more likely to lie closer to
the BSs. Since small cells are usually deployed in the areas
of high user density, this is a much more accurate approach
for modeling HetNets compared to the one where users and
BSs are both modeled as independent PPPs. We model this
scenario by modeling the locations of the users by a PCP with
one small cell deployed at the center of each user cluster. To
maintain generality, we assume that a subset B ⊂ K tiers
out of K tiers have clusters of users around the BSs. In
particular, given the location of a BS in the ith tier acting
as cluster center (i ∈ B), the users of the cluster are assumed
to be symmetrically, independently, and identically distributed
(i.i.d.) around it. Union of all such locations of users around
the BSs of the ith tier forms a PCP [35], [36], denoted by Φui ,
(BS)
where the parent point process of Φui is Φi . To maintain
(i)
generality, we assume that the user location Zu ∈ R2 with
respect to its cluster center follows some arbitrary distribution
with probability density function (PDF) fZ(i) (·), which may
u
not necessarily be the same across tiers. This allows to capture
the fact the cluster size may affect the choice of small cell to
be deployed there. For instance, it may be sufficient to deploy
a low power femtocell to serve a small cluster of users in a
coffee shop, whereas a relatively higher power picocell may
be needed to serve a cluster of users at a big shopping mall
or at an airport. After deriving all the results in terms of the
general distributions, we will specialize them to two cases of
interest where Φui is modeled as: (i) Thomas cluster process
in which the users are scattered according to a symmetric
(BS)
normal distribution of variance σi2 around the BSs of Φi
[37], hence,
1
kzk2
fZ(i) (z) =
exp
−
, z ∈ R2 ,
u
2πσi2
2σi2
and (ii) Matérn cluster process which assumes symmetric
uniform spatial distribution of users around the cluster center
within a circular disc of radius Ri , thus
(
1
if kzk ≤ Ri
2
fZ(i) (z) = πRi
,
u
0
otherwise
(i)
where z is a realization of the random vector Zu . While our
primary interest is in these clustered users, we also consider
users that are homogeneously distributed over the network
independent of the BS locations, for instance, pedestrians and
users in transit. These users are better modeled by a PPP as
done in literature (see [9]–[11], [38], [39] for a small subset).
Thus, in addition to the user clusters modeling users in the
hotspots, we also consider an independent point process of
users Φu(PPP) which is a PPP of density λ(PPP) . Fig. 1 shows
the two-tier HetNet setup with high power macro-BSs overlaid
with an independent PPP of denser but low power small cell
BSs. Fig. 1a illustrates the popular system model used in the
literature where users are modeled as Φu(PPP) [8]–[11], [38].
Fig. 1b highlights the correlated setup where users are only
clustered around small cell BSs. The general scenario i.e. the
mixed user distribution formed by the superposition of PPP
and PCP has been depicted in Fig. 1c.
Since the downlink analysis at the location of a typical
user of Φu(PPP) is well-known, in this paper we will focus
exclusively on the downlink performance of a typical user of
Φui , which is a randomly chosen user from a randomly chosen
4
TABLE I: Summary of Notations
Notation
(BS,o)
Φk
, λk
(BS,c)
Φk
, λ0k
B
Φu
i
Φu(PPP)
y0 , Y 0
(BS,o)
Φ0
Φk , λk
Φ0k , λ0 k
Pk , hk , Vk
yk
xk
Ni
Rk
b(0, r)
Io(j,k)
Ic(k)
Pc (i) , Pc (PPP)
Pc
Description
(BS,o)
serving BS from Φk
can be denoted as:
1
(BS,o)
PPP of BSs of kth open access tier, density of Φk
(BS,c)
PPP of BSs of kth closed access tier, density of Φk
Set of BS tiers that have users clustered around them
(BS)
Point process modeling users clustered around BSs of Φi
Locations of uniformly distributed users modeled as a PPP
Location of cluster center in Euclidean space, Y0 = ky0 k
Tier 0 containing only the cluster center
(BS,o)
Equivalent PPP of Φk
to incorporate shadowing, density of Φk
(BS,c)
Equivalent PPP of Φk
to incorporate shadowing, density of Φ0k
Transmit power, small scale fading gain, shadowing gain
(BS)
Actual location of a BS in Φk
cluster of Φui , also termed as the representative cluster. In
other words, we will primarily focus on the scenario depicted
in Fig. 1b (and then extend our results and insights to scenario
depicted in Fig. 1c). Since the PPPs are stationary, we can
transform the origin to the location of this typical user. Quite
reasonably, we assume that the BS at the center of the representative cluster is in open access mode. This assumption can
be easily relaxed without much effort. Denote the location of
(BS,o)
(BS,o)
the representative cluster center by y0 ∈ Φi
. Now Φi
can be partitioned into two sets: (i) representative cluster center
(BS,o)
y0 , and (ii) the rest of the points Φi
\ y0 . By Slivnyak’s
(BS,o)
theorem, it can be argued that Φi
\ y0 has the same
(BS,o)
distribution as Φi
[37]. For notational simplicity, we form
an additional tier (call it tier 0) consisting of a single point
(BS)
(BS,o)
y0 , i.e., Φ0
≡ Φ0
≡ {y0 }. Then, the set of indices
of all tiers is enriched to K1 = {0} ∪ K = {0, 1, 2, . . . , K}.
The user can either connect to its own cluster center i.e. the
(BS,o)
BS in Φ0
, or to some other BS belonging to one of the
(BS,o)
(BS,o)
tiers Φ1
, . . . , ΦK
. It will be evident in sequel that this
construction will allow us to handle the link from the typical
user to its cluster center separately.
C. Channel Model and User Association
The received power at the location of the typical user at
(BS)
origin from a BS at yk ∈ Φk
(k ∈ K1 ) is modelled
−α
as P (yk ) = Pk hk Vk kyk k , where, α > 2 is the path
loss exponent, hk is the small-scale fading gain and Vk
is the shadowing gain. Under Rayleigh fading assumption,
{hk } is a sequence of i.i.d. exponential random variables
(RVs) with hk ∼ exp(1). For large scale shadowing, we
assume {Vk } to be sequence of i.i.d. log-normal RVs , i.e.,
10 log Vk ∼ N (µk , ηk2 ), with µk and ηk respectively being
the mean and standard deviation (in dB) of the channel power
under shadowing. In this model, we assume average received
power based cell selection in which a typical user connects to
the BS that provides maximum received power averaged over
small-scale fading. The serving BS will be one from the K +1
candidate BSs from each tier. The location of such candidate
yk ∈
(BS,o)
Φk
yk ∈
(BS,o)
Φk
.
Since the 0th tier consists of only a single BS, i.e., the cluster
center, there is only one choice of the candidate serving BS
(BS,o)
from Φ0
, i.e., y0∗ ≡ y0 . The serving BS will be one of
these candidate serving BSs, denoted by
y∗ = arg max∗ Pk Vk kyk−α .
y∈{yk }
−1
Vk α y k )
Location of BS in transformed space (xk =
Average number of users per cluster of Φu
i
Modified distance of nearest BS ∈ Φk , Rk = min kxk k
Disc with radius r centered at origin
Interference from all BSs ∈ Φk when user connects to a BS ∈ Φj
Interference from all BSs ∈ Φ0k
u(PPP)
Coverage probability of a typical user in Φu
i, Φ
Overall coverage probability
−α
yk∗ = arg max Pk Vk kyk k−α = arg max Pk Vk− α kyk k
Using the displacement theorem of PPPs [40, Section 1.3.3]
, it was shown in [41], [42] that if each point in a PPP
(BS,o)
(BS,c)
Φk
(Φk
) is independently displaced such that the
−1
transformed location becomes xk = Vk α yk , then, the resultant point process remains a PPP,
h 2 iby
h 2 iwhich we denote
Φk (Φ0k ) with density λk = λk E Vkα (λ0 k = λ0k E Vkα ).
This transformation is valid for any arbitrary distribution of
2
Vk with PDF fVk (·) as long as E(Vkα ) is finite, which is
indeed true for log-normal distribution. Consequently, we can
express instantaneous received power from a BS ∈ Φk as
Pk hk kxk k−α . Then, the location of candidate serving BS in
Φk can be written as
x∗k = arg max Pk kxk k−α .
xk ∈Φk
For k = 0, we apply similar transformation to the point y0 ∈
(BS,o)
Φ0
and denote the transformed process as Φ0 where x0 ≡
1
−α
V0 y0 . Then, the serving BS at x∗ will be
x∗ = arg max∗ Pk kxk−α .
x∈{xk }
It is worth noting that in the absence of shadowing, the
candidate serving BS from a given tier will be the BS closest
to the typical user from that tier in terms of the Euclidean
distance. This is clearly not true in the presence of shadowing
because of the possibility of a farther off BS providing higher
average received power than the closest BS. However, by
applying displacement theorem, the effect of shadowing gains
has been incorporated at the modified locations xk such that
the strongest BS in the equivalent PPP Φk is also the closest
in terms of the Euclidean distance. As demonstrated in the
literature (e.g., see[41], [42]) and the next two Sections, this
simplifies the coverage probability analysis in the presence
of shadowing significantly. For notational simplicity, let us
define the association event to tier j as SΦj such that
1SΦj = 1(x∗ = x∗j ) (here 1(·) is the indicator function). The
Signal-to-Interference and Noise Ratio (SINR) experienced by
a typical user at origin when 1SΦj = 1 can be expressed as:
SINR(kx∗ k) ≡
N0 +
P
k∈K1
P h kx∗ k−α
P j j
xk ∈Φk ∪Φ0k \{x∗ }
,
Pk hk kxk k−α
(1)
where N0 is the thermal noise power. For quick reference, the
notations used in this paper are summarized in Table I.
Remark 1. While we transform all the PPPs to equivalent
PPPs to incorporate shadowing, the impact of shadowing on
5
the link between the typical user and its cluster center, i.e.,
1
x0 = V0α y0 , needs to be handled separately. For this, we
have two alternatives. First is to find the distribution of x0
as a function of the distributions of V0 and y0 . Second is
to proceed with the analysis by conditioning on shadowing
variable V0 and decondtioning at the last step. We take the
second approach since it gives simpler intermediate results
which can be readily used for no-shadowing scenario by
putting Vk ≡ 1.
Remark 2. In the special case when Φui is a Thomas cluster
process, user coordinates in Cartesian domain are i.i.d. normal RVs with variance σi2 . Then, Y0 is Rayleigh distributed
with PDF and CCDF [34]:
2
−y0
y0
,
y0 ≥ 0, (4a)
PDF:
fY0 (y0 ) = 2 exp
σi
2σ 2
2 i
−y0
CCDF:
,
y0 ≥ 0. (4b)
F Y0 (y0 ) = exp
2σi2
III. A SSOCIATION P ROBABILITY AND S ERVING D ISTANCE
Remark 3. If Φui is a Matérn cluster process, the PDF and
CCDF of Y0 are:
This is the first technical section of the paper, where we
derive the probability that a typical user is served by a given
tier j ∈ K1 , which is usually termed as the association
probability. We will then derive the distribution of kx∗ k
conditioned on SΦj , i.e., the distance from the typical user
to its serving BS conditioned on the the event that it belongs
to the j th open access tier. Recall that the candidate serving
BS located at xk from the equivalent PPP Φk is the one that
is nearest to the typical user located at the origin. Let us
call Rk = kx∗k k as the RV denoting the distance from the
typical user to the nearest point of Φk . Since Φk (k ∈ K)
are independent homogeneous PPPs, the distribution of Rk ,
k ∈ K, is [37]
PDF:
fRk (rk ) =
CCDF:
F Rk (rk ) =
2πλk rk exp(−πλk rk2 )
exp(−πλk rk2 )
rk ≥ 0, (2a)
rk ≥ 0. (2b)
In a similar way, we can define modified distance R0 =
−1
kx0 k = V0 α ky0 k. As noted in Remark 1, we will proceed
with the analysis by conditioning on the shadowing gain V0
and then deconditioning on V0 at the very end. Since R0 is
just a scaled version of ky0 k, it suffices to find the distribution
of Y0 ≡ ky0 k, which we do next.
Recall that the typical user is located at the origin, which
means the relative location of the cluster center with respect
to the typical user, i.e., y0 , has the same distribution as that of
(i)
Zu . Using standard transformation technique from Cartesian
to polar coordinates, we can obtain the distribution of distance
Y0 from the joint distribution of position coordinates (t1 , t2 ),
where y0 = (t1 , t2 ) is in Cartesian domain. Let us denote the
joint PDF of the polar coordinates (Y0 , Θ) as fY0 ,Θ (·). Then
t1 , t2
fY0 ,Θ (y0 , θ) = fy0 (t1 , t2 ) × ∂
,
(3)
y0 , θ
2y0
,
R2i
R2 − y 2
F Y0 (y0 ) = i 2 0 ,
Ri
fY0 (y0 ) =
PDF:
CCDF:
0 ≤ y0 ≤ Ri ,
(5a)
0 ≤ y0 ≤ Ri .
(5b)
A. Association Probability
To derive association probability, let us first characterize the
association event SΦj as: 1SΦj =
\
1(arg max Pk Rk−α = j) =
1 Rk > P̄jk Rj ,
(6)
k∈K1
k∈K1
1/α
where P̄jk = PPkj
and 1(·) is the indicator function of
the random vector R = [R0 , R1 , ..., Rk ]. Note that since the
0th tier is derived from the ith tier, we have P0 ≡ Pi . The
association probability for each tier is now defined as follows.
Definition 1. Association probability, Aj for j th tier, ∀j ∈ K1
is defined as the probability that the typical user will be served
by the j th tier. It can be mathematically expressed as
Aj = P(SΦj ).
(7)
The following Lemma deals with the conditional association
probability to Φj .
Lemma 1. Conditional association probability of the j th tier
given V0 = v0 is
K
1
Q
−α
E
F
(
P̄
v
Y
)
if j = 0;
0
Y0 k=1 Rk 0k 0
K
Aj|v0 =
1
Q
F Rk (P̄jk Rj )
if j ∈ K.
ERj F Y0 (v0α P̄j0 Rj )
k=1
k6=j
(8)
where
∂t1
∂y0
=
∂t2
∂
t1 , t2
y0 , θ
∂y0
∂t1
∂θ
.
∂t2
∂θ
From the joint distribution, the marginal distribution of distance Y0 can now be computed by integrating over θ as
Z 2π
fY0 (y0 ) =
fY0 ,Θ (y0 , θ)dθ.
0
Proof. See Appendix A.
Remark 4. Association probabilities of the j th tier can be
obtained by taking expectation over Aj|V0 with respect to V0 ,
i.e.,
Aj = EV0 (Aj|V0 ).
From Lemma 1, we can obtain the expressions for the
association probabilities to different open access tiers when
Φui is Thomas or Matérn cluster process. The conditional
6
probabilities in these cases can be reduced to closed form
expressions. The results are presented next.
Corollary 1. When Φui is a Thomas cluster process, conditional association probability of the j th tier given V0 = v0
is:
λj
Aj|v0 =
K
P
k=0
, ∀j ∈ K1 ,
(9)
v0α
.
2πσi2
Proof. See Appendix B.
Corollary 2. If Φui is a Matérn cluster process, conditional
association probability of the j th tier given V0 = v0 is:
Aj|v0 =
2
2
−α
v0α
2
R2 Z 1 − exp − v0 Z0 Ri
if j = 0
i 0
2
, (10)
λj π P̄ 2 v0α
Zj R2i
πλ
1
−
exp
−
if
j
∈
K
Zjj − R2j0
2
Z2
where Zj = π
2 vα
P̄j0
0
j
K
P
k=1
2
, ∀ j ∈ K1 .
λk P̄jk
Proof. See Appendix C.
1
if j = 0
,
if j ∈ K
i
k=1
where 0 ≤ wj ≤ Ri . For wj > Ri , fWj |V0 (wj |v0 ) = 0, ∀j ∈
K1 .
2
i
2
K
P
2v0α w0
2
2
λk P̄0k w0
A0|v0 exp −π
R2i
k=1
2
K
P
R2i −v0α wj2
2
2
A 1 2πλj exp −π
wj
λ
P̄
w
k jk j
R2
j|v
0
2 λ
P̄jk
k
where λ0 is defined as λ0 =
Corollary 4. If Φui is Matérn cluster process, the conditional
distribution of serving distance Wj given V0 = v0 can be
expressed as fWj |V0 (wj |v0 ) =
B. Serving Distance Distribution
In this section, we derive the distribution of kx∗ k when
1SΦj = 1 , i.e., the serving distance from the typical user to
its serving BS when it is in Φj . We will call this RV Wj .
Conditioned on SΦj , Wj is simply the distance to the nearest
BS in Φj . Hence Wj is related to Rj as Wj = Rj |SΦj . The
conditional PDF of Wj given V0 = v0 is derived in the next
Lemma.
Lemma 2. Conditional distribution of serving distance Wj at
V0 = v0 is obtained by fWj |V0 (wj |v0 ) =
K
1
1
Q
1
α
α
if j = 0,
A0|ν0 k=1 v0 F Rk P̄0k w0 fY0 (v0 w0 ),
K
1
Q
1
F Y0 (v0α P̄j0 wj )
F Rk (P̄jk wj )fRj (wj ), if j ∈ K.
A
j|ν
0
k=1
k6=j
(11)
Proof. Substituting fY0 (·) for Matérn cluster process from
Eq. 5a and CCDF of Rk from Eq. 2b in Eq. 11 and proceeding
as before, fW0 |V0 (·) can be derived. Similarly, fWj |V0 (·) is
obtained by substituting F Y0 (·) from Eq. 5b. For wj > Ri ,
fWj |V0 (wj |v0 ) = 0, ∀j ∈ K1 as fY0 (·) and F Y0 (·) take zero
value beyond this range.
IV. C OVERAGE P ROBABILITY A NALYSIS
This is the second technical section of the paper where we
use the association probability and the distance distribution
results obtained in the previous section to derive easy-to-use
expressions for the coverage probability of a typical user of
Φui in a user-centric deployment.
According to the association policy, it is easy to deduce
that if the typical user is served by a BS ∈ Φj located at a
distance Wj , there exist no k th tier BSs, ∀k ∈ K1 , within
a disc of radius P̄jk Wj centered at the location of typical
user (origin). We denote this exclusion disc by b(0, P̄jk Wj ).
Assuming association with the j th tier, the total interference
experienced by the typical user originates from two independent sets of BSs: (i) ∪k∈K1 Φk \ b(0, Wj ), the set of
open access BSs lying beyond the exclusion zone b(0, Wj )
and (ii) ∪k∈K Φ0k , the set of closed access BSs. As all the
interferers from the k th open access tier will lie outside
b(0, P̄jk Wj ), we defineP
interference from the k th open-access
−α
tier as Io(j,k) (Wj ) =
. We
xk ∈Φk \b(0,P̄jk Wj ) Pk hk kxk k
express the total contribution of interference from all open
access tiers as
Io(j) (Wj ) =
K
X
Io(j,k) (Wj ).
k=0
It is clear that the interference from the open-access tiers
defined above depends on the serving distance Wj . However,
Further, we obtain closed-form expressions of fWj |V0 (·) it is not the case with the closed access tiers. Recall that since
for Thomas and Matérn cluster processes by putting the the closed access tiers do not participate in the cell selection
corresponding PDFs and CCDFs in the following Corollaries. procedure, there is no exclusion zone in their interference
Corollary 3. If Φui is Thomas cluster process, conditional field. In particular, the closed access BSs may lie closer to
the typical user than its serving
PK BS. We denote the closed
PDF of serving distance given V0 = v0 can be expressed as
access
interference
by
I
=
c
k=1 Ic(k) , where Ic(k) is the
K
2
P
2πλj
interference
from
all
the
BSs
of
the
k th closed access tier Φ0k .
2
fWj |V0 (wj |v0 ) = Aj|v exp − π
P̄jk λk wj wj , ∀j ∈ K1 .
0
Using the variables defined above, we can now express SIR
k=0
(12) defined in Eq. 1 at the typical user when it is served by the
BS located at a distance Wj in a compact form as a function
Proof. See Appendix D.
Proof. See Appendix E.
of the RV Wj as: SIR(Wj ) =
Pj hj Wj−α
Io(j) (Wj )+Ic .
7
A. Coverage Probability
B. Laplace Transform of Interference
A typical user is said to be in coverage if SIR(Wj ) > τ ,
where τ denotes modulation-coding specific SIR threshold
required for successful reception. The coverage probability can
now be formally defined as follows.
As evident from Theorem 1, the Laplace transform of interference from different tiers are the main components of the
coverage probability expression. The following three Lemmas
deal with the Laplace transforms of the interference from
different tiers. We first focus on the interference originating
from all the open access tiers except the interference from the
0th tier (i.e. the BS at cluster center) which requires separate
treatment.
Definition 2 (Coverage probability). Per-tier coverage probability for Φj can be defined as the probability that the typical
user of Φui is in coverage conditioned on the fact that it is
served by a BS from Φj . Mathematically,
(i)
Pc j = E(1(SIR(Wj ) > τ )).
(13)
The total coverage probability Pc (i) can now be defined in
terms of the per-tier coverage probability as
Pc (i) =
K
X
(i)
Aj Pc j ,
(14)
j=0
where Aj is given by Eq. 7.
With the expressions of Aj|V0 and fWj |V0 (·) at hand, we
focus on the derivation of coverage probability Pc (i) . Note that
using the Rayleigh fading assumption along with the fact that
the open access interference terms {Io(j,k) } and the closed
access interference terms {Ic(k) } are all independent of each
other, we can express the per-tier coverage probability in terms
of the product of Laplace transforms of these interference
terms. This result was presented for a special case of Thomas
cluster process in the conference version of this paper [1]
(for K-tier HetNets) as well as in [43] (for single-tier cellular
networks).
Theorem 1 (Coverage probability). Conditional per-tier coverage probability of the typical user from Φui given that the
(i)
serving BS being from the j th tier and V0 = v0 is: Pc j|v0 =
Z
Y
K
α
α
τ wjα
τw
τw
LIo(j,k) Pij LIc(k) Pjj
|v0
Pj
k=1
τ N0 wjα
× exp −
(15)
fWj |V0 (wj |v0 ) dwj ,
Pj
LIo(j,0)|V0
wj >0
and the coverage probability of a typical user from Φui can be
expressed as
K
X
(i)
Pc (i) = EV0
Aj|V0 Pc j|V0 ,
(16)
j=0
where LIo(j,0)|V0 (s|v0 ) is the conditional
Laplace transform
of Io(j,0) , i.e., LIo(j,0)|V0 (s|v
)
≡
E
exp(−sI
0
o(j,0) )|V0 = v0
and
L
(s)
≡
E
exp(−sI
)
; LIc(k) (s) ≡
I
o(j,k)
o(j,k)
E exp(−sIc(k) ) respectively denote the Laplace transforms
of interference of the k th open and closed access tiers (k ∈ K).
Proof. See Appendix F.
Note that the conditioning on V0 appears only in the first
term, i.e. the Laplace transform of Io(j,0) since the interference
from the BS at cluster center is only influenced by V0 while
the other interference terms are independent of V0 .
Lemma 3. Given a typical user of Φui is served by a BS ∈ Φj
(j ∈ K) at a distance Wj = wj , Laplace transform of Io(j,k) ,
τ wα
∀k ∈ K, evaluated at s = Pjj is
α
τ wj
2
LIo(j,k)
λk G(α, τ )wj2 ,
(17)
= exp −π P̄jk
Pj
2τ
2
2
with G(α, τ ) =
; 2 − , −τ , (18)
2 F1 1, 1 −
α−2
α
α
where 2 F1 [a, b, c, t] =
Γ(c)
Γ(b)Γ(c−b)
sian Hypergeometric function.
R1
0
z b−1 (1−z)c−b−1
dz
(1−tz)a
is Gaus-
Proof. The proof follows on the same lines as [10, Theorem
1]. For completeness, the proof is provided in Appendix G.
After dealing with the interference from all open access tiers
Φk ∀k ∈ K, we now focus on the 0th tier, which consists of
only the cluster center.
Lemma 4. Given a typical user of Φui connects to the BS
∈ Φj with j ∈ K at a distance Wj = wj , the Laplace
τ wα
transform of Io(j,0) at s = Pjj conditioned on V0 = v0
τ wα
is: LIo(j,0)|V0 Pjj v0 =
Z
fY0 (y0 )
1
1
y0 >v0α P̄ji wj
1+τ
y0
1
−α
1
dy0 . (19)
F Y0 (v0α P̄ji wj )
v0α P̄ji wj
Proof. Recall that since the 0th tier is created by the BS
at cluster center which actually belongs to the ith openaccess tier, the transmit power P0 ≡ Pi . If the serving
BS ∈ Φj (j ∈ K) lies at a distance Wj = wj , due to
the formation of virtual exclusion zone around the typical
user, the cluster center acting as an interferer will lie outside
b(0, P̄ji wj ). Thus, the PDF of distance from the typical user to
1
cluster center conditioned on Y0 > v0α P̄ji wj is fY0 (y0 |Y0 >
1
1
f (y )
Y01 0
, where y0 > v α P̄ji wj . The conv0α P̄ji wj ) =
0
F Y0 v0α P̄ji wj
ditional Laplace transform LIo(j,0)|V
(s|v0 ) can be expressed
0
as: EY0 Eh0 exp − sPi h0 v0 Y0−α |R0 > P̄ji wj
1
1
(a)
α
= EY0
|Y0 > v0 P̄ji wj
(20)
1 + sPi v0 Y0−α
Z
1
1
α
=
−α fY0 (y0 |Y0 > v0 P̄ji wj ) dy0
1 + sPi v0 y0
1
y0 > v0α P̄ji wj
8
Z
=
1
α
y0 > v0 P̄ji wj
1
fY0 (y0 )
1
dy0 ,
−α
1 + sPi v0 y0 F Y v α P̄ji wj
0
0
Corollary 6 (No shadowing). Under the assumption of no
shadowing, the coverage probability of a typical user belonging to Φui can be expressed as
where (a) follows from h0 ∼ exp(1). This completes the proof.
Pc
Proof. Following from Eq. 20:
LIo(j,0)|V0 (s|v0 ) = EY0 1+sP 1v
1
1 + sPi v0 y0−α
|Y0 > v0 P̄ji wj
y0 →∞
K
X
τ N0 w0α
2
−π
P̄0k
×
P0
k=1
w0 >0
0
λk (G(α, τ ) + 1) + λk H(α, τ ) w02 fY0 (w0 ) dw0 ,
(i)
Pc 0 =
Z
1
A0
−
exp
(25)
(i)
Pc j =
Z
2πλj
Aj
exp
−
τ N0 wjα
K
X
−π
Pj
2
P̄jk
×
k=1
wj >0
λk (G(α, τ ) + 1) + λ0k H(α, τ ) wj2
Z
fY0 (y0 )
dy0 wj dwj ,
×
1 + τ ( P̄jiy0wj )−α
1
.
(26)
given by:
k=0
k6=j
y0 =v0α P̄ji wj
Note that the PDF and CCDF of Yk for k ∈ K can be
obtained by replacing λk (λ0 k ) by λk (λ0k ) in Eq. 2.
C. Bounds on Coverage Probability
0
1
= 1.
1 + sPi v0 y0−α
Lemma 5. Given the typical user of Φui connects to any BS
∈ Φj at a distance Wj = wj , ∀j ∈ K1 , the Laplace transform
τ wα
of Ic(k) at s = Pjj is
α
τ wj
LIc(k)
= exp − πλ0 k H(α, τ )(P̄jk wj )2 , (22)
Pj
In this section, we derive upper and lower bounds on
coverage probability Pc (i) by using the results obtained in
Corollary 5.
Proposition 1 (Bounds on Coverage). The conditional per-tier
(i),L
coverage probability for j ∈ K can be bounded as Pc j|V0 ≤
(i),U
(i)
Pc j|V0 ≤ Pc j|V0 , where
Z
(i),U
Pc j|V0 =
exp
−
τ N0 wjα
Pj
Y
K
LIo(j,k)
× fWj |V0 (wj |v0 )dwj ,
2π
2/α 2π csc( α )
α
.
(23)
Proof. The proof of this fairly well-known result follows in
the same lines as that of Lemma 3, with the only difference
being the fact that Ic(k) is independent of Wj and hence, the
lower limit of the integral in Eq. 36 will be zero. The final form
can be obtained by some algebraic manipulations and using
the properties of Gamma function [44, Eq. 3.241.2].
The expressions of Laplace transforms of interference derived in the above three Lemmas are substituted in Eq. 15
to get the coverage probability. The results for no shadowing
is readily obtained by putting Vk ≡ 1, which omits the final
deconditioning step with respect to V0 .
(i),L
Pc j|V0
τ wjα
Pi
LIc(k)
τ wjα
Pj
k=1
wj >0
where
H(α, τ ) = τ
(24)
with
where Aj is the association probability to Φj
K
Q
F Yk (P̄jk Yj ), ∀k ∈ K1 .
Aj = EYj
τ wα
≤ lim
(i)
Aj Pc j ,
(BS,o)
Substitution of s = Pjj gives the final result. The upper
bound can be obtained by
1
LIo(j,0)|V0 (s|v0 ) =EY0 1+sP 1v Y −α |Y0 > v0α P̄ji wj
i 0
+
K
X
y0 >P̄ji wj
1
α
−α
i 0 Y0
≥
=
(i)
A0 Pc 0
j=1
In the next Corollary, we provide closed form upper and
lower bounds on the Laplace transform of interference from
the BS at 0th tier. The lower bound is obtained by placing
the BS located at the cluster-center of the typical user on the
boundary of the exclusion disc b(0, P̄ji wj ). The upper bound
is found by simply ignoring the interference from this BS.
These bounds will be used later in this section to derive tight
bounds on coverage probability.
Corollary 5. Conditional Laplace transform of Io(j,0) given
τ wα
V0 = v0 at s = Pjj is bounded by
α
τ wj
1
≤ LIo(j,0)|V0
v0 ≤ 1.
(21)
1+τ
Pj
(i)
1
=
1+τ
Z
−
exp
τ N0 wjα
Pj
Y
K
LIo(j,k)
τ wjα
Pi
k=1
wj >0
× LIc(k)
(27)
and
τ wjα
Pj
fWj |V0 (wj |v0 )dwj .
(28)
Hence, from Eq. 16, coverage probability Pc (i) can be bounded
by
Pc (i),L ≤ Pc (i) ≤ Pc (i),U ,
where
(i)
Pc (i),L = EV0 A0|V0 Pc j|V0 +
K
X
j=1
(i),L
Aj|V0 Pc j|V0 ],
9
(i)
Pc (i),U = EV0 A0|V0 Pc j|V0 +
K
X
Proposition 4 (Bounds on Coverage: Matérn cluster process).
For an interference limited network (N0 = 0), when Φui is
Matérn cluster process, Pc (i) can be bounded by
(i),U
Aj|V0 Pc j|V0 ].
j=1
(i)
Proof. Using Corollary 5, bounds on Pc j|V0 can be directly
obtained by substituting the bounds on LIo(j,0)|V0 (·) from
Eq. 21 in Eq. 15.
Remark 5. The intuition behind the upper and lower bound
on coverage probability is underestimating and overestimating
the interference from the BS at cluster center when the typical
user does not connect to it (refer to Corollary 5). Given
that the user connects to some tier j ∈ K, no BS including
that at the cluster center (equivalently the BS of tier 0) must
lie beyond the exclusion disc of radius P̄ji wj , wj being the
serving distance. Upper bound on coverage will be obtained if
the interfering BS at cluster center is pushed away to infinity
and lower bound is obtained if it is assumed to be located on
the boundary of the exclusion disc.
For no shadowing, we can write simpler expressions for the
upper and lower bound of Pc (i) . This result is presented in the
following Proposition.
Proposition 2 (Bounds on Coverage: No Shadowing).
can be bounded by
(i)
Pc j
2πλj
2πλj
(i)
Hj ≤ Pc j ≤
Hj ,
(29)
Aj (1 + τ )
Aj
R
τ N0 wα
where Hj =
exp − Pj j − wj2 F Y0 (P̄j0 wj ) wj dwj .
wj >0
The upper and lower bounds on Pc (i) can be obtained by
(i)
substituting Pc j with its upper and lower bounds in Eq. 24.
It can be readily observed from Proposition 1 and 2 that
the bounds on per-tier coverage probability (for j ∈ K) are
simplified expressions due to the elimination of one integration
by bounding LIo(j,0) (·). In the following Propositions, we
present closed form bounds on coverage probability under no
shadowing for Thomas and Matérn cluster processes in an
interference limited network. The tightness of the proposed
bounds will be investigated in Section V-C.
Proposition 3 (Bounds on Coverage: Thomas cluster process).
For an interference limited network (N0 = 0), when Φui is
Thomas cluster process, Pc (i) can be bounded by
K
K
X
λ0
1 X λj
λj
+
≤ Pc (i) ≤
,
M0
1 + τ j=1 Mj
Mj
j=0
where Mj = λ0 +
K
P
k=1
(i)
Pc 0
(30)
P0 +
K
K
X
1 X
Pj ≤ Pc (i) ≤ P0 +
Pj ,
1+τ
k=1
(31)
k=1
(i)
where Pj = Aj Pc j which can be obtained by replacing Zj by
K
P
2
π
P̄jk
(λk (G(α, τ ) + 1) + λ0k H(α, τ )), λk (λk ) by λk (λ0k )
k=1
and putting V0 ≡ 1 in the expression of conditional association
probability when Φui is a Matérn cluster process (Eq. 10).
Proof. The proof follows the similar lines of the previous one,
except the substitution of fY0 (·) and F Y0 (·) by the PDF and
CCDF of Y0 for Matérn cluster process mentioned in Eq. 5a
and Eq. 5b respectively. The integrals will be exactly in the
similar forms as those appearing in the proof of Corollary 2
and the final result follows on the same line of the proof.
D. Asymptotic Analysis of Coverage
In this section, we examine the limiting behaviour of the
coverage probability expressions with respect to the cluster
size. As the cluster size increases, the typical user is pushed
away from the cluster center, which reduces its association
probability with the BS located at its cluster center. Also the
interference and coverage provided by the BS at the cluster
center will be diminished due to reduced received signal power
from this BS.
As the cluster size increases, let us assume that the distance
of the typical user from the cluster center Y0 is scaled to Z =
ζY0 , where ζ > 1 is the scaling factor. Then, the PDF of Z
is fZ (z) = ζ1 fY0 (z/ζ). The scaling of the distance from the
cluster center and increasing the cluster size are equivalent,
for instance, when Φui is a Matérn cluster process, 0 < Y0 <
Ri implies that 0 < Z < ζRi . When Φui is Thomas cluster
process, since the users have a Gaussian distribution around
cluster center, 99.7% of the total users in cluster will lie within
a disc of radius 3 σi . Thus σi can be treated as the metric of
cluster size and σi scales with ζ. In the following lemma,
we investigate the limiting nature of coverage as cluster size
goes to infinity. To retain the simplicity of expressions, we
restrict the following analysis for interference limited networks
(N0 = 0). However, this can be easily extended for SINRbased coverage probability without much effort.
Lemma 6 (Convergence). If distance of the a typical user and
cluster center Y0 is scaled by ζ (ζ > 1), then the following
limit can be established: lim Pc (i) =
ζ→∞
2
P̄jk
(λk (G(α, τ ) + 1) + λ0k H(α, τ )).
Proof.
can be obtained by substituting fY0 (·) by Eq. 4a
in Eq. 25. This gives the first term λ0 /M0 in the expressions
(i),L
(i),U
of the two bounds. Pc j
and Pc j
can be obtained by
substituting F Y0 (·) from Eq. 4b to Eq. 29. The result follows
from evaluation of the integrals.
∆
Pc (PPP) =
K
X
j=1
K
P
λj
.
P̄jk λk (G(α, τ ) + 1) + λ0 k H(α, τ )
k=1
(32)
Proof. See Appendix H.
10
The results so far are concerned with the users belonging
to Φui . Recall that in our system model we considered that
the users form a mixed point process consisting of Φui (i ∈
B) and Φu(PPP) . So the overall coverage probability will be
a combination of all these individual coverage probabilities
Pc (i) (i ∈ B) and also Pc (PPP) corresponding to the users
in Φu(PPP) , which are distributed independently of the BS
locations. The overall
userpoint process can be expressed as
S u
Φu ≡ Φu(PPP) ∪
Φi . The average number of points of
i∈B
Φu in any given set A ⊂ R2 is given by
X
E(Φu (A)) = E(Φu(PPP) (A)) +
E(Φui (A)),
i∈B
where E(Φu(PPP) (A)) = λ(PPP) A, E(Φui (A)) = Ni λi A. To
avoid notational complication, we use the symbol Φ to denote
a point process as well as the associated counting measure.
Since each point has an equal chance to be selected as location
of the typical user, the probability that a randomly chosen
user from Φu belongs to Φu(PPP) (Φui ), denoted by p0 (pi )
respectively, is
(PPP)
p0 =
Ni λ i
λ
P
and pi = (PPP) P
,
λ(PPP) +
λ
+
Nj λj
Nj λj
j∈B
j∈B
where Ni is the average number of users per cluster of Φui
(i ∈ B). Now, using these probabilities p0 and pi , the overall
coverage probability is formally stated in the next Theorem.
Theorem 2 (Overall Coverage Probability). Overall coverage
probability with respect to any randomly chosen user in a Ktier HetNet with mixed user distribution is:
X
Pc = p0 Pc (PPP) +
pi Pc (i) ,
(33)
i∈B
(PPP)
where Pc
respectively.
and Pc
(i)
are given by Eq. 32 and Eq. 16,
V. N UMERICAL R ESULTS AND D ISCUSSIONS
A. Validation of Results
In this section, the analytical results derived so far are
validated and key insights for the new HetNet system model
with users clustered around BSs are provided. For the sake
of concreteness, we restrict our simulation to two tiers: one
(BS,o)
macrocell tier (Φ1
) with density λ1 with all open access
(BS)
BSs, and one small cell tier (Φ2 ) with a mix of open
(BS)
and closed access BSs. For Φ2 , the open and closed
access BS densities are λ2 and λ02 , respectively. We choose
λ2 = λ02 = 100λ1 = 100 BSs per π(500)2 m2 . We assume
Coverage probability, Pc
E. Overall Coverage Probability
1
0.9
σ2 =10
0.8
σ2 =20
0.7
σ2 =40
0.6
0.5
0.4
0.3
0.2
0.1
0
-50
η2=8dB, Theorem 1
η2=8dB, Simulation
η2=4dB, Theorem 1
η2=4dB, Simulation
η2=0dB, Theorem 1
η2=0dB, Simulation
PPP, Lemma 6
PPP, Simulation
-40
-30
-20
-10
0
10
20
30
40
50
SIR Threshold, τ (in dB)
(a) Users in Thomas cluster process
1
R2=10
0.9
Coverage probability, Pc
Remark 6. In Lemma 6, we formally claim that, irrespective
of the distribution of Y0 , if the size of the cluster is expanded,
the total coverage probability Pc (i) converges to Pc (PPP) , i.e.,
the coverage probability obtained for a typical user under the
assumption of PPP distribution of users independent of the BS
point processes, which is derived in [10].
0.8
R2=20
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-50
R2=40
η2=8dB, Theory
η =8dB, Simulation
2
η2=4dB, Theory
η2=4dB, Simulation
η2=0dB, Theory
η2=0dB, Simulation
PPP, Theory
PPP, Simulation
-40
-30
-20
-10
0
10
20
30
40
50
SIR Threshold (in dB)
(b) Users in Matérn cluster process
Fig. 2: Comparison of coverage probabilities with cluster size
for various shadowing environments. The baseline case when the
user distribution is a PPP is also included. The lines and markers
correspond to the analytical and simulation results, respectively.
the transmit powers are related by P1 = 103 P2 . The user
process is considered to be Φu2 only, i.e., a PCP around
(BS)
Φ2 . For every realization, a BS in the ith tier is randomly
selected and location of a typical user is generated according
to the density function of (i) Thomas cluster process (Eq. 4),
and (ii) Matérn cluster process (Eq. 5). For shadowing, we
have chosen log-normal distribution parameters as µk = 0,
ηk = 8 dB, 4 dB and 0 dB (no shadowing) for all k = 0, 1, 2.
In Fig. 2, the coverage probability (Pc , equivalently Pc (2) ) is
plotted for different values of SIR threshold τ and cluster size
(i.e. different σ2 -s for Thomas and R2 -s for Matérn cluster
processes) for an interference limited network (N0 = 0).
The validity of this assumption will be justified in the next
subsection. It can be observed that the analytically obtained
results exactly match the simulation results. For comparison,
Pc (PPP) , i.e., the coverage probability assuming homogeneity
of users (i.e., independent PPP assumption) is also plotted. The
plots clearly indicate that under clustering, Pc is significantly
higher than Pc (PPP) and increases for denser clusters. Also
the convergence towards Pc (PPP) is evident as cluster size
increases. In Fig. 3, the association probabilities are plotted
for different cluster size with ηk = 4 dB. The figure clearly
11
1
1
Association to Tier 1
Association to Cluster center
Association to Tier 2
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.8
0.7
0.6
0.5
0.4
0.3
0.1
0
10
0.2
20
30
40
50
60
70
80
30
40
50
60
70
80
2
1
1
0.9
0.9
Coverage probability, Pc
Association probability
20
(a) Users in Thomas cluster process
(a) Users in Thomas cluster process
0.8
0.7
0.6
0.4
10
Cluster size (σ )
Cluster Size (σ2)
0.5
Exact, Corollary 6
Lower Bound, Proposition 3
Upper Bound, Proposition 3
Lemma 6
0.9
Coverage probability, Pc
Association probability
0.9
Association to Tier 1
Association to Cluster center
Association to Tier 2
0.3
0.2
Exact, Corollary 6
Lower Bound, Proposition 4
Upper Bound, Proposition 4
Lemma 6
0.8
0.7
0.6
0.5
0.4
0.3
0.1
0
10
20
30
40
50
60
70
80
Cluster Size (R2)
0.2
10 20 30 40 50 60 70 80 90 100 110 120 130
Cluster size (R2 )
(b) Users in Matérn cluster process
(b) Users in Matérn cluster process
Fig. 3: Comparison of the association probabilities to the two tiers
and the cluster center.
illustrates that a user is more likely to be served by its
cluster center if the distribution is more “dense” around the
cluster center. As the cluster expands, association probability
to the BS at cluster center (equivalently the 0th tier) decreases
whereas the association probabilities to the other open access
tiers increase.
B. Effect of Thermal Noise
In this subsection, we investigate the effect of thermal noise
on the coverage probability in the two-tier setup described in
the previous section. In order to do this, we need to first fix
a realistic reference point relative to which the noise variance
N0 will be decided. For that we choose the reference signalto-noise ratio observed at the cell edge of a macrocell. Fixing
this value to say 0 dB we can then calculate the noise variance
N0 using the same procedure that we used in [9, Section VA]. Plugging this value in the theoretical results, we compare
the coverage probability obtained under this setup with its nonoise counterpart under no shadowing in Fig. 5. As expected,
it is observed that the noise does not have any noticeable effect
on the coverage probability due to which we will simply ignore
the effect of noise in the rest of this section.
Fig. 4: Inspection of the proposed closed form bound with variation
of cluster size for constant SIR threshold, τ = 0 dB
C. Tightness of the Bounds
In Proposition 1, we derived upper and lower bounds on
Pc (i) . We found that for no shadowing, these bounds reduce
to closed form expression when Φui is Thomas or Matérn
cluster process (Propositions 3 and 4). In Fig. 4, we plot these
upper and lower bounds on Pc . Recall that the lower bound
was obtained by placing the BS of the cluster-center (in the
representative cluster) on the boundary of the exclusion disc
when the typical user connects to other BSs and the upper
bound was found by simply ignoring the interference from
this BS (see Corollary 5 for details). We observe that the lower
bound becomes loose as the cluster size increases and for large
user clusters, Pc (PPP) becomes tighter lower bound. This is
because the interference from the cluster center is significantly
overestimated by placing the BS of the cluster center at the
boundary of the exclusion zone. The upper bound remains tight
for the entire range of cluster sizes. This can be explained by
looking at the cases of small and large clusters separately. For
small clusters, the typical user will likely connect to the BS at
its cluster center most of the time and hence the interference
term in question (Laplace transform of interference from the
cluster center; see Corollary 5) will not even appear in the
coverage probability expression. On the other hand, for large
12
1
1
0.8
σ 2 =20
σ 2 =40
0.6
0.4
0.2
Simulation
Theorem 1
Theorem 1 (w/o noise)
0
-50 -40 -30 -20 -10
0
10
20
30
40
Coverage probability, P c
Coverage probability, P c
R2 =40
0.4
Simulation
Theorem 1
Theorem 1 (w/o noise)
0
10
20
0.8
σ2=40
PPP
0.7
0.6
1
R2 =20
0
-50 -40 -30 -20 -10
σ2=30
-80
-60
-40
-20
0
20
0
20
(a) Users in Thomas cluster process
R2 =10
0.2
σ2=20
P2 /P 1 (in dB)
1
0.6
0.9
0.5
-100
50
SIR Threshold (in dB)
(a) Users in Thomas cluster process
0.8
σ =10
2
Coverage probability, P c
Coverage probability, P c
σ 2 =10
30
40
50
SIR Threshold (in dB)
(b) Users in Matérn cluster process
Fig. 5: Comparison of coverage probabilities with and without
thermal noise under no shadowing.
clusters, the interference from the BS at the cluster center
of the representative cluster will be negligible compared to
the other interference terms due to large distance between the
typical user and this BS.
D. Power Control of small cell BSs
If Φu is a PPP independent to BS locations, then Pc (PPP) is
independent of the BS transmission power and it predicts that
no further gain in coverage can be achieved by increasing
P2 /P1 (for interference-limited HetNet consisting of open
access BSs under the assumption that the target SIR is the
same for all the tiers) [9]. In the typical two tier HetNet setup
described in Section V-A, we set the density of closed access
tiers, λ01 = λ02 = 0 and vary P1 keeping P2 constant and plot
Pc in Figs. 6a and 6b and fix τ = 0 dB. It is evident that
Pc improves significantly with P2 /P1 . In the figures, we can
identify three regions of Pc : (i) For lower value of P2 /P1 , Pc
is close to Pc (PPP) , (ii) Pc is enhanced as P2 /P1 increases
since the user is likely to be served by the cluster center, (iii)
if P2 /P1 is further increased, Pc is saturated since association
probability to other BSs will diminish. Again, the gain of Pc
is stronger for denser clusters. Thus, coverage gain can be
R2=10
0.9
R2=20
0.8
R2=40
R2=30
PPP
0.7
0.6
0.5
-100
-80
-60
-40
-20
P2 /P 1 (in dB)
(b) Users in Matérn cluster process
Fig. 6: Effect of increasing small cell power on coverage.
harnessed by increasing the transmit powers of small cell BSs
in a certain range.
VI. C ONCLUSION
While random spatial models have been used successfully
to study various aspects of HetNets in the past few years,
quite remarkably all these works assume the BS and user
distributions to be independent. In particular, the analysis is
usually performed for a typical user whose location is sampled
independently of the BS locations. This is clearly not the
case in current capacity-driven user-centric deployments where
the BSs are deployed in the areas of high user density. This
paper presented a comprehensive analysis of such user-centric
HetNet deployments in which the user and BS locations are
naturally correlated. In particular, modeling the user locations
as a general Poisson cluster process, with BSs being the cluster
centers, we have developed new tools leading to tractable
results for the downlink coverage probability of a typical user.
We have specialized the results for the case of Thomas cluster
process in which the users are Gaussian distributed around
BSs, and Matérn cluster process where the users are uniformly
distributed inside a disc centered at the BS. We have also
examined the bounds and the limiting nature of the coverage
13
probability as cluster size goes to infinity. We have derived
the overall coverage probability for a mixed user distribution
containing users uniformly distributed and clustered around
small cell BSs. Overall, this work opens up a new dimension
in the HetNet analysis by providing tools for the analysis of
non-uniform user distributions correlated to the BS locations.
This work has numerous extensions. From the system model
side, one can perform measurement campaigns to characterize
the nature of different user clusters at hotspot locations, such as
restaurants, sports bars, and airports. Various cluster process
models can then be fitted to this real-world data to obtain
accurate user location models, which can then be used for
more accurate performance analyses. From the analytical point
of view, an immediate extension is to perform the rate analysis
and study the effect of traffic offloading from macrocells
to small cells in the current setup. Also, in this work, we
assumed the BS locations to be independent from each other.
This may not always be the case. For instance, small cells,
such as picocells, may not be deployed close to macrocells.
Such dependencies have been modeled recently in [45], [46]
by modeling the BS distribution asusing Poisson Hole Process [47]. Also, the smallcells may be densely deployed in
user hotspot zones and the spatial distribution can be modeled
by PCP [48]. Other considerations, such as device-to-device
(D2D) communication in clusters can also be incorporated in
this model [49].
A PPENDIX
A. Proof of Lemma 1
According to the definition of SΦj in Eq. 6, we can write
from Eq. 7,
\
1 Rk > P̄jk Rj |V0 = v0
Aj|v0 = ER
(a)
= ERj
When j = 0, from Eq. 8, we get,
Z
K
Y
−1
A0|v0 =
F Rk (P̄0k v0 α y0 ) fY0 (y0 ) dy0 .
y0 >0 k=1
−1
Substituting fY0 (y0 ) from Eq. 4a and F Rk (P̄0k v0 α y0 )
from Eq. 2b, we get,
Z∞
K
y0
y2
P
−2
A0|v0 = exp − π
λk P̄0k v0 α y02 2 exp − 02 dy0
σi
2σi
k=1
0
2
v0α
2πσi2
=
2
v0α
2πσi2
+
P Rk > P̄jk Rj |v0 = ERj
k=1
2
vα
k=1
k6=j
0
Z∞
=
0
1
(P̄j0 v0α rj )2
exp −
2σi2
exp(−2πλj rj2 ) rj drj =
!
−π
exp
K
X
2 2
λk P̄jk
rj
2πλj ×
k=1
k6=j
2
P̄j0
2πσi2
+
λj
K
P
k=1
=
2 λ
P̄jk
k
The last step was derived by putting λ0 =
K
Y
F Rk (P̄jk Rj |vo ),
λj
K
P
k=0
2
v0α
2πσi2
.
2 λ
P̄jk
k
.
Similar to Corollary 1, for j = 0, plugging Eq. 2b and
Eq. 5a in Eq. 8, we get,
(34)
A0|v0
where (a) comes from the fact that Φk -s are independent, hence
are Rk -s. For the rest of the proof, we need to consider the
two cases of j = 0 and j 6= 0 separately. Note that only the
RV R0 among all Rj -s is the function of V0 . For case 1: j = 0
Aj|v0 = ER0
.
2 λ
P̄0k
k
0
Putting λ0 = 2πσ
2 , we get the desired result. Note that P̄00 =
i
1. For j 6= 0,
Z∞
K
Y
1
Aj|v0 = F Y0 (v0α P̄j0 Rj )
F Rk (P̄jk rj )fRj (rj ) drj
k=0
k6=j
k=0
k6=j
K
Y
K
P
C. Proof of Corollary 2
k∈K1 \{j}
K
Y
B. Proof of Corollary 1
F Rk (P̄jk R0 |v0 ) = EY0
K
Y
−1
F Rk (P̄0k v0 α Y0 ),
k=1
k=1
and for case 2: j ∈ K,
Y
K
Aj|v0 = ERj
P(Rk > P̄jk Rj |v0 )
k=0
k6=j
ZRi
K
X
2
2y0
2 −α 2
= exp − π
λk P̄0k v0 y0
dy0
R2i
k=1
0
2
α
v
= 20
1 − exp
R i Z0
K
P
where Z0 = π
k=1
k6=j
,
Eq. 2b and Eq. 5b in Eq. 7 and proceeding according to the
proof of Corollary 1, we get,
Ri
1
P̄j0 v0α
Z
K
Y
−1
= ERj P(v0 α Y0 > P̄j0 Rj )
P(Rk > P̄jk Rj )
K
Y
1
= ERj F Y0 (v0α P̄j0 Rj )
F Rk (P̄jk Rj ) .
−
−2
v0 α Z0 R2i
2
λk P̄0k
. Now for j ∈ K, using Eq. 2a,
k=1
exp −π
Aj|v0 = 2πλj
k=1
k6=j
K
X
!
2 2
λk P̄jk
rj
k=1
0
2
=
πλj
−
Zj
2 α
λj π P̄j0
v0
2
2
Ri Zj
where Zj = π
K
X
k=1
Zj R2i
1 − exp −
,
2
2 vα
P̄j0
0
2
λk P̄jk
for j ∈ K.
1
R2i −(P̄j0 v0α rj )2
rj drj
R2i
14
D. Proof of Lemma 2
The conditional CCDF of Wj in this case is,
(i)
P(Rj >wj |V0 ,SΦj |V0 )
P(SΦj |V0 )
P[Wj > wj |V0 ] = P[Rj > wj |SΦj , V0 ] =
(a)
=
F. Proof of Theorem 1
K
Y
P(Pj Rj−α > Pk Rk−α |Rj > wj , V0 ) ,
1
Aj|V0
k=1
k6=j
where (a) follows from Eq. 34. For case 1: when j = 0, given
V0 = v0 , P[W0 > w0 |V0 = v0 ] =
1
K
Y
A0|v0
k=1
=
1
−α
P(P0 v0 Y0−α > Pk Rk−α |v0
k=0
Z∞ Y
K
1
A0|v0
Y0 > w0 )
−1
F Rk (P̄0k v0 α y0 )fY0 (y0 )
dy0 .
k=1
1
Thus, the conditional distribution of W0 is obtained by
d
(1 − P[W0 > w0 |V0 = v0 ])
fW0 |V0 (w0 |v0 ) =
dw0
K
1
Q
F Rk P̄0k w0 fY0 (v0α w0 )
1
.
= v0α k=1
A0|v0
−1
P(v0 α Y0
1
For case 2: when j ∈ K, P[Wj > wj |v0 ] = Aj|v
>
0
QK
−α
−α
P̄j0 Rj ) k=1 k6=j P(Pj Rj > Pk Rk |Rj > wj ). The rest
of the proof continues in the same line of case 2 in Lemma 1.
E. Proof of Corollary 3
The serving distance distribution when the user is served by
its own cluster center is
1
K
1
v0α Y
fW0 |V0 (w0 |v0 ) =
F Rk P̄0k w0 fY0 (v0α w0 ).
A0|v0
k=1
1
Substituting F Rk (P̄0k w0 ) from Eq. 2b and fY0 (v0α w0 ) from
Eq 4a
!
1
K
K
X
v0α Y
2
2
exp −π
λk P̄0k w0
fW0 |V0 (w0 |v0 ) =
A0|v0
k=1
k=1
1
2
v0α w0
v0α w02
.
(35)
×
exp −
σi2
2σi2
Putting λ0 as defined before, we obtain the desired result. For
other open access tiers except the 0th tier we can perform
similar steps to find fWj |V0 (wj |v0 ). Starting from Lemma 2,
fWj |V0 (wj |v0 ) =
1
Aj|v0
1
F Y0 (v0α P̄j0 wj )
K
Q
Aj|v0
exp
2 2
v0α P̄j0
wj
−
2
2σi
× exp(−πλj wj2 )wj =
−
τ N0 wjα
Pj
E exp −
exp
−π
2πλj
exp − π
Aj|v0
2
2
λk P̄jk wj 2πλj
2
2
λk P̄jk
wj wj ,
k=0
where (a) follows from substitution of fRj (·), F Rk (·), F R0 (·)
by Eq. 2a, Eq. 2b and Eq. 4b.
= v0
where (a) follows from hj ∼ exp(1), (b) is due to the
independence of the interference from open and closed access
tiers. Also note that none of the interference components
except Io(j,0) depends on V0 .
Case 2: For j = 0, no contribution due to Io(0,0) will be
accounted for. Hence,
τ N0 w0α
P(SIR(W0 ) > τ |W0 = w0 ) = exp −
Pj
α
K
K
α
Y
τ w0
τ w0 Y
LIc(k)
×
LI0(0,k)
.
Pj
Pj
k=1
k=1
G. Proof of Lemma 3
By definition, the Laplace transform of interference is
LIo(j,k) (s) = E(exp(−sIo(j,k) ))
X
= E exp − s
Pk hk kxk k−α
xk ∈Φk
\b(0,P̄jk Wj )
(a)
= EΦk
Y
xk ∈Φk
\b(0,P̄jk Wj )
Ehk exp −sPk hk kxk k−α
(b)
= EΦk
k=1
k6=j
K
X
k=1
F Rk (P̄jk wj )fRj (wj )
K
X
τ wjα
Pj Io(j,0) |V0
K
K
τ wjα X
τ wjα X
×E exp −
Io(j,k) E exp −
Ic(k)
Pj
Pj
k=1
k=1
α Y
α
K
τ wj
τ N0 wjα
τ wj
LIo(j,0)
|v0
= exp −
LIo(j,k)
Pj
Pj
Pj
k=1
α
K
Y
τ wj
×
LIc(k)
,
Pj
k=1
k6=j
2
1
(b)
= exp
v0α w0
(a)
=
Recalling the definition of Pc j in Eq. 13, we first calculate
the conditional probability, P(SINR(Wj ) > τ |Wj = wj , V0 =
v0 ) ∀j ∈ K1 . The final result can be obtained by taking
expectation with respect to Wj and V0 . For case 1: when
j ∈ K,
Pj hj wj−α
> τ |V0 = v0
P
K
P
Io(j,k) + Ic
N0 +
k=0
!
K
X
τ wjα
(a)
= E exp −
N0 +
Io(j,k) + Ic(k)
|V0 = v0
Pj
(c)
= exp
Y
1
1 + sPk kxk k−α
xk ∈Φk
\b(0,P̄jk Wj )
Z∞
− 2πλk
1
1−
1 + sPk r−α
r dr ,
(36)
P̄jk Wj
where (a) is due to the i.i.d. assumption of hk , (b) follows
from hk ∼ exp(1), (c) follows from the transformation to
15
polar coordinates and probability generating functional of
homogeneous PPP [37]. The final result can be obtained by
using the integral in [44, Eq. 3.194.1].
H. Proof of Lemma 6
As will be evident from the proof, the limiting arguments as
well as the final limit remain the same irrespective of the value
of random variable V0 . Therefore, for notational simplicity,
we provide this proof for the no shadowing scenario, which is
without loss of generality. The Euclidean distance from the
typical user to its cluster center is now Z = ζY0 . In the
(i)
expression of Pc j in Eq. 26, we are particularly interested
in inner integral which comes from the Laplace transform of
interference from the BS at 0th tier derived in Lemma 4. From
Eq. 19 with the substitution P̄ji wj = x and v0 = 1, we can
write:
"
#
Z∞
fZ (z)
1
1
dz = EZ
−α |Z > x
1 + τ ( xz )−α F Z (x)
1 + τ Zx
x
= EY0
Hence, lim
1+τ
ζ→∞
1
ζY0
x
R∞
x
ζ
−α |Y0 >
1
1+τ (
ζY0
x
−α
)
x
.
ζ
fY0 (y0 )
dy0
F Y0 ( x
ζ)
=
R∞
fY0 (y0 )dy0 = 1.
0
(i)
Thus, as ζ → ∞, the inner integral tends to 1. So, lim Pc j =
ζ→∞
K
P
2
λj /
P̄jk
λk (G(α, τ ) + 1) + λ0 k H(α, τ ) . Now we are
k=1
(i)
left with the limit of the first term A0 Pc 0 i.e., the contribution
of the 0th tier in Pc (i) which will obviously go to zero as
ζ → ∞.
R EFERENCES
[1] C. Saha and H. S. Dhillon, “Downlink coverage probability of K-Tier
HetNets with general non-uniform user distributions,” in Proc., IEEE
International Conference on Communications (ICC), May. 2016.
[2] 3GPP, “Further advancements for E-UTRA physical layer aspects,” Tech.
Rep., 2010.
[3] ——, “Consideration of UE Cluster Position and PeNB TX Power in
Heterogeneous Deployment Configuration 4,” Discussion/ Decision R1100477, Jan. 2010, 8.2.1 Relevant scenarios of Heterogeneous Networks.
[4] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for
modeling, analysis, and design of multi-tier and cognitive cellular
wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15,
no. 3, pp. 996 – 1019, 2013.
[5] J. G. Andrews, A. K. Gupta, and H. S. Dhillon, “A primer on cellular
network analysis using stochastic geometry,” arXiv preprint, 2016,
available online: arxiv.org/abs/1604.03183.
[6] S. Mukherjee, Analytical Modeling of Heterogeneous Cellular Networks.
Cambridge University Press, 2014.
[7] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage
and rate in cellular networks,” IEEE Trans. Commun., vol. 59, no. 11,
pp. 3122–3134, Nov. 2011.
[8] H. S. Dhillon, R. K. Ganti, and J. G. Andrews, “A tractable framework
for coverage and outage in heterogeneous cellular networks,” in Proc.,
Information Theory and Applications Workshop (ITA), Feb. 2011.
[9] H. Dhillon, R. Ganti, F. Baccelli, and J. Andrews, “Modeling and
analysis of K-tier downlink heterogeneous cellular networks,” IEEE J.
Sel. Areas Commun., vol. 30, no. 3, pp. 550–560, Apr. 2012.
[10] H.-S. Jo, Y. J. Sang, P. Xia, and J. Andrews, “Heterogeneous cellular
networks with flexible cell association: A comprehensive downlink SINR
analysis,” IEEE Trans. Wireless Commun., vol. 11, no. 10, pp. 3484–
3495, Oct. 2012.
[11] S. Mukherjee, “Distribution of downlink SINR in heterogeneous cellular
networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 575–585,
Apr. 2012.
[12] P. Madhusudhanan, J. G. Restrepo, Y. Liu, and T. X. Brown, “Analysis
of downlink connectivity models in a heterogeneous cellular network via
stochastic geometry,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp.
3895–3907, Jun. 2016.
[13] H. Dhillon, R. Ganti, and J. Andrews, “Load-aware modeling and
analysis of heterogeneous cellular networks,” IEEE Trans. Wireless
Commun., vol. 12, no. 4, pp. 1666–1677, Apr. 2013.
[14] S. Singh, H. Dhillon, and J. Andrews, “Offloading in heterogeneous
networks: Modeling, analysis, and design insights,” IEEE Trans. Wireless
Commun., vol. 12, no. 5, pp. 2484–2497, May 2013.
[15] W. C. Cheung, T. Quek, and M. Kountouris, “Throughput optimization,
spectrum allocation, and access control in two-tier femtocell networks,”
IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 561–574, Apr. 2012.
[16] R. W. Heath, M. Kountouris, and T. Bai, “Modeling heterogeneous
network interference using poisson point processes,” IEEE Trans. Signal
Proc., vol. 61, no. 16, pp. 4114–4126, Aug. 2013.
[17] H. ElSawy, A. Sultan-Salem, M. S. Alouini, and M. Z. Win, “Modeling
and analysis of cellular networks using stochastic geometry: A tutorial,”
IEEE Commun. Surveys Tuts., vol. PP, no. 99, pp. 1–1, 2016.
[18] A. Jaziri, R. Nasri, and T. Chahed, “System level analysis of heterogeneous networks under imperfect traffic hotspot localization,” IEEE
Trans. Veh. Technol., vol. 65, no. 12, pp. 9862–9872, Dec. 2016.
[19] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
“Five disruptive technology directions for 5G,” IEEE Commun. Mag.,
vol. 52, no. 2, pp. 74–80, Feb. 2014.
[20] M. Taranetz and M. Rupp, “Performance of femtocell access point
deployments in user hot-spot scenarios,” in Proc., Telecommunication
Networks and Applications Conference (ATNAC), Nov. 2012.
[21] Q. Ying, Z. Zhao, Y. Zhou, R. Li, X. Zhou, and H. Zhang, “Characterizing spatial patterns of base stations in cellular networks,” in IEEE
International Conference on Communications in China (ICCC), Oct.
2014, pp. 490–495.
[22] C. Chen, R. C. Elliott, and W. A. Krzymien, “Downlink coverage
analysis of n-tier heterogeneous cellular networks based on clustered
stochastic geometry,” in Proc., IEEE Asilomar, Nov. 2013, pp. 1577–
1581.
[23] Y. Zhong and W. Zhang, “Multi-channel hybrid access femtocells: A
stochastic geometric analysis,” IEEE Transa. Commun., vol. 61, no. 7,
pp. 3016–3026, Jul. 2013.
[24] Y. J. Chun, M. O. Hasna, and A. Ghrayeb, “Modeling heterogeneous
cellular networks interference using poisson cluster processes,” IEEE J.
Sel. Areas Commun., vol. 33, no. 10, pp. 2182–2195, Oct. 2015.
[25] Qualcomm Incorporated, “A Comparison of LTE Advanced HetNets and
Wi-Fi,” Tech. Rep., Oct. 2011.
[26] D. Lee, S. Zhou, X. Zhong, Z. Niu, X. Zhou, and H. Zhang, “Spatial
modeling of the traffic density in cellular networks,” IEEE Wireless
Commun., vol. 21, no. 1, pp. 80–88, Feb. 2014.
[27] M. Mirahsan, R. Schoenen, and H. Yanikomeroglu, “HetHetNets:
Heterogeneous traffic distribution in heterogeneous wireless cellular
networks,” IEEE J. Sel. Areas Commun., vol. 33, no. 10, pp. 2252–2265,
Oct. 2015.
[28] Z. Wang, R. Schoenen, H. Yanikomeroglu, and M. St-Hilaire, “The
impact of user spatial heterogeneity in heterogeneous cellular networks,”
in Proc., IEEE Globecom Workshops, Dec. 2014, pp. 1278–1283.
[29] H. Dhillon, R. Ganti, and J. Andrews, “Modeling non-uniform UE
distributions in downlink cellular networks,” IEEE Wireless Commun.
Lett., vol. 2, no. 3, Jun. 2013.
[30] C. Li, A. Yongacoglu, and C. D’Amours, “Mixed spatial traffic modeling of heterogeneous cellular networks,” in 2015 IEEE International
Conference on Ubiquitous Wireless Broadband (ICUWB), Oct. 2015.
[31] M. Taranetz, T. Bai, R. Heath, and M. Rupp, “Analysis of small
cell partitioning in urban two-tier heterogeneous cellular networks,” in
Proc., International Symposium on Wireless Communications Systems
(ISWCS), Aug. 2014, pp. 739–743.
[32] N. Deng, W. Zhou, and M. Haenggi, “Heterogeneous cellular network
models with dependence,” IEEE J. Sel. Areas Commun., vol. 33, no. 10,
pp. 2167–2181, Oct. 2015.
[33] M. Afshang, H. S. Dhillon, and P. H. J. Chong, “Fundamentals of clustercentric content placement in cache-enabled device-to-device networks,”
IEEE Trans. Commun., vol. 64, no. 6, pp. 2511–2526, Jun. 2016.
[34] ——, “Modeling and performance analysis of clustered device-to-device
networks,” IEEE Trans. Wireless Commun., vol. 15, no. 7, pp. 4957–
4972, Jul. 2016.
16
[35] Stochastic Geometry and Its Applications, 2nd ed. Chichester ; New
York: Wiley, Jul. 1996.
[36] R. Ganti and M. Haenggi, “Interference and outage in clustered wireless
ad hoc networks,” IEEE Trans. Inf. Theory, vol. 55, no. 9, pp. 4067–
4086, Sep. 2009.
[37] M. Haenggi, Stochastic Geometry for Wireless Networks. New York:
Cambridge University Press, 2013.
[38] P. Madhusudhanan, J. G. Restrepo, Y. Liu, T. X. Brown, and K. R. Baker,
“Downlink performance analysis for a generalized shotgun cellular
system,” IEEE Trans. Wireless Commun., vol. 13, no. 12, pp. 6684–
6696, Dec. 2014.
[39] M. Di Renzo, A. Guidotti, and G. Corazza, “Average rate of downlink
heterogeneous cellular networks over generalized fading channels: A
stochastic geometry approach,” IEEE Trans. Commun., vol. 61, no. 7,
pp. 3050–3071, Jul. 2013.
[40] F. Baccelli and B. Błaszczyszyn, Stochastic Geometry and Wireless
networks, Volume 1- Theory.
NOW: Foundations and Trends in
Networking, 2009.
[41] H. Dhillon and J. Andrews, “Downlink Rate Distribution in Heterogeneous Cellular Networks under Generalized Cell Selection,” IEEE
Wireless Commun. Lett., vol. 3, no. 1, pp. 42–45, Feb. 2014.
[42] H. P. Keeler, B. Błaszczyszyn, and M. K. Karray, “SINR-based kcoverage probability in cellular networks with arbitrary shadowing,” in
Proc., IEEE International Symposium on Information Theory Proceedings (ISIT), Jul. 2013, pp. 1167–1171.
[43] P. D. Mankar, G. Das, and S. S. Pathak, “Modeling and coverage analysis
of BS-centric clustered users in a random wireless network,” IEEE
Wireless Commun. Lett., vol. 5, no. 2, pp. 208–211, Apr. 2016.
[44] D. Zwillinger, Table of integrals, series, and products. Elsevier, 2014.
[45] G. Nigam, P. Minero, and M. Haenggi, “Spatiotemporal cooperation in
heterogeneous cellular networks,” IEEE J. Sel. Areas Commun., vol. 33,
no. 6, pp. 1253–1265, Jun. 2015.
[46] M. Afshang, Z. Yazdanshenasan, S. Mukherjee, and P. H. J. Chong, “Hybrid division duplex for HetNets: Coordinated interference management
with uplink power control,” in Proc., IEEE International Conference on
Communication Workshop, Jun. 2015, pp. 106–112.
[47] Z. Yazdanshenasan, H. S. Dhillon, M. Afshang, and P. H. J. Chong,
“Poisson hole process: Theory and applications to wireless networks,”
IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7531–7546, Nov.
2016.
[48] M. Afshang and H. S. Dhillon, “Poisson cluster process based analysis of
HetNets with correlated user and base station locations,” arXiv preprint,
2016, available online: http://arxiv.org/abs/1612.07285.
[49] ——, “Spatial modeling of device-to-device networks: Poisson cluster
process meets poisson hole process,” in Proc., IEEE Asilomar, Nov.
2015, pp. 317–321.
| 7cs.IT
|
A Borel maximal cofinitary group
Haim Horowitz and Saharon Shelah
Abstract
arXiv:1610.01344v1 [math.LO] 5 Oct 2016
We construct a Borel maximal cofinitary group.1
Introduction
The study of mad families and their relatives occupies a central place in modern set
theory. As the straightforward way to construct such families involves the axiom of
choice, questions on the definability of such families naturally arise. The following
classical result is due to Mathias:
Theorem ([Ma]): There are no analytic mad families.
In recent years, there has been considerable interest in the definability of several
relatives of mad families, such as maximal eventually different families and maximal
cofinitary groups. A family F ⊆ ω ω is a maximal eventually different family if
f 6= g ∈ F → f (n) 6= g(n) for large enough n, and F is maximal with respect to
this property. The following result was recently discovered by the authors:
Theorem ([HwSh1089]): Assuming ZF , there exists a Borel maximal eventually
different family.
As for maximal cofinitary groups (see definition 1 below), several consistency results
were established on the definability of such groups, for example, the following results
by Kastermans and by Fischer, Friedman and Toernquist:
Theorem ([Ka]): There is a Π11 -maximal cofinitary group in L.
Theorem ([FFT]): b = c = ℵ2 is consistent with the existence of a maximal
cofinitary group with a Π12 −definable set of generators.
Our main goal in this paper is to establish the existence of a Borel maximal cofinitary
group in ZF . We intend to improve the current results in a subsequent paper, and
prove the existence of closed MED families and MCGs.
1
Date: September 8, 2016.
2000 Mathematics Subject Classification: 03E15, 03E25, 20B27
Keywords: Borel, maximal cofinitary groups, mad families, permutation groups.
Publication 10xx of the second author.
Partially supported by European Research Council grant 338821.
1
The main theorem
Definition 1: G ⊆ S∞ is a maximal cofinitary group if G is a subgroup of S∞ ,
|{n : f (n) = n}| < ℵ0 for every Id 6= f ∈ G, and G is maximal with respect to these
properties.
Theorem 2 (ZF ): There exists a Borel maximal cofinitary group.
The rest of the paper will be dedicated for the proof of the above theorem. It will
be enough to prove the existence of a Borel maximal cofinitary group in Sym(U)
where U is an arbitrary set of cardinality ℵ0 .
Convention: Given two sequences η and ν, we write η ≤ ν when η is an initial
segment of ν.
Definition 3: The following objects will remain fixed throughout the proof:
a. T = 2<ω .
b. ū = (uρ : ρ ∈ T ) is a sequence of pairwise disjoint sets such that U = ∪{uρ : ρ ∈
T } ⊆ H(ℵ0 ) (will be chosen in claim 4).
c. <∗ is a linear order of H(ℵ0 ) of order type ω such that given η, ν ∈ T , η <∗ ν iff
lg(η) < lg(ν) or lg(η) = lg(ν) ∧ η <lex ν.
d. For every η ∈ T , Σ{|uν | : ν <∗ η} ≪ |uη |.
ω
e. Borel functions B = B0 and B−1 = B−1
0 such that B : Sym(U) → 2 is injective
ω
with a Borel image, and B−1 : 2 → Sym(U) satisfies B(f ) = η → B−1 (η) = f .
f. Let A1 = {f ∈ Sym(U) : f has a finite number of fixed points}, A1 is obviously
Borel.
g. {fρ,ν : ν ∈ 2lg(ρ) } generate the group Kρ (defined below) considered as a subgroup
of Sym(uρ).
Claim 4: There exists a sequence (uρ, f¯ρ , Āρ : ρ ∈ T ) such that:
a. f¯ρ = (fρ,ν : ν ∈ Tlg(ρ) ).
b. fρ,ν ∈ Sym(uρ) has no fixed points.
c. Āρ = (Aρ,ν : ν ∈ Tlg(ρ) ). We shall denote
∪ Aρ,ν by A′ρ .
ν∈Tlg(ρ)
d. Aρ,ν ⊆ uρ ⊆ H(ℵ0) and Σ{|uη | : η <∗ ρ} ≪ |Aρ,ν |.
e. ν1 6= ν2 ∈ Tlg(ρ) → Aρ,ν1 ∩ Aρ,ν2 = ∅.
f. If ρ ∈ 2n and w = w(..., xν , ...)ν∈2n is a non-trivial group term of length ≤ n then:
1. w(..., fρ,ν , ...)ν∈2n ∈ Sym(uρ) has no fixed points.
2. (w(..., fρ,ν , ...)ν∈2n ( ∪ n Aρ,ν )) ∩ ∪ n Aρ,ν = ∅.
ν∈2
ν∈2
g. {fρ,ν : ν ∈ Tlg(ρ) } generate the group Kρ (whose set of elements is uρ) which is
considered as a group of permutations of uρ .
2
Proof: We choose (uρ , f¯ρ , Āρ ) by <∗ -induction on ρ as follows: Arriving at ρ, we
choose the following objects:
a. n1ρ such that Σ{|uν | : ν <∗ ρ}2lg(ρ)+7 ≪ n1ρ and let n0ρ =
n1ρ
.
2lg(ρ)
b. Let Hρ be the group generated freely by {xρ,ν : ν ∈ Tlg(ρ) }.
c. In Hρ we can find (yρ,n : n < ω) which freely generate a subgroup (we can do it
explicitly, for example, if a and b freely generate a group, then (an bn : n < ω) are
as required), wlog for w1 and w2 as in 4(f) and n1 < n2 we have w1 yρ,n1 6= w2 yρ,n2 .
Now choose A1ρ,ν ⊆ {yρ,n : n < ω} for ν ∈ 2lg(ρ) such that ν1 6= ν2 → A1ρ,ν1 ∩ A1ρ,ν2 = ∅
and n0ρ ≤ |A1ρ,ν |.
d. Let Λρ = {w : w = w(..., xρ,ν , ...)ν∈Tlg(ρ) is a group word of length ≤ lg(ρ)}.
As Hρ is free, it’s residually finite, hence there is a finite group Kρ and an epimorphism φρ : Hρ → Kρ such that φρ ↾ (( ∪ A1ρ,ν ) ∪ Λρ ∪ {wa : w ∈ Λρ ∧ a ∈
∪ A1ρ,ν })
ν∈2lg(ρ)
ν∈2lg(ρ)
is injective (note that there is no use of the axiom of choice as we can
argue in a model of the form L[A]). WLOG Kρ ⊆ H(ℵ0 ) and Kρ is disjoint to
∪{uν : ν <∗ ρ}.
We now define the following objects:
a. uρ = Kρ .
b. Aρ,ν = {φρ (a) : a ∈ A1ρ,ν }.
c. For ν ∈ 2lg(ρ) , let fρ,ν : uρ → uρ be multiplication by φρ (xρ,ν ) from the left.
It’s now easy to verify that (uρ, Āρ , f¯ρ ) are as required, so U = ∪{uρ : ρ ∈ T }.
∗
Definition and claim 5: A. a. Given f ∈ Sym(U), let g = F1 (f ) be gB(f
) , where
∗
ω
∗
for ν ∈ 2 , gν is the permutation of U defined by: gν ↾ uρ = fρ,ν↾lg(ρ) (recall that ū
is a partition of U and each fρ,ν belongs to Sym(uρ), therefore g is well-defined and
belongs to Sym(U)).
b. Let G1 be the subgroup of Sym(U) generated by {gν∗ : ν ∈ 2ω } (which includes
{F1 (f ) : f ∈ Sym(U)}).
c. Let I1 be the ideal on U generated by the sets v ⊆ U satisfying the following
property:
(∗)v For some ρ = ρv ∈ 2ω , for every n, there is at most one pair (a, ν) such that
ν ∈ T , a ∈ v ∩ uν and ρ ∩ ν = ρ ↾ n.
c(1). Note that I1 is indeed a proper ideal: Suppose that v0 , ..., vn are as above and
let ρ0 , ..., ρn witness (∗)vi (i = 0, ..., n). Choose k such that 2k > n + 1 and choose
η ∈ 2k \ {ρi ↾ k : i ≤ n}. For each i ≤ n, there is k(i) ≤ k such that η ∩ρi = ρi ↾ k(i).
For each i ≤ n, let n(i) be the length ν such that (a, ν) witness (∗)vi for k(i). Choose
η ′ above η such that lg(η ′ ) > n(i) for every i, then uη′ ∩ ( ∪ vi ) = ∅.
i≤n
d. Let K1 = {f ∈ Sym(U) : f ix(f ) ∈ I1 } where f ix(f ) = {x : f (x) = x}.
3
e. For η ∈ T , a, b ∈ uη , n = lg(η) < ω and let ya,b = ((fη,ρa,b,l , ia,b,l ) : l < la,b = l(∗))
such that:
1. ρa,b,l ∈ 2n .
2. ia,b,l ∈ {1, −1}.
3. b = (fη,ρa,b,0 )ia,b,0 · · · (fη,ρa,b,l(∗)−1 )ia,b,l(∗)−1 (a).
4. la,b = l(∗) is minimal under 1-3, ya,b is <∗ -minimal under this requirement.
5. il 6= il+1 → ρa,b,l 6= ρa,b,l+1 .
By claim 4(g) and definition 3(c), ya,b is always well-defined.
B. There are Borel functions B1,1 , B1,2 , etc with domain Sym(U) such that:
a. B1,1 (f ) ∈ {0, 1} and B1,1 (f ) = 0 iff |f ix(f )| < ℵ0 .
b. Letting η1 = B(f ), B1,2 (f ) ∈ {0, 1} and B1,2 (f ) = 1 iff B1,1 (f ) = 0 and for
infinitely many n, f ′′ (A′η1 ↾n ) * ∪{uρ : ρ ≤∗ η1 ↾ n} (where A′η1 ↾n is defined in 4(c).
c. B1,3 (f ) ∈ ω such that if B1,1 (f ) = B1,2 (f ) = 0 then for every B1,3 (f ) ≤ n,
f ′′ (A′η1 ↾n ) ⊆ ∪{uρ : ρ ≤∗ η1 ↾ n}.
d. B1,4 (f ) ∈ {0, 1} and B1,4 (f ) = 1 iff B1,1 (f ) = B1,2 (f ) = 0 and {la,f (a) : a ∈ vn
and B1,3 (f ) ≤ n} is unbounded, where vn := {a ∈ A′η1 ↾n ⊆ uη1 ↾n : f (a) ∈ uη1 ↾n }.
e. B1,5 (f ) ∈ ω such that: If B1,4 (f ) = B1,2 (f ) = B1,1 (f ) = 0 then B1,5 (f ) is a
bound of {la,f (a) : a ∈ vn and B1,3 (f ) ≤ n}.
f. B1,6 (f ) ∈ {0, 1} such that: B1,6 (f ) = 1 iff B1,1 (f ) = B1,2 (f ) = B1,4 (f ) = 0 and
for every m there exists n > m such that: There are a1 6= a2 ∈ vn such that for some
l, l < min{la1 ,f (a1 ) , la2 ,f (a2 ) }, ρa1 ,f (a1 ),l 6= ρa2 ,f (a2 ),l and ρa1 ,f (a1 ),l ↾ m = ρa2 ,f (a2 ),l ↾ m.
g. B1,7 (f ) is a sequence (an = an (f ) : n ∈ B1,8 (f )) such that if B1,6 (f ) = 1 then:
1. an ∈ vn
2. B1,8 (f ) ∈ [ω]ω
3. lan ,f (an ) = l(∗) = B1,9 (f )
4. l∗∗ = B2,0 (f ) < l∗
5. (ρan ,f (an ),l∗∗ : n ∈ B1,8 (f )) are pairwise incomparable.
6. For every l < l∗ , the following sequence is constant: (T V (ρan ,f (an ),l ≤ ρak ,f (ak ),l ) :
n < k ∈ B1,8 (f )).
h. B2,1 (f ) is a sequence (An = An (f ) : n ∈ B2,2 (f )) such that if B1,1 (f ) = B1,2 (f ) =
B1,4 (f ) = B1,6 (f ) = 0 then:
1. B2,2 (f ) ∈ [ω]ω
2. An ⊆ A′η1 ↾n (recalling that η1 = B(f ))
3. la,f (a) = l∗ = B2,3 (f ) for n ∈ B2,2 (f ) and a ∈ An
4
4. (il : l < l∗ ) = (ia,f (a),l : l < l∗ ) (recalling definition 5(e)) for every n ∈ B2,2 (f )
and a ∈ An .
5.
1
B′2,3 (f )(n)
≤
|An |
|vn |
where B′2,3 (f )(n) ∈ ω \ {0}, B2,3 (f )(n) ≪ |vn | and vn is defined
in 5(B)(d).
6. B2,4,n (f ) = ρ¯∗n = (ρnl : l < l∗ ) = (ρa,f (a),l : l < l∗ ) for every n ∈ B2,2 (f ) and
a ∈ An .
7. (T V (ρnl ≤ ρm
l ) : n < m ∈ B2,2 (f )) is constantly B2,5,l (f )
i. B2,6 (f ) ∈ {0, 1} is 1 iff B1,1 (f ) = B1,2 (f ) = B1,4 (f ) = B1,6 (f ) = 0 and in (h)(7),
B2,5,l = f alse for some l < l∗ .
j. B′2,6 (f ) ∈ {0, 1} is 0 iff B1,1 (f ) = B1,2 (f ) = B1,4 (f ) = B1,6 (f ) = 0 and B2,6 (f ) =
0
Proof: By the proof of Ramsey’s theorem and the arguments which are implicit in
the proof of claim 7 below. Note that while the statement “there exists an infinite
homogeneous set” is analytic, we can Borel-compute that homogeneous set. See the
proof of claim 6 in [HwSh:1089] for more details.
Definition and claim 6: a. 1. Let H3 be the set of f ∈ Sym(U) such that:
α. B1,1 (f ) = 0
β. If B1,2 (f ) = B1,4 (f ) = B1,6 (f ) = 0 then B′2,6 (f ) = 1
a. 2. H3 is Borel.
b. For f ∈ Sym(U) let Gf be the set of g ∈ Sym(U) such that:
1. If f ∈
/ H3 then Gf = {F1 (f )}.
¯ ē, ν̄) we
2. If f ∈ H3 then Gf be the set of g such that for some (B, η1 , η2 , ā, b̄, c̄, d,
have:
A. B ⊆ ω is infinite.
B. η1 = B(f ) ∈ 2ω and η2 ∈ 2ω .
C. ā = (an : n ∈ B).
D. If n ∈ B then an ∈ A′η1 ↾n = ∪ n Aη1 ↾n,ρ ⊆ uη1 ↾n (recall that we denote ∪ n Aη1 ↾n,ρ
ρ∈2
ρ∈2
by A′η1 ↾n ).
E. b̄ = (bn : n ∈ B) and ν̄ = (νn : n ∈ B), νn ∈ T , such that for each n ∈ B,
bn = f (an ) and bn ∈ uνn . c̄ = (cn : n ∈ B), d¯ = (dn : n ∈ B) and ē = (en : n ∈ B)
are such that bn , cn ∈ uνn and en ∈ uη1 ↾n .
F. For every n ∈ B, g(an ) = f (an ) = bn .
G. For every n ∈ B, g(bn ) = F1 (f )(an ) = en .
H. For every n ∈ B we have cn = F1 (f )−1 (f (an )) and g(cn ) = F1 (f )(f (an )) = dn .
5
I. If b ∈ U is not covered by clauses F-H, then g(b) = F1 (f )(b).
J. g has no fixed points.
K. One of the following holds:
a. For every n ∈ B, η1 ↾ n <∗ νn , lg(η2 ∩ νn ) > max{lg(νm ) : m ∈ B ∩ n} hence
(νn ∩ η2 : n ∈ B) is ≤ −increasing.
b. For every n ∈ B, νn = η1 ↾ n and l(an , f (an ), n) is increasing (see definition 5(e)).
c. For every n ∈ B, νn = η1 ↾ n and in addition, l(an , f (an ), n) = l∗ for every n,
ian ,f (an ),l = il for l < l∗ and for some l∗∗ < l∗ , the elemnts of (ρan ,f (an ),l∗∗ : l∗∗ < l∗ )
are pairwise incomparable.
Claim 7: If f ∈ H3 then there exists g ∈ Sym(U) such that for some (B, η1 , η2 , ā, b̄, ν̄),
g and (B, η1 , η2 , ā, b̄, ν̄) are as required in claim 6(c)(2) (and therefore, there are also
¯ ē) as required there). Moreover, g is unique once (B, η1 , η2 , ā, b̄, ν̄) is fixed.
(c̄, d,
Remark 7A: In claim 9 we need g to be Borel-computable from f , which is indeed
the case by the discussion in the proof of claim 5 and by the proof of claim 6 in
[HwSh:1089].
Proof: f ∈ H3 , so B1,1 (f ) = 0.
We shall first observe that if g is defined as above, then g is a permutation of U
with no fixed points. It’s also easy to see that g is unique once (B, η1 , η2 , ā, b̄, ν̄) has
been chosen. Therefore, it’s enough to find (B, η1 , η2 , ā, b̄, ν̄) as required.
Case I (B1,2(f ) = 1): For infinitely many n, f ′′ (A′η1 ↾n ) * ∪{uρ : ρ ≤∗ η1 ↾ n}.
In this case, let B0 = {n : there is a ∈ A′η1 ↾n such that f (a) ∈
/ ∪{uρ : ρ ≤∗ η1 ↾
n}, and for every n ∈ B0 , let an be the <∗ -first element in A′η1 ↾n witnessing that
n ∈ B0 . Let bn = f (an ) and let νn ∈ T be the sequence for which bn ∈ uνn .
Apply Ramsey’s theorem (we don’t need the axiom of choice, as we can argue in
some L[A]) to get an infinite set B ⊆ B0 such that ck,l ↾ [B]k is constant for every
(k, l) ∈ {(2, 1), (2, 2), (2, 4), (3, 1), (3, 3)}, where for n1 < n2 < n3 :
a) c2,1 (n1 , n2 ) = T V (lg(ν1 ) < lg(ν2 )).
b) c2,2 (n1 , n2 ) = T V (νn2 ∈ {νn : n ≤ n1 }).
c) c3,1 (n1 , n2 , n3 ) = T V (lg(νn2 ∩ νn3 ) > νn1 ).
d) c3,3 (n1 , n2 ) = νn2 (lg(νn1 ∩ νn2 )) ∈ {0, 1,undefined}.
We shall prove now that (lg(νn ) : n ∈ B) has an infinite increasing subsequence:
Choose an increasing sequence n(i) ∈ B by induction on i such that j < i →
lg(νn(j) ) < lg(νn(i) ). Arriving at stage i = j + 1, suppose that there is no such n(i),
then {f (an ) : n ∈ B \ n(j)} ⊆ ∪{uρ : lg(ρ) ≤ lg(νn(j) )}, hence {f (an ) : n ∈ B \ n(j)}
is finite. Similarly, {νn : n ∈ B \ n(j)} is finite, and therefore, there are n1 < n2 ∈
B \ n(j) such that νn1 = νn2 and f (an1 ) = f (an2 ). As f is injective, an1 = an2 , and
by the choice of the an , an1 ∈ uη1 ↾n1 and an2 ∈ uη1 ↾n2 .
This is a contradiction, as uη1 ↾n1 ∩ uη1 ↾n2 = ∅.
6
Therefore, there is an infinite B ′ ⊆ B such that (lg(νn ) : n ∈ B ′ ) is increasing, and
wlog B ′ = B.
Now we shall note that if n1 < n2 < n3 are from B, then lg(νn2 ∩ νn3 ) > lg(νn1 ):
By the choice of B, c3,1 (n1 , n2 , n3 ) is constant for n1 < n2 < n3 , so it suffices to
show that c3,1 ↾ [B]3 = true. Let n1 = min(B) and k = lg(νn1 ) + 1. The sequence
(νn ↾ k : n ∈ B \ {n1 }) is infinite, hence there are n2 < n3 ∈ B \ {n1 } such that
νn2 ↾ k = νn3 ↾ k. Therefore, lg(νn1 ) < k ≤ lg(νn2 ∩ νn3 ), and as c3,1 is constant on
[B]3 , we’re done.
For n < k ∈ B such that k is the successor of n in B, let ηn = νn ∩ νk . Suppose
now that n < k < l are successor elements in B, then lg(ηk ) = lg(νk ∩ νl ) >
lg(νn ) ≥ lg(νn ∩ νk ) = lg(ηn ), and ηn , ηk ≤ νk , therefore, ηn is a proper initial
segment of ηk and η2 := ∪ ηn ∈ 2ω . If n < k ∈ B are successor elements, then
n<ω
lg(ηk ) > lg(νn ) (by a previous claim), therefore, νn ∩ (T \ νk ) is disjoint to η2 , hence
νn ∩ η2 = νn ∩ νk = ηn . Therefore, if n < k ∈ B then νn ∩ η2 is a proper initial
segment of νk ∩ η2 and η2 = ∪ (νn ∩ η2 ).
n<ω
It’s now easy to verify that (B, η1 , η2 , ā, b̄, ν̄) and g are as required.
Case II (B1,2 (f ) = 0 and n1 stands for B1,3 (f )): There is n1 such that for
every n1 ≤ n, f ′′ (A′η1 ↾n ) ⊆ ∪{uρ : ρ ≤∗ η1 ↾ n}.
For each n, recall that vn = {a ∈ A′η1 ↾n ⊆ uη1 ↾n : f (a) ∈ uη1 ↾n }. vn satisfies
|A′η1 ↾n \ vn | ≤ Σ{|uν | : ν <∗ η1 ↾ n}, and as Σ{|uν | : ν <∗ η1 ↾ n} ≪ |A′η1 ↾n |, it
follows that Σ{|uν | : ν <∗ η1 ↾ n} ≪ |vn |. Recall also that for a ∈ vn , as f (a) ∈ uη1 ↾n ,
by definition 5(e), ya,f (a) is well-defined.
We now consider three subcases:
Case IIA (B1,4 (f ) = 1): The set of la,f (a),n for a ∈ vn and n1 ≤ n is unbounded. In this case, we find an infinite B ⊆ [n1 , ω) and an ∈ vn for each n ∈ B
such that (lan ,f (an ),n : n ∈ B) is increasing. Now let η2 := η1 and define b̄, ν̄ and g as
described in Definition 6. It’s easy to verify that (B, η1 , η2 , ā, b̄, ν̄) are as required.
Case IIB (B1,4(f ) = 0 and B1,6 (f ) = 1): Case IIA doesn’t hold, but B1,6 (f ) =
1 and there is an infinite B ⊆ [n1 , ω), l∗∗ < l∗ (see below) and (an ∈ vn : n ∈ B)
(given by B1,8 (f ), B2,0 (f ) and B1,7 (f ), respectively) such that:
a. lan ,f (an ) = l∗ and l∗∗ = B2,0 (f ) < l∗ .
b. ian ,f (an ),l = i∗l for l < l∗ .
c. If n ∈ B and m ∈ B ∩ n, then ρam ,f (am ),l∗∗ * ρan ,f (an ),l∗∗ .
In this case we define b̄, ν̄ and g as in Definition 6 and we let η2 := η1 . It’s easy to
see that (B, η1 , η2 , ā, b̄ν̄) are as required.
Remark: By a routine Ramsey-type argument, it’s easy to prove that if B1,6 (f ) = 1
then the values of B1,7 (f ), B2,0(f ) are well-defined and Borel-computable so the
above conitions hold.
7
Case IIC (B1,4 (f ) = B1,6 (f ) = 0): ¬IIA ∧ ¬IIB. We shall first prove that
B2,1 (f ), B2,2 (f ), B2,3 (f ), (B2,4,n (f ) : n ∈ B2,2 (f )) and (B2,5,l (f ) : l < B2,3 (f )) are
well-define and Borel computable.
Let l(∗) be the supremum of the l(a, f (a)) where n1 ≤ n and a ∈ vn (l(∗) < ω
by ¬2A). We can find l(∗∗) ≤ l(∗) such that B1 := {n ∈ B : vn,1 = {a ∈
vn
vn : l(a, f (a)) = l(∗∗)} has at least l(∗)
elements} is infinite. Next, we can find
i∗ (l) ∈ {1, −1} for l < l(∗∗) such that B2 : {n ∈ B1 : vn,2 = {a ∈ vn,1 :
|vn |
∧ ia,f (a),l = i∗ (l)} has at least 2l(∗∗)
elements} is infinite. For each n ∈ B2 ,
l(∗)
l<l(∗∗)
there are ρn,0 , ..., ρn,l(∗∗)−1 ∈ Tn such that vn,3 = {a ∈ vn,2 :
∧ ρa,f (a),l = ρn,l } has
l<l(∗∗)
|vn |
at least l(∗∗)2l(∗∗)
elements. By Ramsey’s theorem, there is an infinite subset
2nl(∗∗)
B3 ⊆ B2 such that for each l < l(∗∗), the sequence (T V (ρm,l ≤ ρn,l ) : m < n ∈ B3 )
is constant. Therefore, we’re done showing that the above Borel functions are welldefined.
Now if B′2,6 (f ) = 1 then we finish as in the previous case (this time we’re in the
situation of 6(b)(2)(K)(c)). If B′2,6 (f ) = 0, then we get a contradiction to the
assumption that f ∈ H3 , therefore we’re done.
Claim 8: If w(x0 , ..., xk∗ −1 ) is a reduced non-trivial group word, f0 , ..., fk∗∗ −1 ∈ H3
are pairwise distinct, gl ∈ Gfl (l ∈ {0, ..., k∗∗ − 1}), gl = gν∗l where νl ∈ 2ω \ {B(f ) :
f ∈ H3 }, l = k∗∗ , ..., k∗ − 1 and (νl : k∗∗ ≤ l < k∗ ) is without repetition, then
w(g0 , ...gk∗ −1 ) ∈ Sym(U) has a finite number of fixed points.
Notation: For l < k∗∗ , let νl := B(fl ).
i(m−1)
i(0)
Proof: Assume towards contradiction that w(x0 , ..., xk∗ −1 ) = xk(m−1) · ... · xk(0) ,
{f0 , ..., fk∗ −1 } and {g0 , ..., gk∗ −1 } form a counterexample, where i(l) ∈ {−1, 1}, k(l) <
k∗ and k(l) = k(l + 1) → ¬(i(l) = −i(l + 1)) for every l < m. WLOG m = lg(w) is
minimal among the various countrexamples. Let C = {a ∈ U : w(g0, ..., gk∗ −1 )(a) =
a}, this set is infinite by our present assumption. For c ∈ C, define bc,l by induction
on l < m as follows:
1. bc,0 = c.
i(l)
2. bc,l+1 = gk(l) (bc,l ).
Notational warning: The letter c with additional indices will be used to denote the
elements of sequences of the form c̄ from Definition 6(b)(2).
For all but finitely many c ∈ C, (bc,l : l < m) is without repetition by the minimality
of m, so wlog this is true for every c ∈ C.
For every c ∈ C, let ρc,l ∈ T be such that bc,l ∈ uρc,l , and let l1 [c] be such that
ρc,l1 [c] ≤∗ ρc,l for every l < m. We can choose l1 [c] such that one of the following
holds:
1. l1 [c] > 0 and ρc,l1 [c]−1 6= ρc,l1 [c]
2. l1 [c] = 0 and ρc,m−1 6= ρc,0
8
3. ρc,0 = ... = ρc,m−1
We may assume wlog that (l1 [c] : c ∈ C) is constant and that actually l1 [c] = 0 for
every c ∈ C. In order to see that we can assume the second part, for j < m let
i(j−1)
i(0) i(m−1)
i(j)
wj (x0 , ..., xk∗ −1 ) = xk(j−1) ···xk(0)xk(m−1) ···xk(j), then wj (g0 , ..., gk∗ −1 ) ∈ Sym(U) is a
conjugate of w(g0 , ..., gk∗ −1 ). The set of fixed points of wj (g0 , ..., gk∗ −1 ) includes {bc,j :
c ∈ C}, and therefore it’s infinite. For c ∈ C, (bc,j , bc,j+1 , ..., bc,m−1 , bc,0, ..., bc,j−1) and
wj (g0 , ..., gk∗ −1 ) satisfy the same properties that (bc,0 , ..., bc,m−1 ) and w(g0 , ..., gk∗ −1 )
satisfy. Therefore, if (l1 [c] : c ∈ C) is constantly j > 0, then by conjugating and
moving to wj (g0 , ..., gk∗ −1 ), we may assume that (l1 [c] : c ∈ C) is contantly 0.
Let l2 [c] < m be the maximal such that ρc,0 = ... = ρc,l2 [c] , so wlog l2 [c] = l∗ for
every c ∈ C. For l < k∗ let η1,l be B(fl ) if l < k∗∗ and νl if l ∈ [k∗∗ , k∗ ) (we might
also denote it by ρl in this case). As fl ∈ H3 for l <k∗ ∗ , and ρl ∈
/ {B(f ) : f ∈ H3 }
for l ∈ [k∗∗ , k∗), it follows that l1 < k∗∗ ≤ l2 < k∗ → η1,l1 6= η1,l2 . Therefore,
(η1,l : l < k∗ ) is without repetition.
Now let η2,l be defined as follows:
1. If l < k∗∗ , let η2,l be η2 from definition 6(b)(2) for fl and gl .
2. If l ∈ [k∗∗ , k∗ ), let η2,l = η1,l .
Let j(∗) < ω be such that:
a. (η1,l ↾ j(∗) : l < k∗ ) is without repetition.
b. If η1,l1 6= η2,l2 then η1,l1 ↾ j(∗) 6= η2,l2 ↾ j(∗) (l1 , l2 < k∗ ).
c. If η2,l1 6= η2,l2 then η2,l1 ↾ j(∗) 6= η2,l2 ↾ j(∗) (l1 , l2 < k∗ ).
d. j(∗) > 3m, k∗ .
e. j(∗) > n(l1 , l2 ) for every l1 < l2 < k∗ , where n(l1 , l2 ) is defined as follows:
1. If k∗∗ ≤ l1 , l2 , let n(l1 , l2 ) = 0.
2. If l1 < k∗∗ or l2 < k∗∗ , let (νn1 : n ∈ B1 ) and (νn2 : n ∈ B2 ) be as in definition 6(b)(2)
for (fl1 , η1,l1 ) and (fl2 , η1,l2 ), respectively. If there is no νn1 such that νn1 η1,l1 and no
νn2 such that νn2 η1,l2 , let n(l1 , l2 ) = 0. Otherwise, there is at most one n ∈ B1 such
2
that νn1 η1,l1 and νn1 ≤ η1,l2 and there is at most one m ∈ B2 such that νm
η1,l2
2
1
2
1
2
and νm ≤ η1,l1 . If there are νn and νm as above, let n(l1 , l2 ) = lg(νn ) + lg(νm ) + 1.
2
If there is νn1 as above but no νm
as above, let n(l1 , l2 ) = lg(νn1 ) + 1, and similarly
for the dual case.
f. j(∗) > m(l1 , l2 ) for every l1 < l2 < k∗∗ where m(l1 , l2 ) is defined as follows: Let
2
(νn1 : n ∈ B1 ) and (νm
: m ∈ B2 ) be as in definition 6(b)(2) for (fl1 , η1,l1 ) and
2
: m ∈ B2 }| < ℵ0 ,
(fl2 , η1,l2 ), respectively. As η1,l1 6= η1,l2 , |{νn1 : n ∈ B1 } ∩ {νm
let s(l1 , l2 ) be the supremum of the length of members in this intersection and let
m(l1 , l2 ) := s(l1 , l2 ) + 1.
We may assume wlog that lg(ρc,l1 [c] ) > j(∗) for every c ∈ C. We now consider two
possible cases (wlog T V ((ρc,l : l < m) is constant) is the same for all c ∈ C):
9
Case I: For every c ∈ C, (ρc,l : l < m) is not constant.
In this case, for each such c ∈ C, l2 [c] < m − 1 and bc,l2 [c] ∈ uρc,0 = ... = uρc,l2 [c] . Now
i(l [c])
(bc,l2 [c] , bc,l2 [c]+1) ∈ gk(l22 [c]) , and as ρc,l2 [c] 6= ρc,l2 [c]+1 , necessarily k(l2 [c]) < k∗∗ . By the
definition of l1 [c] and the fact that ρc,l1 [c] = ρc,l2 [c] , necessarily ρc,l2 [c] <∗ ρc,l2 [c]+1.
−1
For each l < m − 1, if lg(ρc,l ) < lg(ρc,l+1), then either gk(l) or gk(l)
is as in definition 6(2)(b), so letting n = lg(ρc,l ), (ρc,l , ρc,l+1 ) here correspond to (η1 ↾ n, νn )
there, and there are (al , bl , cl , dl , el ) = (alc , blc , clc , dlc , elc ) in our case that correspond to
(an , bn , cn , dn , en ) in 6(2)(b). In the rest of the proof we shall denote those sequences
by (al , bl , cl , dl , el ), as the identity of the relevant c ∈ C should be clear. In addition,
one of the following holds:
i(l)
1. i(l) = 1 and gk(l) (al ) = bl .
i(l)
2. i(l) = −1 and gk(l) (bl ) = al .
Similarly, for l < m − 1, if lg(ρc,l ) > lg(ρc,l+1) then the above is true modulo the
fact that now (ρc,l , ρc,l+1) correspond to (νn , η1 ↾ n) and one of the following holds:
i(l)
1. i(l) = 1 and gk(l) (bl ) = el .
i(l)
2. i(l) = −1 and gk(l) (bl ) = al .
Therefore, if l = l2 [c] then lg(ρc,l ) < lg(ρc,l+1), so the first option above holds, and
therefore ρc,l is an initial segment of η1,k(l) .
If l = m − 1, then lg(ρc,m−1) > lg(ρc,0) = lg(ρc,m ) and therefore ρc,0 is an initial
segment of η1,k(m−1) . It follows that ρc,0 = ρc,l2 [c] is an initial segment of η1,k(l2 [c]) ∩
η1,k(m−1) . Recalling that lg(ρc,0 ) = lg(ρc,l1 [c] ) > j(∗) and that (η1,l ↾ j(∗) : l < k∗ ) is
without repetition, it follows that k(m − 1) = k(l2 [c]).
We shall now prove that if l2 [c] < m − 1 then l2 [c] = m − 2. Let (al2 [c] , bl2 [c] , ...)
be as above for l = l2 [c], so bc,l2 [c]+1 = bl2 [c] , and as k(l2 [c]) = k(m − 1), we get
bc,l2 [c]+1 = bl2 [c] = bm−1 . In order to show that l2 [c] = m − 2, it suffices to show that
bm−1 = bc,m−1 (as the sequence of the bc,l s is without repetition), which follows from
the fact that ρc,0 <∗ ρc,m−1 .
As we assume that the word w is reduced, and as k(m − 2) = k(l2 [c]) = k(m − 1),
necessarily i(m − 2) = i(m − 1). We may assume wlog that i(m − 2) = i(m − 1) = 1
(the proof for i(m−2) = i(m−1) = −1 is similar, as we can replace w by a conjugate
of its inverse).
i(m−3)
i(0)
Let w ′ = w ′(g0 , ..., gk∗ −1 ) := gk(m−3) · · · gk(0) , by the above considerations and as
i(m−1)
i(m−1)
bm−1 = bc,m−1 , it follows that em−1 = gk(m−1) (bm−1 ) = gk(m−1) (bm−1 ) = gk(m−1) (bc,m−1 ) =
i(m−2)
bc,0 . We also know that gk(m−2) (w ′(bc,0 )) = gk(m−2) (w ′ (bc,0 )) is “higher” than w ′ (bc,0 ).
i(m−2)
Therefore, gk(m−2) (w ′(bc,0 )) = gk(m−2) (w ′ (bc,0 )) = bm−2 = bm−1 . It also follows that
w ′ (bc,0 ) = am−2 = am−1 . Therefore, w ′(em−1 ) = am−1 .
10
We shall now prove that if l < m − 2 then bc,l+1 = (gν∗k(l) )i(l) (bc,l ). Assume that
for some l < m − 2, gk(l) (bc,l ) 6= gν∗k(l) (bc,l ) and we shall derive a contradiction. Let
(am−2 , bm−2 , ...) be as before for gk(m−2) , so (bc,m−2 , bc,0 ) = (am−2 , em−2 ).
Case I (a): k(l) = k(m − 2) = k(m − 1). As the bc,i s are without repetition, if
0 < l < m − 2, then bc,l ∈
/ {bc,m−2 , bc,0} = {am−2 , em−2 } = {al , el }, and of course,
l l
l
bc,l ∈
/ {b , c , d } (as it is a “lower” element). Therefore, gk(l) (bc,l ) = gν∗k(l) (bc,l ), a
i(m−1)
contradiction. If l = 0, then gk(m−1) (bc,m−1 ) = bc,0 and (bc,m−1 , bc,0 ) = (bm−1 , em−1 ).
If i(0) = −i(m − 1), then by conjugating gk(0) , we get a shorter word with infinitely
many fixed points, contradicting our assumption on the minimality of m.
If i(0) = i(m − 1) = 1, then we derive a contradiction as in the case of 0 < l.
Case I (b): k(l) 6= k(m − 2) = k(m − 1). In this case, we know that gk(l) almost
coincides with gν∗k(l) , with the exception of at most {al , bl , cl , dl , el }. Let ρc := ρc,0 =
... = ρc,m−2 , then necessarily ρc ≤ νk(m−1) = η1,k(m−1) (as gk(m−1) moves bc,m−1 to a
lower uρ (namely uρc ), ρc plays the role of η1 ↾ n in Definition 6 for gk(m−1) ). By our
assumption, lg(ρc ) > j(∗) and (η1,l ↾ j(∗) : l < k(∗)) is without repetition, therefore
η1,k(l) ↾ lg(ρc ) 6= ρc , so ρc η1,k(l) . Therefore, when we consider fk(l) and η1,k(l) in
definition 6(b)(2), then ρc has the form νn for some n. By the choice of j(∗), it’s
then impossible to have ρc ≤ η1,k(m−1) , a contradiction.
Therefore, am−2 = w ′ (g0 , ..., gk∗ −1 )(em−2 ) = w ′ (g0 , ..., gk∗ −1 )(bc,0 ) = w ′(gν∗0 , ..., gν∗k∗ −1 )(bc,0 ) =
w ′ (gν∗0 , ..., gν∗k∗ −1 )(em−2 ). In the notation of the claim and definition 6(b)(2), F1 (fk(m−2) )(am−2 ) =
em−2 , therefore, by composing with w ′ , we obtain a word composed of permutation
of uρc,m−2 (in the sense of claim 4(f)) that fixes em−2 ∈ uρc,m−2 , therefore, m − 3 = 0
(or else we get a contradiction by claim 4(f)).
i(0)
i(0)
It follows that w(g0 , ..., gk∗ −1 ) = gk(m−2) gk(m−2) gk(0) and gk(0) (em−2 ) = am−2 . Now,
obviously ρc,m−2 ≤ η1,k(m−2) , so ρc,m−2 η1,k(0) = νk(0) . By the definition, gν∗k(0) ↾
uρc,m−2 = fρc,m−2 ,νk(0) ↾lg(ρc,m−2 ) 6= fρc,m−2 ,ρc,m−2 . Also F1 (fk(m−2) ) ↾ uρc,m−2 = gν∗k(m−2) ↾
i(0)
uρc,m−2 = fρc,m−2 ,ρc,m−2 . Therefore, we get the following: am−2 = gk(0) (em−2 ) =
(gν∗k(0) )i(0) (em−2 ) = (fρc,m−2 ,νk(0) ↾lg(ρc,m−2 ) )i(0) (em−2 ) and em−2 = F1 (fk(m−2) )(am−2 ) =
fρc,m−2 ,ρc,m−2 (am−2 ). In conclusion, we get a contrdiction to claim 4(f), as we have a
short non-trivial word that fixes em−2 .
Case II: (ρc,l : l < m) is constant for every c ∈ C (so l2 [c] = m − 1). Let
i(l)
ρc := ρc,0 = ... = ρc,m−1 . If gk(l) (bc,l ) = (gν∗k(l) )i(l) (bc,l ) for every l < m, then we
get a contradiction to claim 4(f). Therefore, for every c ∈ C, the set vc = {l <
i(l)
m : gk(l) (bc,l ) 6= (gν∗k(l) )i(l) (bc,l )} is nonempty. Without loss of generality, vc doesn’t
depend on c, and we shall denote it by v. For every l ∈ v, if i(l) = 1 then (bc,l , bc,l+1) ∈
{(al , bl ), (bl , el ), (cl , dl )}, if i(l) = −1 then (bc,l , bc,l+1) ∈ {(bl , al ), (el , bl ), (dl , cl )}.
We shall now prove that for some k < k∗∗ , k(l) = k for every l ∈ v. Suppose
not, then for some l1 < l2 ∈ v, k(l1 ) 6= k(l2 ). By the choice of j(∗), each of the
following options in impossible: ρc,l1 ≤ η1,l1 ∧ ρc,l2 ≤ η1,l2 , ρc,l1 ≤ η1,l1 ∧ ρc,l2 η1,l2 ,
11
ρc,l1 η1,l1 ∧ρc,l2 ≤ η1,l2 or ρc,l1 η1,l1 ∧ρc,l2 η1,l2 . Therefore we get a contradiction.
It follows that {k(l) : l ∈ v} is singelton, and we shall denote its only member by
k < k∗∗ .
Note that if l1 ∈ v, l2 ∈ v is the successor of l1 in v, l1 + 1 < l2 and c ∈ C then
bc,l1 +1 6= bc,l2 (recall that (bc,l : l < m) is withut repetition). We shall now arrive at
a contradiction by examining the following three possible cases (in the rest of the
proof, we refer to l(∗) from Definition 5(A)(e) as “the distance between a and b”,
and similarly for any pair of members from some uη ):
Case II (a): gk is as in definition 6(b)(2)(K)(a). In this case, for every l ∈ v, the
only possibilities for (bc,l , bc,l+1) are either of the form (c, d) or (d, c) (and not both,
as we don’t allow repetition). As the distance between c and d is at most 2, we get
a word made of fρ,ν s of length ≤ m + 1 that fixes c, contradicting claim 4(f).
Case II (b): gk is as in definition 6(b)(2)(K)(c). Pick c ∈ C such that lg(ρc ) is alseo
greater than m + l∗ where l∗ is as in definition 6(b)(2)(K)(c) for gk . As the sequence
(bc,l : l < m) is without repetition, necessarily 1 ≤ |v| ≤ 3.
If |v| = 3, then necessarily the sequences (a, b, e) or (e, b, a) occur in (bc,l : l < m),
as well as (c, d) or (d, c). As the distance between a and e is 1 and the distance
between c and d is ≤ 2, we get a contradiction as before.
Suppose that |v| = 2. If the sequence (a, b, e) appears in (bc,l : l < m), we get
a contradiction as above. If (a, e) (or (e, a)) and (c, d) (or (d, c)) appear, we also
get a contradiction as above. If (a, b)/(b, a) and (c, d)/(d, c) appear, as the distance
between a and b is l∗ , we get a word made of fρ,ν s of length ≤ m + l(∗) fixing c, a
contradiction to claim 4(f). Finally, if |v| = 1 we get a contradiction similarly.
Case II (c): gk is as in definition 6(b)(2)(K)(b). As in the previous case, where the
only non-trivial difference is when either |v| ∈ {1, 2} and the sequence (a, b)/(b, a)
appears in (bc,l : l < m), but not as a subsequence of (a, b, e)/(e, b, a). If for some
c this is not the case, then we finish as before, so suppose that it’s the case for
every c ∈ C. As the distance between c and d is ≤ 2, suppose wlog that |v| = 1,
k = k(m − 1) (by conjugating) and the sequence (bc,l : l < m) ends with a and
starts with b or vice versa. Therefore, every c ∈ C is of the form an or bn (where
i(m−2)
i(0)
n ∈ B and B is as in definition 6(b)(2) for gk ) and either gk(0) · · · gkm−2 (an ) = bn
i(0)
i(m−2)
or gk(0) · · · gkm−2 (bn ) = an , so the distance between an and bn is ≤ m − 1. As C is
infinite, the distance between an and bn is ≤ m − 1 for infinitely many n ∈ B. This
is a contradiction to the assumption from definition 6(b)(2)(K)(b) that the distance
between an and bn is increasing.
This completes the proof of claim 8.
Claim 9: There exists a Borel function B4 : U U → U U such that for every f ∈ B1 ,
B4 (f ) ∈ Gf .
Proof: As in [HwSh:1089], and we comment on the main point in the proof of claim
7.
12
Definition 10: Let G be the subgroup of Sym(U) generated by {B4 (f ) : f ∈
H3 } ∪ {gν∗ : ν ∈ 2ω \ {B(f ) : f ∈ H3 }}.
Claim 11: G is a maximal cofinitary group.
Proof: G is cofinitary by claim 8, so it’s enough to prove maximality. Assume
towards contradiction that H is a counterexample and let f∗ ∈ H \G, so B4 (f∗ ) ∈ G,
and we shall denote f ∗ = B4 (f∗ ).
Case I: f∗ ∈ H3 . In this case, by Definition 6, {an : n ∈ B} ⊆ eq(f∗ , f ∗ ) := X (see
thee relevant notation in definition 6), hence it’s infinite. Therefore, f∗−1 f ∗ ↾ X is
the ientity, but f∗−1 f ∗ ∈ H and H is cofinitary, therefore f∗−1 f ∗ = Id so f∗ = f ∗ ∈ G,
a contradiction.
Case II: f∗ ∈
/ H3 . By the definition of H3 , B′2,6 (f∗ ) = 0, so the sequences B2,1 (f∗ ) =
(An = An (f∗ ) : n ∈ B2,2 (f∗ )), ρ¯n∗ = (ρn,i : i < l∗ ) = (ρa,f∗ (a),i : i < l∗ ) and
ī = (il : l < l∗ ) = (ia,f (a),l : l < l∗ ) (n ∈ B2,2 (f∗ ), a ∈ An ) are well-defined, and
for every l < l∗ , (ρn,l : n ∈ B2,2 (f∗ )) is ≤-increasing, so νl :=
∪ ρn,l ∈ 2ω is
n∈B2,2 (f∗ )
well-defined. Let g = (gν∗0 )i0 · · · (gν∗l∗ −1 )il∗ −1 ∈ G1 (we may assume that it’s a reduced
product). Let w1 = {l < l∗ : (∃fl ∈ H3 )(νl = B(fl ))} and w2 = l∗ \ w1 . For l < l∗ ,
define gl as follows:
1. If l ∈ w1 , let gl = B4 (fl ).
2. If l ∈ w2 , let gl = gν∗l .
i(0)
Let g ′ = g0
i(l −1)
∗
· · · gl∗ −1
. By the definition of G, g0 , ..., gl∗ −1 ∈ G, hence g ′ ∈ G.
Again by Definition 6, if l ∈ w1 then gl = F1 (fl ) mod I1 and gl−1 = F1 (fl )−1
mod I1 . Now suppose that g(a) 6= g ′(a), then there is a minimal l < l∗ such
that (gν∗0 )i(0) · · · (gν∗l )i(l) (a) 6= (g0 )i(0) · · · (gl )i(l) (a). Let vl = dif (gν∗l , gl ), then a ∈
i(l−1)
i(l−1)
(g0i0 · · · gl−1 )−1 (vl ). In order to show that (g0i0 · · · gl−1 )−1 (vl ) ∈ I1 it suffices to
observe that for i ∈ w1 , functions of the form gi , gi−1 map elements of I1 to elements
of I1 , therefore it follow that g = g ′ mod I1 . It suffices to show that eq(f∗ , g) ∈
/ I1 ,
′
−1 ′
as it will then follow that eq(f∗ , g ) ∈
/ I1 , so f∗ g = Id on an I1 −positive set, hence
on an infinite set. As f∗−1 g ′ ∈ H and H is cofinitary, f∗−1 g ′ = Id, an therefore
f∗ = g ′ ∈ G, a contradiction.
So let n ∈ B2,2 (f∗ ) and a ∈ An = An (f∗ ), and observe that f∗ (a) = g(a). Indeed,
i −1
by the definition of B2,1 (f∗ ), for every such a, f∗ (a) = ((fηi01 ↾nρn,0 ) · · · (fη↾l∗nρ
n,l∗ −1 ))(a)
(where η1 is as in the definition of B2,1 (f∗ )). It’s now easy to verify that the last
expression equals g(a). It’s also easy to verify that
∪ An ∈
/ I1 , therefore we’re
n∈B2,2 (f∗ )
done.
Claim 12: G is Borel.
Proof: It suffices to prove the following subclaim:
13
Subclaim: There exists a Borel function B5 with domain Sym(U) such that if
im
” for some
g ∈ G then B5 (g) = (g0 , g1 , ..., gm ) such that G |= ”g = g0i0 g1i1 · · · gm
m+1
(i0 , ..., im ) ∈ {−1, 1}
.
Proof: By the definition of G, if g ∈ G then there are m, f0 , ..., fm ∈ A1 (possibly
im
with repetition) and i0 , ..., im ∈ {−1, 1} such that g = g0i0 · · · gm
where each gi
is either of the form B4 (fi ) for fi ∈ H3 (in this case, let νi := B(fi )) or gν∗i for
νi ∈ 2ω \ {B(f ) : f ∈ H3 }.
Now if n is greater than m!, then for some u ⊆ 2n such that |u| ≤ m! <
ρ ∈ 2n \ u we have:
2n
,
2
for every
a. For every l ≤ m, gl ↾ uρ = fρ,νl ↾lg(ρ) .
i0
im
∈ Sym(uρ).
b. g ↾ uρ can be represented as fρ,ν
· · · fρ,ν
m ↾lg(ρ)
0 ↾lg(ρ)
d. By claim 4(f), the above representation of g ↾ uρ is unique.
Therefore, from g we can Borel-compute ((νi ↾ n : n < ω) : i < m) hence (νi : i < m).
As H3 is Borel and B is injective, the sets {B(f ) : f ∈ H3 } and 2ω \ {B(f ) : f ∈ H3 }
are Borel. Now if νi ∈ 2ω \ {B(f ) : f ∈ H3 }, we can Borel compute gi = gν∗i . If
νi ∈ {B(f ) : f ∈ H3 }, then νi = B(fi ) and we can Borel-compute fi (by applying
B−1 from definition 3(e)) hence B4 (fi ).
References
[HwSh1089] Haim Horowitz and Saharon Shelah, A Borel maximal eventuallly different family, arXiv:1605.07123.
[FFT] Vera Fischer, Sy David Friedman, Asger Toernquist, Definable maximal
cofinitary groups, arXiv:1603.02942.
[Ka] Bart Kastermans, The complexity of maximal cofinitary groups, Proc. of the
Amer. Math. Soc., Vol. 137(1), 2009, 307-316
[Ma] A. R. D. Mathias, Happy families, Ann. Math. Logic 12 (1977), no. 1, 59-111.
MR 0491197.
(Haim Horowitz) Einstein Institute of Mathematics
Edmond J. Safra Campus,
The Hebrew University of Jerusalem
Givat Ram, Jerusalem, 91904, Israel.
E-mail address: [email protected]
14
(Saharon Shelah) Einstein Institute of Mathematics
Edmond J. Safra Campus,
The Hebrew University of Jerusalem
Givat Ram, Jerusalem, 91904, Israel.
Department of Mathematics
Hill Center - Busch Campus,
Rutgers, The State University of New Jersey.
110 Frelinghuysen road, Piscataway, NJ 08854-8019 USA
E-mail address: [email protected]
15
| 4math.GR
|
On the Optimal Recovery Threshold of Coded
Matrix Multiplication
Sanghamitra Dutta2∗ , Mohammad Fahim1∗ , Farzin Haddadpour1∗ , Haewon Jeong2∗ , Viveck Cadambe1 , Pulkit Grover2
arXiv:1801.10292v1 [cs.IT] 31 Jan 2018
Abstract
We provide novel coded computation strategies for distributed matrix-matrix products that outperform the recent “Polynomial
code” constructions in recovery threshold, i.e., the required number of successful workers. When m-th fraction of each matrix can
be stored in each worker node, polynomial codes require m2 successful workers, while our MatDot codes only require 2m − 1
successful workers, albeit at a higher communication cost from each worker to the fusion node. Further, we propose “PolyDot”
coding that interpolates between Polynomial codes and MatDot codes. Finally, we demonstrate an application of MatDot codes
to multiplying multiple (> 2) matrices.
I. I NTRODUCTION
As the era of Big Data advances, massive parallelization has emerged as a natural approach to overcome limitations imposed
by saturation of Moore’s law (and thereby of single processor compute speeds). However, massive parallelization leads to
computational bottlenecks due to straggling or faulty nodes, which has been a topic of active recent interest in the emerging area
of “coded computation” (e.g. [1]–[23]), which not only advances on coding approaches in classical works on Algorithm-Based
Fault Tolerance (ABFT), but also provides novel analyses of required computation time (e.g. expected time [1] and deadline
exponents [24]). Perhaps most importantly, it brings an information-theoretic lens to the problem by examining fundamental
limits and comparing them with existing strategies. A broader survey of results and techniques is provided in [25].
This paper focuses on the coded matrix multiplication problem. Matrix multiplication is central to many modern computing
applications, including machine learning and scientific computing. There is a lot of interest in classical ABFT literature (starting
from [26]) and more recently in coded computation literature (e.g. [4], [27]) to make matrix multiplications resilient to faults
and delays. In particular, Yu, Maddah-Ali, and Avestimehr [4] provide novel coded matrix-multiplication constructions called
Polynomial codes that outperform classical work from ABFT literature in terms of the recovery threshold, the minimum number
of successful (non-delayed, non-faulty) processing nodes required for completing the overall matrix multiplication.
Our work advances on those constructions in scaling sense. Concretely, when m-th fraction of each matrix can be stored
in each worker node, Polynomial codes require m2 successful workers, while MatDot codes only require 2m − 1 successful
workers. However, as we note in Section III-A, this comes at an increased per-worker communication cost. We begin with
some motivating examples that show our constructions and a summary of results.
A. Motivating examples and summary of results
Consider the setting where each processing worker node stores N 2 /m linear combinations of the entries of A and N 2 /m
linear combinations of the entries of B. Assume that P worker nodes perform the computation, we evaluate the straggler
tolerance of a technique by its recovery threshold k. In the following, we demonstrate our construction for m = 3. We describe
(i) ABFT matrix multiplication [26] (also called product-coded matrices in [27]), (ii) Polynomial codes [4] and then (iii)
our proposed construction, MatDot codes, each with successively improving, i.e., smaller, recovery threshold. We begin by
describing ABFT matrix multiplication.
Example 1. [ABFT codes [26], Fig. 1, k = 6] Consider two N × N matrices
Ã1
A=
, B = B̃1 B̃2 .
Ã2
Using ABFT, it is possible to compute AB over P nodes such that, (i) each node uses one linear combination of A and one
linear combination of B and (ii) the overall computation is tolerant to P − 6 stragglers, i.e., 6 nodes suffice to recover AB?
ABFT codes use the strategy as per Fig. 1, where 4 of the 9 worker nodes compute Ãi , B̃j , i, j ∈ {1, 2} and the remaining
worker nodes compute Ãi (B̃1 + B̃2 ), (Ã1 + Ã2 )B̃i , (Ã1 + Ã2 )(B̃1 + B̃2 ) for i = 1, 2, respectively. The general principle of
ABFT is to encode the rows of A and the columns of B using systematic MDS codes of dimension m.
A superior recovery threshold was obtained in [4], which gave the following polynomial code construction.
1 Pennsylvania
State University, 2 Carnegie Mellon University,
∗
author ordering in alphabetical order.
Ã1
Fig. 1. ABFT matrix multiplication [26] for P = 9 worker nodes with m = 2, where A =
, B = [B̃1 B̃2 ]. The recovery threshold is 6.
Ã2
Example 2. [Polynomial codes [4], Fig. 2, k = 4] Consider two N × N matrices
Ã1
A=
, B = B̃1 B̃2 .
Ã2
Polynomial codes computes AB over P nodes such that, (i) each node uses N 2 /2 linear combination of the entries of A and
N 2 /2 linear combination of the entries of B and (ii) the overall computation is tolerant to P − 4 straggler, i.e., 4 nodes suffice
to recover AB, Polynomial codes use the following strategy: Node i computes (Ã1 + Ã2 i)(B̃1 + B̃2 i2 ), i = 1, 2, . . . P, so
that from any 4 of the P nodes, the polynomial p(x) = (Ã1 B̃1 + Ã2 B̃1 x + Ã1 B̃2 x2 +Ã1 B̃2 x3 ) can be
interpolated. Having
Ã1 B̃1 Ã1 B̃2
interpolated the polynomial, the coefficient (matrices) can be used to evaluate AB as
.
Ã2 B̃1 Ã2 B̃2
Ã1
Fig. 2. Polynomial Codes [4] with m = 2, where A =
, B = [B̃1 B̃2 ]. The recovery threshold is 4.
Ã2
Our novel MatDot construction obtains a smaller recovery threshold as compared with polynomial codes. Unlike ABFT and
polynomial codes, MatDot divides matrix A column-wise and matrix B row-wise.
Example 3. [MatDot codes, Fig. 3, k = 3]
MatDot codes compute AB over P nodes such that, (i) each node uses N 2 /2 linear combination of the entries of A and
2
N /2 linear combination of the entries of B and (ii) the overall computation is tolerant to P − 3 straggler, i.e., 3 nodes suffice
to recover AB. The proposed MatDot codes use the following strategy: Matrix A is split into submatrices A0 , A1 vertically
each of dimension N ×N/2 and matrix B is split into submatrices B0 , B1 horizontally each of dimension N/2×N as follows:
B0
A = [A0 A1 ] , B =
.
(1)
B1
Now, we define the encoding functions pA (x) and pB (x) as pA (x) = A0 +A1 x and pB (x) = B0 x+B1 . Let x0 , x1 , · · · , xP −1
be distinct elements of field F, the master node sends pA (xl−1 ) and pB (xl−1 ) to the l-th worker node where the l-th worker
node performs the multiplication pA (xl−1 )pB (xl−1 ) and sends the output to the fusion node. The exact computations at each
worker node are depicted in Fig. 3. We can observe that the fusion node can obtain the product AB using the output of any
Fig. 3. An illustration of the computational system with four worker nodes and applying MatDot codes with m = 2. The recovery threshold is 3.
three successful workers as follows: Let the worker nodes 1, 2, and 3 be the first three successful worker nodes, then the fusion
node owns the following three matrices:
pA (x0 )pB (x0 ) = A0 B1 + (A0 B0 + A1 B1 )x0 + A1 B0 x20 ,
pA (x1 )pB (x1 ) = A0 B1 + (A0 B0 + A1 B1 )x1 + A1 B0 x21 ,
pA (x2 )pB (x2 ) = A0 B1 + (A0 B0 + A1 B1 )x2 + A1 B0 x22 .
Since these three matrices can be seen as three evaluations of the matrix polynomial pA (x)pB (x) of degree two at three distinct
evaluation points x0 , x1 , x2 , the coefficients of powers of x in pA (x)pB (x) can be obtained using polynomial interpolation.
This includes the coefficient of x, A0 B0 + A1 B1 , i.e., AB.
Since the worker nodes require to communicate N × N matrices to the fusion node, MatDot incurs higher communication
cost as compared to polynomial codes.
Our main contributions are as follows:
• We present our system model in the next section, and describe MatDot codes in Section III. While polynomial codes have
a recovery threshold of Θ(m2 ), MatDot codes have a recovery threshold of Θ(m).
• We present A systematic version of MatDot codes - where the first operations of the first m worker nodes may be viewed
as multiplication in uncoded form - in Section IV.
• In Section V, we propose “PolyDot codes”, a unified view of MatDot and Polynomial codes that leads to a trade-off
between recovery threshold and communication costs.
• In Section VI, we apply the constructions of Section III to study coded computation for multiplying more than two
matrices.
We note that recent works of Yu, Maddah-Ali, and Avestimehr [28] and Dutta, Bai, Grover, and Low [29] obtain constructions
that outperform PolyDot codes in tradeoffs between communication and recovery threshold (although MatDot codes continue
to have the smallest recovery threshold for given storage constraints). Importantly, Yu et al. [28] also provide interesting
impossibility (converse) results demonstrating approximate optimality of these improved strategies.
II. S YSTEM MODEL AND PROBLEM STATEMENT
A. System model
The system, illustrated in Fig. 4, consists of a master node, multiple worker nodes, and a fusion node that combines outputs
from successful worker nodes. These are defined more formally below.
Definition 1. [Computational system] A computational system consists of (i) a master node that receives computational inputs
and processes them, (ii) memory-constrained worker nodes that perform pre-determined computations on their respective inputs
in parallel, and (iii) a fusion node that receives messages from successful workers and performs decoding operations on them.
For practical utility, it is important to have the amount of processing that the worker nodes perform to be much smaller than
the processing at the master and the fusion node.
Fig. 4. The computational system. Workers are delay/fault prone
We assume that any worker node can fail to complete its computation because of faults or delays. Thus, we define a subset
of all workers as the “successful workers.”
Definition 2. [Successful workers] Workers that finish their computation task successfully and send their output to the fusion
node are called successful workers.
Definition 3. [Successful computation] If the computational system on receiving the inputs produces the correct computational
output, the computation is said to be successful.
Definition 4. [Recovery threshold] The recovery threshold is the worst-case minimum number of successful workers required
by the fusion node to complete the computation successfully.
We will denote the number of worker nodes by P , and the recovery threshold by k. The worst-case in Definition 4 is over
all possible configurations of k successful workers.
B. Problem statement
Compute the matrix-matrix product AB of two N × N square matrices A and B in the computational system specified in
Section II-A. More formally, the goal is to have the system specified in Section II-A receive as inputs matrices A, B ∈ FN ×N
where F is a field (|F| > P ), and produce AB at the output of the fusion node. Each worker can receive at most 2N 2 /m
symbols from the master node, where each symbol is an element of F. For the simplicity, we assume that m divides N . The
computational complexities of the master and fusion nodes, in terms of the matrix parameter N , are strictly less than the
computational complexity at any worker node1 . The goal is to perform this matrix product utilizing faulty/delay prone workers
with as low recovery threshold as possible.
III. M AT D OT C ODES
In this section, we provide a code construction, named MatDot code, for the problem stated in Section II-B that obtains a
recovery threshold of 2m − 1 where m ≥ 2 is any positive integer that divides N .
Theorem 1. For the matrix multiplication problem specified in Section II-B computed on the system defined in Definition 1,
a recovery threshold of 2m − 1 is achievable where m ≥ 2 is any positive integer that divides N .
Before we prove Theorem 1, we first describe the general MatDot code construction.
Construction 1. [MatDot Codes]
Splitting of input matrices: The matrix A is split “vertically” and B is split “horizontally” into m equally sized pieces (of
N 2 /m symbols each) as follows:
B0
B1
A = [A0 |A1 | . . . |Am−1 ] , B =
(2)
,
..
.
Bm−1
where, for i ∈ {0, . . . , m − 1}, and Ai , Bi are N × N/m and N/m × N dimensional
matrices, respectively. P
Pm−1
m−1
Master node (encoding): Let x1 , x2 , . . . , xP be distinct elements of F. Let pA (x) = i=0 Ai xi and pB (x) = j=0 Bj xm−1−j .
The master node sends to the l-th worker evaluations of pA (x), pB (x) at x = xl , that is, it sends pA (xl ), pB (xl ) to the l-th
worker.
Worker nodes: For l ∈ {1, 2, . . . , P }, the l-th worker node computes the matrix product C(xl ) = pA (xl )pB (xl ) and sends
it to the fusion node on successful completion.
1 If the fusion node is allowed to have higher computational complexity, the workers can simply store A, B using maximum distance separable codes to
get a recovery threshold of m; the fusion node simply recovers A, B and then multiplies them, essentially performing the whole operation.
Fusion node (decoding): The fusion node uses outputs of any 2m − 1 successful workers to compute the coefficient of xm−1
in the product C(x) = pA (x)pB (x) (the feasibility of this step will be shown later in the proof of Theorem 1). If the number
of successful workers is smaller than 2m − 1, the fusion node declares a failure.
Notice that in MatDot codes, we have
AB =
m−1
X
Ai Bi ,
(3)
i=0
where, for i ∈ {0, · · · , m − 1}, Ai and Bi are as defined in (2). The simple observation of (3) leads to a different way of
computing the matrix product as compared with Polynomial codes based computation. In particular, to compute the product,
we only require, for each i, the product of Ai and Bi . That is, we do not require products of the form Ai Bj for i 6= j unlike
Polynomial codes, where, after splitting the matrices A, B in to m parts, all m2 cross-products are required to evaluate the
overall matrix product. This leads to a significantly smaller recovery threshold for our construction.
Proof of Theorem 1. To prove the theorem, it suffices to show that in the MatDot code construction described above, the fusion
node is able to construct C from any 2m − 1 worker nodes. Observe that the coefficient of xm−1 in:
! m−1
m−1
X
X
(4)
C(x) = pA (x)pB (x) =
A i xi
Bj xm−1−j
i=0
j=0
Pm−1
is AB = i=0 Ai Bi (from (3)), which is the desired matrix-matrix product. Thus it is sufficient to compute this coefficient at
the fusion node as the computation output for successful computation. Now, because the polynomial C(x) has degree 2m − 2,
evaluation of the polynomial at any 2m − 1 distinct points is sufficient
Pm−1to compute all of the coefficients of powers of x in
pA (x)pB (x) using polynomial interpolation. This includes AB = i=0 Ai Bi , the coefficient of xm−1 . The next section has
a complexity analysis that shows that master and fusion nodes have a lower computational complexity as compared with the
workers.
A. Complexity analyses of MatDot codes
Encoding/decoding complexity: Decoding requires interpolating a 2m − 2 degree polynomial for N 2 elements. Using
polynomial interpolation algorithms of complexity O(k log2 k) [30], where k = 2m − 1, the decoding complexity per matrix
element is O(m log2 m). Thus, for N 2 elements, the decoding complexity is O(N 2 m log2 m).
Encoding for each worker requires performing two additions, each adding m scaled matrices of size N 2 /m, for an overall
encoding complexity for each worker of O(mN 2 /m) = O(N 2 ). Thus, the overall computational complexity of encoding for
P workers is O(N 2 P ).
Each worker’s computational cost: Each worker multiplies two matrices of dimensions N ×N/m and N/m×N , requiring
N 3 /m operations (using straightforward multiplication algorithms2 ). Hence, the computational complexity for each worker is
O(N 3 /m). Thus, as long as P N (and hence m N ), encoding and decoding complexity is much smaller than per-worker
complexity.
Communication cost:
The master node communicates O(P N 2 /m) symbols, and the fusion node receives O(mN 2 ) symbols from the successful
worker nodes. While the master node communication is identical to that in Polynomial codes, the fusion node there only
receives O(m2 N 2 /m2 ) = O(N 2 ) symbols.
B. Why do MatDot exceeds the fundamental limits in [4]
The fundamental limit in [4] concludes that the recovery threshold is Ω(m2 ), whereas our recovery threshold is lower:
2m − 1. To understand why this is possible, one needs to carefully examine the derivation of the fundamental limit in [4],
which uses a cut-set argument to count the number of bits/symbols required for computing the product AB. In doing so, the
authors make the assumption that the number of symbols communicated by each worker to the fusion node is N 2 /m2 , which
is a fallout of a horizontal division of matrix A, and a vertical division of matrix B (the opposite of the division used here).
The bound does not apply to our construction because each worker now communicates N 2 symbols to the fusion node.
Note that while the amount of information in each worker’s transmissions is less, O(N 2 /m) (because the N × N matrices
communicated by the workers have rank N 2 /m), this is still significantly larger than N 2 /m2 assumption made in the
fundamental limits in [4].
From a communication viewpoint, MatDot requires communicating a total of (2m − 1)N 2 symbols, which is larger than the
2
N symbols in the product AB. This is suggestive of a trade-off between minimal number of workers and minimal (sum-rate)
2 More sophisticated algorithms [31] also require super-quadratic complexity in N , so the conclusion will hold if those algorithms are used at workers as
well.
communication from non-straggling workers. Section V describes a unified view of MatDot and Polynomial codes, which
describes the trade-off between worker-fusion communication cost and recovery threshold achieved by our construction.
In practice, whether this increased worker-fusion node communication cost using MatDot codes is worth paying for will
depend on the computational fabric and system implementation choices. Even in systems where communication costs may
by significant, it is possible that more communication from fewer successful workers is less expensive than requiring more
successful workers needed in Polynomial codes. Also note that if P = Ω(m2 ) (e.g. when the system is highly fault prone or
the deadline [24] is very short), communication complexity at the master node will dominate, and hence MatDot codes may
not impose a substantial computing overhead.
IV. S YSTEMATIC C ODE C ONSTRUCTIONS
In this section, we provide a systematic code construction for the problem stated in Section II-B that obtains a recovery
threshold of 2m − 1 where m ≥ 2 is any positive integer that divides N . First, we need to note that a systematic code in
the context of the distributed matrix-matrix multiplication problem is a linear code designed for the problem stated in Section
II-B computed on the system defined in Definition 1 such that the matrices A and B are split as in (2), and the output of the
l-th worker node is the product Al−1 Bl−1 , for all l ∈ {1, · · · , m}. Notice that, in systematic codes, the product AB is the
summation of the output of the first m worker nodes. This follows from noting that the output ofP
the l-th worker node is the
m−1
product Al−1 Bl−1 , for all l ∈ {1, · · · , m}, in the systematic codes along with the fact that AB = i=0 Ai Bi . The presented
systematic code, named systematic MatDot code, is advantageous over MatDot codes in three aspects. First, although both
MatDot and systematic MatDot codes have the same recovery threshold, unlike MatDot codes where for every single instance
of the problem 2m − 1 workers are required to finish before obtaining the result of the computation problem, in many instances
of the systematic MatDot code, the outputs of only m successful workers suffice. For e.g., if the first m worker nodes finish,
then the fusion node can recover AB. Secondly, for the instances of systematic MatDot code in which the outputs of only m
successful workers suffice, the decoding complexity is O(mN 2 ), which is less than the decoding complexity of MatDot codes
O(mN 2 log2 m). Another advantage for systematic MatDot codes over MatDot codes is that the systematic MatDot approach
may be useful for backward-compatibility current-day practice. That is, for systems that are already established and operating
with no straggler tolerance, but do m-way parallelization, it’s easier to apply the systematic approach as the infrastructure
could be appended to additional worker nodes without modifying what the first m nodes are doing.
Theorem 2. For the matrix multiplication problem specified in Section II-B computed on the system defined in Definition 1,
there exists a systematic code, where the product AB is the summation of the output of the first m worker nodes, that solves
this problem with a recovery threshold of 2m − 1, where m ≥ 2 is any positive integer that divides N .
Before we describe the general systematic MatDot code construction that will be used to prove Theorem 2, we first present
the following simple systematic MatDot code example.
Example 4. [Systematic MatDot code, m = 2, k = 3]
Matrix A is split vertically into two submatrices A0 and A1 , each of dimension N ×
into two submatrices B0 and B1 , each of dimension N2 × N as follows:
B0
A = [ A0 | A1 ] , B =
.
B1
N
2
and matrix B is split horizontally
(5)
0
1
1
+ A1 xx−x
and pB (x) = B0 xx−x
+
Now, we define the encoding functions pA (x) and pB (x) as pA (x) = A0 xx−x
0 −x1
1 −x0
0 −x1
x−x0
B1 x1 −x0 , for distinct x0 , x1 ∈ F. Let x2 , · · · , xP −1 be elements of F such that x0 , x1 , x2 , · · · , xP −1 are distinct, the master
node sends pA (xl−1 ) and pB (xl−1 ) to the l-th worker node, l ∈ {1, · · · , P }, where the l-th worker node performs the
multiplication pA (xl−1 )pB (xl−1 ) and sends the output to the fusion node. The exact computations at each worker node are
depicted in Fig. 5.
We can observe that the outputs of the worker nodes 1, 2 are A0 B0 , A1 B1 , respectively, i.e., this code is systematic. In
addition, assuming, for some instance, that these two worker nodes, worker nodes 1, 2, send their computation outputs first, the
fusion node can directly obtain the product AB by performing the summation A0 B0 + A1 B1 , since AB = A0 B0 + A1 B1 .
Otherwise, assume, for some other instance, that the first three worker nodes to finish do not include worker node 1 or worker
node 2. That is, for example, let worker nodes 1, 3, 4 be the first three worker nodes to send their computation outputs to the
fusion node, then the fusion node owns the following three matrices: pA (x0 )pB (x0 ), pA (x2 )pB (x2 ), and pA (x3 )pB (x3 ). Since
these three matrices can be seen as three evaluations of the matrix polynomial pA (x)pB (x) of degree two at three distinct
evaluation points x0 , x2 , x3 , the matrix polynomial pA (x)pB (x) can be obtained using polynomial interpolation. Finally, the
product AB is obtained by summing the evaluations of the interpolated matrix polynomial pA (x)pB (x) at x0 and x1 . That is,
pA (x0 )pB (x0 ) + pA (x1 )pB (x1 ) = A0 B0 + A1 B1 = AB.
Fig. 5. An illustration of the computational system with four worker nodes and applying systematic MatDot codes with m = 2. The recovery threshold is 3.
The following describes the general construction of the systematic codes for matrix-matrix multiplication.
Construction 2. [Systematic MatDot codes]
Splitting of input matrices: A and B are split as in (2).
Pm−1
Master node (encoding): Let x0 , x1 , . . . , xP −1 be arbitrary distinct elements of F. Let pA (x) =
i=0 Ai Li (x) and
Pm−1
pB (x) = i=0 Bi Li (x), such that
Y
x − xj
Li (x) =
,
(6)
xi − xj
j∈{0,··· ,m−1}−{i}
where i ∈ {0, · · · , m − 1}.
The master node sends to the l-th worker evaluations of pA (x), pB (x) at x = xl−1 , that is, it sends pA (xl−1 ), pB (xl−1 ) to
the l-th worker, l ∈ {1, 2, . . . , P }.
Worker nodes: For l ∈ {1, 2, . . . , P }, the l-th worker node computes the matrix product C(xl−1 ) = pA (xl−1 )pB (xl−1 ) and
sends it to the fusion node on successful completion.
Fusion node (decoding): For any k such that m ≤ k ≤ 2m − 1, whenever the outputs of the first k workers to finish
contain the outputs of the worker nodes 1, · · · , m, i.e., {C(xl−1 )}l∈{1,··· ,m} is contained in the set of the first k evaluations
Pm−1
received by the fusion node, the fusion node performs the summation i=0 C(xi ). Otherwise, if {C(xl−1 )}l∈{1,··· ,m} is not
contained in the set of the first 2m − 1 evaluations received by the fusion node, the fusion node performs the following steps:
(i) interpolates the polynomial C(x) = pA (x)pB (x) (the feasibility of P
this step will be shown later in the proof of Theorem
m−1
2), (ii) evaluates C(x) at x0 , · · · , xm−1 , (iii) performs the summation i=0 C(xi ).
If the number of successful worker nodes is smaller than 2m − 1 and the first m worker nodes are not included in the
successful worker nodes, the fusion node declares a failure.
Remark 1. The polynomials Li (x) with i ∈ {0, · · · , m − 1} defined in (6) have the following evaluations at x0 , · · · , xm−1 ,
(
1 if j = i,
Li (xj ) =
(7)
0 if j ∈ {0, · · · , m − 1} − {i},
Lemma 1. For Construction 2, the output of the l-th worker node, for l ∈ {1, · · · , m}, is the product Al−1 Bl−1 . That is,
Construction 2 is a systematic code for matrix-matrix multiplication.
Proof of Lemma 1. The lemma follows from noting that the output of the l-th worker, for l ∈ {1, · · · , m}, can be written as
C(xl−1 ) = pA (xl−1 )pB (xl−1 )
=
m−1
X
Ai Li (xl−1 )
i=0
= Al−1 Bl−1 ,
where the last equality follows from Remark 1.
m−1
X
Bi Li (xl−1 )
i=0
(8)
Now, we proceed with the proof of Theorem 2.
Proof of Theorem 2. Since Construction 2 is a systematic code for matrix-matrix multiplication (Lemma 1), in order to prove
the theorem, it suffices to show that Construction 2 is a valid construction with a recovery threshold k = 2m − 1.
For the first case in which {C(xl−1 )}l∈{1,··· ,m} is contained in the set of the first k evaluations received by the fusion
node, for some k ∈ {m, · · · , 2m − 1}, since Construction 2 is systematic (Lemma 1), the product AB is the summation of
the outputs of the first m workers.
For the other case in which {C(xl−1 )}l∈{1,··· ,m} is not contained in the set of the first 2m − 1 evaluations received by the
fusion node, we need to argue that the fusion node is able to construct AB. P
From (6), observe that the polynomials
Pm−1 Li (x),
m−1
i ∈ {0, · · · , m − 1}, have degrees m − 1 each. Therefore, each of pA (x) = i=0 Ai Li (x) and pB (x) = i=0 Bi Li (x)
has a degree of m − 1 as well. Consequently, C(x) = pA (x)pB (x) has a degree of 2m − 2. Now, because the polynomial
C(x) has degree 2m − 2, evaluation of the polynomial at any 2m − 1 distinct points is sufficient to interpolate C(x) using
polynomial interpolation algorithm. Now, since Construction
2 is systematic (Lemma 1), the product AB is the summation of
Pm−1
the outputs of the first m workers, i.e., AB = i=0 C(xi ). Therefore, after the fusion node interpolates C(x), evaluating
Pm−1
C(x) at x0 , · · · , xm−1 , and performing the summation i=0 C(xi ) yields the product AB.
The next section has a complexity analysis that shows that master and fusion nodes have a lower computational complexity
as compared with the workers.
A. Complexity analyses of the systematic codes
Apart from the encoding/decoding complexity, the complexity analyses of Construction 2 are the same as their MatDot
codes counterparts. In the following, we investigate the encoding/decoding complexity of Construction 2.
Encoding/decoding Complexity: For decoding, first, for the interpolation step, we interpolate a 2m−2 degree polynomial for
N 2 elements. Using polynomial interpolation algorithms of complexity O(k log2 k) [30], where k = 2m − 1, the interpolation
complexity per matrix element is O(m log2 m). Thus, for N 2 elements, the interpolation complexity is O(N 2 m log2 m). For
the evaluation of C(x) at x0 , · · · , xm−1 , each evaluation involves adding 2m − 1 scaled matrices of size N 2 with a complexity
of O(mN 2 ). Hence, for all m evaluations the complexity is O(m2 N 2 ). Finally, the complexity of the final addition of m
matrices of size N 2 is O(mN 2 ). Hence, the overall decoding complexity is O(m2 N 2 ).
Encoding for each worker requires first performing evaluations of polynomials Li (x) for all i ∈ {0, · · · , m − 1}, with each
evaluation requires O(m) operations. This gives O(m2 ) operations for all polynomial evaluations. Afterwards, two additions
are performed, each adding m scaled matrices of size N 2 /m, with complexity O(mN 2 /m) = O(N 2 ). Therefore, the overall
encoding complexity for each worker is O(max(N 2 , m2 )) = O(N 2 ). Thus, the overall computational complexity of encoding
for P workers is O(N 2 P ).
Since the encoding complexity of systematic MatDot codes is O(N 2 P ), and that the computational cost of each worker
in MatDot codes, and hence in systematic MatDot codes, is O(N 3 /m), we conclude that as long as P N (and hence
m N ), encoding and decoding complexity is much smaller than per-worker complexity for Construction 2.
V. U NIFYING M ATDOT AND P OLYNOMIAL C ODES : T RADE - OFF BETWEEN COMMUNICATION COST AND RECOVERY
THRESHOLD
A. PolyDot Codes: A trade-off between communication cost and recovery threshold
In this section, we present a code construction, named PolyDot, that provides a trade-off between communication cost and
recovery threshold. While Polynomial codes [4] have a higher recovery threshold of m2 , and a lower communication cost of
O(N 2 /m2 ) per worker node, MatDot codes have a lower recovery threshold of 2m − 1, but a higher communication cost of
O(N 2 ) per worker node. Here, we present PolyDot codes, which may be viewed as an interpolation of MatDot codes and
Polynomial codes. For the proposed codes, we consider the matrix computation problem specified in Section II-B computed
on the system defined in Definition 1. The next theorem shows that, for any positive integers s, t such that st = m and both
s and t divide N , the product AB can be recovered using outputs from any t2 (2s − 1) worker nodes with a communication
cost from each worker node to the fusion node of O(N 2 /t2 ).
Theorem 3. For the matrix multiplication problem specified in Section II-B computed on the system defined in Definition 1,
there exist codes with a recovery threshold of t2 (2s − 1) and a communication cost from each worker node to the fusion node
bounded by O(N 2 /t2 ) for any positive integers s, t such that st = m and both s and t divide N .
Before describing the general PolyDot code construction that will be used to prove Theorem 3, we first introduce the
following simple PolyDot code example.
Example 5. [PolyDot code, m = 4, s = 2, k = 12]
Matrix A is split into submatrices A0,0 , A0,1 , A1,0 , A1,1 each of dimension N/2 × N/2. Similarly, matrix B is split into
submatrices B0,0 , B0,1 , B1,0 , B1,1 each of dimension N/2 × N/2 as follows:
A0,0 A0,1
B0,0 B0,1
A=
,B =
.
(9)
A1,0 A1,1
B1,0 B1,1
Notice that, from (9), the product AB can be written as
P1
A B
AB = P1i=0 0,i i,0
A
1,i Bi,0
i=0
P1
A B
P1i=0 0,i i,1 .
i=0 A1,i Bi,1
(10)
Now, we define the encoding functions pA (x) and pB (x) as
pA (x) = A0,0 + A1,0 x + A0,1 x2 + A1,1 x3 ,
pB (x) = B0,0 x2 + B1,0 + B0,1 x8 + B1,1 x6 .
Observe the following:
P1
(i) the coefficient of x2 in pA (x)pB (x) is i=0 A0,i Bi,0 ,
P
1
(ii) the coefficient of x8 in pA (x)pB (x) is i=0 A0,i Bi,1 ,
P1
3
(iii) the coefficient of x in pA (x)pB (x) is i=0 A1,i Bi,0 , and
P1
(iv) the coefficient of x9 in pA (x)pB (x) is i=0 A1,i Bi,1 .
Let x0 , · · · , xP −1 be distinct elements of F, the master node sends pA (xl−1 ) and pB (xl−1 ) to the l-th worker node,
l ∈ {1, · · · , P }, where the l-th worker node performs the multiplication pA (xl−1 )pB (xl−1 ) and sends the output to the fusion
node.
Let worker nodes 1, · · · , 12 be the first 12 worker nodes to send their computation outputs to the fusion node, then the fusion
node owns the matrices pA (xl−1 )pB (xl−1 ) for all l ∈ {1, · · · , 12}. Since these 12 matrices can be seen as 12 evaluations of
the matrix polynomial pA (x)pB (x) of degree 11 at 12 distinct evaluation points x0 , · · · , x11 , the coefficients of the matrix
polynomial pA (x)p
includes the coefficients of xr+2+6c for all
PB1(x) can be obtained using polynomial interpolation.
PThis
1
r, c ∈ {0, 1}, i.e., i=0 Ar,i Bi,c for all r, c ∈ {0, 1}. Once the matrices i=0 Ar,i Bi,c for all r, c ∈ {0, 1} are obtained, the
product AB is obtained by (10).
Notice that although only the parameters m and s are sufficient to characterize a PolyDot code, we use parameter t in addition
to m and s to characterize such codes for readability. The following describes the general construction of PolyDot(m, s, t)
codes.
Construction 3. [PolyDot(m, s, t) codes]
Splitting of input matrices: A and B are split both horizontally and vertically:
A0,0
···
A0,s−1
..
..
..
A=
,
.
.
.
At−1,0 · · · At−1,s−1
B=
B0,0
..
.
Bs−1,0
···
..
.
···
B0,t−1
..
.
,
(11)
Bs−1,t−1
where, for i = 0, · · · , s − 1, j = 0, · · · , t − 1, Aj,i ’s are N/t × N/s matrices and Bi,j ’s are N/s × N/t matrices. We choose
s and t such that both s and t divide N and st = m.
Master node (encoding): Define the encoding polynomials as
pA (x, y) =
t−1 X
s−1
X
Ai,j xi y j ,
i=0 j=0
pB (y, z) =
s−1 X
t−1
X
Bk,l y s−1−k z l .
(12)
k=0 l=0
t(2s−1)
The master node sends to the r-th worker evaluations of pA (x, y), pB (y, z) at x = xr , y = xtr , z = xr
distinct for r ∈ {1, 2, . . . , P }.
. xr ’s are all
Fig. 6. An illustration of the trade-off between communication cost (from the workers to the fusion node) and recovery threshold of PolyDot codes by varying
s and t for m = 36. The minimum communication cost is N 2 , corresponding to polynomial codes, that have the largest recovery threshold. It is important
to note here that in the above, we are only including the communication cost from the workers to the fusion node. The communication from the master node
to the workers is not included, and it can dominate in situations when the workers are highly unreliable.
Worker nodes: For r ∈ {1, 2, . . . , P }, the r-th worker node computes the matrix product C(xr , yr , zr ) = pA (xr , yr )pB (yr , zr )
and sends it to the fusion node on successful completion.
Fusion node (decoding): The fusion node uses outputs of any t2 (2s − 1) successful workers to compute the coefficient
of xi−1 y s−1 z l−1 in C(x, y, z) = pA (x, y)pB (y, z). That is, it computes the coefficient of xi−1+(s−1)t+(2s−1)t(l−1) of the
transformed single-variable polynomial. The proof of Theorem 3 shows that this is indeed possible. If the number of successful
workers is smaller than t2 (2s − 1), the fusion node declares a failure.
Remark 2. For PolyDot codes, we transform polynomials in 3 variables x, y, z to a polynomial in one variable x by choosing
y = xt and z = xt(2s−1) . In fact, we show later in this section that this transformation is one-to-one.
Before we prove the theorem, we discuss the utility of PolyDot codes. By choosing different s and t, we can trade off
communication cost and recovery threshold. For s = m and t = 1, PolyDot(m, s = m, t = 1) code is a MatDot code which has
a low recovery threshold but high communication cost. At the other extreme, for s = 1 and t = m, PolyDot(m, s = 1, t = m)
code is a Polynomial code.
√
√
√
√
Now let us consider a code with intermediate
s and t values such as s = m and t = m. A Polydot(m, s = m, t = m)
√
1.5
code has a recovery
of m(2
√ threshold
√ m − 1) = Θ(m ), and the total number of symbols to be communicated to the fusion
node is Θ (N/ m)2 · m1.5 = O( mN 2 ) which is smaller than Θ(mN 2 ) required by MatDot codes, and larger than Θ(N 2 )
required by Polynomial codes. This tradeoff is illustrated in Fig. 6 for m = 36.
To prove Theorem 3, we need the following lemma.
Lemma 2. The following function
f :{0, · · · , t − 1} × {0, · · · , 2s − 2} × {0, · · · , t − 1}
→ {0, · · · , t2 (2s − 1) − 1}
(α, β, γ) 7→ α + tβ + t(2s − 1)γ
(13)
is a bijection.
Proof. Let us assume, for the sake of contradiction, that for some (α0 , β 0 , γ 0 ) 6= (α, β, γ), f (α0 , β 0 , γ 0 ) = f (α, β, γ). Then
f (α, β, γ) mod t = α = f (α0 , β 0 , γ 0 ) mod t = α0 and hence α = α0 . Similarly, f (α, β, γ) mod t(2s−1) = f (α0 , β 0 , γ 0 ) mod t(2s−
1) gives α + tβ = α0 + tβ 0 , and thus β = β 0 (because α = α0 ). Now, because α = α0 and β = β 0 , as we just established,
using f (α, β, γ) = f (α0 , β 0 , γ 0 ), it follows that γ = γ 0 .
Proof of Theorem 3. The product of pA (x, y) and pB (y, z) can be written as follows:
C(x, y, z) = pA (x, y)pB (y, z)
!
t−1 X
s−1
s−1 X
t−1
X
X
=
Ai,j xi y j
Bk,l y s−1−k z l
i=0 j=0
=
X
k=0 l=0
i s−1+j−k l
Ai,j Bk,l x y
z.
(14)
i,j,k,l
Note that the coefficient of xi−1 y s−1 z l−1 in C(x, y, z) is equal to Ci,l =
z = xt(2s−1) we can further simplify C(x, xt , xt(2s−1) ):
Ps−1
k=0
Ai,k Bk,l . By our choice of y = xt and
C(x, y, z) = C(x)
X
=
Ai,j Bk,l xi+t(s−1+j−k)+t(2s−1)l .
(15)
i,j,k,l
The maximum degree of this polynomial is when i = t − 1, j − k = s − 1 and l = t − 1, which is (t − 1) + (2s − 2)t +
t(2s − 1)(t − 1) = t2 (2s − 1) − 1. Furthermore, if we let α = i, β = s − 1 + j − k, γ = l, the function f (α, β, γ) in Lemma 2
is the degreePof x in (15). This implies that for different pairs of (i, j − k, l), we get different powers of x. When j − k = 0,
s−1
we obtain ( k=0 Ai,k Bk,l )xi+t(s−1)+t(2s−1)l = Ci,l xi+t(s−1)+t(2s−1)l which is the desired product we want to recover.
This implies that if we have t2 (2s − 1) successful worker nodes, we can compute all the coefficients in (15) by polynomial
interpolation. Hence, we can recover all Ci,l ’s for i, l = 0, · · · , t − 1.
B. Complexity analyses of PolyDot codes
Encoding/decoding complexity: Decoding requires interpolating a polynomial of degree t2 (2s − 1) − 1 for N 2 elements.
Using polynomial interpolation algorithms of complexity O(k log2 k) [30], [32] where k = t2 (2s−1), the decoding complexity
per matrix element is O(t2 (2s − 1) log2 t2 (2s − 1)). Thus, the overall decoding complexity is O(N 2 t2 (2s − 1) log2 t2 (2s − 1))
Encoding for each worker requires performing two additions, each adding m scaled matrices of size N 2 /m, for an overall
encoding complexity for each worker of O(mN 2 /m) = O(N 2 ). Thus, the overall computational complexity of encoding for
P worker nodes is O(N 2 P ).
3
Each worker’s computational complexity: Multiplication of matrices of size N/t × N/s and N/s × N/t requires O( N
st2 )
computations.
Communication complexity: Master node communicates O(N 2 /ts) = O(N 2 /m) symbols to each worker, hence total
outgoing symbols from the master node will be O(P N 2 /M ). For decoding, each node sends O(N 2 /t2 ) symbols to the fusion
node and recovery threshold is O(t2 (2s − 1)). Total number of symbols communicated to the fusion node is O((2s − 1)N 2 ).
VI. M ULTIPLYING MORE THAN TWO MATRICES
We now present our coding technique for multiplying n matrices. Before introducing our result for multiplying more than
two matrices, we first provide a formal description of the problem.
A. Problem Statement
Qn
Compute the product C = i=1 Ãi of N × N square matrices Ã1 , · · · , Ãn in the computational system specified in
Section II-A. For the ease of readability of this section, we define
(
Adi/2e if i is odd,
Ãi =
(16)
Bi/2
if i is even,
for all i ∈ {1, · · · , n}. Using (16), C can be written as
(Q n
2
i=1 Ai Bi
C = Qb n2 c
n
i=1 Ai Bi Ad 2 e
if n is even,
if n is odd.
(17)
Each worker can receive at most nN 2 /m symbols from the master node, where each symbol is an element of F. Similar to
Section II-B, the computational complexities of the master and fusion nodes, in terms of the matrix parameter N , are required
to be strictly less than the computational complexity at any worker node. The goal is to perform this matrix product utilizing
faulty/delay prone workers with as low recovery threshold as possible. Again, in the sequel, we will assume that |F| > P .
B. Codes for n−matrix multiplication
Theorem 4 (Recovery threshold for multiple matrix multiplications). For the matrix multiplication problem specified in Section
VI-A computed on the system defined in Definition 1, there exists a code with a recovery threshold of
2mn/2 − 1
if n is even,
(18)
k(n, m) =
c
bn
2
− 1 if n is odd.
(m + 1)m
Before we describe the general code construction that will be used to prove Theorem 4, we first present the following simple
examples.
Example 6. [Multiplying 4 matrices m = 2, k = 7]
Here, we give an example of multiplying 4 matrices and show that a recovery threshold of 7 is achievable. For i ∈ {1, 2},
matrix Ai is split vertically into submatrices Ai1 , Ai2 each of dimension N × N2 as follows: Ai = [Ai1 Ai2 ], while, for
i ∈ {1, 2}, matrix Bi is split horizontally into submatrices Bi1 , Bi2 each of dimension N2 × N as follows:
Bi1
Bi =
.
(19)
Bi2
Q2
Notice that the product i=1 Ai Bi can now be written as
2
Y
Ai Bi = (A1 B1 ) (A2 B2 )
i=1
= (A11 B11 + A12 B12 ) (A21 B21 + A22 B22 ) .
(20)
Now, we define the encoding polynomials pAi (x), pBi (x), i ∈ {1, 2} as follows:
pA1 (x) = A11 + A12 x,
pB1 (x) = B11 x + B12 ,
pA2 (x) = A21 + A22 x,
pB2 (x) = B21 x + B22 .
(21)
From (21), we have
pA1 (x)pB1 (x) =
A11 B12 + (A11 B11 + A12 B12 )x + A12 B11 x2 ,
pA2 (x2 )pB2 (x2 ) =
A21 B22 + (A21 B21 + A22 B22 )x2 + A22 B21 x4 .
(22)
From (20) along with (22), we can observe the following:
(i) the coefficient of x in pA1 (x)pB1 (x) is A11 B11 + A12 B12 = A1 B1 ,
(ii) the coefficient of x2 in pA2 (x2 )pB2 (x2 ) is the product A21 B21 + A22Q
B22 = A2 B2 , and
2
(iii) the coefficient of x3 in pA1 (x)pB1 (x)pA2 (x2 )pB2 (x2 ) is the product i=1 Ai Bi .
Let x0 , · · · , xP −1 be distinct elements of F, the master node sends pAi (xil−1 ) and pBi (xil−1 ), for all i ∈ {1, 2}, to the l-th
Q2
worker node, l ∈ {1, · · · , P }, where the l-th worker node performs the multiplication i=1 pAi (xil−1 )pBi (xil−1 ) and sends
the output to the fusion node.
Let worker nodes 1, · · · Q
, 7 be the first 7 worker nodes to send their computation outputs to the fusion node, then the fusion
2
node receives the matrices i=1 pAi (xil−1 )pBi (xil−1 ) for all l ∈ {1, · · · , 7}. Since these 7 matrices can be seen as 7 evaluations
Q2
of the matrix polynomial i=1 pAi (xi )pBi (xi ) of degree 6 at 7 distinct evaluation points x0 , · · · , x6 , the coefficients of the
Q2
matrix polynomial i=1 pAi (xi )pBi (xi ) can be obtained using polynomial interpolation. This includes the coefficient of x3 ,
Q2
i.e., i=1 Ai Bi .
Example 7. [Multiplying 3 matrices, m = 2, k = 5]
Here, we give an example of multiplying 3 matrices and show that a recovery threshold of 5 is achievable. In this example,
we have three input matrices A1 , B1 , and A2 , each of dimension N × N and need to compute the product A1 B1 A2 . First,
the three input matrices are split in the same way as in Example 6. The product A1 B1 A2 can now be written as
A1 B1 A2 = [A1 B1 A21
where A1 B1 = A11 B11 + A12 B12 .
A1 B1 A22 ] ,
(23)
Now, we define the encoding polynomials pA1 (x), pB1 (x), pA2 (x) as follows:
pA1 (x) = A11 + A12 x,
pB1 (x) = B11 x + B12 ,
pA2 (x) = A21 + A22 x.
(24)
From (24), we have
pA1 (x)pB1 (x)pA2 (x2 )
= A11 B12 A21 + (A11 B11 + A12 B12 )A21 x
+ (A12 B11 A21 + A11 B12 A22 )x2
+ (A11 B11 + A12 B12 )A22 x3 + A12 B11 A22 x4 .
(25)
From (25), we can observe the following:
(i) the coefficient of x in pA1 (x)pB1 (x)pA2 (x2 ) is the product A1 B1 A21 , and
(ii) the coefficient of x3 in pA1 (x)pB1 (x)pA2 (x2 ) is the product A1 B1 A22 .
Let x0 , · · · , xP −1 be distinct elements of F, the master node sends pAi (xil−1 ), for all i ∈ {1, 2}, and pB1 (xl−1 ) to the
l-th worker node, l ∈ {1, · · · , P }, where the l-th worker node performs the multiplication pA1 (xl−1 )pB1 (xl−1 )pA2 (x2l−1 ) and
sends the output to the fusion node.
Let worker nodes 1, · · · , 5 be the first 5 worker nodes to send their computation outputs to the fusion node, then the
fusion node receives the matrices pA1 (xl−1 )pB1 (xl−1 )pA2 (x2l−1 ) for all l ∈ {1, · · · , 5}. Since these 5 matrices can be seen
as 5 evaluations of the matrix polynomial pA1 (x)pB1 (x)pA2 (x2 ) of degree 4 at 5 distinct evaluation points x0 , · · · , x4 , the
coefficients of the matrix polynomial pA1 (xl−1 )pB1 (xl−1 )pA2 (x2l−1 ) can be obtained using polynomial interpolation. This
includes the coefficients of x and x3 , i.e., A1 B1 A21 and A1 B1 A22 .
In the following, we present the general code construction for n−matrix multiplication.
Construction 4. [An n−matrix multiplication code]
Splitting of input matrices: for every i ∈ {1, · · · , d n2 e} and j ∈ {1, · · · , b n2 c}, Ai and Bj are split as follows
Bj1
Bj2
Ai = [Ai1 |Ai2 | . . . |Aim ] , Bj = . ,
..
(26)
Bjm
where, for k ∈ {1, . . . , m}, Aik , Bjk are N × N/m and N/m × N dimensional matrices, respectively.
node (encoding): Let x0 , x1 , . . . , xP −1 be arbitrary
distinct elements of F. For i ∈ {1, · · · , d n2 e}, define pAi (x) =
PMaster
Pm
m
n
j−1
, and, for i ∈ {1, · · · , b 2 c}, define pBi (x) = j=1 Bij xm−j . For l ∈ {1, 2, . . . , P }, the master node sends to
j=1 Aij x
j−1
n
n
mi−1
the l-th worker evaluations of pAi (xl−1
) and pBj (xm
l−1 ), for all i ∈ {1, · · · , d 2 e} and j ∈ {1, · · · , b 2 c}.
n
Worker nodes: For i ∈ {1, · · · , d 2 e}, define
(
pAi (x)pBi (x) if i ∈ {1, · · · , b n2 c},
pCi (x) =
(27)
pAi (x)
if n is odd and i = d n2 e.
dne
i−1
2
For l ∈ {1, 2, . . . , P }, the l-th worker node computes the matrix product Πi=1
pCi (xm
l−1 ) and sends it to the fusion node on
successful completion.
n
Fusion node (decoding): If n is even, the fusion node
uses outputs of any 2m 2 − 1 successful workers to compute
n
i−1
mn/2 −1
m
2
the coefficient of x
in the matrix polynomial Πi=1
pCi (x
), nand if n is odd, the fusion node uses outputs of any
b c
2 −1
bn
c
jm
m 2 (m + 1) − 1 successful workers to compute the coefficients of x
, for all j ∈ {1, · · · , m}, in the matrix polynomial
n
d2e
mi−1
) (the feasibility of this step will be shown later in the proof of Theorem 4).
Πi=1 pCi (x
n
n
If the number of successful workers is smaller than 2m 2 − 1 for even n or smaller than mb 2 c (m + 1) − 1 for odd n, the
fusion node declares a failure.
Pm
i
i−1
i−1
Remark 3. The coefficient of xm −m
in pCi (xm ), for any i ∈ {1, · · · , b n2 c}, is j=1 Aij Bij = Ai Bi .
In the following, we show Lemmas 3 and 4 which state properties of coefficients of products of polynomials. Using Lemmas
3 and 4, we show Claims 5 and 6 which demonstrate that the product C is contained in a set of coefficients of the matrix
dne
i−1
2
polynomial Πi=1
pCi (xm ), where pCi (x) is as defined in (27), for i ∈ {1, · · · , dn/2e}. Finally, we provide a proof of
Theorem 4 using Claims 5 and 6.
P2di−1 −2
P2d−2
Lemma 3. If p(x) = j=0
pj xj is a polynomial with degree 2di−1 − 2 for some i ≥ 2, and q(x) = j=0 qj xj is any
i
i−1
other polynomial with degree 2d − 2, then pdi−1 −1 qd−1 is the coefficient of xd −1 in p(x)q(xd ).
Proof. We first expand out p(x) and q(x) as following:
p(x) =
di−1
X−2
i−1
pj xj +pdi−1 −1 xd
−1
i−1
2dX
−2
+
j=0
|
(28)
j=di
{z
}
p̃1 (x)
q(x) =
p j xj
d−2
X
|
qj xj +qd−1 xd−1 +
j=0
2d−2
X
{z
p̃2 (x)
}
q j xj
(29)
j=d
| {z }
{z
|
q̃1 (x)
q̃2 (x)
}
i−1
We show that the term of degree di − 1 in p(x)q(xd ) is only generated by multiplication of the term of degree di−1 − 1
i−1
in p(x) and the term of degree di−1 (d − 1) in q(xd ). For this purpose, we consider following terms:
i−1
1. Consider the multiplication of two lowest degree terms in p̃1 (x) and q̃2 (xd ) of equations (28) and (29). That is,
i
i
i
qdi xd p0 = p0 qdi xd which has higher degree in comparison to xd −1 . Consequently, the degree of any term in the
i−1
multiplication of p̃1 (x) and q̃2 (xd ) will be strictly greater than di − 1.
i−1
2. Consider the multiplication of two highest degree of terms in p̃2 (x) and q̃1 (xd ) of equations (28) and (29),
i−1
qd−2 xd
(d−2)
i−1
p2di−1 −2 x2d
−2
i
= qd−2 p2di−1 −2 xd
−2
i−1
is less than di − 1. Consequently, the degree of any term in the multiplication of p̃2 (x) and q̃1 (xd ) will be strictly less
than di − 1.
i−1
3. Since the degree of any term in the multiplication of p̃2 (x) and q̃1 (xd ) is strictly less than di − 1, and any term in
p̃1 (x) has degree less than the degree of any term in p̃2 (x), we conclude that any term in the multiplication of p̃1 (x) and
i−1
q̃1 (xd ) has degree strictly less than di − 1.
i−1
4. Since the degree of any term in the multiplication of p̃1 (x) and q̃2 (xd ) is strictly greater than di − 1, and any term in
i−1
p̃2 (x) has degree larger than the degree of any term in p̃1 (x), we conclude that any term in p̃2 (x)q̃2 (xd ) has degree
i
strictly greater than d − 1 which completes the proof.
P2di−1 −2
Pd−1
j
i−1
j
Lemma 4. If p(x) = j=0
pj x is a polynomial with degree 2d
− 2 for some i ≥ 2, and q(x) = j=0 qj x is any
i−1
i−1
other polynomial with degree d − 1, then, for 0 ≤ j ≤ d − 1, pdi−1 −1 qj are the coefficients of x(j+1)d −1 in p(x)q(xd ).
Proof. First, we expand out p(x) as in (28), and expand q(x) as follows:
q(x) =
j−1
X
d−1
X
qk xk +qj xj +
k=0
k=j+1
| {z }
|
q̃1 (x)
i−1
qj xj
{z
q̃2 (x)
(30)
}
i−1
In order to prove Lemma 4, we show that the term of degree x(j+1)d −1 in p(x)q(xd ) is produced by the multiplication
i−1
i−1
i−1
of only the term pdi−1 −1 xd −1 in p(x) with the term qj xj(d ) in q(xd ). First, it is clear that the product of the term
di−1 −1
j(di−1 )
di−1
pdi−1 −1 x
in p(x) with the term qj x
in q(x
) has degree (j + 1)di−1 − 1. Thus, to complete the proof, we
i−1
i−1
show that no other terms in p(x) produce a degree of x(j+1)d −1 when multiplied with any term in q(xd ). To do so, we
consider the following terms:
i−1
1. Consider the multiplication of two lowest degree terms in p̃1 (x) and q̃2 (xd ) of equations (28) and (30). That is,
i−1
i−1
p0 qj+1 x(j+1)d
which has higher degree in comparison to x(j+1)d −1 . Consequently, the degree of any term in the
i−1
multiplication of p̃1 (x) and q̃2 (xd ) will be strictly greater than (j + 1)di−1 − 1.
i−1
2. Consider the multiplication of two highest degree of terms in p̃2 (x) and q̃1 (xd ) of equations (28) and (30), the product
i−1
qj−1 x(j−1)d
i−1
p2di −2 x2d
−2
i−1
= qj−1 p2di −2 x(j+1)d
−2
i−1
has degree less than (j + 1)di−1 − 1. Consequently, the degree of any term in the multiplication of p̃2 (x) and q̃1 (xd )
will be strictly less than (j + 1)di−1 − 1.
i−1
3. Since the degree of any term in the multiplication of p̃2 (x) and q̃1 (xd ) is strictly less than (j + 1)di−1 − 1, and the
degree of any term in p̃1 (x) is less than the degree of any term in p̃2 (x), we conclude that any term in the multiplication
i−1
of p̃1 (x) and q̃1 (xd ) has degree strictly less than (j + 1)di−1 − 1.
i−1
4. Since the degree of any term in the multiplication of p̃1 (x) and q̃2 (xd ) is strictly greater than (j + 1)di−1 − 1, and the
i−1
degree of any term in p̃2 (x) is larger than the degree of any term in p̃1 (x), we conclude that any term in p̃2 (x)q̃2 (xd )
has degree strictly greater than (j + 1)di−1 − 1 which completes the proof.
Now, we are able to state the following claims.
Qb n2 c
Qb n2 c
bnc
i−1
Claim 5. The coefficient of xm 2 −1 in i=1
pCi (xm ) is i=1
Ai Bi , where, for i ∈ {1, · · · , b n2 c}, pCi (x) is as defined
in (27).
Proof. We prove the claim iteratively. Since pC1 (x) has degree 2mi−1 − 2 with i = 2, and pC2 (x) has degree 2m − 2, we
2
have, by Lemma 3, that the coefficient of xm −1 in pC1 (x)pC2 (xm ) is the product of the coefficient of xm−1 in pC1 (x)
2
and the coefficient of xm −m in pC2 (xm ). However, from Remark 3, we already know that A1 B1 is the coefficient of xm−1
2
2
in pC1 (x) and that A2 B2 is the coefficient of xm −m in pC2 (xm ). Therefore, A1 B1 A2 B2 is the coefficient of xm −1 in
pC1 (x)pC2 (xm ).
Similarly, consider the two polynomials p0 (x) = pC1 (x)pC2 (xm ) and pC3 (x). Notice that p0 (x) has degree 2mi−1 − 2
3
2
with i = 3, and pC3 (x) has degree 2m − 2, therefore, from Lemma 3, the coefficient of xm −1 in p0 (x)pC3 (xm ) is the
m2 −1
0
m3 −m2
m2
product of the coefficient of x
in p (x) and the coefficient of x
in pC3 (x ). However, from the previous step, we
2
already know that A1 B1 A2 B2 is the coefficient of xm −1 in p0 (x). In addition, from Remark 3, we already know that A3 B3
3
2
2
3
2
is the coefficient of xm −m in pC3 (xm ). Therefore, A1 B1 A2 B2 A3 B3 is the coefficient of xm −1 in p0 (x)pC3 (xm ) =
2
pC1 (x)pC2 (xm )pC3 (xm ).
Qb n2 c
Qb n2 c
bnc
i−1
Ai Bi is the coefficient of xm 2 −1 in i=1
pCi (xm ).
Repeating the same procedure, we conclude that i=1
Q n
bnc
b2c
jm 2 −1
n
in
Claim 6. If n ≥ 3 and odd, then, for any j ∈ {1, · · · , m},
i=1 Ai Bi Ad 2 ej is the coefficient of x
Qd n2 e
n
mi−1
), where, for i ∈ {1, · · · , d 2 e}, pCi (x) is as defined in (27).
i=1 pCi (x
bnc
i−1
2
pCi (xm ) is
Proof. First, notice that since the degree of pCi (x) is 2m − 2 for all i ∈ {1, · · · , b n2 c}, the degree of Πi=1
Pb n2 c i−1
bn
c
(2m − 2) i=1 m
= 2m 2 − 2. In addition, the matrix polynomial pCd n e (x) has degree m − 1. Therefore, from Lemma
4, for 1 ≤ j ≤ m, the product of the coefficient of xm
n
jmb 2 c −1
dn
2e
i−1
bnc
2 −1
bnc
2
2
in Πi=1
pCi (xm
i−1
) and the coefficient of xj−1 in pCd n e (x) is
2
bnc
m 2 −1
in Πi=1 pCi (xm ). However, we already know, from Claim 5, that the coefficient
in
the coefficient of x
Q n of x
n
n
Q
i−1
b2c
b2c
b2c
n
Πi=1
pCi (xm ) is i=1
Ai Bi , also, by definition, the coefficient of xj−1 in pCd n e (x) is Ad n2 ej . Thus,
A
B
A
i
i
d 2 ej
i=1
2
Qd n2 e
bnc
i−1
is the coefficient of xjm 2 −1 in i=1
pCi (xm ).
Now, we prove Theorem 4.
Proof of Theorem 4. To prove the theorem, it suffices to show that for Construction 4, the fusion node is able to construct C
bn
2 c − 1 if n is odd.
from any 2mn/2 − 1 worker nodes if n is even or from any (m + 1)m
Q n2
First, for the case in which n is even, we need
to compute C = i=1 Ai Bi . Notice, from Claim 5, that the desired matrix
Q n2
n/2
i−1
product C is the coefficient of xm −1 in i=1
pCi (xm ). Thus, it is sufficient to compute
this coefficient at the fusion node
Q n2
i−1
as the computation output for successful computation. Now, because the polynomial i=1
pCi (xm ) has degree 2mn/2 − 2,
evaluation
of the polynomial at any 2mn/2 − 1 distinct points is sufficient to compute all of the coefficients of powers of x
Q n2
i−1
n/2
in i=1 pCi (xm ) using polynomial interpolation. This includes C, the coefficient of xm −1 .
Q n
b2c
Now, for the case in which n is odd, we need to compute C =
A
B
. First, notice that C is a concatenation
i i Ad n
i=1
2e
Q n
b2c
n
of the matrices
i=1 Ai Bi Ad 2 ej , j ∈ {1, · · · , m} as follows:
n
b2c
Y
Ai Bi Ad n2 e =
C=
i=1
bn
2c
Y
i=1
bn
2c
Ai Bi Ad n2 e1 · · ·
Y
Ai Bi Ad n2 em .
(31)
i=1
Q n
Qd n2 e
bnc
i−1
b2c
From Claim 6, for all j ∈ {1, · · · , m}, the product
A
B
pCi (xm ).
is the coefficient of xjm 2 −1 in i=1
i i Ad n
i=1
2 ej
Thus, it is sufficient to compute these coefficients, for all j ∈ {1, · · · , m}, at the fusion node as the computation output
Qd n2 e
i−1
n
for successful computation. Now, because the polynomial i=1
pCi (xm ) has degree mb 2 c (m + 1) − 2, evaluation of
bn
c
the polynomial at any m 2 (m + 1) − 1 distinct points is sufficient to compute all of the coefficients
of x in
Q n of powers
n
Qd n2 e
b2c
mi−1
jmb 2 c −1
n
) using polynomial interpolation. This includes the coefficients of x
, i.e.
i=1 pCi (x
i=1 Ai Bi Ad 2 ej , for
all j ∈ {1, · · · , m}.
C. Complexity analyses of Construction 4
n
Encoding/decoding complexity: Decoding requires interpolating a 2mn/2 −2 degree polynomial if n is even or a mb 2 c (m+
1)−2 degree polynomial if n is odd for N 2 elements. Using polynomial interpolation algorithms of complexity O(q log2 q) [30],
n
n
where q = k(n, m), defined in (18), complexity per matrix element is O(md 2 e log2 md 2 e ). Thus, for N 2 elements, the decoding
n
n
2
complexity is O(N 2 md 2 e log md 2 e ).
Encoding for each worker requires performing n additions, each adding m scaled matrices of size N 2 /m, for an overall
encoding complexity for each worker of O(mnN 2 /m) = O(nN 2 ). Thus, the overall computational complexity of encoding
for P workers is O(nN 2 P ).
Each worker’s computational cost: Each worker multiplies n matrices of dimensions N × N/m and N/m × N . For any
worker l with l ∈ {1, · · · , P }, the multiplication can be performed as follows:
Case 1: n is even
n/2−1
m
mn/2−1
In this case, worker l wishes to compute pA1 (xl−1 )pB1 (xl−1 )pA2 (xm
)pBn/2 (xm
).
l−1 )pB2 (xl−1 ) · · · pAn/2 (xl−1
l−1
Worker l does this multiplication in the following order:
i
mi−1
3
2
1. Compute pBi (xl−1
)pAi+1 (xm
l−1 ) for all i ∈ {1, · · · , n/2 − 1} with a total complexity of O(nN /m ).
3
3
2. Compute the product of the output matrices of the previous step with a total complexity of O(nN /m ). Call this product
matrix D. Notice that D has a dimension of N/m × N/m.
3. Compute pA1 (xl−1 )D with complexity O(N 3 /m2 ). Call this product matrix E. Notice that E has a dimension of
N × N/m.
mn/2−1
4. Compute E pBn/2 (xl−1
) with complexity O(N 3 /m).
Hence, the overall computational complexity per worker for even n is O(max(nN 3 /m2 , nN 3 /m3 , N 3 /m2 , N 3 /m)) =
O(max(nN 3 /m2 , N 3 /m)).
Case 2: n is odd
In this case, worker l wishes to compute
(n−3)/2
m
m
pA1 (xl−1 )pB1 (xl−1 )pA2 (xm
l−1 )pB2 (xl−1 ) · · · pA(n−1)/2 (xl−1
(n−3)/2
m
)pB(n−1)/2 (xl−1
(n−1)/2
m
)pA(n+1)/2 (xl−1
).
Worker l does this multiplication in the following order:
i
mi−1
3
2
1. Compute pBi (xl−1
)pAi+1 (xm
l−1 ) for all i ∈ {1, · · · , (n − 1)/2} with a total complexity of O(nN /m ).
3
3
2. Compute the product of the output matrices of the previous step with a total complexity of O(nN /m ). Call this product
matrix D. Notice that D has a dimension of N/m × N/m.
3. Compute pA1 (xl−1 )D with complexity O(N 3 /m2 ).
Hence, the overall computational complexity per worker for odd n is O(max(nN 3 /m2 , nN 3 /m3 , N 3 /m2 )) = O(nN 3 /m2 ).
In conclusion, the computational complexity per worker is O(max(nN 3 /m2 , N 3 /m)) if n is even, and O(nN 3 /m2 ) if n
is odd.
n
Communication cost: The master node communicates O(nP N 2 /m) symbols, and the fusion node receives O(mb 2 c N 2 )
symbols from the successful worker nodes.
VII. C ONCLUSION
We provide the MatDot construction for coded matrix multiplication, which has a recovery threshold of 2m − 1. We also
present the systematic MatDot construction providing the same recovery threshold of MatDot construction. Since a natural
lower bound for the recovery threshold is m, an open question is whether the threshold we obtain is optimal. Another open
question is whether the worker-fusion communication cost can be reduced without increasing the recovery threshold of MatDot.
In addition, we provide the PolyDot construction which allows a trade-off between communication cost and recovery threshold.
Finally, we provide code constructions for multiplying more than two matrices.
Open problem: Tensor product multiplication
We conclude with a discussion on the possibility of extending our ideas to computing tensor products. Consider the problem,
where the goal is to compute A ⊗ B of two N × N square matrices in the distributed system specified in system model section
over P workers with the highest possible recovery threshold. For this problem, an obvious application of Polynomial codes
[4] yields a recovery threshold of m2 .
To see this, let
A = A0 A2 . . . Am−1 ,
B = B0 B2 . . . Bm−1 .
Note that A ⊗ B = [A0 ⊗ B . . . Am−1 ⊗ B] = [A0 ⊗ B0 AP
], and thus it suffices to compute all
0 ⊗ B1 . . . Am−1 ⊗ B
Pm−1
m
m
terms of the form Ai ⊗Bj for computing A⊗B. Let pA (x) = i=1 Ai xi , pB (x) = i=1 Bi xi . Let x1 , x2 , . . . , xP be distinct
scalars. Let pA⊗B (x) be a polynomial of degree m2 − 1 where the coefficient of xi+ml is Ai ⊗ Bl , where 0 ≤ i, l ≤ m − 1.
2
If worker i computes pA (xi ) ⊗ pB (xm
i ) = pA⊗B (xi ), then the fusion node can interpolate pA⊗B (x) from any m worker
nodes. Furthermore, the coefficients of pA⊗B (x) are Ai ⊗ Bj , 0 ≤ i, j ≤ m − 1, Therefore, the fusion node can recover all
matrices in the set {Ai ⊗ Bj , i, j = 0, 1, . . . , m − 1}, and therefore A ⊗ B.
While the above approach obtains a recovery threshold of m2 , note workers do not necessarily carry a greater burden of
computation as compared with the fusion node. Specifically the computational complexity of the fusion node is Θ(N 3 ), which
has the same complexity order as the entire tensor product A ⊗ B, thus in an order sense, the strategy is not gaining in terms
of computational complexity from distributed computing. An open question motivated by our work is the above approach has
any interesting applications.
ACKNOWLEDGMENTS
We thank Mayank Bakshi and Yaoqing Yang for helpful discussions. We acknowledge support from NSF CNS-1702694,
CCF-1553248, CCF-1464336, and CCF-1350314. This work was also supported in part by Systems on Nanoscale Information
fabriCs (SONIC), one of the six SRC STARnet Centers, sponsored by MARCO and DARPA.
R EFERENCES
[1] K. Lee, M. Lam, R. Pedarsani, D. Papailiopoulos, and K. Ramchandran. Speeding up distributed machine learning using codes. In IEEE International
Symposium on Information Theory, pages 1143–1147, 2016.
[2] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr. Coded mapreduce. In Communication, Control, and Computing (Allerton), 2015 53rd Annual Allerton
Conference on, pages 964–971. IEEE, 2015.
[3] R. Tandon, Q. Lei, A. G. Dimakis, and N. Karampatziakis. Gradient coding. In Machine Learning Systems Workshop, Advances in Neural Information
Processing Systems (NIPS), 2016.
[4] Qian Yu, Mohammad Ali Maddah-Ali, and A Salman Avestimehr. Polynomial codes: an optimal design for high-dimensional coded matrix multiplication.
arXiv preprint:1705.10464, 2017.
[5] G. Joshi, Y. Liu, and E. Soljanin. On the delay-storage trade-off in content download from coded distributed storage systems. IEEE Journal on Selected
Areas in Communications, 32(5):989–997, 2014.
[6] S. Dutta, V. Cadambe, and P. Grover. Short-dot: Computing large linear transforms distributedly using coded short dot products. In Advances In Neural
Information Processing Systems, pages 2092–2100, 2016.
[7] N. Azian-Ruhi, S. Avestimehr, F. Lahouti, and B. Hassibi. Consensus-based distributed computing. In Information Theory and Applications Workshop,
2017.
[8] Yaoqing Yang, Pulkit Grover, and Soummya Kar. Fault-tolerant distributed logistic regression using unreliable components. In 54th Annual Allerton
Conference on Communication, Control, and Computing (Allerton), pages 940–947, 2016.
[9] Yaoqing Yang, Pulkit Grover, and Soummya Kar. Computing linear transformations with unreliable components. IEEE Transactions on Information
Theory, 63(6), 2017.
[10] Songze Li, Mohammad Ali Maddah-Ali, Qian Yu, and Amir Salman Avestimehr. A fundamental tradeoff between computation and communication in
distributed computing. CoRR, abs/1604.07086, 2016.
[11] Songze Li, Mohammad Ali Maddah-Ali, and Amir Salman Avestimehr. A unified coding framework for distributed computing with straggling servers.
CoRR, abs/1609.01690, 2016.
[12] Songze Li, Sucha Supittayapornpong, Mohammad Ali Maddah-Ali, and A Salman Avestimehr. Coded terasort. 6th International Workshop on Parallel
and Distributed Computing for Large Scale Machine Learning and Big Data Analytics, 2017.
[13] Songze Li, Mohammad Ali Maddah-Ali, and Amir Salman Avestimehr. Coded distributed computing: Straggling servers and multistage dataflows. 54th
Allerton Conference, Sept. 2016.
[14] Rashish Tandon, Qi Lei, Alexandros G. Dimakis, and Nikos Karampatziakis. Gradient coding: Avoiding stragglers in distributed learning. In Proceedings
of the 34th International Conference on Machine Learning (ICML), pages 3368–3376, 2017.
[15] Netanel Raviv, Itzhak Tamo, Rashish Tandon, and Alexandros G. Dimakis. Gradient coding from cyclic MDS codes and expander graphs. CoRR,
abs/1707.03858, 2017.
[16] Mehmet Fatih Aktas, Pei Peng, and Emina Soljanin. Effective straggler mitigation: Which clones should attack and when? arXiv preprint
arXiv:1710.00748, 2017.
[17] Mehmet Fatih Aktas, Pei Peng, and Emina Soljanin. Straggler mitigation by delayed relaunch of tasks. arXiv preprint arXiv:1710.00414, 2017.
[18] Malihe Aliasgari, Jörg Kliewer, and Osvaldo Simeone. Coded computation against straggling decoders for network function virtualization. arXiv preprint
arXiv:1709.01031, 2017.
[19] Amirhossein Reisizadehmobarakeh, Saurav Prakash, Ramtin Pedarsani, and Salman Avestimehr. Coded computation over heterogeneous clusters. arXiv
preprint arXiv:1701.05973, 2017.
[20] Wael Halbawi, Navid Azizan-Ruhi, Fariborz Salehi, and Babak Hassibi. Improving distributed gradient descent using reed-solomon codes. arXiv preprint
arXiv:1706.05436, 2017.
[21] Can Karakus, Yifan Sun, and Suhas Diggavi. Encoded distributed optimization. In Information Theory (ISIT), 2017 IEEE International Symposium on,
pages 2890–2894. IEEE, 2017.
[22] Yaoqing Yang, Pulkit Grover, and Soummya Kar. Coding method for parallel iterative linear solver. NIPS 2017, to appear, arXiv:1706.00163, 2017.
[23] Amirhossein Reisizadeh and Ramtin Pedarsani. Latency analysis of coded computation schemes over wireless networks. arXiv preprint arXiv:1707.00040,
2017.
[24] Sanghamitra Dutta, Viveck Cadambe, and Pulkit Grover. Coded convolution for parallel and distributed computing within a deadline. In IEEE International
Symposium on Information Theory (ISIT), July 2017.
[25] Viveck Cadambe and Pulkit Grover. Codes for distributed computing: A tutorial. IEEE Information Theory Society Newsletter, 67(4):3, December 2017.
[26] K.-H. Huang and J.A. Abraham. Algorithm-based fault tolerance for matrix operations. IEEE Transactions on Computers, C-33(6):518–528, June 1984.
[27] Kangwook Lee, Changho Suh, and Kannan Ramchandran. High-dimensional coded matrix multiplication. In Information Theory (ISIT), 2017 IEEE
International Symposium on, pages 2418–2422, 2017.
[28] Qian Yu, Mohammad Ali Maddah-Ali, and A Salman Avestimehr. Straggler mitigation in distributed matrix multiplication: Fundamental limits and
optimal coding. arXiv preprint:1801.07487, 2018.
[29] Sanghamitra Dutta, Ziqian Bai, Pulkit Grover, and Tze Meng Low. Coded training of model parallel deep neural networks under soft-errors. Submitted
to ISIT 2018, 2018.
[30] H T Kung. Fast evaluation and interpolation. Technical report, Carnegie Mellon University, 1973.
[31] Volker Strassen. Gaussian elimination is not optimal. Numerische Mathematik, 13(4):354–356, 1969.
[32] Lei Li. On the arithmetic operational complexity for solving vandermonde linear equations. Japan journal of industrial and applied mathematics,
17(1):15–18, 2000.
| 7cs.IT
|
A Faster Construction of Greedy Consensus Trees
Paweł Gawrychowski1 , Gad M. Landau1 , Wing-Kin Sung2 , and Oren Weimann1
1
arXiv:1705.10548v2 [cs.DS] 4 Jul 2017
2
University of Haifa, Israel
National University of Singapore, Singapore
Abstract
A consensus tree is a phylogenetic tree that captures the similarity between a set of conflicting
phylogenetic trees. The problem of computing a consensus tree is a major step in phylogenetic
tree reconstruction. It also finds applications in predicting a species tree from a set of gene trees.
This paper focuses on two of the most well-known and widely used consensus tree methods: the
greedy consensus tree and the frequency difference consensus tree. Given k conflicting trees each
with n leaves, the previous fastest algorithms for these problems were Opkn2 q for the greedy
consensus tree [J. ACM 2016] and Õpmintkn2 , k 2 nuq for the frequency difference consensus tree
[ACM TCBB 2016]. We improve these running times to Õpkn1.5 q and Õpknq respectively.
1
Introduction
A phylogenetic tree describes the evolutionary relationships among a set of n species called taxa. It
is an unordered rooted tree whose leaves represent the taxa and whose inner nodes represent their
common ancestors. Each leaf has a distinct label from rns. The inner nodes are unlabeled and have
at least two children.
Numerous phylogenetic trees, reconstructed from data sources like fossils or DNA sequences, have
been published in the literature since the early 1860s. However, the phylogenetic trees obtained from
different data sources or using different reconstruction methods result in conflicts (similar though not
identical phylogenetic trees over the same set rns of leaf labels). The conflicts between phylogenetic
trees are usually measured by their difference in signatures: The signature of a phylogenetic tree
T is the set tLpuq : u P T u where Lpuq denotes the set of labels of all leaves in the subtree rooted
at node u of T (the set Lpuq is sometimes called a cluster). To deal with the conflicts between k
phylogenetic trees in a systematic manner, the concept of a consensus tree was invented. Informally,
the consensus tree is a single phylogenetic tree that summarizes the branching structure (signatures)
of all the conflicting trees. Consensus trees have been widely used in two applications:
1. Constructing a phylogenetic tree: First, by sampling the dataset, we generate k different
datasets (for some constant k that can be as large as 10, 000). Then, we reconstruct one
phylogenetic tree for each dataset. Finally, we build the consensus tree of these k trees.
2. Constructing a species tree: First, a phylogenetic tree (called a gene tree) is reconstructed
for each individual gene. Then, the species tree is created by building the consensus tree of
all k gene trees.
Many different types of consensus trees have been proposed in the literature. For almost all
of them, optimal or near-optimal Õpknq time constructions are known. These include Adam’s
consensus tree [1], strict consensus tree [27], loose consensus tree [4, 13], majority-rule consensus
tree [13, 17], majority-rule (+) consensus tree [11], and asymmetric median consensus tree [20, 21]1 .
Two of the most notable exceptions are the frequency difference consensus tree [10] and the greedy
consensus tree [5, 9] whose running time remains quadratic in either k or n. In particular, the
former can be constructed in Õpmintkn2 , k 2 nuq time [11] and the later in Opkn2 q time [13]. For
more details about different consensus trees and their advantages and disadvantages see the survey
in [5], Chapter 30 in [8], and Chapter 8.4 in [31].
In this paper we propose novel algorithms for the frequency difference consensus tree problem
and the greedy consensus tree problem. First, we present an Opkn log2 nq time deterministic labeling
method. The labelling method counts the frequency (number of occurrences) of every cluster S in
the input trees. Based on this labeling method, we obtain an Opkn log2 nq time construction of
the frequency difference consensus tree. Then, for the greedy consensus tree, we present our main
technical contribution: a method that uses micro-macro decomposition to verify if a cluster S is
compatible with a tree T in Opn0.5 log nq time and, if so, modify T to include S in Opn0.5 log nq
amortized time. Using this procedure, we obtain an Opkn1.5 log nq time construction of the greedy
consensus tree.
1
Constructing the asymmetric median consensus tree was proven to be NP-hard for k ą 2 [20] and solvable in
Õpnq time for k “ 2 [21].
1
The frequency difference consensus tree. The frequency f pSq of a cluster S (a set of labels of
all leaves in some subtree) is the number of trees that contain S. A cluster is said to be compatible
with another cluster if they are either disjoint or one is included in the other. A frequent cluster
is a cluster that occurs in more trees than any of the clusters that are incompatible with it. The
frequency difference consensus tree is a tree whose signature is exactly all the frequent clusters.
The frequency difference consensus tree was initially proposed by Goloboff et al. [10], and its
relationship with other consensus trees was studied in [7]. In particular, it can be seen as a refinement
of the majority-rule consensus tree [13, 17]. Moreover, it is known to give less noisy branches than
the greedy consensus tree defined below. Steel and Velasco [30] concluded that “the frequency
difference method is worthy of more widespread usage and serious study”. A naive construction
of the frequency difference consensus tree takes Opk 2 n2 q time. The free software TNT [10] has
implemented a heuristics method to construct it more efficiently. However, its time complexity
remains unknown.
Recently, Jansson et al. [11] presented an Opmintkn2 , k 2 n ` kn log2 nuq time construction (implemented in the FACT software package [12]). Their algorithm first computes the frequency f pSq
of every cluster S with non-zero frequency. This is done in total Opmintkn2 , k 2 nuq time. They
then show that given these computed frequencies, the frequency difference consensus tree can be
computed in additional Opkn log2 nq time. In Section 2 we show how to compute all frequencies in
total Opkn log2 nq time leading to the following theorem:
Theorem 1. The frequency difference consensus tree of k phylogenetic trees T1 , T2 , . . . , Tk on the
same set of leaves rns can be computed in Opkn log2 nq time.
To prove the above theorem, we first develop an Opkn log2 nq time algorithm for assigning a
number idpuq P rkns to every u P Ti such that idpuq “ idpu1 q iff Lpuq “ Lpu1 q. With these numbers in
hand, we can then compute the frequencies of all clusters in Opknq time using counting sort (since
there are only kn clusters with non-zero frequencies, and each was assigned an integer bounded by
kn). Notice that this also generates a sorted list of all clusters with non-zero frequencies.
The greedy consensus tree. We say that a given collection C of subsets of rns is consistent if
there exists a phylogenetic tree T such that the signature of T is exactly C. The greedy consensus
tree is defined by the following procedure: We begin with an initially empty C and then consider
all clusters S in decreasing order of their frequencies. In this order, for every S, we check if C Y tSu
is consistent, and if so we add S to C.
The greedy consensus tree is one of the most well-known consensus trees. It has been used
in numerous papers such as [2, 3, 6, 14–16, 18, 19, 23, 24, 26, 29, 33] to name a few. For example,
in a recent landmark paper in Nature [23], it was used to construct the species tree from 1000
gene trees of yeast genomes, and in [6] it was asserted that “The greedy consensus tree offers some
robustness to gene-tree discordance that may cause other methods to fail to recover the species tree.
In addition, the greedy consensus method outperformed our other methods for branch lengths outside
the too-greedy zone.”.
The greedy consensus tree is an extension of the majority-rule consensus tree, and is sometimes
called the extended majority-rule consensus (eMRC) tree. It is implemented in popular phylogenetics software packages like PHYLIP [9], PAUP* [32], MrBayes [22], and RAxML [28]. A naive
construction of the greedy consensus tree requires Opkn3 q time [5]. To speed this up, these software packages use some forms of randomization methods. For example, PHYLIP uses hashing to
2
improve the running time. Even with randomization, the time complexities of these solutions are
not known. Recently, Jansson et al. [13] gave the best known provable construction with an Opkn2 q
deterministic running time (their implementation is also part of the FACT package). In Section 3
we present our main contribution, a deterministic Õpkn1.5 q construction as stated by the following
theorem:
Theorem 2. The greedy consensus tree of k phylogenetic trees T1 , T2 , . . . , Tk on the same set of
leaves rns can be computed in Opkn1.5 log nq time.
To prove the above theorem, we develop a generic procedure that takes any ordered list of
clusters S1 , S2 , . . . , S` Ď rns and tries adding them one-by-one to the current solution C. We assume
that every cluster Si is specified by providing a tree Ti and a node ui P Ti such that Si “ Lpui q. Our
procedure requires Opn0.5 log nq time per cluster (to add this cluster to C or assert that it cannot
be added) and needs not to assume anything about the order of the clusters. In particular, it does
not rely on the clusters being sorted by frequencies.
2
Computing the Identifiers
We process the nodes of every Ti in the bottom-up order. For every node u P Ti , we compute the
identifier idpuq by updating the following structure called the dynamic set equality structure:
Lemma 1. There exists a dynamic set equality structure that supports: (1) create a new empty
structure in constant time, (2) add x P rns to the current set in Oplog2 nq time, (3) return the
identifier of the current set in constant time, and (4) list all ` elements of the current set in Op`q
time. The structure ensures that the identifiers are bounded by the total number of update operations
performed so far, and that two sets are equal iff their identifiers are equal.
Proof. To allow for listing all elements of the current set S, we store them in a list. Before adding
the new element x to the list, we need to check if x P S. This will be done using the representation
described below.
Conceptually, we work with a complete binary tree B on n leaves labelled with 0, 1, . . . , n ´ 1
when read from left to right (without losing generality, n “ 2k ), where every node u corresponds
to a set Dpuq Ď rns defined by the leaves in its subtree (note that Dpuq “ ti, i ` 1, . . . , ju, where
0 ď i ď j ă n). Now, any set S is associated with a binary tree B, where we write 1 in a leaf
if the corresponding element belongs to S and 0 otherwise. Then, for every node we define its
characteristic vector by writing down the values written in the leaves of its subtree in the natural
order (from left to right). Clearly, the vector of an inner node is obtained by concatenating the
vector of its children. We want to maintain identifiers of all nodes, so that the identifiers of two
nodes are equal iff their characteristic vectors are identical. If we can keep the identifiers small,
then the identifier of the current set can be computed as the identifiers of the root of B.
Assume that we have already computed the identifiers of all nodes in B and now want to add x
to S. This changes the value in the leaf u corresponding to x and, consequently, the characteristic
vectors of all ancestors of u. However, it does not change the characteristic vectors of any other
node. Therefore, we traverse the ancestors of u starting from u and recompute their identifiers.
Let v be the current node. If we have never seen the characteristic vector of v before, we can set
the identifier of v to be the largest already used identifier plus one. Otherwise, we have to set
the identifier of v to be the same as the one previously used for a node with such a characteristic
3
vector. As mentioned above, the characteristic vector of an inner node v is the concatenation of the
characteristic vectors of its children v` and vr . We maintain a dictionary mapping a pair consisting
of the identifier of v` and the identifier of vr to the identifier of v. The dictionary is global, that
is, shared by all instances of the structure. Then, assuming that we have already computed the
up-to-date identifiers of v` and vr , we only need to query the dictionary to check if the identifier of
v should be set to the largest already used identifier plus one (which is exactly when the dictionary
does not contain the corresponding pair) or retrieve the appropriate identifier. Therefore, adding
x to B reduces to log n queries to the dictionary. By implementing the dictionary with balanced
search trees, we therefore obtain the claimed Oplog2 nq time for adding an element.
We are not completely done yet, because creating a new complete binary tree B takes Opnq time
and therefore the initialization time is not constant yet. However, we can observe that it does not
make sense to explicitly maintain a node u of B such that S X Dpuq “ H, because we can assume
that the identifier of such an u is 0. In other words, we can maintain only the part of B induced
by the leaves corresponding to S. Adding an element x P S is implemented as above, except that
we might need to create (at most Oplog nq) new nodes on the leaf-to-root path corresponding to x
(if such a leaf already exists, we terminate the procedure as x P S already) and then recompute the
identifiers on the whole path as described above.
Armed with Lemma 1, we process every Ti bottom-up. Consider an inner node v P Ti and
let v1 , v2 , . . . , vd be its children ordered so that |Lpv1 q| “ maxj |Lpvj q|, that is, the subtree rooted
at v1 is the largest. Assuming that we have already stored every Lpvj q in a dynamic set equality
structure, we construct a dynamic set equality structure storing Lpvq by simply inserting all elements
of Lpv2 q Y Lpv3 q Y ¨ ¨ ¨ Y Lpvd q into the structure of Lpv1 q. This takes Oplog2 nq time per element.
Then, we set idpuq to be the identifier of the obtained structure. By a standard argument (heavy
path decomposition), every leaf of Ti is inserted into at most log n structures and therefore the
whole Ti is processed in Opn log3 nq time. This gives us the claimed Opkn log3 nq total time.
We now proceed with a faster Opkn log2 nq total time solution. While this is irrelevant for our
Opkn1.5 log nq time construction of the greedy consensus tree, it implies a better complexity for
constructing the frequency difference consensus tree.
We start with a high-level intuition. Lemma 1 is, in a sense, more than we need, as it is not
completely clear that we need to immediately compute the identifier of the current set. Indeed,
applying heavy path decomposition we can partially delay computing the identifiers by proceeding
in Oplog nq phases. In each phase, we can then replace the dynamic dictionary used to store the
mapping with a radix sort. Intuitively, this shaves one log from the time complexity. We proceed
with a detailed explanation.
Theorem 3. The numbers idpuq can be found for all nodes of the k phylogenetic trees T1 , T2 , . . . , Tk
in Opkn log2 nq total time.
Proof. For a node v P Ti , define its level levelpvq to be `, such that 2` ď |Lpvq| ă 2``1 . Thus, the
levels are between 0 and log n, level of a node is at least as large as the levels of its children, and a
node on level ` has at most one child on the same level. We work in phases ` “ 0, 1, . . . , log n. In
phase `, we assume that the numbers idpvq are already known for all nodes v, such that levelpvq ă `,
and want to assign these numbers to all nodes v, such that levelpvq “ `. We will show how to achieve
this in Opkn log nq time, thus proving the theorem.
Consider all nodes v, such that levelpvq “ `. Because every such v has at most one child at
the same level, all level-` nodes in Ti can be partitioned into maximal paths of the form p “
4
v1 ´ v2 ´ . . . ´ vs , where the level of the parent of v1 is larger than ` (or v1 is the root of Ti ),
and the levels of all children of vj (except for vj`1 , if defined) are smaller than `. v1 is called the
head of p and denoted headppq. Now, our goal is to find idpvj q with the required properties for
every j “ 1, 2, . . . , s. We will actually achieve a bit more. The sets Lpheadppqq are disjoint in every
tree Ti , and thus we can define, for every i, a partition Pi “ tPi p1q, Pi p2q, . . . , Pi pti qu of the set of
leaves rns, where every Pi pzq corresponds to a level-` path p “ v1 ´ v2 ´ . . . ´ vs in Ti , such that
Lpheadppqq “ Pi pzq. The elements of Pi pzq are then ordered, and we think that Pi pzq is a sequence
of length |Pi pzq|. The ordering is chosen so that, for every j “ 1, 2, . . . , s, the set Lpvj q corresponds
to some prefix of Pi pzq. Pi pzqr1..rs denotes the prefix of Pi pzq of length r. We will assign identifiers
to all such prefixes Pi pzqr1..rs, for every i “ 1, 2, . . . , k, z “ 1, 2, . . . , ti and r “ 1, 2, . . . , |Pi pzq|, with
the property that the identifiers of two prefixes are equal iff the sets of leaves appearing in both
of them are equal. Then, we can extract the required idpvj q in constant time each by taking the
identifiers of some Pi pzqr1..rs.
Recall that in the slower solution we worked with a complete binary tree B on n leaves. For
every set S in the collection and every u P B, we computed an identifier of the set S X Dpuq.
This was possible, because if u` and ur are the left and the right child of u, respectively, then the
identifier of S X Dpuq can be found using the identifiers of S X Dpu` q and S X Dpur q. We need to
show that retrieving these identifiers can be batched.
Fix a node u P B and, for every i “ 1, 2, . . . , k and z “ 1, 2, . . . , ti , consider all prefixes Pi pzqr1..rs
for r “ 1, 2, . . . , |Pi pjq|. We create a version of u for every such prefix. The version corresponds to the
set containing all elements of Dpuq occurring in the prefix Pi pzqr1..rs. We want to assign identifiers
to all versions of u. First, observe that we only have to create a new version if Pi pzqrrs P Dpuq,
as otherwise the set is the same as for r ´ 1. Thus, the total number of required versions, when
summed over all nodes u P B on the same depth in B, is only kn, as a leaf of Ti creates exactly new
version for some u. For every node u P B, we will store a list of all its versions. A version consists
of its identifier (such that the identifier of two versions is the same iff the corresponding sets are
equal) together with the indices i, z and r. We describe how to create such a list for every node
u P B at the same depth d given the lists for all nodes at depth d ` 1 next.
Let u1 and u2 be the left and the right child of u P B, respectively. Then, we need to create
a new version of u for every new version of u1 and every new version of u2 , because for the set
corresponding to u to change either the set corresponding to u1 or the set corresponding to u2 must
change, and every change is adding one new element. Fix i and z and consider all versions of u1
corresponding to i and z sorted according to r. Let the sorted list of their r’s be a1 ă a2 ă . . ..
Similarly, consider all versions of ur corresponding to i and z sorted according to r, and let the
sorted list of their r’s be b1 ă b2 ă . . .. For every x P ta1 , a2 , . . .u Y tb1 , b2 , . . .u, we create a new
version of u corresponding to i, z, and r equal to x. This is done by retrieving the version of u1 with
r equal to ap , such that ap ď x and p is maximized, and the version of u2 with r equal to bq , such
that bq ď x and q is maximized. Then, the identifier of the new version of u can be constructed
from the pair consisting of the identifiers of these versions of u1 and u2 (this is essentially the same
reasoning as in the slower solution). We could now use a dictionary to map these pairs to identifiers.
However, we can also observe that, in fact, we have reduced finding the identifiers of all versions of
all nodes u P B at the same depth d to identifying duplicates on a list of kn pairs of numbers from
rkns. This can be done by radix sorting all pairs in linear time (more precisely, Opknq time and
Opknq space), and then sweeping through the sorted list while assigning the identifiers. This takes
only Opknq time for every depth d, so Opkn log nq for every level as claimed.
5
The proof of Theorem 1 follows immediately from Theorem 3.
3
Simulating the Greedy Algorithm
We consider k trees T1 , . . . , Tk on the same set of leaves rns, and assume that every node u has
an identifier idpuq such that idpuq “ idpu1 q iff Lpuq “ Lpu1 q. We next develop a general method for
maintaining a solution C (i.e., a set of compatible identifiers) so that, given any node u P Ti , we are
able to efficiently check if Lpuq is compatible with C, meaning that C Y Lpuq is consistent, and if so
add Lpuq to C. Our method does not rely on the order in which the sets arrive and in particular
can be used to run the greedy algorithm.
We represent C with a phylogenetic tree Tc such that C “ tLpuq : u P Tc u. Tc is called the current
consensus tree. By Lemma 2.2 of [13], S is compatible with C iff there exists a node v P T such that
for every child v 1 of v either Lpv 1 q X S “ H or Lpv 1 q Ď S. Also, adding Lpuq to C can be done by
creating a new child w of v and reconnecting every original child v 1 of v such that Lpv 1 q Ď S to the
new w. This is illustrated in Figure 1.
v
v
w
a
b
c
d
e
f
g
h
c
i
a
b
g
h
d
e
f
i
Figure 1: Adding ta, b, g, h, iu to S.
Initially, Tc consists only of n leaves attached to the common root (which corresponds to C “
ttxu : x P rnsu). Our goal is to to maintain some additional information so that given any node
u P Ti , we can check if Lpuq is compatible with C in Opn0.5 log nq time. After adding Lpuq to C
the information will be updated in amortized Opkn0.5 log nq time. To explain the intuition, we first
show how to check if Lpuq is compatible with C in roughly Op|Lpuq|q time.
Let Lpuq “ t`1 , `2 , . . . , `s u and let ui be the leaf of Tc labelled with `i . Then, v must be an
ancestor of every ui . We claim that, in fact, v should be chosen as the lowest common ancestor of
u1 , u2 , . . . , us , because if all ui ’s are in the same subtree rooted at a child v 1 of v then we can as
well replace v with v 1 . So, we can find v by asking s ´ 1 lca queries: we start with u1 and then
iteratively jump to the lca of the current node and ui . Assuming that we represent Tc in such a
way that an lca query can be answered efficiently, this takes roughly Opsq time. Then, we need to
decide if for every child v 1 of v it holds that Lpv 1 q Ď Lpuq or Lpv 1 q X Lpuq “ H. This can be done by
computing, for every such v 1 , how many ui ’s belong to the subtree rooted at v 1 , and then checking
if this number is either 0 or |Lpv 1 q|. To compute these numbers, we maintain a counter for every
v 1 . Then, for every ui we retrieve the child v 1 of v such that ui belongs to the subtree rooted at
v 1 and increase the counter of v 1 . Assuming that we represent Tc so that such v 1 can be retrieved
efficiently, this again takes roughly Opsq time. Finally, we iterate over all ui again, retrieve the
6
v
3
m n
0
5
o
c
a
b
g
e
f
h
i
d
j
`
k
Figure 2: Checking if S “ tm, n, o, b, g, h, i, ku is compatible with Tc . Leaves corresponding to the
elements of S are shown in red and their lca is v. S is not compatible with Tc because the counter
of the middle child of v is equal to 5 yet there are 7 leaves in its subtree.
corresponding v 1 and check if its counter is equal to |Lpv 1 q| (so our representation of Tc should also
allow retrieving the number of leaves in a subtree). If not, then Lpuq is not compatible with C, see
Figure 2. Otherwise, we create the new node w and reconnect to w all children v 1 of v, such that
the counter of v 1 is equal to |Lpv 1 q|.
We would like to avoid explicitly iterating over all elements of Lpuq. This will be done by
maintaining some additional information, so that we only have to iterate over up to n0.5 elements.
To explain what is the additional information we need the (standard) notion of a micro-macro
decomposition. Let b be a parameter and consider a binary tree on n nodes. We want to partition
it into Opn{bq node-disjoint subtrees called micro trees. Each micro tree is of size at most b and
contains at most two boundary nodes that are adjacent to nodes in other micro trees. One of these
boundary nodes, called the top boundary node, is the root of the whole micro tree, and the other
is called the bottom boundary node. Such a partition is always possible and can be found in Opnq
time.
We binarize every Ti to obtain Ti1 . Then, we find a micro-macro decomposition of Ti1 with
b “ n0.5 . By properties of the decomposition we have the following:
Proposition 1. For any u P Ti such that |Lpuq| ą n0.5 , there exists a boundary node v P Ti1 such
that Lpuq can be obtained by adding at most n0.5 elements to Lpvq. Furthermore, v and these up to
n0.5 elements can be retrieved in Opn0.5 q time after Opnq preprocessing.
The total number of boundary nodes is only Opkn0.5 q. For each such boundary node u, we
maintain a pointer to a node fingerpuq P Tc called the finger of u. fingerpuq is a node v P Tc such
that Lpuq Ď Lpvq but, for every child vi of v, Lpuq Ę Lpvi q.
Proposition 2. The node fingerpuq is the lowest common ancestor in Tc of all leaves with labels
belonging to Lpuq.
Additionally, the children of fingerpuq are partitioned into three groups: (1) vi such that Lpvi q Ď
Lpvq, (2) vi such that Lpvi q X Lpvq “ H, and (3) the rest. We call them full, empty, and mixed,
respectively (with respect to u). For each group we maintain a list storing all nodes in the group,
7
u
e
d
g
a
b
c
f
v
j
k `
h i
m no
p
q
Figure 3: A schematic illustration of the micro-macro decomposition. v is a boundary node and
Lpvq “ th, i, j, k, `, m, n, o, p, qu. Then, Lpuq “ ta, b, c, f, g, h, i, j, k, `, m, n, o, p, qu so Lpuq “ Lpvq Y
ta, b, c, f, gu.
every node knows its group, and the group knows it size. Additionally, every group knows the total
number of leaves in all subtrees rooted at its nodes.
We also need to augment the representation Tc to allow for efficient extended lca queries. The
lowest common ancestor (lca) of u and v is the leafmost node w that is an ancestor of both u and
v. An extended lca query, denoted lca_extpu, vq, returns the first edge on the path from the lca of
u and v to u, and -1 if u is an ancestor of v. For example, in Figure 2, lca_extpv, kq “ ´1 whereas
lca_extpn, kq is the edge between v and its leftmost child.
Lemma 2. We can maintain a collection of rooted trees under: (1) create a new tree consisting of
a single node, (2) make the root of one tree a child of a node in another tree, (3) delete an edge
from a node to its parent, (4) count leaves in the tree containing a given node, and (5) extended lca
queries, all in Oplog nq amortized time, where n is the total size of all trees in the collection.
Proof. We apply the link/cut trees of Sleator and Tarjan [25] to maintain the collection. This
immediately gives us the first three operations. To implement computing the size and lca_extpu, vq
queries we need to explain the internals of link/cut trees. Each tree is partitioned into node-disjoint
paths consisting of preferred edges. Each node has at most one such edge leading to its preferred
child. For each maximal path consisting of preferred edges, called a preferred path, we store its
nodes in a splay tree, where the left-to-right order on the nodes of the splay tree corresponds to
8
the top-bottom order on the nodes in the rooted tree. Each such splay tree stores a pointer to
the topmost node of its preferred path. Additionally, each node of the tree stores a pointer to its
current parent. All operations on a link/cut tree use the access procedure. Its goal is to change the
preferred edges so that there is a preferred path starting at the root and ending at v. This is done
by first shortening the preferred path containing v so that it ends at v. Then, we iteratively jump
to the topmost node u of the current preferred path and make u the preferred child of its parent.
Whenever the preferred child of a node changes, we need to update the splay tree representing the
nodes of the preferred path. Even though the number of jumps might be Ωpnq, it can be shown
that all these updates take Oplog nq amortized time.
Now we can explain how to implement lca_extpu, vq. First, we access node v. This gives us a
preferred path starting at the root and ending at v. Second, we access node u while keeping track of
the topmost nodes of the visited preferred paths. If u is on the same preferred path as v, then u is an
ancestor of v. Otherwise, let p be the preferred path visited just before reaching the preferred path
starting at the root of the whole tree. Then the topmost node of p (before changing the preferred
child of its parent) should be returned as lca_extpu, vq. Thus, the complexity of lca_extpu, vq is the
same as the complexity of access.
To compute the size of a tree, we augment the splay trees. Every node of a preferred path stores
the total number of leaves in all subtrees attached to it through non-preferred edges (plus one if
the node itself is a leaf). Additionally, every node of a splay tree stores the sum of the numbers
stored in its subtree, or in other words the total number of leaves in all subtrees attached to its
corresponding contiguous fragment of the preferred path through non-preferred edges. The sums
stored at the nodes of the splay tree are easily maintained during rotations. We also need to update
the total number of leaves after making a preferred edge non-preferred or vice versa. This is easily
done by accessing the sum stored at the root of the splay tree. To access the number the leaves in
the tree containing v, we need to access v. This makes all of v’s children non-preferred and makes
v the root of its splay tree. Hence, the number stored at v is the total number of leaves in the tree
containing v.
We next show how to efficiently check for any u if Lpuq is compatible with C. By the following
lemma, this can be done in Opn0.5 log nq time, assuming we have stored the required additional
information. Recall that this additional information includes:
1. The tree Tc maintained using Lemma 2.
2. For every boundary node w, we store fingerpwq.
3. For every boundary node w, we store three lists containing the full, the mixed, and the empty
children of w respectively. Each list also stores the total number of leaves in all subtrees rooted
at its nodes.
Lemma 3. Assuming access to the above additional information, given any node u P Ti we can
check if Lpuq is compatible with C in Opn0.5 log nq time.
Proof. By Lemma 2.2 of [13], to check if Lpuq is compatible with C we need to check if there exists
a node v such that for every child v 1 of v either Lpv 1 q X Lpuq “ H or Lpv 1 q Ď Lpuq. First, observe
that v can be chosen as the lowest common ancestor of all leaves with labels belonging to Lpuq. By
properties of the micro-macro decomposition, we can retrieve a boundary node w and a set S of up
to n0.5 labels such that Lpuq “ Lpwq Y S (if |Lpuq| ă n0.5 , there is no w). Then, the lowest common
9
ancestor of all leaves with labels belonging to Lpuq is the lowest common ancestor of fingerpwq and
all leaves with labels belonging to S. Therefore, v can be found with |S| lca queries in Opn0.5 log nq
time. Second, to check if Lpvi q X Lpuq “ H or Lpvi q Ď Lpuq for every child vi of v we distinguish two
cases:
If v is a proper ancestor of fingerpwq we can calculate |Lpvi q X Lpuq| for every vi in Op|S| log nq “
Opn0.5 log nq time as follows. Every edge has its associated counter. We assume that all counters
are set to zero before starting the procedure and will make sure that they are cleared at the end.
First, we use an lca_extpw, vq query to access the edge leading to the subtree containing w and set
its counter to |Lpwq|. Then, we iterate over all ` P S, retrieve the leaf u of Tc labelled with `, and
use an lca_extpu, vq query to access the edge leading to the subtree of v containing u and increase
its counter by one. Additionally, whenever we access an edge for the first time (in this particular
query), we add it to a temporary list Q. After having processed all ` P S, we iterate over pv, vi q P Q
and check if the counter of pv, vi q is equal to the number of leaves in the subtree rooted at vi (which
requires retrieving the number of leaves). If this condition holds for every pv, vi q P Q then Lpuq is
compatible with C and furthermore, the nodes vi such that pv, vi q P Q are exactly the ones that
should be reconnected. Finally, we iterate over the edges in Q again and reset their counters.
If v “ fingerpwq the situation is a bit more complicated because we might not have enough time
to explicitly iterate over all children of v that should be reconnected. Nevertheless, we can use a
very similar method. Every edge has its associated counter (again, we assume that the counter are
set to zero before starting the procedure and will make sure that they are cleared at the end). We
also need a global counter g, that is set to the total number of leaves in all subtrees rooted at either
full or mixed children of v decreased by |Lpwq|. g can be initialized in constant time in the first step
of the procedure due to the additional information stored with every list of children. Intuitively, g is
how many leaves not belonging to Lpwq we still have to see to conclude that indeed Lpvi qXLpuq “ H
or Lpvi q Ď Lpuq for every child vi of v. We iterate over ` P S and access the edge pv, vi q leading to
the subtree containing u labelled with `. We decrease g by one and, if vi is an empty child of v
and this is the first time we have seen vi (in this query) then we add the number of leaves in the
subtree rooted at vi to g. If, after having processed all ` P S, g “ 0 then we conclude that Lpuq is
compatible with C. The whole process takes Op|S| log nq “ Opn0.5 log nq time.
Before explaining the details of how to update the additional information, we present the intuition. Recall that adding Lpuq to C is done by creating a new child v 1 of v and reconnecting some
children of v to v 1 . Let the set of all children of v be C and the set of children that should be
reconnected be Cr . Note that if |Cr | “ 1 or |C| “ |Cr | then we do not have to change anything in
Tc . Otherwise, updating Tc can be implemented using two different methods:
1. Delete edges from nodes in Cr to v. Create a new tree consisting of a single node v 1 and make
it a child of v. Then, make all nodes in Cr children of v 1 .
2. Delete edges from nodes in CzCr to v. Delete the edge from v to its parent w. Create a new
tree consisting of a single node v 1 and make it a child of w. Then, make v a child of v 1 and
also make all nodes in CzCr children of w. See Figure 4.
Thus, by using Cr or CzCr , the number of operations can be either Op|Cr |q or Op|C| ´ |Cr |q. We
claim that by choosing the cheaper option we can guarantee that the total time for modifying the
link-cut tree representation of Tc is Opn log2 nq. Intuitively, every edge of the final consensus tree
participates in Oplog nq operations, and there are at most n such edges. This is formalized in the
following lemma.
10
Lemma 4. mint|Cr |, |C| ´ |Cr |u summed over all updates of Tc is n log n.
Proof. We assume that 2 ď |Cr | ă |C| in every update, as otherwise there is nothing to change in
Tc . Then, there are at most n updates, as each of them creates a new inner node and there are
never any nodes with degree 1 in Tc .
We bound the sum of mint|Cr |, |C| ´ |Cr |u by assigning credits to inner nodes of Tc . During
the execution of the algorithm, a node u with b siblings should have log b credits. Thus, whenever
we create a new inner node we need at most log n new credits, thus the total number of allocated
credits is n log n. It remains to argue that, whenever we create a new child v 1 of v and reconnect
some of its children, the original credits of v can be used to pay for the update and make sure that
all children of v and v 1 have enough credits after the update.
Denoting x “ |Cr | and y “ |C| ´ |Cr |, the cost of the update is mintx, yu. The total number of
credits of all children of v before the update is px ` yq logpx ` y ´ 1q. After the update, the number
of credits of all children of v is py ` 1q log y ď y log y ` log n and the number of credits of all children
of v 1 is x logpx ` 1q. Ignoring the log n new credits allocated to v 1 , the number of available credits
is thus:
px ` yq logpx ` y ´ 1q ´ y log y ´ x logpx ´ 1q “ x logp1 ` y{px ´ 1qq ` y logp1 ` px ´ 1q{yq
which is at least mintx, yu for x ě 2, so enough to pay mint|Cr |, |C| ´ |Cr |u for the update. Hence,
the sum is at most n log n.
w
v
v1 v2 v3 v4 v5 v6 v7 v8 v9
w
v
w
w
v0
v0
v
v
v1 v2 v3 v4 v5 v6 v7 v8 v9
v1 v2 v3 v4 v5 v6 v7 v8 v9
v1 v2 v4 v8 v3 v5 v6 v7 v9
Figure 4: Reconnecting children v3 , v5 , v6 , v7 , v9 of v using the second method.
Before presenting the whole update procedure, we need one more technical lemma.
Lemma 5. The procedure for checking if Lpuq is compatible with C can be requested to return Cr in
Op|Cr | ` n0.5 q time or CzCr in Op|C| ´ |Cr | ` n0.5 q time.
Proof. By inspecting the proof of Lemma 3, we see that there are two cases depending on whether
v is a proper ancestor of fingerpwq or not.
1. If v is a proper ancestor of fingerpwq then Cr can be obtained from Q. More precisely,
for every pv, vi q P Q we add vi to Cr in Op|Cr |q total time. We can also obtain CzCr in
Op|C|q “ Op|CzCr | ` |S|q “ Op|C| ´ |Cr | ` n0.5 q time.
2. If v “ fingerpwq then, while iterating over ` P S, if this is the first time we have seen vi then
we add vi to Cr . Additionally, we add all full children of v to Cr . Thus, Cr can be generated
in Op|Cr |q time. Similarly, CzCr consists of all empty children of v without the nodes vi seen
when iterating over ` P S, and so can be generated in Op|CzCr | ` |S|q “ Op|C| ´ |Cr | ` n0.5 q
time.
11
Thus, we can always generate Cr in Op|Cr | ` n0.5 q time and CzCr in Op|C| ´ |Cr | ` n0.5 q time.
To add Lpuq to C, we will need to iterate over either Cr or CzCr (depending on which is smaller).
After paying additional Opn0.5 q time we can assume that we have access to a list of the elements in
the appropriate set. The additional time sums up to Opn1.5 q, because there can be only n distinct
new sets added to C.
Lemma 6. If Lpuq is compatible with C then, after adding Lpuq to C and modifying Tc we can update
all additional information in amortized Opkn0.5 log nq time assuming that we add n such sets.
Proof. Recall that Tc is maintained using the data structure from Lemma 2, and adding Lpuq to
C is implemented by creating a new child v 1 of v and reconnecting some of the children of v to v 1 .
C is the set of all children of v and Cr is the set of children of v that are reconnected to v 1 . If
|Cr | ď |C| ´ |Cr | we iterate over Cr and reconnect them one-by-one. If |Cr | ą |C| ´ |Cr | we iterate
over CzCr and reconnect them to a new node w that is inserted between v and its parent. To
iterate over either Cr or CzCr , we extend the query procedure as explained in Lemma 5. This adds
Opn0.5 to the time complexity, but then we can assume that the requested set can be generated in
time proportional to its size. To unify the case of |Cr | ď |C| ´ |Cr | and |Cr | ą |C| ´ |Cr |, we think
that v is replaced with two nodes v 1 and v 2 , where v 1 is the parent of v 2 . All nodes in Cr become
children of v 2 while all nodes of CzCr become children of v 1 after iterating over either Cr or CzCr ,
depending on which set is smaller, so by Lemma 4 in the whole process we iterate over sets of total
size n log n, so only amortized log n assuming that we add n sets Lpuq.
Consider a boundary node u. If fingerpuq ‰ v then there is no need to update the additional
information concerning u. If fingerpuq “ v then we need to decide if the finger of u should be set
to v 1 or v 2 and update the partition of the children of fingerpuq accordingly. fingerpuq should be set
to v 1 exactly when, for any w P CzCr , Lpwq X Lpuq “ H or, in other words, all nodes in CzCr are
empty with respect to u. The groups should be updated as follows:
1. If fingerpuq is set to v 2 then we should remove all nodes in CzCr from the list of empty nodes
with respect to u (as they are no longer children of fingerpuq). Other groups remain unchanged.
2. If fingerpuq is set to v 1 then we should remove all nodes in Cr from the lists. Additionally,
we need to insert v 2 into the appropriate group: full if all nodes in Cr were full, empty if all
nodes in Cr were empty, and mixed otherwise.
We need to show that all these conditions can be checked by either iterating over the nodes of C or
over the nodes of CzCr , because we want to iterate over the smaller of these. This then guarantees
that the amortized cost of updating the additional information for a boundary node is only Oplog nq,
so amortized Opkn0.5 log nq overall.
To check if all nodes in CzCr are empty with respect to u, we can either iterate over the nodes
in CzCr or iterate over all nodes in Cr and check if all nodes in C that are full or empty in fact
belong to Cr (this is possible because we also keep the total number of full and empty nodes in C).
Thus, we can check if fingerpuq should be set to v 1 .
If fingerpuq is set to v 1 we need to decide where to put v 2 . We only explain how to decide if all
nodes in Cr are full, as the procedure for empty is symmetric. We can either iterate over all nodes
in Cr and check that they are full or iterate over all nodes in CzCr and check that all nodes in C
that are empty or mixed in fact belong to CzCr (and thus do not belong to Cr , so all nodes in Cr
12
are full). Finally, we add the number of leaves in the subtree rooted at v 2 (extracted in Oplog nq
time) to the appropriate sum.
It remains to describe how to remove all unnecessary nodes from the lists. Here we do not worry
about having to iterate over the smaller set, because there are only Opnq new edges created during
the whole execution of the algorithm, so we can afford to explicitly iterate over the nodes that
should be removed, that is, over C or CzCr . For every removed node, we also subtract the number
of leaves in its subtree (extracted in Oplog nq time) from the appropriate sum. Overall, this adds
Opn log nq per boundary node to the time complexity, so only amortized Opkn0.5 log nq overall.
References
[1] E. N. Adams III. Consensus techniques and the comparison of taxonomic trees. Systematic
Zoology, 21(4):390–397, 1972.
[2] M. Bayzid, S. Mirarab, B. Boussau, and T. Warnow. Weighted statistical binning: enabling
statistically consistent genome-scale phylogenetic analyses. PLOS One, page e0129183, 2015.
[3] M. S. Bayzid and T. J. Warnow. Naive binning improves phylogenomic analyses. Bioinformatics, 29(18):2277–2284, 2013.
[4] K. Bremer. Combinable component consensus. Cladistics, 6(4):369–372, 1990.
[5] D. Bryant. A classification of consensus methods for phylogenetics. In Bioconsensuss, DIMACS
Series in Discrete Mathematics and Theoretical Computer Science, volume 61, pages 163–184.
2003.
[6] J. H. Degnan, M. DeGiorgio, D. Bryant, and N. A. Rosenberg. Properties of consensus methods
for inferring species trees from gene trees. Systematic Biology, 58(1):35–54, 2009.
[7] J. Dong, D. Fernández-Baca, F. R. McMorris, and R. C. Powers. Majority-rule (+) consensus
trees. Mathematical Biosciences, 228(1):10–15, 2010.
[8] J. Felsenstein. Inferring Phylogenies. Sinauer Associates, Inc., Sunderland, Massachusetts,
2004.
[9] J. Felsenstein. PHYLIP, version 3.6. Software package, Department of Genome Sciences,
University of Washington, Seattle, U.S.A., 2005.
[10] P. A. Goloboff, J. S. Farris, and K. C. Nixon. TNT, a free program for phylogenetic analysis.
Cladistics, 24(5):774–786, 2008.
[11] J. Jansson, R. Rajaby, C. Shen, and W.-K. Sung. Algorithms for the majority rule (+) consensus tree and the frequency difference consensus tree. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2016.
[12] J. Jansson, C. Shen, and W.-K. Sung. Fast Algorithms for Consensus Trees (FACT). http:
//compbio.ddns.comp.nus.edu.sg/~consensus.tree, 2013.
[13] J. Jansson, C. Shen, and W.-K. Sung. Improved algorithms for constructing consensus trees.
Journal of the ACM, 63(3):1–24, 2016.
13
[14] E. D. Jarvis et al. Whole-genome analyses resolve early branches in the tree of life of modern
birds. Science, 346(6215):1320–1331, 2014.
[15] L. Liu, L. Yu, and S. Edwards. A maximum pseudo-likelihood approach for estimating species
trees under the coalescent model. BMC Evolutionary Biology, 10:302, 2010.
[16] L. Liu, L. Yu, L. Kubatko, D. K. Pearl, and S. V. Edwards. Coalescent methods for estimating
phylogenetic trees. Molecular Phylogenetics and Evolution, 53(1):320 – 328, 2009.
[17] T. Margush and F. R. McMorris.
43(2):239–244, 1981.
Consensus n-trees.
Bulletin of Mathematical Biology,
[18] S. Mirarab, S. Bayzid, and T. Warnow. Evaluating summary methods for multilocus species
tree estimation in the presence of incomplete lineage sorting. Systematic Biology, 65(3):366–380,
2016.
[19] J. B. Pease, D. C. Haak, M. W. Hahn, and L. C. Moyle. Phylogenomics reveals three sources
of adaptive variation during a rapid radiation. PLOS Biology, 14(2):1–24, 2016.
[20] C. Phillips and T. J. Warnow. The asymmetric median tree — a new model for building
consensus trees. Discrete Applied Mathematics, 71(1-3):311–335, 1996.
[21] R. Rajaby and W.-K. Sung. Computing asymmetric median tree of two trees via better bipartite
matching algorithm. In IWOCA, 2017.
[22] F. Ronquist and J. P. Huelsenbeck. MrBayes 3: Bayesian phylogenetic inference under mixed
models. Bioinformatics, 19(12):1572–1574, 2003.
[23] L. Salichos and A. Rokas. Inferring ancient divergences requires genes with strong phylogenetic
signals. Nature, 497:327–331, 2013.
[24] L. Salichos, A. Stamatakis, and A. Rokas. Novel information theory-based measures for quantifying incongruence among phylogenetic trees. Molecular Biology and Evolution, 31(5):1261–
1271, 2014.
[25] D. D. Sleator and R. E. Tarjan. A data structure for dynamic trees. J. Comput. Syst. Sci.,
26(3):362–391, 1983.
[26] J. V. Smith, E. L. Braun, and R. T. Kimball. Ratite nonmonophyly: independent evidence
from 40 novel loci. Systematic Biology, 62(1):35–49, 2013.
[27] R. R. Sokal and F. J. Rohlf. Taxonomic congruence in the Leptopodomorpha re-examined.
Systematic Zoology, 30(3):309–325, 1981.
[28] A. Stamatakis. Raxml version 8: a tool for phylogenetic analysis and post-analysis of large
phylogenies. Bioinformatics, 30(9):1312–1313, 2014.
[29] A. Stamatakis, P. Hoover, J. Rougemont, and S. Renner. A rapid bootstrap algorithm for the
raxml web servers. Systematic Biology, 57(5):758, 2008.
[30] M. Steel and J. D. Velasco. Axiomatic opportunities and obstacles for inferring a species tree
from gene trees. Systematic Biology, 63(5):772–778, 2014.
14
[31] W.-K. Sung. Algorithms in Bioinformatics: A Practical Introduction. Chapman & Hall/CRC,
Boca Raton, Florida, 2010.
[32] D. L. Swofford. PAUP*, version 4.0. Software package, Sinauer Associates, Inc., Sunderland,
Massachusetts, 2003.
[33] J. Yang and T. J. Warnow. Fast and accurate methods for phylogenomic analyses. BMC
Bioinformatics, 12(S-9):S4, 2011.
15
| 8cs.DS
|
arXiv:1503.03685v3 [math.RA] 27 May 2016
MONOMIAL RIGHT IDEALS AND THE HILBERT SERIES OF
NONCOMMUTATIVE MODULES
ROBERTO LA SCALA∗
Per Armando, in memoriam.
Abstract. In this paper we present a procedure for computing the rational
sum of the Hilbert series of a finitely generated monomial right module N
over the free associative algebra Khx1 , . . . , xn i. We show that such procedure
terminates, that is, the rational sum exists, when all the cyclic submodules
decomposing N are annihilated by monomial right ideals whose monomials
define regular formal languages. The method is based on the iterative application of the colon right ideal operation to monomial ideals which are given by an
eventual infinite basis. By using automata theory, we prove that the number
of these iterations is a minimal one. In fact, we have experimented efficient
computations with an implementation of the procedure in Maple which is the
first general one for noncommutative Hilbert series.
1. Introduction
It is difficult to list the plenty of objects that are represented as a finite sequence
of symbols, that is, by a word or a string. Let us think for example to our own dna,
a message in a code, a word in a natural or formal language, the base expansion
of a number, a path in a graph or a transition of states in a machine, etc. From
the perspective of the algebraist these elements are just monomials of the free
associative algebra F = Khx1 , . . . , xn i, that is, the algebra of polynomials in the
noncommutative variables xi . In fact, any finitely generated (associative) algebra
is isomorphic to a quotient F/I where I ⊂ F is the two-sided ideal of the relations
satisfied by the generators of the algebra. Another important generalization is the
notion of finitely generated right F -module which corresponds to a quotient F r /M
of the free right F -module F r = F ⊕ · · · ⊕ F with respect to some submodule
M ⊂ F r . It is probably useless to mention the importance of the noncommutative
structures, for instance, in physics but let us just refer to the recent book [15].
Fundamental data about algebras and modules consist in their dimension over
the base field K or in their “growth” in case they are infinite dimensional. By
definition, the growth or the affine Hilbert function of N = F r /M is the map
r
d 7→ dim F≤d
/M≤d where F≤d ⊂ F is the subspace of the polynomials of degree at
r
most d and M≤d = M ∩ F≤d
. This sequence is naturally encoded by its generating
function which is called the affine Hilbert series of N .
Unlike the commutative case, owing to non-Noetherianity of the free associative
algebra one has that this series has generally not a rational sum. Nevertheless,
finitely generated free right modules have clearly a rational Hilbert series and the
2010 Mathematics Subject Classification. Primary 16Z05. Secondary 16P90, 05A15.
Key words and phrases. Hilbert series, noncommutative modules, automata.
Partially supported by Università di Bari.
1
2
R. LA SCALA
same happens in fact for all finitely presented modules. This is because F is a free
ideal ring [2] which implies that any right submodule M ⊂ F r is in fact free. It
is then natural to ask if there are other cases when the Hilbert series is rational.
One possible approach consists in reducing the problem to the monomial cyclic
case that is to right modules C = F/I where I ⊂ F is a monomial right ideal.
This can be done by using a suitable monomial ordering of F r and by proving a
noncommutative analogue of a well-known theorem of Macaulay. We present these
results in Section 2.
A short exact sequence for a monomial cyclic right module C = F/I is presented
in Section 3 allowing the possibility to reduce the computation of the Hilbert series
of C to the one of n cyclic modules defined by the colon right ideals of I with
respect to each variable xi . By iterating this exact sequence one obtains in principle
a rational sum for the Hilbert series of C but the main problem is to provide that
one has only a finite number of iterations. Under such assumption, in Section 4
we present all details of the corresponding algorithm which is based on the explicit
description of a monomial basis for the colon right ideals of a monomial ideal with
respect to a monomial of F .
In Section 5 we explain that the set of the colon right ideals of I which are involved in the computation of the Hilbert series of the monomial cyclic right module
C = F/I are in fact the states of a minimal deterministic automaton recognizing
the formal language given by the monomials of I. By the Kleene’s theorem this
implies that the proposed procedure has termination if and only if such language
is a regular one, that is, all the monomials of I can be obtained from finite subsets by applying a finite number of elementary operations. For the two-sided case,
one finds in [9, 13] the idea that the rationality of the Hilbert series of an algebra
A = F/I is provided by the existence of a deterministic finite automaton recognizing the normal monomials modulo I or, equivalently, the monomials of I. In this
paper we present a complete and effective direct method for costructing a minimal
such automaton. The roots of our approach can be traced back to the formal languages notion of Nerode equivalence that we have enhanced for right and two-sided
monomial ideals by means of the tools of algebra.
The proposed ideas results in a feasible algorithm which has been illustrated in
Section 6 for an infinitely related monomial algebra that we obtain from an Artin
group. In Section 7, by means of an experimental implementation that we have
developed with Maple, we present some data and timings for the computation of
the Hilbert series of a couple of Hecke algebras. Finally, some conclusions and ideas
for further research directions are proposed in Section 8.
2. Hilbert functions and series
Let K be any field and let X = {x1 , . . . , xn } be any finite set. We denote by
F = KhXi the free associative algebra which is freely generated by X. In other
words, the elements of F are noncommutative polynomials in the variables xi . We
define then W = Mon(F ) the set of all monomials of F , that is, the elements of W
are words over the alphabet X. We consider F as a graded algebra by means of
the standard grading which defines deg(w) as the length of a word w ∈ W . For any
integer r > 0, denote by F r the (finitely generated) free right F -module of rank
r
r. If {ei } is the canonical basis
are the
P of F then the elements of such module
right F -linear combinations i ei fi (fi ∈ F ). We endow the module F r with the
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
3
standard grading, that is, we put deg(ei w) = deg(w), for any 1 ≤ i ≤ r and for each
w ∈ W . This implies
that F r is also a P
filtered module by means of the canonical
S
r
r
r
where F≤d
= 0≤k≤d Fkr , for any d ≥ 0. We recall now
filtration F = d≥0 F≤d
the notion of Hilbert function and series for finitely generated right F -modules.
Definition 2.1. Let M be a right submodule of F r and
module N = F r /M . For all d ≥ 0, we define HFa (N )(d)
r
M≤d = M ∩ F≤d
. Moreover, we put
HFa (N )(0)
HF(N )(d) =
HFa (N )(d) − HFa (N )(d − 1)
define the quotient right
r
= dimK F≤d
/M≤d where
if d = 0;
otherwise.
We call the sequences HFa (N ) = {HFa (N )(d)}d≥0 and HF(N ) = {HF(N )(d)}d≥0
respectively the affine Hilbert function and the (generalized) Hilbert function of the
finitely generatedP
right F -module N . Observe that if M (hence N ) is a graded module, that is, M = d≥0 Md with Md = M ∩Fdr , then HF(N )(d) = dimK Fdr /Md . We
denote by HSa (N ), HS(N ) the generating series corresponding to HFa (N ), HF(N ),
namely
X
X
HSa (N ) =
HFa (N )(d)td , HS(N ) =
HF(N )(d)td .
d≥0
d≥0
We call such series respectively the affine Hilbert series and the (generalized) Hilbert
series of N . From the definition it follows immediately that HS(N ) = (1−t)HSa (N ).
In literature one may find different names for the above functions and series,
especially in the case when M ⊂ F is a two-sided ideal and hence N = F/M is
an algebra. For instance, it is quite common to call HFa (N ) the growth function
of N . Sometimes one has also the name of Henri Poincaré together with the one
of David Hilbert to denominate all these functions and series. It is well-known
that any Hilbert function can be obtained as the one which is defined by a suitable
monomial module. To this purpose we introduce the following notions.
Definition 2.2. Let F r be a free right module endowed with the canonical basis
{ei }. We define W (r) = {ei }W = {ei w | 1 ≤ i ≤ r, w ∈ W } which is a canonical linear basis of F r . For this reason, the elements of W (r) may be called the
monomials of F r . Let now ≺ be a well-ordering of W (r). We call ≺ a monomial
ordering of F r if ei w ≺ ej w′ implies that ei wv ≺ ej w′ v, for all ei w, ej w′ ∈ W (r)
and v ∈ W . In particular, we say that ≺ is a graded ordering if deg(w) < deg(w′ )
implies that ei w ≺ ej w′ , for any ei w, ej w′ ∈ W (r).
From now on, we assume that F r is endowed with a graded monomial ordering.
For instance, for all ei w, ej w′ ∈ W (r) one may define that ei w ≺ ej w′ if and only
if w < w′ in some graded lexicographic ordering or w = w′ and i < j.
P
Definition 2.3. Let f = k eik wk ck ∈ F r with eik wk ∈ W (r) and ck ∈ K, ck 6= 0.
If eil wl = max≺ {eik wk } then we put lm(f ) = eil wl and lc(f ) = cl . Moreover,
if M ⊂ F r is a right submodule then one defines the monomial right submodule
LM(M ) = hlm(f ) | f ∈ M, f 6= 0i ⊂ F r . We call LM(M ) the leading monomial
module of M .
The following is a generalization of a well-know theorem of Macaulay for commutative modules (see, for instance, [11]).
4
R. LA SCALA
Theorem 2.4. Let M ⊂ F r be a right submodule and consider the right modules
N = F r /M and N ′ = F r /LM(M ). Then, one has that HFa (N ) = HFa (N ′ ) and
hence HF(N ) = HF(N ′ ).
r
Proof. Put M ′ = LM(M ) and denote W (r)≤d = W (r) ∩ F≤d
, for any d ≥ 0. Since
′
r
r
′
M ⊂ F is a monomial submodule, one has clearly that a linear basis of F≤d
/M≤d
′
′
is given by the set {ei w + M≤d | ei w ∈ W (r)≤d \ M }. Consider now any element
f ∈ F r , f 6= 0. If lm(f ) ∈ M ′ then there exists g ∈ M such that lm(f ) = lm(g)v,
)
for some v ∈ W . By putting f1 = f − gv lc(f
lc(g) we obtain that f ≡ f1 mod M with
f1 = 0 or lm(f ) ≻ lm(f1 ). In the last case, we can repeat this division step for the
element f1 ∈ F r , f1 6= 0. Since ≺ is a well-ordering, we conclude that f ≡ f2 mod
M , for some f2 ∈ F r such that f2 = 0 or lm(f ) ≻ lm(f2 ) ∈
/ M ′ . If f2 6= 0 then we
consider the element f3 = f2 −lm(f2 )lc(f2 ) and hence f −lm(f2 )lc(f2 ) ≡ f3 mod M .
Note that one has f3 = 0 or lm(f2 ) ≻ lm(f3 ). P
By iterating the division process we
finally obtain that f ≡ f ′ mod M where f ′ = k eik wk ck with eik wk ∈ W (r) \ M ′
and ck ∈ K. Moreover, one has clearly that f ′ ∈ M if and only if f ′ = 0. Recall
now that ≺ is a graded ordering and note that lm(f ) ≻ eik wk when ck 6= 0. We
r
r
conclude that if f ∈ F≤d
then also f ′ ∈ F≤d
.
For the computation of the Hilbert series of a monomial module we are reduced
to the cyclic case. Precisely, it holds immediately the following result.
Proposition 2.5. Let M ⊂ F r be a monomial right submodule, L
that is, M is
generated by some set {eik wk } ⊂ W (r). Then, one has that M = i ei Ii where
Ii ⊂ F is the monomial right ideal which is generated byP{wk | eik = ei } ⊂ W . In
particular, if N = F r /M and Ci = F/Ii then HS(N ) = i HS(Ci ).
Let {wk } ⊂ W be a right basis of a monomial right ideal I ⊂ F . Observe that
{wk } is a minimal basis, that is, for all j 6= k there is no v ∈ W such that wk = wj v
if and only if {wk } is a free right basis. In other words, the ideal I is a free right
submodule of F according to the general property that the algebra F is a free ideal
ring [2, 5].
Proposition 2.6. Let {wk } ⊂ W be a finite free basis of a (finitely generated)
monomial right ideal I ⊂ F and define C = F/I the corresponding monomial cyclic
right module. Then, one has that
P
1 − k td k
.
HS(C) =
1 − nt
Proof. Put dk = deg(wk ) and denote by F [−dk ] the algebra F endowed with the
grading induced by the weight dk , that is, we define F [−dk ]d = 0 if d < dk and
F [−dk ]dL
= Fd−dk otherwise. In a similar way, one defines the graded free right F module k F [−dk ]. If {ek } is the canonical basis of such module then by definition
deg(ek ) = δk , for all k. We consider now the graded right F -module homomorphism
M
X
X
F [−dk ] → F,
ek fk 7→
wk fk
k
k
k
which implies the following short exact sequence
M
0→
F [−dk ] → F → C → 0.
k
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
5
Since the Hilbert series of F [−dk ] is clearly tdk /(1 − nt) we obtain that
P
X td k
1 − k td k
1
−
=
.
HS(C) =
1 − nt
1 − nt
1 − nt
k
Note that by Proposition 2.5 one obtains also the rationality of the Hilbert series
of a finitely presented (monomial) right module. Unfortunately, the free associative
algebra F = KhXi is not a right Noetherian ring. Consider, for instance, the right
ideal I = hx1 , x2 x1 , x22 x1 , . . .i. Then, Proposition 2.6 does not generally apply and
we have to find a different approach to the computation of the Hilbert series of
infinitely presented monomial cyclic right modules.
3. A key short exact sequence
Let I ⊂ F be any monomial right ideal and consider the corresponding monomial
cyclic right module C = F/I. If F [−1] is the algebra F endowed with the grading
induced by the weight 1 then we define I[−1] as the right ideal I endowed with
the grading of F [−1] and we put C[−1] = F [−1]/I[−1]. Moreover, we denote
by F [−1]n the graded free right F -module which is the direct sum of n copies of
F [−1]. Denote by {ei } the canonical basis of F [−1]n and therefore deg(ei ) = 1, for
all 1 ≤ i ≤ n. We consider the graded right F -module homomorphism
X
X
xi fi .
ei fi 7→
ϕ : F [−1]n → C,
i
i
The image Im ϕ is clearly the graded right submodule B = hx1 , . . . , xn i ⊂ C. One
has immediately that the cokernel C/B is either 0 when C = 0 or it is isomorphic
to the base field K otherwise.
We need to describe the kernel Ker ϕ which is a graded right submodule of
F [−1]n . For this purpose we make use of the following notion (see, for instance,
[10]). For each element f ∈ F , one defines the colon right ideal
(I :R f ) = {g ∈ F | f g ∈ I}.
This set is in fact a right ideal of F which generally does not contain I. Nevertheless,
one has that I ⊂ (I :R f ) when I is a two-sided ideal. It is clear that (I :R f ) = h1i
if and only if f ∈ I. Because I is a monomial ideal, we have that (I :R w) is
also a monomial ideal, for any w ∈ W . In Proposition 4.9 and Proposition 4.10
we will provide methods to compute a monomial basis of (I :R w) starting from a
monomial basis of I.
Since we are in the monomialL
case, one has immediately that ker ϕ is isomorphic
to the graded right submodule i (I :R xi ) ⊂ F [−1]n . To simplify notations, we
put Ixi = (I :R xi ) and we denote by Cxi = F/Ixi the corresponding monomial
cyclic right module. One obtains the following short exact sequence of graded right
F -module homomorphisms
M
(1)
0→
Cxi [−1] → C → C/B → 0.
1≤i≤n
It is useful to introduce the following number
1 if I = h1i,
γ(I) =
0 otherwise.
6
R. LA SCALA
In other words, we have that dimK (C/B) = 1 − γ(I). By the exact sequence (1)
one obtains the following key formula for the corresponding Hilbert series
X
(2)
HS(C) = 1 − γ(I) + t
HS(Cxi ).
1≤i≤n
The above formula suggests that one may compute a rational form for the sum
of the Hilbert series HS(C) by successively transforming a monomial right ideal I
into its colon right ideals (I :R xi ), for all variables xi ∈ X. We analize this method
in the next section under the assumption that there are only a finite number of
transforms to be performed. When such condition is satisfied will be characterized
in Section 5.
4. Computing the Hilbert series
It is useful to think the colon right ideal operation as a mapping on the set of
(monomial) right ideals of F .
Definition 4.1. Denote by mi the set of all monomial right ideals of F . For all
1 ≤ i ≤ n, we define the function Txi : mi → mi such that Txi (I) = (I :R xi ), for
any monomial right ideal I ∈ mi.
Definition 4.2. Let I ∈ mi. Define OI ⊂ mi the minimal subset containing I
such that Txi (OI ) ⊂ OI , for all 1 ≤ i ≤ n. We call OI the orbit of I. One has
clearly that OJ ⊂ OI , for any J ∈ OI .
Definition 4.3. We call I ∈ mi a regular (monomial) right ideal if its orbit OI is
a finite set and we denote by ri the set of all such ideals. Since Txi (I) ∈ OI observe
that Txi maps the set ri into itself, for each 1 ≤ i ≤ n.
Definition 4.4. Let I ∈ ri with OI = {I1 , . . . , Ir }. We define AI = (akl ) the
integer r-by-r matrix such that
akl = #{1 ≤ i ≤ n | Txi (Ik ) = Il }.
Moreover, we denote by pI (t) ∈ Z[t] the characteristic polynomial of AI and we
consider the 0-1 column r-vector
CI = (1 − γ(I1 ), . . . , 1 − γ(Ir ))T .
We call AI , pI (t) and CI respectively the adjacency matrix, characteristic polynomial and constant vector of (the orbit of) I.
In the above definition one has clearly to consider OI as an ordered set. Note
also that in the vector CI there is a single entry equal to 0 corresponding to the
trivial ideal F = h1i. In fact, such ideal has to belong to the orbit OI since for any
monomial w = xi1 · · · xid ∈ I one has that
(Txid · · · Txi1 )(I) = (((I :R xi1 ) · · · ) :R xid ) = (I :R w) = h1i.
Theorem 4.5. Let I ∈ ri and define C = F/I the corresponding monomial cyclic
right module. Then, the Hilbert series HS(C) is a rational function with integer
coefficients.
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
7
Proof. Let OI = {I1 , . . . , Ir } and put Ck = F/Ik (1 ≤ k ≤ r). Moreover, we put
Hk = HS(Ck ) and ck = 1 − γ(Ik ). By applying the formula (2) for each cyclic
module Ck , one obtains r linear equations in the r unknowns Hk , namely
X
Hk − t
Hlki = ck (1 ≤ k ≤ r)
1≤i≤n
where the index 1 ≤ lki ≤ r is defined such that Ilki = Txi (Ik ). Note that these
equations are with coefficients in the rational function field Q(t). By definition of
the adjacency matrix AI and the constant vector CI , one has therefore that the
column vector H = (H1 , . . . , Hr )T is a solution of the matrix equation
(I − tAI )H = CI .
(3)
Observe now that det(I − tAI ) = tr pI (1/t) 6= 0 since pI (t) = det(tI − AI ) is
the characteristic polynomial of the adjacency matrix AI . We conclude that the
equation (3) has a unique solution H = (I − tAI )−1 CI .
An immediate consequence of the above result is the following one.
Corollary 4.6. Let M ⊂ F r be a monomial right submodule and denote by N =
F r /M the corresponding
finitely generated monomial right module. Moreover, deL
note by M = i ei Ii the decomposition of M into monomial right ideals according
to Proposition 2.5. If all the ideals Ii are regular ones then the Hilbert series HS(N )
is a rational function with integer coefficients.
The numerator and the denominator of the Hilbert series of a regular right ideal
can be explicitly obtained in the following way.
Definition 4.7. Let I ∈ ri and assume OI = {I1 , . . . , Ir } with I1 = I. Denote by
BI the column r-vector which coincides with the first column of the cofactor matrix
of the r-by-r matrix I − tAI . Note that the entries of BI are polynomials with
integer coefficients. We call BI the polynomial vector of (the orbit) of I.
The matrix equation (3) implies immediately the following result.
Corollary 4.8. Let I ∈ ri and consider C = F/I. The rational function HS(C) =
f (t)/g(t) (f (t), g(t) ∈ Z[t]) is such that
(4)
f (t) = BTI CI , g(t) = tr pI (1/t)
where r = #OI = deg(pI (t)).
Observe in the above result that the entry of BI corresponding to the zero entry
of CI (h1i ∈ OI ) does not clearly contribute to the formation of the numerator
f (t). We present now a simple algorithm to compute the data defining the matrix
equation (3) corresponding to a regular right ideal.
8
R. LA SCALA
Algorithm 4.1 OrbitData
Input: I ∈ ri.
Output: (OI , AI , CI ).
N := {I};
O := {I};
A = (akl ) := 0 (square matrix);
C = (ck )T := 0 (column vector);
while N 6= ∅ do
choose J ∈ N ;
N := N \ {J};
k := the position of J in O;
if J 6= h1i then
ck := 1;
end if ;
for all 1 ≤ i ≤ n do
J ′ := Txi (J);
if J ′ ∈
/ O then
N := N ∪ {J ′ };
O := O ∪ {J ′ };
end if ;
l := the position of J ′ in O;
akl := akl + 1;
end for;
end while;
return (O, A, C).
Note explicitly that A and C are in fact a square matrix and a vector of increasing
dimension. When the algorithm stops, such dimension is exactly the cardinality of
O. Then, the Hilbert series corresponding to the regular right ideal I is obtained by
using the formula (4) or equivalently by solving the matrix equation (I −tA)H = C.
Observe that the matrix (I − tA) results sparse if the number of ideals in the orbit
O is large with respect to the number of variables xi .
We describe now how the mapping Txi is actually defined in terms of a monomial
basis of I.
Proposition 4.9. Let {wj } ⊂ W
and consider a monomial w ∈ W .
1
vj
(5)
wj′ =
0
be a right basis of a monomial right ideal I ⊂ F
For all j, we define the element
if w = wj vj (vj ∈ W ),
if wj = wvj ,
otherwise.
Then, a right basis of (I :R w) is given by the set {wj′ } ⊂ W . In particular, if
w = xi , that is, (I :R w) = Txi (I) then the right basis {wj′ } is defined according to
the following rule
1 if wj = 1 or wj = xi ,
vj if wj = xi vj (vj ∈ W ),
(6)
wj′ =
0 otherwise.
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
9
Proof. If a monomial t ∈ W belongs to (I :R w) then by definition wt ∈ I and
therefore wt = wj t′ , for some t′ ∈ W and some index j. One has two cases. If w =
wj vj (hence vj t = t′ ) for some vj ∈ W , then w ∈ I and therefore (I :R w) = h1i.
Otherwise, there is vj ∈ W such that wj = wvj and t = vj t′ .
An important case is when I is a monomial two-sided ideal, that is, C = F/I is
a monomial algebra. Recall that a two-sided ideal is always contained in the colon
right ideals it defines.
Proposition 4.10. Let {wj } ⊂ W be a two-sided basis of a monomial two-sided
ideal I ⊂ F and consider w ∈ W . Assume that w ∈
/ I, that is, (I :R w) 6= h1i. For
all j, we define the (finitely generated) monomial right ideal
Iw (wj ) = hvjk | ujk wj = wvjk , ujk , vjk ∈ W, deg(vjk ) < deg(wj )i.
P
Then, one has that (I :R w) = j Iw (wj ) + I.
Proof. If {wj } is a monomial two-sided basis of I then we have that the set
W {wj } = {uwj | u ∈ W, j any} is a monomial right basis of I. By Proposition 4.9
one has therefore that a right basis of (I :R w) is given by the monomials vjk ∈ W
such that ujk wj = wvjk , for some ujk ∈ W . Moreover, if deg(vjk ) ≥ deg(wj ) then
we have clearly that vjk ∈ I.
From the above results it is clear that if a monomial ideal I ⊂ F is finitely
generated in the right or two-sided case then the mappings Txi can be effectively
performed and the orbit OI is finite. Observe explicitly that if I is a finitely
generated two-sided ideal then I is generally infinitely generated as a right ideal
and Proposition 2.6 does not apply. In this case, we can apply instead OrbitData
which is an effective procedure that can be easily implemented in any computer
algebra system. We will provide in fact some implementation and application of this
algorithm in Section 7. Note that commutative ideals are always finitely generated
and hence OrbitData can be used also for computing commutative Hilbert series.
We remark finally that the rationality of the Hilbert series of a finitely presented
monomial algebra was first proved by Govorov [6] and further investigated by Anick
[1] and Ufnarovski [12, 13, 14].
Owing to non-Noetherianity of the free associative algebra, the monomial ideal
I may be infinitely generated also in the two-sided case and hence one can object
about the possibility to compute the transforms Txi . As a matter of fact, we will
show in the next section that the case when I is regular, that is, the orbit OI is
finite is also the case when the monomials of I can be represented by finitely many
“regular expressions” which make possible the computation of the Txi . We will
apply this approach to a concrete (infinitely generated) regular two-sided ideal in
Section 6.
5. Automata of monomial right ideals
To the aim of understanding which monomial right ideals are regular, it is convenient to introduce here the concept of deterministic automaton which is a fundamental one in information theory and engineering. Recall that we have denoted
by W the monoid of all monomials of F = KhXi, that is, the elements of W are
words over the alphabet X = {x1 , . . . , xn }. Another usual notation for this monoid
is X ∗ . Moreover, we denote by W op the opposite monoid of W .
10
R. LA SCALA
Definition 5.1. Let Q be any set. Fix an element q0 ∈ Q and a subset A ⊂ Q.
By denoting F (Q) the monoid of all functions Q → Q, we consider also a monoid
homomorphism T : W op → F (Q). In other words, the mapping T defines a monoid
right action of W on the set Q. By definition, a deterministic automaton is any
5-tuple (Q, X, T, q0 , A). This is called a deterministic finite automaton, briefly a
DFA, when Q is a finite set. The elements of Q are usually called the states of the
automaton where q0 is the initial state. A state q ∈ A is called an accepting state.
The homomorphism T is called the transition function of the automaton. For this
map we make use of the notation w = xi1 · · · xid 7→ Tw = Txid · · · Txi1 .
Strictly related to the notion of automaton are the following concepts.
Definition 5.2. Any subset L ⊂ W is called a (formal) language. A deterministic
automaton (Q, X, T, q0 , A) recognizes the language L if L = {w ∈ W | Tw (q0 ) ∈ A}.
A language is called regular if it is recognized by a FDA. A FDA which recognizes
a regular language L is called minimal if the number of its states is minimal with
respect of any other FDA recognizing L.
Clearly, if the automaton (Q, X, T, q0 , A) recognizes L then (Q, X, T, q0 , Ac ) recognizes Lc where Ac , Lc are the complements of A, L in Q, W , respectively. It
follows immediately that L is a regular language if and only if Lc is such. For
any language L ⊂ W , observe that the monoid structure of W canonically defines
an infinite automaton recognizing L. Precisely, such automaton is (W, X, T, 1, L)
where Tv (w) = wv, for all v, w ∈ W .
Definition 5.3. Let L ⊂ W and define the set w−1 L = {v ∈ W | wv ∈ L}, for
all w ∈ W . The Nerode equivalence defined by L is by definition the following
equivalence relation on W
w ∼ w′ (w, w′ ∈ W ) if and only if w−1 L = w′−1 L.
Observe that the quotient set W/ ∼ is in one-to-one correspondence with the set
{w−1 L | w ∈ W }. Moreover, we have that the relation ∼ is right invariant, that is,
w ∼ w′ implies that wv ∼ w′ v, for all v, w, w′ ∈ W . Then, an induced automaton
(W/ ∼, X, T̄ , [1], [L]) recognizing L is defined by putting T̄v ([w]) = [wv], for any
v, w ∈ W and [L] = {[w] | w ∈ L}. This is called the Nerode automaton recognizing
L.
Fundamental results in automata theory are the following ones (see, for instance,
[3]).
Theorem 5.4 (Myhill-Nerode). A language L ⊂ W is regular if and only if the
quotient set W/ ∼ is finite. In this case, the Nerode automaton is a minimal FDA
recognizing L.
Given two languages L, L′ ⊂ W , one can define their set-theoretic union L ∪ L′
and their product LL′ = {ww′ | w ∈ L, w′ ∈ L′ }. For any d ≥ 0, we haveSalso the
power Ld = {w1 · · · wd | wi ∈ L} (L0 = {1}) and the star operation L∗ = d≥0 Ld .
In other words, the language L∗ is the (free) submonoid of W which is generated by
L. The union, the product and the star operation are called the rational operations
over the languages.
Theorem 5.5 (Kleene). A language L ⊂ W is regular if and only if it can be
obtained from finite languages by applying a finite number of rational operations.
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
11
Example 5.6. The language L = {(x1 xi2 )j | i, j ≥ 0} ∪ {(x2 x1 )k | k ≥ 0} is regular
since it is obtained from the finite languages L1 = {x1 }, L2 = {x2 } by rational
operations, namely L = (L1 L∗2 )∗ ∪ (L2 L1 )∗ .
We show now how automata theory applies to our approach to compute noncommutative Hilbert series. Recall that mi denotes the set of all monomial right
ideals of F and for each variable xi ∈ X we have map Txi : mi → mi such that
Txi (I) = (I :R xi ), for any I ∈ mi. We have therefore a monoid homomorphism
T : W op → F (mi) such that w = xi1 . . . xid 7→ Tw = Txid · · · Txi1 . For each
monomial right ideal I ∈ mi, we have clearly that
Tw (I) = (I :R w) = {f ∈ F | wf ∈ I}.
Fix now I ∈ mi. One has immediately that OI = {Tw (I) | w ∈ W } and hence
we can restrict the right action of W that T defines on the set mi to an action
on the orbit OI . By abuse of notation, we will denote the corresponding monoid
homomorphism as T : W op → F (OI ). By considering the singleton {h1i} ⊂ OI , one
has then a deterministic automaton (OI , X, T, I, {h1i}). The language L which is
recognized by this automaton is by definition the set of monomials w ∈ W such that
Tw (I) = (I :R w) = h1i, that is, w ∈ I. In words, we have that L = I ∩ W . Observe
that the complementary automaton (OI , X, T, I, OI \ {h1i}) recognizes exactly the
language L = W \ I of the normal monomials modulo I.
Recall now that the states of the Nerode automaton recognizing L are in one-toone correspondence with the sets w−1 L = {v ∈ W | wv ∈ L}, for all w ∈ W . Since
L = I ∩ W , it is clear now that w−1 L = (I :R w) ∩ W and hence one obtains the
following result.
Proposition 5.7. The monomial right ideal I ⊂ F is regular if and only if the corresponding language L = I ∩ W is regular. In this case, the algorithm OrbitData
computes a FDA (OI , X, T, I, {h1i}) recognizing L which is a minimal one.
By the above proposition it follows that, in the regular case, the algorithm
OrbitData is an optimal one for computing the rational Hilbert series of a monomial cyclic right module C = F/I by means of the exact sequence (1). Moreover,
we have the Kleene’s theorem to verify if the monomial right ideal I is a regular
one.
6. An illustrative example
For illustrating how the algorithm OrbitData works in practice with infinitely
generated but regular monomial ideals, we consider an algebra which is associated
to the presentation of an Artin group defined by the following Coxeter matrix
− ∞ 3
C = ∞ − 2 .
3 2 −
In other words, we consider the algebra A = F/J where F = Khx, y, zi and J ⊂ F
is the two-sided ideal generated by the two binomials yz − zy, xzx − zxz. The base
field K may be any. To the aim of computing the Hilbert series of A, we calculate
a two-sided Gröbner basis G of J with respect to the graded left lexicographic
monomial ordering of F with x ≻ y ≻ z. It is immediate to verify that G is the
12
R. LA SCALA
following infinite set
G = {yz − zy, xzx − zxz} ∪ {xz 2 z d xz − zxz 2 xxd | d ≥ 0}.
This implies that the leading monomial two-sided ideal I = LM(J) is as follows
I = hyz, xzxi + hxz 2 z d xz | d ≥ 0i.
If X = {x, y, z} and W = X ∗ then the language L = I ∩ W of the monomials of I
can be clearly written as
L = X ∗ {yz, xzx}X ∗ ∪ X ∗ {xz 2 }{z}∗{xz}X ∗.
By the Kleene’s theorem we have that L is a regular language and hence the orbit
OI is a finite set and the Hilbert series HS(A) is a rational function. Recall that
to calculate OI we have to compute the colon right ideals Iw = Tw (I) = (I :R w)
(w ∈ W ). Moreover, by Proposition 4.10 one has a method to compute the colon
right ideals that are defined by a monomial two-sided ideal. Precisely, for computing
Ix = (I :R x) we have to consider the following monomial right ideals
Ix (yz) = 0, Ix (xzx) = hzxi, Ix (xz 2 z d xz) = hz 2 z d xzi (d ≥ 0).
One obtains therefore that
Ix = hzxi + hz 2 z d xz | d ≥ 0i + I.
In a similar way, we compute that Iy = hzi + I and we have that {I, Ix , Iy } ⊂ OI .
Moreover, one has immediately that Iz = I. We compute now
Ix2 (yz) = 0, Ix2 (xzx) = hzxi, Ix2 (xz 2 z d xz) = hz 2 z d xzi (d ≥ 0)
and one concludes that Ix2 = Ix . By considering the ideals
Ixy (yz) = hzi, Ixy (xzx) = 0, Ixy (xz 2 z d xz) = 0 (d ≥ 0)
we obtain also that Ixy = Iy . In a similar way, one has that
Ixz = hxi + hzz d xz | d ≥ 0i + I
and hence {I, Ix , Iy , Ixz } ⊂ OI . It is immediate to calculate now Iyx = Ix , Iy2 = Iy
and Iyz = h1i because yz ∈ I. One obtains then {I, Ix , Iy , Ixz , Iyz } ⊂ OI . Clearly
Ixzx = h1i and we have that Ixzy = Iy . One computes also that
Ixz2 = hz d xz | d ≥ 0i + I
and therefore {I, Ix , Iy , Ixz , Iyz , Ixz2 } ⊂ OI . It is clear that the ideal Iyz = h1i is
invariant under the maps Tx , Ty , Tz and hence we apply them to the ideal Ixz2 . One
computes the following monomial right ideals
2
Ixz2 x (yz) = 0, Ixz2 x (xzx) = hzxi,
Ixz2 x (xz xz) = hzi, Ixz2 x (xz 2 z d xz) = hz 2 z d xzi (d > 0)
and we have therefore that Ixz2 x = Iy . In a similar way, one has that Ixz2 y = Iy
and Ixz3 = Ixz2 and we conclude that
OI = {I, Ix , Iy , Ixz , Iyz , Ixz2 }.
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
Then, for the regular ideal I one has
1
0
0
AI =
0
0
0
13
the following adjacency matrix
1 1 0 0 0
1 1 1 0 0
1 1 0 1 0
0 1 0 1 1
0 0 0 3 0
0 2 0 0 1
and the constant vector CI = (1, 1, 1, 1, 1, 0, 1)T . Denote now by H the column
vector whose entries are the sums of the Hilbert series of the cyclic right modules
defined by the right ideals in the orbit OI . From Theorem 4.5 it follows that H
is obtained by solving over the field Q(t) the linear system corresponding to the
matrix equation (3). In particular, we obtain easily that
1
.
(t − 1)(t2 + 2t − 1)
√
Since the roots of the denominator are 1, −1 ± 2, note finally that the algebra
A = F/J has an exponential growth with exponential dimension (see, for instance,
[14])
p
1
√ .
lim sup d dimK Ad =
−1 + 2
d→∞
HS(A) =
7. Experiments
By means of an experimental implementation of OrbitData that we have developed using the language of Maple, we propose now the computation of the (generalized) Hilbert series of a couple of Hecke algebras which have a finite Gröbner
basis for the two-sided ideal of the relations satisfied by their generators. Precisely,
consider the following two Coxeter matrices of order 4
1323
1332
3132
3133
′
C =
2313 , C = 3313.
3231
2331
Let F = Khx, y, z, vi be the free associative algebra with four generators and define
the following two-sided ideals of F
J = h(x − q)(x + 1), (y − q)(y + 1), (z − q)(z + 1), (v − q)(v + 1),
xyx − yxy, xz − zx, xvx − vxv, yzy − zyz, yv − vy, zvz − vzvi;
J ′ = h(x − q)(x + 1), (y − q)(y + 1), (z − q)(z + 1), (v − q)(v + 1),
xyx − yxy, xzx − zxz, xv − vx, yzy − zyz, yvy − vyv, zvz − vzvi.
Then, the Hecke algebras corresponding to the matrices C, C ′ are by definition
the quotient algebras A = F/J, A′ = F/J ′ . Observe that the coefficients of the
generators of J, J ′ are ±1 together with the parameter “q”. In fact, one has the
same situation for the Gröbner bases of these ideals, that is, the base field K may
be any field containing q. For the graded left lexicographic monomial ordering of
F with x ≻ y ≻ z ≻ v, one computes (see for instance [7, 8]) the following finitely
14
R. LA SCALA
generated leading monomial ideals
LM(J) = hx2 , xz, y 2 , yv, z 2 , v 2 , xyx, xvx, yzy, zvz, xyzx, xvzx, xvzv, yzvy,
xvyxy, xvyxv, xyzvxv, xvzyxy, xvyzxyz, xvyzxvz, xvzyxvy, xvyzvxvz,
xvzyzxyz, xvyzxvyzv, xvzyxvzyz, xvzyzxvyzv, xvzyzxvzyzvi;
LM(J ′ ) = hx2 , xv, y 2 , z 2 , v 2 , xyx, xzx, yzy, yvy, zvz, xyvx, xzvx, xyzxz,
xzyxy, yzvyv, yvzyz, xzvyxy, xyvzxz, xzyxzyz, yvzyvzv, xzyvxyv,
xyzvxzv, xzvyxzyz, xzvyvxyv, xyvzvxzv, xzyxzyvzv, xzvyxzyvzvi.
By performing OrbitData for the monomial two-sided ideals I = LM(J) and
I ′ = LM(J ′ ), we obtain the following Hilbert series
HS(A) =
(1 + t)(1 + t2 )
(1 + t)(1 + t2 )(1 + t + t2 )
, HS(A′ ) =
.
3
(1 − t)
(1 − t)(1 − t − t2 − t3 − t4 )
From the roots of the denominators of such series it follows that the algebra A has
a polynomial growth but the growth of A′ is an exponential one. The number of
elements in the orbits OI , OI ′ are respectively 36 and 33. This implies that the
linear system corresponding to the matrix equation (3) is quite sparse and hence
easy to solve for both the orbits. In other words, the solving time is irrelevant with
respect to the computation of the orbit. With our experimental implementation
in the language of Maple, the total computing times for A, A′ are respectively 0.9
and 0.7 sec. We obtain such timings on a server running Maple 16 with a four core
Intel Xeon at 3.16GHz and 64 GB RAM.
8. Conclusions and future directions
The above experiments clearly show that the proposed method for computing
noncommutative Hilbert series is really feasible. Beside the ease of implementing
it by using Proposition 4.9 and 4.10, we believe that this good behaviour relies on
the property that the deterministic finite automaton which is constructed by the
algorithm OrbitData is a minimal one according to Proposition 5.7. Since the
monomial exact sequence that we use in Section 3 is similar to the ones which are
generally used for the computation of commutative Hilbert series (see, for instance,
[11]), one possible further investigation may consist in applying our method to the
commutative case and in comparing it with other existing strategies. Of course, this
can be done properly only when OrbitData will be implemented in the kernel of a
computer algebra system. Such implementation will make also available the Hilbert
series of involved noncommutative algebras and modules which are of interest in
mathematics and physics.
One drawback of these monomial methods consists in the necessity of computing
Gröbner bases in order to obtain leading monomial ideals. In the noncommutative
case these bases may be infinite but in fact we have proved in Section 5 and illustrated in Section 6 that our method works also in the infinitely generated but
regular monomial case. Nevertheless, one may have, for instance, a finite minimal
presentation of a graded algebra and hence we may prefer to use this finite data
to try to determine its Hilbert series. This eventually leads to the computation of
the homology of the algebra which is of course of great interest in itself. We hope
therefore in the next future to provide new methods also for such computations.
MONOMIAL RIGHT IDEALS AND HILBERT SERIES . . .
15
9. Acknowledgments
The author would like to thank Victor Ufnarovski for introducing him to automata theory. He is also grateful to the research group of Singular [4] for the
courtesy of allowing access to their servers for performing computational experiments. Finally, the author would thank the anonymous referees for the opportunity
to apply the Occam’s razor to the paper once more.
References
[1] Anick, D.J., Noncommutative graded algebras and their Hilbert series. J. Algebra, 78 (1982),
120–140.
[2] Cohn P.M., Free Ideal Rings and Localization in General Rings, New Mathematical Monographs, 3. Cambridge University Press, Cambridge, 2006.
[3] de Luca, A., Varricchio, S., Finiteness and Regularity in Semigroups and Formal Languages,
Monographs in Theoretical Computer Science. An EATCS Series. Springer-Verlag, Berlin,
1999.
[4] Decker, W.; Greuel, G.-M.; Pfister, G.; Schönemann, H.: Singular 4-0-2 — A computer
algebra system for polynomial computations (2015). http://www.singular.uni-kl.de.
[5] Drensky, V., Defining relations of noncommutative algebras, Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, Sofia. http://www.fmi.unisofia.bg/algebra/drensky/DrenskyPalermo.pdf.
[6] Govorov, V.E., Graded algebras. (Russian) Mat. Zametki, 12 (1972), 197–204.
[7] La Scala, R.; Levandovskyy, V., Letterplace ideals and non-commutative Gröbner bases. J.
Symbolic Comput., 44 (2009), 1374–1393.
[8] La Scala, R., Extended letterplace correspondence for nongraded noncommutative ideals and
related algorithms. Internat. J. Algebra Comput., 24 (2014), 1157–1182.
[9] Mansson, J., On the computation of Hilbert series and Poincaré series for algebras with
infinite Gröbner bases, Comput. Sci. J. Mold., 8, (2000), 42–63.
[10] Mora, T., A primer on ideal theoretical operation in non-commutative polynomial rings, J.
Algebra Appl., 14, (2015), 16pp.
[11] Kreuzer, M.; Robbiano, L., Computational commutative algebra 2. Springer-Verlag, Berlin,
2005.
[12] Ufnarovski, V.A., On the use of graphs for calculating the basis, growth and Hilbert series
of associative algebras. (Russian) Mat. Sb., 180 (1989), 1548–1560; translation in Math.
USSR-Sb., 68, (1991), 417–428.
[13] Ufnarovski, V.A, Calculations of growth and Hilbert series by computer. Computational
algebra (Fairfax, VA, 1993), 247–256, Lecture Notes in Pure and Appl. Math., 151, Dekker,
New York, 1994.
[14] Ufnarovski, V.A., Combinatorial and asymptotic methods in algebra. Algebra VI, 1–196,
Encyclopaedia Math. Sci., 57, Springer, Berlin, 1995.
[15] van Suijlekom, W.D., Noncommutative geometry and particle physics. Mathematical Physics
Studies. Springer, Dordrecht, 2015.
∗ Dipartimento di Matematica, Università di Bari, Via Orabona 4, 70125 Bari, Italy
E-mail address: [email protected]
| 0math.AC
|
1
Minimum Probabilistic Finite State Learning
Problem on Finite Data Sets: Complexity, Solution
and Approximations
arXiv:1501.01300v2 [cs.FL] 25 Feb 2017
Elisabeth Paulson and Christopher Griffin
Abstract—In this paper, we study the problem of determining
a minimum state probabilistic finite state machine capable of
generating statistically identical symbol sequences to samples
provided. This problem is qualitatively similar to the classical
Hidden Markov Model problem and has been studied from a
practical point of view in several works beginning with the work
presented in: Shalizi, C.R., Shalizi, K.L., Crutchfield, J.P. (2002)
An algorithm for pattern discovery in time series. Technical Report
02-10-060, Santa Fe Institute. arxiv.org/abs/cs.LG/0210025. We
show that the underlying problem is NP-hard and thus all
existing polynomial time algorithms must be approximations on
finite data sets. Using our NP-hardness proof, we show how
to construct a provably correct algorithm for constructing a
minimum state probabilistic finite state machine given data and
empirically study its running time.
Index Terms—probabilistic finite state machine, minimum
state, estimation, clique covering, mixed integer programming,
Markov chain
I. I NTRODUCTION
The goal of the field of computational mechanics is to
use statistical methods to produce models that explain an
observed, and possibly probabilistic, behavior. In [1], Shalizi
and Crutchfield give an excellent overview of computational
mechanics. These models can be used to simulate the behavior
of, predict future performance for, and gain insights into the
underlying processes that govern the observed behavior [1].
Crutchfield broadly discusses the idea of pattern recognition
and the importance of discerning signal and noise in [2]. This
is the underlying motivation behind computational mechanics–
attempting to produce the most meaningful models to capture
the important features of a behavior.
Perhaps the most common and well-known models for pattern recognition are Hidden Markov models. Hidden Markov
models (HMMs) are a pattern recognition tool used to construct a Markov model when the state space is unknown.
HMMs are used for a wide variety of purposes, such as speech
recognition [3], handwriting recognition [4], [5], and tracking
[6]. Given a process which produces some string of training
data, there are many algorithms that are widely used to infer a
Hidden Markov model for the process. In this paper, we focus
on the approach developed by Shalizi [7] for producing ǫmachines from a string of input data, which can be thought of
as a kind of HMM. This work is extended in [8], [9] and [10].
E. Paulson is with the Operations Research Center, Massachusetts Institute
of Technology, Cambridge, MA, 02142, [email protected]
C. Griffin is with the Mathematics Department, United States Naval
Academy, Annapolis, MD, 21402, [email protected]
A model like the one discussed in this paper is also assumed
in [11]–[13]
Shalizi’s approach to constructing a HMM is to find statistically significant groupings of the training data which then
correspond to causal states in the HMM. In this formulation,
each state is really an equivalence class of unique strings.
This algorithm groups unique strings based on the conditional
probabilities of the next symbol as a window slides over the
training data. The window gradually increases in length up to
a maximum length L, which is the maximum history length
that contains predictive power for the process. This approach
is called the Causal State Splitting and Reconstruction (CSSR)
algorithm. The result of the CSSR algorithm is a Markov
model where each state consists of a set of histories of up
to length L that all share the same conditional probability
distributions on the next symbol.
In this paper we merge states based on a multiple comparison test of the conditional probability distributions of
each state; however, it should be noted that there are many
accepted state merging techniques. Stolcke and Omohundro
propose a state merging technique for Hidden Markov Models
that uses Bayesian posterior probabilities to merge similar
states [14]. Another known merging technique is the subtree
merging algorithm [15]. This is a very similar technique to that
proposed in this paper states are merged together based on the
conditional probabilities of the next symbol. However, due to
the iterative nature of this procedure, it has the same flaw as
the CSSR algorithm in that it is not guaranteed to produce a
minimal state machine. A new approach to estimating epsilonmachines uses a finite sample to generate a set of candidate
unifilar hidden Markov Models, and uses Bayesian Structural
Inference to determine the posterior probability of each model
topology [16]. [17] eschews unifilar models to obtain an
algorithm with linear running time using a Euclidean metric
on conditional symbol distributions.
Assuming an input string of infinite length where the conditional probability distributions converge to their true values,
the CSSR algorithm produces a minimal-state ǫ-machine for
the given process. That is, any time two states could be
combined while still maintaining the desired properties of
the ǫ-machine, they are. As the length of the input string
approaches infinity, the CSSR algorithm is correct in the limit
as a result of the Glivenko-Cantelli Theorem [18]. However,
for input sequences of finite length, the CSSR algorithm does
not guarantee a minimal-state machine due to the technique by
which the strings are grouped. Consequently, state explosion
2
can occur. This paper seeks to develop a new approach, based
on the CSSR algorithm, which always guarantees a minimalstate HMM. We use a very similar idea to the CSSR algorithm,
but formulate it as an integer programming problem whose
goal is to minimize the number of states. We then show that
solving this integer program is NP-hard.
It should be noted that there are many hardness results for
problems that are similar to ours. In [19], Abe et al. show that
polynomial time learnability of stochastic rules, in particular
probabilistic concepts, with respect to the Kullback-Leibler
divergence is equivalent to the same notion using quadratic
distance. Abe et al. give a polynomial time learning algorithm
for a class of convex linear combinations of stochastic rules.
Pitt and Warmuth also consider a slightly different problem
than this paper by considering DFAs whose goal is to learn a
set of positive and negative words and then classify incoming
words into these categories [20]. It was shown by Gold that
determining the smallest possible consistent DFA for this
problem is NP-hard [21], and Pitt and Warmuth show that
determining an approximately small DFA is also hard [20].
In [22] the authors study the problem of distribution free
learning of Boolean functions. In particular, they cover the
case of learning unifilar acyclic finite automata from string
samples. In particular, the results presented in this paper are
analogous and consistent with our results. Where we differ
is in considering probabilistic generating systems leading to
probabilistic finite state machines. Additionally, the inclusion
of the state minimization problem means we must consider
cyclic automata.
The constraints of this integer programming problem ensure
that we obtain a model that is statistically consistent with, and
able to generate, the input string. Thus, it accomplishes the
same generalizability and consistency as the CSSR algorithm.
As both Shalizi and Crutchfield explain, the best model is the
one with minimum size that also minimizes the amount of
error and maximizes predictive power [1], [2]. In this paper,
we make strides at accomplishing that goal by producing
a consistent, generalizable model, that also guarantees state
minimization. We then provide a reformulation of the integer
programming problem as a minimal clique covering problem,
which yields a faster algorithm (in practice) for finding the
minimal-state HMM.
II. N OTATION AND P RELIMINARIES
Let A be a finite set of symbols or alphabet representing the
observable events of a given system. Let y be the symbolic
output of our system, so y is a sequence composed of the
elements in A. We will write y = y1 y2 ...yn to denote the
individual elements in y, so yi ∈ A.
Let A∗ be the set of all sequences composed of symbols
from A, and let λ be the empty sequence. If y is a sequence,
then x is a subsequence if
1) x is a sequence with symbols in A and
2) There are integers i and k such that x = yi yi+1 · · · yi+k .
The subsequences of y of length k constitute the sliding
data windows of length k over y.
For us, a hidden Markov model is a tuple G = hQ, A, δ, pi,
where Q is a finite set of states, A is a finite alphabet, δ ⊆
Q × A × Q is a transition relation, and p : δ → [0, 1] is a
probability function such that
X
p(q, a, q ′ ) = 1 ∀q ∈ Q
(1)
a∈A,q′ ∈Q
This is slightly different than the standard definition of an
HMM as observed in [3] because we are particularly interested
in the state transition relation and associated probabilities,
rather than the symbol probability distribution for a given state.
The Baum-Welch Algorithm is the standard expectation
maximization algorithm used to determine the transition relation and probability function of a hidden Markov model.
However, this algorithm requires an initial estimate of the
transition structure, so some initial knowledge of the structure
of the Markov process governing the dynamics of the system
must be known.
Given a sequence y produced by a stationary process, the
Causal State Splitting and Reconstruction (CSSR) Algorithm
infers a set of causal states and a transition structure for a
hidden Markov model that provides a maximum likelihood
estimate of the true underlying process dynamics. In this
case, a causal state is a function mapping histories to their
equivalence classes, as in [23].
The states are defined as equivalence classes of conditional
probability distributions over the next symbol that can be
generated by the process. The set of states found in this manner
accounts for the unifilar behavior of the process while tolerating random noise that may be caused by either measurement
error or play in the system under observation. The CSSR
Algorithm has useful information-theoretic properties in that
it attempts to maximize the mutual information among states
and minimize the remaining uncertainty (entropy) [23].
The CSSR Algorithm is straightforward. We are given a
sequence y ∈ A∗ and know a priori the value L ∈ Z+ .1
For values of i increasing from 0 to L, we identify the set
of sequences W that are subsequences of y and have length
i. (When i = 0, the empty string is considered to be a
subsequence of y.) We compute the conditional distribution
on each subsequence x ∈ W and partition the subsequences
according to these distributions. These partitions become states
in the inferred HMM. If states already exist, we compare the
conditional distribution of subsequence x to the conditional
distribution of an existing state and add x to this state, if
the conditional distributions are ruled identical. Distribution
comparison can be carried out using either a KolmogorovSmirnov test or a χ2 test with a specified level of confidence.
The level of confidence chosen affects the Type I error rate.
Once state generation is complete, the states are further split to
ensure that the inferred model has a unifilar transition relation
δ. Algorithm 1 provide pseudocode for the CSSR Algorithm.
The complexity of CSSR is O(k 2L+1 ) + O(N ), where k is
the size of the alphabet, L is the maximum subsequence length
considered, and N is the size of the input symbol sequence
[23]. Given a stream of symbols y, of fixed length N , from
alphabet A, the algorithm is linear in the length of the input
data set, but exponential in the size of the alphabet.
1 There
are some heuristics for choosing L found in [23]
3
Algorithm 1 – CSSR Algorithm
Input: Observed sequence y; Alphabet A, Integer L
Initialization:
1: Define state q0 and add λ (the empty string) to state q0 . Set Q = {q0 }.
2: Set N := 1.
Splitting (Repeat for each i ≤ L)
1: Set W = {x|∃q ∈ Q(x ∈ q ∧ |x| = i − 1)} {The set of strings in states of
the current model with length equal to i − 1}
2: Let N be the number of states.
3: for each x ∈ W do
4:
for each a ∈ A do
5:
if ax is a subsequence of y then
6:
Determine fax|y : A → [0, 1], the probability distribution over the next
input symbol.
7:
Let fqj |y : A → [0, 1] be the unified state conditional probability
distributions; that is, the probability given the system is in state qi , that the
next symbol observed will be a. For each j, compare fqj |y with fax|y
using an appropriate statistical test with confidence level α. Add ax to
the state that has the most similar probability distribution as measured by
the p-value of the test. If all tests reject the null hypothesis that fqj |y
and fax|y are the same, then create a new state qN +1 and add ax to
it. Set N := N + 1.
8:
end if
9:
end for
10: end for
Reconstruction
1: Let N0 = 0.
2: Let N be the number of states.
3: while N0 6= N do
4:
for each i ∈ 1, . . . , N do
5:
Set k := 0
6:
Let M be the number of sequences in state qi . Choose a sequence x0 from
state qi .
7:
Create state pik and add x0 to it
8:
for all sequences xj (j > 0) in state qi do
9:
For each a ∈ A xj a produces a sequence that is resident in another
state qk . Let δ(xj , a) = qk .
10:
For l = 0, . . . , k, choose x from pik . If δ(xj , a) = δ(x, a) for all
a ∈ A, then add xj to pik . Otherwise, create a new state pik+1 and
add xj to it. Set k := k + 1.
11:
end for
12:
end for
13:
Reset Q = {pik }; recompute the state conditional probabilities fq|y for q ∈
Q and assign transitions using the δ functions defined above.
14:
Set N0 = N .
15:
Set N to be the number of states in our current model.
16: end while
17: The model of the system has state set Q and transition probability function computed
from the δ functions and state conditional probabilities.
A. A Practical Problem with CSSR
As observed in Shalizi and Shalizi (2004), the CSSR converges to the minimum state estimator asymptotically [23],
however it does not always result in a correct minimum state
estimator for a finite sample. With an infinitely long string, all
of the maximum likelihood estimates converge to their true
value and the CSSR algorithm works correctly. That is, as
N → ∞, the probability of a history being assigned to the
wrong equivalence class approaches zero. A formal proof of
this is given in [7] and relies on large deviation theory for
Markov chains.
The error in estimating the conditional probabilities with
strings of finite length can cause the CSSR algorithm to
produce a set of causal states that is not minimal.
The following example will clarify the problem. Consider
A = {0, 1} and y defined as follows:
yi = 0 1 ≤ i ≤ 518
[yi yi+1 yi+2 yi+3 ] = 1100 518 < i ≤ 582
[yi yi+1 yi+2 ] = 100 582 < i ≤ 645
[yi yi+1 yi+2 ] = 101 645 < i ≤ 648
(2)
Without loss of generality, we consider exclusively strings
of length two. The strings, in order of appearance, are
{00, 01, 11, 10}. The conditional probability distribution on
the next element for each string is shown in Table I:
TABLE I
C ONDITIONAL P ROBABILITIES
string
00
01
11
10
P r(0|string)
0.9314
0.5789
1.0
0.9737
P r(1|string)
0.0686
0.4211
0
0.0263
The p-values associated with comparing each string’s conditional probability distribution to that of every other string
are shown in Table II:
TABLE II
P - VALUES
string
00
01
11
10
00
1.0000
0
0.0067
0.0944
01
–
1.0000
0
0
11
–
–
1
0.7924
10
–
–
–
1
We now step through the Splitting phase of the CSSR
algorithm:
1) String 00 is put into its own state, q1
2) We use Pearson’s Chi-Squared test to test if the 01 and
00 have the same conditional distribution. This results in
a p-value of approximately 0, as seen Table II so string
01 is put into a new state, q2
3) We compare the conditional distribution of 11 to string
00. This results in a p-value of 0.0067, and comparing
it to 01 results in a p-value of 0, so string 11 is put into
a new state, q3
4) We compare the conditional probability distribution of
10 to that of 00. This results in a p-value of 0.0944.
Comparing it to 00 results in a p-value of 0, and
comparing it to 11 results in a p-value of 0.7924. Since
0.7924 is greater than 0.0944 and greater than the chosen
significance level of 0.05, the string 10 is clustered with
string 11, so q3 = {11, 10}.
Thus, at the end of the Splitting step of CSSR, we are left we
three different states:
q1 = {00}
q2 = {01}
q3 = {11, 10}
We now go on to the Reconstruction step. Let xij be a
state in qi where j ∈ 1, 2, ..., |qi |. The Reconstruction step
checks whether, for each a ∈ A, δ(xi1 , a) = δ(xi2 , a) = ... =
qk . Recall that δ(xij , a) = qk means that xij a produces a
sequence that resides in state qk . If this is not satisfied, then
state qi must be broken up into two or more states until we
have a unifilar transition relation function.
In our example, the first two states do not need to be checked
since they each only consist of one string. We check the third
4
state: δ(11, 0) = q3 since string 10 ∈ q3 , and δ(10, 0) = q1
since string 0 ∈ q1 . Thus, determinism is not satisfied and
state q3 must be split into two states. The result of the CSSR
algorithm is a four-state Markov chain where each string
resides in its own state.
We now show why this result is not the minimal state
Markov generator for this input sequence. Suppose that during
the Splitting step, string four was put into state q1 instead of
q3 . This could have been done soundly from a statistical point
of view, since 0.0944 is also greater than our alpha-level of
0.05. However, by the CSSR heuristic, the fourth string was
grouped according to the highest p-value. If the fourth string
were put in q1 , then after the Splitting step we would have
obtained the following states:
an even k, the integer program will result in a machine with
k2
k2
and an odd k will result in a machine with
2 + ⌈ 6 ⌉ states,
(k−1)k
(k2 −3k)
+
⌈
⌉
+ k states. A figure showing a comparison
2
6
between CSSR and the integer program for varying values of
k can be seen in Figure 1.
q1 = {00, 10}
q2 = {01}
q3 = {11}
Now we move on to the Reconstruction step. This time,
we only need to consider state q1 . Notice that δ(10, 0) =
δ(00, 0) = q1 and δ(00, 1) = δ(10, 1) = q2 . We see that
state q1 does satisfy determinism, and does not need to be
split. Thus, our final number of states is only three, which is
minimal.
This provides us with motivation to reformulate the CSSR
algorithm so that it always produces a the minimal number of
states, even when given a small sample size. The remainder of
this paper seeks to address this issue, proposes a reformulation
of the CSSR algorithm which always results in minimal state
models even with a finite sample, and discusses the computational complexity of the minimum state unifilar probabilistic
finite state automaton modeling problem.
B. State Explosion with CSSR
Although in the preceding example, the number of states
that the CSSR algorithm returned was only one larger than the
integer program, the differences become larger as we increase
the size of the alphabet, k. In this section we generalize the
preceding example to show that state explosion can occur
quadratically as a function of k, while keeping L = 2.
In the sequel, we treat individual L-length strings as vertices
in a graph whose edge set consists of pair of vertices with statistically equivalent next-symbol distributions. In the example,
the problem with CSSR arose from a connected component
of this graph that contained three vertices of the form xy zy
and uv (00, 10 and 11 in the example). The CSSR algorithm
clustered xy and uv together, resulting in the state being split,
whereas the integer program clustered xy and zy together,
satisfying the unifilar property. The “worst case” string for the
CSSR algorithm is one that maximizes the number of these
special three-vertex components, and then chooses the wrong
clustering for each of them.
This worst case scenario results in CSSR producing a
machine with k 2 states. The integer program we discuss
in Section III, however, would choose correctly for each
component so that no states are split during the splitting
phase. With some simple calculation it can be seen that, for
Fig. 1. The number of states produced by CSSR and the integer program for
the ”worst case” strings
It remains to see that there are indeed strings which have
this “worst case” property of their graph counterpart having
the maximum possible number of special three-vertex components. In order to show this, we argue that given any conditional probability distribution on all strings of length L, we can
build a sequence whose strings of length L have statistically
equivalent distributions to the specified distribution.
A formal proof of this is a part of the proof of our general
NP-hardness result. Intuitively, this can be accomplished by
constructing a Markov chain where each string of length L
corresponds to a state and creating probabilistic transitions as
needed. We then use this chain to produce a sequence. A long
enough sequence will have statistically equivalent properties to
a desired distribution, and can thus be constructed as a “worst
case” string.
III. I NTEGER P ROGRAMMING F ORMULATION
Much of this section has been adapted from [24]. Suppose
we have a string y ∈ A∗ and we assume that L is known.
Let W be the set of distinct substrings of length L in y, and
let |W | = n. When we cluster the elements of W into states
we must be sure of two things: 1) That all strings in the same
state have the same conditional distributions, and 2) that the
resulting transition function is unifilar. Our goal is to formulate
an integer programming problem that minimizes the number
of states and preserves the two conditions above. Define the
following binary variables:
(
1 string i maps to state j
xij =
(3)
0 else
Remark 1. We assume implicitly that there are n states, since n
is clearly the maximum number of states that could be needed.
If, for some j, xij = 0 ∀ i, then that state is unused, and our
true number of states is less than n.
Let:
5
zilσ =
(
1 string i maps to string l upon symbol σ
0 else
(4)
For example, zilσ = 1 if i = hx1 , ..., xL i and l =
hx2 , ..., xL , σi. Assuming j and k are state indices while i
σ
and l are string indices, we also define the variable yjk
by the
following logical rule:
σ
(xij = 1) ∧ (zilσ = 1) ∧ (xlk = 1) =⇒ (yjk
= 1)
This variable ensures the encoded transition relation maps
each state to the correct next state given a certain symbol.
Specifically, if string i in clustered in state j and string i
transforms to string l given symbol σ, and string l is clustered
in state k, then (j, σ, k) must be the transition relation. This
can be written as the constraint
σ
(1 − xij ) + (1 − zilσ ) + (1 − xlk ) + yjk
≥ 1 ∀i, j, k, l, σ
In order for the automata to be unifilar, we must have the
condition that
X
σ
≤ 1 ∀j, σ
yjk
k
This condition ensures that for a given symbol, each state can
transition to only one other state.
We ensure that the strings in each state have the same conditional probability distributions using a parameter µ, where
(
1 strings i and l have identical conditional distributions
µil =
0 else
(5)
In order to determine if two strings have identical distributions an appropriate statistical test (like Pearson’s χ2 or
Kolmogorov-Smirnov) must be used. The variables must satisfy
(xij = 1) ∧ (xlj = 1) =⇒ (µil = 1)
This states that strings i and l can only belong to the same
state if they have statistically indistinguishable distributions.
This can be written as the following constraint
(1 − xij ) + (1 − xlj ) + µil ≥ 1 ∀i, l, j
Finally, we define the variable p to be used in our objective
function. From Remark 1, this variable simply enumerates up
the number of “used” states out of n. That is:
( P
0
i xij = 0
pj =
(6)
1 else
This can be enforced as the following constraint
P
xij
, pj ∈ {0, 1}
pj ≥ i
n
P
x
Note that in ij ≤ 1. P
These constraints are equivalent to
0, and when
Equation
6
because
when
i xij = 0, pj will beP
P
x
>
0,
p
will
be
1,
when
we
minimize
ij
j
j pj in our
i
optimization problem. This addition extends the work done in
[24], in which a series of optimization problems had to be
solved.
Definition 1 (Minimum State Deterministic pFSA Problem).
The following binary integer programming problem, which, if
feasible, defines a mapping from strings to states and a probability transition function between states which satisfies our two
requirements and is called the Minimum State Deterministic
pFSA (MSDpFSA) Problem.
X
pj
min
j∈{1,...,n}
σ
σ
s.t. (1 − xij ) + (1 − zil ) + (1 − xlk ) + yjk ≥ 1,
∀i, l ∈ W, j, k ∈ {1, ..., n}, σ ∈ A
X
σ
yjk ≤ 1 ∀j, k ∈ {1, ..., n}, σ ∈ A
k
(1 − xij ) + (1 − xlj ) + µil ≥ 1,
P (N ) =
∀i, l ∈ W, j ∈ {1, ..., n}
P
xij
∀i ∈ W, j ∈ {1, ..., n}
pj ≥ i
n
X
xij = 1 ∀i ∈ W
j
pj ∈ {0, 1} ∀j ∈ {1, ..., n}
xij ∈ {0, 1} ∀i ∈ W, j ∈ {1, ..., n}
σ
yjk
∈ {0, 1} ∀j, k ∈ {1, ..., n}, σ ∈ A
(7)
We note that in Problem 7, the constraint (1 − xij ) + (1 −
xlj ) + µil ≥ 1 becomes redundant µil = 1, and similarly the
σ
constraint (1−xij )+(1−zilσ )+(1−xlk )+yjk
≥ 1 is redunant
when zil = 1.
The following proposition is clear from the construction of
Problem 7:
Proposition III.1. Any optimal solution to MSDpFSA yields
an encoding of a minimum state probabilistic finite machine
capable of generating the input sequence(s) with statistically
equivalent probabilities. Further,
1) There is a starting state and set of transitions that
produces the input sequence exactly
2) Additional strings produced by a solution to MSDpFSA
are statistically equivalent to the input string
3) The number of states in any solution to MSDpFSA is
bounded above by the number of unique strings of length
L in the input sequence
Remark 2. Because this formulation does not rely on the
assumption of infinite string length to correctly minimize the
number of states, it succeeds where the CSSR algorithm fails.
Instead of clustering strings into the state with the most identical conditional probability distributions, this optimization
problem uses the identical distributions as a condition with
the goal of state minimization. Thus, in the example, even
though the fourth string had a higher p-value when compared
to the third state than the first state, since both of the p-values
are greater than 0.05 this algorithm allows the possibility of
clustering the fourth string into either state, and chooses the
state that results in the smallest number of final states after
reconstruction. Unlike CSSR, this algorithm does not rely on
asymptotic convergence of the conditional probabilities to their
6
IV. C OMPUTATIONAL C OMPLEXITY
TABLE III
I NTEGER P ROGRAM VARIABLES AND PARAMETERS
–
00
01
11
10
(a) x values
q1
q2
q3
1
0
0
0
1
0
0
0
1
1
0
0
q4
0
0
0
0
–
00
01
11
10
(b) µ values
00
01
11
1
0
0
0
1
0
0
0
1
1
0
1
10
1
0
1
1
–
00
01
11
10
(c) z 0 values
00
01
11
1
0
0
0
0
0
0
0
0
1
0
0
10
0
1
1
0
–
00
01
11
10
(d) z 1 values
00
01
11
0
1
0
0
0
1
0
0
1
0
1
0
10
0
0
0
0
(e) y 0 values
–
q1
q2
q3
q1
1
0
0
q2
1
0
0
q3
1
0
0
(f) y 1 values
–
q1
q2
q3
q1
0
1
0
q2
0
0
1
q3
0
0
1
(g) p values
1
1
1
0
OF
MSD P FSA
In this section we assert that the integer programming
problem given by Problem 7 is NP-hard.
We show that MSDpFSA is NP-hard by reduction to the
minimum clique covering problem. Recall given a graph, a
minimal clique covering problem is to identify a set of cliques
in the graph so that each vertex belongs to at least one clique
and so that this set of cliques is minimal in cardinality. We construct this proof in stages: (i) To illustrate the idea behind the
reduction, we show that the Minimum State Non-Deterministic
pFSA Problem, in which we drop the determinism constraint,
is NP-hard by reduction to the minimum clique covering
problem. (ii) We show for an aribtrary graph structure, we
can create a data gadget that transforms the minimum clique
covering problem into MSDpFSA in polynomial time.
A. Minimum State Non-Deterministic pFSA Problem
TABLE IV
S TATE T RANSITION P ROBABILITIES
–
q1
q2
q3
q1
0.9341
0.5789
1
q2
0.0659
0
0
q3
0
0.4211
0
true value. However, it is clear that as the sample size increases
the probability distributions will become more exact, leading
to better values of µ.
A. Resolution to the Example Problem
Using the integer program formulation to solve the example
previously presented does result in the minimum state estimator of the process represented by the given input string. The
integer program finds the following solution for the variables
and parameters z, x, y, µ, and p seen in Table III. Our objective
function is to minimize the sum of the variables pj . As can
be seen in Table III, there is only one nonzero column in the
x matrix (the last column). Thus, only p4 = 0, and the the
value of our objective function is 3.
We can also recover a matrix of transition probabilities
shown in Table IV. Our final reduced state space machine
is shown in Figure 2.
While the CSSR algorithm attempts to find a minimal state
unifilar finite-state automata, being unifilar is not a necessary property for accurate reconstruction. For a given input
sequence, the construction of a finite-state automata which
is not necessarily unifilar is less computationally intensive
than solving the MSDpFSA problem, and is discussed by
Schmiedekamp et al. (2006) [17]. Like the CSSR algorithm,
the algorithm presented in [17] is a heuristic which does not
always results in a minimal state probabilistic FSA as we show
in the sequel. The following definition provides an integer
programming problem for optimally solving a minimal state
pFSA which is not necessarily unifilar. This is identical to
Problem 7 except that we discard the constraints that ensure
the automata is unifilar.
Definition 2 (Minimum State Non-Deterministic pFSA Problem). The following binary integer programming problem,
which, if feasible, defines a mapping from strings to states
and a probability transition function between states which is
not necessarily unifilar and is called the Minimum State NonDeterministic pFSA (MSNDpFSA) Problem.
X
pj
min
j∈{1,...,n}
s.t. (1 − xij ) + (1 − xlj ) + µil ≥ 1,
∀i, l ∈ W, j ∈ {1, ..., n}
P
x
ij
P ′ (N ) =
∀i ∈ W, j ∈ {1, ..., n} (8)
pj ≥ i
n
X
xij = 1 ∀i ∈ W
j
pj ∈ {0, 1} ∀j ∈ {1, ..., n}
xij ∈ {0, 1} ∀i ∈ W, j ∈ {1, ..., n}
Theorem IV.1. MSNDpFSA is NP-Hard.
Fig. 2. The reduced state space machine derived from the input given by
Equation 2
Proof: Let G = (V, E) be the graph on which we
want to find the minimal clique covering. Assume that V =
{1, 2, ..., n} and let I be a n×n matrix such that Iij = 1 ⇐⇒
there is an edge connecting vertices i and j. We reduce
Problem 8 by letting n be the number of unique strings of
length L in y, so we can let each string correspond to a vertex
7
of G. Let Iij = 1 ⇐⇒ µij = 1. This means that two strings
are connected if and if only if they have identical conditional
probability distributions. We show that Problem 8P
is equivalent
to finding a minimal clique cover of G where j pj is the
number of cliques. Let the set of cliques corresponding to the
minimal clique cover of G be C = {C1 , ..., Cm }, where Cj
is a specific clique, and Cj = Vj ⊂ V .
We can define the variables x by xij = 1 ⇐⇒ Vi ∈
Cj . Thus, the constraint that (1 − xij ) + (1 − xlj ) + µil ≥
1 ∀i, l, j simply means that if two vertices are in the same
clique then there must be an edge between them. We also
have that pj = 0 iff there is at least one vertex in clique
j, so the set of j such that pj = 1 corresponds to the nonempty cliques, i.e., is identical to C (since
P C only consists
of non-empty cliques). Thus, minimizing j pj is equivalent
to minimizing the
P number of cliques needed to cover G. The
constraint that j xij = 1 ∀i simply means that each vertex
belongs to exactly one clique. Thus, it is clear that the integer
programming problem given in Problem 8 is equivalent to a
minimal clique covering and is NP-hard.
We illustrate the equivalence of Problem 8 and the minimal
clique covering through an example. Using the same example
as before, the resulting graph G is shown in Figure 3(a). The
string 00 and 10 have an edge in common because they have
identical conditional distributions as noted by Table II. By the
same reasoning, 10 and 11 also have an edge in common.
However, 00 and 11 do not share an edge, and 01 has no
incident edges.
The integer programming problem given by Equation 8
produces either of the following two results:
Q = {q1 = {00, 10}, q2 = {11}, q3 = {01}} or
Q′ = {q1 = {11, 10}, q2 = {00}, q3 = {01}}
Both of these two groupings results in one of two minimal
clique coverings. Let V1 = 00, V2 = 10, V3 = 11, V4 = 01.
C = {C1 = {1, 2}, C2 = {3}, C3 = {4}} or
C ′ = {C1 = {2, 3}, C2 = {1}, C3 = {4}}
(a)
(b)
Fig. 3. (a) The graph G corresponding to the string y given in Equation 2.
(b) The graph G considered by the CSSR algorithm.
When the unifilar constraint is added to the integer programming formulation, the result is Q′ (or equivalently C ′ )
instead of Q (or C). The solver recognizes that Q (or C)
would have to be split in order to satisfy the unifilar property,
so Q′ (or C ′ ) is chosen instead. If we think of the CSSR algorithm in terms of our graph equivalency, the CSSR algorithm
does not consider the existence of an edge between 00 and 10.
Let T be Table II. For a vertex i, CSSR only places an edge
between i and argmaxj6=i {T (i, j) : T (i, j) > 0.05}.Thus, by
the CSSR algorithm, the graph is really the image shown in
Figure 3(b). In this formulation, there is only one minimal
state clustering (minimal clique covering), G (or C) which
does not satisfy the unifilar property, so G is split into four
states.
We now state and prove the main result of this paper.Like
the proof of Theorem IV.1, the proof of the following theorem
relies on reduction to the Minimum Clique Cover problem, but
the reduction gadget is more complex to construct.
Theorem IV.2. MSDpFSA is NP-hard.
Proof: We show that for any arbitrary graph G there is
an input sequence y so that the solution to the Minimum State
Deterministic pFSA (MSDpFSA) problem generated from this
string solves the minimum clique covering problem for G. It
follows at once that MSDpFSA is NP-hard.
Let G = (V, E) be an arbitrary graph with vertices V
and edges E. Since G is arbitrary, it may be composed of
several components. Two vertices v1 , v2 ∈ V are in the same
component of G if there is a walk from v2 to v1 .
As before, we will map the vertices of G to unique substrings of length L taken from the input sequence y. Let n0
be the cardinality of the largest component in G, let c(G)
be the number of components and let N = max{n0 , c(G)}.
Let L = 2 and assume we have alphabet A = {0, . . . , N }.
There are n0 + 1 strings of length 2 with form {s0 : s ∈
{0, . . . , n0 − 1}}. For example, the strings 00,10, 20 etc. all
satisfy this pattern. To each vertex in the largest component
of G (with size n0 ) assign one of these strings.
Let n1 be the size of the second largest component and
repeat this process of string assignment but with the set of
length 2 strings {s1 : s ∈ {0, . . . , n1 − 1}. It is clear we can
repeat this process until each vertex in V has been assigned a
length 2 string.
Since L = 2 and our alphabet has N + 1 symbols, there are
(N + 1)2 unique strings. For each possible string that was not
already assigned to a vertex, add a vertex to G corresponding
to this string to create a new graph G′ . Call these added
vertices V ′ . Clearly G is a subgraph of G′ and any minimum
clique covering of G produces a minimum clique covering of
G′ since the additional isolated vertices are necessarily in their
own clique.
Notice any clique in G′ not made up of vertices in V ′ must
be a subset of a component in G. Let δ : V ∪V ′ ×A → V ∪V ′
be the induced transition function on the graph G′ when we
associate each vertex with its corresponding string and let c(v)
denote the component in which v resides. By construction,
δ has the property that if σ ∈ A and v1 , v2 ∈ V and
c(v1 ) = c(v2 ), then δ(v1 , σ) = δ(v2 , σ). For example every
string (vertex) in the set {00, 10, · · · (n0 − 1)0} transitions to
the string 0σ on symbol σ. Therefore by construction, every
possible clique on G′ must transition to one and only one
other clique. Thus, any minimal clique covering is guaranteed
to generate a unifilar transition function. In particular, on σ,
this component will map either to an extra string or to the
next σ th component.
8
It remains to show that we can construct a sequence
polynomial time using the alphabet A so that the conditional
probabilities of each string of length two ensure that we obtain
the correct edges to produce the graph G′ and thus G. If
so, then MSDpFSA reduces tot the minimum clique cover
problem.
Because the equivalence of two probability distributions in
the CSSR algorithm is defined through a statistical test, the
distributions only need to match within a predefined value ǫ
in order to be deemed statistically equivalent. This value of ǫ
will vary by statistical test, sample size, etc. but it exists.
To construct the string, define a Markov chain so that each
state corresponds to the vertex set of V ∪ V ′ , where vertices
are treated as their length 2-string assignments. Suppose v corresponds to string α1 α2 , we enforce a transition structure that
ensures only transitions of the form α1 α2 7→ α2 α3 may have
non-zero probability p3|12 . The resulting substring produced
by traversing these states is α1 α2 α3 and asymptotically will
approach p3|12 as the Markov chain is used to generate the
string. By judiciously building the transition probabilities, we
may ensure that in an appropriate statistical test, the string y
that is generated by the Markov chain will produce conditional
probabilities satisfying the following constraints:
|pσ|v1 − pσ|v2 | ≤ ǫ ∀v1 , v2 ∈ E, ∀σ ∈ A
X
|pσ|v1 − pσ|v2 | ≥ |A|ǫ ∀v1 , v2 6∈ E
σ∈A
X
pσ|v = 1 ∀v ∈ V
σ∈A
pσ|v ≥ 0 ∀v ∈ V, σ ∈ A
The second constraint ensures there is at least one symbol
σ ∈ A so that |pσ|v1 − pσ|v2 | ≥ ǫ.
For small enough ǫ (implying a potentially large sample
length), the feasible region is non-empty and thus there is
an assignment of probabilities satisfying these constraints.
We now show that this assignment can be accomplished in
polynomial time. Note first, we can re-write these constraints
as:
pσ|v1 − pσ|v2 ≤ ǫ
∀v1 , v2 ∈ E, ∀σ ∈ A
σ∈A
pσ|v = 1
∀v ∈ V
σ∈A
pσ|v ≥ 0 ∀v ∈ V, σ ∈ A
eliminating one non-linear term and leaving only the (now)
third (non-linear) constraint. For all v1 , v2 6∈ E and σ ∈ A,
define:
+
−
− zσ|v
pσ|v1 − pσ|v2 = zσ|v
1 v2
1 v2
σ∈A
+
−
− zσ|v
pσ|v1 − pσ|v2 = zσ|v
,
1 v2
1 v2
∀v1 , v2 6∈ E, ∀σ ∈ A
X
pσ|v = 1 ∀v ∈ V
σ∈A
+
−
pσ|v , zσ|v
, zσ|v
≥0
1 v2
1 v2
∀v ∈ V, σ ∈ A
(10)
We know that Equation 9 must hold. Further, the objective
+
function is minimized if and only if either zσ|v
> 0 or
1 v2
−
+
−
zσ|v1 v2 > 0, but not both. Therefore, (zσ|v1 v2 + zσ|v1 v2 ) =
|pσ|v1 −pσ|v2 |. Thus, the constraints of the linear programming
problem (Expression 10) are equivalent to the constraints
required by our probability assignment problem. Linear programming problems are solvable in polynomial time by [25].
Furthermore, the interior point methods that are known to
be polynomial will return feasible points very close to extreme points, implying that we can find a pσ|v > 0. (It is
actually sufficient to execute the optimization only until the
original constraints are satisfied.) After transition probability
assignment, remove any states that do not have in-transitions,
these are unreachable. The fact that pσ|v > 0 ensures the
irreducibility of the Markov chain. For any given state xy,
there is non-zero probability of transitioning to yz, where z
is an arbitrary element of A. Furthermore, yz has a non-zero
probability of transitioning to zw, where w is also an arbitrary
element of A. Thus, the chain is irreducible because every
state can be reaching by every other state. Strings produced by
the resulting Markov chain will have statistical properties that
yield a sub-string statistical equivalence graph like G′ . Thus
we have found a polynomial time reduction of MSDpFSA
to the Minimum Clique Covering Problem. Therefore, the
MSDpFSA is NP-hard.
V. M INIMAL C LIQUE C OVERING R EFORMULATION
CSSR A LGORITHM
pσ|v2 − pσ|v1 ≤ ǫ ∀v1 , v2 ∈ E, ∀σ ∈ A
X
|pσ|v1 − pσ|v2 | ≥ |A|ǫ ∀v1 , v2 6∈ E
X
problem:
X
+
−
min
zσ|v
+
z
v
σ|v
v
1 2
1 2
v1 ,v2 6∈E,σ∈A
s.t. pσ|v1 − pσ|v2 ≤ ǫ ∀v1 , v2 ∈ E, ∀σ ∈ A
pσ|v2 − pσ|v1 ≤ ǫ ∀v1 , v2 ∈ E, ∀σ ∈ A
X
+
−
≥ |A|ǫ ∀v1 , v2 6∈ E
zσ|v
+ zσ|v
1 v2
1 v2
(9)
+
−
where zσ|v
, zσ|v
≥ 0. Consider the linear programming
1 v2
1 v2
OF
In this section we present an algorithm for determining
the minimal state hidden Markov model using a minimal
clique vertex covering reformulation. As shown in the previous
section, the two problem formulations are equivalent if we
exclude the unifilar constraint. Let W = {S1 , ..., Sn } be the
set of unique strings of length L in A, so W is the set of
vertices in our graph formulation. In the revised algorithm,
we first use the CSSR Algorithm, Algorithm 1, to find an
upper bound on the number of cliques needed. We then use
Bron-Kerbosch algorithm to enumerate all maximal cliques,
C = {C1 , C2 , ..., Cm }, given as Algorithm 2 [26].
Define H := argminR {|R| : R ⊂ C ∧ ∪i Ri = W }.
Thus H is the set of maximal cliques of minimal cardinality
such that every vertex belongs to at least one clique. This
9
can be found using a simple binary integer programming
problem given in Algorithm 3. We can think of Algorithm
3 as defining a mapping from every string to the set of all
maximal cliques. A clique is “activated” if at least one string
is mapped to it. The goal, then, is to activate as few cliques
as possible. Of course, each string can only be mapped to a
clique in which it appears. We define a few variables, which
are present in our integer programming formulation:
yi =
(
1 Ci ∈ H
0 else
(11)
The variable y can be thought of as whether or not clique Ci
is activated.
(
1 Si ∈ Cj ∈ H
Iij =
(12)
0 else
(
1 Si is mapped to Cj ∈ H
sij =
(13)
0 else
To distinguish between I and s, notice that each string is only
mapped to one clique, but each string could appear in more
than one clique.
Proposition V.1. The set H is a minimal clique vertex
covering.
Proof: Suppose there is a minimal clique vertex covering
of a graph G with k cliques such that every clique is not
maximal. Choose a non-maximal clique and expand the clique
until it is maximal. Continue this procedure until every clique
is maximal. Let our set of cliques be called R. Clearly |R| = k
since no new cliques were created. Also note that R ⊂ C
since R consists of maximal cliques. Further, ∪i Ri = W since
we started with a clique vertex covering. Finally, notice that
R = argminH {|H| : H ⊂ C ∧ ∪i Hi = W } because |R| = k
and k is the clique covering number of G, so it is impossible
to find an H with cardinality less than k.
Proposition V.2. Any solution to Q(C) results in a minimal
clique vertex covering.
Proof: This is clear by noting that argmin Q(C) results
in the subset, H, of C = {C1 , C2 , ..., Cm } where H is defined
as argminR {|R| : R ⊂ C and ∪i Ri = W } in conjunction
with Proposition V.1.
Before discussing the rest of the algorithm, which concerns
the unifilar property, we first state an important remark about
Algorithm 3. It is possible that the set H is such that some
vertices are in more than one maximal clique. However, we are
actually interested in the set of minimal clique vertex coverings
for which each vertex only belongs to one clique, which can
be extracted from H. See Figures 4 and 5 for an example.
The set H is shown in Figure 4. From this set H, we can
deduce the two minimal clique vertex coverings shown in 5.
In initializing Algorithm 4, which imposes the unifilar property
on each covering, we find the set of all minimal coverings for
which each vertex belongs to one clique.
Algorithm 2 –Finding all maximal cliques
Input: Observed sequence y; Alphabet A, Integer L
Setup:
1: Call the set of strings that appear in y W = {S1 , ..., Sn }, so W is our
set of vertices
2: Determine fx|y (a) for each unique string
3: Generate an incident matrix, U ∈ Rnxn , Uij = uij , where uij is
defined in 5
4: P ← [1, 2, ..., n] {P holds the prospective vertices connected to all
vertices in R}
5: R ← ∅ {Holds the currently growing clique}
6: X ← ∅ {Holds the vertices that have already been processed}
function Bron-Kerbosch(R,P,X)
1: k ← 0
2: if P = ∅ and X = ∅ then
3:
report R as a maximal clique, R = Ck
4:
k ←k+1
5: end if
6: for each vertex v in P do
7:
Bron-Kerbosch(R ∪ v, P ∩ N (v), X ∩ N (v)) {where N (v) are the
neighbors of v}
8:
P ←P \v
9:
X ←X ∪v
10: end for
11: return {C1 , C2 , ..., Cm }
Algorithm 3 –Minimal Clique Vertex Covering
Input: Observed sequence y; Alphabet A, Integer L
Set-up:
1. CSSR
run the CSSR Algorithm (Algorithm 1)
return k, the number of cliques found by CSSR
2. Find all maximal cliques
run Algorithm 2
return {C1 , C2 , ..., Cm }
Finding all minimal clique coverings:
Let Iij = 1 denote string Si belonging to clique Cj
Let sij = 1 denote string Si being mapped to clique Cj
X
min
yj
j∈{1,...,m}
{yj indicates whether clique Cj is being used or not}
P
i sij
s.t. yj ≥
∀i ∈ {1, ..., n}, j ∈ {1, ..., m}
|Cj |
Q(C) =
X
yj ≤ k ∀j ∈ {i, .., m}
j
sij ≤ Iij ∀i ∈ {1, ..., n}, j ∈ {1, ..., m}
X
sij = 1 ∀i ∈ {1, ..., n}, j ∈ {1, ..., m}
j
Once we have determined the set of minimal clique covers
of our original graph where each vertex only belongs to one
clique, we select a final clique covering which is minimal
and unifilar. This is done by considering each minimal clique
covering individually and then restructuring it when necessary
to be unifilar. This corresponds to the reconstruction part of
Algorithm 1. Note in Algorithm 4 each vertex corresponds to
a string of length L, which came from an initial string y. Also
recall that we have previously defined z as
zilσ =
(
1 string i maps to string j upon symbol σ
0 else
where σ is any element of our alphabet A. The following
proposition is clear from the construction of Algorithms 3 and
10
Algorithm 4 – Minimal Clique Covering Reconstruction
4.
Proposition V.3. Algorithm 3 along with Algorithm 4 produces an encoding of a minimum state probabilistic finite
machine capable of generating the input sequence.
Fig. 4. An example output from Algorithm 3
Fig. 5. Minimal clique coverings where each vertex is only in one clique
VI. C OMPARING
RUN TIMES OF MODIFIED
Input: Observed sequence y; Alphabet A, Integer L;
Set-up:
Perform Algorithm 2 on y and obtain a minimal clique covering
From this initial clique covering, find the set of all minimal coverings such
that each vertex belongs to exactly one clique, an example of which is seen
in Figure 5. Let this set of minimal coverings be T = {T1 , ..., Tl }
From y, A, and the set {S1 , ..., Sn } as found in Algorithm 2, determine
the matrix z.
Minimal Clique Covering Reconstruction:
1: for h = 1 to l do
2:
i=1
3:
N = the number of cliques in each Th {n}ote that all Th have the
same number of cliques because they are all minimal
4:
M =N
5:
while i ≤ M do
6:
N=M
7:
F = the set of all vertices (strings) which are in clique i
8:
if length(F ) > 1 then
9:
for j = 2, 3, ..., #{F } do
10:
Use z to find what clique F (j) maps to for each σ ∈ A
11:
if F (j) does not map to the same clique as F (1) upon each
σ then
12:
if M > N and F (j) maps to the same clique as some
F (k) upon each σ, k ∈ [N + 1, M ] then
13:
Add F (i) to the clique containing F (k)
14:
else
15:
Create a new clique containing F (i)
16:
M = M +1
17:
end if
18:
end if
19:
end for
20:
end if
21:
i =i+1
22:
end while
23: end for
24: FinalCovering = min{Ch |Ch has the minimum number of columns ∀h}
return FinalCovering
CSSR
ALGORITHMS
This paper has discussed three different approaches for
determining minimal state hidden Markov models from a
given input sequence: the CSSR algorithm, CSSR as an integer programming problem, and the minimal clique covering
reformulation. We now compare the run times of all three
algorithms. We find that the CSSR algorithm has the fastest
run time, however it does not always produce a minimal state
model as we have seen in a prior example. The minimal clique
covering reformulation is relatively fast and always results in
a minimal state model. The integer programming formulation
is extremely slow, but also always results in a minimal state
model. This makes CSSR a useful heuristic for solving the
NP-hard MSDpFSA problem.
All computational experiments in this section were performed using MATLAB, and the binary integer programming
problems were solved using MATLAB’s bintprog function.
The formulations were each tested using randomly generated
strings in order to observe their behavior on a diverse set
of inputs. The CSSR integer programming formulation, using only a two-character alphabet and a randomly generated
sequence of length ten as input, takes about 100 seconds to
run. We can see in Figure 6 that the minimal clique covering reformulations takes less than 0.2 seconds for randomly
generated two-character sequences up to length 10,000. For
only two character alphabets, the run time appears to be a
nearly linear function of the sequence length, with spikes
occurring presumably in cases where more reconstruction was
needed. In Figure 6 we see the run time of the minimal clique
covering reformulation for a randomly generated string with a
three-character alphabet. For sequences up to length 100, the
algorithm took no more than 120 seconds, which is remarkably
better than the integer programming problem.
For two-, three-, and 4-character alphabets, we also compare
the minimal clique covering formulation time to that of the
original CSSR algorithm. This can be seen in Figure 7 where
we take the average of the trial run times for the CSSR and
clique covering formulation to compare the two. Note that with
a two-character alphabet, the clique covering formulation is
slightly faster, but for the three- and four-character alphabets
the CSSR algorithm is significantly faster. This is due to the
fact that the CSSR algorithm does not actually guarantee a
minimal state Markov model.
VII. C ONCLUSION
AND
F UTURE D IRECTIONS
In this paper we illustrated the problem stated by Shalizi and
Crutchfield [7] of determining a minimal state probabilistic
finite state representation of a data set is NP-hard for finite
data sets. As a by-product, we formally proved this to be the
case for the non-unifilar case as studied in [17]. As such,
this shows that both the CSSR algorithm of [7] and CSSA
algorithm of [17] can be thought of as low-order polynomial
approximations of NP-hard problems.
11
Fig. 6. Clique covering reformulation run times. Each graph shows three
different trial runs of randomly generated strings.
Future work in this area includes studying, in detail, these
approximation algorithms for the problems to determine what
their approximation properties are. A specific area of interest
is placing bounds on the number of states given by a solution
to Equation 7 based on properties of the input string. We plan
to investigate other linear approximation algorithms like those
found in [27], [28].
In addition to this work, determining weaknesses in this approach to modeling behavior are planned. We are particularly
interested in the affects deception can have on models learned
in this way. For example, we are interested in formulating a
problem of constrained optimal deception in which a learner,
using an algorithm like the one described here or the CSSR
or CSSA approximations, is optimally confused by an input
stream that is subject to certain constraints in its incorrectness.
ACKNOWLEDGEMENTS
Portions of Dr. Griffin’s and Ms. Paulson’s work were
supported by the Army Research Office under Grant W911NF11-1-0487.
A PPENDIX
The following formulas can be used to compute fqi |y and
fax|y in Algorithm 1. Let #(x, y) be the number of times the
sequence x is observed as a subsequence of y.
fx|y (a) = Pr(a|x, y) =
P
#(xa, y)
#(x, y)
x∈q
fqi |y (a) = Pr(a|qi , y) = P i
#(xa, y)
x∈qi
#(x, y)
(14)
(15)
Fig. 7. Minimal clique covering run times compared to CSSR
R EFERENCES
[1] C. Shalizi and J. Crutchfield, “Computational mechanics: Pattern and
prediction, structure and simplicity,” J. Statistical Physics, vol. 104, no.
3/4, 2001.
[2] J. P. Crutchfield, “Between order and chaos,” Nature Physics, vol. 8, pp.
17–24, 2012.
[3] L. R. Rabiner, “A tutorial on hidden markov models and selected
applications in speech recognition,” Proceedings of the IEEE, vol. 77,
no. 2, pp. 257–286, 1989.
[4] H. Xue and V. Govindaraju, “Hidden markov models combining discrete
symbols and continuous attributes in handwriting recognition,” IEEE
Trans. Pattern Anal. Machine Intell., vol. 28, no. 3, pp. 458–462, 2006.
[5] J. Hu, M. K. Brown, and W. Turin, “Mm-based online handwriting
recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 10,
pp. 1039–1045, 1996.
[6] S. Lefevre, E. Bouton, T. Brouard, and N. Vincent, “A new way to
use hidden markov models for object tracking in video sequences,” in
Process. III-117-III-120, 2003.
[7] C. Shalizi, K. Shalizi, and J. Crutchfield, “An algorithm for pattern
discovery in time series,” arXiv:cs.LG/0210025 v3, November 2002.
[8] R. R. Brooks, J. Schwier, C. Griffin, and S. Bukkapatnam, “Zero knowledge hidden markov model inference,” Pattern Recognition Letters,
vol. 30, pp. 1273–1280, 2009.
[9] J. Schwier, R. R. Brooks, and C. Griffin, “Methods to window data to
differentiate between markov models,” IEEE Trans. Sys. Man and Cyber.
B, vol. 41, no. 3, pp. 650–663, 2010.
[10] Y. Lu, J. Schwier, R. Craven, R. R. Brooks, and C. Griffin, “Inferring
statistically significant hidden markov models,” IEEE Transactions on
Knowledge and Data Engineering, vol. 25, no. 7, pp. 1548–1558, 2013.
[11] R. R. Brooks, J. Schwier, and C. Griffin, “Behavior detection using
confidence intervals of hidden markov models,” IEEE Transactions on
Systems, Man and Cybernetics, Part B, vol. 39, no. 6, pp. 1484–1492,
December 2009.
[12] H. Bhanu, J. Schwier, R. Craven, I. Ozcelik, C. Griffin, and R. R. Brooks,
“Noise tolerant symbolic learning of markov models of tunneled pro-
12
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
tocols.” Submitted to IWCMC2011 TRaffic Analysis and Classification
(TRAC) Workshop, 2011.
C. Lu, J. Schwier, R. M. Craven, R. R. Brooks, and C. Griffin., “A
normalized statistical metric space for markov models,” IEEE Trans.
Sys. Man and Cyber. B, vol. 43, no. 3, pp. 806–819, 2013.
A. Stolcke and S. Omohundro, “Hidden markov model induction by
bayesian model merging,” Advances in Neural Information Processing
Systems, vol. 5, pp. 11–18, 1993.
J. P. Crutchfield and K. Young, “Inferring statistical complexity,” Physical Review Letters, vol. 63, pp. 105–108, 1989.
C. Strelioff and J. Crutchfield, “Bayesian structural inference for hidden
processes,” Physical Review E, vol. 89, p. 042119, 2014.
M. Schmiedekamp, A. Subbu, and S. Phoha, “The clustered causal
state algorithm: Efficient pattern discovery for lossy data-compression
applications,” Comput. Sci. Eng., vol. 8, no. 5, 2006.
H. G. Tucker, “A generalization of the glivenko-cantelli theorem,”
Ann. Math. Statist., vol. 30, no. 3, pp. 828–830, 09 1959. [Online].
Available: http://dx.doi.org/10.1214/aoms/1177706212
J. T. N. Abe and M. Warmuth, “Polynomial learnability of stochastic
rules with respect to the kl-divergence and quadratic distance,” IEICE
Transactions on Information and Systems, vol. E84-D, no. 3, pp. 299–
316, 2001.
L. Pitt and M. Warmuth, “The minimum consistent dfa problem cannot
be approximated by any polynomial,” Journal of the Association for
Computing Machinery, vol. 40, no. 1, pp. 95–142, 1993.
E. Gold, “Complexity of automation identification from given data,”
Information and Control, vol. 37, pp. 302–320, 1978.
M. Kearns and L. Valiant, “Cryptographic limitations on learning
boolean formulae and finite automata,” Journal of the ACM, vol. 41,
pp. 67–95, 1994.
C. R. Shalizi and K. L. Shalizi, “Blind construction of optimal nonlinear
recursive predictors for discrete sequences,” in Proc. ACM International
Conference on Uncertainty in Artificial Intelligence, 2004.
C. Cluskey, “A linear integer program for the causal state splitting and
reconstruction algorithm,” Honor’s Thesis, Schreyer Honors College,
Penn State Univeristy, Spring, 2013.
N. Karmarkar, “A new polynomial-time algorithm for linear programming,” in Proceedings of the Sixteenth Annual ACM Symposium on
Theory of Computing, New York, NY, USA, 1984, pp. 302–311.
C. Bron and J. Kerbosch, “Algorithm 457: finding all cliques of an
undirected graph,” Comm. ACM, vol. 16, no. 9, pp. 575–577, 1973.
R. C. Carrasco and J. Oncina, “Learning stochastic regular grammars by
means of a state merging method.” Springer-Verlag, 1994, pp. 139–152.
C. de la Higuera, Grammatical Inference: Learning Automata and
Grammars. New York, NY, USA: Cambridge University Press, 2010.
| 8cs.DS
|
1
Cross-Mode Interference Characterization in
Cellular Networks with Voronoi Guard Regions
arXiv:1702.05018v3 [cs.IT] 20 Oct 2017
Stelios Stefanatos, Antonis G. Gotsis, and Angeliki Alexiou
Abstract—Advances in cellular networks such as device-todevice communications and full-duplex radios, as well as the inherent elimination of intra-cell interference achieved by networkcontrolled multiple access schemes, motivates the investigation
of the cross-mode interference properties under a guard region
corresponding to the Voronoi cell of an access point (AP). By
modeling the positions of interfering APs and user equipments
(UEs) as Poisson distributed, analytical expressions for the
statistics of the cross-mode interference generated by either APs
or UEs are obtained based on appropriately defined density
functions. The considered system model and analysis are general
enough to capture many operational scenarios of practical interest, including conventional downlink/uplink transmissions with
nearest AP association, as well as transmissions where not both
communicating nodes lie within the same cell. Analysis provides
insights on the level of protection offered by a Voronoi guard
region and its dependence on type of interference and receiver
position. Numerical examples demonstrate the validity/accuracy
of the analysis in obtaining the system coverage probability for
operational scenarios of practical interest.
Index Terms—Stochastic geometry, cellular networks, guard
region, D2D communications, full-duplex radios, interference.
I. I NTRODUCTION
Characterization of the interference experienced by the
receivers of a wireless network is of critical importance for
system analysis and design [1]. This is especially the case
for the future cellular network, whose envisioned fundamental
changes in its architecture, technology, and operation will have
significant impact on the interference footprint [2]. Interference characterization under these new system features is of
the utmost importance in order to understand their potential
merits as well as their ability to co-exist.
Towards increasing the spatial frequency reuse, two of the
most prominent techniques/features considered for the future
cellular network are device-to-device (D2D) communications
[3] and full-duplex (FD) radios [4]. Although promising,
application of these techniques introduces additional, crossmode interference. For example, an uplink transmission is
no longer affected only by interfering uplink transmissions
but also by interfering D2D and/or downlink transmissions.
Although it is reasonable to expect that the current practice of
eliminating intra-cell interference by employing coordinated
transmissions per cell will also hold in the future [5], the
continually increasing density of APs and user equipments
(UEs) suggest that inter-cell interference will be the major
limiting factor of D2D- and/or FD-enabled cellular networks,
rendering its statistical characterization critical.
The authors are with the Department of Digital Systems, University of
Piraeus, Piraeus, Greece. Email: {sstefanatos, agotsis, alexiou}@unipi.gr.
A. Previous Work
Stochastic geometry is by now a well accepted framework
for analytically modeling interference in large-scale wireless
networks [6]. Under this framework, most of the numerous
works on D2D-enabled cellular networks (without FD radios)
consider the interfering D2D nodes as uniformly distributed
on the plane, i.e., there is no spatial coordination of D2D
transmissions (see, e.g., [7], [8], [9], [10]). Building on the
approach of [11], various works consider the benefits of
employing spatially coordinated D2D transmissions where, for
each D2D link in the system, a circular guard region (zone)
is established, centered at either the receiver or transmitter,
within which no interfering transmissions are performed [12],
[13]. However, when the D2D links are network controlled
[5], a more natural and easier to establish guard region is the
(Voronoi) cell of a coordinating AP. Under a non-regular AP
deployment [14], this approach results in a random polygon
guard region, which makes the interference characterization a
much more challenging task.
Interference characterization for this type of guard region
has only been partially investigated in [15], [16], [17] and only
for the case of conventional uplink transmissions with nearest
AP association and one active UE per cell (with no crossmode interference). As the positions of the interfering UEs
are distributed as a Voronoi perturbed lattice process (VPLP)
in this case [18], [19], which is analytically intractable, an
approximation based on a Poisson point process (PPP) model
with a heuristically proposed equivalent density is employed.
This approach of approximating a complicated system model
by a simpler one with appropriate parameters (in this case,
by a PPP of a given density) was also used in [20] for the
characterization of downlink systems (with no cross-mode
interference as well). Reference [19] provides a rigorous
characterization of the equivalent density of the UE-generated
interference both from the “viewpoint” of an AP as well as
its associated UE, with the latter case of interest in case
of cross-mode interference. The analysis reveals significant
differences in the densities corresponding to these two cases
suggesting that the equivalent density is strongly dependent on
the considered receiver position. Interference characterization
for the case of arbitrary receiver position that may potentially
lie outside the Voronoi guard region as, e.g., in the case of
cross-cell D2D links [21], has not been investigated in the
literature, let alone under cross-mode interference conditions.
Investigation of interference statistics under a cell guard
region is also missing in the (much smaller) literature on FDenabled cellular networks, which typically assumes no spatial
2
coordination for the UE transmissions (see e.g., [22], [23],
[24]).
B. Contributions
This paper considers a stochastic geometry framework for
modeling the cross-mode interference power experienced at
an arbitrary receiver position due to transmissions by APs
or UEs and under the spatial protection of a Voronoi guard
region. Modeling the positions of interfererers (APs or UEs)
as a Poisson point process, the statistical characterization of
the cross-mode interference is pursued via computation of its
Laplace transform. The main contributions of the paper are
the following.
• Consideration of a general system model, which allows
for a unified analysis analysis of cross-mode interference
statistics. By appropriate choice of the system model
parameters, the interference characterization is applicable
to multiple operational scenarios, including conventional
downlink/uplink communications with nearest AP association and (cross-cell) D2D links where the transmitterreceiver pair does not necessarily lie in a single cell.
• Exact statistical characterization of AP-generated crossmode interference, applicable, e.g., in the case where
interference experienced at an AP is due to transmissions
by other FD-enabled APs. An equivalent interferer density is given in a simple closed form, allowing for an
intuitive understanding of the interfernece properties and
its dependence on the position of the receiver relative
to the position of the AP establishing the Voronoi guard
region.
• Determination of a lower bound of the Laplace transform in case of UE-generated cross-mode interference,
applicable, e.g., in the case where a UE experiences
interference due to other FD-enabled or D2D-operating
UEs. The properties of the corresponding equivalent
density of interferers is studied in detail, providing insights for various operational scenarios, including the
standard uplink communication scenario (with no crossmode interfernece) where a rigorous justification of why
the heuristic approaches previously proposed in [16], [17]
are accurate.
Simulated examples indicate the accuracy of the analytical
results also for cases where the positions of interferering UEs
are VPLP distributed, suggesting their use as a basis for
determination of performance as well as design of optimal
resource allocation algorithms for future, D2D- and/or FDenabled cellular networks.
C. Notation
The origin of the two-dimensional plane R2 will be denoted
as o. The Euclidean norm of x ∈ R2 is denoted as |x| with
operator | · | also used to denote the absolute value of a scalar
or the Lebesgue measure (area) of a bounded subset of R2 .
The polar form representation of x ∈ R2 will be denoted as
(|x|, ∠x) or |x|∠x, where ∠x is the principal branch of the
angular coordinate of x taking values in [−π, π). The open ball
in R2 , centered at x ∈ R2 and of radius R > 0, is denoted
as B(x, R) , {y ∈ R2 : |y − x| < R}, whereas its boundary
is denoted as C(x, R) , {y ∈ R2 : |y − x| = R}. I(·) is the
indicator (0 − 1) operator, P(·) is the probability measure,
and E(·) is the expectation operator. Functions arccos(·) :
[−1, 1] → [0, π] and arcsin(·) : [−1, 1] → [−π/2, π/2]
are the principal branches of the inverse cosine and sine,
respectively. The Laplace transform of a random variable z
equals Lz (s) , E(e−sz ), for all s ∈ R for which the
expectation exists.
II. S YSTEM M ODEL
A large-scale model of a cellular network with APs and
UEs positioned over R2 is considered. The positions of APs
are modeled as a realization of a homogeneous PPP (HPPP)
Φa ⊂ R2 of density λa > 0. All the communication scenarios
considered in this paper involve three nodes, namely,
• a, so called, typical node, which, without loss of generality (w.l.o.g.), will be assumed that its position coincides
with the origin o. The typical node is an AP if o ∈ Φa
or a UE, otherwise;
∗
• the closest AP to the typical node, located at x
,
arg minx∈Φa |x| (x∗ = o in case o ∈ Φa , i.e., the typical
node is an AP);
• a (typical) receiver located at an arbitrary position xR ∈
R2 . Typical node is an AP if xR ∈ Φa or a UE, otherwise.
The case xR = o is also allowed, meaning that the typical
node is also the typical receiver.
Let V ∗ denote the Voronoi cell of the AP at x∗ , generated by
the Poisson-Voronoi tessellation of the plane from Φa , i.e.,
V ∗ , y ∈ R2 : |y − x∗ | < |y − x|, for all x ∈ Φa \ {x∗ } .
The goal of this paper is to characterize the interference power
experienced at xR due to transmissions from nodes (APs or
UEs) that lie outside V ∗ , i.e., with V ∗ effectively forming a
spatial guard region within which no interference is generated.
Let Φu ⊂ R2 denote the point process representing the
positions of the UEs in the system other than the typical
node and typical receiver, and IxR ,a and IxR ,u denote the
interference power experienced at xR due to transmissions
by all APs and UEs in the system, respectively. Under the
Voronoi guard region scheme described above, the standard
interference power model is adopted in this paper, namely [1],
X
IxR ,k ,
Pk gx |x − xR |−αk , k ∈ {a, u},
(1)
x∈Φk \V ∗
where gx ≥ 0 is the channel gain of a transmission generated
by a node at x ∈ R2 , assumed to be an independent of x,
exponentially distributed random variable of mean 1 (Rayleigh
fading), αk > 2 is the path loss exponent and Pk > 0 is
the transmit power, which, for simplicity, is assumed fixed
and same for all nodes of the same type. Figure 1 shows an
example realization of V ∗ and the positions of the interferers
for the case xR 6= x∗ 6= o. Note that V ∗ is a random guard
region as it depends positions of the APs, therefore, there is no
guarantee that xR lies within V ∗ , i.e., it holds P(xR ∈ V ∗ ) <
3
TABLE I
S PECIAL C ASES OF S YSTEM M ODEL
V∗
|x∗ |
condition
xR 6= x∗ ,
x∗ 6= o
xR = o,
x∗ 6= o
xR = x∗ ,
x∗ 6= o
Fig. 1. Random realization of the system model. APs and UEs are shown as
triangle and circle markers, respectively. The typical node in this example
is also a UE shown as a square. The shaded polygon area indicates the
Voronoi guard region V ∗ within which no node (indicated by open marker)
transmits. All nodes outside V ∗ (filled markers) generate interference that is
experienced by the receiver of interest whose position (xR ) is indicated by
an open diamond marker. This scenario may correspond to a (cross-cell) D2D
link between typical node and receiver or to the receiver acting as relay aiding
the donwlink or uplink communication of the typical node with its nearest
AP.
1, unless xR is a point on the line segment joining the origin
with x∗ (inclusive).
The above model is general enough so that, by appropriately
choosing xR and/or x∗ , IxR ,a and IxR ,u correspond to many
practical instances of (cross-mode) interference experienced
in D2D/FD-enabled as well as conventional cellular networks.
For example, for the case where x∗ 6= xR 6= o, Eq. (1) may
correspond to the AP- and UE-generated interference power
experienced by the receiver of a, potentially cross-cell, D2D
link between the typical nodes at o and xR . Various other
scenarios of practical interest are captured by the model, some
of which are described in Table I. Note that from the scenarios
identified in Table I, only the standard downlink and uplink
scenarios have been considered previously in the literature
[15], [16], [17], [25], however, without consideration of crossmode interference, i.e., UE-generated and AP-generated interference for downlink and uplink transmissions, respectively.
Reference [19] considers the UE-generated cross-mode interference experienced at the typical node in downlink mode,
however, the results provided cannot not be straightforwardly
generalized to the general case.
For characterization of the performance of the communication link as well as obtaining insights on the effectiveness of
the guard region V ∗ for reducing interference, it is of interest to
describe the statistical properties of the random variables IxR ,a
and IxR ,u . This is the topic of the following section, where
the marginal statistics of IxR ,a and IxR ,u conditioned on x∗
and treating xR as a given, but otherwise free, parameter, are
investigated in detail. Characterization of the joint distribution
of IxR ,a and IxR ,u is left for future work.
Note that the respective unconditional interference statistics
can be simply obtained by averaging over the distribution
of x∗ , which corresponds to a uniformly distributed
∠x∗
√
∗
and a Rayleigh distributed |x | of mean 1/(2 λa ) [27].
However, results conditioned on x∗ are also of importance
xR =
6 o,
x∗ = o
xR = x∗ ,
x∗ = o
scenario
General case. Corresponds to (a) a D2D link when the
typical node sends data to the typical receiver or (b) to a
relay-assisted cellular communication where the typical
receiver acts a relay for the link between the typical
node and it nearest AP (either downlink or uplink).
Typical receiver is not guaranteed to lie within V ∗ .
Typical node is a UE in receive mode and lies within
V ∗ . Represents a standard downlink when the nearest
AP is transmitting data to the typical node/receiver.
Nearest AP to the typical node is in receive mode with
the typical node lying in V ∗ . Represents a standard
uplink when the AP receives data from the typical node.
Typical node is an AP. When the typical node/AP is the
one transmitting to the typical receiver, a non-standard
downlink is established as xR is not necessarily lying
within V ∗ .
Typical node is an AP in receive mode; the position of
the corresponding transmitter is unspecified and may as
well lie outside V ∗ , thus modeling a non-standard uplink.
on their own as they can serve as the mathematical basis
for (optimal) resource allocation algorithms given network
topology information, e.g., decide on whether the typical node
employs D2D or cellular mode given knowledge of x∗ .
For tractability purposes, the following assumption on the
statistics of Φu will be considered throughout the analysis.
Assumption. The positions of the (potentially) interfering
UEs, Φu , is a realization of an HPPP of density λu > 0,
that is independent of the positions of the APs, typical node,
and typical receiver.
Remark: In general, this assumption is not exact since
resource allocation and scheduling decisions over the entire
network affect the distribution of interfering UE positions.
For example, in the conventional uplink scenario with at least
one UE per cell, the transmitting UE positions correspond
to a VPLP process of density equal to λu = λa [17],
[19]. However, as also shown in [17], [19] for the standard
uplink scenario, the HPPP assumption of the UE point process
allows for a tractable, yet accurate approximation of the actual
performance, which, as will be demonstrated in Sec. V, is also
the case for other operational scenarios as well. For generality
purposes, an arbitrary value λu 6= λa is also allowed in
the analysis, which is actually the case, e.g., when one UE
is only allowed to transmit per cell but there exist empty
cells and, therefore, λu < λa (ultra dense networks [26]),
and also serves as an approximate model for the case when
some arbitrary/undefined coordination scheme is employed by
other cells in the system (if at all), that may as well result in
λu > λa .
III. I NTERFERENCE C HARACTERIZATION
Towards obtaining a tractable characterization of the distribution of the (AP or UE) interference power, or, equivalently,
its Laplace transform, the following standard result in PPP
theory is first recalled [27].
P
Lemma 1. Let I , x∈Φ̃ P hx |x − z|−α , z ∈ R2 , P > 0,
with Φ̃ an inhomogeneous PPP of density λ : R2 → [0, ∞),
4
and {hx }x∈Φ i.i.d. exponential random variables of mean 1.
The Laplace transform of I equals
Z
−α
LI (s) = exp −
λ(x)γ(sP |x − z| )dx
(2)
R2
Z ∞
= exp −2π
λz (r)rγ(sP r−α )dr ,
(3)
0
1
1+t ,
and the second equality holds only
where γ(t) , 1 −
in the case of a circularly-symmetric density function w.r.t.
point z, i.e., λ(z + x) = λz (|x|), for an appropriately defined
function λz : [0, ∞) → [0, ∞).
The above result provides a complete statistical characterization in integral form of the interference experienced at
a position z ∈ R2 due to Poisson distributed interferers.
Of particular interest is the case of a circularly-symmetric
density, which only requires a single integration over the radial
coordinate and has previously led to tractable analysis for
various wireless network models of interest such as mobile ad
hoc [1] and downlink cellular [25]. In the following it will
be shown that even though the interference model of Sec.
II suggests a non circularly-symmetric interference density
due to the random shape of the guard region, the Laplace
transform formula of (3) holds exact for IxR ,a and is a (tight)
lower bound for IxR ,u with appropriately defined circularlysymmetric equivalent density functions.
A. AP-Generated Interference
Note that considering the Voronoi cell of a random AP at
x ∈ Φa as a guard region has no effect on the distribution of
the interfering APs. This is because the Voronoi cell of any
AP is a (deterministic) function of Φa , therefore, it does not
impose any constraints on the realization of Φa , apart from
implying that the AP at x does not produce interference. By
well known properties of HPPPs [27], conditioning on the
position x of the non-interfering AP, has no effect on the
distribution of the interfering APs. That is, the interfering AP
positions are still distributed as an HPPP of density λa as in
the case without imposing any guard region whatsoever.
However, when it is the Voronoi cell of the nearest AP
to the originthat is considered as a guard region, an implicit
guard region is formed. Indeed, the interfering APs, given the
AP at x∗ is not interfering, are effectively distributed as an
inhomogeneous PPP with density [25]
λ̃a (x) = λa I(x ∈
/ B(o, |x∗ |)), x ∈ R2 ,
(4)
i.e., an implicit circular guard zone is introduced around the
typical node (see Fig. 1), since, if this was not the case,
another AP could be positioned at a distance smaller than
|x∗ | from the origin, which is impossible by assumption. This
observation may be used to obtain the Laplace transform of
the AP-generated interference experienced at xR directly from
the formula of (2). However, the following result shows that
the two-dimensional integration can be avoided as the formula
of (3) is also valid in this case with an appropriately defined
equivalent density function.
Proposition 2. The Laplace transform, LIxR ,a (s | x∗ ) ,
E(e−sIxR ,a | x∗ ), of the AP-generated interference power
IxR ,a , conditioned on x∗ , equals the right-hand side of (3)
with P = Pa , α = αa and λz (r) = λxR ,a (r), r ≥ 0, where
λa
, r > |x∗ | + |xR |
(
, |x∗ | > |xR |,
0
, r ≤ ||x∗ | − |xR ||
λxR ,a (r) ,
λa , |x∗ | < |xR |,
λa /2
, |x∗ | = |xR |
λa 1 − π1 arccos (d) , otherwise,
(5)
r 2 −(|x∗ |2 −|xR |2 )
with d ,
, for xR 6= o, and
2r|xR |
λxR ,a (r) = λa I(r ≥ |x∗ |),
(6)
for xR = o.
Proof: See Appendix A.
The following remarks can be made:
1) The interference power experienced at an arbitrary position xR 6= o under the considered guard region
scheme is equal in distribution to the interference power
experienced at the origin without any guard region and
with interferers distributed as an inhomogeneous PPP of
an equivalent, xR -depednent density given by (5).
2) Even though derivation of the statistics of IxR ,a was
conditioned on xR and x∗ , the resulting equivalent inteferer density depends only on their norms |xR | and |x∗ |.
Although the indepedence from ∠xR might have been
expected due to the isotropic property of Φa [1], there is
no obvious reason why one would expect independence
also from ∠xR .
3) λxR ,a (r) is a decreasing function of |x∗ |, corresponding
to a (statistically) smaller AP interference power due to
an increased guard zone area.
4) For xR = o, corresponding to a standard downlink with
interference generated from other donwlink transmissions, Prop. 2 coincides with the analysis of [25], as
expected.
5) The Laplace transform of the interference in the case
of xR 6= x∗ 6= o was examined previously in [28].
However, the corresponding formulas appear as twodimensional integrals that offer limited insights compared to the simpler and more intuitive equivalent density formulation given in Prop. 2.
6) The case xR = x∗ 6= o corresponds to the nearestneighbor transmission scenario considered in [29] where
the validity of λxR ,a (r) in (5) as an equivalent density
function for computation of the Laplace transform of the
interference power was not observed.
7) In [30], the Laplace transform of the interference power
experienced at xR ∈ R2 due to a Poisson hole process,
i.e., with interferers distributed as an HPPP over R2
except in the area covered by randomly positioned disks
(holes), was considered. The holes were assumed to not
include xR and a lower bound for the Laplace transform
was obtained by considering only a single hole [30,
Lemma 5], which coincides with the result of Prop.
1
1
0.8
0.8
0.6
λx∗ ,a (r)/λa
λxR ,a (r)/λa
5
0.4
√
|xR | × 2 λa
√
|xR | × 2 λa
√
|xR | × 2 λa
√
|xR | × 2 λa
0.2
0
0
0.5
1
1.5
2
2.5
√
r × 2 λa
3
= 1/2
=1
= 3/2
=2
3.5
0.6
0.4
√
|x∗ | × 2 λa
√
∗
|x | × 2 λa
√
∗
|x | × 2 λa
√
|x∗ | × 2 λa
0.2
4
Fig. 2. Equivalent radial density of interfering
APs experienced at various
√
xR ∈ R2 , conditioned on |x∗ | = 1/(2 λa ).
2. This is not surprising as the positions of the APs,
conditioned on x∗ , are essentially distributed a singlehole Poisson process. Note that Prop. 2 generalizes [30,
Lemma 5] by allowing xR to be covered by the hole
and considers a different proof methodology.
Figure 2 shows the normalized density function λxR ,a (r)/λ
√ a
for various values of |xR | and assuming that |x∗ | = 1/(2 λa ),
the expected distance from the nearest AP. It can be seen that
the presence of the implicit circular guard region results in
reducing the equivalent interferer density in certain intervals
of the radial coordinate r, depending on the value of |xR |.
In particular, when |xR | < |x∗ |, it is guaranteed that no APs
exist within a radius |x∗ | − |xR | > 0 from xR . In contrast,
when |xR | > |x∗ | there is no protection from APs in the close
vicinity of xR .
The case |xR | = |x∗ | is particularly interesting since it
corresponds to the case when the receiver of interest is the
AP at x∗ , experiencing interference from other AP, e.g., when
operating in FD. For x∗ 6= o, corresponding to an uplink
transmission by the typical node to its nearest AP, it can be
easily shown that it holds
1
r
λx∗ ,a (r) = λa
+
+ O(r3 ), r → 0,
(7)
2 2π|x∗ |
i.e., the guard region results in the serving AP experiencing
about half of the total interfering APs density in its close vicinity, which is intuitive as for asymptotically small distances
from x∗ the boundary of the circular guard region can be
locally approximated as a line that divides the plane in two
halves, one with interferer density λa and one with interferer
density 0. Figure 3 shows the normalized λx∗ ,a (r) for various
values of |x∗ |, where its linear asymptotic behavior as well as
the advantage of a larger |x∗ | are clearly visible.
B. UE-Generated Interference
The positions of interfering UEs, conditioned on the realization of Φa , are distributed as an inhomogeneous PPP of
0
0
0.5
1
1.5
2
2.5
√
r × 2 λa
3
= 1/2
=1
= 3/2
=2
3.5
4
Fig. 3. Equivalent radial density of interfering APs experienced at x∗ .
density
λ̃u (x|Φa ) = λu I(x ∈
/ V ∗ ), x ∈ R2 .
(8)
Although the expression for λ̃u is very similar to the that of
the density λ̃a of interfering APs given in (4), the random
Voronoi guard region V ∗ appearing in (8) is significantly
more complicated than the deterministic circular guard zone
B(o, |x∗ |), which renders the analysis more challenging. To
simplify the following exposition, it will be assumed that a
rotation of the Cartesian coordinate axes is performed such
that ∠x∗ = 0. Note that this rotation has no effect in the
analysis due to the isotropic property of the HPPP [1] and
immediately renders the following results independent of the
value of ∠x∗ in the original coordinate system.
Since it is of interest to examine the UE interference
statistics conditioned only on x∗ , a natural quantity to consider,
that will be also of importance in the interference statistics
analysis, is the probabilistic cell area (PCA) function pc (x |
x∗ ), x ∈ R2 . This function gives the probability that a point
x ∈ R2 lies within V ∗ conditioned only on x∗ , i.e.,1
pc (x | x∗ ) , P(x ∈ V ∗ | x∗ ).
Lemma 3. For all x ∈ R2 , the PCA function equals
(
1
, |x| ≤ |x∗ |, ∠x = 0,
∗
pc (x | x ) =
e−λa |A| < 1 , otherwise,
where A , B(x, |x − x∗ |) \ B(o, |x∗ |). For all x 6= x∗ , it holds
(
r2 (|∠x|+θ∗ )−|x∗ |2 |∠x|+|x||x∗ | sin(|∠x|) , x∗ 6= o,
|A| = ∗ 2
π|x|
, x∗ = o,
p
with r∗ , |x − x∗ | = |x|2 + |x∗ |2 − 2|x||x∗ | cos(∠x) and
π − arcsin |x| sin(|∠x|)
, |x| cos(∠x) > |x∗ |,
r
∗
θ∗ ,
arcsin |x| sin(|∠x|)
, |x| cos(∠x) ≤ |x∗ |.
r∗
1 Recall that, under the system model, x∗ is also the AP closest to the
origin.
6
2.4
2.2
2
λa E(|V ∗ | | x∗ )
0.5
0.9
0.1
|x∗ |
1.8
1.6
1.4
1.2
Fig. 4. Contours of pc (x | x∗ ) (solid lines),
corresponding to probabilities
√
0.1 to 0.9 in steps of 0.1, for |x∗ | = 1/ λa . The points of the line segment
∗
joining o to x are the only ones with pc (x|x∗ ) = 1. The corresponding
∗ 2
contours of e−λa π|x−x | are also shown (dashed lines).
Proof: The probability that the point x ∈ R2 belongs
to V is equal to the probability that there does not exist a
point of Φa \ {x∗ } within the set A, which equals e−λa π|A|
[27]. For |x| ≤ |x∗ | and ∠x = 0, it is a simple geometrical
observation that A = Ø, therefore, |A| = 0. When x∗ = o,
A = B(x, |x|) \ B(o, 0) = B(x, |x|), and, therefore, |A| =
π|x|2 . For all other cases, |A| can be computed by the same
approach as in the proof of [31, Theorem 1]. The procedure
is straightforward but tedious and is omitted.
A simple lower bound for pc (x | x∗ ) directly follows by
noting that A ⊆ B(x, |x − x∗ |) for all x ∈ R2 .
1
analysis
simulation
0
0.5
1
1.5
√
|x∗ | × 2 λa
2
2.5
3
Fig. 5. Average area of V ∗ conditioned on x∗ as a function of the
(normalized) distance of x∗ from the origin.
∗
Corollary 4. The PCA function is lower bounded as
pc (x | x∗ ) ≥ e−λa π|x−x
with equality if and only if x∗ = o.
∗ 2
|
, x ∈ R2 ,
(9)
Remark: The right-hand side of (9) equals the probability
that x belongs to the Voronoi cell of the AP positioned at x∗
when the latter is not conditioned on being the closest AP to
the origin or any other point in R2 .
Figure
4 depicts pc (x | x∗ ) for the case where |x∗ | =
√
1/ λa (behavior is similar for other values of |x∗ | > 0). Note
that, unless x∗ = o, pc (x | x∗ ) is not circularly symmetric
w.r.t. any point in R2 , with its form suggesting that points
isotropically distributed in the vicinity of x∗ are more probable
to lie within V ∗ than points isotropically distributed in the
vicinity of o.
∗ 2
The lower bound e−λa π|x−x | , x ∈ R2 , is also shown in
Fig. 4, clearly indicating the probabilistic expansion effect of
the Voronoi cell of an AP, when the latter is conditioned on
being the closest to the origin. This cell expansion effect is
also demonstrated in Fig. 5 where the conditional average cell
area, equal to
Z
∗
∗
∗
∗
E(|V | | x ) = E
I(x ∈ V )dx x
R2
Z
=
pc (x|x∗ )dx,
(10)
R2
is plotted as a function of the distance |x∗ |, with the integral of
(10) evaluated numerically using Lemma 3. Simulation results
are also depicted serving as a verification of the validity of
Lemma 3. It can be seen that the average area of the guard
region increases with |x∗ |, which implies a corresponding
increase of the average number of UEs that lie within the
region. However, as will be discussed in the following, even
though resulting in more UEs silenced on average, an increasing |x∗ | does not necessarily imply improved protection from
UE interference, depending on the receiver position.
The interference statistics of the UE-generated interference
power are given in the following result.
Proposition 5. The Laplace transform LIxR ,u (s | x∗ ) ,
E(e−sIxR ,u | x∗ ) of the UE-generated interference power
IxR ,u , conditioned on x∗ , is lower bounded as the right-hand
side of (3) with P = Pu , α = αu and λz (r) = λxR ,u (r), r ≥
0, where
Z 2π
1
∗
λxR ,u (r) , λu 1 −
pc ((r, θ) + xR | x ) dθ ,
2π 0
(11)
with pc (x | x∗ ) as given in Lemma 3.
Proof: See Appendix B.
The following remarks can be made:
1) The statistics of the interfering UE point process when
averaged over V ∗ (equivalently, over Φa \ {x∗ }) do not
correspond to a PPP, even though this is the case when a
fixed realization of V ∗ is considered. Therefore, it cannot
be expected that Lemma 1 applies for LIxR ,u (s | x∗ ).
However, as Prop. 5 shows, when the interfering point
process is treated in the analysis as an PPP of an appropriately defined equivalent density function λxR ,u (r), a
tractable lower bound for LIxR ,u (s | x∗ ) is obtained,
which will be shown to be tight in the numerical results
section.
2) For x∗ 6= o, the non circularly-symmetric form of the
PCA function results in an integral-form expression for
the equivalent density λxR ,u (r) that depends in general
on ∠xR , in contrast to the case for λxR ,a (r) (see Sec.
7
III.A, Remark 2). A closed form expression for the
equivalent density is only available for x∗ = o discussed
below.
3) When the receiver position is specified as |xR |∠xR with
∠xR uniformly distributed over [−π, π), a straightforward extension of the proof of Prop. 5 results in the
same lower bound for the conditioned Laplace transform
of IxR ,u as in Prop. 5, with λxR ,u (r) replaced by
Z π
1
λx ,u (r)d∠xR ,
(12)
λ|xR |,u (r) ,
2π −π R
which is independent of ∠xR .
Although λxR ,u (r) is not available in closed form when x∗ 6=
o, the numerical integration over a circular contour required
in (11) is very simple. With λxR ,u (r), r ≥ 0, pre-computed,
evaluation of the lower bound for LIxR ,u (s | x∗ ) is of the same
(small) numerical complexity as the evaluation of LIxR ,a (s |
x∗ ). Moreover, a closed-form upper bound for λxR ,u (r) is
available, which can in turn be used to obtain a looser lower
bound for LIxR ,u (s | x∗ ) that is independent of ∠xR . The
bound is tight in the sense that it corresponds to the exact
equivalent density when x∗ = o.
Proposition 6. The equivalent density λxR ,u (r) of Prop. 5 is
upper bounded as
h
i
∗ 2
2
2
λxR ,u (r) ≤ λu 1 − e−λa π(|x | +|xR | +r ) I0 (λa 2πx̄r) ,
(13)
where x̄ , max(|x∗ |, |xR |) and I0 (·) denotes the zero-order
modified Bessel function of the first kind. Equality holds only
when x∗ = o.
Proof: See Appendix C.
An immediate observation of the above result is that, given
xR , the equivalent interfering UE density decreases for any
x∗ 6= o compared to the case x∗ = o. This behavior is expected
due to the cell expansion effect described before. However, as
the bound of (13) is independent on ∠xR , it does not offer
clear information on how the interferering UE density changes
when different values of x∗ 6= o are considered for a given
xR . In order to obtain futher insights on the UE-generated
interference properties, λxR ,u (r) is investigated in detail in the
following section for certain special instances of xR and/or x∗ ,
which are of particular interest in cellular networks.
IV. A NALYSIS OF S PECIAL C ASES OF UE-G ENERATED
I NTERFERENCE
A. xR = x∗
This case corresponds to an standard uplink cellular transmission with nearest AP association, no intra-cell interference,
and interference generated by UEs outside V ∗ operating in
uplink and/or D2D mode. This case was previously investigated in [16], [17], with heuristically introduced equivalent
densities considering only the average performance over x∗ .
In this section, a more detailed investigation of the equivalent
density properties is provided.
For this case, a tighter upper bound than the one in Prop. 6
is available for λx∗ ,u (r), which, interestingly, is independent
of |x∗ |, in the practical case when x∗ 6= o.
Lemma 7. For xR = x∗ , the equivalent density λxR ,u (r) =
λx∗ ,u (r) of Prop. 5 is upper bounded as
2
λx∗ ,u (r) ≤ λu 1 − e−λa πr ,
(14)
with equality if and only if x∗ = o.
Proof:
Follows
by
replacing
the
term
pc ((r, θ) + xR | x∗ ) = pc ((r, θ) + x∗ | x∗ ) appearing in
(11) with its bound given in Cor. 4, which evaluates to
2
e−λa πr .
The bound of (14) indicates that λx∗ ,u (r) tends to zero
at least as fast as O(r2 ) for r → 0, irrespective of x∗ .
The following exact statement on the asymptotic behavior of
λx∗ ,u (r) shows that λx∗ ,u (r) ∼ cr2 , r → 0, with the value of
c independent of |x∗ | when x∗ 6= o.
Proposition 8. For xR = x∗ , the equivalent density
λxR ,u (r) = λx∗ ,u (r) of Prop. 5 equals
λx∗ ,u (r) = λu λa bπr2 + O(r3 ), r → 0,
(15)
with b = 1 for x∗ = o and b = 1/2 for x∗ 6= o.
Proof: For x∗ = o, the result follows directly from
Lemma 7. For x∗ 6= o, it can be shown by algebraic manipulation based on Lemma 3, that the term pc ((r, θ) + xR | x∗ ) =
pc ((r, θ) + x∗ | x∗ ) appearing in (11) equals
λu λa (qπ + (−1)q arccos(sin(θ)) + cos(θ) sin(θ)) r2 +O(r3 ),
for r → 0, where q = 0 for θ ∈ [−π, −π/2] ∪ [π/2, π) and
q = 1 for θ ∈ (−π/2, π/2). Substituting this expression in
(11) and performing the integration leads to the result.
Equation (15) indicates that imposing a guard region V ∗
when x∗ 6= o reduces the interfering UE density by half in
the close vicinity of x∗ , compared to the case x∗ = o. This
effect is similar to the behavior of λx∗ ,a (r) discussed in Sec.
III. A. However, λx∗ ,a (0) = λa /2 , whereas λx∗ ,u (0) = 0,
clearly demonstrating the effectiveness of the guard region for
reducing UE-generated interference in the vicinity of x∗ .
For non-asymptotic values of r, the behavior of λx∗ ,u (r)
can only be examined by numerical evaluation of (11). Figure 6 depicts the normalized equivalent density λx∗ ,u (r)/λu
for various values
of |x∗ | in the range of practical
in√
√
∗
terest [0, 3/(2
λ
)]
(note
that
P(|x
|
>
3/(2
λ
))
=
a
a
√
2
e−λa π(3/(2 λa )) ≈ 8.5 × 10−4 ). It can be seen that for nonasymptotic values of r, λx∗ ,u (r) is a decreasing function of
|x∗ | for all r > 0, corresponding to a reduced UE-generated
interference. This is expected since, as evident from Fig. 4, the
cell expansion effect with increasing |x∗ | can only improve the
interference protection at x∗ .
Interestingly, the dependence of λx∗ ,u (r) on |x∗ |, although
existing, can be seen to be rather small. This observation along
with Prop. 8 strongly motivates the consideration of a single
curve as an approximation to the family of curves in Fig. 6.
A natural approach is to consider the upper bound of (14),
which has the benefit of being available in closed form. This
is essentially the approach that was heuristically employed in
[17] without observing that it actually corresponds to a (tight)
bound for the equivalent interferer density.
1
1
0.8
0.8
λxR ,u (r)/λu
λx∗ ,u (r)/λu
8
0.6
0.4
√
exact (|x∗ | × 2 λa = 0, 0.5, 1, 1.5, 2, 2.5, 3)
piecewise-linear upper bound
piecewise-linear approx. in [14]
0.2
0
0
0.5
1
1.5
2
2.5
√
r × 2 λa
3
0.6
0.4
0.2
3.5
0
4
Fig. 6. Equivalent density function λx∗ ,u (r) for various |x∗ | (solid curves),
with the direction of the arrow corresponding to increasing |x∗ |. A piecewiselinear approximation and bound are also shown.
Fig. 7.
|xR |.
√
|xR | × 2 λa = 0, 0.5, 1, 1.5, 2, 2.5, 3
0
0.5
1
1.5
2
2.5
√
r × 2 λa
3
3.5
4
Equivalent density function λxR ,u (r) when x∗ = o for various
B. x∗ = o
Another approach is to consider a piecewise-linear curve,
which, in addition to having a simple closed form, results
in a closed form approximate expression for the bound of
LIx∗ ,u (s | x∗ ) as given below (details are straightforward and
omitted).
Lemma 9. With the piecewise-linear approximation
λx∗ ,u (r) ≈ λu min (δr, 1) , r ≥ 0,
(16)
for some δ > 0, the lower bound for LIx∗ ,u (s | x∗ ) equals
1 2
exp −λu π C(Pu s)2/α + 2
F̃ (3) − F̃ (2)
δ
3
where
,
(2π/α)/
sin(2π/α) and F̃ (x)
C
x
x
−1
with 2 F1 (·) denoting
2 F1 1, α , 1 + α , sPu δ α
hypergeometric function.
,
the
A piecewise-linear approximation as in (16) was heuristically proposed in [16] for a slightly more general system
model than the one considered here. A closed form formula
for δ was provided, only justified as obtained by means of
curve fitting without providing any details on the procedure.
For the system model √
considered in this paper, this formula
results in δ ≈ 0.82687 λa and the corresponding piecewiselinear approximation of the equivalent density is also depicted
in Fig. 6. It can be seen that this approximation is essentially
an attempt to capture the average behavior of λx∗ ,u (r) over
|x∗ |, thus explaining its good performance reported in [16].
However, it is not clear why this particular value of δ is
more preferable than any other (slightly) different value. A
more rigorous approach for the selection of δ is to consider
the tightest piecewise-linear upper bound for λx∗ ,u (r), which
leads to another (looser) lower bound for LIxR√
(s | x∗ ). This
,u
bound corresponds to a value of δ ≈ 1.13118 λa , found by
numerical optimization, and is also shown in Fig. 6.
This case corresponds to the typical node being an AP, i.e.,
a typical AP. Note that, in contrast to the cases when x∗ 6= o,
the Voronoi cell of the typical AP is not conditioned to cover
any point in R2 , therefore, no cell expansion effect is observed
and, for xR 6= o, it is possible that the receiver lies outside
the guard region. This is a different operational scenario from
the common downlink with nearest AP association [25] where
the intended receiver by default lies within the Voronoi cell
of the serving AP. It may arise, for example, in case of D2D
communications, where both communicating UEs receive data
(e.g., control information) from the same AP.
The equivalent interfering UE density is given in Prop. 6 and
is depicted in Fig. 7 for various values of |xR |. Note that the
case xR = o corresponds to the typical AP in receive mode,
whereas for xR 6= o the typical AP transmits. It can be seen
that increasing |xR | effectively results in reduced protection
in the close vicinity of the receiver since the latter is more
probable to lie near the edge or even outside the Voronoi guard
region.
C. xR = o
This case corresponds to the typical node being at receive
mode. One interesting scenario which corresponds to this case
is downlink cellular transmission with nearest AP association,
where interference is due to transmitting UEs (e.g., in uplink
or D2D mode) that lie outside the serving cell.
When x∗ = o, the setup matches the case xR = x∗
∗
with
above, and it holds λo,u (r|x∗ ) =
x = o, discussed
−λa πr 2
∗
λu 1 − e
. For x 6= o, the asymptotic behavior of
∗
λo,u (r|x ) can be obtained by the same approach as in the
proof of Prop. 8.
Proposition 10. For x∗ 6= o the equivalent density function
λxR ,u (r) = λo,u (r|x∗ ) of Prop. 5 equals
8
λo,u (r) = λu λa |x∗ |r + O(r2 ), r → 0.
π
9
1
λo,u (r)/λu
0.8
0.6
0.4
0.2
0
√
|x∗ | × 2 λa = 0, 0.5, 1, 1.5, 2, 2.5, 3
0
0.5
1
1.5
2
2.5
√
r × 2 λa
3
3.5
4
Fig. 8. Equivalent density function λo,u (r) for various |x∗ |, with the
direction of the arrow corresponding to increasing |x∗ |.
In stark contrast to the behavior of λx∗ ,u (r), λo,u (r) tends
to 0 only linearly (instead of quadradically) as r → 0 when
x∗ 6= o, and with a rate that is also proportional to (instead
of independent of) |x∗ |. This implies that the typical node
is not as well protected from interfering UEs as its nearest
AP, experiencing an increased equivalent interferer density in
its close vicinity as |x∗ | increases. This can be understood by
noting that with increasing |x∗ |, the origin, although contained
in a guard region of (statistically) increasing area, is more
probable to lie near the cell edge (Fig. 1 demonstrates such a
case.)
The strong dependence of λo,u (r) on |x∗ | observed for r →
0 also holds for all values of r. This is shown in Fig. 8 where
the numerically evaluated λo,u (r) is depicted for the same
values of |x∗ | as those considered in Fig. 6. It can be seen
that λo,u (r) increases with |x∗ | for all r up to approximately
√
1 − 1.5 times the average value of |x∗ | (equal to 1/(2 λa )),
whereas it decreases with |x∗ | for greater r .
This behavior of λo,u (r), especially for r → 0, does not
permit the use of single function to be used as a reasonable
approximation or bound for λo,u (r), irrespective of |x∗ |, as
was done for the case for λx∗ ,u (r). Noting that the upper
bound of (13) has the same value whenever |x∗ | or |xR | are
zero, it follows that the curves shown in Fig. 7 with |xR |
replaced by |x∗ | are upper bounds for the corresponding curves
of Fig. 8. Unfortunately, it can be seen that the bound is very
loose for small r and/or large |x∗ |.
D. |xR | ∈ (0, |x∗ |), ∠xR = 0
The previous cases demonstrate how different the properties of the UE-generated interference become when different
receiver positions are considered. This observation motivates
the question of which receiver position is best protected from
UE interfernce given the positions o and x∗ 6= o of the typical
node and nearest AP, respectively. This question is relevant in,
e.g., relay-aided cellular networks where the communication
between a UE and its nearest AP is aided by another node (e.g.,
an inactive UE) [32]. Although the performance of a relayaided communication depends on multiple factors including
the distances among the nodes involved and considered transmission scheme [33], a reasonable choice for the realy position
is the one experiencing less interference.
By symmetry of the system geometry, this position should
lie on the line segment joining the origin with x∗ (inclusive).
However, as can be seen from Figs. 6 and 8, the shape of
the equivalent interferer density does not allow for a natural
ordering of the different values of xR . For example, λx∗ ,u (r) is
smaller than λo,u (r) for small r, whereas the converse holds
for large r. Noting that the receiver performance is mostly
affected by nearby interference [1], a natural criterion to order
the interfering UE densities for varying xR is their asymptotic
behavior as r → 0. For |xR | = |x∗ | and |xR | = o, this is
given in Props. 8 and 10, respectively. For all other positions
the asymptotic density is given by the following result.
Proposition 11. For |xR | ∈ (0, |x∗ |), ∠xR = 0, and x∗ 6= o,
the equivalent density function λxR ,u (r) of Prop. 5 equals
λxR ,u (r) = λu λa
8 (|x∗ |/|xR |)2 3
r + O(r4 ), r → 0. (17)
9π |x∗ | − |xR |
It follows that the optimal receiver position must have
|xR | ∈ (0, |x∗ |) since, in this case, the density scales as O(r3 )
instead of O(r) and O(r2 ) when |xR | = 0 and |xR | = |x∗ |,
respectively. Its value can be easily obtained by minimization
of the expression in (17) w.r.t. |xR |.
Corollary 12. The receiver position that experiences the
smallest equivalent UE interferer density in its close vicinity
∗
lies on the line
√ segment joining the origin to x and is at dis∗
tance |x |/ 2 from the origin. The corresponding equivalent
32 3
r + O(r4 ), r → 0.
density equals λxR ,u (r) = λu λa 9π
V. N UMERICAL E XAMPLES
In this section, numerical examples will be given, demonstrating the performance of various operational scenarios in
cellular networks that can be modeled as special cases of the
analysis presented in the previous sections. The system performance metric that will be considered is LIxR ,k (ραk θ|x∗ ),
with k ∈ {a, u} depending on whether AP- or UE-generated
interference is considered, with ρ > 0, θ > 0 given. Note that
this metric corresponds to the coverage probability that the
signal-to-interference ratio (SIR) at xR is greater than θ, when
the distance between transmitter and receiver is ρ, the direct
link experiences a path loss exponent ak and Rayleigh fading,
and the transmit power is equal to Pk [1]. For simplicity and
w.l.o.g., all the following examples correspond to λa = 1 and
αa = αu = 4.
The analytical results of the previous sections will be
compared against simulated system performance where, in
a single experiment, the distribution of interfering APs and
UEs is generated as follows. Given the AP position closest
to the origin, x∗ , the interfering AP positions are obtained as
a realization of a PPP of density as in (4). Afterwards, an
independent realization of a HPPP of density λu is generated,
which, after removing the points lying within V ∗ , represents
the interferering UE positions.
10
1
1
0.8
coverage probability
0.8
coverage probability
simulation
analysis
1
4
1
2
0.6
3
4
0.4
1
|xR |
|x∗ |
0.2
0
−20
=
5
4
−10
0.6
=
1
4
1
2
3
4
0.4
1
0.2
sim. (PPP)
sim. (VPLP)
analysis
−15
|xR |
|x∗ |
5
4
−5
0
5
10
15
20
θ (dB)
(a)
0
−20
−15
−10
−5
0
5
10
15
20
θ (dB)
(b)
Fig. 9. Coverage probability of the link between the typical node and an isotropically positioned receiver under UE interference for (a) λu = λa , (b)
λu = 10λa .
In addition, for cases where the UE-generated interference
is of interest, a VPLP process of density λu = λa is also
simulated, corresponding to scenarios where one UE from each
Voronoi cell in the system transmits (as in the standard uplink
scenario with nearest AP association).
A. D2D Link Not Necessarily Contained in a Single Cell
In this example, UE-generated inteference is considered at
a receiver position xR 6= o with ∠xR uniformly distributed
in [−π, π) and ρ = |xR |. This case may model a D2D link
where the typical node is a UE that directly transmits to a
receiver isotropically distributed at a distance |xR | > 0 . A
guard region V ∗ is established by the closest AP to the typical
node, however, it is not guaranteed to include xR , i.e., the D2D
nodes may not be contained within the same cell, which is a
common scenario that arises in practice, referred to as crosscell D2D communication [21].
With interference generated from other UEs in D2D and/or
uplink mode, a lower bound for LIxR ,u (ραu θ|x∗ ) can be
obtained using Prop. 5 with an equivalent density function
λ|xR |,u (r) as given in (12). Figure 9 shows this lower bound
for λu = λa (Fig. 9a) and λu = 10λa (Fig. 9b). The
position x∗ √is assumed to be isotropically distributed with
|x∗ | = 1/(2 λa ) 2 and various values of the ratio |xR |/|x∗ |
are considered. It can be seen that, compared to the simulation
under the PPP interference assumption, the quality of the
analytical lower bound depends on λu . For λu = λa , it
is very close to the exact coverage probability, whereas for
λu = 10λa , it is reasonably tight and able to capture the
behavior of the exact coverage probability over varying |xR |.
Also, for λu = λa and a VPLP interference, the analytical
expression provides a reasonably accurate approximation for
the performance of this analytically intractable case. As expected, performance degradation is observed with increasing
2 Recall that this value is the average distance of the closest AP from the
typical node.
|xR |, due to both increasing path loss of the direct link as well
as increased probability of the receiver lying outside V ∗ .
B. AP-to-D2D-Receiver Link
In this example, AP-generated inteference is considered at
a receiver position xR 6= o with ∠xR uniformly distributed
in [−π, π) and ρ = |x∗ − xR |. This case is similar to the
previous, however, considering AP-generated interference and
the receiver at xR receiving data from the AP node at x∗ . This
case models the scenario where the typical node establishes a
D2D connection with a node at xR , and the AP at x∗ sends
data to xR via a dedicated cellular link (e.g., for control purposes or for implementing a cooperative transmission scheme).
Note that, in contrast to the previous case, the link distance
ρ is random and equal to ρ = |x∗ | + |xR | − 2|x∗ ||xR | cos ψ,
with ψ uniformly distributed in [−π, π). The exact coverage
probability of this link can be obtained using Prop. 2 followed
by a numerically computed expectation over ψ.
αa
Figure 10 shows E(LIxR ,a (ρ√
θ|x∗ )) for an isotropically
∗
∗
distributed x with |x | = 1/(2 λa ), and for various values
of the ratio |xR |/|x∗ |. Note that the case |xR | = 0 corresponds
to the standard downlink transmission model with nearest
AP association [25]. Monte Carlo evaluation of the coverage
probability (not shown here) perfectly matches the analytical
curves. It can be seen that increasing |xR | reduces the coverage
probability for small SIR but increases it for high SIR. This is
in direct proportion to the behavior of the equivalent interferer
density λxR ,a (r) with increasing |xR | shown in Fig. 2.
C. Uplink with Nearest AP Association
In this example, UE-generated interference is considered
at a receiver position xR = x∗ with ρ = |x∗ | and λu = λa
corresponding to the conventional uplink transmission scenario
with nearest AP association and one active UE per cell (no
cross-mode interference). Figure 11 shows the lower bound
11
1
1
|xR |
|x∗ |
0.6
5
4
0.4
1
1
3 2
4
0.8
=0
coverage probability
coverage probability
0.8
1
4
0.6
0.4
√
0.2 |x∗ | × 2 λa = 2
0.2
−15
−10
−5
0
5
10
15
0
−20
20
θ (dB)
Fig. 10. Coverage
probability for the link between the AP positioned at x∗
√
(|x∗ | = 1/(2 λa )) and an isotropicaly positioned receiver.
of LIxR ,u (ραu θ|x∗ ) obtained by Prop. 5 as well as the
looser, but more easily computable, lower bound obtained
using the closed
form expression given in Lemma 7 with
√
δ = 1.13118 λa (resulting in the tightest bound possible with
a piecewise-linear equivalent density function). The coverage
√
probability is computed for an AP at a distance |x∗ |=c/(2 λa )
from the typical node with c = 1/2, 1, 2, roughly corresponding to a small, average, and large uplink distance (performance
is independent of ∠x∗ ).
As was the case in Fig. 9a, Prop. 5 provides a very tight
lower bound for the actual coverage probability (under the PPP
model for the interfering UE positions). The bound of Lemma
7, although looser, is nevertheless a reasonable approximation
of the performance, especially for smaller values of |x∗ |.
Compared to the VPLP model, the PPP model provides an
optimistic performance prediction that is, however, reasonably
tight, especially for large |x∗ |. This observation motivates its
usage as a tractable approximation of the actual inteference
statistics as was also reported in [16], [17]. Interestingly,
the bound of Lemma 7 happens to provide an even better
approximation for√the VPLP performance for |x∗ | close to or
smaller than 1/(2 λa ).
−15
−10
1
2
1
−5
0
5
10
15
20
θ (dB)
Fig. 11. Coverage probability of the uplink between the typical node its
nearest AP (λu = λa ).
1
Case A
Case B
Case C
0.8
coverage probability
0
−20
sim. (PPP)
sim. (VPLP)
Prop. 5
Lemma 7
0.6
0.4
0.2
√
ρ × 2 λa = 2
0
−20
−15
−10
1
2
1
−5
0
5
10
15
20
θ (dB)
Fig. 12. Coverage probability under UE interference for the links corresponding to the cases described in Sec. V. D.
D. Effect of Guard Region on UE-Generated Interference
Protection
Case C (receiver imposes a Voronoi guard region): This
case corresponds to xR = x∗ = o modeling an nonstandard uplink transmission (see also description in
Table I)
In order to see the effectiveness of a Voronoi guard region
for enhancing link quality under UE-generated interference,
the performance under the following operational cases is
examined.
• Case A (no guard region is imposed): This results in the
standard transmission model under an HPPP of interferers
positions of density λu [1]. The exact coverage probability is well known (see, e.g., [1, Eq. 3.29]).
• Case B (transmitter imposes a Voronoi guard region):
This case corresponds to x∗ = o, xR 6= o, modeling an
non-standard downlink transmission (see also description
in Table I).
Cases B and C correspond to the analysis considered in Sec.
V. B. Assuming the same link distance ρ for all cases, Fig. 12
shows the analytically obtained coverage probability (exact for
Case A, lower bound for Cases B and C), assuming λu = λa .
It can be seen that imposing a Voronoi guard region (Cases B
and C) is always beneficial, as expected. However, for large
link distances, Case B provides only marginal gain as the
receiver is very likely to be located at the edge or even outside
the guard region. In contrast, the receiver is always guaranteed
to be protected under case C, resulting in the best performance
and significant gains for large link distances.
•
12
VI. C ONCLUSION
This paper considered the analytical characterization of
cross-mode inter-cell interference experienced in future cellular networks. By employing a stochastic geometry framework,
tractable expressions for the interference statistics were obtained that are exact in the case of AP-generated interference
and serve as a tight lower bound in the case of UE-generated
interference. These expressions are based on appropriately
defined equivalent interferer densities, which provide an intuitive quantity for obtaining insights on how the interference
properties change according to the type of interference and
receiver position. The considered system model and analysis
are general enough to capture many operational scenarios
of cellular networks, including conventional downlink/uplink
transmissions with nearest AP association as well as D2D
transmissions between UEs that do not necessarily lie in the
same cell. The analytical expressions of this paper can be
used for sophisticated design of critical system aspects such
as mode selection and resource allocation towards getting the
most out D2D- and/or FD-enabled cellular communications.
Interesting topics for future research is investigation of the
joint properties of AP- and UE-generated interference as well
as extension of the analysis to the case of heterogeneous,
multi-tier networks.
A PPENDIX A
P ROOF OF P ROPOSITION 2
Let Φ̃a , Φa \ {x∗ } denote the interfering AP point
process. Given x∗ , Φ̃a is an inhomogeneous PPP of density
λ̃a (x), x ∈ R2 , defined in (4). Density λ̃a (x) is circularly
symmetric w.r.t. xR = o, which, using Lemma 1, directly
leads to the Laplace transform expression of (3) with radial
density as in (6). Considering the case xR 6= o, it directly
follows from Lemma 1 and (2) that
Z
LIxR ,a (s) = exp −
λ̃a (x + xR )γ(sPa |x|−αa )dx
R2
by a change of integration variable. By switching to polar
coordinates (centered at o) for the integration, the right-hand
side of (3) results with P = Pa , α = αa and λxR (r) =
λxR ,a (r), where
Z 2π
1
λ̃a ((r, θ) + xR )dθ
λxR ,a (r) =
2π 0
Z
λa 2π
=
I {(r, θ) + xR ∈
/ B(o, |x∗ |)} dθ
2π 0
Z
λa 2π
=
I {(r, θ) ∈
/ B(−xR , |x∗ |)} dθ
2π 0
Z
λa 2π
=
I {(r, θ) ∈
/ B((|xR |, 0), |x∗ |)} dθ,
2π 0
with xR representing the polar coordinates of the receiver in
the first equality, with a slight abuse of notation. The last
equality follows by noting that the integral does not depend on
the value of ∠xR . Evaluation of the final integral is essentially
the evaluation of an angle (see Fig. 13), which can be easily
obtained by elementary Euclidean geometry, resulting in the
expression of (5).
B((|xR |, 0), |x∗ |)
C(o, r)
|xR |
θ0
Fig.
the angle θ0 , which is equal to quantity
R 2π 13. Geometrical figure depicting
/ B((|xR |, 0), |x∗ |)) dθ appearing in the proof of Prop. 2.
0 I ((r, θ) ∈
A PPENDIX B
P ROOF OF P ROPOSITION 5
It holds
LIxR ,u (s | x∗ )
=E E e−sIxR ,u | Φa | x∗
Z
(a)
= E exp −
λ̃u (x|Φa )γ(sPu |x − xR |−αu )dx x∗
R2
Z
∗
−αu
≥ exp −
E λ̃u (x + xR |Φa ) | x γ(sPu |x|
)dx ,
R2
where (a) follows by noting that, conditioned on Φa , the interfering UE point process equals Φu \V ∗ , which is distributed
as an inhomogeneous PPP of density λ̃u (x|Φ1 ), x ∈ R2
as given in (8), and using Lemma 1. The inequality is an
application of Jensen’s inequality with a change
of integration
variable. Result follows by noting that E λ̃u (x|Φa ) | x∗ =
λu [1 − pc (x | x∗ )] and switching to polar coordinates for the
integration.
A PPENDIX C
P ROOF OF P ROPOSITION 6
Replacing the PCA function in (11) with its lower bound
as per Cor. 4 results in the equivalent density bound
λxR ,u (r)
Z 2π
1
−λa π|(r,θ)+xR −x∗ |2
e
dθ
≤λu 1 −
2π 0
!
∗ 2 Z 2π
(a)
e−λa π|x |
−λa π|(r,θ)+xR |2
≤ λu 1 −
e
dθ
2π
0
!
∗ 2 Z 2π
2
e−λa π|x |
(b)
=λu 1 −
e−λa π|(r,θ)+(|xR |,0)| dθ
2π
0
!
∗ 2
2
2 Z
e−λa π(|x | +|xR | +r ) 2π λa π2r|xR | cos θ
(c)
= λu 1 −
e
dθ ,
2π
0
where (a) follows by application of the triangle inequality,
(b) by noting that the integral is independent of ∠xR and
(c) since |(r, θ) + (|xR |, 0)|2 = r2 + |xR |2 − 2r|xR | cos(π −
θ) (cosine law). The integral in the last equation is equal to
2πI0 (λa 2π|xR |r). By exchanging the roles x∗ and xR in (a)
and following the same reasoning, another upper bound of the
form of (c) results, with the only difference that the integrand
has |x∗ | instead of |xR | and evaluates to 2πI0 (λa 2π|x∗ |r).
Considering the minimum of these two bounds leads to (13).
13
R EFERENCES
[1] M. Haenggi and R. K. Ganti, “Interference in Large Wireless Networks,”
Found. Trends Netw., vol. 3, no. 2, pp. 127–248, 2008.
[2] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
“Five disruptive technology directions for 5G,” IEEE Commun. Mag.,
vol. 52, pp. 74–80, Feb. 2014.
[3] K. Doppler, M. Rinne, C. Wijting, C. Ribeiro, and K. Hugl, “Deviceto-Device Communication as an Underlay to LTE-Advanced Networks,”
IEEE Commun. Mag., vol. 50, no. 3, pp. 170–177, Mar. 2012.
[4] A. Sabharwal, P. Schniter, D. Guo, D. Bliss, S. Rangarajan, and R. Wichman, “In-band full-duplex wireless: Challenges and opportunities,” IEEE
J. Select. Areas Commun., vol. 32, pp. 1637–1652, Sept. 2014.
[5] L. Lei and Z. Zhong, “Operator controlled device-to-device communications in LTE-Advanced networks,” IEEE Wireless Commun., vol. 19,
no. 3, pp. 96–104, Jun. 2012.
[6] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M.
Franceschetti, “Stochastic geometry and random graphs for the analysis
and design of wireless networks,” IEEE J. Select. Areas Commun., vol.
27, pp. 1029-1046, Sept. 2009.
[7] Q. Ye, M. Al-Shalash, C. Caramanis, and J. G. Andrews, “Resource optimization in device-to-device cellular systems using time-frequency hopping,” IEEE Trans. Wireless Commun., vol. 13, no. 10, pp. 5467–5480,
Oct. 2014.
[8] X. Lin, J. G. Andrews, and A. Ghosh, “Spectrum sharing for deviceto-device communication in cellular networks,” IEEE Trans. Wireless
Commun., vol. 13, no. 12, pp. 6727–6740, Dec. 2014.
[9] N. Lee, X. Lin, J. G. Andrews, and R. W. Heath, Jr., “Power control for
D2D underlaid cellular networks: Modeling, algorithms and analysis,”
IEEE J. Sel. Areas Commun., vol. 33, no. 1, pp. 1–13, Jan. 2015.
[10] S. Stefanatos, A. Gotsis, and A. Alexiou, “Operational region of D2D
communications for enhancing cellular network performance,” IEEE
Trans. Wireless Commun., vol. 14, no. 11, pp. 5984–5997, Nov. 2015.
[11] A. Hasan and J. G. Andrews, “The guard zone in wireless ad hoc
networks,” IEEE Trans. Wireless Commun., vol. 6, no.3, pp. 897–906,
Mar. 2007.
[12] G. George, R. K. Mungara, and A. Lozano, “An Analytical Framework for Device-to-Device Communication in Cellular Networks,” IEEE
Trans. Wireless Commun., vol. 14, no. 11, pp. 6297–6310, Nov. 2015.
[13] Z. Chen and M. Kountouris, “Decentralized opportunistic access for
D2D underlaid cellular networks,” Jul. 2016. [Online]. Available:
http://arxiv.org/abs/1607.05543.
[14] J. G. Andrews, R. K. Ganti, N. Jindal, M. Haenggi, and S. Weber, “A
primer on spatial modeling and analysis in wireless networks,” IEEE
Commun. Mag., vol. 48, no. 11, pp. 156–163, Nov. 2010.
[15] H. ElSawy and E. Hossain, “On stochastic geometry modeling of cellular
uplink transmission with truncated channel inversion power control,”
IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4454–4469, Aug.
2014.
[16] H. Lee, Y. Sang, and K. Kim, “On the uplink SIR distributions in
heterogeneous cellular networks,” IEEE Commun. Lett., vol. 18, no. 12,
pp. 2145–2148, Dec. 2014.
[17] S. Singh, X. Zhang, and J. G. Andrews, “Joint rate and SINR coverage
analysis for decoupled uplink-downlink biased cell associations in HetNets,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5360–5373,
Oct. 2015.
[18] B. Błaszczyszyn and D. Yogeshwaran, “Clustering comparison of point
processes with applications to random geometric models,” in Stochastic
Geometry, Spatial Statistics and Random Fields, vol. 2120, Lecture
Notes in Mathematics. Cham, Switzerland: Springer-Verlag, 2015, pp.
31–71.
[19] M. Haenggi, “User point processes in cellular networks,” IEEE Wireless
Commun. Lett., vol. 6, no. 2, pp. 258–261, Apr. 2017.
[20] M. Di Renzo, W. Lu, and P. Guan, “The intensity matching approach:
A tractable stochastic geometry approximation to system-level analysis
of cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 9,
pp. 5963–5983, Sep. 2016.
[21] H. Zhang, Y. Ji, L. Song, and H. Zhu “Hypergraph based resource
allocation for cross-cell device-to-device communications,” in Proc.
IEEE Int. Conf. Commun. (ICC), Kuala Lumpur, Malaysia, May 2016,
pp. 1–6.
[22] Z. Tong and M. Haenggi, “Throughput analysis for full-duplex wireless
networks with imperfect self-interference cancellation,” IEEE Trans.
Commun., vol. 63, no. 11, pp. 4490–4500, Nov. 2015.
[23] C. Psomas, M. Mohammadi, I. Krikidis, and H. A. Suraweera, “Impact
of directionality on interference mitigation in full-duplex cellular net-
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
works,” IEEE Trans. Wireless Commun., vol. 16, no. 1, pp. 487–502,
Jan. 2017.
K. S. Ali, H. ElSawy, and M.-S. Alouini, “Modeling cellular networks
with full-duplex D2D communication: a stochastic geometry approach,”
IEEE Trans. Commun., vol. 64, no. 10, pp. 4409–4424, Oct. 2016.
J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
no. 11, pp. 3122–3134, Nov. 2011.
S. Stefanatos and A. Alexiou, “Access Point density and bandwidth
partitioning in ultra-dense wireless networks,” IEEE Trans. Commun.,
vol. 62, no. 9, pp. 3376–3384, Sept. 2014.
F. Baccelli and B. Błaszczyszyn, Stochastic Geometry and Wireless
Networks. Delft, The Netherlands: Now Publishers Inc., 2009.
Z. Chen and M. Kountouris, “Guard zone based D2D underlaid cellular
networks with two-tier dependence,” 2015 IEEE International Conference on Communication Workshop (ICCW), London, 2015, pp. 222-227.
M. Haenggi, “The local delay in poisson networks,” IEEE Trans. Inf.
Theory, vol. 59, pp. 1788–1802, Mar. 2013.
Z. Yazdanshenasan, H. S. Dhillon, M. Afshang, and P. H. J. Chong,
“Poisson hole process: theory and applications to wireless networks,”
IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7531–7546, Nov.
2016.
S. Sadr and R. S. Adve, “Handoff rate and coverage analysis in multitier heterogeneous networks,” IEEE Trans. Wireless Commun., vol. 14,
no. 5, pp. 2626–2638, May 2015.
H. E. Elkotby and M. Vu, “Uplink User-Assisted Relaying Deployment
in Cellular Networks,” IEEE Trans. on Wireless Commun., vol. 14, no.
10, pp. 5468–5483, Oct. 2015.
A. E. Gamal and Y.-H. Kim, Network Information Theory, Cambridge
University Press, 2012.
| 7cs.IT
|
arXiv:1705.07686v1 [cs.PL] 22 May 2017
On the computational complexity of dynamic
slicing problems for program schemas
Michael R. Laurence c Sebastian Danicic a Robert M. Hierons b
a Department
b
c
of Computing, Goldsmiths College, University of London, New
Cross, London SE14 6NW UK
Department of Information Systems and Computing, Brunel University,
Uxbridge, Middlesex, UB8 3PH.
Department of Computer Science, Regent Court, 211 Portobello, Sheffield, S1
4DP, UK.
Abstract
Given a program, a quotient can be obtained from it by deleting zero or more
statements. The field of program slicing is concerned with computing a quotient
of a program which preserves part of the behaviour of the original program. All
program slicing algorithms take account of the structural properties of a program
such as control dependence and data dependence rather than the semantics of its
functions and predicates, and thus work, in effect, with program schemas. The
dynamic slicing criterion of Korel and Laski requires only that program behaviour
is preserved in cases where the original program follows a particular path, and that
the slice/quotient follows this path. In this paper we formalise Korel and Laski’s
definition of a dynamic slice as applied to linear schemas, and also formulate a less
restrictive definition in which the path through the original program need not be
preserved by the slice. The less restrictive definition has the benefit of leading to
smaller slices. For both definitions, we compute complexity bounds for the problems
of establishing whether a given slice of a linear schema is a dynamic slice and whether
a linear schema has a non-trivial dynamic slice and prove that the latter problem
is NP-hard in both cases. We also give an example to prove that minimal dynamic
slices (whether or not they preserve the original path) need not be unique.
Key words: program schemas, program slicing, NP-completeness, Herbrand
domain, linear schemas
Preprint submitted to Elsevier
23 May 2017
u := h();
if p(w)
then v := f (u);
else v := g();
Fig. 1. A schema
1
Introduction
A schema represents the statement structure of a program by replacing real functions
and predicates by symbols representing them. A schema, S, thus defines a whole class
of programs which all have the same structure. A schema is linear if it does not contain
more than one occurrence of the same function or predicate symbol. As an example,
Figure 1 gives a schema S; and Figure 2 shows one of the programs obtainable from
the schema of Figure 1 by interpreting its function and predicate symbols.
The subject of schema theory is connected with that of program transformation and
was originally motivated by the wish to compile programs effectively[1]. Thus an
important problem in schema theory is that of establishing whether two schemas are
equivalent; that is, whether they always have the same termination behaviour, and give
the same final value for every variable, given any initial state and any interpretation of
function and predicate symbols. In Section 1.2, the history of this problem is discussed.
Schema theory is also relevant to program slicing, and this is the motivation for the
main results of this paper. We define a quotient of a schema S to be any schema
obtained by deleting zero or more statements from S. A quotient of S is non-trivial
if it is distinct from S. Thus a quotient of a schema is not required to satisfy any
semantic condition; it is defined purely syntactically. The field of program slicing
is concerned with computing a quotient of a program which preserves part of the
behaviour of the original program. Program slicing is used in program comprehension
[2,3], software maintenance [4,5,6,7], and debugging [8,9,10,11].
All program slicing algorithms take account of the structural properties of a program
such as control dependence and data dependence rather than the semantics of its
functions and predicates, and thus work, in effect, with linear program schemas. There
are two main forms of program slicing; static and dynamic.
u := 1;
if w > 1 then v := u + 1;
else v := 2;
Fig. 2. A program defined from the schema of Figure 1
2
• In static program slicing, only the program itself is used to construct a slice. Most
static slicing algorithms are based on Weiser’s algorithm[12], which uses the data
and control dependence relations of the program in order to compute the set of
statements which the slice retains. An end-slice of a program with respect to a
variable v is a slice that always returns the same final value for v as the original
program, when executed from the same input. It has been proved that Weiser’s algorithm gives minimal static end-slices[13] for linear, free, liberal program schemas.
This result has recently been strengthened by allowing function-linear schemas, in
which only predicate symbols are required to be non-repeating[14].
• In dynamic program slicing, a path through the program is also used as input.
Dynamic slices of programs may be smaller than static slices, since they are only
required to preserve behaviour in cases where the original program follows a particular path. As originally formulated by Korel and Laski [15], a dynamic slice
of a program P is defined by three parameters besides P , namely a variable set
V , an initial input state d and an integer n. The slice with respect to these parameters is required to follow the same path as P up to the nth statement (with
statements not lying in the slice deleted from the path through the slice) and give
the same value for each element of V as P after the nth statement after execution from the initial state d. Many dynamic slicing algorithms have been written
[16,17,18,19,20,15,21,22]. Most of these compute a slice using the data and control
dependence relations along the given path through the original program. This produces a correct slice, and uses polynomial time, but need not give a minimal or
even non-trivial slice even where one exists.
Our definition of a path-faithful dynamic slice (PFDS) for a linear schema S comprises
two parameters besides S, namely a path through S and a variable set, but not an
initial state. This definition is analogous to that of Korel and Laski, since the initial
state included in their parameter set is used solely in order to compute a path through
the program in linear schema-based slicing algorithms. We prove, in effect, that it is
decidable in polynomial time whether a particular quotient of a program is a dynamic
slice in the sense of Korel and Laski, and that the problem of establishing whether a
program has a non-trivial path-faithful dynamic slice is intractable, unless P=NP. This
shows that there does not exist a tractable dynamic slicing algorithm that produces
correct slices and always gives a non-trivial slice of a program where one exists.
The requirement of Korel and Laski that the path through the slice be path-faithful
may seem unnecessarily strong. Therefore we define a more general dynamic slice (DS),
in which the sequence of functions and predicates through which the path through
the slice passes is a subsequence of that for the path through the original schema,
but the path through the slice must still pass the same number of times through the
program point at the end of the original path. For this less restrictive definition, we
prove that it is decidable in Co-NP time whether a particular slice of a program is a
dynamic slice, and the problem of establishing whether a program has a non-trivial
dynamic slice is NP-hard.
3
We also give an example to prove that unique minimal dynamic slices (whether or
not path-faithful) of a linear schema S do not always exist.
The results of this paper have several practical ramifications. First, we prove that
the problem of deciding whether a linear schema has a non-trivial dynamic slice is
computationally hard and clearly this result must also hold for programs. In addition,
since this decision problem is computationally hard, the problem of producing minimal
dynamic slices must also be computationally hard. Second, we define a new notion
of a dynamic slice that places strictly weaker constraints on the slice than those
traditionally used and thus can lead to smaller dynamic slices. In Section 4 we explain
why these (smaller) dynamic slices can be appropriate, motivating this through a
problem in program testing. Naturally, this weaker notion of a dynamic slice is also
directly applicable to programs. Finally, we prove that minimal dynamic slices need
not be unique and this has consequences when designing dynamic slicing algorithms
since it tells us that algorithms that identify and then delete one statement at a time
can lead to suboptimal dynamic slices.
It should be noted that much theoretical work on program slicing and program analysis, including that of Müller-Olm’s study of dependence analysis of parallel programs [23], and on deciding validity of relations between variables at given program points [24,25] only considers programs in which branching is treated as nondeterministic, and is thus more ‘approximate’ than our own in this respect, in that
we take into account control dependence as part of the program structure.
1.1 Different classes of schemas
Many subclasses of schemas have been defined:
Structured schemas, in which goto commands are forbidden, and thus loops must
be constructed using while statements. All schemas considered in this paper are
structured.
Linear schemas, in which each function and predicate symbol occurs at most once.
Free schemas, where all paths are executable under some interpretation.
Conservative schemas, in which every assignment is of the form
v := f (v1 , . . . , vr ); where v ∈ {v1 , . . . , vr }.
Liberal schemas, in which two assignments along any executable path can always
be made to assign distinct values to their respective variables by a suitable choice
of domain.
It can be easily shown that all conservative schemas are liberal.
Paterson [26] gave a proof that it is decidable whether a schema is both liberal and
free; and since he also gave an algorithm transforming a schema S into a schema T
such that T is both liberal and free if and only if S is liberal, it is clearly decidable
4
whether a schema is liberal. It is an open problem whether freeness is decidable for
the class of linear schemas. However he also proved, using a reduction from the Post
Correspondence Problem, that it is not decidable whether a schema is free.
1.2 Previous results on the decidability of schema equivalence
Most previous research on schemas has focused on schema equivalence. All results
on the decidability of equivalence of schemas are either negative or confined to very
restrictive classes of schemas. In particular Paterson [27] proved that equivalence
is undecidable for the class of all schemas containing at least two variables, using a
reduction from the halting problem for Turing machines. Ashcroft and Manna showed
[28] that an arbitrary schema, which may include goto commands, can be effectively
transformed into an equivalent structured schema, provided that statements such
as while ¬p(u) do T are permitted; hence Paterson’s result shows that any class of
schemas for which equivalence can be decided must not contain this class of schemas.
Thus in order to get positive results on this problem, it is clearly necessary to define
the relevant classes of schema with great care.
Positive results on the decidability of equivalence of schemas include the following; in
an early result in schema theory, Ianov [29] introduced a restrictive class of schemas,
the Ianov schemas, for which equivalence is decidable. This problem was later shown
to be NP-complete [30,31]. Ianov schemas are characterised by being monadic (that
is, they contain only a single variable) and having only unary function symbols; hence
Ianov schemas are conservative.
Paterson [26] proved that equivalence is decidable for a class of schemas called progressive schemas, in which every assignment references the variable assigned by the
previous assignment along every legal path.
Sabelfeld [32] proved that equivalence is decidable for another class of schemas called
through schemas. A through schema satisfies two conditions: firstly, that on every
path from an accessible predicate p to a predicate q which does not pass through
another predicate, and every variable x referenced by p, there is a variable referenced
by q which defines a term containing the term defined by x, and secondly, distinct
variables referenced by a predicate can be made to define distinct terms under some
interpretation.
It has been proved that for the class of schemas which are linear, free and conservative,
equivalence is decidable [33]. More recently, the same conclusion was proved to hold
under the weaker hypothesis of liberality in place of conservatism [34,35].
5
1.3 Organisation of the paper
In Section 2 we give basic definitions of schemas. In Section 3 we define path-faithful
dynamic slices and in Section 4 we define general dynamic slices. In Section 5 we give
an example to prove that unique minimal dynamic slices need not exist. In Section 6
we prove complexity bounds for problems concerning the existence of dynamic slices.
Lastly, in Section 7, we discuss further directions for research in this area.
2
Basic Definitions of Schemas
Throughout this paper, F , P, V and L denote fixed infinite sets of function symbols,
predicate symbols, variables and labels respectively. A symbol means an element of
F ∪P in this paper. For example, the schema in Figure 1 has function set F = {f, g, h},
predicate set P = {p} and variable set V = {u, v}. We assume a function
arity : F ∪ P → N.
The arity of a symbol x is the number of arguments referenced by x, for example in
the schema in Figure 1 the function f has arity one, the function g has arity zero,
and p has arity one.
Note that in the case when the arity of a function symbol g is zero, g may be thought
of as a constant.
The set Term(F , V) of terms is defined as follows:
• each variable is a term,
• if f ∈ F is of arity n and t1 , . . . , tn are terms then f (t1 , . . . , tn ) is a term.
For example, in the schema in Figure 1, the variable u takes the value (term) h();
after the first assignment is executed and if we take the true branch then the variable
v ends with the value (term) f (h()).
We refer to a tuple t = (t1 , . . . , tn ), where each ti is a term, as a vector term. We call
p(t) a predicate term if p ∈ P and the number of components of the vector term t is
arity(p).
Schemas are defined recursively as follows.
• skip is a schema.
• Any label is a schema.
• An assignment y := f (x); for a variable y, a function symbol f and an n-tuple x of
variables, where n is the arity of f , is a schema.
6
• If S1 and S2 are schemas then S1 S2 is a schema.
• If S1 and S2 are schemas, p is a predicate symbol and y is an m-tuple of variables,
where m is the arity of p, then if p(y) then S1 else S2 is a schema.
• If T is a schema, q is a predicate symbol and z is an m-tuple of variables, where m
is the arity of q, then the schema while q(z) T is a schema.
If no function or predicate symbol, or label, occurs more than once in a schema S, we
say that S is linear. If a schema does not contain any predicate symbols, then we say
it is predicate-free. If a linear schema S contains a subschema if p(y) then S1 else S2 ,
then we refer to S1 and S2 as the T-part and F-part respectively of p in S. For example
in the schema in Figure 1 the predicate p has T-part v := f (u); and F-part v := g();. If
a linear schema S contains a subschema while q(z) T , then we refer to T as the body
of q in S.
Quotients of schemas are defined recursively as follows; skip is a quotient of every
schema; if S ′ is a quotient of S then S ′ T is a quotient of ST and T S ′ is a quotient of
T S; if T ′ is a quotient of T , then while q(y) T ′ is a quotient of while q(y) T ; and if T1
and T2 are quotients of schemas S1 and S2 respectively, then if p(x) then T1 else T2 is a
quotient of if p(x) then S1 else S2 . A quotient T of a schema S is said to be non-trivial
if T 6= S.
Consider the schema in Figure 1. Here we can obtain a quotient by replacing the first
statement by skip or by replacing the if statement by skip. It is also possible to replace
either or both parts of the if statement by skip or any combination of these steps.
2.1 Paths through a schema
We will express the semantics of schemas using paths through them; therefore the definition of a path through a schema has to include the variables assigned or referenced
by successive function or predicate symbols.
The set of prefixes of a word (that is, a sequence) σ over an alphabet is denoted by
pre(σ). For example, if σ = x1 x3 x2 over the alphabet {x1 , x2 , x3 }, then the set pre(σ)
consists of the words x1 x2 x3 , x1 x2 , x1 and the empty word. More generally, if Ω is a
set of words, then we define pre(Ω) = {pre(σ)| σ ∈ Ω}.
For each schema S there is an associated alphabet alphabet(S) consisting of all elements of L and the set of letters of the form y := f (x) for assignments y := f (x); in
S and p(y), Z for Z ∈ {T, F}, where if p(y) or while p(y) occurs in S. For example,
the schema in Figure 1 has no labels and has alphabet
{y := h(), v := f (u), v := g(), p(w), T, p(w), F}. The set Π(S) of terminating paths through
S, is defined recursively as follows.
• Π(l) = l, for any l ∈ L.
7
•
•
•
•
•
Π(skip) is the empty word.
Π(y := f (x); ) = y := f (x).
Π(S1 S2 ) = Π(S1 ) Π(S2 ).
Π( if p(x) then S1 else S2 ) = p(x), T Π(S1 ) ∪ p(x), F Π(S2 ).
Π( while (q(y) T )) = (q(y), T Π(T ))∗ q(y), F.
We sometimes abbreviate q(y), Z to q, Z and y := f (x) to f .
We define Πω (S) to be the set containing Π(S), plus all infinite words whose finite
prefixes are prefixes of terminating paths. A path through S is any (not necessarily
strict) prefix of an element of Πω (S). As an example, if S is the schema in Figure 1,
which has no loops, then Π(S) = Πω (S). In fact, Π(S) in this case contains exactly
two paths, defined by p(w) taking the true or false branches, and every path through
S is a prefix of one of these paths.
If S ′ is a quotient of a schema S, and ρ ∈ pre(Π(S)) (that is, ρ is a path through S),
then proj S ′ (ρ) is the path obtained from ρ by deleting all letters having function or
predicate symbols not lying in S ′ and all labels not occurring in S ′ . It is easily proved
that proj S ′ (Π(S)) = Π(S ′ ) in this case.
2.2 Semantics of schemas
The symbols upon which schemas are built are given meaning by defining the notions
of a state and of an interpretation. It will be assumed that ‘values’ are given in a
single set D, which will be called the domain. We are mainly interested in the case in
which D = Term(F , V) (the Herbrand domain) and the function symbols represent
the ‘natural’ functions with respect to Term(F , V).
Definition 1 (states, (Herbrand) interpretations and the natural state e)
Given a domain D, a state is either ⊥ (denoting non-termination) or a function
V → D. The set of all such states will be denoted by State(V, D). An interpretation
i defines, for each function symbol f ∈ F of arity n, a function f i : D n → D, and for
each predicate symbol p ∈ P of arity m, a function pi : D m → {T, F}. The set of all
interpretations with domain D will be denoted Int(F , P, D).
We call the set Term(F , V) of terms the Herbrand domain, and we say that a function
from V to Term(F , V) is a Herbrand state. An interpretation i for the Herbrand
domain is said to be Herbrand if the functions f i : Term(F , V)n → Term(F , V) for
each f ∈ F are defined as
f i (t1 , . . . , tn ) = f (t1 , . . . , tn )
for all n-tuples of terms (t1 , . . . , tn ).
8
We define the natural state e : V → Term(F , V) by e(v) = v for all v ∈ V.
In the schema in Figure 1 the natural state simply maps variable u to the name u,
variable v to the name v, and variable w to the name w. The program in Figure 2 can
be produced from this schema through the interpretation that maps h(); to 1, p(w)
to w > 1, f (u) to u + 1, and g() to 2; clearly this is not a Herbrand interpretation.
Observe that if an interpretation i is Herbrand, this does not restrict the mappings
pi : (Term(F , V))m → {T, F} defined by i for each p ∈ P.
It is well known [36, Section 4-14] that Herbrand interpretations are the only ones
that need to be considered when considering many schema properties. This fact is
stated more precisely in Theorem 8. In particular, our semantic slicing definitions
may be defined in terms of Herbrand domains.
Given a schema S and a domain D, an initial state d ∈ State(V, D) with d 6= ⊥ and
an interpretation i ∈ Int(F , P, D) we now define the final state M[[S]]id ∈ State(V, D)
and the associated path πS (i, d) ∈ Πω (S). In order to do this, we need to define the
predicate-free schema associated with the prefix of a path by considering the sequence
of assignments through which it passes.
Definition 2 (the schema schema(σ))
Given a word σ ∈ (alphabet(S))∗ for a schema S, we recursively define the predicatefree schema schema(σ) by the following rules; schema(skip) = skip, schema(l) = l
for l ∈ L, schema(σv := f (x)) = schema(σ) v := f (x); and schema(σp(x), X) =
schema(σ).
Consider, for example, the path of the schema in Figure 1 that passes through the true
branch of p. Then this defines a word σ = u := h()p(w), Tv := f (u) and schema(σ) =
u := h()v := f (u).
Lemma 3 Let S be a schema. If σ ∈ pre(Π(S)), the set {m ∈ alphabet(S)| σm ∈
pre(Π(S))} is one of the following; a label, a singleton containing an assignment letter
y := f (x), a pair {p(x), T, p(x), F} for a predicate p of S, or the empty set, and if
σ ∈ Π(S) then the last case holds.
Proof. [14, Lemma 6].
Lemma 3 reflects the fact that at any point in the execution of a program, there is
never more than one ‘next step’ which may be taken, and an element of Π(S) cannot
be a strict prefix of another.
Definition 4 (semantics of predicate-free schemas) Given a state d 6= ⊥, the
final state M[[S]]id and associated path πS (i, d) ∈ Πω (S) of a schema S are defined as
follows:
9
• M[[skip]]id = d and πskip (i, d) is the empty word.
• M[[l]]id = d and πl (i, d) = l for l ∈ L.
d(v)
if v 6= y,
• M[[y := f (x);]]id (v)
(where the vector term d(x) =
=
f i (d(x)) if v = y
(d(x1 ), . . . , d(xn )) for x = (x1 , . . . , xn )), and
πy := f (x); (i, d)
=
y := f (x).
• For sequences S1 S2 of predicate-free schemas, M[[S1 S2 ]]id
πS1 S2 (i, d)
=
=
M[[S2 ]]iM[[S1 ]]i and
d
πS1 (i, d)πS2 (i, M[[S1 ]]id ).
This uniquely defines M[[S]]id and πS (i, d) if S is predicate-free. In order to give the
semantics of a general schema S, first the path, πS (i, d), of S with respect to interpretation, i, and initial state d is defined.
Definition 5 (the path πS (i, d)) Given a schema S, an interpretation i, and a state,
d 6= ⊥, the path πS (i, d) ∈ Πω (S) is defined by the following condition; for all
σ p(x), Z ∈ pre(πS (i, d)), the equality pi (M[[schema(σ)]]id (x)) = Z holds.
In other words, the path πS (i, d) has the following property; if a predicate expression
p(x) along πS (i, d) is evaluated with respect to the predicate-free schema consisting
of the sequence of assignments preceding that predicate in πS (i, d), then the value
of the resulting predicate term given by i ‘agrees’ with the value given in πS (i, d).
Consider, for example, the schema given in Figure 1 and the interpretation that gives
the program in Figure 2. Given a state d in which w has a value greater than one, we
obtain the path u := h() p(w), T v := f (u).
By Lemma 3, this defines the path πS (i, d) ∈ Πω (S) uniquely.
Definition 6 (the semantics of arbitrary schemas) If πS (i, d) is finite, we define
M[[S]]id = M[[schema(πS (i, d))]]id
(which is already defined, since schema(πS (i, d)) is predicate-free) otherwise πS (i, d)
is infinite and we define M[[S]]id = ⊥. In this last case we may say that M[[S]]id is not
terminating.
For convenience, if S is predicate-free and d : V → Term(F , V) is a state then we
define unambiguously M[[S]]d = M[[S]]id ; that is, we assume that the interpretation i
is Herbrand if d is a Herbrand state. Also, if ρ is a path through a schema, we may
write M[[ρ]]e to mean M[[schema(ρ)]]e .
Observe that M[[S1 S2 ]]id = M[[S2 ]]iM[[S1 ]]i and
d
πS1 S2 (i, d) = πS1 (i, d)πS2 (i, M[[S1 ]]id )
10
hold for all schemas (not just predicate-free ones).
Given a schema S and µ ∈ pre(Π(S)), we say that µ passes through a predicate term
p(t) if µ has a prefix µ′ ending in p(x), Y for y ∈ {T, F} such that M[[schema(µ′)]]e (x) =
t holds. In this case we say that p(t) = Y is a consequence of µ. For example, the
path u := h() p(w), T v := f (u) of the schema in Figure 1 passes through the predicate
term p(w) since this path has no assignments to w before p.
Definition 7 (path compatibility and executability) Let ρ be a path through a
schema S. Then ρ is executable if ρ is a prefix of πS (j, d) for some interpretation j and
state d. Two paths ρ, ρ′ through schemas S, S ′ are compatible if for some interpretation
j and state d, they are prefixes of πS (j, d) and πS ′ (j, d) respectively.
The justification for restricting ourselves to consideration of Herbrand interpretations
and the state e as the initial state lies in the fact that Herbrand interpretations are
the ‘most general’ of interpretations. Theorem 8, which is virtually a restatement of
[36, Theorem 4-1], expresses this formally.
Theorem 8 Let χ be a set of schemas, let D be a domain, let d be a function from
the set of variables into D and let i be an interpretation using this domain. Then there
is a Herbrand interpretation j such that the following hold.
(1) For all S ∈ χ, the path πS (j, e) = πS (i, d).
(2) If S1 , S2 ∈ χ and v1 , v2 are variables and ρk ∈ pre(πSk (j, e)) for k = 1, 2 and
M[[ρ1 ]]e (v1 ) = M[[ρ2 ]]e (v2 ), then also M[[ρ1 ]]id (v1 ) = M[[ρ2 ]]id (v2 ) holds.
As a consequence of Part (1) of Theorem 8, it may be assumed in Definition 7 that
d = e and the interpretation j is Herbrand without strengthening the Definition. In
the remainder of the paper we will assume that all interpretations are Herbrand.
3
The path-faithful dynamic slicing criterion
In this section we adapt the notion of a dynamic program slice to program schemas.
Dynamic program slicing is formalised in the original paper by Korel and Laski [15].
Their definition uses two functions, F ront and DEL, in which F ront(T, i) denotes the
first i elements of a trajectory 1 T and DEL(T, π) denotes the trajectory T with all
elements that satisfy predicate π removed. A trajectory is a path through a program,
where each node is represented by a line number and so for path ρ we have that ρ̂ is
the corresponding trajectory.
1
A trajectory is a path in which we do not distinguish between true and false values for a
predicate. There is a one-to-one correspondence between paths and trajectories unless there
is an if statement that contains only skip.
11
Korel and Laski use a slicing criterion that is a tuple c = (x, I q , V ) in which x is the
program input being considered, I q denotes the execution of statement I as the qth
statement in the path taken when p is executed with input x, and V is the set of
variables of interest.
The following is the definition provided 2 :
Definition 9 Let c = (x, I q , V ) be a slicing criterion of a program p and T the
trajectory of p on input x. A dynamic slice of p on c is any executable program p′ that
is obtained from p by deleting zero or more statements such that when executed on
input x, produces a trajectory T ′ for which there exists an execution position q ′ such
that
• (KL1) F ront(T ′ , q ′ ) = DEL(F ront(T, q), T (i) 6∈ N ′ ∧ 1 ≤ i ≤ q),
• (KL2) for all v ∈ V , the value of v before the execution of instruction T (q) in T
equals the value of v before the execution of instruction T ′ (q ′ ) in T ′ ,
• (KL3) T ′ (q ′ ) = T (q) = I,
where N ′ is a set of instructions in p′ .
In producing a dynamic slice all we are allowed to do is to eliminate statements. We
have the requirement that the slice and the original program produce the same value
for each variable in the chosen set V at the specified execution position and that the
path in p′ up to q ′ followed by using input x is equivalent to that formed by removing
from the path T all elements not in the slice. Interestingly, it has been observed that
this additional constraint, that F ront(T ′ , q ′ ) = DEL(F ront(T, q), T (i) 6∈ N ′ ∧ 1 ≤
i ≤ q), means that a static slice is not necessarily a valid dynamic slice [37].
We can now give a corresponding definition for linear schemas.
Definition 10 (path-faithful dynamic slice) Let S be a linear schema containing
a label l, let V be a set of variables and let ρ l ∈ pre(Π(S)) be executable. Let S ′ be
a quotient of S containing l. Then we say that S ′ is a (ρl, V )-path-faithful dynamic
slice (PFDS) of S if the following hold.
(1) Every variable in V defines the same term after proj S ′ (ρ) as after ρ in S.
(2) Every maximal path through S ′ which is compatible with ρ has proj S ′ (ρ) as a
prefix.
If the label l occurs at the end of S, so that S = T l for a schema T , and S ′ is a
(ρl, V )-dynamic slice of S, so that S ′ = T ′ l, then we simply say that T ′ is a (ρ, V )path-faithful dynamic end slice of T .
Theorem 11 Let S be a linear schema, let ρl ∈ pre(Π(S)) be executable, let V be a
2
Note that this almost exactly a quote from [15] and is taken from [37]
12
set of variables and let S ′ be a quotient of S containing l. Then S ′ is a (ρl, V )-PFDS
of S if and only if M[[ρ]]e (v) = M[[proj S ′ (ρ)]]e (v) for all v ∈ V and every expression
p(t) = X which is a consequence of proj S ′ (ρ) is also a consequence of ρ.
Proof. This follows immediately from the two conditions in Definition 10.
As an example of a path-faithful dynamic end slice, consider the linear schema of
Figure 3. We assume that V = {v} and the path
ρ = ( p, T g f q, T h H )2 p, F
which passes twice through the body of p, in each case passing through q, T, and then
leaves the body of p. Thus the value of v after ρ is f (h(u)). Thus any ({v}, ρ)-DPS S ′
of S must contain f and h in order that (1) is satisfied, and hence contains p and q.
By Theorem 11, S ′ would also have to contain g, since otherwise p(w) = F would be a
consequence of proj S ′ (ρ), whereas p(w) = F is not a consequence of ρ. Also, S ′ would
contain the function symbol H, since otherwise q(g(w), t) = T would be a consequence
of proj S ′ (ρ), but not of ρ. Thus S itself is the only ({v}, ρ)-PFDS of S. Observe that
the inclusion of the assignment t := H(t); has the sole effect of ensuring that for every
interpretation i for which πS ′ (i, e) = ρ, πS ′ (i, e) passes through q, T instead of q, F
during its second passing through the body of p, and so deleting t := H(t); does not
alter the value of v after πS ′ (i, e). This suggests that our definition of a dynamic slice
may be unnecessarily restrictive, and this motivates the generalisation of Definition
14.
while p(w) {
w := g(w);
v := f (u);
if q(w, t) then u := h(u);
t := H(t);
}
Fig. 3. A linear schema with distinct minimal dynamic and path-faithful dynamic slices
4
A New Form of Dynamic Slicing
Path-faithful dynamic slices of schemas correspond to dynamic program slices and
in order to produce a dynamic slice of a program we can produce the path-faithful
dynamic slice of the corresponding linear schema. In this section we show how this
notion of dynamic slicing can be weakened, to produce smaller slices, for linear schemas
and so also for programs.
13
Consider the schema in Figure 3, the path ρ = p, T g f q, T h H p, T g f q, T h H p, F
and variable v. It is straightforward to see that a dynamic slice has to retain the
predicate p since it controls a statement (u := h(u)) that updates the value of u and
this can lead to a change in the value of v on the next iteration of the loop. Thus, a
dynamic slice with regards to v and ρ must retain predicate q. Further, the assignment
t := H(t) affects the value of t and so the value of q on the second iteration of the loop
in ρ and so a (path-faithful) dynamic slice must retain this assignment.
We can observe that in ρ the value of the predicate q on the last iteration of the loop
does not affect the final value of v. In addition, in ρ the assignment t := H(t) only
affects the value of q on the last iteration of the loop and this assignment does not
influence the final value of v. In this section we define a type of dynamic slice that
allows us to eliminate this assignment. At the end of this section we describe a context
in which we might be happy to eliminate such assignments.
Proposition 12 Let S be a linear schema and let ρ be a path through S.
(1) Let q be a while predicate in S and let µ be a terminal path in the body of q in
S. Then a word αq, Tµq, Fγ is a path in S if and only if αq, Fγ is a path in S.
(2) Let q be an if predicate in S, let Z ∈ {T, F} and let µ, µ′ be terminal paths in the
Z-part and ¬Z-part respectively of q in S. Then a word αq, Zµγ is a path in S
if and only if αq, ¬Zµ′ γ is a path in S.
Furthermore, in both cases, one path is terminal if and only if the other is terminal.
Proof. Both assertions follow straightforwardly by structural induction from the definition of Π(S) in Section 2.1.
Definition 13 Let S be a linear schema, let l be a label and let ρ, ρ′ be paths through
S. Then we say that ρ is simply l-reducible to ρ′ if ρ′ can be obtained from ρ by one
of the following transformations, which we call simple l-reductions.
(1) Replacing a segment p, T σ p, F within ρ by p, F, where σ is a terminal path in
the body of a while predicate p which does not contain l in its body.
(2) Replacing a segment p, Z σ within ρ by p, ¬Z, where σ is a terminal path in the
Z-part of an if predicate p, l does not lie in either part of p and the ¬Z-part of
p is skip.
If ρ′ can be obtained from ρ by applying zero or more l-reductions, then we say that ρ
is l-reducible to ρ′ . If the condition on the label l is removed from the definition then
we use the terms reduction and simple reduction.
By Proposition 12, the transformations given in Definition 13 always produce paths
through S. Observe that if ρ is l-reducible to ρ′ , then the sequence of function and
predicate symbols through which ρ′ passes is a subsequence of that through which ρ
14
passes, ρ and ρ′ pass through the label l the same number of times, and the length of
ρ′ is not greater than that of ρ.
Definition 14 (dynamic slice) Let S be a linear schema containing a label l, let V
be a set of variables and let ρ l ∈ pre(Π(S)) be executable. Let S ′ be a quotient of S
containing l. Then we say that S ′ is a (ρl, V )-dynamic slice (DS) of S if every maximal
path through S ′ compatible with ρ has a prefix ρ′ to which proj S ′ (ρ) is l-reducible
and such that every variable in V defines the same term after ρ′ as after ρ in S.
If the label l occurs at the end of S, so that S = T l for a schema T , and S ′ is a (ρl, V )dynamic slice of S, so that S ′ = T ′ l, then we simply say that T ′ is a (ρ, V )-dynamic
end slice of T .
Consider again the schema in Figure 3 and path ρ = p, T g f q, T h H p, T g f q, T h H p, F.
Here the quotient T obtained from S by deleting the assignment t := H(t); is a (ρ, v)dynamic end slice of S, since the path
ρ′ = p, T g f q, T h H p, T g f q, F p, F
is simply reducible from proj S ′ (ρ) and gives the correct final value for v, and ρ′ and
proj S ′ (ρ) are the only maximal paths through S ′ that are compatible with ρ. This
shows that a DS of a linear schema may be smaller than a PFDS.
One area in which it is useful to determine the dependence along a path in a program is
in the application of test techniques, such as those based on evolutionary algorithms,
that automate the generation of test cases to satisfy a structural criterion. These
techniques may choose a path to the point of the program to be covered and then
attempt to generate test data that follows the path (see, for example, [38,39,40,41]).
If we can determine the inputs that are relevant to this path then we can focus on
these variables in the search, effectively reducing the size of the search space. Current
techniques use static slicing but there is potential for using dynamic slicing in order to
make the dependence information more precise and, in particular, the type of dynamic
slice defined here.
5
A linear schema with two minimal path-faithful dynamic slices
Given a linear schema, a variable set V and a path ρ through S, we wish to establish
information about the set of all (ρ, V )-dynamic slices, which is partially ordered by
set-theoretic inclusion of function and predicate symbols. In particular, it would be
of interest to obtain conditions on S which would ensure that minimal slices were
unique since under such conditions it may be feasible to produce minimal slices in
an incremental manner, deleting one statement at a time until no more statements
can be removed. As we now show, however, this is false for arbitrary linear schemas,
whether or not slices are required to be path-faithful. To see this, consider the schema
15
while P (v) {
if Q(v)
then {
if q(v) then
{
x := ggood ();
v := Ggood (x, v);
}
else
{
x := gbad ();
v := Gbad (x, v);
}
if s1 (v) then
x := g1 ();
if s2 (v) then
x := g2 ();
if t(x)
then v := H(v);
}
else skip;
v := J(v);
}
Fig. 4. A linear schema with distinct minimal path-faithful dynamic slices
S of Figure 4 and the slicing criterion defined by the variable v and the terminal path
ρ which enters the body of P 5 times as follows.
1st time; ρ passes through ggood and H, but not through either gi .
2nd time; ρ passes through ggood , g1 and H, but not through g2 .
3rd time; ρ passes through ggood , g2 and H, but not through g1 .
4th time; ρ passes through gbad , g1 , g2 and H.
5th time; ρ passes through Q, F.
Define the quotient S1 of S by deleting the entire if statement guarded by s2 and define
S2 analogously by interchanging the suffices 1 and 2. By Theorem 11, S1 and S2 are
both (ρ, v)-PFDS’s of S, since t(x) will still evaluate to T over the path proj S1 (ρ) or
proj S2 (ρ) on paths 2–4. On the other hand, if the if statements guarded by s1 and
16
s2 are both deleted, then on the 4th path, t(x) may evaluate to F, since gbad never
occurs in the predicate term defined by t(x) along ρ, hence the final value of v may
contain fewer occurrences of H in the slice than after ρ. Furthermore, every (ρ, v)-DS
of S must contain the function symbols J, H, Ggood and Gbad and hence ggood and
gbad , since the final term defined by v contains these symbols, and so S1 and S2 are
minimal (ρ, v)-DS’s, and are also both path-faithful.
6
Decision problems for dynamic slices
In this section, we establish complexity bounds for two problems; whether a quotient
S ′ of a linear schema S is a dynamic slice, and whether a linear schema S has a nontrivial dynamic slice. We consider the problems both with and without the requirement
that dynamic slices be path-faithful.
Definition 15 (maximal common prefix of a pair of words) The maximal common prefix of words σ, σ ′ is denoted by maxpre(σ, σ ′ ). For example, the maximal common prefix of the words x1 x2 x3 x4 and x1 x2 yx4 over the five-word alphabet {x1 , x2 , x3 , x4 , y}
is x1 x2 ; that is, maxpre(x1 x2 x3 x4 , x1 x2 yx4 ) = x1 x2 .
Lemma 16 Let S be a linear schema containing a label l and let ρ, ρ′ be paths through
S. Suppose ρ is l-reducible to ρ′ . Then there is a sequence ρ1 = ρ, . . . , ρn = ρ′ such
that each ρi is simply l-reducible to ρi+1 , and maxpre(ρi , ρi+1 ) is always a strict prefix
of maxpre(ρi+1 , ρi+2 ).
Proof. This follows from the fact that the two transformation types commute. Since
ρ is l-reducible to ρ′ , there is a sequence ρ1 = ρ, . . . , ρn = ρ′ such that each ρi is
obtained from ρi−1 by a simple l-reduction, and we may assume that n is minimal.
Thus for each i < n, and using the definition of a simple l-reduction, we can write
ρi = αi pi , Zi βi γi and ρi+1 = αi pi , ¬Zi γi . If every αi is a strict prefix of αi+1 , then the
sequence of paths ρi already satisfies the required property. Thus we may assume that
for some minimal i, αi is not a strict prefix of αi+1 .
We now compare the two ways of writing
ρi+1 = αi pi , ¬Zi γi = αi+1 pi+1 , Zi+1βi+1 γi+1 .
Clearly αi+1 is a prefix of αi . We consider three cases.
(1) Suppose that αi = αi+1 . Thus the first letter of ρi+1 after αi is pi , ¬Zi = pi+1 , Zi+1.
If pi = pi+1 were a while predicate, then Zi = T would follow from the fact that ρi
is l-reducible to ρi+1 , and Zi+1 = T would follow similarly from the pair ρi+1 , ρi+2 ,
giving a contradiction, hence pi must be an if predicate and so the ¬Zi -part and
the Zi = ¬Zi+1 -part of p is skip from the definition of l-reduction and hence
ρi+2 = ρi holds, contradicting the minimality of n.
17
(2) Assume that αi+1 is a strict prefix of αi and that αi pi , ¬Zi is a prefix of
αi+1 pi+1 , Zi+1 βi+1 . Thus pi , ¬Zi occurs in βi+1 , and we can write
αi = αi+1 pi+1 , Zi+1 δ1 , βi+1 = δ1 pi , ¬Zi δ2 and since ρi can be obtained by replacing pi , ¬Zi by pi , Zi βi after αi in ρi+1 ,
ρi = αi+1 pi+1 , Zi+1 δ1 pi , Zi βi δ2 γi+1
follows. By our assumption on the pair (ρi , ρi+1 ), βi is a terminal path in the body
or Zi -part of pi and so by Proposition 12, δ1 pi , Zi βi δ2 is a terminal path in the
body or Zi+1 -part of pi+1 and so ρi+2 is obtainable from ρi by a simple l-reduction,
by replacing pi+1 , ¬Zi+1 δ1 pi , Zi βi δ2 by pi+1 , ¬Zi+1 in ρi , again contradicting the
minimality of n.
(3) Lastly, assume that αi+1 is a strict prefix of αi and that αi pi , ¬Zi is not a prefix of αi+1 pi+1 , Zi+1 βi+1 . Thus we can write αi = αi+1 pi+1 , Zi+1 βi+1 δ. We now
change the order of the two reductions by replacing ρi+1 in the sequence by
ρ̂i+1 = αi+1 pi+1 , ¬Zi+1 δpi , Zi βi γi , which by two applications of Proposition 12,
is a path through S. In effect we are replacing pi+1 , Zi+1 βi+1 by pi+1 , ¬Zi+1 before replacing pi , Zi βi by pi , ¬Zi , instead of in the original order. Since ρi+2 =
αi+1 pi+1 , ¬Zi+1 δpi , ¬Zi γi , maxpre(ρi , ρ̂i+1 ) is a strict prefix of maxpre(ρ̂i+1 , ρi+2 ).
Thus, by the minimality of i, after not more than n − i such replacements, the
maximal common prefixes of consecutive paths in the resulting sequence will be
strictly increasing in length, as required.
Theorem 17 Let S be a linear schema, let l be a label and let ρ, ρ′ ∈ pre(Π(S)).
Then it is decidable in polynomial time whether ρ is l-reducible to ρ′ .
Proof. By Lemma 16, ρ is l-reducible to ρ′ if and only if ρ can be simply l-reduced to
some ρ2 ∈ pre(Π(S)) such that ρ2 is l-reducible to ρ′ and maxpre(ρ, ρ2 ) is a strict prefix
of maxpre(ρ2 , ρ′ ) and hence maxpre(ρ, ρ2 ) = maxpre(ρ, ρ′ ). Thus ρ2 exists satisfying
these criteria if and only if ρ and ρ′ have prefixes τ σ and τ σ ′ respectively such that
σ ′ is obtained from σ by either of the transformations given in Definition 13, and ρ2
is obtained from ρ by replacing σ by σ ′ . Thus σ can be computed in polynomial time
if it exists, and this procedure can be iterated using ρ2 in place of ρ. The number of
iterations needed is bounded by the number of letters in ρ′ , thus proving the Theorem.
Theorem 18 Let S be a linear schema containing a label l, let ρl ∈ pre(Π(S)) be
executable, let V be a set of variables and let S ′ be a quotient of S containing l.
(1) The problem of deciding whether S ′ is a (ρl, V )-path-faithful dynamic slice of S
lies in polynomial time.
(2) The problem of deciding whether S ′ is a (ρl, V )-dynamic slice of S lies in co-NP.
Proof. (1) follows immediately from the conclusion of Theorem 11, since given any
18
predicate-free schema T and any variable v, the term M[[T ]]e (v) is computable in
polynomial time.
To prove (2), we proceed as follows. Any path ρ′ l through S ′ such that ρ′ is l-reducible
from proj S ′ (ρ) has length ≤ |proj S ′ (ρ)l|. We compute a path τ through S ′ of length
≤ |proj S ′ (ρ)l|, with strict inequality if and only if τ is terminal. This can be done
in NP-time by starting with the empty path and successively appending letters to it
until a terminal path, or one of length |proj S ′ (ρ)l| is obtained. We then test whether
τ is compatible with ρ and does not have a prefix ρ′ l through S ′ such that ρ′ is lreducible from proj S ′ (ρ) and M[[ρ]]e (v) = M[[ρ′ ]]e (v) for all v ∈ V . By Theorem 17,
this can be done in polynomial time. If no such prefix exists for the given τ , then no
longer path through S ′ having prefix τ has such a prefix either, and hence S ′ is not a
(ρl, V )-dynamic slice of S. Conversely, if S ′ is not a (ρl, V )-dynamic slice of S, then
a path τ can be computed satisfying the conditions given, proving (2).
Theorem 19 Let S be a linear schema, let ρl ∈ pre(Π(S)) be executable and let V
be a set of variables.
(1) The problem of deciding whether there exists a non-trivial (ρl, V )-path-faithful
dynamic slice of S is NP-complete.
(2) The problem of deciding whether there exists a non-trivial (ρl, V )-dynamic slice
of S lies in PSPACE and is NP-hard.
Proof. To prove membership in NP for Problem (1), it suffices to observe that a
quotient S ′ of S can be guessed in NP-time, and using Theorem 11, it can be decided
in polynomial time whether S ′ is a non-trivial (ρl, V )-path-faithful dynamic slice of S.
Membership of Problem (2) in PSPACE follows similarly from Part (2) of Theorem
18 and the fact that co-NP⊆PSPACE=NPSPACE.
To show NP-hardness of both problems, we use a polynomial-time reduction from
3SAT, which is known to be an NP-hard problem [42]. An instance of 3SAT comprises
V
a set Θ = {θ1 , . . . , θn } and a propositional formula α = m
k=1 αk1 ∨ αk2 ∨ αk3 , where
each αij is either θk or ¬θk for some k. The problem is satisfied if there exists a
valuation δ : Θ → {T, F} under which α evaluates to T. We will construct a linear
schema S containing a variable v and a terminal path ρ through S such that S has
a non-trivial (ρ, v)-dynamic end slice if and only if α is satisfiable, in which case this
quotient is also a (ρ, v)-path-faithful dynamic end slice. The schema S is as in Figure
5.
We say that the function symbol gi corresponds to θi and gi′ corresponds to ¬θi . The
terminal path ρ passes a total of 4 + 3n + 6n(n − 1) + m times through the body of S,
and then leaves the body. The paths within the body of S are of fourteen types, and
are listed as follows, in the order in which they occur along ρ; note that only those of
19
while p(v) {
v := H(v);
if qgood (v) then x := ggood ();
if qbad (v) then x := gbad ();
if qlink (v)
then b := glink (x);
if qreset (v)
then b := greset();
if Qlink/reset (v) then v := Flink/reset(b, v);
if q1 (v)
then x := g1 (b);
if q1′ (v)
..
.
then x := g1′ (b);
if qn (v)
then x := gn (b);
if qn′ (v)
then x := gn′ (b);
if Qtest (v)
then if qtest (x) then v := Ftest (v);
}
Fig. 5.
type (5) depend on the value of α. The total number of paths of each type is given in
parentheses at the end.
(0)
(0.1) ρ passes through ggood , glink , and Flink/reset, and through no other assignment
apart from H.
(0.2) ρ passes through greset, and Flink/reset, and through no other assignment apart
from H.
(0.3) ρ passes through gbad , glink , and Flink/reset, and through no other assignment
apart from H.
(3 paths)
(1) ρ passes through ggood and Ftest , and through no other assignment apart from H.
(1 path)
(2) For each i ≤ n, ρ passes through ggood , greset, gi and Ftest and through no other
assignment apart from H. (n paths)
(2′ ) As for type (2), but with gi′ in place of gi . (n paths)
(3) For each i ≤ n, ρ passes through ggood , glink , gi′ and Ftest and through no other
20
assignment apart from H. (n paths)
(4) For each i 6= j ≤ n, ρ passes 3 times consecutively through the body of S, as
follows;
(4.1) The first time, it passes through ggood , greset , and gi , but not through qtest or
any other assignment apart from H.
(4.2) The 2nd time, it passes through glink and gi′ , but not through qtest or any
other assignment apart from H.
(4.3) The 3rd time, it passes through greset and gj and Ftest , but through no other
assignment apart from H.
(3n(n − 1) paths)
(4.1′), (4.2′), (4.3′ ) As for types (4.1),(4.2),(4.3), but with gj′ in place of gj . (3n(n − 1)
paths)
(5) For each i ≤ m, ρ passes through gbad and greset, and then through the 3 function
symbols corresponding to the implicants αi1 , αi2 , αi3 , and then through Ftest and
through no other assignment apart from H. (m paths)
Before continuing with the proof, we first record the following facts about the terminal
path ρ.
(a) ρ passes through all three assignments to v and through both assignments to b.
(b) All three assignments to v in S also reference v, and hence if there exists a
terminal path σ through any slice T of S such that M[[ρ]]e (v) = M[[σ]]e (v), then
the following hold;
(b0) By (a), T contains H, Ftest , Flink/reset and hence glink , greset, ggood and gbad
because of the type (0) paths, and thus contains the predicates controlling
these function symbols.
(b1) By (a), σ passes through all the assignments to v in S in the same order as
ρ does.
(b2) σ and ρ enter the body of p the same number of times, namely the depth of
the nesting of H in the term M[[ρ]]e (v).
(b3) For any function symbol f in S assigning to v and for all k ≥ 0, v defines
the same term after the kth occurrence of f in ρ and σ, since this term is
the unique subterm of M[[ρ]]e (v) containing k nested occurrences of f whose
outermost function symbol is f .
(b4) For any predicate q in T and for all k ≥ 0, σ and ρ pass the same way
through q at the kth occurrence of q. For q 6= qtest , this follows from (b3)
applied to H or Flink/reset. For q = qtest , it follows from (b1) and (b4) applied
to Qtest .
(b5) proj T (ρ) = σ. For assume σ ′ q, Z ∈ pre(σ), whereas σ ′ q, ¬Z ∈ pre(proj T (ρ)),
where σ ′ q, Z contains k q’s; this contradicts (b4) immediately, and hence
proj T (ρ) = σ follows from Lemma 3 and the fact that proj T (ρ) and σ are
both terminal paths through T .
(c) ρ never passes through the predicate terms qtest (gbad ()) or qtest (gi′ (glink (gi (greset())))).
(d) For any prefix ρ′ of ρ, the term M[[ρ′ ]]e (v) does not contain any gi or gi′ ; for these
symbols, which do not occur on the type (0) paths, assign to x, whereas Flink/reset,
21
which does not occur on ρ after the type (0) paths, is the only assignment to v
referencing a variable other than v.
• (⇒). Let T be a non-trivial (ρ, v)-DS of S. By (b5), T is a (ρ, v)-PFDS of S and
by (b0), T contains all symbols in S apart possibly from some of those of the form
gi , gi′ and the if predicates qi , qi′ controlling them. Thus it remains only to show that
α is satisfiable.
We first show that if T does not contain a symbol gj , then for all i 6= j, it cannot
contain both gi and gi′ . Consider the type (4.3) path for the values i, j. If T contains
gi and gi′ , but not gj , then when qtest is reached on path (4.3), the predicate term thus
defined, built up over paths (4.1), (4.2), (4.3), is qtest (gi′ (glink (gi (greset())))), which
does not occur along the path ρ, contradicting Theorem 11. By considering type
(4.3′) paths the same assertion holds for the symbols gj′ . Since T 6= S holds, this
implies that T contains at most one element of each set {gi , gi′ }.
We now show that for each i ≤ m, T contains at least one symbol corresponding
to an element in {αi1 , αi2 , αi3 }. If this is false, then the predicate term qtest (gbad ()),
which does not occur along the path ρ, would be defined on the ith type (5) path,
contradicting Theorem 11.
Thus α is satisfied by any valuation δ such that for all i ≤ n, T contains gi
⇒ δ(θi ) = T and T contains gi′ ⇒ δ(¬θi ) = T; since T contains at most one element
of each set {gi , gi′ }, such a valuation exists.
• (⇐). Conversely, suppose that α is satisfiable by a valuation δ : Θ → {T, F}, and
let T be the quotient of S which contains each gi and qi if and only if δ(θi ) = T, and
containing gi′ and qi′ otherwise, and contains all the other symbols of S. We show
that T is a (ρ, v)-DPS of S. By Theorem 11, it suffices to show that all predicate
terms occurring along proj T (ρ) also occur along ρ with the same associated value
from {T, F}, since by (d), M[[proj T (ρ)]]e (v) = M[[ρ]]e (v).
By (d), all predicate terms occurring in proj T (ρ) but not ρ must occur at qtest
rather than at a predicate referencing v. We consider each path type separately and
show that no such predicate terms exist.
(0) These paths do not pass through qtest .
(1) proj T (ρ) defines qtest (ggood ()), which also occurs along ρ in the type (1) path.
(2) If T does not contain gi then proj T (ρ) defines qtest (ggood ()), which occurs along ρ
in the type (1) path. Otherwise proj T (ρ) defines qtest (gi (greset())), which occurs
along ρ in a type (2) path.
(2′ ) Similar to type (2).
(3) If T does not contain gi′ then proj T (ρ) defines qtest (ggood ()), which occurs along
ρ in the type (1) path. Otherwise proj T (ρ) defines qtest (gi′ (glink (ggood ()))), which
also occurs along ρ in a type (3) path.
(4) If T contains gi but not gj or gi′ then proj T (ρ) defines qtest (gi (greset())), which
occurs along ρ in a type (2) path. If T contains gi′ but not gj or gi then proj T (ρ)
defines qtest (gi′ (glink (ggood ()))), which also occurs along ρ in a type (3) path. Lastly,
if T contains gj then proj T (ρ) defines qtest (gj (greset ())), which occurs along ρ in a
type (2) path.
(4′ ) Similar to type (4).
22
(5) Since the valuation δ satisfies α, for each k ≤ m, T contains at least one of
the 3 function symbols corresponding to the implicants αk1 , αk2 , αk3 , and hence
proj T (ρ) defines qtest (gi (greset())) or qtest (gi′ (greset())) for some i ≤ n, which occur
along ρ in a type (2) or (2′ ) path.
Since the schema S and the path ρ can clearly constructed in polynomial time from
the formula α, this concludes the proof of the Theorem.
7
Conclusion and further directions
We have reformulated Korel and Laski’s definition of a dynamic slice of a program as
applied to linear schemas, which is the normal level of program abstraction assumed by
slicing algorithms, and have also given a less restrictive slicing definition. In addition,
we have given P and co-NP complexity bounds for the problem of deciding whether
a given quotient of a linear schema satisfies them. We conjecture that the problem
of whether a quotient S ′ of a linear schema S is a general dynamic slice with respect
to a given path and variable set is co-NP-complete. Future work should attempt to
resolve this.
We have also shown that it is not possible to decide in polynomial time whether a
given linear schema has a non-trivial dynamic slice using either definition, assuming
P6=NP. It is possible that this NP-hardness result can be strengthened to PSPACEhardness for general dynamic slices, since in this case the problem does not appear to
lie in NP.
We have also shown that minimal dynamic slices (whether or not path-faithful) are
not unique. Placing further restrictions on either the schemas or the paths may ensure
uniqueness of dynamic slices or lower the complexity bounds proved in Section 6, and
this should be investigated.
Schemas correspond to single programs/methods and so results regarding schemas
cannot be directly applied when analysing a program that has multiple procedures
and thus the results in this paper do not apply to inter-procedural slicing. It would be
interesting to extend schemas with procedures and then analyse both dynamic slicing
and static slicing for such schemas.
These results have several practical ramifications. First, since the problem of deciding
whether a linear schema has a non-trivial dynamic slice is computationally hard this
result must also hold for programs. A further consequence is that the problem of producing minimal dynamic slices must also be computationally hard. We also defined
a new notion of a dynamic slice for linear schemas (and so for programs) that places
strictly weaker constraints on the slice and so can lead to smaller dynamic slices.
23
Finally, the fact that minimal dynamic slices need not be unique suggests that algorithms that identify and then delete one statement at a time can lead to suboptimal
dynamic slices.
References
[1] S. Greibach, Theory of program structures: schemes, semantics, verification, Vol. 36 of
Lecture Notes in Computer Science, Springer-Verlag Inc., New York, NY, USA, 1975.
[2] A. De Lucia, A. R. Fasolino, M. Munro, Understanding function behaviours through
program slicing, in: 4th IEEE Workshop on Program Comprehension, IEEE Computer
Society Press, Los Alamitos, California, USA, Berlin, Germany, 1996, pp. 9–18.
[3] M. Harman, R. M. Hierons, S. Danicic, J. Howroyd, C. Fox, Pre/post conditioned
slicing, in: IEEE International Conference on Software Maintenance (ICSM’01), IEEE
Computer Society Press, Los Alamitos, California, USA, Florence, Italy, 2001, pp. 138–
147.
[4] G. Canfora, A. Cimitile, A. De Lucia, G. A. D. Lucca, Software salvaging based on
conditions, in: International Conference on Software Maintenance (ICSM’96), IEEE
Computer Society Press, Los Alamitos, California, USA, Victoria, Canada, 1994, pp.
424–433.
[5] A. Cimitile, A. De Lucia, M. Munro, A specification driven slicing process for identifying
reusable functions, Software maintenance: Research and Practice 8 (1996) 145–178.
[6] K. B. Gallagher, Evaluating the surgeon’s assistant: Results of a pilot study, in:
Proceedings of the International Conference on Software Maintenance, IEEE Computer
Society Press, Los Alamitos, California, USA, 1992, pp. 236–244.
[7] K. B. Gallagher, J. R. Lyle, Using program slicing in software maintenance, IEEE
Transactions on Software Engineering 17 (8) (1991) 751–761.
[8] H. Agrawal, R. A. DeMillo, E. H. Spafford, Debugging with dynamic slicing and
backtracking, Software Practice and Experience 23 (6) (1993) 589–616.
[9] M. Kamkar, Interprocedural dynamic slicing with applications to debugging and
testing, PhD Thesis, Department of Computer Science and Information Science,
Linköping University, Sweden, available as Linköping Studies in Science and Technology,
Dissertations, Number 297 (1993).
[10] J. R. Lyle, M. Weiser, Automatic program bug location by program slicing, in: 2nd
International Conference on Computers and Applications, IEEE Computer Society
Press, Los Alamitos, California, USA, Peking, 1987, pp. 877–882.
[11] M. Weiser, J. R. Lyle, Experiments on slicing–based debugging aids, Empirical studies
of programmers, Soloway and Iyengar (eds.), Molex, 1985, Ch. 12, pp. 187–197.
[12] M. Weiser, Program slicing, IEEE Transactions on Software Engineering 10 (4) (1984)
352–357.
24
[13] S. Danicic, C. Fox, M. Harman, R. Hierons, J. Howroyd, M. R. Laurence, Static program
slicing algorithms are minimal for free liberal program schemas, The Computer Journal
48 (6) (2005) 737–748.
[14] M. R. Laurence, Characterising minimal semantics-preserving slices of function-linear,
free, liberal program schemas, Journal of Logic and Algebraic Programming 72 (2)
(2005) 157–172.
[15] B. Korel, J. Laski, Dynamic program slicing, Information Processing Letters 29 (3)
(1988) 155–163.
[16] H. Agrawal, J. R. Horgan, Dynamic program slicing, in: Proceedings of the ACM
SIGPLAN ’90 Conference on Programming Language Design and Implementation,
Vol. 25, White Plains, NY, 1990, pp. 246–256.
URL citeseer.ist.psu.edu/agrawal90dynamic.html
[17] A. Beszédes, T. Gergely, Z. M. Szabó, J. Csirik, T. Gyimóthy, Dynamic slicing method
for maintenance of large C programs, in: Proceedings of the Fifth European Conference
on Software Maintenance and Reengineering (CSMR 2001), IEEE Computer Society,
2001, pp. 105–113.
[18] R. Gopal, Dynamic program slicing based on dependence graphs, in: IEEE Conference
on Software Maintenance, 1991, pp. 191–200.
[19] M. Kamkar, N. Shahmehri, P. Fritzson, Interprocedural dynamic slicing, in: PLILP,
1992, pp. 370–384.
[20] M. Kamkar, Application of program slicing in algorithmic debugging, in: M. Harman,
K. Gallagher (Eds.), Information and Software Technology Special Issue on Program
Slicing, Vol. 40, Elsevier, 1998, pp. 637–645.
[21] B. Korel, Computation of dynamic slices for programs with arbitrary control flow, in:
M. Ducassé (Ed.), 2nd International Workshop on Automated Algorithmic Debugging
(AADEBUG’95), Saint–Malo, France, 1995.
[22] B. Korel, J. Rilling, Dynamic program slicing methods, in: M. Harman, K. Gallagher
(Eds.), Information and Software Technology Special Issue on Program Slicing, Vol. 40,
Elsevier, 1998, pp. 647–659.
[23] M. Mller-Olm, Precise interprocedural dependence analysis of parallel programs,
Theoretical Computer Science (TCS) 31 (1) (2004) 325–388.
[24] M. Mller-Olm, H. Seidl, Computing polynomial program invariants, Information
Processing Letters (IPL) 91 (5) (2004) 233–244.
[25] M. Mller-Olm, H. Seidl, Precise interprocedural analysis through linear algebra, in:
Proceedings of Principles of Programming Languages (POPL’04), Venice, Italy, 2004.
[26] M. S. Paterson, Equivalence problems in a model of computation, Ph.D. thesis,
University of Cambridge, UK (1967).
[27] D. C. Luckham, D. M. R. Park, M. S. Paterson, On formalised computer programs, J.
of Computer and System Sciences 4 (3) (1970) 220–249.
25
[28] E. A. Ashcroft, Z. Manna, Translating program schemas to while-schemas, SIAM
Journal on Computing 4 (2) (1975) 125–146.
[29] Y. I. Ianov, The logical schemes of algorithms, in: Problems of Cybernetics, Vol. 1,
Pergamon Press, New York, 1960, pp. 82–140.
[30] J. D. Rutledge, On Ianov’s program schemata, J. ACM 11 (1) (1964) 1–9.
[31] H. B. Hunt, R. L. Constable, S. Sahni, On the computational complexity of program
scheme equivalence, SIAM J. Comput 9 (2) (1980) 396–416.
[32] V. K. Sabelfeld, An algorithm for deciding functional equivalence in a new class of
program schemes, Journal of Theoretical Computer Science 71 (1990) 265–279.
[33] M. R. Laurence, S. Danicic, M. Harman, R. Hierons, J. Howroyd, Equivalence of
conservative, free, linear program schemas is decidable, Theoretical Computer Science
290 (2003) 831–862.
[34] M. R. Laurence, S. Danicic, M. Harman, R. Hierons, J. Howroyd, Equivalence
of linear, free, liberal, structured program schemas is decidable in polynomial
time, Tech. Rep. ULCS-04-014, University of Liverpool, electronically available at
http://www.csc.liv.ac.uk/research/techreports/ (2004).
[35] S. Danicic, M. Harman, R. Hierons, J. Howroyd, M. R. Laurence, Equivalence of linear,
free, liberal, structured program schemas is decidable in polynomial time, Theoretical
Computer Science 373 (1-2) (2007) 1–18.
[36] Z. Manna, Mathematical Theory of Computation, McGraw–Hill, 1974.
[37] D. Binkley, S. Danicic, T. Gyimóthy, M. Harman, Á. Kiss, B. Korel, Theoretical
foundations of dynamic program slicing, Theoretical Computer Science 360 (1–3) (2006)
23–41.
[38] M. Harman, C. Fox, R. M. Hierons, L. Hu, S. Danicic, J. Wegener, Vada: A
transformation-based system for variable dependence analysis, in: SCAM, IEEE
Computer Society, 2002, pp. 55–64.
[39] B. F. Jones, H.-H. Sthamer, D. E. Eyres, Automatic structural testing using genetic
algorithms, The Software Engineering Journal 11 (5) (1996) 299–306.
[40] J. Wegener, K. Grimm, M. Grochtmann, H. Sthamer, B. F. Jones, Systematic testing
of real-time systems, in: 4th International Conference on Software Testing Analysis and
Review (EuroSTAR 96), 1996.
[41] J. Wegener, H. Sthamer, B. F. Jones, D. E. Eyres, Testing real-time systems using
genetic algorithms, Software Quality 6 (2) (1997) 127–135.
[42] S. A. Cook, The complexity of theorem-proving procedures, in: STOC ’71: Proceedings
of the third annual ACM symposium on Theory of computing, ACM, New York, NY,
USA, 1971, pp. 151–158.
26
| 6cs.PL
|
DETERMINISTIC PARALLEL ALGORITHMS FOR FOOLING
POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
arXiv:1610.03383v6 [cs.DS] 29 Jan 2018
DAVID G. HARRIS1
Abstract. Many randomized algorithms can be derandomized efficiently using either the method
of conditional expectations or probability spaces with low (almost-) independence. A series of papers, beginning with Luby (1988) and continuing with Berger & Rompel (1991) and Chari et al.
(1994), showed that these techniques can be combined to give deterministic parallel algorithms
for combinatorial optimization problems involving sums of w-juntas. We improve these algorithms
through derandomized variable partitioning, reducing the processor complexity to essentially independent of w and time complexity to linear in w.
As a key subroutine, we give a new algorithm to generate a probability space which can fool
a given set of neighborhoods, each of size at most w. Schulman (1992) gave an NC algorithm to
do so for w ≤ O(log n). Our new algorithm is NC1 , with essentially optimal time and processor
complexity, when w = O(log n); it remains NC up to w = polylog(n). This answers an open
problem of Schulman.
One major application of these algorithms is an NC algorithm for the Lovász Local Lemma.
Previous NC algorithms, including the seminal algorithm of Moser & Tardos (2010) and the work
of Chandrasekaran et. al (2013), required that (essentially) the bad-events could span only O(log n)
variables; we relax this to polylog(n) variables. We use this for an NC2 algorithm for defective vertex
coloring, which works for arbitrary degree graphs.
This is an extended version of a paper appearing in the Proceedings of the 28th ACM-SIAM
Symposium on Discrete Algorithms (SODA) 2017.
1. Introduction
Many algorithms can be formulated
as optimization problems, in which we seek to maximize a
P
function of the form S(x) = j fj (x) over x ∈ {0, 1}n ; we refer to the summands fj as objective
functions. These may correspond to a scoring function measuring how good our solution is, or they
might be indicators for bad events that we wish to avoid. We will consider cases in which each fj
depends on at most w coordinates of x; this is known as a w-junta.
This often leads to randomized algorithms with the following structure:
if X is drawn from a
P
suitable distribution (say independent fair coins), then E[S(X)] = j E[fj (X)] = S0 . Obviously,
there exists some x ∈ {0, 1}n with the property S(x) ≥ S0 . Usually we can find such an x with a
randomized algorithm, since a “typical” vector x has the property S(x) ≈ S0 .
A key derandomization problem is thus to find such x deterministically. There are two main
paradigms to do so: conditional expectations and low-independence probability spaces. To use
conditional expectations, we gradually assign the bits of X to 0 or 1, ensuring that at each step the
conditional expectation E[S(X)] increases. To solve this by low-independence, we draw the random
variables X from a probability space which has (nearly) the same w-wise marginal distributions as
the independent space {0, 1}n .
Each of these methods has disadvantages. The method of conditional expectations is inherently
sequential: decisions about some xi depend on the assignment of x1 , . . . , xi−1 . The method of lowindependence can easily be parallelized, but doing so requires a separate processor for each element
of the probability space; this leads to large processor counts. A hybrid approach was proposed by
1
Department of Computer Science, University of Maryland, College Park, MD 20742. Research supported in part
by NSF Awards CNS-1010789 and CCF-1422569. Email: [email protected].
1
2
DAVID G. HARRIS
Luby in [11], which can combine parallelism with low processor complexity. The key observation
was that probability spaces with polylog-wise-independence can be described as linear codes over
GF (2), of length polylog(n). One can then apply the method of conditional expectations to the
code itself, not the solution vector directly. This method was formalized and extended by Berger
& Rompel in [4].
The main limitation of this hybrid algorithm is that, at least in its simplest form, it has processor
complexity which is exponential in w. Berger & Rompel describe a limited number of problemspecific techniques to overcome this. In this paper, we will investigate a more general method
of dealing with this computational bottleneck, based on a derandomization of random variable
partitioning.
1.1. Alternate derandomization approaches. We mention three other general approaches to
derandomization, and the ways in which they fail to cover some key applications. The first approach
is to use a probability space which is ǫ-approximately w-wise-independent (see Definition 3.4).
Such a space is significantly smaller than a fully-independent space. If the objective functions
fj were simply monomial functions, or more generally had small decision tree complexity, then
their expectation would differ only slightly between an ǫ-approximately independent and a fully
independent space. However, in many applications, fj may be significantly more complex and the
overall bias can become as large as 2w ǫ — requiring ǫ to be super-polynomially small, and requiring
the probability space to be too large to explore exhaustively.
A second approach is the derandomization technique of Sivakumar [18], when the functions fj
can be computed via compact automata with a polynomial state-space. One can build a relatively
small probability distribution which fools a polynomial number of such automata. However, one
critical aspect of Sivakumar’s method is that the multiple automata all must process the input bits
in the same order. Many applications lack this property, most notably the algorithms for the Lovász
Local Lemma. Another disadvantage of Sivakumar’s method is its high processor complexity (on
the order of n20 or more).
The final approach is one of Schulman [17] for generating a probability space which fools a
given list of neighborhoods (see Section 2.1 for a formal definition). If Ω fools the neighborhoods
corresponding to each fj , then there is guaranteed to exist some x ∈ Ω with S(x) ≥ S0 ; if Ω has
small support then this leads to an efficient algorithm. Although the space Ω can be significantly
smaller support than a fully w-wise-independent space, it is still super-polynomial for w > ω(log n)
and so this approach does not give NC algorithms. Fooling neighborhoods will nonetheless be a
key algorithmic building block of our algorithms, and we will consider it more fully in this paper.
1.2. Our contributions and overview. In Section 2, we present a new algorithm to produce
probability spaces fooling a list of neighborhood or a list of Fourier characters over GF (2) (a closely
related problem). These are important subroutines needed for the algorithmic approach of Berger
& Rompel [4]. The algorithm we develop has essentially optimal time and processor complexity
(polylogarithmic and linear, respectively, in the time to read the input). In particular, when the
neighborhood size w is w = O(log n), then we obtain an NC1 algorithm and when w = polylog(n)
we obtain an NC algorithm.
In addition to their use in the Berger-Rompel framework, these algorithms can be used to solve
some other derandomization problems. For instance, we use these for a near-optimal algorithm to
find a codeword of Hamming weight at least L/2 in a length-L binary code, a toy derandomization
problem introduced by [14].
In Section 3, we consider fooling sums of w-juntas. As we have discussed, the main bottleneck
in the algorithm of [4] is the exponential processor dependency on w. We give an algorithm based
on random variable partitioning, which is then derandomized. This approach makes the processor
complexity independent of w while giving a linear time dependency on w. This allows us to handle,
for the first time in NC, many applications with w = polylog(n).
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
3
We describe a sample application in Section 4 to rainbow hypergraph coloring. Given a d-uniform
hypergraph, we wish to d-color the vertices so that at least md!/dd edges see all d colors (as is expected in a uniform random coloring). This was an example application given in [4]. Although [4]
did not provide concrete complexity bounds, their algorithm appears to require O(log4 m) time and
O(m1+ln 2 ) processors. We reduce this to roughly Õ(log2 m) time and O(m) processors. This illustrates how our derandomization procedure has been optimized for processor and time complexity,
so that it can be beneficial even for applications with prior NC algorithms.
In Section 5 we consider the seminal Moser-Tardos algorithm for the Lovász Local Lemma [12].
In this setting, one seeks to avoid a set of “bad events”, which are boolean functions of the variables
x1 , . . . , xn . There have been some NC versions of this algorithm, appearing in the original paper of
Moser & Tardos along with some extensions in [6, 10]. These algorithms are somewhat limited in
the types of problems they can handle, with restrictive conditions on the decision-tree-complexity
of the bad-events. We greatly expand the scope of these algorithms to give NC algorithms in almost
any application where the bad-events are w-juntas for w = polylog(n).
In Section 6, we apply our LLL derandomization to two graph theory applications. First, defective
vertex coloring: given a graph G of maximum degree ∆, we achieve an Õ(log2 n)-time algorithm
for a k-defective vertex coloring with c = O(∆/k) colors. Notably, although our general LLL
algorithm only applies to bad-events which span a polylogarithmic number of variables (in particular
here, when ∆ ≤ polylog(n)), our coloring algorithm works for arbitrary values of ∆ and k. The
second application is to domatic partition; here we only get an NC algorithm for graphs of degree
k = polylog(n).
1.3. Notation and conventions. All our algorithms will be described in the deterministic EREW
PRAM model. In this model, we say an algorithm A has complexity (C1 , C2 ) if it runs in C1 and
C2 processors. In order to focus on the leading-order terms, we often use a looser metric which
we refer to as quasi-complexity. We say A has quasi-complexity (C1 , C2 ) if it has complexity
1+o(1)
(Õ(C1 ), C2
), where we define Õ(t) = t(log t)O(1) .1
We let [n] denote the set {1, . . . , n}. For any collection of sets E ⊆ 2[n] , we define the width of
E to be width(E) = maxe∈E |e|.
For a probability space Ω, we use the notation |Ω| to mean the cardinality of the support of Ω,
and we define the seed-length of Ω as log2 |Ω|. Given a set X, we write x ∼ X to mean that x
is drawn from the uniform probability distribution on X. For a boolean predicate P , we use the
Iverson notation where [P ] is one if P is true and zero otherwise.
2. Fooling neighborhoods
2.1. Fourier characters, neighborhoods, and codes. Many probability spaces satisfying (approximate) independence conditions are built on top of codes over GF (2). These are closely related
to Fourier characters over GF (2). We begin by reviewing some definitions and basic results on
these objects.
n
Definition 2.1.
P A Fourier character over GF (2) is a function χe : GF (2) → {−1, 1} defined by
x
χe (x) = (−1) i∈e i , for some e ⊆ [n]
For e 6= ∅, one can see easily that EX∼{0,1}n [χe (X)] = 0. We thus say that a probability space
Ω is unbiased for e 6= ∅ if EX∼Ω [χe (X)] = 0. (For e = ∅, we will say that every probability space
is unbiased for e.) We say Ω fools e if Ω is unbiased for every subset f ⊆ e. Likewise, we say Ω
1We note that it is very difficult to obtain estimates which are finer than this; small changes in the computational
model or the input data (for example, whether the registers contain log n bits or single bits, the precisions of the
real-valued weights, which arithmetic operations can be performed atomically, etc.) can change the runtime by
hard-to-track polyloglog factors.
4
DAVID G. HARRIS
is unbiased for (respectively, fools) a list E = {e1 , . . . , em } if Ω is unbiased for (respectively, fools)
each e1 , . . . , em .
Our notation and definitions will differ slightly from the standard use in coding theory. Given a
list of vectors A(1), . . . , A(n) ∈ GF (2)L , we refer to the list A = A(1), . . . , A(n) as a code of length
L and size n. We use the following
notational shortcut throughout: if x ∈ {0, 1}n and e ⊆ [n] is a
L
of n binary vectors and e ⊆ [n], then A(e)
set, then we define x(e) = i∈e xi . If A is a collection
L
is the binary vector defined coordinatewise by i∈e A(i).
Proposition 2.2. For any boolean function g : {0, 1}n → R, there are weights γe ∈ R, where e
ranges over all 2n subsets of [n], such that for all x ∈ {0, 1}n we have
X
g(x) =
χe (x)
e⊆[n]
The weights γ can be determined with quasi-complexity (n, 2n ).
P
Proof. This is the Discrete Fourier Transform over GF (2). Define γe = 2−n y∈{0,1}n χe (y)g(y);
these weights can be computed efficiently using the Fast Walsh-Hadamard Transform, a well-known
algorithm.
Proposition 2.3. If a probability space Ω fools e ⊆ [n], then for all z ∈ {0, 1}n we have
^
Px∼Ω ( xi = zi ) = 2−|e|
i∈e
V
Proof. Let g(x) = [ i∈e xi = zP
i ]. By Proposition 2.2, there exist weights γf , where f ranges over
subsets of e, such that g(x) = f ⊆e γf χf (x). Then
X
X
EX∼Ω [g(X)] =
γf EX∼Ω [χf (X)] =
γf EX∼{0,1}n [χf (X)] = EX∼{0,1}n [g(X)] = 2−|e| .
f ⊆e
f ⊆e
The main connection between codes, Fourier characters, and fooling neighborhoods comes from
the following construction:
Definition 2.4. Given a code A(1), . . . , A(n) of length L, define the probability space ΩA as follows:
draw a vector y ∼ GF (2)L , and set Xi = A(i) · y for i = 1, . . . , n, where · is the inner product over
GF (2)L . Note that |ΩA | = 2L .
Definition 2.5 (E-unbiased code). The code A is an E-unbiased code if A(e) 6= ~0 for all non-empty
sets e ∈ E.
Proposition 2.6. If A is an E-unbiased code, then ΩA is unbiased for every e ∈ E.
Proof. Let e ∈ E with e 6= ∅. We have:
X
EX∼ΩA [χe (X)] = 2−L
y∈GF (2)L
P
(−1)
i∈e
A(i)·y
= 2−L
X
(−1)A(e)·y = 0
y∈GF (2)L
2.2. Unbiased codes and fooling neighborhoods. We begin with an algorithm to construct a
code which is unbiased for a given set E ⊆ 2[n] ; we later extend this to fool neighborhoods. This
algorithm has two phases: first, we show how to find a code which is unbiased for most of a given
set E; we then bootstrap this to be unbiased on all of E.
We begin with a simple result about multi-variate polynomials over a finite field:
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
5
Proposition 2.7. Let p(z1 , . . . , zk ) be a non-zero polynomial over GF (2s ), with degree ≤ d in each
variable separately. For any α ∈ GF (2s ), note that p(α, z2 , . . . , zk ) is a k − 1 variable polynomial
over GF (2s ). If α ∼ GF (2s ), then p(α, z2 , . . . , zk ) ≡ 0 with probability at most d/2s .
Proof. Factor p as
p(z1 , . . . , zk ) =
X
qt2 ,...,tk (z1 )z2t2 . . . zktk ,
t2 ,...,tk
where t2 , . . . , tk range over non-negative integers. Each such polynomial q has degree d, and they
are not all zero (else p ≡ 0). Let t2 , . . . , tk be such that qt2 ,...,tk 6= 0; with probability ≤ d/2s we have
qt2 ,...,tk (α) = 0. But if qt2 ,...,tk (α) 6= 0, then p(α, z2 , . . . , zk ) has a non-zero coefficient of z2t2 . . . zktk ,
hence p(α, z2 , . . . , zk ) 6≡ 0.
Proposition 2.8. Let E ⊆ 2[n] . Given integer parameters k ≥ 1, s ≥ 0, there is an algorithm
to construct a code A = A(1), . . . , A(n) of length s, such that at most kn1/k 2−s |E| sets e ∈ E
have A(e)P
= 0. This procedure has quasi-complexity (k(s + log(mn)), 2s W ), where m = |E| and
W = n + e∈E |e|.
Proof. Let Z be the set of formal monomials of the form z1u1 . . . zkuk in the ring GF (2s )[z1 , . . . , zk ],
where u1 , . . . , uk ∈ {0, . . . , d} and d = ⌈n1/k − 1⌉. Enumerate Z (in some arbitrary order) as
µ1 , . . . , µℓ where ℓ ≥ n.
For i ∈ [n], we will define A(i) to be the binary representation of µi (α1 , . . . , αk ), where α1 , . . . , αk
will be chosen suitably from GF (2s ).
P
For any e ⊆ [n], define the polynomial µe =
i∈e µi . By linearity, A(e) is the binary representation of µe (α1 , . . . , αk ). So we need to select α1 , . . . , αk so that there are few sets e ∈ E
with µe (α1 , . . . , αk ) = 0. We select α1 , . . . , αk sequentially, according to the following rule. For
i = 1, . . . , k + 1 let us define
Ei = {e ∈ E | µe (α1 , α2 , . . . , αi−1 , zi , zi+1 , . . . , zk ) 6≡ 0}
By Proposition 2.7, if αi is chosen uniformly at random, then in expectation at most (d/2s )|Ei |
sets e ∈ Ei satisfy µe (α1 , α2 , . . . , αi , zi+1 , . . . , zk ) ≡ 0. By enumerating over all possible values of
αi ∈ GF (2s ) to maximize the size of |Ei+1 | we ensure that
|Ei+1 | ≥ (1 − d/2s )|Ei |
As E1 = E, at the end of this process we have |Ek+1 | ≥ |E|(1 − d/2s )k ≥ |E|(1 − kd/2s ) ≥
|E|(1 − kn1/k /2s ), as desired.
This procedure requires k separate stages, each using time O(s + log mn) (to count the number
of good sets e ∈ E). For each choice of αi ∈ GF (2s ), this requires W 1+o(1) processors to count
|Ei+1 |, thus giving a total processor complexity of 2s W 1+o(1) .
Theorem 2.9. Let E ⊆ 2[n] . There is an algorithm with quasi-complexity (log mn, W ) to find an
P
log m
E-unbiased code of length L = log2 m + O( log log
e∈E |e|.
log m ), where m = |E| and W = n +
Proof. We will first discuss our algorithm under the assumption that m ≥ n.
We form the code A by concatenating r separate codes A1 , . . . , Ar , each of length s, i.e.,
A(ℓ) = A1 (ℓ)(1), . . . , A1 (ℓ)(s), A2 (ℓ)(1), . . . , A2 (ℓ)(s), . . . , Ar (ℓ)(1), . . . , Ar (ℓ)(s)
The resulting code A has length L = rs. We form A1 , . . . , Ar sequentially, according to the
following rule. For i = 1, . . . , r + 1 define
Ei = {e ∈ E | A1 (e) = · · · = Ai−1 (e) = 0}
Note that E1 = E. Select each Ai in turn by applying Proposition 2.8 to the set Ei , so that
ln m
m
|Ei+1 | ≤ ǫ|Ei | for ǫ = kn1/k 2−s , where we set k = ⌈ln ln m⌉, s = ⌈ ln ln
ln ln m ⌉ and r = ⌈1 + ln(1/ǫ) ⌉.
6
DAVID G. HARRIS
At the end of this process, we have |Er+1 | ≤ ǫr |E| < 1, and hence Er+1 = ∅, and hence the
resulting code A is unbiased for E.
Using the fact that n ≤ m, we may compute r as
r ≤2+
ln m
ln m
ln ln ln m
≤
≤2−
+ O(1)
1
ln m
−
ln(1/ǫ)
ln 2
ln 2 ln ln ln m (ln ln m + 1)n ln ln m
m
s
So L = rs ≤ log2 m+O( log log
log log m ). Overall, this procedure has quasi-complexity (rk log m, 2 W ) =
(log(mn), W ).
Next, we discuss how to modify this procedure when m < n. In that case, with a simple preprocessing step of quasi-complexity (log mn, W ), we can identify for each e ∈ E a coordinate ve ∈ e.
Let V ′ = {ve | e ∈ E}, and define E ′ = {e ∩ V ′ | e ∈ E}. Using the above procedure we find a code
log m
′
A′ of length L ≤ log2 m + O( log log
log m ) which is unbiased for E . We finish by setting
(
A′ (i) if i ∈ V ′
A(i) =
0
otherwise
2.3. Fooling neighborhoods. If we wish to fool a list E ⊆ 2[n] , we could apply Theorem 2.9 to
the set E ′ = {f | f ⊆ e ∈ E}. This is the approach taken by the algorithm of [17]. However,
even forming E ′ directly might require exponential work. Instead, we can modify our algorithm to
construct an E ′ -unbiased code, without needing to list E ′ explicitly.
As in the proof of Proposition 2.8, we associate to each i ∈ [n] a distinct non-zero monomial µi
1/k
over GF (2s )[z
P1 , . . . , zk ], wherein each indeterminate zi has degree at most d = ⌈n − 1⌉. We also
define µe = i∈e µi for any e ⊆ [n].
To construct the code A, we will find appropriate values αi,j ∈ GF (2s ), where i = 1, . . . , r and
j = 1, . . . , k, and form A as
A(ℓ) = µℓ (α1,1 , . . . , α1,k ), µℓ (α2,1 , . . . , α2,k ), . . . , µℓ (αr,1 , . . . , αr,k )
(where here we identify elements of GF (2s ) with binary vectors of length s). We also define
α(i) = (αi,1 , . . . , αi,k ), and we write α as short-hand for α1,1 , . . . , αr,k .
For i = 1, . . . , r and j = 1, . . . , k and e ∈ E, let us define the potential function
i
Xh
Fi,j,e (α) =
µf (α(1) ) = · · · = µf (α(i−1) ) = 0 ∧ µf (αi,1 , . . . , αi,j , zj+1 , . . . , zk ) ≡ 0
f ⊆e
f 6=∅
The function Fi,j,e (α) only depends on α1,1 , . . . , α1,k , α2,1 , . . . , α2,k , . . . , αi,1 , . . . , αi,j . Note here
that µf (α(1) ), . . . , µf (α(i−1) ) are elements of GF (2s ) while µf (αi,1 , . . . , αi,j , zj+1 , . . . , zk ) is a polynomial in k−j variables over GF (2s ). This function Fi,j,e (α) can be regarded as a type of pessimistic
estimator for the number of sets f ⊆ e for which the code A will be biased.
Proposition 2.10. For any values i, j, e, the function Fi,j,e can be computed with complexity
(polylog(|e|, r, s), poly(|e|, r, s)).
Proof. If we associate the collection of subsets of e with binary vectors of length w, then the set of
all f ⊆ e which satisfy the given constraint is a linear subspace U , and so Fi,j,e (α) has the value
(2rank(U ) − 1). Thus, we need to compute the rank of the vectors
X = {µℓ (α(1) ), . . . , µℓ (α(i) ), µℓ (αi,1 , . . . , αi,j , zj+1 , . . . , zk ) | ℓ ∈ e}.
Here µℓ (αi,1 , . . . , αi,j , zj+1 , . . . , zk ) is regarded as a listing of coefficients.
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
7
The number of vectors is |X| = |e|; next let us count the length of each such vector. Each term
µℓ (α(t) ) is an entry of GF (2s ), hence has length s. Each value of ℓ corresponds to a distinct monomial µℓ (αi,1 , . . . , αi,j , zj+1 , . . . , zk ), so over all we need to keep track of at most |e| distinct monomials
for the polynomial µℓ (αi,1 , . . . , αi,j , zj+1 , . . . , zk ), for which each coefficient also has length s. In total, the length of a vector x ∈ X is at most (r + |e|)s. There is an NC algorithm to compute matrix
rank [13]; thus, this rank calculation has overall complexity (polylog(|e|, r, s), poly(|e|, r, s)).
Theorem 2.11. Let E ⊆ 2[n] , where m = |E| and width(E) = w. There is a deterministic
algorithm to produce a code A of length L ≤ (1 + o(1))(w + log2 m), such that ΩA fools E. This
algorithm has quasi-complexity (w + log(mn), (m + n)wO(1) ).
Proof. We assume n ≤ mw, as we can simply ignore all coordinates which do not appear in E. For
i = 1, . . . , r and j = 1, . . . , k let us define
X
Hi,j (α) =
Fi,j,e (α)
e∈E
We also define H0,k =
P
|e|
e∈E (2
Hi,k
− 1), so that for i = 0, . . . , k, we have
XX
[µf (α(1) ) = · · · = µf (α(i) ) = 0]
=
e∈E f ⊆e
f 6=∅
If Hr,k (α) = 0, then the code A is unbiased for every f ⊆ e ∈ E. Our strategy will be to loop
over i = 1, . . . , r and then j = 1, . . . , k, selecting αi,j at each stage to minimize Hi,j (α).
We now make a few observations on the sizes of Fi,j,e (α). First, H0 (α) ≤ 2w m. Also, since
µf (z1 , . . . , zk ) 6≡ 0, we always have Fi,0,e (α) = 0. Next, observe that if αi,j ∼ GF (2s ), then for any
f with µf (αi,1 , . . . , αi,j−1 , zj , zj+1 , . . . , zk ) 6≡ 0, Proposition 2.7 gives
P (µf (αi,1 , . . . , αi,j , zj+1 , . . . , zk ) ≡ 0) ≤ d/2s ≤ n1/k /2s
Consequently, when αi,j ∼ GF (2s ) and we condition on α1,1 , . . . , αi,1 , . . . , αi,j−1 , we have
E[Fi,j,e (α)] ≤ Fi,j−1,e (α) + (n1/k /2s )(Fi−1,k,e (α) − Fi,j−1,e (α))
By selecting αi,j to minimize Hi,j (α), we thus ensure that
Hi,j (α) ≤ Hi,j−1 (α) + (n1/k /2s )(Hi−1,k (α) − Hi,j−1 (α)) ≤ Hi,j−1 (α) + (n1/k /2s )Hi−1,k (α)
Since Fi,0,e (α) = 0, this in turn ensures that Hi,k (α) ≤ (kn1/k /2s )Hi−1,k (α), so Hr (α) <
(2w m)(kn1/k /2s )r . Thus, for
l log(2w m) m
,
r=
log(2s /(kn1/k ))
the code A will fool all of E.
Now set k = ⌈ln ln(mw)⌉ and s = ⌈ ln ln(mw)
ln ln(mw) ⌉. Using the fact that n ≤ mw, calculations similar
to Theorem 2.9 show that that r ≤ (1 + o(1))(w + log2 m) ln ln ln(wm)/ ln(wm). The code A has
length L = rs ≤ (1 + o(1))(w + log2 m).
Next let us examine the complexity of this process. In each iteration, we must evaluate Fi,j,e for
every e ∈ E and every αi,j ∈ GF (2s ). By Proposition 2.10, each evaluation of Fi,j,e has complexity
(polylog(w, log mn), poly(w, log mn)). Over all possible values α ∈ GF (2s ) and e ∈ E, this gives a
total complexity of (polylog(w, log mn), 2s m poly(w, log mn)).
w+log m
m
There are rk ≤ Õ( w+log
log(mw) ) iterations, so the overall complexity of this process is Õ( log(mw) ) ×
Õ(log(mn) + (log w)O(1) ) time and (m + n)1+o(1) wO(1) processors. As n ≤ mw, this simplifies to
Õ(w + log(mn)) time.
8
DAVID G. HARRIS
2.4. Comparison with previous algorithms. Let us briefly compare Theorem 2.11 with previous algorithms for fooling neighborhoods. The simplest approach to fool a list E ⊆ 2[n] , is to select
a code A whose dual code has weight w + 1, where w = width(E). The resulting probability space
ΩA is then w-wise-independent. There are algebraic constructions to do so efficiently; for example,
the use of BCH codes in this context was first applied to derandomization in [3]. Such codes have
length roughly (w/2) log 2 n.
An algorithm of Schulman [17] can reduce the code-length significantly to O(w + log |E|), by
generating a code which is unbiased for the set E ′ = {f ⊆ e | e ∈ E}. A similar approach is
used in [7], which combines a number of other algorithmic steps with the generation of the code.
The basic idea of both these works is to form Ω as a product of many independent copies of
an ǫ-approximately-independent probability space, where ǫ is constant. In [17], the underlying
ǫ-approximately-independent probability space was based on a construction of [14] using ReedSolomon codes; these have a particularly nice form for derandomizing part of the random seed.
These algorithms have high processor complexity (approximately O(mn2w )), and there are two
main reasons for this. First, simply enumerating the set E ′ requires a large processor count,
exponential in w. Second, and more subtly, these algorithms must test all possibilities for the seed
of the underlying Reed-Solomon code, and this requires a processor complexity exponential in the
seed-length of that code.
Theorem 2.11 thus improves in two ways over the previous algorithms. First, it has reduced time
and processor complexity; in particular, it answers an open problem posed by Schulman [17] in
giving an NC1 algorithm for w ≤ O(log(mn)), and it gives an NC algorithm for w = polylog(mn).
Second, its seed-length is smaller: it gives L ≤ (1+o(1))(w+log 2 m) whereas the previous algorithms
only guarantee L ≤ O(w + log m).
3. Fooling sums of juntas
We say that a function f : {0, 1}n → R is a w-junta if there exists a set Y = {y1 , . . . , yw } ⊆ [n],
such that
f (x1 , . . . , xn ) = f ′ (xy1 , . . . , xyw )
for some function f ′ : {0, 1}w → R.
In this section, we consider a function S : {0, 1}n → R of the form
(1)
S(x) =
m
X
fj (x)
j=1
where each fj is a w-junta whose value is determined by a variable subset Yj ⊆ [n]. Our goal is
to find some x ∈ {0, 1}n with the property that S(x) ≥ EX∼{0,1}n [S(X)] = S0 . Our algorithm has
four main components, which we will describe in turn:
(1) We show how to apply conditional expectations when the objective functions are Fourier
characters.
(2) We show how to use Fourier decomposition to the sum (1), thus reducing a sum of w-juntas
to a sum of Fourier characters. As we have discussed earlier, this step implemented directly
has an exponential processor dependence on w.
(3) We use (derandomized) random variable partitioning to break the overall sum into w/w′
subproblems involving w′ -juntas, where w′ = o(log n).
(4) We introduce an object we refer to as partial-expectations oracle (PEO), which allows us to
use conditional expectations to solve these subproblems sequentially. This requires O(w)
′
time, but only 2w = no(1) processors.
Berger & Rompel discuss a few alternate strategies to mitigate the exponential dependence on
w, for example when the underlying variables are drawn from {0, 1}b for b = polylog(n), or when
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
9
fj are indicators of affine functions. But these strategies are not as general as we need for many
applications. Our overall algorithm will handle all these situations as special cases.
3.1. Conditional expectations for sums of Fourier characters and sums of w-juntas. Our
approach begins with a subroutine for optimization problems involving sums of Fourier characters.
This idea has been used in a number of deterministic algorithms, starting with [11] and more
extensively used in [4] and [7]. We present a slightly optimized form here.
Theorem 3.1. Let E ⊆ 2[n] be given along with associated weights γe for every e ∈ E. There is
an algorithm with quasi-complexity (log mn, W ) to find x ∈ {0, 1}n such that
X
γe χe (x) ≥ γ∅ ,
e∈E
where m = |E| and W = n +
P
e∈E
|e|.
Proof. First, use Theorem 2.9 to construct the E-unbiased code A of length L = O(log m), using
quasi-complexity P
(log mn, W ).
Define G(x) = e∈E γe χe (x). Now, observe that when X ∼ ΩA , we have E[χe (X)] = 0 for e 6= ∅
and E[χe (X)] = 1 for e = ∅. So EX∼ΩA [G(X)] = γ∅ . In particular, there exists some x supported
on ΩA with G(x) ≥ γ∅ . We can write x ∈ ΩA as xi = Ai · y for i = 1, . . . , n, where y ∼ GF (2)L .
We thus define
X
G′ (y) =
γe (−1)A(e)·y
e∈E
[G′ (y)]
= γ∅ . So we want to find y ∈ GF (2)L with S ′ (y) ≥ γ∅ .
and note that Ey∼GF (2)L
We do so by using conditional expectation: we guess chunks of t = logloglogmn
mn bits of y at a time,
′
to ensure that the expected value of E[S (y)] increases. For each such guess, we will compute
in parallel the resulting expected value E[S ′ (y)], when certain bits of y are fixed and the rest
remain independent fair coins. We may compute the conditional expectations of a term (−1)A(e)·y ,
using the following observation: suppose that y1 , . . . , yk are determined while yk+1 , . . . , yL remain
independent fair coins. Then E[(−1)A(e)·y ] = 0 if A(e)∩{k +1, . . . , L} =
6 ∅, and otherwise (−1)A(e)·y
is determined by y1 , . . . , yk .
log mn
This process requires ⌈L/t⌉ ≤ O(1 + L log
log mn ) rounds. For each possible value for a t-bit chunk
of y, evaluating S ′ has complexity (log(mn), W 1+o(1) ). Thus the overall processor complexity is
2t W 1+o(1) ≤ W 1+o(1) and overall time complexity is ˜( log mn).
Lemma 3.2. Suppose we have a full listing of the truth-table of each fj . There is an algorithm to
find x ∈ {0, 1}n satisfying S(x) ≥ S0 , using quasi-complexity (w + log mn, 2w m + n).
P
Proof. Using Proposition 2.2, transform each fj as fj (x) = e⊆Yj γj,e χe (x). This step has quasicomplexity (w + log mn, 2w m). We thus have:
XX
X
X
S(x) =
γj,e χe (x) =
χe (x)(
γj,e )
j
e⊆Yj
e
j
P
and j γj,∅ = S0 .
P
Next apply Theorem 3.1 to the set E = {e | e ⊆ Yj } and associated weights
j γj,e . Since
w
w
w
|E| ≤ 2 m and width(E) ≤ w, this procedure has quasi-complexity (log(2 mn), 2 m + n).
When w = polylog(n), this means that Lemma 3.2 gives quasi-NC algorithms. When w =
Θ(log n) then Lemma 3.2 gives NC algorithms; however, the processor complexity (while polynomial) may be quite large, depending on the size of w.
As a side application of Theorem 3.1 (which is not needed for our overall derandomization
approach), let us consider the heavy-codeword problem. We are given a code of length L and size n,
10
DAVID G. HARRIS
presented as a L × n generator matrix. Our goal is to find a codeword whose weight is at least the
expected weight of a randomly-chosen codeword. This was introduced as a toy derandomization
problem by [14]; this work also gave an algorithm with complexity roughly (log Ln, L2 n2 ). This
was later improved by [7] to complexity (log Ln, Ln2 ). We improve this further to nearly optimal
time and processor complexities.
Corollary 3.3. There is an algorithm with quasi-complexity (log Ln, Ln) to find a heavy codeword.
Proof. Let A be the generator matrix. We suppose without loss of generality that no row of A is
all zero. In this case, the expected weight of a codeword is L/2. Letting y1 , . . . , yL denote the rows
of A, we wish to find a vector x ∈ {0, 1}n such that yj · x = 1 for at least L/2 values of j.
P
yj ·x . Observe that if G(x) ≥ 0 then x is orthogonal to at least half of
Define G(x) = − L
j=1 (−1)
y1 , . . . , yL as desired. Furthermore, as y1 , . . . , yL are non-zero we have E[G(X)] = 0. So we apply
Theorem 3.1, noting that W ≤ Ln.
3.2. Derandomized variable partitioning. This step is based on a derandomization technique
of [1] using symmetric polynomials and approximately-independent probability spaces (also known
as small-bias probability spaces). We begin by defining and quoting some results on approximate
independent probability spaces.
Definition 3.4. A probability space Ω over {0, 1}n is t-wise, ǫ-approximately independent, if for
any indices 1 ≤ i1 < i2 < · · · < it ≤ n, and any bits y1 , . . . , yt ∈ {0, 1}t , we have
Px∼Ω (xi1 = y1 ∧ · · · ∧ xit = yt ) ≤ (1 + ǫ)2−t
Theorem 3.5 ([14]). For any integer t ≥ 1 and ǫ > 0, there is a t-wise, ǫ-approximately independent probability space Ω of support size |Ω| ≤ 2O(t+log(1/ǫ)+log log n) . The space Ω can be constructed
with quasi-complexity (t + log(1/ǫ) + log n, 2O(t+log(1/ǫ)+log log n) )
Lemma 3.6. Let E ⊆ 2[n] , where m = |E| and w = width(E). One can construct a partition of
log mn)5
), satisfying
[n] into R parts [n] = T1 ⊔ T2 ⊔ · · · ⊔ TR , for R = O(1 + w(log
log mn
|f ∩ Tk | ≤ O(
log mn
) for all f ∈ E, k ∈ [R]
log log log mn
This algorithm has quasi-complexity (log w log(mn), wO(1) (m + n)).
5
log mn)
⌉, where C is a sufficiently large constant. We will construct
Proof. Let r = ⌈log2 Cw(log
log mn
n
binary vectors y1 , . . . , yr ∈ {0, 1} and then define for each k ∈ {0, 1}r , ℓ ∈ {0, . . . , r} the sets
Tkℓ ⊆ [n] by
Tkℓ = {i ∈ [n] | y1 (i) = k(1) ∧ y2 (i) = k(2) ∧ · · · ∧ yℓ (i) = k(ℓ)}
We will finish by setting R = 2r and forming the sets T1 , . . . , TR by Tk = Tkr where k ranges over
{0, 1}r . For each k ∈ {0, 1}r , ℓ ∈ {0, . . . , r}, f ∈ E we define Hℓ (f, k) = |f ∩ Tkℓ |. We will achieve
the goal of the theorem if we select y1 , . . . , yr so that every f ∈ E, k ∈ {0, 1}r has Hr (f, k) ≤ t for
mn
t = ⌈ log log
log log mn ⌉.
For each ℓ = 0, . . . , r let us define the potential function
X Hℓ (f, k)
Qℓ =
t
f,k
Observe that Qr is an integer; thus, if Qr < 1, then it follows that Qr = 0 and so Hr (f, k) < t for
all f, k as desired.
Let Ω be a probability distribution over GF (2) which is t-wise, ǫ-approximately independent,
where ǫ = 1/r, according to Definition 3.4. By Theorem 3.5, we have |Ω| ≤ (mn)o(1) wO(1) ; furthermore, the complexity of generating Ω will be negligible for the overall algorithm.
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
11
When yℓ ∼ Ω, then note that each t-tuple of elements in f ∩ Tkℓ has a probability of at most
2−t (1 + ǫ) of surviving to Tkℓ+1 . This implies that that any f, k has
E[ Hℓ+1t(f,k) | y1 , . . . , yℓ ] ≤ (1 + ǫ)2−t Hℓ (f,k)
t
and similarly E[Qℓ+1 | y1 , . . . , yℓ ] ≤ 2−t (1 + ǫ)Qℓ . We thus select y1 , . . . , yr sequentially in order to
minimize Qℓ+1 at each stage ℓ. This ensures that for ℓ = 1, . . . , r we have Qℓ ≤ 2−t (1 + ǫ)Qℓ−1 . At
the end of this process, we have
X H0 (f, k))
w
1−t
r −tr
r −t
≤ eR m
≤ eR1−t m(ew/t)t
Qr ≤ (1 + ǫ) 2 Q0 = (1 + 1/r) R
t
t
f,k
Simple calculations now show that Qr < 1 for C a sufficiently large constant.
We now examine the complexity of this algorithm. There are r stages; in each stage, we must
search the probability space Ω and compute Qℓ . The potential function Qℓ can be computed with
quasi-complexity (log mn, mwR). Note now that mwR ≤ (mn)o(1) mwO(1) . As |Ω| ≤ (mn)o(1) wO(1) ,
this costs (mn)o(1) wO(1) (m + n) processors and Õ(r log mn) = Õ(log w log mn) time.
3.3. The partial-expectations oracle. As we have discussed, we need implicit access to fj in
order to avoid the exponential dependence on w. A key idea of Berger & Rompel [4] to achieve this
is an algorithm capable of determining certain conditional expectations for the objective functions.
For our purposes, we define this as follows.
Definition 3.7. An algorithm A is a partial-expectations oracle (PEO) for the functions fj , if
it is capable of the following operation. Given any X ′ ∈ {0, 1, ?}n , the algorithm A computes
Fj = EX∼Ω [fj (X)] for j = 1, . . . , m, where the probability distribution Ω is defined by drawing each
bit Xi independently, such that if Xi′ = ? then Xi is Bernoulli-1/2 and if Xi′ 6= ? then Xi = Xi′ .
We note that this form of PEO is simpler than that used by Berger & Rompel: the latter requires
evaluating the conditional expectation of fj (X) given that X is confined to an affine subspace, while
our PEO only requires computing this conditional expectation when individual bits of X are fixed.
This can be a much easier task, as it allows one to take advantage of independence among the
entries of X.
We now combine all the ingredients to obtain our conditional expectations algorithm.
P
n
Theorem 3.8. Suppose S(x) = m
j=1 fj (x) for x ∈ {0, 1} , where each fj is a w-junta. Suppose
we have a PEO for the functions fj with complexity (C1 , C2 ).
Then there is an algorithm to find a vector x satisfying
S(x) ≥ EX∼{0,1}n S(X),
wC1
O(1) C ).
with quasi-complexity ( log
2
mn + C1 + log w log mn, w
Proof. We assume C1 ≥ Ω(log mn) and C2 ≥ Ω(m + n) as it requires this complexity to take as
input the values j, X ′ and output Fj . We similarly assume that n ≤ mw, as variables not involved
in any objective function may be ignored.
log mn)5
),
First apply Lemma 3.6 to determine a partition [n] = T1 ⊔ · · · ⊔ TR for R = O(1 + w(log
log mn
mn
such that |Yj ∩ Tk | ≤ w′ for w′ ≤ O( log log
log log mn ). This stage has quasi-complexity (log w log mn,
wO(1) (m + n)).
Next, for r = 1, . . . , R, we seek to determine the bits {xi | i ∈ Tr }. Define the function fj′ (z)
to be the expected value of fj (X), when the entries Xi for i ∈ Tr are set to zi , the variables Xi
for i ∈ T1 , . . . , TR are set to xi , and the remaining entries of X (for i ∈ Tr+1 , . . . , TR ) remain fair
′
coins. Each fj′ is a w′ -junta and we can determine its truth-table fj′ using 2w invocations of our
12
DAVID G. HARRIS
PEO, where we define Xi′ = ? for i ∈ Tr+1 ∪ · · · ∪ TR and Xi′ 6= ? otherwise. This in turn requires
′
′
C1 + Õ(log(2w mn)) ≤ Õ(C1 ) time and (mn)o(1) 2w C2 ≤ (mn)o(1) C2 processors.
Next, apply Lemma 3.2 to determine a value for the relevant variables in Tr ; this step takes
′
Õ(w′ + log mn) ≤ Õ(log mn) time and (n + 2w m)1+o(1) ≤ (mn)o(1) (m + n) processors.
wC1
Over all R stages, the total time for this algorithm is Õ(RC1 ) ≤ Õ( log
mn + w + C1 ).
We emphasize the low time and processor complexity of this algorithm. For example, if w =
polylog(mn) and C1 = Õ(log mn) (which are typical parameters), then this has quasi-complexity
(w, C2 ). Even if w = Θ(log mn), this can lead to greatly reduced complexities as compared to the
algorithm of [4].
In order to apply the algorithm, however, we must have an appropriate PEO. These are problemspecific, and there is no general recipe for constructing them. One simple class of objective functions,
which was one of the main cases considered by Berger & Rompel, is an indicator function for an
affine space. PEO’s for such functions can be derived by a rank calculation. This simple class of
functions already gives numerous non-trivial results.
We can also construct PEO’s when the functions fj are computed via a read-once branching
program (ROBP). In this model, each fj is represented as a directed acyclic graph on M states;
at each node v, a single variable xv is read and the program branches to two possible destinations
depending on the variable xv . There is a designated starting vertex and at some designated sink
vertices, a real number is output. In addition, every variable label appears at most once on each
directed path. This is a quite general class of functions, which includes log-space statistical tests
as used by Sivakumar’s derandomization [18]. See [5] for further details.
Proposition 3.9. If a function fj can be represented as a ROBP on M states, then it has a PEO
with quasi-complexity (log w log M n, M 3 wO(1) ).
Proof. See Theorem 3.14 for a proof of a stronger result.
3.4. Non-binary variables. Let us now consider a slightly more general type of problem: we
have n variables x1 , . . . , xn , each of which is an integer in the range {0, . . . , 2b − 1}. Our objective
function is again a sum of w-juntas, that is, each fj (x) depends on at most w coordinates of x.
This can easily be reduced to the model we have discussed earlier: we replace each variable xi with
b separate binary variables xi1 , . . . , xib . Now each fj depends on wb bits of the expanded input,
and so is a wb-junta.
However, there is a complication. In order to apply Theorem 3.8, we need a PEO for the
functions fj . Thus we need to compute the expected value of fj (x), given that certain bits of x
are fixed to specific values. This can be somewhat awkward, as restricting arbitrary bits of xi does
not necessarily have any natural interpretation when xi is an integer in the range {0, . . . , 2b − 1}.
It is often easier to use the strategy of [4], which fixes the bit-levels of x1 , . . . , xn one at a time.
This allows us to use a simpler type of PEO where the pattern of known/unknown bits is more
controlled.
For the purposes of the algorithm, we identify the integer set {0, . . . , 2b − 1} with the set of
P
i
length-b binary vectors; a value x0 , . . . , xb−1 corresponds to the integer b−1
i=0 2 xb−1−i . Note here
b
that x0 is the most-significant bit. Let us define Mb as the set {0, . . . , 2 − 1} equipped with this
bit-based interpretation. Likewise, if x ∈ Mnb , we let x(i, j) denote the j th -most-significant-bit of
the integer value xi .
Definition 3.10. We say that X ′ ∈ {0, 1, ?}nb is graded if there is some integer ℓ ∈ {0, . . . , b − 1}
such that
(1) X ′ (i, j) ∈ {0, 1} for j = 0, . . . , ℓ − 1
(2) X ′ (i, j) = ? for j = ℓ + 1, . . . , b − 1
We say that X ′ is fully-graded if X ′ satisfies for some integer ℓ ∈ {0, . . . , b} the stronger condition
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
13
(1) X ′ (i, j) ∈ {0, 1} for j = 0, . . . , ℓ − 1
(2) X ′ (i, j) = ? for j = ℓ, . . . , b − 1
An algorithm A is a graded PEO (respectively fully-graded PEO) for the functions fj if it is a
PEO, but only for queries X ′ which are graded (respectively, fully-graded).
P
n
Theorem 3.11. Suppose that S(x) = m
j=1 fj (x) for x ∈ Mb , where each function fj depends on
at most w coordinates of x. Suppose we have a graded PEO for the functions fj with complexity
(C1 , C2 ).
wC1
Then we can find a vector x satisfying S(x) ≥ EX∼Mnb [S(X)], using quasi-complexity (b( log(mn)
+
C1 + w + log w log mn), wO(1) C2 ).
Proof. We will determine the bits of x in b separate stages; at the ℓth , we determine the bit-level
ℓ of each entry xi . For ℓ = 1, . . . , b, consider the following process. Define the function fℓ,j (z) to
be the expected value of fj (X), when the bit-levels 1, . . . , ℓ − 1 of X are taken from the alreadydetermined vector x; when the bit-level ℓ of X is set to z; and when the bit-levels ℓ + 1, . . . , b of X
are independent fair coins.
Each fℓ,j is a w-junta, and the graded PEO for the functions fj yields a PEO for
P the funcn
tions fℓ,j . Therefore, using Theorem 3.8, we select z ∈ {0, 1} so as to ensure that j fℓ,j (z) ≥
P
EZ∼{0,1}n [ j fℓ,z (Z)].
One important application of non-binary variables concerns derandomizing biased coins. For
a vector of probabilities p ∈ [0, 1]n , consider the probability space with n independent variables
X1 , . . . , Xn , wherein each Xi is Bernoulli-pi . We write this more compactly as X ∼ p. Most of our
derandomization results we have proved earlier have assumed that the underlying random bits are
independent fair coins (i.e. with p1 , . . . , pn = 1/2).
Definition 3.12. An algorithm A is a continuous partial-expectations oracle (PEO) for the functions fj , if it is capable of the following operation. Given any vector q ∈ [0, 1]n , whose entries
are rational number with denominator bounded by some integer 2b , the algorithm A computes
Fj = EX∼q [fj (X)] for j = 1, . . . , m.
Note that a PEO can be regarded as a special case of a continuous PEO, in which the probability
vector q is restricted to the entries {0, 1/2, 1}.
Pm
n
Theorem 3.13. Suppose that S(x) =
j=1 fj (x) for x ∈ {0, 1} , where each function fj is a
w-junta. Suppose we have a continuous PEO for the functions fj with complexity (C1 , C2 ). Let
p ∈ [0, 1]n be a vector of probabilities, wherein each entry pi is a rational number with denominator 2b . Then we can find a vector x satisfying S(x) ≥ EX∼p [S(X)], using quasi-complexity
(b(wC1 / log(mn) + C1 + w + log w log mn), (wb)O(1) C2 )).
Proof. Consider the function fj′ (y1 , . . . , yn ) : Mnb → R defined by
fj′ (y1 , . . . , yn ) = fj [y1 /2b ≤ p1 ], . . . , [yn /2b ≤ pn ]
Each function fj′ depends on w coordinates of y. Furthermore, if certain top bit-levels of y are
fixed to a certain value and the remaining lower bit-levels of y are independent fair coins, then each
term [yi /2b ≤ pi ] is a Bernoulli-qi variable, where qi depends on the fixed values of yi . Therefore,
the given continuous PEO for fj provides a graded PEO for the functions fj′ , with a complexity of
(log(nb) + C1 , C2 + nb).
Finally, observe that when Y ∼ Mnb , each term [Yi /2b ≤ pi ] is Bernoulli-pi ; therefore, we have
X
X
fj′ (Y )] = EX∼p [
fj (X)]
EY ∼Mnb [
j
j
14
DAVID G. HARRIS
So Theorem 3.11 produces y1 , . . . , yn ∈ Mnb with
vector x ∈ {0, 1}n defined by xi = [yi /2b ≤ pi ].
P
j
P
fj′ (y) ≥ EX∼p [ j fj (X)]. Output the
Proposition 3.14. If a function fj can be represented as a ROBP on M states, then it has a
continuous PEO with quasi-complexity (log b log w log M n, bM 3 wO(1) ).
Proof. We must calculate the expected value of fj , given that the variables X1 , . . . , Xn are independent Bernoulli-qi . Now observe that, for any states s1 , s2 , the probability that s1 goes to s2 in
at most h time-steps is the sum over intermediate states s of the probability that s1 goes to s in at
most h/2 time-steps and that s goes to s2 in at most h/2 time-steps; this follows from the definition
of an ROBP. Using this relation, one may recursively build the transition matrix for pairs of states
s1 → s2 over time horizons h = 1, 2, 4, . . . , w. Each such iteration takes time Õ(log b log M n) and
there are Õ(log w) iterations.
4. Rainbow hypergraph coloring
Consider a d-regular hypergraph G, with m edges and n vertices. Given a vertex coloring of G, we
say that an edge e is rainbow if all its vertices receive distinct colors. A challenge for a deterministic
algorithm is to find a coloring with m dd!d rainbow edges, which is the expected number in a uniform
random coloring. In [3], an NC algorithm was given in the case d = O(1). This was extended by
[4] to arbitrary d; although [4] did not give any concrete time or processor complexity, it would
appear that the complexity would be roughly (log4 mn, n + m1+ln 2 ) We will significantly improve
both the time and processor costs here.
Theorem 4.1. There is an NC algorithm to find a vertex coloring with at least m dd!d rainbow edges,
using quasi-complexity (log2 mn, md + n).
Proof. We begin with simple pre-processing steps. First, when d ≥ ln m + Ω(log log m), then
it suffices to rainbow-color a single edge, which may be done easily. Second, when n ≤ md,
then some vertex is not used; we may delete it from the graph. Hence we assume m ≥ nd and
d ≤ ln m + O(log log m).
Given a binary vector x ∈ Mnb , we define the associated d-coloring φ : V → {0, . . . , d − 1} by
φx (v) = ⌊(d/2b )xv ⌋. For each edge P
e ∈ G let fe (x) be the indicator function that e is rainbow on
the coloring φx , and define S(x) = e fe (x).
. FurAs shown in [4], by taking b = Õ(log mn), we can ensure that EX∼Mnb [S(X)] > (md!−1)
dd
thermore, since S(x) is an integer and
(md!−1)
is a rational number with denominator dd , when
dd
md!
. So, if we find x ∈ Mnb with S(x) ≥ EX∼Mnb [S(X)],
dd
we ensure that S(x) ≥
S(x) > (md!−1)
dd
then this will yield our desired coloring.
Each fe depends on d input coordinates and d ≤ ln m + O(log log m). So applying Theorem 3.11
wC1
with w = d gives a total quasi-complexity of O(b( log
mn +C1 +w+log w log mn), n+C2 ) to construct
φx , where (C1 , C2 ) is the complexity of PEO for the functions fe .
Observe that w ≤ O(log m) and b ≤ Õ(log mn). In Proposition 4.2 we show that C1 ≤ Õ(log mn)
and C2 ≤ (m + n)1+o(1) , so this simplifies to quasi-complexity of (log2 mn, m + n).
Proposition 4.2. Let fe be the indicator function that edge e is rainbow. Then the collection of
functions fe has a graded PEO with overall quasi-complexity (log mn, m + n).
Proof. It suffices to compute the probability that a given edge e will be rainbow on the coloring
φx for some graded u ∈ {0, 1, ?}db (here u represents the projection of the overall partially-graded
x ∈ {0, 1, ?}nb to the vertices in e). Since db = (mn)o(1) , the processor complexity of this task can
be an arbitrarily polynomial in b, d.
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
15
We first describe how to do so if u is fully-graded; we then modify it to allow u to be merely
graded. Suppose the most-significant ℓ ≤ b bit-levels of the vector y have been determined and the
least-significant b−ℓ bit-levels of u remain fair coins. We may write u in the form uv = (yv , ?, . . . , ?),
where yv ∈ Mℓ . For each c ∈ Mℓ let Sc denote the set of vertices v ∈ e with yv = c.
For each k ∈ {0, . . . , d − 1}, let us define Gk to be the set of values c ∈ {0, . . . , 2ℓ − 1} such that
some vertex v ∈ Sc could (depending on the lower order bits of y) be assigned color k. Specifically,
Gk = {c ∈ {0, . . . , 2ℓ − 1} | F (2b−ℓ c) ≤ k ≤ F (2b−ℓ (c + 1) − 1)}
where function F : {0, 2b − 1} → {0, . . . , d − 1} is defined by F (x) = ⌊(d/2b )x⌋.
Observe that if y ≥ x + (d/2b ), then we must have F (y) > F (x). Using this fact, we claim that
each Gk is either a singleton set, or a set of two adjacent elements {c, c + 1}. For, suppose not;
then there must exist c1 , c2 ∈ Gk with c2 > c1 + 1 and
F (2b−ℓ c1 ) ≤ k ≤ F (2b−ℓ (c1 + 1) − 1)
But, note that in this case
F (2b−ℓ c2 ) ≤ k ≤ F (2b−ℓ (c2 + 1) − 1)
(2b−ℓ c2 ) − (2b−ℓ (c1 + 1) + 1) = 2b−ℓ (c2 − c1 − 1) − 1 ≥ 2b−ℓ + 1 ≥ 1 ≥ (d/2b )
and so F (2b−ℓ c2 ) > F (2b−ℓ (c1 + 1) − 1), a contradiction.
Also, we claim that for each value of c, there is at most one value k such that Gk = {c, c + 1}.
For, if not, then there would be values k1 < k2 with
F (2b−ℓ (c + 1)) ≤ k1 ≤ F (2b−ℓ (c + 2) − 1)
But note then that
F (2b−ℓ c) ≤ k2 ≤ F (2b−ℓ (c + 1) − 1)
F (2b−ℓ (c + 1) − 1) ≥ k2 ≥ k1 + 1 ≥ F (2b−ℓ (c + 1)) + 1 > F (2b−ℓ (c + 1)) ≥ F (2b−ℓ (c + 1) − 1),
a contradiction.
Thus, for each c ∈ {0, . . . , 2ℓ − 1}, let us define Wc to be the set of values k ∈ {0, . . . , d − 1} such
that Gk = {c, c + 1}. We observe that |Wc | ≤ 1.
Now consider the random experiment of assigning independent Bernoulli-1/2 values to the loworder b − ℓ bit-levels of u. Let Zc be a random variable, which is the number of vertices in Sc
assigned a value k ∈ Wc . (If Wc = ∅, then Zc = 0 necessarily.) In order for e to be rainbow, every
c ∈ Mℓ must have Zc ∈ {0, 1}.
For any integers 0 ≤ c0 < c1 ≤ 2ℓ and values z0 , z1 ∈ {0, 1}, let us thus define the function g by
g(c0 , c1 , z0 , z1 ) =
P (the vertices in Sc0 , Sc0 +1 , . . . , Sc1 −1 , Sc1 receive distinct colors and Zc1 = z1 | Zc0 −1 = z0 )
The overall probability that the random experiment results in a rainbow coloring of e is given
by g(0, 2ℓ , 0, 0). With a little thought, one can see that g satisfies the recurrence:
g(c0 , c1 , z0 , z1 ) = g(c0 , c2 − 1, z0 , 0)g(c2 , c1 , 0, z1 ) + g(c0 , c2 − 1, z0 , 1)g(c2 , c1 , 1, z1 )
for c2 = (c0 + c1 )/2.
If Sc = ∅, then the value c is not relevant to this calculation; thus, during this calculation, we can
skip all such entries. As there are at most d values of c with Sc 6= ∅, we can recursively compute
g(0, 2ℓ , 0, 0) using poly(db) processors and using Õ(log mn) time. (The base cases can be computed
using simple arithmetic as functions of |Sc |.)
We next discuss how to modify this to graded PEO. Here, the top ℓ − 1 bits of each yv are
completely known, while the lowest-order bit is in {0, 1, ?}. Now suppose we want to calculate
g(0, 1, z0 , z1 ); in this case, some vertices are known to correspond to the sets S0 , S1 and some
vertices (for which bit at level ℓ is unspecified) have a 1/2 probability of going into S0 and a 1/2
16
DAVID G. HARRIS
′ . Now |S ′ | is not
probability of going into S1 . Let us denote the three cases as S0′ , S1′ , and S0/1
0/1
known exactly but has a binomial distribution; we integrate over it to determine |S0 |, |S1 | and hence
to compute g(0, 1, z0 , z1 ).
5. The Lovász Local Lemma with complex bad-events
The Lovász Local Lemma (LLL) is a keystone principle in probability theory which asserts that
if one has a probability space Ω and and a set B of “bad-events” in Ω, then under appropriate
“local” conditions there is a positive probability that no event in B occurs. The LLL has numerous
applications to combinatorics, graph theory, routing, etc. Although the LLL applies to general
probability spaces, in most applications a simpler bit-based form suffices. In this setting, the space
Ω is a product space over n binary variables x1 , . . . , xn , which are independent Bernoulli-1/2. Each
bad-events B ∈ B is a boolean function fB on a subset YB of the variables. We say that bad-events
B, B ′ affect each other (and write B ∼ B ′ ) if YB ∩YB ′ 6= ∅. We let m = |B|. We say that x ∈ {0, 1}n
avoids B if fB (x) = 0 for all B ∈ B.
The simplest “symmetric” form of the LLL states that if each bad-event B ∈ B has probability
PΩ (B) ≤ p, and each bad-event B affects at most d bad-events (including itself), then if epd ≤ 1
there is a positive probability of avoiding B.
In a seminal paper [12], Moser & Tardos introduced the following simple randomized algorithm,
which we refer to as the MT algorithm, giving efficient randomized constructions for nearly all
applications of the LLL:
Algorithm 1 The Moser-Tardos algorithm
1: Generate x1 , . . . , xn as independent fair coins.
2: while some bad-event is true on x do
3:
Arbitrarily select some true bad-event B
4:
For each i ∈ YB , draw xi as an independent fair coin. (We refer to this as resampling B.)
Under nearly the same conditions as the probabilistic LLL, the MT algorithm terminates in
polynomial expected time. Moser & Tardos also gave a parallel (RNC) variant of this algorithm,
requiring a slack compared to the LLL criterion.
There are two key techniques introduced by [12] to analyze their MT algorithm. The first is
the idea of a resampling table. In the MT algorithm as we have presented it, the new values for
each variable are drawn in an on-line fashion. Instead, one can imagine a fixed table R. This table
records, for each variable i, an infinite list of values R(i, 1), R(i, 2), . . . , for that variable, which are
all independent fair coins. When the MT algorithm begins, it sets xi = R(i, 1) for each variable i;
if a variable xi needs to be resampled, one sets xi = R(i, 2), and so forth. Once we have fixed a
resampling table R, the MT algorithm can be executed deterministically.
The other key idea introduced by Moser & Tardos is the witness tree, which represents a possible
execution path for the MT algorithm leading to a given resampling. This is explained in great
detail in [12], which we recommend as an introduction. As a brief summary, suppose we want to
explain why some bad-event B was resampled at time t. We form a witness tree τ by first placing
a root node labeled by B, and then going in time through the execution log from time t − 1 to
time 1. For each event B we encounter at time s < t, we look in τ to find if there is some node v ′
labeled by B ′ ∼ B. If so, we place a node v labeled by B in the tree as a child of v ′ ; if there are
multiple choices for v ′ , we always choose the one of greatest depth (if there are multiple choices at
greatest depth, we break the tie arbitrarily).
For any witness tree τ and any node v ∈ τ , we let L(v) ∈ B denote the label of v.
Definition 5.1 (Weight and size of witness tree). For a witness tree
Q τ , we define the size of τ as
the number of nodes in τ and we define the weight of τ as w(τ ) = v∈τ PΩ (L(v)).
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
17
The most important result of [12], which explains why the MT algorithm works, is the Witness
Tree Lemma:
Lemma 5.2 ([12]). The probability that a witness tree τ appears during the execution of the MT
algorithm is at most w(τ ).
To prove this Lemma, [12] shows τ imposes certain conditions on the resampling table R.
Lemma 5.3 ([12]). Let τ be a witness tree with t nodes. Then there is a set of subsets Wv ⊆ [n]×Z+ ,
indexed by nodes v ∈ τ , such that
(A1) For v 6= v ′ we have Wv ∩ Wv′ = ∅.
(A2) A necessary condition for τ to occur is that fL(v) (Wv (R)) = 1 for every v ∈ τ .
(A3) The sets Wv can be determined from τ with quasi-complexity (log nt, nt).
(A4) Every v ∈ τ has |Wv | = |YL(v) |.
Proof. For each node v ∈ τ and each i ∈ [n], let ui,v denote the number of nodes v ′ which are at
greater depth than v and which have i ∈ YL(v′ ) . Define Wv = {(i, ui,v + 1) | i ∈ YL(v) }.
Lemma 5.2 follows from Lemma 5.3; as the entries of R are fair coins, the probabilities of each
event fL(v) (Wv (R)) = 1 is PΩ (L(v)); furthermore, since the sets Wv are non-intersecting, these
events are all independent.
5.1. Derandomizing Moser-Tardos. The original paper of Moser & Tardos gave a sequential
deterministic algorithm that only worked for a very limited class of LLL instances, for example
when d was constant. An NC algorithm was later given in [6], covering a slightly larger class
of bad-events. This algorithm required satisfying the LLL criterion with a slack, in particular it
required
epd1+ǫ ≤ 1
3
for some ǫ > 0, and had a complexity of roughly ( log ǫmn , mO(1/ǫ) ). An alternative NC algorithm
was provided in [10], which is slightly faster than [6].
These latter algorithms have numerous conditions on the functions fB ; roughly speaking, they
require fB to have decision-tree complexity of order roughly log d. The clearest example of this
type of problem is that of k-SAT, in which bad-event corresponds to a clause being violated. So
each bad-event is defined by xi1 = j1 ∧ · · · ∧ xik = jk , a monomial in k variables.
Many other LLL applications, particularly those in which the bad-events are determined by
sums of random variables, do not fit into this paradigm. Some of these problems are discussed in
Section 6. The hallmark of these types of problems is that the bad-events are complex boolean
functions; our focus here will be to give NC algorithms for such problems.
The analysis of [10] is based on an extension of the witness tree to a more general object referred
to as a collectible witness DAG (CWD). Again, the collection of these objects represent in a sense
all the ways the MT algorithm could require a long execution time. This requires a great deal of
notation to define properly, but the important point for us is that each CWD τ satisfies Lemma 5.3
as well. We will not discuss in detail the (technical) differences between witness trees and CWD’s.
If some CWD τ and resampling table R satisfy condition (A2) of Lemma 5.3, then we say that
τ is compatible with R. We denote by T R the set of all CWD’s compatible with R. We summarize
some key results on [10] which are relevant to us.
Lemma 5.4 ([10]). Suppose that epd1+ǫ ≤ 1 for ǫ > 0, and suppose that the functions fB can be
evaluated with complexity (U, poly(m, n)).
for some constant c > 0. There is a set T of CWD’s with the following
Let K = c log(mn/ǫ)
ǫ log d
properties:
(T1) |T | ≤ (mn/ǫ)O(1/ǫ)
(T2) Each τ ∈ T has size at most 2K.
18
DAVID G. HARRIS
(T3) Suppose there is no τ ∈ T R of size ≥ K. Then an assignment avoiding B can be found with
complexityP(KU + K log(mn|T R |) + log2 |T R |, poly(m, |T R |)).
(T4) We have τ ∈T ,|τ |≥K w(τ ) < 1/2.
P
(T5) We have τ ∈T w(τ ) < O(m).
(T6) The set T can be enumerated with quasi-complexity ( log
2
(mn/ǫ)
, (mn/ǫ)O(1/ǫ) ).
ǫ
Now consider drawing a resampling table R(i, j) where i = 1, . . . , n and j = 1, . . . , 2K. Given
any CWD τ , let f (τ, R) denote the indicator function that τ is compatible with R. We also define
X
1 X
(2)
S(R) =
f (τ, R)
f (τ, R) +
Cm
τ ∈T
τ ∈T
|τ |≥K
for some constant C > 0.
When the entries of R are drawn as independent fair coins, properties (T4) and (T5) ensure that
E[S(R)] < 1 for sufficiently large C. Also, property (T3) ensures that if S(R) < 1 then one can find
a satisfying assignment in quasi-complexity (KU + K log(mn/ǫ) + log2 (mn/ǫ), (mn)O(1/ǫ) ). Thus,
the problem of finding a satisfying assignment is reduced to the problem of minimizing S(R), which
is a sum of juntas.
Theorem 5.5. Suppose that epd1+ǫ ≤ 1 for some ǫ > 0. Suppose that each bad-event B ∈ B is
a w-junta (i.e. |YB | ≤ w). Suppose that we have a PEO for the collection of functions fB which
has complexity (C1 , poly(m, n)); namely, given any B ∈ B as well as a partial assignment to the
variables in YB , it computes the corresponding probability that fB = 1.
Then we can find x ∈ {0, 1}n avoiding B, using quasi-complexity (C1 (w + logǫmn ), (mn)O(1/ǫ) ).
Proof. We suppose that ǫ > 1/m, as otherwise we exhaustively search the space using 2n ≤
(mn)O(1/ǫ) processors. Also, we assume that C1 ≥ Ω(log mn), as it requires this much time to
simply read in the input.
The function S(R), given in (2), is by (T1) a sum of m′ = (mn)O(1/ǫ) functions f (τ, R). By (T2),
each τ ∈ T has size at most 2K; property (A4) ensures that each term fL(v) (Wv (R)) depends on
at most w bits of R, so in all each function f (τ, R) is a w′ -junta for w′ ≤ O(ǫ−1 w log mn). The
total number of variables determining R is n′ = 2nK ≤ O(ǫ−1 n log(mn)).
We claim next that we can form a PEO for the functions f (τ, R) with quasi-complexity (C1′ , C2′ ),
where C1′ = C1 + logǫmn and C2′ = (mn)O(1/ǫ) . For suppose we are given a partial assignment
query R′ . We can map this onto the sets W1 , . . . , Wt determining each τ using Õ( logǫmn ) time.
Next, the probability that any τ is compatible with R′ is simply the product of the probabilities of
fBi (Wi ). The PEO for fB allows us to compute these probabilities in parallel with quasi-complexity
((log K)C1 , (mn)O(1/ǫ) ). Finally, we multiply the probabilities together in Õ( logǫmn ) time.
w′ C ′
Now apply Theorem 3.8 to find R with S(R) ≤ E[S(R)] < 1 using quasi-complexity Õ( log m′1n′ +
2
C1′ + log w′ log(m′ n′ ), (w′ )O(1) C2′ ). We can simplify this as (wC1 + log ǫmn , (mn)O(1/ǫ) ).
Once we have found R, we use (T3) to find x avoiding B. The PEO can be used to check whether
2
mn
a given bad-event is true, so U ≤ C1 and this step requires O( C1 log
+ log ǫmn ) time.
ǫ
6. Applications of the LLL
6.1. Defective vertex coloring. A k-defective vertex c-coloring of a graph G = (V, E), is an
assignment of colors to the vertices such that every vertex v has at most k neighbors with the same
color as v. This generalizes proper vertex coloring, in that a proper vertex coloring is a 0-defective
coloring. In this section, we give an algorithm which gives a k-defective c-coloring of a graph G of
maximum degree ∆ with c = O(∆/k), for any choice of k in the range {1, . . . , ∆}. The main idea,
inspired by a similar randomized distributed algorithm of [8], is a degree-splitting step; when ∆ is
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
19
small, this can be achieved efficiently using our deterministic LLL algorithm and when ∆ is large
then we can use an alternate algorithm based on simple Chernoff bounds.
Lemma 6.1. There is an absolute constant K with the following property. Given a graph G of
∆
maximum degree ∆ and an integer parameter j ≤ log2 ( K log
∆ ), there is an algorithm with quasi2
j
complexity (∆
p log n + log n, poly(n)) to 2 -color the vertices, so that each vertex v has at most
j
∆/2 (1 + K log(∆)2j /∆)) neighbors with the same color as v.
Proof. Color applying the LLL to the random process in which each vertex independently and
uniformly selects a color (represented as a j-bit string). Each vertex v has a bad-event Bv that it
has too many neighbors of its own color. Thus there are m = n bad-events, and each bad-event
involves at most j∆ variables. The number of neighbors of each color is a Bernoulli variable with
mean at most ∆2−j . Note that Bv ∼ Bw iff v and w are at distance at most 2 in G. So in sense of
2
the LLL we have
p d≤∆ .
j
Let δ = c log(∆)2 /∆ for some constant c. For K sufficiently large, we ensure that δ ≤ 1.
2
Therefore, by the Chernoff bound, Bv has probability at most e−µδ /3 , which is smaller than ∆−4
for an appropriate choice of c. So, in the sense of the LLL, we have p ≤ ∆−4 . These parameters
satisfy Theorem 5.5 with ǫ = 1/2. Each bad-event Bv is a boolean function on at most j∆ binary
variables, and a PEO can be constructed with quasi-complexity (log n, poly(n)). Thus Theorem 5.5
gives the desired goal in quasi-complexity (j∆ log n + log2 n, poly(n)). Note that j ≤ O(log ∆), and
so it can be dropped from the quasi-complexity bounds.
Theorem 6.2. Let G be a graph with maximum degree ∆ and k ∈ {1, . . . , ∆}. Then there is an
NC algorithm running in time Õ(log2 n) to obtain a k-defective vertex c-coloring with c = O(∆/k).
Proof. When ∆ ≤ log n, we can achieve this using iterated applications of Lemma 6.1. Each
iteration reduces the degree of the residual graphs by a logarithmic factor, and so the overall
running time is close to the running time of a single application of Lemma 6.1. We defer the full
proof to Appendix A, as the construction is technical and similar to that of [8].
Finally, let us discuss what to do when ∆ ≥ log n. In this case, consider the random process of
∆
assigning every vertex a color uniformly selected from log
n ; a simple Chernoff bound shows that,
with high probability, this ensures that each vertex has at most O(log n) neighbors of each color
class. This can be derandomized by the method of Sivakumar (among other methods) [18], as there
are a polynomial number of “statistical tests” (in this case, the degree of each vertex with respect
to each color class) which can be computed in logspace. After this first coloring step, which can
be executed in O(log2 n) time, we get a multiple subgraphs with maximum degree ∆ = log n, and
apply the above process.
6.2. Domatic partition. A domatic partition of a graph is a c-coloring of the vertices of G with
the property that each vertex of G sees all c-colors in its neighborhood (including itself). That is,
for any color ℓ = 1, . . . , c, the color-ℓ vertices form a dominating set of G. An algorithm was given
in [9] using the LLL to find a domatic partition with a large number of colors. For simplicity, we
specialize their algorithm to k-regular graphs.
Theorem 6.3. Let η > 0 be any fixed constant. There is some constant K = Kη with the following
property. If G is k-regular with k > K, then G has a domatic partition of size c ≥ (1 − η) lnkk ,
which can be found using Õη (k log n + log2 n) time and nOη (1) processors.
Proof. We follow the iterated LLL construction of [9], in which the color of each vertex is an
ordered pair χ(v) = (χ1 (v), χ2 (v)); here χ1 is chosen from c1 = k/ ln3 k colors, and χ2 is chosen
20
DAVID G. HARRIS
from c2 = (1 − η) ln2 k colors. In the first phase of the LLL, we will select χ1 and the second phase
will select χ2 . Each vertex chooses its colors uniformly at random among [c1 ], [c2 ] respectively.2
Now consider the phase I coloring. For each vertex v and each color j ∈ [c1 ], define Nj (v) to the
set of neighbors w with χ1 (w) = j and let Xv,j = |Nj (v)|. The expected value of Xv,j is µ ≥ ln3 k.
For each vertex v and each color j ∈ [c1 ], we have a bad-event Bv,j that Xv,j ≤ t0 or Xv,j ≥ t1 ,
where t0 = µ − φ ln2 k and t1 = µ + φ ln2 k and φ is a large constant.
For φ sufficiently large, the Chernoff bounds shows that Bv,j has probability at most p ≤ k−5 .
Furthermore, each bad-event Bv,j affects Bv′ ,j ′ only if dist(v, v ′ ) ≤ 2, so in the sense of the LLL we
have d ≤ k4 . One can construct PEO’s for these bad-events with running time C1 = Õ(log mn).
Apply Theorem 5.5 to find χ1 with quasi-complexity (k log n + log2 n, nO(1) ).
For each vertex v, each j ∈ [c1 ], and each j ′ ∈ [c2 ], we have a bad-event Bv,j,j ′ that there
is no w ∈ Nj (v) with χ2 (w) = j ′ ; if all such bad-events are avoided then the resulting coloring
(χ1 (v), χ2 (v)) gives a domatic partition. The only dependencies now are between bad-events Bv,j,j ′
and Bw,j,j ′′ where v, w share a neighbor u with χ1 (u) = j, so d ≤ t1 kc2 and p ≤ (1 − 1/c2 )t0 .
Set ǫ = η/2, φ = 10. It is straightforward to verify that that the criterion epd1+ǫ ≤ 1 is satisfied
when k is sufficiently large. Thus, Theorem 5.5 gives a coloring avoiding the phase-II bad-events
using Õη (k log n + log2 n) time and nOη (1) processors.
7. Acknowledgments
Thanks to Aravind Srinivasan, Vance Faber, and anonymous referees for helpful comments and
discussion.
References
[1] Alon, N., Srinivasan, A.: Improved parallel approximation of a class of integer programming problems. Proceedings
of the 23rd International Colloquium on Automata, Languages, and Programming (ICALP), pp. 562-573 (1996)
[2] Alon, N., Goldreich, O., Haståd, J., Peralta, R.: Simple constructions of almost k-wise independent random
variables. Random Structures and Algorithms 3-3, pp. 289-303 (1992)
[3] Alon, N., Babai, L., Itai, A.: A fast and simple randomized parallel algorithm for the maximal independent set
problem. Journal of Algorithms 7, pp. 567-583 (1986)
[4] Berger, B., Rompel, J.: Simulating (logc n)-wise independence in NC. Journal of the ACM 38-4, pp. 1026-1046.
(1991)
[5] Borodin, A., Cook, S.: A time-space tradeoff for sorting on a general sequential model of computation. SIAM
Journal on Computing 11-2, pp. 287-297 (1982)
[6] Chandrasekaran, K., Goyal, N., Haeupler, B.: Deterministic algorithms for the Lovász local lemma. SIAM Journal
on Computing 42-6, pp. 2132-2155 (2013)
[7] Chari, S., Rohatgi, P, Srinivasan, A.: Improved algorithms via approximations of probability distributions. Journal
of Computer and System Sciences 61-1, pp. 81-107 (2000)
[8] Fischer, M., Ghaffari, M.: Sublogarithmic distributed algorithms for Lovász local lemma, and the complexity
hierarchy. Proceedings of the 31st International Symposium on Distributed Computing (DISC), pp. 18:1-18:16
(2017)
[9] Feige, U., Halldórsson, M. M., Kortsarz, G., Srinivasan, A.: Approximating the domatic number. SIAM Journal
on Computing 32-1, pp. 172-195 (2002)
[10] Haeupler, B., Harris, D.G.: Parallel algorithms and concentration bounds for the Lovász Local Lemma via
witness-DAGs. ACM Transactions on Algorithms 13-4, Article #53 (2017)
[11] Luby, M.: Removing randomness in parallel computation without a processor penalty. Journal of Computer and
System Sciences 47-2, pp. 250-286 (1993)
[12] Moser, R., Tardos, G.: A constructive proof of the general Lovász Local Lemma. Journal of the ACM 57-2, pp.
11:1-11:15 (2010)
2If c , c are not powers of two, then we cannot directly represent this in our bit-based LLL formulation. However,
1 2
we can simulate it by drawing values u1 , u2 from [2r1 , 2r2 ] for ri = log2 ci + O(log k), and projecting uniformly down
to [ci ]. This changes the probabilities of the bad-events by a negligible factor of polylog(1/k). We omit further details
for simplicity.
FOOLING POLYLOGARITHMIC JUNTAS AND THE LOVÁSZ LOCAL LEMMA
21
[13] Mulmuley, K. A fast parallel algorithm to compute the rank of a matrix over an arbitrary field. Combinatorica
7-1, pp. 101-104 (1987).
[14] Naor, J., Naor, M.: Small-bias probability spaces: efficient construction and applications, SIAM Journal of
Computing 22-4, pp. 835-856 (1993)
[15] Nisan, N.: Pseudorandom generator for space-bounded computation. Combinatorica 12-4, pp. 449-461 (1992)
[16] Nisan, N.: RL ⊆ SC. Computational Complexity 4-1, pp. 1-11 (1994)
[17] Schulman, L.: Sample spaces uniform on neighborhoods. Proceedings of the 24th ACM Symposium on Theory
of Computing (STOC), pp. 17-25 (1992)
[18] Sivakumar, D.: Algorithmic derandomization via complexity theory. Proceedings of the 34th ACM Symposium
on Theory of Computing (STOC), pp. 619-626 (2002)
Appendix A. Full proof of Theorem 6.2
We assume here that ∆ ≤ log n, and that k is larger than any needed constant. The case of large
∆ has already been discussed.
We will build the coloring gradually over stages i = 0, . . . , r + 1; at stage i, the vertices have a
ti -coloring, in which every vertex has at most ∆i neighbors of its own color class. At stage i, and
for any integer ℓ ∈ {1, . . . , ti }, let us define Gi,ℓ to be the subgraph of G induced on vertices with
color ℓ. So Gi,ℓ has maximum degree ∆i . We will apply Lemma 6.1 with parameter ji to each
Gi,ℓ ; this is done in parallel across all values of ℓ. At the end of this process, we thereby obtain a
∆r+1 -defective coloring with tr+1 colors.
We next need to define the sequence of values ji , ∆i which will be valid for the degree splitting
procedure. We do so recursively by setting ∆0 = ∆, and
(
√
⌈log2 ∆i / log2 ∆i ⌉ i < r
ji
−1/2
∆i+1 = ∆i /2 (1 + K 2(log
∆i ))
ji =
⌈log2 ∆i /k ⌉
i=r
where r is a parameter to be determined. Our coloring will now have t0 = 1, ti+1 = ti 2ji for
i = 0, . . . , r.
We need to show that these parameters satisfy the conditions given by Lemma 6.1. We just need
to ensure that Gi has maximum degree ∆i and
ji ≤ log2 (
(3)
∆i
)
K log ∆i
holds for i = 0, . . . , r.
In order to carry out this analysis, let us define a sequence of real numbers by
b0 = ∆,
bi+1 =
1
2
log2 bi
Let r be maximal such that br ≥ k; we stop this procedure at stage r + 1. We easily see that
r ≤ O(log∗ ∆). We claim that for i = 0, . . . , r we have k ≤ bi ≤ ∆i ≤ 4bi , and we show this by
induction on i. The bound bi ≥ k is immediate from the definition of r. The bound on ∆0 is
immediate. For i < r, the lower bound is shown by
∆i+1 ≥ ∆i /2ji ≥ ∆i /2⌈log 2 ∆i / log
2
∆i ⌉
≥ ∆i /21+log2 ∆i / log
2
∆i
=
1
2
log2 ∆i ≥
1
2
log2 bi = bi+1
For the upper bound, we have
√
√
∆i+1 = ∆i /2ji (1 + K 2 log−1/2 ∆i ) ≤ (1 + K 2 log−1/2 ∆i ) log2 ∆i
√
for bi ≥ k and k sufficiently large
≤ (1 + K 2 log−1/2 bi ) log2 (4bi ) ≤ 1.01 log 2 (bi ) ≤ 4bi+1
We can now show (3). For i < r, we have
2ji ≤
2∆i
∆i
2
∆i
2
∆i
≤
×
≤
× ≤
2
log ∆i log ∆i
log ∆i k
K log ∆i
log ∆i
22
DAVID G. HARRIS
For i = r, we note that br+1 ≤ k and so
2jr ≤ 2∆r /k ≤ (8br )/k ≤ 16br / log2 br ≤ 16∆r / log2 ∆r ≤ ∆r /(K log ∆r )
As condition (3) holds, Lemma 6.1 ensures that for i < r, the graph Gi+1,ℓ has maximum degree
at most
q
q
∆i /2ji (1 + K log(∆i )2ji /∆i ) ≤ (∆i /2ji )(1 + K log(∆i ) × 2∆i / log2 ∆i × ∆−1
i )
√
≤ (∆i /2ji )(1 + K 2 log−1/2 ∆i ) = ∆i+1
Finally, when i = r, Lemma 6.1 ensures that the graphs Gr+1,ℓ have maximum degree
∆r+1 ≤ ∆r /2jr (1 + O(log−1/2 ∆r )) ≤ k(1 + O(log−1/2 ∆r )) ≤ O(k)
Thus, the overall coloring we obtain is indeed O(k)-defective. Our next task is to count the
number of colors used. Let us define ai = ∆i ti . We want to show that tr+1 ≤ O(∆/k). As
∆r+1 ≥ k/2, it suffices to show that ar+1 ≤ O(∆). We may bound ai inductively as
√
√
√
ai+1 ≤ ai /2ji (1 + K 2 log−1/2 ∆i ) × (ti 2ji ) ≤ ai (1 + K 2 log−1/2 ∆i ) ≤ ai (1 + K 2 log−1/2 bi )
Therefore, we see that
ar+1 ≤ a0
r
Pr
Y
√
−1/2
bi )
≤ O(∆)
(1 + K 2 log−1/2 bi ) ≤ ∆eO( i=0 log
i=0
where the last inequality follows by noting that the sequence log bi is decreasing super-exponentially.
Now that we have finished the construction in the case ∆ ≤ log n, let us discuss the runtime. Each
iteration requires (∆i log n + log2 n, poly(n)) quasi-complexity. We see easily that ∆i ≤ O(∆) ≤
O(log n), so this is Õ(log2 n) time per iteration. There are r ≤ O(log∗ ∆) iterations, giving a total
runtime again of Õ(log2 n).
| 8cs.DS
|
Under review as a conference paper at ICLR 2018
A C LASSIFICATION –BASED P ERSPECTIVE ON
GAN D ISTRIBUTIONS
arXiv:1711.00970v3 [cs.LG] 19 Nov 2017
Shibani Santurkar, Ludwig Schmidt, Aleksander Madry
˛
Massachusetts Institute of Technology
{shibani,ludwigs,madry}@mit.edu
A BSTRACT
A fundamental, and still largely unanswered, question in the context of Generative Adversarial Networks (GANs) is whether GANs are actually able to capture
the key characteristics of the datasets they are trained on. The current approaches
to examining this issue require significant human supervision, such as visual inspection of sampled images, and often offer only fairly limited scalability. In this
paper, we propose new techniques that employ a classification–based perspective
to evaluate synthetic GAN distributions and their capability to accurately reflect
the essential properties of the training data. These techniques require only minimal human supervision and can easily be scaled and adapted to evaluate a variety
of state-of-the-art GANs on large, popular datasets. Our analysis indicates that
GANs have significant problems in reproducing the more distributional properties
of the training dataset. In particular, when seen through the lens of classification,
the diversity of GAN data is orders of magnitude less than that of the original data.
1
I NTRODUCTION
Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) have garnered a significant
amount of attention due to their ability to learn generative models of multiple natural image datasets
(Radford et al., 2015; Denton et al., 2015; Zhang et al., 2016). Since their conception, a fundamental
question regarding GANs is to what extent they truly learn the underlying data distribution. This
is a key issue for multiple reasons. From a scientific perspective, understanding the capabilities of
common GANs can shed light on what precisely the adversarial training setup allows the GAN to
learn. From an engineering standpoint, it is important to grasp the power and limitations of the
GAN framework when applying it in concrete applications. Due to the broad potential applicability
of GANs, researchers have investigated this question in a variety of ways.
When we evaluate the quality of a GAN, an obvious first check is to establish that the generated
samples lie in the support of the true distribution. In the case of images, this corresponds to checking
if the generated samples look realistic. Indeed, visual inspection of generated images is currently
the most common way of assessing the quality of a given GAN. Individual humans can performs
this task quickly and reliably, and various GANs have achieved impressive results for generating
realistic-looking images of faces and indoor scenes (Salimans et al., 2016; Denton et al., 2015).
Once we have established that GANs produce realistic-looking images, the next concern is that the
GAN might simply be memorizing the training dataset. While this hypothesis cannot be ruled out
entirely, there is evidence that GANs perform at least some non-trivial modeling of the unknown
distribution. Previous studies show that interpolations in the latent space of the generator produce
novel and meaningful image variations (Radford et al., 2015), and that there is a clear disparity
between generated samples and their nearest neighbors in the true dataset (Arora & Zhang, 2017).
Taken together, these results provide evidence that GANs could constitute successful distribution
learning algorithms, which motivates studying their distributions in more detail. The direct approach
is to compare the probability density assigned by the generator with estimates of the true distribution
(Wu et al., 2016). However, in the context of GANs for high-dimensional image distributions, this
is complicated by two factors. First, GANs do not naturally provide probability estimates for their
samples. Second, estimating the probability density of the true distribution is a challenging problem
1
Under review as a conference paper at ICLR 2018
itself (the adversarial training framework specifically avoids this issue). Hence prior work has only
investigated the probability density of GANs on simple datasets such as MNIST (Wu et al., 2016).
Since reliably computing probability densities in high dimensions is challenging, we can instead
study the behavior of GANs in low-dimensional problems such as two-dimensional Gaussian mixtures. Here, a common failure of GANs is mode collapse, wherein the generator assigns a disproportionately large mass to a subset of modes from the true distribution (Goodfellow, 2016). This raises
concerns about a lack of diversity in the synthetic GAN distributions, and recent work shows that
the learned distributions of two common GANs indeed have (moderately) low support size for the
CelebA dataset (Arora & Zhang, 2017). However, the approach of Arora & Zhang (2017) heavily
relies on a human annotator in order to identify duplicates. Hence it does not easily scale to comparing many variants of GANs or asking more fine-grained questions than collision statistics. Overall,
our understanding of synthetic GAN distributions remains blurry, largely due to the lack of versatile
tools for a quantitative evaluation of GANs in realistic settings. The focus of this work is precisly to
address this question:
Can we develop principled and quantitative approaches to study synthetic GAN distributions?
To this end, we propose two new evaluation techniques for GAN distributions. Our methods are
inspired by the idea of comparing moments of distributions, which is at the heart of many methods in classical statistics. Although simple moments of high-dimensional distributions are often not
semantically meaningful, we can extend this idea to distributions of realistic images by leveraging
image statistics identified using convolutional neural networks. In particular, we train image classifiers in order to construct test functions corresponding to semantically meaningful properties of the
distributions. An important feature of our approach is that it requires only light human supervision
and can easily be scaled to evaluating many GANs and large synthetic datasets.
Using our new evaluation techniques, we study five state-of-the-art GANs on the CelebA and LSUN
datasets, arguably the two most common testbeds for advanced GANs. We find that most of the
GANs significantly distort the relative frequency of even basic image attributes, such as the hair
style of a person or the type of room in an indoor scene. This clearly indicates a mismatch between
the true and synthetic distributions. Moreover, we conduct experiments to explore the diversity of
GAN distributions. We use synthetic GAN data to train image classifiers and find that these have
significantly lower accuracy than classifiers trained on the true data set. This points towards a lack
of diversity in the GAN data, and again towards a discrepancy between the true and synthetic distributions. In fact, our additional examinations show that the diversity in GANs is only comparable to
a subset of the true data that is 100× smaller.
2
U NDERSTANDING GAN S THROUGH THE L ENS OF C LASSIFICATION
When comparing two distributions, a common first test is to compute low-order moments such as
the mean and the variance. If the distributions are simple enough, these quantities provide a good
understanding for how similar they are. Moreover, low-order moments have a precise definition and
are usually quick to compute. On the other hand, low-order moments can also be misleading for
more complicated, high-dimensional distributions. As a concrete example, consider a generative
model of digits (such as MNIST). If a generator produces digits that are shifted by a significant
amount yet otherwise perfect, we will probably still consider this as a good approximation of the
true distribution. However, the expectation (mean moment) of the generator distribution can be very
different from the expectation of the true data distribution. This raises the question of what other
properties of high-dimensional image distributions are easy to test yet semantically meaningful.
In the next two subsections, we describe two concrete approaches to evaluate synthetic GAN data
that are easy to compute yet capture relevant information about the distribution. The common theme
is that we employ convolutional neural networks in order to capture properties of the distributions
that are hard to describe in a mathematically precise way, but usually well-defined for a human (e.g.,
what fraction of the images shows a smiling person?). Automating the process of annotating images
with such high-level information will allow us to study various aspects of synthetic GAN data.
2.1
Q UANTIFYING M ODE C OLLAPSE
Mode collapse refers to the tendency of the generator to concentrate a large probability mass on a
few modes of the true distribution. While there is ample evidence for the presence of mode-collapse
2
Under review as a conference paper at ICLR 2018
in GANs (Goodfellow, 2016; Arora & Zhang, 2017; Metz et al., 2016), elegant visualizations of this
phenomena are somewhat restricted to toy problems on low-dimensional distributions (Goodfellow,
2016; Metz et al., 2016). For image datasets, it is common to rely on human annotators and therein
derived heuristics (see Section 2.3). While these methods have their merits, they are restrictive both
in the scale and granularity of testing. Here we propose a classification-based tool to assess how
good GANs are at assigning the right mass across broad concepts/modes. To do so, we use a trained
classifier as an expert “annotator” that labels important features in synthetic data, and then analyze
the resulting distribution. Specifically, our goal is to investigate if a GAN trained on a well-balanced
dataset (i.e., contains equal number of samples from each class) can learn to reproduce this balanced
N
structure. Let D = (X, Y ) = {(xi , yi )}i=1 represent a dataset of size N with C classes, where
(xi , yi ) denote an image-label pair drawn from true data. If the dataset D is balanced, it contains
N/C images per class. The procedure for computing class distribution in synthetic data is:
1. Train an annotator (a multi-class classifier) using the dataset D.
2. Train an unconditional GAN on the images X from dataset D, without using class labels.
3. Create a synthetic dataset by sampling N images from a GAN and labeling them using the
annotator from Step 1.
The annotated data generated via the above procedure can provide insight into the GAN’s class
distribution at the scale of the entire dataset. Moreover, we can vary the granularity of mode analysis
by choosing richer classification tasks, i.e., more challenging classes or a larger number of them. In
Section 3.3, we use this technique to visualize mode collapse in several state-of-the-art GANs on the
CelebA and LSUN datasets. All the studied GANs show significant mode collapse and the effect
becomes more pronounced when the granularity of the annotator is increased (larger number of
classes). We also investigate the temporal aspect of the GAN setup and find that the dominant mode
varies widely over the course of training. Our approach also enables us to benchmark and compare
GANs on different datasets based on the extent of mode collapse in the learned distributions.
2.2
M EASURING DIVERSITY
Our above method for inspecting distribution of modes in synthetic data provides a coarse look at the
statistics of the underlying distribution. While the resulting quantities are semantically meaningful,
they capture only simple notions of diversity. To get a more holistic view on the sample diversity
in the synthetic distribution, we now describe a second classification-based approach for evaluating
GAN distributions. The main question that motivates it is: Can GANs recover the key aspects of
real data to enable training a good classifier? We believe that this is an interesting measure of sample diversity for two reasons. First, classification of high-dimensional image data is a challenging
problem, so a good training dataset will require a sufficiently diverse sample from the distribution.
Second, augmenting data for classification problems is one of the proposed use cases of GANs (e.g.,
see the recent work of Shrivastava et al. (2017)).
If GANs are truly able to capture the quality and diversity of the underlying data distribution, we
expect almost no gap between classifiers trained on true data and synthetic data from a GAN. A
generic method to produce data from GANs for classification is to train separate GANs for each
class in the dataset D.1 Samples from these class-wise GANs can then be pooled together to get a
labeled synthetic dataset. Note that the labels are trivially determined based on the class modeled
by the particular GAN from which a sample is drawn. We perform the following steps to assess the
classification performance of synthetic data vs. true data:
1. Train a classifier on the true data D (from Section 2.1) as a benchmark for comparison.
2. Train C separate unconditional GANs, one per class in dataset D.
3. Generate a balanced synthetic labeled dataset of size N by consolidating an equal number of
samples drawn from each of these C GANs. The labels obtained by aggregating samples from
per-class GANs are designated as “default" labels for the synthetic dataset. Note that by design,
both true and synthetic datasets have N samples, with N/C examples per class.
4. Use synthetic labeled data from Step 3 to train classifier with the same architecture as Step 1.
Comparing the classifiers from Steps 1 and 4 can then shed light on the disparity between the two
distributions. Radford et al. (2015) conducted an experiment similar to Step 2 on the MNIST dataset
1
We tried the alternate approach of using class-conditional GANs to get labeled datasets. This method did
not yield good samples since most common GANs have not been designed with a conditional structure in place.
3
Under review as a conference paper at ICLR 2018
using a conditional GAN. They found that samples from their DCGAN performed comparably to
true data on nearest neighbor classification. We obtained similar good results on MNIST, which
could be due to the efficacy of GANs in learning the MNIST distribution or due to the ease of
getting good accuracy on MNIST even with a small training set (Rolnick et al., 2017). To clarify
this question, we restrict our analysis to more complex datasets, specifically CelebA and LSUN.
We evaluate the two following properties in our classification task:
(i) How well can the GANs recover nuances of the decision boundary, which is reflected by how
easily the classifier can fit the training data?
(ii) How does the diversity of synthetic data compare to that of true data when measured by classification accuracy on a hold-out set of true data?
We observe that all the studied GANs have very low diversity in this metric. In particular, the
accuracy achieved by a classifier trained on GAN data is comparable only to the accuracy of a
classifier trained on a 100× (or more) subsampled version of the true dataset. Even if we draw more
samples from the GANs to produce a training set several times larger than the true dataset, there is
no improvement in performance. Looking at the classification accuracy gives us a way to compare
different models on a potential downstream application of GANs. Interestingly, we find that visual
quality of samples does not necessarily correlate with good classification performance.
2.3
R ELATED W ORK
In GAN literature, it is common to investigate performance using metrics that involve human supervision. Arora & Zhang (2017) proposed a measure based on manually counting duplicates in GAN
samples as a heuristic for the support or diversity of the learned distribution. In Wu et al. (2016),
manual classification of a small sample (100 images) of GAN generated MNIST images is used as
a test for the GAN is missing certain modes. Such annotator-based metrics have clear advantages
in identifying relevant failure-modes of synthetic samples, which explains why visual inspection
(eyeballing) is still the most popular approach to assess GAN samples.
There have also been various attempts to build good metrics for GANs that are not based on manual
heuristics. Parzen window estimation can be used to approximate the log-likelihood of the distribution, though it is known to work poorly for high-dimensional data (Theis et al., 2016). Wu et al.
(2016) develop a method to get a better estimate for log-likelihood using annealed importance sampling. Salimans et al. (2016) propose a metric known as Inception Score, where the entropy in the
labels predicted by a pre-trained Inception network is used to assess the diversity in GAN samples.
3
E XPERIMENTS
In the following sub-sections we describe the setup and results for our classification-based GAN
benchmarks. Additional details can be found in Section 5 in the Appendix.
3.1
DATASETS
GANs have shown promise in generating realistic samples, resulting in efforts to apply them to a
broad spectrum of datasets. However, the Large-scale CelebFaces Attributes (CelebA) (Liu et al.,
2015) and Large-Scale Scene Understanding (LSUN) (Yu et al., 2015) datasets remain the most
popular and canonical ones in developing and evaluating GAN variants. Conveniently, these datasets
also have rich annotations, making them particularly suited for our classification–based evaluations.
Details on the setup for classification tasks for these datasets are given in the Appendix.
3.2
M ODELS
Using our framework, we perform a comparative study of several popular variants of GANs:
1. Deep Convolutional GAN (DCGAN): Convolutional GAN trained using a Jensen–Shannon
divergence–based objective (Goodfellow et al., 2014; Radford et al., 2015).
2. Wasserstein GAN (WGAN): GAN that uses a Wasserstein distance–based objective (Arjovsky
et al., 2017).
3. Adversarially Learned Inference (ALI): GAN that uses joint adversarial training of generative
and inference networks (Dumoulin et al., 2017).
4
Under review as a conference paper at ICLR 2018
Female
+
Mouth Closed
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
Female
+
Mouth Open
Male
+
Mouth Closed
Male
+
Mouth Open
4-class dataset from CelebA for attributes Male, Mouth open.
Figure 1: Illustration of a dataset used in the proposed classification benchmarks. Shown alongside
are images sampled from various unconditional GANs trained on this dataset. Labels for the GAN
samples are obtained using a pre-trained classifier as an annotator.
4. Boundary Equilibrium GAN (BEGAN): Auto-encoder style GAN trained using Wasserstein distance objective (Berthelot et al., 2017).
5. Improved GAN (ImGAN): GAN that uses semi-supervised learning (labels are part of GAN
training), with various other architectural and procedural improvements (Salimans et al., 2016).
All the aforementioned GANs are unconditional, however, ImGAN has access to class labels as a
part of the semi-supervised training process. We use standard implementations for each of these
models, details of which are provided in the Appendix (Section 5). We also used the prescribed
hyper-parameter settings for each GAN, including number of iterations we train them for. Our
analysis is based on 64 × 64 samples, which is a size at which GAN generated samples tend to be of
high quality. We also use visual inspection to ascertain that the perceptual quality of GAN samples
in our experiments is comparable to those reported in previous studies. We demonstrate sample
images in Figures 1, 4 and 5. BEGAN did not converge in our experiments on the LSUN dataset
and hence is excluded from the corresponding analysis.
In our study, we use two types of classification models:
1. ResNet: 32-Layer Residual network He et al. (2016). This model is chosen as it is a standard
classifier in vision and yields high accuracy on various datasets, making it a reliable baseline.
2. Linear Model: This is a network with one-fully connected layer between the input and output (no
hidden layers) with a softmax non-linearity. If the dimensions of input x and output ŷ, are D and
C (number of classes) respectively, then linear models implement the function ŷ = σ(W T x+b),
where W is a D × C matrix, b is a C × 1 vector and σ(·) is the softmax function. Due to it’s
simplicity, this model will serve as a useful baseline in some of our experiments.
We always train the classifiers to convergence, with decaying learning rate and no data augmentation.
3.3
E XAMINATION OF M ODE C OLLAPSE
Experimental results for quantifying mode collapse through classification tasks, described in Section 2.1, are presented below. Table 1 gives details on datasets (subsets of CelebA and LSUN)
used in our analysis, such as size (N ), number of classes (C), and accuracy of the annotator, i.e.,
a classifier pre-trained on true data, which is then used to label the synthetic, GAN-generated data.
Figure 2 presents class distribution in the synthetic data, as determined by using these annotators.
The left panel compares the relative distribution of modes in true data (uniform) with that in various
GAN-generated datasets. Each of these datasets is created by drawing N samples from the GAN
after it was trained on the corresponding true dataset. The right panel illustrates the evolution of
class distributions in various GANs over the course of training2 .
Results: These visualization lead to the following findings:
2
Temporal evaluation of ALI class distribution is absent in the analysis due to absence of periodic checkpointing provisions in the code.
5
Under review as a conference paper at ICLR 2018
Dataset
CelebA: Makeup, Smiling
CelebA: Male, Mouth Open
CelebA: Bangs, Smiling
LSUN: Bedroom, Kitchen, Classroom
LSUN: Bedroom, Conference Room, Dining Room,
Kitchen, Living Room
N
102,436
115,660
45,196
150,000
250,000
C
4
4
4
3
5
Annotator’s Accuracy (%)
90.9, 92.4
97.9, 93.5
93.9, 92.4
98.7
93.7
Table 1: Details of CelebA and LSUN subsets used for the studies on mode collapse in Section 3.3.
Here, we use a classifier trained on true data as an annotator that let’s us infer label distribution for
the synthetic, GAN-generated data. N is the training set size and C is the number of classes for both
the true and synthetic datasets. Annotator’s accuracy refers to the accuracy of the classifier on a test
set of true data. For CelebA, we use a combination of attribute-wise binary classifiers as annotators
due their higher accuracy compared to a single classifier trained jointly on all the four classes.
CelebA: Makeup, Smiling
CelebA: Male, Mouth Open
CelebA: Bangs, Smiling
LSUN: Bedroom, Kitchen, Classroom
LSUN: Bedroom, Conference Room, Dining Room, Kitchen, Living Room
Figure 2: Visualizations of mode collapse in the synthetic, GAN-generated data produced after training on our chosen subsets of CelebA and LSUN datasets. Left panel shows the relative distribution
of classes in samples drawn from synthetic datasets extracted at the end of the training process, and
compares is to the true data distribution (leftmost plots). On the right, shown is the evolution of analogous class distribution for different GANs over the course of training. BEGAN did not converge
on the LSUN tasks and hence is excluded from the corresponding analysis.
6
Under review as a conference paper at ICLR 2018
• All GANs seem to suffer from significant mode-collapse. This becomes more apparent when the
annotator granularity is increased, by considering a larger set of classes. For instance, one should
compare the relatively balanced class distributions in the 3-class LSUN task to the near-absence
of some modes in the 5-class task.
• Mode collapse is prevalent in GANs throughout the training process, and does not seem to recede
over time. Instead the dominant mode(s) often fluctuate wildly over the course of the training.
• For each task, often there is a common set of modes onto which distinct GANs exhibit collapse.
In addition to viewing our method as an approach to analyze the mode collapse, we can also use it
as a benchmark for GAN comparison. From this perspective, we can observe that ALI consistently
shows lowest mode collapse amongst all the studied GANs. The behavior of the other GANs appears
to vary based on the dataset – on CelebA, DCGAN learns a somewhat balanced distributions, while
WGAN, BEGAN and ImGAN show prominent mode collapse. This is in contrast to the results
obtained LSUN, where, for example, WGAN exhibit relatively small mode collapse, while DCGAN
and ImGAN show significant mode collapse. This points to a general challenge in real world applications of GANs: they often perform well on the datasets they were designed for (e.g. WGAN on
LSUN), but extension to new datasets is not straightforward. Temporal analysis of mode-collapse
shows that there is wide variation in the dominant mode for WGAN and Improved GAN, whereas
for BEGAN, the same mode(s) often dominates the entire training process.
3.4
D IVERSITY E XPERIMENTS
Using the procedure outlined in Section 2.2, we perform a quantitative assessment of sample diversity in GANs on the CelebA and LSUN datasets. We restrict our experiments to binary classification
as we find they have sufficient complexity to highlight the disparity between true and synthetic data.
Results for classification-based evaluation of GANs are presented in Table 2 and Figure 3.
As a preliminary check, we inspect the quality of our labeled GAN datasets. For this, we use
high-accuracy annotators from Section 2.1 to predict labels for GAN generated data and measure
consistency between the predicted and default labels (label correctness). We also inspect confidence
scores, defined as the softmax probabilities for predicted class, of the annotator. The motivation
behind these metrics is that if the classifier can correctly and with high-confidence predict labels
for labeled GAN samples, then it is likely that they are convincing examples of that class, and
hence of good “quality". Empirical results for label agreement and annotator confidence of GAN
generated datasets are shown in Table 2, and Figure 6 in the Appendix. In Table 2, we also report
an equivalent Inception Score (Salimans et al., 2016), similar to that described in Section 2.3. Using
the Inception network to get the label distribution may not be meaningful for face or scene images.
Instead, we compute the Inception Score using the label distribution predicted from the annotator
networks. Score is computed as exp(Ex [KL(p(y|x))||p(y)]), where y refers to label predictions
from the annotators 3 .
Next, we train classifiers using the true and labeled GAN-generated datasets and study their performance in terms of accuracy on a hold-out set of true data. ResNets (and other deep variants) yield
good classification performance on true data, but suffer from severe overfitting on the synthetic data,
leading to poor test accuracy (see Table 2). This already indicates a possible problem with GANs
and the diversity of the data they generate. But to highlight this problem better and avoid the issues
that stem from overfitting, we also look for a classifier which does not always overfit on the synthetic
data. We, however, observed that even training simple networks, such as one fully connected layer
with few hidden units, led to overfitting on synthetic data. Hence, we resorted to a very basic linear
model described in Section 3.2. Table 2 shows results from binary classification experiments using
linear models, with the training and test accuracies of the classifier on various datasets.
Finally, to get a better understanding of the underlying "diversity" of synthetic datasets, we train
linear models using down-sampled versions of true data (no augmentation), and compare this to
the performance of synthetic data, as shown in Table 2. Down-sampling the data by a factor of
M , denoted as ↓M implies selecting a random N/M subset of the data D. Visualizations of how
GAN classification performance compares with (down-sampled) true data are in Figure 3. Here,
the bold curve shows test accuracy of a classifier trained on true data as a function of true dataset
3
Code from https://github.com/openai/improved-gan
7
Under review as a conference paper at ICLR 2018
Task
CelebA
Male (Y/N)
# Images: 136522
CelebA
Smiling (Y/N)
# Images: 156160
CelebA
Black Hair (Y/N)
# Images: 77812
LSUN
Bedroom/Kitchen
# Images: 200000
Data Source
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True↓512
True ↓1024
True ↓2048
True ↓4096
DCGAN
WGAN
ALI
Improved GAN
Classification Performance
Label
Accuracy (%)
Inception Score
Correctness
Linear model
(%)
(µ ± σ)
↑1
↑10
Train Test Train Test
88.1 88.8
89.6 88.7
97.9
1.98 ± 0.0033 91.6 86.9
96.3 83.8
100.0 83.1
98.2
1.97 ± 0.0013 100.0 79.2 100.0 79.6
98.3
1.97 ± 0.0013 96.7 84.0 96.7 83.9
99.2
1.99 ± 0.0008 95.8 86.7 95.8 86.7
99.3
1.99 ± 0.0006 97.9 78.0 98.0 78.2
99.8
1.99 ± 0.0004 100.0 75.6 100.0 71.0
85.7 85.6
87.6 85.0
92.4
1.69 ± 0.0074 91.5 82.4
93.7 80.0
95.0 76.2
96.1
1.67 ± 0.0028 100.0 77.1 100.0 77.1
98.2
1.68 ± 0.0031 96.8 83.4 96.8 83.5
93.3
1.71 ± 0.0027 94.5 80.1 95.0 82.4
93.5
1.74 ± 0.0028 98.5 69.5 98.5 69.6
98.4
1.88 ± 0.0021 100.0 70.2 100.0 70.1
76.4 76.5
79.7 75.4
84.5
1.68 ± 0.0112 86.3 72.6
89 68.7
100.0 65.4
86.7
1.68 ± 0.0040 100.0 70.9 100.0 70.5
76.0
1.60 ± 0.0055 94.4 73.7 94.3 73.4
79.4
1.63 ± 0.0028 94.9 71.0 94.9 70.2
87.6
1.74 ± 0.0028 94.1 67.6 94.1 67.7
86.7
1.64 ± 0.0045 100.0 70.3 100.0 69.1
64.7 64.1
64.7 64.0
98.2
1.94 ± 0.0217 65.2 64.0
98.7 56.2
100.0 55.1
92.7
1.85 ± 0.0036 90.8 56.5 91.2 56.3
87.8
1.70 ± 0.0023 86.2 58.2 96.3 54.1
94.1
1.86 ± 0.0021 82.5 60.0 82.0 59.7
84.2
1.68 ± 0.0030 91.6 55.9 90.8 56.5
ResNet
↑1
Test
97.9
92.9
89.8
82.6
81.4
56.4
50.0
58.9
55.4
71.7
92.4
87.8
82.1
77.8
71.2
63.3
65.3
55.8
64.1
61.6
84.5
80.0
75.8
73.9
72.7
53.4
58.5
55.7
67.2
70.2
99.1
76.4
66.9
56.5
55.1
51.2
55.7
56.2
51.2
Table 2: Results from a comparative study on classification performance of true data vs. GAN generated synthetic data for the CelebA and LSUN datasets. Label correctness measures the agreement
between default labels for the synthetic datasets, and those predicted by the annotator, a classifier
trained on the true data. Shown alongside are the equivalent inception scores computed using labels
predicted by the annotator (instead of the Inception Network). Training and test accuracies for a
linear model classifier on the various true and synthetic datasets are reported. Also presented are the
corresponding accuracies for a linear model trained on down-sampled true data (↓M ) and oversampled synthetic data (↑L ). Test accuracy for ResNets trained on these datasets is also shown (training
accuracy was always 100%), though it is noticeable that deep networks suffer from issues when
trained on synthetic datasets.
8
Under review as a conference paper at ICLR 2018
CelebA: Male/Female
CelebA: Smiling/Not Smiling
LSUN: Bedroom/Kitchen
CelebA: Black Hair/Not Black Hair
Figure 3: Illustration of the classification performance of true data vs. GAN-generated synthetic
data. Classification is performed using a basic linear model, and performance is reported in terms of
accuracy on the test set of true data. The bold curve shows variation of the classification performance
of models trained on true data with the size of the training set (maximum size is N ). Here, training
sets of various sizes are obtained by down-sampling the true dataset at random. Dashed lines show
the performance of classifiers trained on various GAN-generated datasets of size N .
size (maximum size being N ). The dashed curves show test accuracy of classifiers trained on GAN
datasets, obtained by drawing N samples from GANs at the culmination of the training process.
A natural argument in the defense of GANs is that we can oversample them, i.e. generate datasets
much larger than the size of training data. Results for linear models trained using a 10-fold oversampling of GANs (drawing 10N samples), denoted by ↑10 , are show in Table 2.
Results: The major findings from these experiments are:
• Based on Table 2, and Figure 6, we see strong agreement between annotator labels and true
labels for synthetic data, on par with the scores for the test set of true data. It is thus apparent that
the GAN images are of high-quality, as expected based on the visual inspection. These scores
are lower for LSUN than CelebA, potentially due to lower quality of generated LSUN images.
From these results, we can get a broad understanding of how good GANs are at producing
convincing/representative samples from different classes across datasets. This also shows that
simple classification-based benchmarks can highlight relevant properties of synthetic datasets.
• The equivalent inception score is not very informative and is similar for the true (test set) and
synthetic datasets. This is not surprising given the simple nature of our binary classification task
and the fact that the true and synthetic datasets have almost a uniform distribution over labels.
• It is evident from Table 2 that there is a large performance gap between true and synthetic data
on classification tasks. Inspection of training accuracies shows that linear models are able to
nearly fit the synthetic datasets, but are grossly underfitting on true data. Given the high scores
of synthetic data on the previous experiments to assess dataset ‘quality’ (Table 2 and Figure 6),
it is likely that the poor classification performance is more indicative of lack of ‘diversity’.
• Comparing GAN performance to that of down-sampled true data reveals that the learned distribution, which was trained on datasets that have around hundred thousand data points exhibits
diversity that is on par with what only mere couple of hundreds of true data samples constitute! This shows that, at least from the point of view of classification, the diversity of the GAN
generated data is severely lacking.
9
Under review as a conference paper at ICLR 2018
• Oversampling GANs by 10-fold to produce larger datasets does not improve classification performance. The disparity between true and synthetic data remains nearly unchanged even after
this significant oversampling, further highlighting the lack of diversity in GANs.
In terms of the conclusions of relative performance of various GANs, we observe that WGAN and
ALI perform consistently better than the other GANs on both the CelebA and LSUN datasets. While
BEGAN samples have good perceptual quality (see Figures 1 and 4), it performs badly on our
classification tasks. On the other hand, WGAN samples have relatively poor visual quality but seem
to outperform most other GANs in classification tasks. This is a strong indicator of the need to
consider other metrics, such as the ones proposed in this paper, in addition to visual inspection to
study GANs. For LSUN, the gap between true and synthetic data is much larger, with the classifiers
getting near random performance on all the synthetic datasets. Note that these classifiers get poor
test accuracy on LSUN but are not overfitting on the training data. In this case, we speculate the
lower performance could be due to both lower quality and diversity of LSUN samples.
In summary, our key experimental finding is that even simple classification–based tests can hold
tremendous potential to shed insight on the learned distribution in GANs. This not only helps us
to get a deeper understanding of many of the underlying issues, but also provides with a more
quantitative and rigorous platform on which to compare different GANs. Our techniques could,
in principle, be also applied to assess other generative models such as Variational Auto-Encoders
(VAEs) Kingma & Welling (2014). However, VAEs have significant problems in generating realistic
samples on the datasets used in our analysis in the first place – see Arora & Zhang (2017).
4
C ONCLUSIONS
In this paper, we put forth techniques for examining the ability of GANs to capture key characteristics of the training data, through the lens of classification. Our tools are scalable, quantitative and
automatic (no need for visual inspection of images). They thus are capable of studying state-ofthe-art GANs on realistic, large-scale image datasets. Further, they serve as a mean to perform a
nuanced comparison of GANs and to identify their relative merits, including properties that cannot
be discerned from mere visual inspection.
We used the developed techniques to perform empirical studies on popular GANs on the CelebA and
LSUN datasets. Our examination showed that mode collapse is indeed a prevalent issue for GANs.
Also, we observed that synthetic GAN-generated datasets have significantly reduced diversity, at
least when examined from a classification perspective. In fact, the diversity of such synthetic data is
often few orders of magnitude smaller than that of the true data. Furthermore, this gap in diversity
does not seem to be bridged by simply producing much larger datasets by oversampling GANs.
Also, our methods can be viewed as benchmarks for GANs. From this perspective, we observed
that ALI consistently outperforms all the other studied GANs, both in terms of the extent of modecollapse, as well as, the diversity of the learned distribution when assessed from the classification
point-of-view. Finally, we noticed that good perceptual quality of samples does not necessarily correlate – and might sometimes even anti-correlate – with distribution diversity. These findings suggest
that we need to go beyond the visual inspection–based evaluations and look for more quantitative
tools for assessing quality of GANs, such as the ones presented in this paper.
10
Under review as a conference paper at ICLR 2018
R EFERENCES
Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks.
In International Conference on Machine Learning (ICML), pp. 214–223, 2017.
Sanjeev Arora and Yi Zhang. Do GANs actually learn the distribution? An empirical study. arXiv
preprint arXiv:1706.08224, 2017.
David Berthelot, Tom Schumm, and Luke Metz. BEGAN: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017.
Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a Laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems
(NIPS), pp. 1486–1494, 2015.
Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martin Arjovsky, Olivier Mastropietro, and Aaron Courville. Adversarially learned inference. In International Conference on Learning Representations (ICLR), 2017.
Ian Goodfellow.
Nips 2016 tutorial:
arXiv:1701.00160, 2016.
Generative adversarial networks.
arXiv preprint
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NIPS), pp. 2672–2680, 2014.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference
on Learning Representations (ICLR), 2014.
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.
In International Conference on Computer Vision (ICCV), 2015.
Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial
networks. arXiv preprint arXiv:1611.02163, 2016.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive
label noise. arXiv preprint arXiv:1705.10694, 2017.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training GANs. In Advances in Neural Information Processing Systems
(NIPS), pp. 2234–2242, 2016.
Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb.
Learning from simulated and unsupervised images through adversarial training. In Computer
Vision and Pattern Recognition (CVPR), July 2017.
L. Theis, A. van den Oord, and M. Bethge. A note on the evaluation of generative models. In
International Conference on Learning Representations (ICLR), 2016.
Yuhuai Wu, Yuri Burda, Ruslan Salakhutdinov, and Roger Grosse. On the quantitative analysis of
decoder-based generative models. arXiv preprint arXiv:1611.04273, 2016.
Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. LSUN: Construction
of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint
arXiv:1506.03365, 2015.
Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaolei Huang, Xiaogang Wang, and Dimitris
Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial
networks. arXiv preprint arXiv:1612.03242, 2016.
11
Under review as a conference paper at ICLR 2018
5
A PPENDIX
5.1
E XPERIMENTAL S ETUP
5.1.1
DATASETS FOR C LASSIFICATION TASKS
To assess GAN performance from the perspective of classification, we construct a set of classification tasks on the CelebA and LSUN datasets. In the case of the LSUN dataset, images are annotated
with scene category labels, which makes it straightforward to use this data for binary and multiclass classification. On the other hand, each image in the CelebA dataset is labeled with 40 binary
attributes. As a result, a single image has multiple associated attribute labels. Here, we can construct classification tasks by considering binary combinations of an attribute(s) (examples are shown
in Figures 1 and 4). Attributes used in our experiments were chosen such that the resulting dataset
was large, and classifiers trained on true data got high-accuracy so as to be good annotators for the
synthetic data.
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
No Makeup
+
Not Smiling
No Makeup
+
Smiling
Makeup
+
Not Smiling
Makeup
+
Smiling
(a) 4-class dataset from CelebA for attributes Makeup, Smiling.
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
No Bangs
+
Not Smiling
No Bangs
+
Smiling
Bangs
+
Not Smiling
Bangs
+
Smiling
(b) 4-class dataset from CelebA for attributes Bangs, Smiling.
Figure 4: Illustration of datasets from CelebA used in proposed classification-based benchmarks to
evaluate GANs. Shown alongside are images sampled from various unconditional GANs trained on
this dataset. Labels for the GAN samples are obtained using a pre-trained classifier as an annotator.
5.1.2
M ODELS
Benchmarks were performed on standard implementations •
•
•
•
•
DCGAN: https://github.com/carpedm20/DCGAN-tensorflow
WGAN: https://github.com/martinarjovsky/WassersteinGAN
ALI: https://github.com/IshmaelBelghazi/ALI
BEGAN :https://github.com/carpedm20/BEGAN-tensorflow
Improved GAN: https://github.com/openai/improved-gan
12
Under review as a conference paper at ICLR 2018
True Data
DCGAN
WGAN
ALI
Improved GAN
Bedroom
Classroom
Kitchen
(a) 3-class dataset from LSUN for Bedroom, Classroom, Kitchen.
True Data
DCGAN
WGAN
ALI
Improved GAN
Bedroom
Conference
Room
Dining
Room
Kitchen
Living
Room
(b) 5-class dataset from LSUN for Bedroom, Conference Room, Dining Room, Kitchen, Living Room.
Figure 5: Illustration of datasets from LSUN used in proposed classification-based benchmarks to
evaluate GANs. Shown alongside are images sampled from various unconditional GANs trained on
this dataset. Labels for the GAN samples are obtained using a pre-trained classifier as an annotator.
• ResNet Classifier: Variation of the standard TensorFlow ResNet https://github.com/
tensorflow/models/blob/master/research/resnet/resnet_model.py
5.2
5.2.1
A DDITIONAL E XPERIMENTAL R ESULTS
S AMPLE Q UALITY
For each of our benchmark experiments, we ascertain that the visual quality of samples produced
by the GANs is comparable to that reported in prior work. Examples of random samples drawn for
multi-class datasets from both true and synthetic data are shown in Figures 1 and 4 for the CelebA
dataset, and in Figure!5 for the LSUN dataset.
5.2.2
D IVERSITY E XPERIMENTS
Sections 2.2 and 3.4 describe techniques to study classification performance of labeled GANgenerated datasets against that of true data. In order to assess the quality of our GAN-generated
synthetic datasets, we evaluate label agreement between the default labels of the synthetic datasets,
13
Under review as a conference paper at ICLR 2018
Figure 6: Histograms of annotator confidence (softmax probability) during label prediction on true
data (test set) and synthetic data for tasks on the CelebA and LSUN datasets (see Section 3.4).
and that obtained using the annotators described in Section 2.1 as shown in Table 2. In Figure 6,
we inspect the confidence scores, defined as the softmax probabilities for predicted class, of the annotator while making these predictions. As can be seen from these results, the annotator confidence
for the synthetic data is comparable to that on the test set of true data. Thus, it seems likely that the
GAN generated samples are of good quality and are truly representative examples of their respective
classes, as expected based on visual inspection.
14
| 9cs.NE
|
arXiv:1701.01223v1 [cs.SY] 5 Jan 2017
Consensus as a Nash Equilibrium of a Dynamic
Game
Muhammad Umar B. Niazi
Arif Bülent Özgüler
Aykut Yıldız
Department of Electrical and
Electronics Engineering,
Bilkent University,
Ankara, Turkey.
Email: [email protected]
Department of Electrical and
Electronics Engineering,
Bilkent University,
Ankara, Turkey.
Email: [email protected]
Department of Electrical and
Electronics Engineering,
Bilkent University,
Ankara, Turkey.
Email: [email protected]
Abstract—Consensus formation in a social network is modeled
by a dynamic game of a prescribed duration played by members
of the network. Each member independently minimizes a cost
function that represents his/her motive. An integral cost function
penalizes a member’s differences of opinion from the others as
well as from his/her own initial opinion, weighted by influence
and stubbornness parameters. Each member uses its rate of
change of opinion as a control input. This defines a dynamic
non-cooperative game that turns out to have a unique Nash equilibrium. Analytic explicit expressions are derived for the opinion
trajectory of each member for two representative cases obtained
by suitable assumptions on the graph topology of the network.
These trajectories are then examined under different assumptions
on the relative sizes of the influence and stubbornness parameters
that appear in the cost functions.
Index Terms—Opinion dynamics, consensus, social network,
dynamic games, Nash equilibrium, game theory.
I. I NTRODUCTION
How gossip spreads in a small community, how a political
leader reaches or fails to reach voters, and how some students
learn faster than others among those with comparable intellectual capacity are three questions that fall into the study of
social opinion dynamics. It is no surprise that the research
question has attracted the attention of many disciplines in a
short span of time and a sizable penetrating literature has
been accumulated. We refer to the survey papers [1], [2]
and [13] for only a partial panorama. These publications can
roughly be divided into those that take a Bayesian perspective
such as [4] and those that put forward non-Bayesian models
such as [6]. Yet another classification is that while most
of the research focuses on formation of a consensus [14],
there are also those that study disagreement as in the case of
Hegselmann and Krause model [10], [7] or as in [5]. The study
of consensus has several engineering applications including
multi-agent coordination [17], information fusion in sensor
networks [19], consensus in small-world networks [12] and
distributed optimization algorithms [18].
We study consensus formation via Nash equilibrium in a
dynamic game of a prescribed duration played by members
This work is supported by the Science and Research Council of Turkey
(TÜBİTAK) under the project EEEAG-114E270.
in a social network. Each member (player or agent) independently minimizes a cost function that represents “its” (can be
read as “his/her”) motive. An integral cost function penalizes
its differences of opinion from its neighbors as well as from its
own initial opinion, weighted by influence and stubbornness
parameters. Each member uses its rate of change of opinion as
a control input. This defines a dynamic non-cooperative game
that turns out to have a unique Nash equilibrium. For two
representative cases obtained by suitable assumptions on the
information structure (graph topology), we are able to obtain
explicit analytic expressions for the opinion trajectories of
all members in the Nash solution. These trajectories are then
examined under different assumptions on the relative sizes of
influence and stubbornness parameters.
Nash equilibrium is only one among a wide range of
equilibrium concepts in games. One interpretation in [15]
suggests that if the same game is played several times with no
strategic dependencies between consecutive plays, then a Nash
equilibrium is most likely reached. This is for static games but
one can extend the interpretation to dynamic games as well.
The point of the matter is that it is a very useful construct (and
presently the only rigorous one) if the research objective is to
examine under what conditions, from independent motives of
agents, a pattern of collective behavior emerges.
In [9], a static game of opinion dynamics is posed and the
best response function in a Nash solution is used to postulate
an update scheme. The convergence of this dynamic scheme to
a consensus is examined. One can view our game model here
as a dynamic version of [9]. The optimal control of consensus
model and control through a leader model in [2] also use
integral cost functions and has similarities to our model except
that the objective in their case is control of consensus via
external actions. The non-cooperative dynamic game model
here is inspired by the foraging biological swarm models in
[16], [20], and [21].
In the next section we pose the opinion dynamics game in
its most generality. In Section 3, we study two specialized
versions and obtain explicit Nash solutions for these two
games that represent extreme cases of information structure.
Section 4 contains a number of simulation results for the
games of Section 2 and 3. The last section is on conclusions.
II. A G AME OF O PINION DYNAMICS
We represent a social network of n agents by a weighted
directed graph G = (N, E, wij ), where N = {1, ..., n} is
the set of all nodes (agents), E ⊆ N × N is the set of all
ordered pairs of connected nodes, and wij is the influence of
agent j on agent i when (i, j) ∈ E. One-sided or two-sided
connection between the nodes indicate one-sided or two-sided
interaction between the agents. The neighborhood of agent
i is defined to be the set of all agents with whom agent i
interacts, i.e., ηi := {j ∈ N : (i, j) ∈ E}. The reason for
a directed graph representation is because we can interpret
the weight on the edges to be the influence of an agent on
its neighbor or the value its neighbor gives to the opinion of
an agent. Thus, two neighbors can have different levels of
influence on each other. Let xi (t) be the opinion at time t
of agent i and let it be normalized so that for every t in the
interval [0, T ], xi (t) ∈ [0, 1]. Each agent has an initial opinion
xi (0) = xi0 ∈ [0, 1] about a certain issue, where the values
0 and 1 indicate the extreme cases. For example, 0 may be
interpreted as strong disagreement and 1 as strong agreement
cases. Let x(t) = [x1 (t) ... xn (t)]′ ∈ [0, 1]n denote the opinion
profile at time t in the network of n agents, where ‘prime’
denotes transpose. The cost functional of agent i is postulated
to be
Z T X
2
1
i
i
i
wij xi (t) − xj (t)
L (x, x0 , u ) =
2
0
j∈ηi
2 1 i 2
1 i
i
dt, (1)
u (t)
+ ki x (t) − x (0) +
2
2
where wij ∈ [0, ∞) is the parameter that weighs the susceptibility of agent j to influence agent i, ki ∈ [0, ∞) weighs
the stubbornness of agent i or the reluctance of i to divert
from its initial opinion. The control of agent i is assumed to
be ui (t) = ẋi (t), so that agent i controls the rate of change
of its opinion. The coefficient of the control term in the cost
is normalized to 1, without loss of generality. The integral in
the time interval [0, T ] indicates that the agent penalizes the
cumulative effect in each of the three terms in the integrand.
Considering the first term, for instance, what it penalizes as
part of the cost is the sum total of the divergence from the
opinions of the neighbors, not the instantaneous differences
from their opinions. This cost functional, which should be
viewed to be a model of the motive of agent i towards a
prevailing social issue, is prompted by [9], in which a static
model for the motives of agents in a social network was used
and by [16], in which a similar cost functional modeled the
motives of members in a foraging biological swarm. If each
agent in the social network minimizes its cost, then we have
a non-cooperative dynamic (or, differential) game played by
n agents
min
{Li } subject to ẋi (t) = ui (t)
i
u
∀i ∈ N.
(2)
A solution to such a game, if it exists, is a Nash solution, or
a Nash equilibrium of the game. Note that although x(0) is
specified as x0 ∈ [0, 1]n , its final value x(T ) is left free. Thus,
the optimization each agent carries out is one of free terminal
condition, [11]. The game (2) lies within the framework of
Theorem 6.11 in [3] and is in fact a quadratic game as we
show in the Appendix so that a unique Nash equilibrium exists
by Theorem 6.12 of [3]. Instead of using this result (after
transforming the problem to the set up of [3]), it is easier to
use the necessary conditions provided by Theorem 6.11 of [3].
We thus state those necessary conditions in the set up of our
game (2) first.
Let So be a trajectory or opinion space {x(t), 0 ≤ t ≤ T }
and Γi be a strategy space of agent i so that its every mapping
γ i : [0, T ]×So → Γi is a permissible strategy for agent i. And
define g i (x, xi0 , ui ) to be the integrand of the cost functional
(1),
Lemma 1. For an n-agent dynamic game of prescribed fixed
duration [0, T ], let
(i) ui (t) be continuously differentiable on R, ∀t ∈ [0, T ],
(ii) g i (x, xi0 , ui ) be continuously differentiable on R, ∀t ∈
[0, T ], i ∈ N .
If {γ i∗ (t, xi0 ) = ui∗ (t); i ∈ N } provides a unique openloop Nash equilibrium solution, and {x∗ (t), 0 ≤ t ≤ T } is the
corresponding opinion trajectory, then there exist n costate
functions pi (t) : [0, T ] → R, i ∈ N , such that the following
relations are satisfied:
ẋi∗ (t) = ui∗ (t),
ṗi (t) = − ∂Hi ,
∂xi
i∗
i
γ
(t,
x
)
≡
ui∗ (t) = arg minui ∈Γi Hi (pi , x, xi0 , ui ),
0
xi∗ (0) = xi0 ∈ [0, 1], pi (T ) = 0, i ∈ N,
(3)
where
Hi (pi , x, xi0 , ui ) = g i (x, xi0 , ui ) + pi (t)ui (t), t ∈ [0, T ]. (4)
Here we note that the terminal condition of the costate
functions is a consequence of the fact that the game has free
terminal conditions. Defining a Hamiltonian as in (4) and using
the relations in (3), we can combine the state and costate
equations into the following equation,
ẋ(t)
x(t)
x(0)
=A
+ K̂
,
(5)
ṗ(t)
p(t)
p(0)
where
A=
0
−W
−I
0
, K̂ =
0
K
0
0
where I is the identity matrix of size n
[p1 (t) ... pn (t)]′ , K = diag [k1 , ..., kn ]. Here,
q1
−w12 . . . −w1n
−w21
q2
. . . −w2n
W =
..
..
..
..
.
.
.
.
P
−wn1
−wn2
...
qn
,
and p(t) =
,
where qi = j∈ηi wij + ki . Notice that the matrix W is a
Laplacian-like matrix of a weighted directed graph G. Every
ij-th element in the off-diagonal, i 6= j, shows the weight of
the edge that is directed from i to j, and the diagonal elements
consist of the sum of all the weights associated with every
node and its stubbornness parameter. Solving the differential
equation (5) gives,
Z t
x(t)
x(0)
At
A(t−τ )
= e +
e
dτ.K̂
,
(6)
p(t)
p(0)
0
where Φ(t) = eAt = L−1 (sI − A)−1 and Ψ(t) =
R t A(t−τ )
e
dτ . Since
0
s(s2 I − W )−1
−(s2 I − W )−1
(sI−A)−1 =
, (7)
−W (s2 I − W )−1 s(s2 I − W )−1
one can correspondingly get the natural partitions
φ11 (t) φ12 (t)
ψ11 (t) ψ12 (t)
Φ(t) =
, Ψ(t) =
.
φ21 (t) φ22 (t)
ψ21 (t) ψ22 (t)
The diagonalizability assumption, although not necessary, is a
simplifying assumption .
Proposition 1. Suppose W is diagonalizable so that W =
V ΛV −1 , where Λ = diag [λ1 , λ2 , ..., λn ] and V is the matrix
whose columns are the corresponding linearly independent
eigenvectors. Then, a Nash equilibrium of the game (2) exists
and is unique. The opinion trajectory of the Nash solution is
given by
−1
x(t) = ζ11 (t) − ζ12 (t)ζ22
(T )ζ21 (T ) x(0),
(8)
where
ζ11 (t) = φ11 (t) + ψ12 (t)K,
ζ12 (t) = φ12 (t),
ζ21 (t) = φ21 (t) + ψ22 (t)K,
ζ22 (t) = φ22 (t),
and
φ11 (t)
φ12 (t)
φ21 (t)
φ22 (t)
ψ12 (t)
ψ22 (t)
with
= V diag [π1 , π2 , . . . , πn ] V −1 ,
= −V diag [π̂1 , π̂2 , . . . , π̂n ] V −1 ,
= W φ12 (t),
= φ11 (t),
= −V diag [π̃1 , π̃2 , . . . , π̃n ] V −1 ,
= −φ12 (t),
√
p
sinh λi t
√
λi t , π̂i =
,
πi = cosh
λi
√
cosh λi t − 1
π̃i =
, i ∈ N.
λi
III. G AMES WITH AN E XPLICIT NASH S OLUTION
The equation (8) in Proposition 1 will yield explicit expressions for opinion trajectories only if one can compute
the eigenvalues and the eigenvectors of W explicitly. In this
section, we present two typical situations in which analytic
expressions of the opinion trajectories are derived.
We will say that a full consensus is reached in the network
at the terminal time whenever the Nash solution of the game
(2) is such that x1 (T ) = ... = xn (T ). Of course, the equality
may hold only for a subset of N , which will then indicate a
partial consensus.
A. Consensus in a complete information structure
In a network where all agents are connected to each other,
i.e., ηi = N \ {i}, the opinion of agent i will be influenced
by all other agents and one may expect that a consensus
will eventually be reached. But, due to the presence of some
stubborn agents, a full consensus may still not be reached. The
present special game investigates this issue.
For simplicity and in order to get explicit solutions, we
assume equal parameters for all agents, i.e., ki = k, wij =
wji = w, ∀i ∈ N and (i, j) ∈ E.
Theorem 1. For a network of complete information structure, and where all the agents have equal parameters, the
unique Nash equilibrium is such that the opinion dynamics of
agent i is given by
n
n
1X j
1X j
x0 + γ(t)(xi0 −
x ),
(9)
n j=1
n j=1 0
√
cosh( λ1 (T −t))
√
where γ(t) = λk1 + nw
and λ1 = k + nw.
λ1
cosh( λ1 T )
The opinion dynamics x(t) with the i-th entry (9) has the
following properties:
(i) A full consensus is never achieved but the opinion
dynamics will progressively converge to
xi (t) =
n
lim lim xi (t) =
T →∞ t→T
n
1X j
1X j
k
x0 + (xi0 −
x ).
n j=1
λ1
n j=1 0
(10)
(ii) The Nash equilibrium will be a full
P consensus and the
opinions will converge to the average n1 nj=1 xj0 of the initial
opinions if and only if there are no stubborn agents, i.e., k = 0.
(iii) The opinion distance between any two agents at time
t ∈ [0, T ] is given by
|∆xij (t)| = γ(t)|∆xij
0 |,
(11)
j
i
where ∆xij (t) = xi (t) − xj (t) and ∆xij
0 = x0 − x0 .
Remark 1. The opinion trajectory of every agent has two
parts. The first term on the right hand side of (9) is the
average of initial opinions of all agents in the network, and
the second term depends on the difference between the initial
opinion of agent i and that average. The weight of the latter
is a coefficient that gets progressively closer to k/λ1 as time
passes.
Remark 2. Since we are able to derive explicit expressions
for the opinion trajectories, it is a simple matter to compute
the time it takes the network to reach a consensus within ǫvicinity of the average opinion. Or, to determine the individual
influence of each parameter on the ǫ-closeness to a full
consensus.
Remark 3. A fast convergence to average opinion obviously
requires a large λ1 , since the opinion distance as the terminal
time T → ∞ is
lim
T →∞
|xi (t) − xj (t)|
|xi0
−
xjo |
=
k
nw −√λ1 t
+
e
.
λ1
λ1
The degree of closeness to full consensus at the steady state
is decreased if k → 0 or if w ≫ k. A higher convergence rate
requires a large λ1 , which will be the case if any one of k, n,
w is large. Note that in case of a larger network population,
a quick consensus gets more likely because each agent experiences more social pressure in a complete information graph
topology.
IV. S OME S IMULATIONS OF THE G ENERAL O PINION
DYNAMICS G AME
(ii) For i ∈ N \ {1}, opinion of agent i will converge to the
leader’s opinion as T → ∞ if and only if ki = 0.
(iii) The opinion distance of any agent to the leader is given
by
ki
|∆xi1 (t)| =
+ ξi (t) |∆xi1
(14)
0 |.
λi
i
1
where ∆xi1 (t) = xi (t) − x1 (t) and ∆xi1
0 = x0 − x0 .
Remark 4. Note that the consensus in the long run is a
convex combination of the initial opinions of agent i and the
leader. In this convex rivalry, a stubborn agent will stand alone.
Remark 5. It is possible to determine the time in which
agent i is ǫ-close to the opinion maintained by the leader.
Similarly, it is straightforward to examine the sensitivity of an
ǫ-consensus to each parameter value wi1 , ki .
Remark 6. A fast convergence to the opinion of the leader
requires a large λi . This will be the case if the value of wi1
3
5
2
6
1
10
7
9
(13)
4
The leader (agent 1) in this network can be considered as
some political analyst who can influence the opinions of other
agents through electronic media. Therefore, the leader can
influence the opinions of other agents, but not the other way
round, based on the value of their influence and stubbornness
parameters. Due to that influence, they tend to adjust their
opinions closer to leader’s opinion. The network is represented
by a directed graph where the edges are directed from agents
towards the leader. Thus η1 = ∅, ηi = {1}, ∀i ∈ N \ {1}. It
follows that in this special game (2), wij 6= 0 only if j = 1.
The question we investigate is whether the leader’s opinion
will prevail under all parameter values given enough time. One
of course expects that a full consensus may not be achieved in
a finite duration whenever stubborn agents exist but if some
agent i is not stubborn, i.e., ki = 0, then that agent will make
consensus with the leader.
Theorem 2. For a network in which all agents are unilaterally connected to the leader (agent 1), the unique Nash
equilibrium is such that the opinion dynamics of agents are
given by
1
1
x (t) =x0 ,
i
1
(12)
xi (t) = ki x0 + wi1 x0 + ξi (t) xi0 − x10 ,
λi
cosh √λ (T −t)
( √i
)
and λi = ki + wi1 , ∀i ∈
where ξi (t) = wλi1i
cosh( λi T )
N \ {1}. The opinion dynamics x(t) with the i-th entry (9)
has the following properties:
(i) The leader never changes its initial opinion and the
opinions of other agents i ∈ N \ {1}, converge to
8
B. Consensus under a leader
ki xi0 + wi1 x10
.
lim lim xi (t) =
T →∞ t→T
λi
or ki is large. Although the convergence time is increased, the
property (i) shows that final value of the opinion will incline
towards either the leader’s or agent’s initial opinion depending
on whether wi1 > ki or wi1 < ki , respectively.
(a)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
3
3.5
4
4.5
5
(b)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
(c)
Fig. 1: Complete Information Structure
We have simulated a number of network structures, focusing on those with diagonalizable W matrix, and investigated the effect of some parameters on opinion dynamics.
Due to space limitations, we present three simulations that
illustrate in Figures 1, 2, and 3 the results of Theorems 1,
2 and Proposition 1. Here, we confine our investigation of
1
2
2
3
4
5
6
8
7
9
4
3
1
10
5
6
7
8
9
10
(a)
(a)
1
1
Leader
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
Leader-1
Leader-10
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0
0.5
1
1.5
2
(b)
2.5
3
3.5
4
4.5
5
(b)
1
1
Leader
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
Leader-1
Leader-10
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
(c)
(c)
Fig. 2: A network with one leader (agent 1)
Fig. 3: A network with two leaders (agent 1 and 10)
parameter effects to only see what happens if the control
term in the cost function is dominant or not dominant. The
simulation results in Figures 1 and 2 coincide with the plots
obtained by the analytic expressions of the opinion trajectories from Theorems 1 and 2. In all simulations, number
of agents is n = 10 and terminal time is T = 5 units.
Initial opinion levels of the agents are chosen as x(0) =
[0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95]′.
Fig. 1(a) illustrates a complete information structure of
Theorem 1 and Fig. 2(a), the one leader network of Theorem
2, respectively. In Fig. 1(b) and Fig. 2(b), w = 2 and k = 0.2
for all agents. In Fig. 1(c) and Fig. 2(c), we set w = 0.4 and
k = 0.04. The reduction of the weights of both the influence
and the stubbornness parameters have the effect of bringing
forth the penalization of the control term in the cost function.
This results in slowing down the convergence rate although
the opinions converge to the same values in both cases.
To illustrate the result of Proposition 1, we present a 2leader network (Fig. 3(a)) in which agent 1 and agent 10
are two leaders. The opinion trajectories are obtained via
the expression (8) after computing Φ(t) and Ψ(t) through
MATLAB. It is assumed that half of the followers support each
leader. The followers of leader-1 can be named as followers1, and of leader-10 as followers-10. The followers also have
influence among themselves in a society, of course, but an
agent can be assumed to have more impact on his fellow
supporters. We set ki = 0.2 ∀i ∈ N . The matrix W in (5)
is such that w1i = wni = 0, ∀i ∈ N and n = 10. In Fig.
3(b), we assume that the influences of leader-1 and leader10 on their followers is 10; the social impact of followers-1
and followers-10 among themselves is 2; the cross impact of
followers-1 on followers-10, and vice versa, is 0.2; and the
influence that followers take from other leader is 100 times less
than their own leader’s influence. In this case, agents follow
their respective leaders. However, if we assume that followers1 are more loyal to their leader and they run a good campaign
in order to attract the followers-10, then they are able to steal
followers-10 from their leader. For Fig. 3(c), we increase the
influence of leader-1 to 20 and the cross impact of followers-1
on followers-10 to 10, while all other parameters are the same.
It can be seen that rather than following leader-10, followers10 tend to follow followers-1 due to social impact.
V. C ONCLUSION
The main conclusion of this study is that a consensus, Nash
equilibrium, can spontaneously be reached from independent
motives of agents in a social network in the long run. How
unanimous this consensus is depends on the initial differences
of opinion, on the susceptibility of agents to influence, and
on the stubbornness of agents. If one member is singled out
as a leader who firmly sticks to its opinion, then a consensus
about the leader’s opinion is again formed in the long run. If
an opinion game is played in a finite interval a full consensus
is never reached in the presence of stubborn members.
The game with leader can also be viewed as a game of
learning in which the leader is the teacher and the others,
students. Here, it may be more instructive to examine the
situation when learning is poorly achieved. It is clear that
initial ignorance of the subject and reluctance to learn, all
contribute to this. But, a low willingness to update is also a
negative factor.
The game studied can be extended in several directions. The
opinion on a single issue is not essential and one can consider
each agent having opinions on several issues as e.g., in [10].
The technicality such an extension requires is that xi (t) is
no longer a scalar but a vector with each entry representing
the level of opinion on one issue. One can also extend the
game considered here to a network in which agents can have
different “types” of motives. Then, the integrands of the cost
functions the agents use will not have a uniform structure,
e.g., non-quadratic cost functions may be employed along with
quadratic ones.
A PPENDIX
A1. Proof of Proposition 1
We first note that the cost functional Li of (1) can be
transformed to a quadratic functional by
Z
1 T i ′ i i
z (t) G z (t) + [ui (t)]2 dt,
Li (zi , ui ) =
2 0
where
zi = ∆xi1 , ..., ∆xi
i−1
, ∆xi
i+1
′
, ..., ∆xin , xi − xi0 ,
and
Gi = diag [wi1 , ..., wi
i−1 , wi i+1 , ..., win , ki ]
≥ 0.
This fact allows one to employ Theorem 6.12 of [3] and the
opinion trajectories in a unique Nash solution can be derived.
However, because the transformation above is not a simple
one, it is much easier to directly obtain the Nash solutions
through the necessary conditions of Lemma 1. This is the
approach used in this Appendix. The uniqueness of a Nash
solution when one exists is, however, a direct consequence of
the above transformation and will not be separately addressed.
Let us write (6) as
x(t)
ζ11 (t) ζ12 (t)
x(0)
=
,
(15)
p(t)
ζ21 (t) ζ22 (t)
p(0)
and note that the expressions for Φ(t), Ψ(t), and their partitions are obtained by the inverse Laplace transform of (7)
via a matrix partial fraction expansion (see e.g. Lemma A.3
in [20] for a similar procedure). By Lemma 1, p(T ) = 0
so that from equation (15) evaluated at t = T we get
p(T ) = ζ21 (T )x(0) + ζ22 (T )p(0) = 0. Since, by its expression in Proposition 1, φ22 (T ) = ζ22 (T ) is nonsingular, we
−1
obtain p(0) = −ζ22
(T )ζ21 (T )x(0). Substituting into (15),
the solution (8) is obtained.
A2. Proof of Theorem 1
We have K = kI and W = qI − w(I − I); where q =
w(n − 1) + k, I is the identity matrix and I is the matrix of
all ones. For a matrix W , we can easily find the eigenvalues
and the eigenvectors, [16]. They are λ1 = q+w = k+nw with
multiplicity n − 1 and λ2 = q + w − nw = k with multiplicity
1.
Computing the corresponding eigenvectors, we obtain W =
V ΛV −1 , where Λ = diag [λ1 , ..., λ1 , λ2 ], V = V̂1 V̂2 ,
Ṽ1
, where V̂2 and Ṽ2 are given by
and V −1 =
Ṽ2
1
1
1 . . . 1 1×n ,
, Ṽ2 =
V̂2 = ...
n
1 n×1
because they are, respectively, right and left eigenvectors of O
associated with λ2 . Since V̂1 Ṽ1 + V̂2 Ṽ2 = I, we have V̂2 Ṽ2 =
1
1
n I and V̂1 Ṽ1 = I − n I. Also note that, in the notation of
Proposition 1, λi = λ1 for i = 1, ..., n − 1 and λn = λ2 . This
significantly simplifies the expressions for φij (t) and ψij (t)
of Proposition 1. For instance,
αI 0
1
Ṽ1
= αI + (β −α)I,
φ11 (t) = V̂1 V̂2
0 β
Ṽ2
n
√
√
where α = cosh( λ1 t), β = cosh( λ2 t). Thus, φ11 (t) is
a matrix with diagonal entries all equal to n1 (β + (n − 1)α)
and off-diagonal entries all equal to n1 (β − α). Simplifying
all partition matrices of Proposition 1 with this procedure and
substituting in (8), one arrives at
1
[1 + (n− 1)γ(t)]I + (1 − γ(t))(I − I) x(0), (16)
x(t) =
n
√
cosh( λ1 (T −t)
√
where γ(t) = λk1 + nw
. The i-th row of the
λ1
cosh( λ1 T )
right hand side of (16) simplifies to the right hand side of (9).
A3. Proof of Theorem 2
Given the information structure of this game, we get a lower
triangular matrix
q1
−w21 q2
−w31 0 q3
W =
,
.. . .
..
..
.
.
.
.
−wn1
0
...
0
qn
where q1 = k1 , qi = ki + wi1 ∀i ∈ N \ {1}. It turns
out that W is diagonalizable with W = V ΛV −1 . Here,
Λ = diag[q1 , ..., qn ] and the matrix V and its inverse are lower
triangular in the form
1
v21 1
v31 0 1
V (vi1 ) :=
.
..
.. . .
.
.
.
.
.
.
vn1 0 . . . 0 1
i1
where V = V (νi1 ), V −1 = V (−νi1 ) with νi1 = qiw−q
∀i ∈
1
N \ {1}. In the notation of Proposition1, λi = qi , ∀i ∈ N .
∗
0
Also exploiting the common structure
of the
∗ Q or I
matrices W, V, V −1 , where Q = diag[q2 , ..., qn ], the matrices
φij and ψij of Proposition 1 can all be simplified and (8) can
be found as
1
ρ2 (t) σ2 (t)
ρ3 (t)
0
σ
(t)
3
x(t) =
x(0), (17)
..
..
.
.
.
.
.
.
.
.
ρn (t)
0
...
0 σn (t)
where
ρi (t) =
wi1
− ξi (t),
qi
σi (t) =
ki
+ ξi (t).
qi
and
The right hand side of (12) is obtained by simplifying the ith
row of (17).
R EFERENCES
[1] D. Acemoglu and A. Ozdaglar, “Opinion Dynamics and Learning in
Social Networks,” Dyn. Games Appl., vol. 1, no. 1, pp. 3-49, 2010.
[2] G. Albi, L. Pareschi, G. Toscani, M. Zanella, “Recent advances in opinion
modeling: control and social influence,” arXiv:1607.05853 [physics.socph], 2016.
[3] T. Basar and G. J. Olsder, “Dynamic Noncooperative Game Theory,” in
SIAM Series in Classics in Applied Mathematics, 1999.
[4] S. Bikchandani, D. Hirshleifer, and I. Welch, “A theory of fads, fashion,
custom, and cultural change as information cascades,” Journal of Political
Economy, vol. 100, pp. 992-1026, 1992.
[5] D. Bindel, J. Kleinberg, and S. Oren, “How bad is forming your own
opinion?,” Games Econ. Behav., vol. 92, pp. 248-265, 2015.
[6] M. H. DeGroot, “Reaching a Consensus,” J. Am. Stat. Assoc., vol. 69,
no. 345, pp. 118-121, 1974.
[7] S. R. Etesami and T. Basar, “Game-Theoretic Analysis of the
Hegselmann-Krause Model for Opinion Dynamics in Finite Dimensions,”
IEEE Trans. Automat. Contr., vol. 60, no. 7, pp. 1886-1897, 2015.
[8] D. Ferraioli, P. W. Goldberg, and C. Ventre, “Decentralized Dynamics for
Finite Opinion Games,” CoRR, arXiv:1311.1610, 2013.
[9] J. Ghaderi and R. Srikant, “Opinion dynamics in social networks: A local
interaction game with stubborn agents,” Am. Control Conf. (ACC), vol.
50, no. 12, pp. 1982-1987, 2013.
[10] R. Hegselmann and U. Krause, “Opinion dynamics and bounded confidence models, analysis, and simulations,” Journal of Artifical Societies
and Social Simulation (JASSS), vol. 5, no. 3, 2002.
[11] D. E. Kirk, “Optimal Control Theory: An Introduction,” Courier Dover
Publications, Mineola, NY, USA, 2012.
[12] R. Olfati-Saber, “Ultrafast consensus in small-world networks,” Proc.
2005, Am. Control Conf., pp. 2371-2378, 2005.
[13] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1, pp.
215-233, 2007.
[14] A. Olshevsky and J. N. Tsitsiklis, “Convergence speed in distributed
consensus and averaging,” SIAM Journal on Control and Optimization,
vol. 48, no. 1, pp. 33-55, 2009.
[15] M. J. Osborne, and A. Rubinstein, “A Course in Game Theory,” The
MIT Press, Cambridge, Massachusetts, 1994.
[16] A. B. Özgüler and A. Yıldız, “Foraging swarms as Nash equilibria of
dynamic games,” IEEE Trans. Cybern., vol. 44, no. 6, pp. 979-987, 2014.
[17] W. Ren, R. W. Beard, and E. M. Atkins, “A survey of consensus
problems in multi-agent coordination,” Proc. 2005, Am. Control Conf.,
pp. 1859-1864 vol. 3, 2005.
[18] K. I. Tsianos, S. Lawlor, and M. G. Rabbat, “Consensus-based distributed optimization: Practical issues and applications in large-scale
machine learning,” 50th Annual Allerton Conference on Communication,
Control, and Computing, pp. 1543-1550, 2012.
[19] L. Xiao, S. Boyd, and S. J. Kim, “Distributed average consensus with
least-mean-square deviation,” J. Parallel Distrib. Comput., vol. 67, no. 1,
pp. 33-46, 2007.
[20] A. Yıldız and A. B. Özgüler, “Partially informed agents can form a
swarm in a Nash equilibrium,” IEEE Trans. Automat. Conr., vol. 60, no.
11, pp. 3089-3094, Nov. 2015.
[21] A. Yıldız and A. B. Özgüler, “Foraging motion of swarms wih leaders
as Nash equilibria,” Automatica, to appear.
| 3cs.SY
|
Definitions of distance function in radial basis function approach
W. Chen
Department of Mechanical System Engineering, Shinshu University, Wakasato 4-17-1, Nagano City,
Nagano 380-8533, Japan
(E-mail: * [email protected] and [email protected])
Key words: radial basis function, distance function, time-space RBF, nonlinear problem.
1. Introduction
ξ = f1 (x, y) , η = f2 (x, y ) ,
Very few studies involve how to construct
the efficient RBFs by means of problem
features. Recently the present author presented
general solution RBF (GS-RBF) methodology
to create operator-dependent RBFs successfully
[1]. On the other hand, the normal radial basis
function (RBF) is defined via Euclidean space
distance function or the geodesic distance [2].
This purpose of this note is to redefine distance
function in conjunction with problem features,
which
include
problem-dependent
and
time-space distance function.
The corresponding distance function of the
Euclidean norm is given by
r=
r=
The general second order partial differential
system with varying coefficient can be stated as
∂2u
∂2u
∂2u
R 2 +S
+ T 2 = 0,
∂x
∂xy
∂y
2
2
j
j
.
[f (x, y) − f (x , y )]
+[f (x, y ) − f (x , y )]
(3)
2
1
1
j
j
2
2
2
j
(4)
j
Obviously, the above definition of the distance
function in terms of variable x and y are
different from the direct use of the Euclidean
norm.
In the following we illustrate two important
special cases. Let us consider [3]
ym
2.1 Varying parameter problems
(ξ − ξ ) + (η − η )
Substituting Eq. (2) into Eq. (3), we define the
distance function on the original independent
variables of x and y as
2. Problems with varying parameter,
time-dependent or nonlinear terms
The GS-RBF [1] is constructed based on the
canonical form of some operators such as the
known Laplace or Helmholtz operators. Many
important cases in engineering do not possess
such standard form of operators. This section
will show that some cares may be taken to
handle these problems.
(2)
∂ 2u ∂ 2u
+
= 0,
∂x 2 ∂y 2
( y ≥ 0, m f −2) ,
(5)
its general solution is
( )
u = r2−2 β w r 2 ,
(6)
(1)
where R, S and T are continuous functions of x
and y. We can translate it into the canonical
form by a suitable change of independent
variables [2]
where β=m/2(m+2), w is the hypergoemetric
functions,
r=
r1
,
r2
(7a)
r1 =
r2 =
(x − x ) + (m +4 2 )
y
(x − x ) + (m +4 2 )
y
2
j
2
2
j
2
m+ 2
2
−y
m+ 2
2
+y
Let
ξ = x,
we have
2
y
m +2
η=
m+ 2
2
j
2
,
m+ 2
2
j
m +2
2
(7b)
2.2 Nonlinear problems
2
,
∂2u ∂2u
+
=0.
∂ξ 2 ∂η 2
(7c)
(8)
(9)
Substituting Eq. (8) into Eq. (3) also yields the
distance function Eq. (7b).
For another example, let us consider
∂2u
∂2u
∂u
+
y
= 0 , (y ≥ 0) ,
2
2 +α
∂x
∂y
∂y
(10)
where α is coefficients. Its solution is
−2 β′
2
u=r
( ),
wr
2
(11)
where w is also the hypergoemetric function,
β′ =1 −β = 1 2 −α,
r1 =
(x − x ) + 4(y
r2 =
(x − x ) + 4(y
2
12
j
2
j
r=
12
)
12 2
j
−y
The GS-RBF approach is applicable to the
nonlinear problems. It is usually not easy to get
the fundamental solution of general nonlinear
problems. In some cases, however, we can
transfer the nonlinear operator into the linear
one via a variable transformation [4]. For
example, consider the stationary heat transfer
through the nonlinear materials where the
conductivity K is often a function of
temperature u. The governing equation is given
by
∂ ∂ u ∂ ∂u
∂u
−
K + K − vx
∂ x ∂ x ∂y ∂y
∂x
∂u
∂u
vy
− u = ρc ,
∂y
∂t
,
)
r1
,
r2
(12d)
By
using
the
independent
variable
transformation approach as in the previous case,
we can get the distance functions of the RBF as
shown in Eqs. (12a-d).
This hints us to define the distance
functions simultaneously for the RBF ψ (r1, r2 )
ψ (r1, r2 ) = ϕ (r2 )φ (r1 r2 ) ,
u − u0
K = K 0 1 + β
.
u0
(12b)
(12c)
(13)
The above examples show that for one certain
(15)
where u is the temperature, vx and vy are the
components of the velocity vector v, ρand c
are density and the specific heat of the materials,
respectively.
The
coefficient
K
is
time-dependent conductivity coefficient
(12a)
12 2
j
+y
problem, we can use multiple distance variables
and RBFs simultaneously rather than one
distance function and one RBF.
(16)
If we intend to solve this problem using the
domain-type RBF, one of the reasonable
choices of the RBF should be the TPS-type
RBF due to the presence of Laplace-like
operator. However, the brutal use of the TPS is
discouraged in this case. By using the Kirchhoff
transformation [4],
U=
∫
u
u0
K (u)du ,
(17)
we have
2
u − u0
Ku
K 0u0
U = 0 0 1+ β
.
−
2β
2β
u0
(18)
The nonlinear Laplace operator is transformed
1
z(r, t) = cos(ct ) + sin (ct ) J 0 (r)
c
to the linear Laplace one
∂ 2U ∂ 2U
∂U
+
− Vx
−
∂x 2 ∂ z 2
∂x
∂U
∂U
Vy
− f (U ) = g(U )
,
∂y
∂t
(19)
where Vx, Vy, f(U) and g(U) are the functions of
U. The TPS RBF r2mln(r) is recommended to
solve Eq. (19). However, our purpose is to
directly solve Eq. (15). According to Eq. (18),
we have
u = u0 +
u0
2βU
1+
β K 0 u0
12
−
u0
β
u
2βr 2 m ln(r )
ψ (r ) = u0 + 0 1 +
β
K 0 u0
uxx =
1
utt + f (x, t) .
c2
(20)
(25)
12
−
u0
β
, (21)
The free symmetrical vibration of a very
large membrane are governed by the equation
(22)
with z=f(r), ∂z ∂t = g(r ) when t=0. We have the
solution
∞
z(r, t) = ∫ ξf (ξ )cos(ξct )J 0 (ξr )dξ +
(26)
uxx + uss = f (x, t) .
(23)
where upper-dashed f and g are the zero-order
Hankel transforms of f(r) and g(r), respectively.
According to the GS-RBF, we have
(27)
By analogy with the Euclidean definition of
distance variable, the time-space RBFs (TSR) is
defined
rj =
=
2.3. Time-space RBFs
∂ 2 z 1 ∂z 1 ∂ 2 z
+
=
∂ r 2 r ∂ r c 2 ∂t 2
s = ict ,
where i means unit imaginary number, we have
which should be used in the RBF collocation of
the original governing equation (15).
From the above analysis, we conclude that
the use of RBF should fully use the features of
the problems.
0
The present author introduced concept of
time-space distance variable and corresponding
RBFs for time-dependent problems [1]. Here
we hope to further improve its definition.
Consider the equation governing wave
propagation
Let
.
The corresponding TPS RBF for u is
1 ∞
g (ξ )sin (ξct )J 0 (ξr)dξ
c ∫0
(24)
(x − x ) + (s − s )
2
2
j
j
(x − x ) − c (t − t )
2
2
2
j
.
(28)
j
However, such definition can lead to imaginary
number of distance variable. Thus, it is safe to
use
rj =
(x − x ) + c (t − t ) .
2
j
2
2
j
(29)
The above definition of r differs from the
standard radial basis function in that the time
variable is included into the distance function
and is handled equally as the other space
variables. We can construct the RBF by means
of the GS-RBF in such a way that
time-dependence is naturally eliminated. In
general, hyperbolic and elliptic equations have
solutions whose arguments have the form x+at
and x+ibt respectively, where a and b are real.
Namely,
u(x,t ) = f (p)
(30)
where p is some unknown function of x and t.
For 3D case,
p = lx + my + nz + µt
(31)
This provides some theoretical support to use
time-space distance functions (28) and (29).
However, the above situations do not exist for
diffusion equation
∇ 2u =
1
ut + f (x,t ).
k
(32)
Its fundamental solution is given by
u* =
(t
1
j
−t
)
d2
r2
H tj − t , (33)
exp −
4k t − t
j
(
(
)
)
where d is the space dimensionality, H() is the
Heaviside function. Therefore, we give the
corresponding RBF
φ (r ) = h(r)u * (r, t, tj ) ,
(34)
where h(r) is chosen according to problem
feature. It is stressed that the response and
source nodes must be totally differentially
placed to avoid singularity in this case. This
strategy of node placement is called as response
source points staggering.
On the other hand, Chen and Tanaka [5]
proposed time-space non-singular general
solution
u* = Ae
( )
− k t− tj
φ (r ) ,
(35)
for diffusion problem and
*
[ ( ( ))+ Dsin(c(t − t ))]φ (r )
u = C cos c t − t j
j
(36)
for wave problems, where φ (r) is the zero
order Bessel function of the first kind for 2D
problems and sin(r)/r for 3D problems.
Substituting these non-singular general solution
into Eq. (34) produces the time-space RBF
without singularity.
In addition, it is stressed that the time-space
distance function and corresponding RBFs are
applicable to transient data processing such as
motion picture and movie.
References
1. Chen, W. and Tanaka, M., New Insights in
Boundary-only and Domain-type RBF
Methods, Inter. J. Nonlinear Sci. Numer.
Simulation, 1(2000), 145-152.
2. Golberg, M.A. and Chen, C.S., A
bibliography on radial basis function
approximation,
Boundary
Element
Communications, 7(1996), 155-163.
3. Cmnphob, M.M., Examples of Partial
Differential Equations, (Japanese version
translated by M. Sayigo), Morikata Press,
Tokyo, 1982.
4. Bialecki, R. and Nowak, A. J., Boundary
value problems for non-linear materials and
non-linear boundary conditions, Applied
Mathematics Modelling, 5 (1981), 417-421.
5. Chen, W. and Tanaka, M., Boundary knot
method:
A
meshless,
exponential
convergence,
integration-free,
and
boundary-only RBF technique, (submitted),
2000.
| 5cs.CE
|
Finding HeavyPaths in Weighted Graphs and a Case-Study
on Community Detection
Mohammad Khabbaz
Personal Business Development
arXiv:1512.04010v1 [cs.DS] 13 Dec 2015
[email protected]
ABSTRACT
A heavy path in a weighted graph represents a notion of
connectivity and ordering that goes beyond two nodes. The
heaviest path of length ` in the graph, simply means a sequence of nodes with edges between them, such that the sum
of edge weights is maximum among all paths of length `. It is
trivial to state the heaviest edge in the graph is the heaviest
path of length 1, that represents a heavy connection between
(any) two existing nodes. This can be generalized in many
different ways for more than two nodes, one of which is finding the heavy weight paths in the graph. In an influence
network, this represents a highway for spreading information from a node to one of its indirect neighbors at distance
`. Moreover, a heavy path implies an ordering of nodes. For
instance, we can discover which ordering of songs (tourist
spots) on a playlist (travel itinerary) is more pleasant to a
user or a group of users who enjoy all songs (tourist spots)
on the playlist (itinerary). This can also serve as a hard
optimization problem, maximizing different types of quantities of a path such as score, flow, probability or surprise,
defined as edge weight. Therefore, if one can solve the Heavy
Path Problem (HPP) efficiently, they can as well use HPP
for modeling and reduce other complex problems to it.
More precisely, we aim at finding k heaviest (top-k) paths
of a given length `. The weight of a path is defined as
a monotone aggregation of individual edge weights using
functions such as sum or product. We argue finding simple
paths is way more practical than finding paths with cycles
in applications such as: routing, playlist recommendation,
itinerary planning and influence maximization. Avoiding cycles results in lists without repeated items and with better
diversity. Simple paths are also expected to have a higher
utility in practice. This makes HPP NP-hard and inapproximable. We propose an efficient algorithm that despite its
exponential (theoretical) worst case running time, achieves
the exact answer of the NP-hard problem in many useful
cases and study the problem from different perspectives.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
We compare our main algorithm, Repeated Sorted Access
(RSA), against baseline and state-of-the-art algorithms. We
show with experiments that following our approach is significantly more scalable than existing algorithms for solving
HPP. We conduct a comprehensive set of experiments on
four graphs that inherit characteristics of real life applications. This supports our arguments regarding scalability of
RSA in practice. We make all of our implementations as well
as graphs publicly available to enable reproducibility of our
experiments. Our focus is on solving the technical problem
in the first segment of the paper. Following our technical
presentation, we provide a novel case study on community
detection to illustrate the usefulness of solving HPP with a
novel application.
1.
INTRODUCTION
Graphs are convenient data structures for modeling real
life data such as social networks, citation networks, the web,
phone call and email histories, purchase/transaction records
and user actions [1]. It is easy to model data with graphs.
However, we need sophisticated tools to make sense of such
data. In order to do so, we need to study notions such as
connectivity [1], centrality [6] and density [28]. Graph algorithms are at the core of community detection, for discovering various types of groups based on previous interactions
of individuals [39]. Random walk based methods have been
used in many mining and graph applications. Mining interesting multi-relational patterns is solved as discovering
maximal pseudo cliques in k-partite graphs [41].
A problem that has received relatively less attention in the
research community is that of finding k cycle-free heaviest
paths of length ` in a weighted graph (HPP). Sum and product are examples of monotone aggregation functions that
define the aggregate weight of a path. Generating lists of
items of interest is a great practical example. Suppose edge
probability w(u,v) = P (v|u), the conditional transition probability that a random user would like to listen to u (as next
song) given that v is the current song. Also assume there is
equal probability for any user to like any song. Using product as aggregation function, the weight of a path is directly
proportional to the probability that a cycle free ”chain” of
songs appears one after another in a listening session. A
heavy path in such graph, represents a playlist with high
overall transition probability between songs. The graph can
be created according to user(s) listening history. We can
imagine other kinds of lists such as reading lists of books or
papers that show flows of ideas through publications.
In influence networks, recently there have been attempts
for sparcification [42]. A path with high probability is a
highway for spreading influence in an influence network with
transition probabilities on edges. A more principled approach takes into account only heavy paths rather than
technicalities and complex problem definitions. A complete
study of all of HPP applications goes beyond the scope of
this paper. We need the preliminary work done in this paper,
in order to aim at creating a toolbox that uses heavy paths
for network analysis. Here, we show a possible novel application for community detection in research communities in a
co-authorship network. We use maximal frequent sub-paths
in top-k heavy paths in order to discover the cores and further ”grow” each core with peripheral nodes using existing
edges in top paths. We enhance our observations with metadata, and show that the core communities found using our
approach unanimously report highly influential researchers.
We make the following contributions in the subsequent
sections:
• We formalize and motivate the HPP problem and show
it is NP-hard and inapproximable. We discuss four
exact algorithms, and provide a comparison of classical
vs. novel algorithms for solving HPP. We provide a
comprehensive experimental study of HPP algorithms
presented in Section 7.
• We present a case study to discover small but influential networks of people in research communities. We
do this analysis on a sub-graph of DBLP co-authorship
graph, and report the results in Section 6.
• We show how a well studied top-k query processing
problem in the database literature can be converted
to HPP on graphs where there is no database. We
solely focus on algorithmic aspects, and propose a more
scalable algorithm that overcomes the shortcommings
of the Rankjoin [25] algorithm.
2.
present using sum. We define weight of P = (v0 , ..., v` ), as
P`−1
P.weight =
j=0 w(j,j+1) . Furthermore, we know we can
convert product to sum using log values and this does not
change the order of the list of top-k heaviest paths.
Finding the heaviest path for a given parameter ` is equivalent to the `-TSP problem, defined as follows: Given a
weighted graph, find a path of minimum weight that passes
through any ` + 1 nodes. It is easy to see that for a given
length `, a path P is a solution to `-TSP iff it is a solution to HPP (with k = 1) on the same graph but with edge
weights modified as follows: let w(u,v) be the weight of an
edge (u, v) in the `-TSP instance; then the edge weight in
0
the HPP instance is w(u,v)
= C − w(u,v) , where C is any
constant. This reduction works in both ways given that we
use the same constant C. We are not concerned with negative edge weights since we deal with paths without cycles;
but, we focus on cases where edges are either all negative
or all positive for simplicity. It is well known that `-TSP
is NP-hard (as a version of TSP). Furthermore, `-TSP does
not have any bounded approximation in the general case i.e.
no triangle inequality on edge weights etc [2, 3].
PROPOSITION 1. HPP does not have any bounded
approximation in polynomial time.
Proof. We know `-TSP is not approximable. Remember the
C−w reduction between `-TSP(G1) and HPP (G2) introduced earlier.
Suppose C = 0 and k = 1, if P is the answer to HPP (the heaviest
path) on G2, it is the answer to `-TSP on G1 after −w transformation
on edge weights. The weight of P in G1 (−Wmax ) is the negative of
its weight in G2 (Wmax ). If we find an α (0 < α < 1) approximation
to P on G2 as a result of finding path Q in polynomial time we
have Q.W ≥ α × Wmax . On G1 graph equivalently we can say,
−Q.W ≤ α × −Wmax . Since −Wmax is the optimal answer on G1
graph (i.e. minimum weight among all paths), it is guaranteed that
−Q.W is in the range −Wmax ≤ −Q.W ≤ α × −Wmax and is a true
approximate answer to `-TSP. We know `-TSP is not approximable.
According to the above argument, approximability of HPP results in
approximability of `-TSP. This is in contradiction with the known
fact that `-TSP and TSP are not approximable in the general case.
TECHNICAL PROBLEM DEFINITION
The most similar well studied problem to HPP in the literature is finding top-k heavy paths on directed `-partite
graphs. A nice example is k`-Stable Clusters problem [7].
Absence of cycles makes the studied problem considerably
simpler than HPP. Furthermore, practical applications of
HPP are not limited to `-partite graphs as described in Section 1. We show HPP is NP-hard and propose a practical algorithmic framework for solving it in the general case.
It is common to use exponential time but efficient algorithms for mining purposes. Great examples are Apriori
and FP-Growth [43], that is used for frequent item-set mining. Heavy paths also represent lists of highly connected
items that can be used in similar ways to frequent item-sets
and also for ordering of the items in the set.
Given a weighted graph G(V, E, W ), where weight w(u,v)
represents the edge weight between u and v, we aim at finding the k heaviest simple paths of length `, with highest
overall path weights. We call this the Heavy Path Problem
(HPP).
A simple path of length ` is a sequence of nodes (v0 , . . . , v` )
such that (vi , vi+1 ) ∈ E, 0 ≤ i < `, and there are no cycles, i.e., the nodes vi are distinct. Unless otherwise specified, in the rest of the paper, we use the term path to
mean simple path. We focus on sum and product, as aggregation functions that define the weight of a path. For
simplicity of presentation and practical considerations, we
In Figure 1(a), the heaviest simple path of length 4, is
obtained by visiting the nodes in the order 6–1–2–3–4 and
has a weight of 3.35. In contrast, a different permutation of
the nodes 4–3–6–1–2 has a weight of 3.08. The higher weight
of a path can represent a sequence with more strength or
better quality. Different permutations result in paths with
different weights simply because paths consist of different
edges most likely with different edge weights.
3.
3.1
ALGORITHMS
Baselines
An obvious algorithm for finding the heaviest paths of
length ` is performing a depth-first search (DFS) from each
node, with the search limited to a depth of `, while maintaining the k heaviest paths and avoiding cycles. This is an
exhaustive algorithm and is not expected to scale. A somewhat better approach is dynamic programming. Held and
Karp [36] proposed a dynamic programming algorithm for
TSP which we adapt to HPP as follows. Since path lengths
required in our problem are typically much smaller than the
total number of nodes in the graph, we replace the notion of
”allowed nodes” with ”avoiding nodes”. The idea is to find
the heaviest J-avoiding path of length ` − 1 that ends in one
of the neighbors of J. We can repeat this recursively during
(a)
Figure 1:
(b)
(a) A complete graph with 6 nodes and sorted list of edges. (b) An example of random access is shown on the right.
the execution using smaller path lengths and larger avoidance sets as parameters. The base case is where the input
parameter assigned to length is 1, in which we only need to
consider the paths created by extending the current path by
one hop, making sure we do not extend by any nodes that
belong to the (` − 1) size avoidance set. The rest is similar
to DFS. We repeat this for every node(J) and keep track of
the heaviest path. Dynamic Programming algorithm aggregates and discards many short path segments early on by
choosing the heaviest one that ends at a certain node. Both
DFS and Dynamic Programming can be easily generalized
to report top-k paths which is not the focus of this paper.
nects dummy1 to its neighborhood with a directed edge. Every restaurant (l2 ), is modeled as a directed edge that connects its neighborhood to dummy3. Edge weight is defined
as item score. Heaviest paths of length 2 are top combinations. Rankjoin can be converted to heavy path on a tripartite
graph.
We design an adaptation of Rankjoin that works with selfjoins (i.e. all li are identical) for solving HPP on general
graphs. Direct usage of Rankjoin results in creating many
paths with cycles that need to be pruned later on. SAA
(Algorithm 1), avoids creating cycles early on. SAA algoAlgorithm 1 SAA (E, `, k)
3.2
Sorted Access Algorithm (SAA) for HPP
The methods discussed in the previous section lack the
ability to prune the search space using edge weights. It
may be more promising to first look at those parts of the
graph with heavier edge weights. If they connect, they can
create heavy paths. This way, we solve HPP only for a
smaller sub-graph or at least explore the search space more
efficiently. This can make a considerable difference to the
exponentially increasing running time of the problem using
smaller input size. There exists a body of work on top-k
algorithms in the database community started by Fagin et
al. [34]. These algorithms aim at finding top-k items without
searching the whole search space using sorted access. The
framework is such that partial item scores from multiple
sorted lists (l1 ...l` ), are aggregated using a monotone aggregation function. Following their idea more aggressive top-k
algorithms were proposed in order to extend their framework
with probabilistic guarantees such as [29]. We can argue the
weight of a path (item), is the aggregation of several partial item scores (edges). This makes following top-k query
processing ideas promising for solving HPP. In particular,
in Rank Join [26, 4] problem, partial scores in l1 can join
more than one partial scores in l2 . For instance, suppose
l1 provides a sorted list of movie theaters and l2 provides
a sorted list of restaurants sorted by their popularity score.
One may be interested in watching a movie and going to a
restaurant in the same neighborhood. The score of a combination is defined using the sum of the individual scores and
items of interest are (movie theater, restaurant) combinations. There may be many movie theaters and restaurants
in the same neighborhood.
Problem Conversion: Define a tripartite graph. The
first column of nodes is a single dummy node (dummy1) as
well as the third column(dummy3). Every node in the middle column represents a neighborhood (such as downtown).
Every movie theater (l1 ), is modeled as an edge that con-
Input: Sorted edge list E, path length `, number of paths k
Output: top-k heaviest paths of length `
1: topKBuf f er = // empty sorted set
2: θ = ` × Wmax // edge weights ∈ [0, 1]
3: ScannedEdges = // empty hash table
4: e = E.getN extEdge()
5: while topKBuf f er.BottomScore < θ do
6: P 1 = e
7: for l = 1 to ` − 1 do
8:
P i+1 = Join(P i , ScannedEdges) // avoid cycles
9: U pdate(topKBuf f er, P ` )
10:
11:
12:
13:
14:
θ = e.w + (` − 1) × Wmax
ScannedEdges.put(e.start, e)
ScannedEdges.put(e.end, e)
e = E.getN extEdge()
return topKBuf f er
rithm scans the sorted list of edges and reads one edge at
a time. Once a new edge (e) is read under sorted access,
it is joined with the list of heavier edges (ScannedEdges)
` − 1 times. The Join operation used in line 8, takes two
input parameters: 1) a list of paths of length i, P i ; 2) a list
of edges, ScannedEdges. It extends all paths of length i
in P i with matching edges in ScannedEdges (i.e. smaller
search space) and ensures no cycles are created. This results in paths of length i + 1. In case P i+1 is empty because there are no matching nodes to join without cycles,
naturally, no path will be created. For instance no path of
length l > 1 is created after joining the heaviest edge and
the empty set of scanned edges. This is how Join is defined
as an operation. Join is one of the operations defined in
Relational Algebra. Therefore, it is not scary to use Join,
it has been used in practice as part of the DBMS operations for a long time. We do an enhanced implementation
for self-joins that prunes cycles as well. Paths of length
` will be created after enough edges are scanned in sorted
order. This operation is repeated ` − 1 times, until all possible paths of length ` including e and heavier edges are
produced. At this point (line 10), SAA knows that no path
can be created during the rest of execution which is heavier
than θ = e.w + (` − 1) × Wmax . Therefore, if we have already
constructed k paths heavier than θ, we know that these are
the correct top-k paths of length `. In order to speed up the
Join operation, ScannedEdges is maintained as a hash table that maps nodes (keys) to edges (values). Consider the
example graph in Figure 1(a). Suppose we are interested
in the heaviest path of length 3. The SAA algorithm proceeds by scanning the edge list in sorted order of the edge
weights. After reading 4 edges at depth d = 4, the edge
weight (w4 ) is 0.77 and wmax = 0.93. We can calculate the
threshold (upperbound) as: θ(2.63) = 0.77 + (3 − 1) ∗ 0.93.
At depth 5, SAA is able to construct the path (6,1,2,3) and
θ(2.62) = 0.76 + 2 ∗ 0.93. Since θ is equal to the weight of
(6,1,2,3), we know it is at least one of the heaviest paths in
the graph and can terminate without processing the rest of
the edges in the sorted list. This is what we call search space
pruning by Sorted Access.
Figure 2:
4.
LIMITATION 2. SAA threshold is conservative compared to what can be easily obtained during the execution.
LIMITATIONS OF SAA AND OPTIMIZATIONS
LIMITATION 1. Let P be a path of length ` and suppose e is the lightest edge on P , i.e., its weight is the least
among all edges in P . Then until e is seen under sorted
access, the path P will not be constructed by SAA.
Limitation 1, simply follows the fact that paths are created
using only sorted access to the list of edges. If we run SAA
on the graph of Fig. 2, we won’t create the heaviest path
(a,b,c,d) until the edge (c,d) is scanned. However, (c,d) is
the lightest edge in the graph. This means following only
sorted access, may delay the construction of the heaviest
path until many irrelevant paths are created and discarded.
This observation motivates the following optimization for
finding heaviest paths.
OPTIMIZATION 1. We should try to avoid delaying
the production of a heavy path of a certain length until the
lightest edge on the path is seen under sorted access. One
possible way of making this happen is via random access.
However, random accesses have to be done with care in order
to keep the overhead low.
In the example of Fig. 2, suppose we have constructed
(a,b,c) and know it is the heaviest path of length 2. It may
be a good heuristic to extend (a,b,c) with new edges such as
(c,d) and create paths of length 3. Since (a,b,c) is the heaviest path of length 2, it is very likely to create heavy paths of
length 3 by extending (a,b,c). Making access to graph edges
such as (c,d), regardless of where they appear on the sorted
list of edges, is what we refer to as ”random access”. An
example of random access is presented by Figure 1(b). Making random access to the edge list simply means reading an
edge without caring about where the edge is located on the
sorted list of edges (through the adjacency matrix). This
can help produce heavy paths containing low weight edges
earlier. Our case study in Section 6, shows one practical
example where a low weight edge on a heavy path, can have
a meaningful interpretation. The discussion is provided in
Section 6, for Gui-Rong Xue’s edges with others. Following this optimization, suppose we use random accesses to
find “matching” edges with which to extend heaviest paths
Example instance of HPP. Graph has one heavy
path and n lighter paths.
of length ` − 1 to length `. This way we won’t delay the creation of heavy paths until their lightest edge is visited under
sorted access. However, this may not be enough for early termination of the algorithm. Another decisive factor in early
termination is the upperbound threshold (θ) value on weight
of paths we have not created yet using the ScannedEdges.
SAA stops when θ gets smaller than the weight of the
heaviest path of length ` discovered so far. Suppose there is
an instance in which the heaviest path of length ` is lighter
than wmin + (` − 1) × wmax . In this case, SAA will produce every path of length ` before reporting the heaviest
path, resulting in no pruning. In Fig. 1(a), this happens
when trying to find the heaviest path of length 4 and in
Fig. 2, when trying to find the heaviest path of length 3.
The following lemma highlights a natural possibility during
the execution of SAA for obtaining a tighter threshold that
results in earlier termination.
LEMMA 1. Let P be the heaviest path of length `. When
SAA terminates, every path of length ` − 1 that has weight
no less than P.weight − wmax , will be created. This includes
the heaviest path of length ` − 1.
Proof. Suppose SAA finds P at depth d. This means P.weight ≥
wd + (` − 1) × wmax , and, P.weight − wmax ≥ wd + (` − 2) × wmax .
Notice that P.weight − wmax is a lower bound on the weight of the
heaviest path of length ` − 1. Therefore, if there is a path of length
` − 1 that is heavier than P.weight − wmax , it will be produced by
SAA by depth d.
One way of making the threshold tighter is by keeping
track of shorter paths. For example, if we know P is a
heaviest path of length ` − 1, we can infer that the heaviest
path of length ` cannot be heavier than P.weight + wmax , a
bound often much tighter than wd + (` − 1)wmax .
For this, we need to keep track of the heaviest paths of
length ` − 1 one by one to lower the threshold more aggressively, gradually and smoothly. Pursuing this idea recursively leads to a framework where we maintain and release
heaviest paths of each length i, 2 ≤ i ≤ `, at the right time
and make the next threshold tighter. More precisely, we
can perform the following optimization. For this purpose,
we can use Buffers as sorted sets (priority queue) in order
to store (insert) and release heavy paths (remove top).
Figure 3 schematically describes the idea of using buffers
for different path lengths. Buffer B1 is the sorted edge list
for the graph in Figure 2. Buffers B2 and B3 store paths
of length 2 and 3 respectively. When random accesses are
performed to extend paths of length l − 1 to those of length
l, the buffers can be used to store these intermediate paths.
Algorithm 3 NextHeavyPath (E, l)
Input: Sorted list of edges E, and path length l
Output: Next heaviest path of length l
1: if l = 1 then
2: P 1 ← ReadEdge(E)
3: θ2 = 2 × P 1 .weight
4: return P 1
5: while Bl .topScore ≤ θl do
6: P l−1 ← NextHeavyPath (E, l − 1) // recursion
7: s, t ← EndNodes(P l−1 )
8: for all y ∈ V | (y, s) ∈ E do
9:
Bl ← Bl ∪ ((y, s) + P l−1 ) // avoiding cycles
10: for all z ∈ V | (t, z) ∈ E do
11:
Bl ← Bl ∪ (P l−1 + (t, z)) // avoiding cycles
12: P l ← RemoveTopPath(Bl )
13: if l < ` then
14: θl+1 = max(Bl .topScore, θl ) + wmax
15: return P l
Figure 3:
RSA example for finding the heaviest path of
length 3 on graph of Figure 2
Algorithm 2 RSA (E, `, k)
Input: Sorted edge list E, path length `, number of paths k
Output: top-k heaviest paths of length `
1: for l = 2 to ` do
2: Bl ← ∅ // empty sorted set
3: θl = wmax × l
4: topP aths ← ∅ // empty sorted set
5: while | topP aths |< k do
6: topP aths ← topP aths ∪ NextHeavyPath (E, `)
For instance, when the heaviest path of length 1 is seen, say
edge (a, b) is seen, it can be extended to paths of length 2 by
accessing edges connected to its end points, as represented
by B2 in Figure 3. Similarly, the heaviest path of length 2
can be extended by an edge using random access to obtain
path(s) in buffer B3 . Obviously, we do not extend all shorter
path segments and extend only the one that is known to be
the next heaviest path of shorter length in its own buffer.
We describe this framework in more detail when we present
the final algorithm next.
LIMITATION 3. SAA tends to produce the same subpath multiple times. For example, when SAA is required to
produce the heaviest path of length 3 for the graph in Figure 2, for paths (a0 , b0 , c0 , d0i ), i ∈ [1, n], it produces the length
2 sub-path (a0 , b0 , c0 ) n times as it does not maintain shorter
path segments. This results in more edge reads and significantly increases the running time.
5.
REPEATED SORTED ACCESS(RSA) ALGORITHM
We start this section by providing an outline of our main
algorithm for solving HPP. Our algorithm maintains a buffer
Bi for storing paths of length i explored and not released
(P.w < θ), where 2 ≤ i ≤ `. Let threshold θi denote an
upper bound on the weight of any path of length i that
has never been inserted into Bi . Algorithm 2 describes the
overall approach. It takes as input a list of edges E sorted
in non-increasing order of edge weights, and parameters `
and k. It calls the NextHeavyPath method (Algorithm 3)
repeatedly until the top-k heaviest paths of length ` are
found.
Algorithm 31 describes the NextHeavyPath method. It
takes as input a list of edges E sorted in non-increasing order of edge weights, and the desired path length l, l ≥ 2.
It is a recursive algorithm that produces heaviest paths of
shorter lengths on demand, and extends them with edges to
produce paths of length `. The base case for this recursion
is when l = 1 and the algorithm reads the next edge from
the sorted list of edges. The ReadEdge method returns the
heaviest unseen edge in E (sorted access, line 2). If l < `,
the path of length l obtained as a result of the recursion
is extended by one hop to produce paths of length l + 1.
Specifically, a path of length l < ` is extended using edges
(random access, lines 8 and 10) that can be appended to either one of its ends (returned by method EndNodes). The
“+” operator for appending an edge to a path is defined in
a way that guarantees no cycles are created. The threshold
θl is updated aggressively when the next heaviest path of
length l − 1 is released from Bl−1 . This is done by calling
the method RemoveTopPath for buffer Bl−1 and returning the resulting path. If there is no path already in the
buffer that beats θl−1 , this results in more recursion using
smaller l, until this condition becomes true at some point
during the execution. At any point during the execution if
i < ` and the next heaviest path of length i (P) has been obtained (Bi .topScore > θi ), θi+1 is more intuitively updated
as θi+1 = P.W + Wmax 2 .
THEOREM 1. Algorithm RSA correctly finds top-k heaviest paths of length `.
Proof. The proof is by induction. The base case is for going
from edges to paths of length 2. Given that all of the edges above
depth d are extended, the heaviest path that can be created from
them is already in B2 . The weight of the heaviest path that can be
created from lighter edges is at most 2 × wd . If the heaviest path in
B2 is heavier than 2 × wd , then it must be the heaviest path of length
2. Assuming the heaviest paths of length l are produced correctly
in sorted order, we show the heaviest path of length l + 1 is found
correctly. Suppose P , the heaviest path of length l + 1, is created
for the first time by extending Q, which is the nth heaviest path
of length l. The next heaviest path of length l is either already in
Bl or has not been created yet. Therefore, max(θl , Bl .topScore) is
an upper bound on the next heaviest path of length l that has not
been identified yet. Any such path can be extended by an edge of
weight at most wmax . Suppose when the mth heaviest path of length
l is seen, max(θl , Bl .topScore) + wmax is updated to a value smaller
than P.weight. It is guaranteed that P is already in Bl+1 and has
the highest weight in that buffer. In other words, when the threshold
1
In order to avoid StackOverflow we do a non-recursive implementation. We use the recursive pseudo code because it is more intuitive.
2
We use a bound tighter threshold, line 14 of Algorithm 3.
2. Finding cores of communities: Find all maximal frequent sub-paths in top-k paths. Frequent sub-paths
are those that appear in more number of paths than
the popularity parameter.
3. Aggregate top-k paths and form the aggregate graph of
top paths
4. Given a core and the aggregate graph of top paths, do
the following: add any edges that connect any other
researchers in the aggregate graph of top paths to the
core. This shapes a community of core researchers
that appear on many heavy paths. It also adds those
other researchers who are strongly connected to the
core community.
(a) Community 1 (Chinese University of Hong
Kong, Aliyun.com, MSRA)
5. Communities can grow using larger values for k that
adds more nodes to the picture. In a way, k can be
used to zoom in and zoom out of communities.
(b) Community 2 (Yahoo!, Google)
Figure 4:
Two sample communities along with their cores,
found using parameters ` = 5, k = 100 and support = 10.
is smaller than P.weight, the difference between the weight of P and
next heaviest path of length l is more than wmax . Now, paths of
length l + 1 that can be created from heavier paths of length l are
already in the buffer, and no unseen path of length l can be extended
to create a path heavier than P . Therefore, P is guaranteed to be
the heaviest path. The preceding arguments hold for top-k heaviest
paths where k > 1.
6.
COMMUNITY DETECTION CASE-STUDY
Suppose we have a co-authorship graph that connects researchers according to their publications. The edge weight
is defined as the sum of ”pairwise collaboration credits” that
researchers get from each paper they publish together. Pairwise collaboration credit for a pair of authors in a paper, is
1/authorN um. Where authorN um is the number of authors of a paper. Given a collection of publications, we define an edge between two researchers if they have published
at least one paper together. Then use the sum of ”pairwise
collaboration credits” they get from all of the papers they
have published together as the edge weight. Single author
papers do not contribute to edge weights. We further normalize edge weights by the maximum so that they are all in
[0, 1]. Using this graph, and given parameters ` and k and
a popularity 3 parameter, we do the following,
1. Find top-k heaviest paths of length l
3
Same as support in frequent pattern mining
Using DBLP data [27], we constructed this graph using papers published between 2008 and 2011 (inclusive),
presented in the main program one of these conferences:
”SDM”, ”PKDD”, ”ICDM”, ”SIGKDD”, ”CIKM”, ”SIGIR”,
”ICML”, ”NIPS”, ”WWW” and ”WSDM”. Figure 4, shows
two of the communities we find using parameters l = 5,
k = 100 and popularity = 10. We provide this graph
(DBLPG), and the top-100 paths of length 5, to enable the
reader to reproduce the communities described 4 . We notice that the core of community 1 (Figure 4(a)) contains
the pair of researchers with the highest edge weight in the
graph. This shows a long lasting collaboration. We did a
quick google search about each of the core communities in
DBLPG. Irwin King and Michael R. Lyu work for the Chinese University of Hong Kong. They are connected to GuiRong Xue who is the senior director at Aliyun.com and he is
connected to a principal researcher from Microsoft research
Asia (Zheng Chen).
Most of the researchers in community 2 (Figure 4(a)) have
worked for Yahoo! Research Labs, including the three researchers in the core. Andrei Z. Broder is currently a distinguished scientist at Google. He has previously worked at
Yahoo!, IBM and Altavista. We also found on his Wikipedia
page that he is the inventor of MinHash locality sensitive
hashing. Evgeniy Gabrilovich, is a senior staff research scientist at Google who also used to have a central role at
Yahoo! research. Vanja Josifovski apparently still works for
Yahoo! and is a Principal Research Scientist and the Lead
of the Performance Advertising Group at Yahoo! Research.
Most of the other researchers in community 2 work for yahoo! as their emails appear on their publications. Some may
have moved to Google recently!
It is worth noting that we discovered three other cores as
well in our output. Due to space limitations, we only report
the cores: core 3 = {Andrei Z. Broder, Evgeniy Gabrilovich,
Xuerui Wang }(weight = 0.73), core 4 = {Shuicheng Yan,
Ning Liu, Jun Yan, Zheng Chen } and core 5 = {Ning Liu,
Jun Yan, Zheng Chen, Weizhu Chen }. Xuerui Wang is
also a scientist at Yahoo! labs. Cores 4 and 5 have size
4. We notice that a sub-path of length 3 is shared between
these two, with all researchers from Microsoft Research Asia.
Community 1 is connected to two of these three researchers
4
http://ucalgary.ca/∼mkhabbaz/DBLPG.zip
through Gui-Rong Xue. This shows a potential for collaboration between these two cores in the future! who are the
Chinese University of Hong Kong and Microsoft Research
Asia. Changing the parameters l, k and popularity, changes
the output in the form presented in Figure 6. It can also
help with creating a visualization tool that adapts its output
with the current parameter values and be used for visualizing hierarchies and relationships of communities at different
scales. Such a tool needs to run RSA repeatedly with different parameter values in order to change its view. This highlights the importance of scalability and efficiency for HPP
problem. We leave the continuation and implementation of
this idea to our future work.
The case study presented in this section, is also an example of an application where finding the exact solution is
required. This is because we focus on 100 heaviest paths
and frequent pattern mining in order to discover the cores
of communities. Heuristic or any sort of approximate solutions may result in paths that do not necessarily go through
the cores that form the top communities. Top-k heaviest
paths, rely on their cores to be heavy!
7.
7.1
EXPERIMENTAL ANALYSIS
Reproducibility
We use four graphs in our experiments. We create graphs
using Cora5 , last.fm6 and DBLP datasets. Cora(70 nodes1580 edges): nodes represent research topics and edge
weights are defined using the average fraction of citations between the two topics. Last.fm(40k nodes-183k edges):
we crawled the existing collection of playlists. Nodes represent songs and edges represent co-occurrence in playlists.
Edge weight is defined using the Dice coefficient7 . BAY(321k
nodes-400k edges): it represents the road network in San
Francisco Bay area8 . For this graph we perform the C − W
(C=1) transformation described in Section 2, and solve `TSP because distance minimization makes more sense for
road networks. DBLPG(4.5k nodes-9k edges): Section 6 describes how the graph is created. In all cases, the
graph we work with has edge weights in [0 − 1] and the input graph is a text file that stores one edge per line in this
format: ”source destination weight”. You can access a .zip
file containing all the preprocessed graphs9 . We do not use
graphs with millions of edges since the problem is NP-hard
and we focus on finding exact answers. On the other hand,
all of the graphs we use in experiments deal with practical
real life applications. One interesting fact we discovered in
our experiments is that the smallest of the graphs (Cora) we
use, is one of the most challenging scenarios. This is due to
the fact that its average node degree is higher than others
and this results in an exponentially larger number of longer
paths when ` increases. We do not report any results regarding the DFS algorithm because it is inferior to Dynamic
Programing in all cases. Source code of all algorithms, along
with instructions for running can be downloaded10 . All experiments were performed on a Linux machine with 64GB
5
6
7
8
http://www.cs.umass.edu/~mccallum/data
of main memory and 2.93GHz-8Mb Cache CPU. To be consistent, we allocated 12GB of memory for each run. In all
of the cases where RSA11 runs out of memory, other algorithms either also run out of the 12GB allocated memory, or
do not terminate after running for a couple of days. Our
implementation is in Java. For all algorithms other than
RSA, we provide implementation such that the time spent
on Garbage Collection can be measured accurately using the
Jstat tool12 . All of the figures that present running time
use log scale on Y-axis. We further analyzed GC time and
found it is negligible compared to the total running time in
all cases as supposed to be since Java is one of the most
reliable programming languages.
7.2
Empirical Evaluation
Figure 5(a-c) show the running time for finding the top-1
path of various lengths for Cora, last.fm and Bay graphs.
Average node degree plays a key role in the empirical hardness of HPP. This is due to the fact that the complexity of
the problem increases exponentially with node degree.
The running time increases with ` for all algorithms, as
supposed to be. SAA does not manage to go beyond short
path lengths in most cases, the reasons being those limitations mentioned in Section 4. Beyond some depth down the
sorted edge list, SAA becomes inefficient due to the increasing complexity of the problem and the fact that it either does
not manage to construct heavy paths early on or the fact
that the threshold decays slowly. For shorter path lengths
SAA terminates earlier than Dynamic Programing because
it does pruning and uses a threshold for termination while
Dynamic Programing explores a larger search space. RSA
performs considerably more efficient than other algorithms
except for one case on Cora graph for ` = 3 where it is
slightly less efficient than SAA. We believe this is due to the
fact that RSA uses a combination of sorted and random access. Random access results in fundamental improvements
in the running time as we see in most cases. However, in
some rare cases SAA may get lucky because it avoids random access and also the edge weights are such that result in
fast termination. Albeit, at the cost of losing all the other
experiments. All in all, our findings show that RSA is a way
more scalable and reliable algorithm. Dynamic Programing
behaves consistently and manages to achieve pruning by aggregating shorter path segments early on. However, it finds
the heaviest path ending at every node and although it scales
to longer lengths, it is orders of magnitude slower than RSA.
We would like to highlight again that we use log scale on Yaxis and the difference in the running times is considerable.
All in all, there is no question about the surprising scalability of RSA in all experiments.
Figures 5(d-f), compare the running times of SAA and
RSA for varying k and fixed ` = 4, since these two are the algorithms designed to work as top-k algorithms. In all cases,
RSA is more efficient and it spends a smaller marginal running time compared to SAA for finding top-k. We observe
sudden jumps in the running time of SAA when k changes
while RSA continues to work reliably for values of k less
than 100 in these experiments and scales more smoothly.
www.last.fm
dice(i, j) =
2|i∩j|
|i|+|j|
http://www.dis.uniroma1.it/
9
http://ucalgary.ca/∼mkhabbaz/Graphs.zip
10
http://ucalgary.ca/∼mkhabbaz/HPPCase.zip
11
We use the name HeavyPath for our implementation of the algorithm.
12
Java does garbage collection automatically using GC algorithms.
http://docs.oracle.com/javase/1.5.0/docs/tooldocs/share/jstat.html
(a) Vary `, k = 1, Cora
(b) Vary `, k = 1, last.fm
(c) Vary `, k = 1, Bay
(d) ` = 4, vary k, Cora
(e) ` = 4, vary k, last.fm
(f) ` = 10, vary k, Bay
Figure 5:
Running time comparisons for exact algorithms with different parameter settings.
(a) Vary `, k = 1, Cora
Figure 6:
(b) Vary `, k = 1, last.fm
Comparing the number of edge reads between SAA and RSA for different `
We further compare the number of edge reads of SAA
and RSA in Figure 6, that is a system independent notion
of the running time. We observe patterns very similar to
Figures 5(a-c). This highlights the fact that the number of
edges read during the execution is the main factor determining the running time. We observe in some cases (for small
`), SAA reads fewer edges but spends slightly more time for
execution. This is due to the fact that RSA is performing
random access using the adjacency matrix of the graph to
construct longer paths. However, SAA performs the Join
operation and looks up edges in a hash table which itself
requires few milliseconds of time overall.
Finally, Figure 7 shows the running times and the number
of edge reads on the DBLPG graph for finding the heaviest
paths of different lengths. We notice on this Graph, SAA algorithm performs more poorly compared with the Dynamic
Programming algorithm. Of course, RSA beats all algorithms in all cases and finds the exact solution almost 10
times faster than the rest. Figures 7(b) and 8, illustrate
this.
8.
(c) Vary `, k = 1, Bay
DISCUSSIONS
Our theoretical results prove that HPP is np-hard and
inapproximable in the general case. While it makes sense
to seek more tractable problem definitions by adding constraints, still we show experimentally that we can achieve
scalability practical enough for our main algorithm to be
used in a real time fashion and be used in smart software
technologies. We use a branch and bound solution for finding
the exact solution, inspired by top-k algorithms. We use a
constraint `, on edge weights and this makes our comparison
for finding heavy paths, representing significantly important
sequences in the graph more fair. Comparing path weights
of paths with different lengths may not make enough sense
in the general case. We show exact solution even for small
`, if obtained, can be used in practice effectively, while other
algorithms fail to report even for small ` in a timely manner.
It is obviously nice to scale to longer paths and this requires
smarter strategies for reducing θ during the execution that
leads to faster search space pruning. We make a proposal
for designing a network visualization tool-box with zoom-in
and zoom-out functionalities with changing k. We find our
presented case study practical but obviously as mentioned
in Section 1, applications are not restricted to this and there
are other applications such as those in bioinformatics. We
can use high probability sequences in order to fix technology
related errors such as replacing inaccurate or suspicious entries or filling in missing values in biological sequences. We
(a) Vary `, k = 1, DBLPG: Running Time(ms) (b) Vary `, k = 1, DBLPG: Number of Edge Reads
Figure 7:
Figure 8:
Scalability results on DBLPG dataset
Heaviest paths of ` = 1 to ` = 5 on DBLPG
Graph. SAA terminates for ` = 2 before processing all edges
Figure 9:
at depth 28 (out of 9060 edges).
Facebook and DBLPG using Egonetwork circles.
The minimum threshold
Heaviest path weights of ` = 1 to ` = 5 for Twitter,
value when SAA terminates is 1.4338 and smaller than the
weight of the heaviest path of ` = 2(1.4388). For ` > 2, SAA
can not terminate without processing all edges i.e. almost
equivalent to the DFS algorithm. As can be seen in Figure 7,
SAA terminates earlier than Dynamic Programing for ` = 2
but when ` > 2, Dynamic Programing becomes more efficient
since SAA behaves as naive as DFS.
can also use these high probability sequences for feature extraction from DNA which is an extremely long sequence of
symbols in bioinformatics for classification and other learning tasks. One reason we choose visualizing bibliographic
networks is presenting interpretable results for computer science audience. While there is a variety of solutions to be
used in these applications, we want to propose a solution to
a well-defined problem in computer science similar to TSP,
that can be significantly leveraged for creating technologies
in software industry. Another way to add constraints to
the problem and make pruning more possible is through designing reliable heuristic algorithms with constraints on the
distribution of edge weights. Of course, this requires careful
testing for estimating distributions and we can focus on this
extension in the future.
Recently, there have been advancements in the field of social networks described as Egonetworks 13 . Egonetworks are
typically presented as groups of people involved in different
”circles”. Our bibliographic network described in Section 6,
is similar to the notion of circles in Egonetwork where each
list of authors represents a circle. In an extreme scenario
there may be many nodes involved in each circle in domains
13
http://snap.stanford.edu/data/index.html
such as concrete social networks. Any problem definition for
discovering dense or significant subgraphs according to such
networks can result in finding significant sub-structures. We
propose an effective approach for constructing such a graph.
Based on the data available we prefer to make no judgement
about the significance of individuals. Our proposed solution
based on finding top-k heavy paths is proven to be scalable by experiments. On the other hands we can manage to
change the output using simple parameters such as k, that
enables us to zoom-in and zoom-out. Our argument for reliability of our output is the following observation. We notice
many heavy paths share the same exact shorter sub-paths.
This emphasizes the importance of some collaborations in
the bibliographic network around which an important and
successful research community is formed. This led to our
definition of community detection approach that makes use
of maximal frequent subpaths for discovering such frequent
short paths shared among many top-k heavy paths. We follow the same approach for constructing graphs from Facebook and Twitter Egonetworks. In particular, we only use
circles and convert mutual presence of nodes in circles to define edges. We find the same exact method described in Section 6, in order to define edge weights. Figure 8, reports the
heaviest path weights of ` = 1 to ` = 5, and compares them
with those obtained from DBLP data. We use the union
of all circles first and then construct the graph. Since edge
weights are all scaled to (0 − 1), we expect heaviest paths
of different length to be in the same range for all datasets.
In comparison, we notice in Figure 9, that all graphs follow
a similar curve for increasing `. We notice that the slope
decreases with `, and it is quite likely path weight converges
quickly for large `. This increases the chances of finding reliable heuristic algorithms with output close enough to the
exact answer of the np-hard problem. Facebook and Twitter datasets result in paths with slightly heavier weights and
we believe this is because in real life networks people with
strong connections belong to many active networks while
researchers may not be as socially active as ordinary people. Although the difference in path weights is small and
almost negligible. The most notable facts to highlight are
the similar shape of curves as well as the decreasing slope.
9.
RELATED WORK
In [10], Hansen and Golbeck make a case for recommending playlists and other collections of items. The AutoDJ system of Platt et al. [11] is one of the early works on playlist
generation that uses acoustic features of songs rather than
a graph based representation. Random walk methods were
used in [8], in order to generate lists of videos in a co-view
graph. We formalized and presented our view of generating
lists using random walks in introduction for recommending
lists as an ordered package of items to users.
The HPP in a special case was studied recently by [7] and
they defined a notion of Stable Clusters for the specific application of finding persistent chatter in the blogosphere. Unlike our setting, the graph associated with their application
is `-partite and acyclic. Absence of cycles makes their technical problem more tractable than ours. The most efficient
algorithm presented in their work is a BFS based method
that follows a dynamic programming approach.
We followed dynamic programming algorithm that also avoids cycles and found this approach extremely inefficient compared
to RSA.
Rankjoin was first proposed by Ilyas et al. [26, 25, 24,
30, 33] to produce top-k join tuples in a relational database.
We show how to convert Rankjoin to HPP. Furthermore, no
one has studied Rankjoin with self-joins to the best of our
knowledge. In [4], authors provide a comprehensive summary of this algorithm and do optimality analysis as done
in the theoretical database literature. Their conclusion is
that beyond ` = 2, we can not guarantee any optimality results for Rankjoin. This also matches our results regarding
HPP being np-hard and inapproximable.
10.
CONCLUSIONS AND FUTURE WORK
Finding the top-k heaviest paths in a graph is an interesting problem with many practical applications. We discuss
the hardness of this problem. We focus on developing practical exact algorithms for this problem. We use an innovative
top-k query processing algorithm. We motivate the problem from several different perspectives and discuss possible
applications. We present a case study on core community
detection. Our findings suggest that our case study can be
further extended to a community detection tool that discovers the key influential people in a network represented by
a weighted graph. Our future work will focus on more applications of RSA, specially those in network analysis, community detection, recommender systems and bioinformatics.
We would also like to investigate more scalable algorithms
and threshold update strategies within the same framework.
We also intend to design probabilistic and robust heuris-
tic algorithms that work under memory and other types of
constraints such as edge distribution.
11.
REFERENCES
[1] Andrei Broader, Ravi Kumar, Farzin Maghoul,
Prabhakar Raghavan, Sridhar Rajagopalan, Raymie
Stata, Andrew Tomkins, Janet Wiener Graph
Structure in the Web, Computer Networks, 2000.
[2] N. Garg. et al. S. Arora and G. Karakostas. A 2 +
epsilon approximation algorithm for the k-MST
problem., Math. Prog., 2006.
[3] N. Garg. et al. a 2-approximation for the k-mst
problem in graphs., STOC, 2005.
[4] Schnaitter, Karl and Polyzotis, Neoklis. Optimal
algorithms for evaluating rank joins in database
systems, TODS, 2010.
[5] Schnaitter, Karl and Polyzotis, Neoklis. Evaluating
rank joins with optimal cost, PODS, 2008.
[6] U Kang, Spiros Papadimitriou, Jimeng Sun,
Hanghang Tong. Centralities in Large Networks:
Algorithms and Observations, SDM, 2011.
[7] Bansal, Nilesh and Chiang, Fei and Koudas, Nick and
Tompa, Frank Wm. Seeking stable clusters in the
blogosphere, VLDB, 2007.
[8] Shumeet Baluja, Rohan Seth, D. Sivakumar, Yushi
Jing, Jay Yagnik, Shankar Kumar, Deepak
Ravichandran, Mohamed Aly, Video suggestion and
discovery for youtube: taking random walks through
the view graph, WWW, 2008.
[9] Fields, Ben and Rhodes, Christophe and d’Inverno,
Mark, Using Song Social Tags and Topic Models to
Describe and Compare Playlists, Workshop on Music
Recommendation and Discovery, 2010.
[10] Derek L. Hansen and Jennifer Golbeck, Mixing it up:
recommending collections of items, CHI, 2009.
[11] John C. Platt, Christopher J. C. Burges, S. Swenson,
C. Weare, A. Zheng. Learning a Gaussian Process
Prior for Automatically Generating Music Playlists,
NIPS, 2001.
[12] Hauver, D.B. and French, J.C. Web Delivering of
Music, Flycasting: using collaborative filtering to
generate a playlist for online radio, 2001.
[13] Conor Hayes and Padraig Cunningham and Patrick
Clerkin and Marco Grimaldi, Programme Driven
Music Radio, ECAI, 2002.
[14] Luke Barrington and Reid Oda and Gert Lanckriet,
SMARTER THAN GENIUS? HUMAN
EVALUATION OF MUSIC RECOMMENDER
SYSTEMS, International Society for Music
Information Retrieval, 2009.
[15] Toni Cebrián and Marc Planagumà and Paulo Villegas
and Xavier Amatriain, Music recommendations with
temporal context awareness, RecSys, 2010.
[16] Michael D. Ekstrand and Praveen Kannan and James
A. Stemper and John T. Butler and Joseph A.
Konstan and John Riedl, Automatically building
research reading lists, RecSys, 2010.
[17] Cai-Nicolas Ziegler and Sean M. McNee and Joseph A.
Konstan and Georg Lausen, Improving
recommendation lists through topic diversification,
WWW, 2005.
[18] X. Zhang and J. Chomicki, Preference Queries over
Sets, IEEE International Conference on Data
Engineering (ICDE), 2011.
[19] Holger Bast and Debapriyo Majumdar and Ralf
Schenkel and Martin Theobald and Gerhard Weikum,
IO-Top-k: Index-access Optimized Top-k Query
Processing, VLDB, 2006.
[20] Measures of the Amount of Ecologic Association
Between Species, Lee R. Dice, Ecology, 1945.
[21] Hu, Yifan and Koren, Yehuda and Volinsky, Chris,
Collaborative Filtering for Implicit Feedback Datasets,
ICDM, 2008.
[22] Yehuda Koen, Tutorial on recent progress in
collaborative filtering, RecSys, 2008.
[23] desJardins, M. and Eaton, E. and Wagstaff, K. L.,
Learning user preferences for sets of objects, ICML
2006.
[24] Li, Chengkai and Chang, Kevin Chen-Chuan and
Ilyas, Ihab F. and Song, Sumin, RankSQL: Query
Algebra and Optimization for Relational Top-k
Queries, SIGMOD, 2005.
[25] Ilyas, Ihab F. and Shah, Rahul and Aref, Walid G.
and Vitter, Jeffrey Scott and Elmagarmid, Ahmed K.,
Rank-aware query optimization, SIGMOD, 2004.
[26] Ihab F. Ilyas and Walid G. Aref and Ahmed K.
Elmagarmid, Supporting Top-k Join Queries in
Relational Databases, VLDB, 2003.
[27] Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang,
and Zhong Su. ArnetMiner, Extraction and Mining of
Academic Social Networks., SIGKDD, 2008.
[28] Gibson, David and Kumar, Ravi and Tomkins,
Andrew, Discovering large dense subgraphs in massive
graphs, VLDB, 2005.
[29] M. Theobald, Gerhard Weikum, Ralf Schenkel, Top-k
Query Evaluation with Probabilistic Guarantees.,
VLDB, 2004.
[30] Jonathan Finger, Neoklis Polyzotis, Robust and
Efficient Algorithms for Rank Join Evaluation,
SIGMOD, 2009.
[31] Johnson, David and Krishnan, Shankar and Chhugani,
Jatin and Kumar, Subodh and Venkatasubramanian,
Suresh, Compressing large boolean matrices using
reordering techniques, VLDB, 2004.
[32] Lappas, Theodoros and Liu, Kun and Terzi, Evimaria,
Finding a team of experts in social networks, KDD,
2009.
[33] Davide Martinenghi and Marco Tagliasacchi,
Cost-Aware Rank Join with Random and Sorted
Access, TKDE, 2011.
[34] Ronald Fagin and Amnon Lotem and Moni Naor,
Optimal Aggregation Algorithms for Middleware, J. of
Comp. Sys. Sc., 2003.
[35] Sanjeev Arora and George Karakostas, A 2 +
approximation algorithm for the k-MST problem,
Math. Prog., 2006.
[36] A Dynamic Programming Approach to Sequencing
Problems, Michael Held and Richard M. Karp, 1962.
[37] Naveen Garg, Saving an epsilon: a 2-approximation
for the k-MST problem in graphs, STOC, 2005.
[38] M. Choudhury, M. Feldman, S. Amer-Yahia, N.
Golbandi, R. Lempel, and C. Yu., Automatic
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
construction of travel itineraries using social
breadcrumbs., HT, 2010.
Community Detection in Graphs, Santo Fortunato,
Physics Reports, 2009.
Mohsen Jamali, Martin Ester TrustWalker: a random
walk model for combining trust-based and item-based
recommendation., KDD, 2009.
Eirini Spyropoulou, Tijl De Bie, Interesting
Multi-Relational Patterns., ICDM, 2011.
Michael Mathioudakis, Francesco Bonchi, Carlos
Castillo, Aristides Gionis, Antti Ukkonen,
Sparsification of Influence Networks., KDD, 2011.
Jiawei Han, Jian Pei, Yiwen Yin, Mining frequent
patterns without candidate generation., SIGMOD,
2000.
Gerhard Weikum, Database Researchers: Plumbers or
Thinkers?, EDBT, 2011.
Mohammad Khabbaz, Smriti Bhagat, Laks V. S.
Lakshmanan, Finding Heavy Paths in Graphs: A
Rank Join Approach., CoRR abs/1112.1117, 2011.
Hansen, Derek L. and Golbeck, Jennifer, Mixing it up:
recommending collections of items, CHI 2009.
Mohammad Khabbaz and Smriti Bhagat and Laks V.
S. Lakshmanan, Finding Heavy Paths in Graphs: A
Rank Join Approach, Dept. of CS, UBC.
Chandra Chekuri and Martin PÂt’al, A recursive
greedy algorithm for walks in directed graphs, FOCS,
2005.
| 8cs.DS
|
A Language-theoretic View on Network Protocols
Pierre Ganty1 , Boris Köpf1 , and Pedro Valero1,2
1
IMDEA Software Institute, Madrid, Spain
Universidad Politécnica de Madrid, Spain
{pierre.ganty,boris.koepf,pedro.valero}@imdea.org
arXiv:1610.07198v2 [cs.FL] 10 Jul 2017
2
Abstract. Input validation is the first line of defense against malformed
or malicious inputs. It is therefore critical that the validator (which is
often part of the parser) is free of bugs.
To build dependable input validators, we propose using parser generators
for context-free languages. In the context of network protocols, various
works have pointed at context-free languages as falling short to specify
precisely or concisely common idioms found in protocols. We review those
assessments and perform a rigorous, language-theoretic analysis of several
common protocol idioms. We then demonstrate the practical value of our
findings by developing a modular, robust, and efficient input validator
for HTTP relying on context-free grammars and regular expressions.
1
Introduction
Input validation, often carried out during parsing, is the first line of defense
against malformed or maliciously crafted inputs. As the following reports demonstrate, bugs make parsers vulnerable, hence prone to attacks: a bug in the URL
parser enabled attackers to recover user credentials from a widely used password
manager [14], a bug in the RTF parser led to a vulnerability in Word 2010 [19],
and a lack of input validation in the Bash shell has been used for privilege
escalation by a remote attacker [13] just to cite a few. To stop the flow of such
reports improved approaches for building input validators are needed.
To build dependable input validators, an approach is to rely on mature parsing
technologies. As a candidate, consider parser generators for context-free languages
(hereafter CFLs). Their main qualities are:
1. The code for parsing is synthesized automatically from a grammar specification, shifting the risk of programming errors away from the architect of the
validator to the designer of the parser generator;
2. CFL is the most expressive class of languages supported by trustworthy implementation of parser generators. Here by trustworthy we mean an implementation that either stood the test of time like Flex, Bison, or ANTLR, or that
has been formally verified like the certified implementations of Valiant’s [3]
and CYK [11] algorithms.
Relying on established CFL technology is an asset compared to existing solutions
which are either programmed from scratch [8] or generated from ad hoc parser
generators [2, 21].
2
Pierre Ganty, Boris Köpf, and Pedro Valero
However, even though CFL is the most expressive language class with trustworthy parser generators, previous works suggest CFLs are not enough for network
messages. Specifically, they claim the impossibility or difficulty to specify precisely
the various idioms found in network messages using CFLs. For example, some
authors argue that CFL are insufficiently expressive because “data fields that are
preceded by their actual length (which is common in several network protocols)
cannot be expressed in a context-free grammar” [7]. Yet other authors suggest
that going beyond CFLs is merely required for conciseness of expression, because
“it is possible to rewrite these grammars to [. . . ] be context-free, but the resulting
specification is much more awkward” [4]. Surprisingly, the arguments made are
not backed up by any formalization or proof.
In this paper we formally analyze which idioms can and which cannot be
(concisely) specified using CFLs, and we turn the results into practice by building
an input validator for HTTP messages entirely based on CFL technology.
As the main contributions of our analysis we find out that:
– Length fields of bounded size are finite and hence form a regular language.
However, while they can be concisely represented in terms of a context-free
grammar, every finite automaton that recognizes them grows exponentially
with the bound. In contrast, length fields of unbounded size cannot even be
expressed as a finite intersection of CFLs.
– Equality tests between words of bounded size again form a regular language
but, as opposed to length fields, they cannot be compactly represented in
terms of a context-free grammar. They do, however, allow for a concise
representation in terms of a finite intersection of context-free grammars.
Specifically, we show that both the grammar and the number of membership
checks grows only linearly with the size bound, which has interesting practical
implications, see below. As in the case of length fields, equality checks between
words of unbounded size cannot be expressed as a finite intersection of CFLs.
We consider finite intersections of CFLs because they are (a) strictly more
expressive than CFLs, and (b) checking membership in the intersection of CFLs
is equivalent to checking membership in each individual CFL.
These results lead to a principled and modular approach to input validation:
several CFL parsers are run on the input and their boolean results (whether the
input belongs or not to the CFL) are combined following a predefined logic to
decide whether or not the input message conforms to the standard (that specifies
what valid messages are).
We demonstrate that this approach is practical by implementing a proof of
concept input validator for a large subset of the HTTP protocol, covering a
significant number of the idioms found in network messages. Our input validator,
called HTTPValidator [24] draws inspiration from HTTPolice [8], a state of the
art input validator for HTTP messages built from scratch by the open source
community. HTTPValidator is close to achieve feature parity (in terms of checks)
with HTTPolice and offers competitive performance.
Summary of Contributions. In summary, our contributions are both foundational
and applied. On the foundational side we perform a language-theoretic analysis
A Language-theoretic View on Network Protocols
3
of important protocol idioms, making a step towards more rigor in the field. On
the applied side we show how to implement an input validator for HTTP using
off-the-shelf parser generators.
Paper structure. Section 2 introduces basic language-theoretic and input validation concepts, Section 3 discusses the case of length fields, including the chunked
messages, whereas Section 4 considers the case of comparisons. We show the
practicality of our approach in Section 5, before concluding with related work
(Section 6), conclusions and future work (Section 7).
2
Preliminaries
Language Theory. We begin by introducing the language-theoretic context needed
for our development. An alphabet Σ is a nonempty finite set of symbols. A word
w is a finite sequence of symbols of Σ where the empty sequence is denoted ε.
A language is a set of words and the set of all words over Σ is denoted Σ ∗ . We
denote by |w| the length of w. Further define (w)i as the i-th symbol of w if
1 ≤ i ≤ |w| and ε otherwise. Given a nonempty subset X of Σ and i ∈ N define
X i as {w ∈ X ∗ | |w| = i}.
We assume the reader is familiar with common operations on languages such
as concatenation and boolean combinations. Likewise, we count on the reader’s
familiarity with regular languages and finite-state automata. Yet we next give a
description of context-free grammars, which are the formal basis of our work.
A context-free grammar (or grammar for short) is a tuple G = (V, Σ, S, R)
where V is a finite set of variables (or non-terminals) including the start variable
S; Σ is an alphabet (or set of terminals), R ⊆ V × (Σ ∪ V )∗ is a finite set of rules.
We often write X → w for a rule (X, w) ∈ R. We define a step as the binary
relation ⇒ on (V ∪ Σ)∗ given by u ⇒ v if there exists a rule X → w of G such
that u = α X β and v = α w β for some α, β ∈ (V ∪ Σ)∗ . Define u ⇒∗ v if there
exists a n ≥ 0 steps sequence u0 ⇒ u1 ⇒ . . . ⇒ un such that u0 = u and un = v.
A step sequence u ⇒∗ w is called a derivation whenever u = S and w ∈ Σ ∗ .
Define L(G) = {w ∈ Σ ∗ | S ⇒∗ w} and call it the language generated by G. A
language L is said to be context-free, or CFL, if there exists a grammar G such
that L = L(G). The size of a grammar
is the sum of the sizes of its production
P
rules R, that is, it is given by (X,w)∈R 1+|w|.
Input Validation. In this paper, validating an input means checking whether it
belongs to a language. In particular, no data structure is filled and no information
is extracted from the input other than its membership status. Thus, validating
an input w for a language L means deciding whether w is a member of L which
is also known as the membership problem. To specify L, we use context-free
grammars or regular expressions.
4
3
Pierre Ganty, Boris Köpf, and Pedro Valero
Formal Analysis of Length Fields
Length fields, whose role is to specify the length of subsequent fields, are commonly
found in network protocols such as HTTP [9], SIP [23], DNS [20] and UDP [22].
As an example, consider the following HTTP POST message:
POST /1/notification/list HTTP/1.1\r\n
Content-Length: 47\r\n\r\n
{"header":{},"query":{"count":100},"answer":{}}
The length field begins after the keyword Content-Length: and terminates
before the carriage return/newline \r\n. Its content, i.e. 47, describes the length
of the message body, which is the string coming after the double \r\n.
In this section, we characterize length fields from the point of view of formal
language theory. We begin with a formalization aiming to capture their essence,
and then characterize the class of languages specifying them in the bounded and
unbounded cases. We consider both cases because some protocols, such as DNS
and UDP, require length fields to have fixed size, while others, such as HTTP
and SIP, have no such restriction. We conclude by leveraging these results to
analyze chunked transfer encoding.
3.1
Modeling Length Fields
To model length fields, we will work with formal languages over an alphabet Σ.
For the example of HTTP, Σ would be the ASCII character set.
Fixed Size. To describe length fields of finite size n > 0 we define the language
Llen (n) over Σ = B ∪ W as follows:
Llen (n) = {x w | x ∈ B n , w ∈ W ∗ , |w| =
def
Pn−1
i=0
(x)i+1 · bi }
where B = {0, . . . , b − 1} for an integer b > 1. Intuitively, Llen (n) represents
the same number twice, using two different encodings: first b-ary as x and
then unary
Pn−1 as w, where the relationship between both encodings is given by
|w| = i=0 (x)i+1 · bi . For example, let n = 3, B = {0, 1} and W = {a, b, c} the
word 110abc consists of the binary representation of 3 = (1 · 20 ) + (1 · 21 ) + (0 · 22 )
followed by a word (abc) of length 3 and, therefore, 110abc ∈ Llen (3). We choose
this unconventional “least significant digit first” to keep notation simple. The
results of this section stay valid for the “most significant digit first” convention.
Unbounded Size. For describing length fields of unbounded size, observe that
any overlap between the alphabets W and B for describing the body of the
message and its length, respectively, introduces ambiguity as to where the length
field ends. A common approach to remove such ambiguities is to use a delimiter,
which is a special symbol ] not occurring in x whose aim is to separate explicitly
A Language-theoretic View on Network Protocols
5
the length field from the body of the message. We extend the definition of Llen (n)
to account for such delimiters:
L]len (n) = {x ] w | x ∈ B n , w ∈ W ∗ , |w| =
def
Pn−1
i=0
(x)i+1 · bi } .
We are now in position to define a language for describing length fields of
arbitrary and unbounded size:
L]len =
def
S
i>0
L]len (i) .
Results shown in this section remain valid when there is no overlap between
alphabets W and B. In such case the delimiter is no longer needed and removing
it from the results in Section 3.2 has no effect on them.
3.2
Unbounded Length Fields
The following theorem shows that length fields of unbounded size cannot be
specified using intersection of finitely many CFLs. This means that we need to
impose restrictions, such as size bounds, in order to specify length fields using
CFLs. We will study fixed size length fields in Section 3.3.
Theorem 1. L]len is not a finite intersection of CFLs.
To prove this result, we begin by defining the following subset of L]len :
L∠ = L]len ∩ 1∗ ]∗ a∗ .
def
Lemma 2. L∠ is not a finite intersection of CFLs. Moreover, no infinite subset
of L∠ is a finite intersection of CFLs.
The proof argument relies on semilinear sets which we recall next: a subset
of Nk , with k > 0, is called semilinear, if it can be specified as a finite union of
linear sets. A set S ⊆ Nk is called linear if there exists b ∈ Nk and a finite subset
{p1 , . . . , pm } of Nk such that S = {b + λ1 p1 + · · · + λm pm | λ1 , . . . , λm ∈ N} .
Let w̄ = hw1 , . . . , wk i be a tuple of k > 0 words, define a mapping fw̄ : Nk →
∗
w1 . . . wk∗ by fw̄ (i1 , . . . , ik ) = w1i1 . . . wkik , that is, the output of fw̄ is a word in
which the i-th component of w̄ is repeated a number of times that corresponds
to the i-th input to fw̄ . We define the preimage of fw̄ and liftings of fw̄ from
elements to subsets of Nk in the natural way.
The following result by Latteux [15] establishes a fundamental correspondence
between languages given by finite intersection of CFLs and semilinear sets.
Proposition 3 ([15, Prop 7]). Let w̄ = hw1 , . . . , wk i, k > 0, and L ⊆
w1∗ . . . wk∗ : fw̄−1 (L) is semilinear if and only if L is a finite intersection of CFLs.
Now we meet the requirements to prove Lemma 2.
6
Pierre Ganty, Boris Köpf, and Pedro Valero
Proof (Sketch). The proof of Lemma 2 relies on the observation that
L∠ = {1n ] aval | val =
Let w̄ = h1, ], ai, since
Pn−1
i=0
bi =
fw̄−1 (L∠ ) =
bn −1
b−1
n
Pn−1
i=0
bi } .
for all b > 1, we obtain:
o
i
−1
|i ∈N .
i, 1, bb−1
(1)
Next, we show this set is not semilinear using the facts that (a) the third
component grows exponentially in i, and (b) fw̄−1 (L∠ ) is an infinite set. The
definition of semilinear set then shows that by taking two elements in (b) we can
obtain a third one. We then show that those three elements violate (a) unless
they all coincide. The same reasoning remains valid when considering an infinite
subset of L∠ . The details of the proof are given in Appendix A.
t
u
Once Lemma 2 is proved, the proof of Theorem 1 easily follows.
Proof (of Theorem 1). Assume L]len is a finite intersection of CFLs. Since 1∗ ]∗ a∗
is a CFL, L∠ is also a finite intersection of CFLs contradicting Lemma 2.
t
u
Our definitions of Llen (n) and L]len do not put any constraints on the structure
of the word w that follows the length field and the delimiter (if any). In practice,
however, the word w may need to satisfy constraints beyond those on its length,
such as containment in a specific language.
Theorem 4. The language L]len ∩ {x ] w | x ∈ B ∗ , w ∈ LC } is a finite intersection of CFLs for no infinite CFL LC ⊆ W ∗ .
The proof of this Theorem follows the same argument used to prove Theorem
1. Hence, we begin by defining a subset of L]len ∩ {x ] w | x ∈ B ∗ , w ∈ LC } for
which Proposition 3 holds.
Let S be the start symbol of the grammar generating language LC . Since
language LC is infinite the following must hold: for some non terminal A and
αi ∈ W ∗ , we have S ⇒∗ α1 Aα5 ; A ⇒∗ α2 Aα4 ; A ⇒∗ α3 with α2 6= or α4 6= .
It follows that {α1 α2i α3 α4i α5 | i ≥ 0} ⊆ LC and, thus,
Li = L]len ∩ {x ] w | x ∈ B ∗ , w ∈ LC } ∩ 1∗ ]∗ α1∗ α2∗ α3∗ α4∗ α5∗
def
is an infinite language contained in 1∗ ]∗ α1∗ α2∗ α3∗ α4∗ α5∗ .
Lemma 5. Language Li is not a finite intersection of CFLs. Moreover, no
infinite subset of Li is a finite intersection of CFLs.
This Lemma is similar to Lemma 2 and so is the proof, whose details are given
in Appendix A. Finally, we proceed to prove Theorem 4 by contradiction.
Proof (of Theorem 4). Assume that L]len ∩ {x ] w | x ∈ B ∗ , w ∈ LC } is a finite
intersection of CFLs. Since 1∗ ]∗ α1∗ α2∗ α3∗ α4∗ α5∗ is context-free, Li is also a finite
intersection of CFLs, which contradicts Lemma 5.
t
u
A Language-theoretic View on Network Protocols
3.3
7
Fixed Size Length Fields
In this section we sidestep the negative results of Section 3.2 by assuming an
upper bound on the length field which indeed occurs in some network protocols.
Such is the case of the IP, UDP and DNS protocols, whose specifications [1, 20, 22]
define 16-bit fields containing the length of the data in terms of bytes. In some
cases, assuming an upper bound on the length field, even if it is not defined by
the standard, yields no loss of generality for all practical purposes. It is the case
for HTTP where the majority of implementations do assume a bound on the size
of length fields (e.g. major web browsers all do).
We start with the family of languages Llen (n) where the length field is n
symbols long. It is easy to see that each language of this family is finite, hence
regular. Now we turn to the size of specifications for Llen (n). In terms of finite
state automata, all automata specifying Llen (n) grow exponentially in n. Let b > 1
be the base in which the length is encoded, then there are bn possible encodings
for the length. By the pigeonhole principle, having less than bn reachable states
after reading the first n symbols implies that two distinct length encodings end
up in the same state, making them indistinguishable for the automaton. Hence,
it cannot decide Llen (n). However, when Llen (n) is specified using context-free
grammars, we show that it admits a more compact description.
Theorem 6. Let Σ be fixed alphabet and n > 0, there exists a context-free
grammar Glen (n) of size O(n) such that L(Glen (n)) = Llen (n).
Proof. For simplicity of presentation we assume that length fields are encoded in
binary, i.e. b = 2 in the definition of Llen (n). The generalization to any b > 2 is
tedious but straightforward.
The grammar Glen (n) is defined by the alphabet Σ, variables {S} ∪ {Xi |
0 ≤ i ≤ n} ∪ {Fi | 0 ≤ i ≤ n − 1} and the following rules:
{S → X0 }
{Xn → ε}
{Xi → 0 Xi+1 | 0 ≤ i < n}
{Xi → 1 Xi+1 Fi | 0 ≤ i < n}
{Fj → Fj−1 Fj−1 | 1 ≤ j ≤ n − 1}
{F0 → c | c ∈ W }
It follows by construction that L(Glen (n)) = Llen (n). A closer look reveals
that, since the alphabet is fixed and therefore so is |Σ| ≥ |W |, the size of the
rules of each set is bounded and independent from n while there are 3n+2+|W |
rules so the size of Glen (n) is O(n).
t
u
Next, we show that 110abc ∈ Llen (3) is also contained in L(Glen (3)).
S ⇒ X0 ⇒ 1X1 F0 ⇒ 11X2 F1 F0 ⇒ 110X3 F1 F0
⇒ 110F1 F0 ⇒ 110F0 F0 F0 ⇒∗ 110abc
3.4
Chunked Messages
Closely related to length fields are chunked messages, a feature found in the HTTP
protocol. According to the standard, the header Transfer-Encoding: chunked
8
Pierre Ganty, Boris Köpf, and Pedro Valero
signals that the body of the message is divided into chunks, each of which has its
size defined by a variable size length field as shown below:
HTTP/1.1 200 OK\r\n
Transfer-Encoding: chunked\r\n\r\n
12\r\nThe file is \r\n
16\r\n3,400 bytes long\r\n
0\r\n\r\n
Relying on previous definitions we model chunked messages by defining the
+
+
def
def
languages L]chunk = L]len {]} and L]chunk (n) = Llen (n) {]} for unbounded
and fixed (given by n) length field size, respectively. We further assume ] ∈
/W
and Σ = B ∪W ∪{]} to recognize the end of each chunk and thus avoid ambiguity.
Next, we turn to the claims found in the literature [7] about the impossibility
of specifying chunked messages using CFLs. The proofs, which are slight variations
of proofs for length fields, can be found in Appendix A.
Theorem 7. L]chunk is not a finite intersection of CFLs.
Theorem 8. Let Σ be a fixed alphabet and n > 0. The language L]chunk (n) is
regular and can be specified by a context-free grammar of size O(n).
4
Formal Analysis of (In)equalities
Input validation sometimes requires comparing different parts of a message, e.g.,
to check that two subwords are identical or that the first one represents a number
or a date that is greater than the second one. For instance, an HTTP GET message
is valid only if the field last-byte-pos is greater than first-byte-pos.
4.1
Equality Check
Consider the case of HTTP when a client is asking for a transition to some
other protocol. As the standard of mandates, equality should hold between the
Upgrade fields of the request and its response.
======== REQUEST ========
GET /example HTTP/1.1\r\n
Upgrade: h2c\r\n
======== RESPONSE ========
HTTP/1.1 101 Switching Protocols\r\n
Connection: Upgrade\r\n
Upgrade: h2c\r\n
Modeling Equality Check. We begin our study of comparisons with the case
of two contiguous subwords compared for equality. To this end consider the
following language over alphabet Σ given by
def
L]= = {x ] y | x = y} .
This language consists of twice the same word with ‘]’ in between. Again, we
assume ] occurs in x for no x.
A Language-theoretic View on Network Protocols
9
When the size of the words x and y is fixed, the delimiter is no longer needed.
Thus, we define L= (n)
def
L= (n) = {x y | x, y ∈ Σ n ∧ x = y} .
Unbounded Size. We now consider the case where the length of the subwords
to compare is unbounded. The example at the top of the section requires, when
validating a request-response pair of HTTP messages, to check equality across
Upgrade fields.
This situation is described by the language L]= . Next, we recall results by
Liu and Weiner [17] and Brough [5] enabling us to show that L]= is not a finite
intersection of context-free languages.
Proposition 9 ([5, Prop 2.1]). For every k > 0, the set of languages that
are an intersection of k CFLs is closed under (i) inverse GSM mappings, and
(ii) union with context-free languages.
Theorem 10 ([17, Thm 8]). Let a1 , . . . ak be k > 0 distinct symbols. Then
def
L(k) = {ai11 ai22 . . . aikk ai11 ai22 . . . aikk | ij ≥ 0 for all j} is an intersection of `
context-free languages for no ` < k.
We are now in position to prove our impossibility result about L]= .
Theorem 11. L]= is not a finite intersection of CFLs.
Proof (Sketch.). For the proof sketch we deliberately ignore the delimiter. Details
about how to deal with it can be found in Appendix A.
Assume L= (the delimiterless version of L]= ) is an intersection of m CFLs.
Now observe that L(k) = L= ∩ a∗1 a∗2 . . . a∗k a∗1 a∗2 . . . a∗k This implies L(k) is an
intersection of m + 1 context-free languages, which contradicts Theorem 10 for
k > m + 1.
t
u
Fixed Size. Because of the negative result of Theorem 11 we turn back again
to the restriction assuming an upper bound on the length of the subwords to
compare. We argue next that, in practice, such a restriction is reasonable.
Consider the following HTTP message.
HTTP/1.1 200 OK\r\n
Date: Sat, 25 Aug 2012 23:34:45 GMT\r\n
Warning: 112 - "Net down" "Sat, 25 Aug 2012 23:34:45 GMT"\r\n
The RFC mandates that the date in the Warning header be equal to Date.
Since date formats have bounded length we immediately have an upper bound of
the length of the subwords to compare.
Another example is given by the MIME protocol which allows to split messages
into multiple parts provided they are flanked by a user-defined delimiter string.
Let us consider an example:
10
Pierre Ganty, Boris Köpf, and Pedro Valero
MIME-Version: 1.0\r\n
Content-type: multipart/mixed; boundary="Mydelimiter"\r\n\r\n
PREAMBLE to be ignored\r\n--Mydelimiter\r\n
Plain ASCII text.\r\n--Mydelimiter\r\n
Plain ASCII text.\r\n--Mydelimiter--\r\n
EPILOGUE to be ignored.\r\n
Observe that the delimiter is first declared, boundary="Mydelimiter", and
then Mydelimiter is used three times, the first two times as --Mydelimiter
the last time as --Mydelimiter--.
Equality checks can ensure each part is flanked with the same delimiter. In
the case of MIME, the standard [12] imposes a maximum length of 69 symbols
for the delimiter giving us an upper bound.
Equality checks for a fixed number n of symbols are specified by L= (n). For
every n, the language L= (n) is finite, hence regular. Nonetheless Theorem 12,
due to Filmus [10], states that this language has no “compact” specification as
a grammar.3 Still it can be represented “compactly” as a finite intersection of
CFLs, as shown by Theorem 13. In this section we will study the size of different
grammars assuming the alphabet Σ is fixed and, thus, |Σ| is a constant.
Theorem 12 ([10, Thm 7]). Let |Σ| > 2, every context-free grammar for
L= (n) has size
!
n/4
|Σ|
√
Ω
.
2n
Recall that f (n) = Ω(g(n)) means that f is bounded from below4 by g for
sufficiently large n, which implies that context-free grammars for L= (n) exhibit
exponential growth in n.
Our next theorem based on the observation that x = y iff (x)i = (y)i for all i
allows to capture L= (n) as a intersection of n CFLs.
Theorem 13. Let the alphabet Σ be fixed, the language L= (n) is an intersection
of n CFLs, each of which is specified by a grammar of size O(n).
Proof. Given i ∈ {1, . . . , n}, define the language L=i (n) over the alphabet Σ
given by
def
L=i (n) = {x y | x, y ∈ Σ n , (x)i = (y)i } .
Clearly, for every word u we have u ∈ L= (n) iff u ∈ L=i (n) for all i ∈
{1, . . . , n}. Next, define G=i as the grammar for L=i (n) with variables S and T ,
alphabet Σ and the rules:{S → T i−1 c T n−1 c T n−i | c ∈ Σ}, {T → c | c ∈ Σ}.
It is routine to check that the size of the grammar is O(n).
t
u
Above, we studied specification of equality checks for two contiguous subwords.
In practice, however, comparisons are often more general. In the previous HTTP
3
4
This implies it has no “compact” specification by a finite state automaton either.
f (n) = Ω(g(n)) iff ∃ positive c, n0 s.t. ∀n > n0 , f (n) ≥ c · g(n)
A Language-theoretic View on Network Protocols
11
example, the dates to compare for equality are not necessarily contiguous. Also,
to specify the split messages of MIME using equality checks we need to generalize
to the cases where equality covers more than two subwords (each of the multiple
parts is flanked with the same delimiter) and those are not necessarily contiguous
(some parts are non empty).
We show this generalization of equality checks can still be specified concisely
by a finite intersection of CFLs. For the sake of space, results are deferred to
Appendix B.
4.2
Inequality Checks
Thus far, we have focused on languages whose words consist of two equal subwords.
However, comparisons sometimes require that the first subword represents a lower
number than the second or an earlier date. The following request is asking for
bytes of BigBuckBunny.mp4 between offsets 2833 and 7026. To be valid the
requested range should describe a non empty set.
GET /BigBuckBunny.mp4\r\n
Range: bytes=2833-7026\r\n
Modeling Inequality Checks. Let define a total order on Σ. We extend
to Σ ∗ and denote it ∗ as follows. We first define ∗ when its arguments have
equal length, then we proceed with the general case.
Given x, y ∈ Σ ∗ of equal length, let p be the least position such that (x)p 6=
(y)p . Then (x)p (y)p iff x ∗ y. Otherwise (no such position p exists) we also
have x ∗ y since the two words are equal.
Let us now proceed to the case where x and y have different length and
assume x is the shortest of the two words (the other case is treated similarly).
Then we have x ∗ y iff x0 ∗ y where x0 = min (Σ)|y|−|x| x, that is x0 is the
result of padding x with the minimal element of Σ so that the resulting word
and y have equal length. For instance, 5 ∗ 21 because 05 ∗ 21 where Σ is the
set of all digits and is defined as expected. It is an easy exercise to check that
∗ is a total order (hint: is a total order).
Unbounded Size. Let us turn back to the Range field example at the top of
the section. To specify the language of valid ranges, since the two subwords are
unbounded, a delimiter is needed to indicate the end of the first word. In our
example the delimiter is the dash symbol.
Next we define L] , the language deciding unbounded size inequality check
using ] as a delimiter, as follows:
L] = {x ] y | x, y ∈ Σ ∗ , x ∗ y} .
def
Theorem 14. The language L] is not a finite intersection of CFLs.
12
Pierre Ganty, Boris Köpf, and Pedro Valero
Proof. We begin by defining the order over Σ as −1 and define ∗ by replacing
with in the definition of ∗ . Clearly a b iff b a holds, hence there exists
a permutation γ on Σ. Indeed, we can write Σ as the set {a1 , . . . , an } such that
ai aj iff i ≤ j. Now define γ : ai 7→ an+1−i . It follows that a b iff γ(a) γ(b).
The previous equivalence naturally lifts to words (∗ and ∗ ), e.g. v ∗ w iff
γ(v) ∗ γ(w).
Next define
def
L] = {x ] y | x, y ∈ Σ ∗ , x ∗ y} .
Notice that the following equality holds: L] = {γ(x) ] γ(y) | x ] y ∈ L] }. Stated
equivalently, γ(L] ) = L] where γ is lifted to be a language homomorphism and
also L] = γ −1 (L] ) since γ is a bijection.
Following Proposition 9 (i) finite intersections of CFLs are closed under
inverse GSM mapping. This implies that they are also closed under inverse
homomorphism such as γ −1 .
Assume L] is a finite intersection of CFLs. It follows from above that so is
L] . Finally, consider the equivalence v = w iff v ∗ w and v ∗ w. Lifted to the
languages the previous equivalence becomes: L]= = L] ∩ L] .
Since both L] and L] are finite intersection of CFLs we conclude that so is
]
L= which contradicts Theorem 11.
t
u
Fixed Size. With the same motives as for equality checks we turn to the case
in which the size of the words to be compared is fixed, say n. As opposed to the
unbounded case, we can discard the delimiter because n – the last position of the
first word – is known. The message below illustrates an inequality check between
fixed size subwords.
HTTP/1.1 304 Not Modified\r\n
Date: Tue, 29 Mar 2016 09:05:57 GMT\r\n
Last-Modified: Wed, 24 Feb 2016 15:23:38 GMT\r\n
To ensure that this response is valid the Last-Modified field must contain
a date earlier than the Date field.
Let n > 0, we define L (n) to be:
L (n) = {x y | x, y ∈ Σ n , x ∗ y} .
def
Theorem 15. Let the alphabet Σ be fixed and n > 0, L (n) is a boolean combination of 2n languages each one specified by a grammar of size O(n).
Proof. Let G=i be the grammars used in the proof of Theorem 13 and let Gi
def
a grammar for the language Li (n) = {x y | x, y ∈ Σ n , (x)i (y)i }. Then, by
n
definition of the order over Σ , we write
w ∈ L (n) ⇔ w ∈ L=1..n (n)
n
_
i=1
w ∈ L=1..i−1 (n) ∧ w ∈
/ L=i (n) ∧ w ∈ Li (n)
A Language-theoretic View on Network Protocols
13
Ti
where w ∈ L=1..i (n) is equivalent to w ∈ j=1 L=j (n).
The size of each grammar G=i was shown to be O(n). On the other hand,
each grammar Gi is defined by the alphabet Σ, variables S, T and {Ta | a ∈ Σ}
and the rules:
{S → T i−1 a T n−1 Ta T n−i | a ∈ Σ}
{Ta → c | c ∈ Σ, a c}
It is routine to check that the size of the grammar is O(n).
{T → c | c ∈ Σ} .
t
u
The language L (n) can be extended to describe the situation in which x
and y represent dates and means “earlier than”. To this end, whenever the
month is given by its name instead of the number thereof we should read it as a
single symbol, considering each one as an element of the alphabet. Otherwise, a
comparison between numbers as described in proof of Theorem 15 will work. Once
we know how to compare the years, months, and days of two dates, combining
them to construct the language comparing two dates is straightforward.
5
Practical Evaluation
The results given in Sections 3 and 4 characterize the extent to which (intersections
of) CFL can be used to specify common idioms of network protocols. In this
section, we demonstrate that the positive theoretical results can be turned into
practical input validators for real-world network protocols. We begin by discussing
practical encoding issues, before we present an input validator for HTTP.
5.1
Encoding Real-World Protocols as CFG
Encoding effort. The manual effort of translating protocol specification into
grammars is facilitated by the RFC format: Protocol RFCs usually consist of
a grammar accompanied by a list of additional constraints written in English.
This grammar is typically given in ABNF format [6] which easily translates to a
context-free grammar. The additional constraints translate to regular expressions
or CFGs, along the lines described in this paper. Then the set of valid messages
of the protocol is described by a boolean combination of small CFLs.
Encoding size. The grammars required to perform the validation against the
idioms discussed in this paper remain small even for real-world protocols:
Length Fields. The CFG for Llen (n) consists of 3n + 2 rules, i.e. it grows linearly
in the size of the length field. This implies that it grows only logarithmically
with the size of the message body, which makes the CFG encoding practical for
real-world scenarios.
Comparisons. To compare two strings of length n we need 2n grammars each
with no more than 3|Σ| rules where |Σ| is the size of the alphabet. In practice,
n is small because it is the length of the encoding of a position within a file, a
timestamp, a hash value,. . .
14
Pierre Ganty, Boris Köpf, and Pedro Valero
5.2
An Input Validator for HTTP
Next we report on HTTPValidator [24], a proof of concept implementation to
validate HTTP messages based on mere CFGs and regular expressions, without
using attributes nor semantic actions.
Why HTTP? First, HTTP contains almost all of the features that have been
used in the literature [4, 21] to dispute the suitability of CFLs for parsing network
protocols. Second, HTTP is a widely used and complex protocol, making it an
ideal testbed for our approach. Finally, HTTPolice [8] is a lint for HTTP messages
which checks them for conformance to standards and best practices and provides
a reference for comparison.
HTTP as CFG. The ABNF described by the standard [9] is translated into a
single CFG while constraints such as “A client MUST send a Host header field in
all HTTP/1.1 request messages.” and “A client MUST NOT send the chunked
transfer coding name in TE” are translated into regular expressions and CFGs.
Implementation. Regular expressions and grammars are compiled with Flex and
Bison respectively. We avoid conflicts altogether by relying on the %glr-parser
declaration, which forces Bison to produce a generalized LR parser 5 that copes
with unresolved conflicts without altering the specified language. Finally, a script
runs all these validators sequentially and combines their boolean outputs to
conclude the validation. Table 1 describes the sizes of each separate element of
our validator. Further details can be found in the repository [24].
Table 1. Sizes of the formal languages required to validate an HTTP message
Feature
Size
HTTP ABNF as a CFG
1013 grammar rules
Decimal length field of size up to 80 871 grammar rules
Comparison of version numbers
3 grammars with 13 grammar rules each
Constraints (91 different ones)
260 regular expressions
Evaluation. We evaluate HTTPValidator on messages obtained from real-world
traffic (using Wireshark) and on messages provided with HTTPolice as test cases.
In total we thus obtain 239 test cases of which HTTPolice classifies 116 as valid
and 123 as invalid HTTP. We run HTTPValidator on these test cases obtaining
the same classification as HTTPolice but for two false positives. These errors
are due to well-formedness checks on message bodies in JSON and XML format,
which we currently do not consider in HTTPValidator but HTTPolice does.
5
https://www.gnu.org/software/bison/manual/html_node/
GLR-Parsers.html
A Language-theoretic View on Network Protocols
15
The time required for evaluating all test cases 6 is 16.1s for HTTPValidator
and 60s for HTTPolice, i.e. we achieve a 4-fold speedup. Note that this comparison
is slightly biased towards HTTPValidator because HTTPolice relies on interpreted
Python code whereas the parsers in HTTValidator are compiled to native code.
Moreover, we store each of the test cases in a single file, forgoing HTTPolice’s
ability to process several HTTP messages in a single file. On the other hand, we
have put ease of implementation before performance so no parallelization has
been implemented so far.
Overall, the experimental evaluation shows that, on our testbed, HTTPValidator achieves coverage and performance that is competitive with the state-of-the-art
in the field, thereby demonstrating the practicality of our approach.
6
Related Work
We discussed related work on language theory and input validation in the paper
body. Here we focus on discussing recent efforts for building parser generators
for network protocols.
In recent years, a number of parser generators for network protocols have
emerged. They are often parts of larger projects, but can be used in a standalone fashion. Important representatives are BinPac [21], which is part of the
Bro Network Security Monitor 7 , UltraPac [16], which is part of the NetShield
Monitor, Gapa [4], FlowSifter [18], and Nail [2]. The difference to our approach
is that they are all are built from scratch, whereas we rely on established CFG
parsing technology. Moreover, they rely on user-provided code for parsing idioms
such as length fields, whereas we specify everything in terms of (intersections
of) CFG. However, we emphasize that the focus of our approach lies on the
task of input validation, whereas those approaches deal with parsing, i.e. they
additionally fill a data structure.
Among the previous parser generators, Gapa and Nail stand out in terms
of their safety features. Gapa achieves a degree of safety by generating parsers
in a memory-safe language. Note that this does not prevent runtime error, e.g.,
dividing by zero still remains possible. Nail also aims at safety by providing some
automated support for filling user-defined data structure therefore reducing the
risk of errors introduced by user-defined code. In contrast, we do not rely on any
user-provided code.
Another line of work [7] relies on the use of the so-called attribute grammars,
an extension of context-free grammar that equips rules with attributes that can
be accessed and manipulated. For the parser generator, the authors use Bison and
encode the attribute aspect of grammars through user-defined C code annotating
the grammar rules which, as we argued before, augments the risk of errors.
6
7
We run our experiments on an Intel Core i5-5200U CPU 2.20GHz with 8GB RAM.
https://www.bro.org/
16
7
Pierre Ganty, Boris Köpf, and Pedro Valero
Conclusions and Future Work
Input validation is an important step for defending against malformed or malicious
inputs. In this paper we perform the first rigorous, language theoretic study of the
expressiveness required for validating a number of common protocol idioms. We
further show that input validation based on formal languages is practical and build
a modular input validator for HTTP from dependable software components such
as off-the-shelf parser generators for context-free languages. Our experimental
result shows that our approach is competitive with the state-of-the-art input
validator for HTTP in terms of coverage and speed.
There are some promising avenues for extending our work. For instance our
approach can be generalized to boolean closures of CFLs, which are known to be
strictly more expressive than the finite intersection we deal with in this paper [5].
Besides, our approach can be extended with a notion of state that is shared
between protocol participants which will allow us to implement, e.g., stateful
firewalls using our approach.
Acknowledgments We thank Juan Caballero for feedback on an early version
of this paper. This work was supported by Ramón y Cajal grant RYC-201416766, Spanish projects TIN2015-70713-R DEDETIS, TIN2015-71819-P RISCO,
TIN2012-39391-C04-01 StrongSoft, and Madrid regional project S2013/ICE-2731
N-GREENS.
References
[1] Internet Protocol. RFC 791 (Proposed Standard), Sept. 1981.
[2] J. Bangert and N. Zeldovich. Nail: A practical tool for parsing and generating
data formats. In 11th USENIX Symposium on Operating Systems Design and
Implementation, 2014.
[3] J.-P. Bernardy and P. Jansson. Certified context-free parsing: A formalisation of
valiant’s algorithm in agda. Logical Methods in Computer Science, 12(2), 2016.
[4] N. Borisov, D. Brumley, H. J. Wang, J. Dunagan, P. Joshi, and C. Guo. Generic
application-level protocol analyzer and its language. In NDSS 2007, 2007.
[5] T. Brough. Groups with poly-context-free word problem. Groups Complexity
Cryptology, 6(1), Jan 2014.
[6] P. D. Crocker, Ed. Brandenburg InternetWorking. Augmented BNF for Syntax
Specifications: ABNF. RFC 5234 (Proposed Standard), Jan. 2008.
[7] D. Davidson, R. Smith, N. Doyle, and S. Jha. Protocol normalization using
attribute grammars. In ESORICS 2009, LNCS, 2009.
[8] V. Faronov. HTTPolice. https://github.com/vfaronov/httpolice, 2017.
[9] R. Fielding and J. Reschke. Hypertext Transfer Protocol (HTTP/1.1): Message
Syntax and Routing. RFC 7230 (Proposed Standard), June 2014.
[10] Y. Filmus. Lower bounds for context-free grammars. Information Processing
Letters, 111(18), 2011.
[11] D. Firsov and T. Uustalu. Certified CYK parsing of context-free languages. Journal
of Logical and Algebraic Methods in Programming, 83(5-6), 2014.
[12] N. Freed and N. Borenstein. Multipurpose internet mail extensions (mime). RFC
1341 (Proposed Standard), Nov. 1996.
A Language-theoretic View on Network Protocols
17
[13] J. Graham-Cumming. Inside shellshock: How hackers are using it to exploit
systems. https://blog.cloudflare.com/inside-shellshock/.
https://blog.lastpass.com/2016/07/
[14] Lastpass security updates.
lastpass-security-updates.html/.
[15] M. Latteux. Intersections de langages algébriques bornés. Acta Informatica, 11(3),
1979.
[16] Z. Li, G. Xia, H. Gao, Y. Tang, Y. Chen, B. Liu, J. Jiang, and Y. Lv. NetShield:
massive semantics-based vulnerability signature matching for high-speed networks.
In ACM SIGCOMM 2010, 2010.
[17] L. Y. Liu and P. Weiner. An infinite hierarchy of intersections of context-free
languages. Mathematical systems theory, 7(2), 1973.
[18] C. R. Meiners, E. Norige, A. X. Liu, and E. Torng. Flowsifter: A counting automata
approach to layer 7 field extraction for deep flow inspection. In IEEE INFOCOM
2012, 2012.
[19] Microsoft releases security advisory 2953095. https://technet.microsoft.
com/library/security/2953095.
[20] P. Mockapetris. Domain Names - Implementation and Specification. RFC 1035
(Proposed Standard), Nov. 1987.
[21] R. Pang, V. Paxson, R. Sommer, and L. L. Peterson. binpac: a yacc for writing
application protocol parsers. In ACM SIGCOMM IMC 2006, 2006.
[22] J. Postel. User Datagram Protocol. RFC 768 (Proposed Standard), Aug. 1980.
[23] J. Rosenberg, H. Schulzrinne, G. Camarillo, A. Johnston, J. Peterson, R. Sparks,
M. Handley, and E. Schooler. SIP: Session Initiation Protocol. RFC 3261 (Proposed
Standard), June 2002.
[24] P. Valero. HTTPValidator. https://github.com/pevalme/HTTPValidator,
Apr 2017.
A
Deferred Proofs
A.1
Proof of Lemma 2
Proof (of Lemma 2). According to Proposition 3, if L∠ is a finite intersection of
CFLs then the set fw−1 (L∠ ) is semilinear, which implies that it is a finite union of
linear sets. Since the set is infinite, at least one of the linear sets must be infinite.
We call this set S.
Any element in S will have the form defined in 1 so define x, y ∈ S as follows:
bi x − 1
bi y − 1
x = ix , 1,
, y = iy , 1,
.
b−1
b−1
Let δ = iy − ix , we write
y=
ix + δ, 1,
bix +δ − 1
b−1
.
Without loss of generality, assume δ > 0 (for otherwise swap x and y). Next
define
bix (bδ − 1)
∆ = y − x = δ, 0,
.
b−1
18
Pierre Ganty, Boris Köpf, and Pedro Valero
Let z = x + 2∆:
bix − 1 + 2bix (bδ − 1)
bix (2bδ − 1) − 1
z = ix + 2δ, 1,
= ix + 2δ, 1,
. (2)
b−1
b−1
Since the set S is linear we find that z ∈ S so we can write:
biz − 1
z = iz , 1,
b−1
which means that iz = ix + 2δ obtaining:
bix +2δ − 1
z = ix + 2δ, 1,
.
b−1
(3)
Let us now derive a contradiction using (2) and (3). For this, we start by
deriving the following equivalence:
bix (2bδ − 1) = bix +2δ iff 2bδ − 1 = b2δ .
Applying the change of variable t = bδ we obtain
2t − 1 = t2 iff t2 − 2t + 1 = 0 .
Solving the equation give us t = 1 iff δ = 0.
So the vector z obtained basing on the linearity of the set S will belong to S
if and only if z = y = x.
We conclude that the set S is not infinite, in fact, if it is linear then it contains
only one element. Thus the set fw−1 (L∠ ) can not be written as a finite union of
linear set so the language L∠ is not a finite intersection of CFLs.
t
u
A.2
Proof of Lemma 5
Proof (of Lemma 5). This proof is conceptually identical to the previous one
but requires a slight modification in the notation used. Now, vectors in S have
7 components and the linear combination of the last five ones equals to the old
third component.
Thus, vector x ∈ S should be defined as:
(2)
(3)
(4)
(5)
x = ix , 1, n(1)
,
n
,
n
,
n
,
n
x
x
x
x
x
5
with
bi x − 1 X
=
|αk | × n(k)
.
x
b−1
k=1
Having this definition, the proof follows that of Lemma 2, considering the sum
of the last five components rather than their concrete values. Assuming Li is a
finite intersection of context-free languages, there exist three vectors x, y, z ∈ S
such that z = x + 2∆ where ∆ = y − x. For the sum of the last five components
of vectors z and x + 2∆ to be equal we require x, y and z to have the same first
component (δ = 0).
A Language-theoretic View on Network Protocols
19
Sharing the first component implies that each linear set, S, relates (by fw−1 )
only with vectors with the same first component. Therefore assuming fw−1 (Li )
is as a finite union of linear sets requires Li to only contain words with a finite
number of different lengths, which contradicts the fact that it is infinite. By
Proposition 3 we conclude that Li is not a finite intersection of CFLs.
t
u
A.3
Proof of Theorem 7
Proof (of Theorem 7). Let a ∈ Σ be such that a 6= 1 and a =
6 ]. One can easily
see that
def
L∠ {]} = L]chunk ∩ 1∗ ]∗ a∗ ]∗ .
Assuming L]chunk is a finite intersection of CFLs then we find that so is L∠ {]}.
Let w̄ = h1, ], a, ]i, then
o
n
i
−1
,1 | i ∈ N .
fw̄−1 (L∠ {]}) = i, 1, bb−1
From there, it is straightforward to extend the result of Lemma 2 to show that
L∠ {]} is not a finite intersection of CFLs, hence derive a contradiction.
t
u
A.4
Proof of Theorem 8
Proof (of Theorem 8). Recall the grammar Glen (n) for the language Llen (n).
The grammar Gchunk (n) is defined by adding to Glen (n) a fresh start variable
Z and the following two rules: {Z → S ] Z, Z → S ]}. Clearly, L(Gchunk (n)) =
L]chunk (n) and the size of Gchunk (n) is O(n) since the size of Glen (n) is O(n) due
to Theorem 6
t
u
A.5
Proof of Theorem 11
Proof (of Theorem 11). We first set up languages and a GSM mapping to take
proper care of the delimiter. Then we proceed with the proof of theorem. Let
L](k) be the language obtained by adding the delimiter to L(k) , that is
L](k) = {ai11 ai22 . . . aikk ]ai11 ai22 . . . aikk | ij ≥ 0 for all j} .
Now given k > 0 define8
Laux = L](k) ∩ {aji ]z | 1 ≤ i ≤ k, j ≥ 0, z ∈ Σ ∗ } .
(4)
Intuitively, Laux is obtained from L](k) by filtering out the words where at most
one symbol occurs to the left of the delimiter, e.g. a3 a3 #a1 a2 a3 is such a word
8
Given L over Σ, let L denotes the complement of L, that is Σ ∗ \ L.
20
Pierre Ganty, Boris Köpf, and Pedro Valero
whereas a1 a2 # is not. Observe that the set {aji ]z | 1 ≤ i ≤ k, j ≥ 0, z ∈ Σ ∗ } is
regular, and so is its complement.
Next, we define the GSM mapping π which intuitevely inserts the delimiter
at the “right” place. Unless stated otherwise, when describing π, we assume the
GSM outputs exactly what it reads.
The GSM reads the first symbol, say ai , and moves to a state qi keeping track
of that symbol. Notice that the first symbol need not be the delimiter following
equation (4). Then π keeps reading the same ai symbol without leaving qi . Upon
reading some symbol x =
6 ai , π moves to a different state wherein it keeps reading
symbols x =
6 ai until it reads ai again. At that point, π has reached the middle
of the word and so it outputs ] (besides ai ) and keeps processing the remainder
of the word.
From above, we find that π −1 (Laux ) coincides with L(k) minus the words
where at most one symbol occurs, i.e. {(ai ai )j | 1 ≤ i ≤ k, j ≥ 0}.
Now we reinstate these words and obtain the following equality:
L(k) = π −1 (Laux ) ∪ {(ai ai )j | 1 ≤ i ≤ k, j ≥ 0} .
We are in position to prove the statement of the theorem.
It is easy to see that the following equality holds:
L](k) = L]= ∩ a∗1 a∗2 . . . a∗k ]a∗1 a∗2 . . . a∗k
Now assume L]= is a finite intersection of m context-free languages. Then Laux
is a finite intersection of m + 2 CFLs by equation 4. Proposition 9 (i) shows that
π −1 (Laux ) is a finite intersection of m + 2 CFLs, hence we find that L(k) is an
intersection of m + 2 CFLs by Proposition 9 (ii). Notice that this holds for all
values of k and in particular for k > m+2 which contradicts Theorem 10.
t
u
B
General equality checks
To specify general equality checks we consider the following language over Σ
where n is the upper bound on the size of word to check equality for:
^
def
Ld (n) = {w x w y1 . . . w yk w z | |w| = n ∧ x ∈ Lfm ∧
yj ∈ Lmm ∧ z ∈ Llm }
j
where Lfm , Llm and Lmm are CFLs.
It is routine to check that Ld (n) is context-free for every value of n. However,
Theorem 12 shows there is no concise specification for it. Therefore, we study
the size of specifications given as a finite intersection of CFLs.
Theorem 16. Assuming a fixed size grammar specification for the languages
Lfm , Lmm and Llm , the language Ld (n) over fixed size alphabet Σ is a finite
intersection of n CFLs, each one specified by a grammar of size O(n).
A Language-theoretic View on Network Protocols
21
Proof. Given i ∈ {1, . . . , n}, define the language Ld=i (n) over alphabet Σ given
by:
{w0 x w1 y1 . . . wk yk wk+1 z |
^
(|w` | = n ∧ (w` )i = (w`+1 )i )
`
∧ x ∈ Lfm ∧
^
yj ∈ Lmm ∧ z ∈ Llm } .
j
It is easily seen that give a word u we have u ∈ Ld (n) iff u ∈ Ld=i (n) for all
i ∈ {1, . . . , n}.
Let Sfm , Smm and Slm be the start symbols of the grammars defining languages Lfm , Lmm and Llm respectively. Next, define Gd=i (n) as the grammar for
Ld=i (n) with alphabet Σ, set of variables {S, T, P, Q, Sfm , Smm , Slm } and rules:
{S → P Sfm Q P Slm }
{P → T
i−1
cT
n−i
| c ∈ Σ}
{T → c | c ∈ Σ}
{Q → P Smm Q}
{Q → P Smm } .
Gd=i (n) also contains all variables and rules of the grammars specifying
languages Lfm , Lmm and Llm . Because the size of the grammars defining the
languages Lfm , Lmm and Llm is fixed and independent of n, it is routine to check
that the size of Gd=i (n) is O(n).
t
u
| 6cs.PL
|
Difference Target Propagation
Dong-Hyun Lee1 , Saizheng Zhang1 , Asja Fischer1 , and
Yoshua Bengio1,2
arXiv:1412.7525v5 [cs.LG] 25 Nov 2015
1
Université de Montréal, Quebec, Canada
2
CIFAR Senior Fellow
Abstract. Back-propagation has been the workhorse of recent successes of deep
learning but it relies on infinitesimal effects (partial derivatives) in order to perform credit assignment. This could become a serious issue as one considers
deeper and more non-linear functions, e.g., consider the extreme case of nonlinearity where the relation between parameters and cost is actually discrete. Inspired by the biological implausibility of back-propagation, a few approaches
have been proposed in the past that could play a similar credit assignment role. In
this spirit, we explore a novel approach to credit assignment in deep networks that
we call target propagation. The main idea is to compute targets rather than gradients, at each layer. Like gradients, they are propagated backwards. In a way that
is related but different from previously proposed proxies for back-propagation
which rely on a backwards network with symmetric weights, target propagation
relies on auto-encoders at each layer. Unlike back-propagation, it can be applied
even when units exchange stochastic bits rather than real numbers. We show that
a linear correction for the imperfectness of the auto-encoders, called difference
target propagation, is very effective to make target propagation actually work,
leading to results comparable to back-propagation for deep networks with discrete and continuous units and denoising auto-encoders and achieving state of
the art for stochastic networks. 3
1
Introduction
Recently, deep neural networks have achieved great success in hard AI tasks [2, 11,
13, 18], mostly relying on back-propagation as the main way of performing credit assignment over the different sets of parameters associated with each layer of a deep net.
Back-propagation exploits the chain rule of derivatives in order to convert a loss gradient on the activations over layer l (or time t, for recurrent nets) into a loss gradient
on the activations over layer l − 1 (respectively, time t − 1). However, as we consider
deeper networks– e.g., consider the recent best ImageNet competition entrants [19]
with 19 or 22 layers – longer-term dependencies, or stronger non-linearities, the composition of many non-linear operations becomes more strongly non-linear. To make this
concrete, consider the composition of many hyperbolic tangent units. In general, this
3
This paper was accepted in ECML/PKDD 2015. Please cite like following : Dong-Hyun Lee,
Saizheng Zhang, Asja Fischer and Yoshua Bengio, Difference Target Propagation, in Machine
Learning and Knowledge Discovery in Databases, pages 498-515, Springer International Publishing, 2015
2
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
means that derivatives obtained by back-propagation are becoming either very small
(most of the time) or very large (in a few places). In the extreme (very deep computations), one would get discrete functions, whose derivatives are 0 almost everywhere,
and infinite where the function changes discretely. Clearly, back-propagation would fail
in that regime. In addition, from the point of view of low-energy hardware implementation, the ability to train deep networks whose units only communicate via bits would
also be interesting.
This limitation of back-propagation to working with precise derivatives and smooth
networks is the main machine learning motivation for this paper’s exploration into an
alternative principle for credit assignment in deep networks. Another motivation arises
from the lack of biological plausibility of back-propagation, for the following reasons:
(1) the back-propagation computation is purely linear, whereas biological neurons interleave linear and non-linear operations, (2) if the feedback paths were used to propagate credit assignment by back-propagation, they would need precise knowledge of
the derivatives of the non-linearities at the operating point used in the corresponding
feedforward computation, (3) similarly, these feedback paths would have to use exact
symmetric weights (with the same connectivity, transposed) of the feedforward connections, (4) real neurons communicate by (possibly stochastic) binary values (spikes), (5)
the computation would have to be precisely clocked to alternate between feedforward
and back-propagation phases, and (6) it is not clear where the output targets would come
from.
The main idea of target propagation is to associate with each feedforward unit’s
activation value a target value rather than a loss gradient. The target value is meant to
be close to the activation value while being likely to have provided a smaller loss (if that
value had been obtained in the feedforward phase). In the limit where the target is very
close to the feedforward value, target propagation should behave like back-propagation.
This link was nicely made in [15, 16], which introduced the idea of target propagation
and connected it to back-propagation via a Lagrange multipliers formulation (where the
constraints require the output of one layer to equal the input of the next layer). A similar
idea was recently proposed where the constraints are relaxed into penalties, yielding a
different (iterative) way to optimize deep networks [9]. Once a good target is computed,
a layer-local training criterion can be defined to update each layer separately, e.g., via
the delta-rule (gradient descent update with respect to the cross-entropy loss).
By its nature, target propagation can in principle handle stronger (and even discrete)
non-linearities, and it deals with biological plausibility issues (1), (2), (3) and (4) described above. Extensions of the precise scheme proposed here could handle (5) and (6)
as well, but this is left for future work.
In this paper, we describe how the general idea of target propagation by using
auto-encoders to assign targets to each layer (as introduced in an earlier technical report [4]) can be employed for supervised training of deep neural networks (section
2.1 and 2.2). We continue by introducing a linear correction for the imperfectness of
the auto-encoders (2.3) leading to robust training in practice. Furthermore, we show
how the same principles can be applied to replace back-propagation in the training
of auto-encoders (section 2.4). In section 3 we provide several experimental results
on rather deep neural networks as well as discrete and stochastic networks and auto-
Difference Target Propagation
3
encoders. The results show that the proposed form of target propagation is comparable
to back-propagation with RMSprop [21] - a very popular setting to train deep networks
nowadays- and achieves state of the art for training stochastic neural nets on MNIST.
2
Target Propagation
Although many variants of the general principle of target propagation can be devised,
this paper focuses on a specific approach, which is based on the ideas presented in an
earlier technical report [4] and is described in the following.
2.1
Formulating Targets
Let us consider an ordinary (supervised) deep network learning process, where the training data is drawn from an unknown data distribution p(x, y). The network structure is
defined by
hi = fi (hi−1 ) = si (Wi hi−1 ), i = 1, . . . , M
(1)
where hi is the state of the i-th hidden layer (where hM corresponds to the output of
the network and h0 = x) and fi is the i-th layer feed-forward mapping, defined by a
non-linear activation function si (e.g. the hyperbolic tangents or the sigmoid function)
and the weights Wi of the i-th layer. Here, for simplicity of notation, the bias term of
the i-th layer is included in Wi . We refer to the subset of network parameters defining
i,j
the mapping between the i-th and the j-th layer (0 ≤ i < j ≤ M ) as θW
= {Wk , k =
i + 1, . . . , j}. Using this notion, we can write hj as a function of hi depending on
i,j
i,j
parameters θW
, that is we can write hj = hj (hi ; θW
).
0,M
Given a sample (x, y), let L(hM (x; θW ), y) be an arbitrary global loss function
0,M
) for the target y,
measuring the appropriateness of the network output hM (x; θW
e.g. the MSE or cross-entropy for binomial random variables. Then, the training ob0,M
jective corresponds to adapting the network parameters θW
so as to minimize the
0,M
expected global loss Ep {L(hM (x; θW )y)} under the data distribution p(x, y). For
i = 1, . . . , M − 1 we can write
0,i
i,M
0,M
L(hM (x; θW
), y) = L(hM (hi (x; θW
); θW
), y)
(2)
to emphasize the dependency of the loss on the state of the i-th layer.
Training a network with back-propagation corresponds to propagating error signals
through the network to calculate the derivatives of the global loss with respect to the
parameters of each layer. Thus, the error signals indicate how the parameters of the network should be updated to decrease the expected loss. However, in very deep networks
with strong non-linearities, error propagation could become useless in lower layers due
to exploding or vanishing gradients, as explained above.
To avoid this problems, the basic idea of target propagation is to assign to each
0,i
hi (x; θW
) a nearby value ĥi which (hopefully) leads to a lower global loss, that is
which has the objective to fulfill
i,M
0,i
i,M
L(hM (ĥi ; θW
), y) < L(hM (hi (x; θW
); θW
), y) .
(3)
4
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
Such a ĥi is called a target for the i-th layer.
Given a target ĥi we now would like to change the network parameters to make
hi move a small step towards ĥi , since – if the path leading from hi to ĥi is smooth
enough – we would expect to yield a decrease of the global loss. To obtain an update
direction for Wi based on ĥi we can define a layer-local target loss Li , for example by
using the MSE
0,i 2
Li (ĥi , hi ) = ||ĥi − hi (x; θW
)||2 .
(4)
Then, Wi can be updated locally within its layer via stochastic gradient descent, where
ĥi is considered as a constant with respect to Wi . That is
(t+1)
Wi
(t)
= Wi
− ηfi
0,i
)
∂Li (ĥi , hi ) ∂hi (x; θW
∂Li (ĥi , hi )
(t)
= Wi − ηfi
,
∂Wi
∂hi
∂Wi
(5)
where ηfi is a layer-specific learning rate.
Note, that in this context, derivatives can be used without difficulty, because they
correspond to computations performed inside a single layer. Whereas, the problems
with the severe non-linearities observed for back-propagation arise when the chain rule
is applied through many layers. This motivates target propagation methods to serve as
alternative credit assignment in the context of a composition of many non-linearities.
However, it is not directly clear how to compute a target that guarantees a decrease
of the global loss (that is how to compute a ĥi for which equation (3) holds) or that at
least leads to a decrease of the local loss Li+1 of the next layer, that is
Li (ĥi+1 , fi (ĥi )) < Li (ĥi+1 , fi (hi )) .
(6)
Proposing and validating answers to this question is the subject of the rest of this paper.
2.2
How to assign a proper target to each layer
Clearly, in a supervised learning setting, the top layer target should be directly driven
from the gradient of the global loss
ĥM = hM − η̂
∂L(hM , y)
,
∂hM
(7)
where η̂ is usually a small step size. Note, that if we use the MSE as global loss and
η̂ = 0.5 we get ĥM = y.
But how can we define targets for the intermediate layers? In the previous technical
report [4], it was suggested to take advantage of an “approximate inverse”. To formalize
this idea, suppose that for each fi we have a function gi such that
fi (gi (hi )) ≈ hi
or
gi (fi (hi−1 )) ≈ hi−1 .
(8)
Then, choosing
ĥi−1 = gi (ĥi )
(9)
would have the consequence that (under some smoothness assumptions on f and g)
minimizing the distance between hi−1 and ĥi−1 should also minimize the loss Li of
Difference Target Propagation
5
the i-th layer. This idea is illustrated in the left of Figure 1. Indeed, if the feed-back
mappings were the perfect inverses of the feed-forward mappings (gi = fi−1 ), one gets
Li (ĥi , fi (ĥi−1 )) = Li (ĥi , fi (gi (ĥi ))) = Li (ĥi , ĥi ) = 0 .
(10)
But choosing g to be the perfect inverse of f may need heavy computation and instability, since there is no guarantee that fi−1 applied to a target would yield a value that
is in the domain of fi−1 . An alternative approach is to learn an approximate inverse
gi , making the fi / gi pair look like an auto-encoder. This suggests parametrizing gi as
follows:
gi (hi ) = s̄i (Vi hi ), i = 1, ..., M
(11)
where s̄i is a non-linearity associated with the decoder and Vi the matrix of feed-back
weights of the i-th layer. With such a parametrization, it is unlikely that the auto-encoder
will achieve zero reconstruction error. The decoder could be trained via an additional
auto-encoder-like loss at each layer
Linv
= ||gi (fi (hi−1 )) − hi−1 ||22 .
i
(12)
Changing Vi based on this loss, makes g closer to fi−1 . By doing so, it also makes
fi (ĥi−1 ) = fi (gi (ĥi )) closer to ĥi , and is thus also contributing to the decrease of
Li (ĥi , fi (ĥi−1 )). But we do not want to estimate an inverse mapping only for the concrete values we see in training but for a region around the these values to facilitate the
computation of gi (ĥi ) for ĥi which have never been seen before. For this reason, the
loss is modified by noise injection
Linv
= ||gi (fi (hi−1 + )) − (hi−1 + )||22 ,
i
∼ N (0, σ) ,
(13)
which makes fi and gi approximate inverses not just at hi−1 but also in its neighborhood.
As mentioned above, a required property of target propagation is, that the layerwise parameter updates, each improving a layer-wise loss, also lead to an improvement
of the global loss. The following theorem shows that, for the case that gi is a perfect
inverse of fi and fi having a certain structure, the update direction of target propagation
does not deviate more then 90 degrees from the gradient direction (estimated by backpropagation), which always leads to a decrease of the global loss.
Theorem 1. 4 Assume that gi = fi−1 , i = 1, ..., M , and fi satisfies hi = fi (hi−1 ) =
Wi si (hi−1 ) 5 where si can be any differentiable monotonically increasing elementwise function. Let δWitp and δWibp be the target propagation update and the backpropagation update in i-th layer, respectively. If η̂ in Equation (7) is sufficiently small,
then the angle α between δWitp and δWibp is bounded by
0<
4
5
1 + ∆1 (η̂)
≤ cos(α) ≤ 1
+ ∆2 (η̂)
λmax
λmin
See the proof in the Appendix.
This is another way to obtain a non-linear deep network structure.
(14)
6
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
Fig. 1. (left) How to compute a target in the lower layer via difference target propagation.
fi (ĥi−1 ) should be closer to ĥi than fi (hi−1 ). (right) Diagram of the back-propagation-free
auto-encoder via difference target propagation.
Here λmax and λmin are the largest and smallest singular values of (JfM . . . Jfi+1 )T ,
where Jfk is the Jacobian matrix of fk and ∆1 (η̂) and ∆2 (η̂) are close to 0 if η̂ is
sufficiently small.
2.3
Difference target propagation
From our experience, the imperfection of the inverse function leads to severe optimization problems when assigning targets based on equation (9). This brought us to propose
the following linearly corrected formula for target propagation which we refer to as
“difference target propagation”
ĥi−1 = hi−1 + gi (ĥi ) − gi (hi ) .
(15)
Note, that if gi is the inverse of fi , difference target propagation becomes equivalent
to vanilla target propagation as defined in equation (9). The resulting complete training
procedure for optimization by difference target propagation is given in Algorithm 1.
In the following, we explain why this linear corrected formula stabilizes the optimization process. In order to achieve stable optimization by target propagation, hi−1
should approach ĥi−1 as hi approaches ĥi . Otherwise, the parameters in lower layers
continue to be updated even when an optimum of the global loss is reached already
by the upper layers, which then could lead the global loss to increase again. Thus, the
condition
hi = ĥi ⇒ hi−1 = ĥi−1
(16)
greatly improves the stability of the optimization. This holds for vanilla target propagation if gi = fi−1 , because
hi−1 = fi−1 (hi ) = gi (ĥi ) = ĥi−1 .
(17)
Difference Target Propagation
7
Algorithm 1 Training deep neural networks via difference target propagation
Compute unit values for all layers:
for i = 1 to M do
hi ← fi (hi−1 )
end for
Making the first target: ĥM −1 ← hM −1 − η̂ ∂h∂L
, (L is the global loss)
M −1
Compute targets for lower layers:
for i = M − 1 to 2 do
ĥi−1 ← hi−1 − gi (hi ) + gi (ĥi )
end for
Training feedback (inverse) mapping:
for i = M − 1 to 2 do
Update parameters for gi using SGD with following a layer-local loss Linv
i
Linv
= ||gi (fi (hi−1 + )) − (hi−1 + )||22 , ∼ N (0, σ)
i
end for
Training feedforward mapping:
for i = 1 to M do
Update parameters for fi using SGD with following a layer-local loss Li
Li = ||fi (hi−1 ) − ĥi ||22 if i < M , Li = L (the global loss) if i = M .
end for
Although the condition is not guaranteed to hold for vanilla target propagation if gi 6=
fi−1 , for difference target propagation it holds by construction, since
ĥi−1 − hi−1 = gi (ĥi ) − gi (hi ) .
(18)
Furthermore, under weak conditions on f and g and if the difference between hi
and ĥi is small, we can show for difference target propagation that if the input of the
i-th layer becomes ĥi−1 (i.e. the i − 1-th layer reaches its target) the output of the i-th
layer also gets closer to ĥi . This means that the requirement on targets specified by
equation (6) is met for difference target propagation, as shown in the following theorem
Theorem 2. 6 Let the target for layer i − 1 be given by Equation (15), i.e. ĥi−1 =
hi−1 + gi (ĥi ) − gi (hi ). If ĥi − hi is sufficiently small, fi and gi are differentiable, and
the corresponding Jacobian matrices Jfi and Jgi satisfy that the largest eigenvalue of
(I − Jfi Jgi )T (I − Jfi Jgi ) is less than 1, then we have
||ĥi − fi (ĥi−1 )||22 < ||ĥi − hi ||22 .
(19)
The third condition in the above theorem is easily satisfied in practice, because gi is
learned to be the inverse of fi and makes gi ◦ fi close to the identity mapping, so that
(I − Jfi Jgi ) becomes close to the zero matrix which means that the largest eigenvalue
of (I − Jfi Jgi )T (I − Jfi Jgi ) is also close to 0.
6
See the proof in Appendix.
8
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
2.4
Training an auto-encoder with difference target propagation
Auto-encoders are interesting for learning representations and serve as building blocks
for deep neural networks [10]. In addition, as we have seen, training auto-encoders is
part of the target propagation approach presented here, where they model the feedback
paths used to propagate the targets.
In the following, we show how a regularized auto-encoder can be trained using difference target propagation instead of back-propagation. Like in the work on denoising
auto-encoders [22] and generative stochastic networks [6], we consider the denoising
auto-encoder like a stochastic network with noise injected in input and hidden units,
trained to minimize a reconstruction loss. This is, the hidden units are given by the
encoder as
h = f (x) = sig(W x + b) ,
(20)
where sig is the element-wise sigmoid function, W the weight matrix and b the bias
vector of the input units. The reconstruction is given by the decoder
z = g(h) = sig(W T (h + ) + c), ∼ N (0, σ) ,
(21)
with c being the bias vector of the hidden units. And the reconstruction loss is
L = ||z − x||22 + ||f (x + ) − h||22 , ∼ N (0, σ) ,
(22)
where a regularization term can be added to obtain a contractive mapping. In order to
train this network without back-propagation (that is, without using the chain rule), we
can use difference target propagation as follows (see Figure 1 (right) for an illustration):
at first, the target of z is just x, so we can train the reconstruction mapping g based on
the loss Lg = ||g(h) − x||22 in which h is considered as a constant. Then, we compute
the target ĥ of the hidden units following difference target propagation where we make
use of the fact that f is an approximate inverse of g. That is,
ĥ = h + f (ẑ) − f (z) = 2h − f (z) ,
(23)
where the last equality follows from f (ẑ) = f (x) = h. As a target loss for the hidden
layer, we can use Lf = ||f (x + ) − ĥ||22 , where ĥ is considered as a constant and
which can be also augmented by a regularization term to yield a contractive mapping.
3
Experiments
In a set of experiments we investigated target propagation for training deep feedforward
deterministic neural networks, networks with discrete transmissions between units,
stochastic neural networks, and auto-encoders.
For training supervised neural networks, we chose the target of the top hidden layer
(number M −1) such that it also depends directly on the global loss instead of an inverse
M ,y)
mapping. That is, we set ĥM −1 = hM −1 − η̃ ∂L(h
∂hM −1 , where L is the global loss (here
the multiclass cross entropy). This may be helpful when the number of units in the
Difference Target Propagation
9
output layer is much smaller than the number of units in the top hidden layer, which
would make the inverse mapping difficult to learn, but future work should validate that.
For discrete stochastic networks in which some form of noise (here Gaussian) is
injected, we used a decaying noise level for learning the inverse mapping, in order to
stabilize learning, i.e. the standard deviation of the Gaussian is set to σ(e) = σ0 /(1 +
e/e0 ) where σ0 is the initial value, e is the epoch number and e0 is the half-life of this
decay. This seems to help to fine-tune the feedback weights at the end of training.
In all experiments, the weights were initialized with orthogonal random matrices
and the bias parameters were initially set to zero. All experiments were repeated 10
times with different random initializations. We put the code of these experiments online
(https://github.com/donghyunlee/dtp).
3.1
Deterministic feedforward deep networks
As a primary objective, we investigated training of ordinary deep supervised networks
with continuous and deterministic units on the MNIST dataset. We used a held-out validation set of 10000 samples for choosing hyper-parameters. We trained networks with
7 hidden layers each consisting of 240 units (using the hyperbolic tangent as activation
function) with difference target propagation and back-propagation.
Training was based on RMSprop [21] where hyper-parameters for the best validation error were found using random search [7]. RMSprop is an adaptive learning rate
algorithm known to lead to good results for back-propagation. Furthermore, it is suitable
for updating the parameters of each layer based on the layer-wise targets obtained by
target propagation. Our experiments suggested that when using a hand-selected learning
rate per layer rather than the automatically set one (by RMSprop), the selected learning
rates were different for each layer, which is why we decided to use an adaptive method
like RMSprop.
Fig. 2. Mean training cost (left) and train/test classification error (right) with target propagation
and back-propagation using continuous deep networks (tanh) on MNIST. Error bars indicate the
standard deviation.
The results are shown in Figure 2. We obtained a test error of 1.94% with target
propagation and 1.86% with back propagation. The final negative log-likelihood on the
training set was 4.584 × 10−5 with target propagation and 1.797 × 10−5 with back
10
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
propagation. We also trained the same network with rectifier linear units and got a test
error of 3.15% whereas 1.62% was obtained with back-propagation. It is well known
that this nonlinearity is advantageous for back-propagation, while it seemed to be less
appropriate for this implementation of target propagation.
In a second experiment we investigated training on CIFAR-10. The experimental
setting was the same as for MNIST (using the hyperbolic tangent as activation function) except that the network architecture was 3072-1000-1000-1000-10. We did not
use any preprocessing, except for scaling the input values to lay in [0,1], and we tuned
the hyper-parameters of RMSprop using a held-out validation set of 1000 samples. We
obtained mean test accuracies of 50.71% and 53.72% for target propagation and backpropagation, respectively. It was reported in [14], that a network with 1 hidden layer of
1000 units achieved 49.78% accuracy with back-propagation, and increasing the number of units to 10000 led to 51.53% accuracy. As the current state-of-the-art performance on the permutation invariant CIFAR-10 recognition task, [12] reported 64.1%
but when using PCA without whitening as preprocessing and zero-biased auto-encoders
for unsupervised pre-training.
3.2
Networks with discretized transmission between units
To explore target propagation for an extremely non-linear neural network, we investigated training of discrete networks on the MNIST dataset. The network architecture
was 784-500-500-10, where only the 1st hidden layer was discretized. Inspired by biological considerations and the objective of reducing the communication cost between
neurons, instead of just using the step activation function, we used ordinary neural net
layers but with signals being discretized when transported between the first and second
layer. The network structure is depicted in the right plot of Figure 3 and the activations
of the hidden layers are given by
h1 = f1 (x) = tanh(W1 x) and
h2 = f2 (h1 ) = tanh(W2 sign(h1 ))
(24)
where sign(x) = 1 if x > 0, and sign(x) = 0 if x ≤ 0. The network output is given
by
p(y|x) = f3 (h2 ) = sof tmax(W3 h2 ) .
(25)
The inverse mapping of the second layer and the associated loss are given by
g2 (h2 ) = tanh(V2 sign(h2 )) ,
Linv
= ||g2 (f2 (h1 + )) − (h1 + )||22 ,
2
∼ N (0, σ) .
(26)
(27)
If feed-forward mapping is discrete, back-propagated gradients become 0 and useless
when they cross the discretization step. So we compare target propagation to two baselines. As a first baseline, we train the network with back-propagation and the straightthrough estimator [5], which is biased but was found to work well, and simply ignores
the derivative of the step function (which is 0 or infinite) in the back-propagation phase.
As a second baseline, we train only the upper layers by back-propagation, while not
changing the weight W1 which are affected by the discretization, i.e., the lower layers
do not learn.
Difference Target Propagation
11
The results on the training and test sets are shown in Figure 3. The training error
for the first baseline (straight-through estimator) does not converge to zero (which can
be explained by the biased gradient) but generalization performance is fairly good. The
second baseline (fixed lower layer) surprisingly reached zero training error, but did
not perform well on the test set. This can be explained by the fact that it cannot learn
any meaningful representation at the first layer. Target propagation however did not
suffer from this drawback and can be used to train discrete networks directly (training
signals can pass the discrete region successfully). Though the training convergence was
slower, the training error did approach zero. In addition, difference target propagation
also achieved good results on the test set.
Fig. 3. Mean training cost (top left), mean training error (top right) and mean test error (bottom
left) while training discrete networks with difference target propagation and the two baseline
versions of back-propagation. Error bars indicate standard deviations over the 10 runs. Diagram
of the discrete network (bottom right). The output of h1 is discretized because signals must be
communicated from h1 to h2 through a long cable, so binary representations are preferred in
order to conserve energy. With target propagation, training signals are also discretized through
this cable (since feedback paths are computed by bona-fide neurons).
3.3
Stochastic networks
Another interesting model class which vanilla back-propagation cannot deal with are
stochastic networks with discrete units. Recently, stochastic networks have attracted
attention [3, 20, 5] because they are able to learn a multi-modal conditional distribution
12
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
P (Y |X), which is important for structured output predictions. Training networks of
stochastic binary units is also biologically motivated, since they resemble networks
of spiking neurons. Here, we investigate whether one can train networks of stochastic
binary units on MNIST for classification using target propagation. Following [17], the
network architecture was 784-200-200-10 and the hidden units were stochastic binary
units with the probability of turning on given by a sigmoid activation:
hpi = P (Hi = 1|hi−1 ) = σ(Wi hi−1 ), hi ∼ P (Hi |hi−1 ) ,
(28)
hpi .
that is, hi is one with probability
As a baseline, we considered training based on the straight-through biased gradient
estimator [5] in which the derivative through the discrete sampling step is ignored (this
method showed the best performance in [17].) That is
δhpi−1 = δhpi
∂hpi
≈ σ 0 (Wi hi−1 )WiT δhpi .
∂hpi−1
(29)
With difference target propagation the stochastic network can be trained directly, setting
the targets to
ĥp2 = hp2 − η
∂L
∂h2
and
ĥp1 = hp1 + g2 (ĥp2 ) − g2 (hp2 )
(30)
where gi (hpi ) = tanh(Vi hpi ) is trained by the loss
Linv
= ||gi (fi (hi−1 + )) − (hi−1 + )||22 ,
i
∼ N (0, σ) ,
(31)
and layer-local target losses are defined as Li = ||ĥpi − hpi ||22 .
Method
Test Error(%)
Difference Target-Propagation, M=1
1.54%
Straight-through gradient estimator [5] + backprop, M=1
as reported in Raiko et al. [17]
1.71%
as reported in Tang and Salakhutdinov [20], M=20
3.99%
as reported in Raiko et al. [17], M=20
1.63%
Table 1. Mean test Error on MNIST for stochastoc networks. The first row shows the results of
our experiments averaged over 10 trials. The second row shows the results reported in [17]. M
corresponds to the number of samples used for computing output probabilities. We used M=1
during training and M=100 for the test set.
For evaluation, we averaged the output probabilities for a given input over 100 samples, and classified the example accordingly, following [17]. Results are given in Table
1. We obtained a test error of 1.71% using the baseline method and 1.54% using target propagation, which is – to our knowledge – the best result for stochastic nets on
MNIST reported so far. This suggests that target propagation is highly promising for
training networks of binary stochastic units.
Difference Target Propagation
3.4
13
Auto-encoder
We trained a denoising auto-encoder with 1000 hidden units with difference target
propagation as described in Section 2.4 on MNIST. As shown in Figure 4 stroke-like
filters can be obtained by target propagation. After supervised fine-tuning (using backpropagation), we got a test error of 1.35%. Thus, by training an auto-encoder with target
propagation one can learn a good initial representation, which is as good as the one obtained by regularized auto-encoders trained by back-propagation on the reconstruction
error.
Fig. 4. Filters learned by the back-propagation-free auto-encoder. Each filter corresponds to the
hidden weights of one of 100 randomly chosen hidden units. We obtain stroke filters, similar to
those usually obtained by regularized auto-encoders.
4
Conclusion
We introduced a novel optimization method for neural networks, called target propagation, which was designed to overcome drawbacks of back-propagation and is biologically more plausible. Target propagation replaces training signals based on partial
derivatives by targets which are propagated based on an auto-encoding feedback loop.
Difference target propagation is a linear correction for this imperfect inverse mapping
which is effective to make target propagation actually work. Our experiments show
that target propagation performs comparable to back-propagation on ordinary deep networks and denoising auto-encoders. Moreover, target propagation can be directly used
on networks with discretized transmission between units and reaches state of the art
performance for stochastic neural networks on MNIST.
Acknowledgments We would like to thank Junyoung Chung for providing RMSprop code,
Caglar Gulcehre and Antoine Biard for general discussion and feedback, Jyri Kivinen for discussion of backprop-free auto-encoder, Mathias Berglund for explanation of his stochastic networks.
14
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
We thank the developers of Theano [8, 1], a Python library which allowed us to easily develop
a fast and optimized code for GPU. We are also grateful for funding from NSERC, the Canada
Research Chairs, Compute Canada, and CIFAR.
References
[1] Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I.J., Bergeron, A.,
Bouchard, N., Bengio, Y.: Theano: new features and speed improvements. Deep
Learning and Unsupervised Feature Learning NIPS 2012 Workshop (2012)
[2] Bengio, Y.: Learning deep architectures for AI. Now Publishers (2009)
[3] Bengio, Y.: Estimating or propagating gradients through stochastic neurons. Tech.
Rep. arXiv:1305.2982, Universite de Montreal (2013)
[4] Bengio, Y.: How auto-encoders could provide credit assignment in deep networks
via target propagation. Tech. rep., arXiv:1407.7906 (2014)
[5] Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients
through stochastic neurons for conditional computation. arXiv:1308.3432 (2013)
[6] Bengio, Y., Thibodeau-Laufer, E., Yosinski, J.: Deep generative stochastic networks trainable by backprop. In: ICML’2014 (2014)
[7] Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization. J. Machine Learning Res. 13, 281–305 (2012)
[8] Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G.,
Turian, J., Warde-Farley, D., Bengio, Y.: Theano: a CPU and GPU math expression
compiler. In: Proceedings of the Python for Scientific Computing Conference
(SciPy) (Jun 2010), oral Presentation
[9] Carreira-Perpinan, M., Wang, W.: Distributed optimization of deeply nested systems. In: AISTATS’2014, JMLR W&CP. vol. 33, pp. 10–19 (2014)
[10] Erhan, D., Courville, A., Bengio, Y., Vincent, P.: Why does unsupervised pretraining help deep learning? In: JMLR W&CP: Proc. AISTATS’2010. vol. 9, pp.
201–208 (2010)
[11] Hinton, G., Deng, L., Dahl, G.E., Mohamed, A., Jaitly, N., Senior, A., Vanhoucke,
V., Nguyen, P., Sainath, T., Kingsbury, B.: Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine 29(6), 82–97 (Nov
2012)
[12] Konda, K., Memisevic, R., Krueger, D.: Zero-bias autoencoders and the benefits
of co-adapting features. Under review on International Conference on Learning
Representations (2015)
[13] Krizhevsky, A., Sutskever, I., Hinton, G.: ImageNet classification with deep convolutional neural networks. In: NIPS’2012 (2012)
[14] Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images.
Master’s thesis, University of Toronto (2009)
[15] LeCun, Y.: Learning processes in an asymmetric threshold network. In: FogelmanSoulié, F., Bienenstock, E., Weisbuch, G. (eds.) Disordered Systems and Biological Organization, pp. 233–240. Springer-Verlag, Les Houches, France (1986)
[16] LeCun, Y.: Modèles connexionistes de l’apprentissage. Ph.D. thesis, Université
de Paris VI (1987)
Difference Target Propagation
15
[17] Raiko, T., Berglund, M., Alain, G., Dinh, L.: Techniques for learning binary
stochastic feedforward neural networks. NIPS Deep Learning Workshop 2014
(2014)
[18] Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural
networks. Tech. rep., arXiv:1409.3215 (2014)
[19] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. Tech. rep.,
arXiv:1409.4842 (2014)
[20] Tang, Y., Salakhutdinov, R.: A new learning algorithm for stochastic feedforward
neural nets. ICML’2013 Workshop on Challenges in Representation Learning
(2013)
[21] Tieleman, T., Hinton, G.: Lecture 6.5-rmsprop: Divide the gradient by a running
average of its recent magnitude. COURSERA: Neural Networks for Machine
Learning 4 (2012)
[22] Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P.A.: Stacked denoising autoencoders: Learning useful representations in a deep network with a local
denoising criterion. J. Machine Learning Res. 11 (2010)
Appendix
A
Proof of Theorem 1
Proof. Given a training example (x, y) the back-propagation update is given by
δWibp = −
0,M
∂L(x, y; θW
)
∂L
= −JfTi+1 . . . JfTM
(si (hi−1 ))T ,
∂Wi
∂hM
∂hk
where Jfk = ∂h
= Wi · Si0 (hk−1 ), k = i + 1, . . . , M . Here Si0 (hk−1 ) is a
k−1
diagonal matrix with each diagonal element being element-wise derivatives and Jfk
is the Jacobian of fk (hk−1 ). In target propagation the target for hM is given by
∂L
. If all hk ’s are allocated in smooth areas and η̂ is sufficiently
ĥM = hM − η̂ ∂h
M
small, we can apply a Taylor expansion to get
ĥi = gi+1 (. . . gM (ĥM ) . . . ) = gi+1 (. . . gM (hM ) . . . )− η̂Jgi+1 . . . JgM
∂L
+o(η̂) ,
∂hM
where o(η̂) is the remainder satisfying limη̂→0 o(η̂)/η̂ = 0. Now, for δWitp we have
δWitp = −
∂||hi (hi−1 ; Wi ) − ĥi ||22
∂Wi
= −(hi − (hi − η̂Jf−1
. . . Jf−1
i+1
M
= −η̂Jf−1
. . . Jf−1
i+1
M
∂L
+ o(η̂)))(si (hi−1 ))T
∂hM
∂L
(si (hi−1 ))T + o(η̂)(si (hi−1 ))T .
∂hM
16
D-H. Lee, S. Zhang, A. Fischer and Y. Bengio
We write
∂L
∂hM
as l , si (hi−1 ) as v and JfM . . . Jfi+1 as J for short. Then the inner
production of vector forms of δWibp and δWitp is
hvec(δWibp ), vec(δWitp )i = tr((J T lvT )T (η̂J −1 lvT + o(η̂)vT ))
= η̂tr(vlT JJ −1 lvT ) − tr(vlT Jo(η̂)vT ) = η̂||v||22 ||l||22 − hJ T l, o(η̂)i||v||22 .
For ||vec(δWibp )||2 and ||vec(δWitp )||2 we have
q
||vec(δWibp )||2 = tr((−J T lvT )T (−J T lvT )) = ||v||2 ||J T l||2 ≤ ||v||2 ||J T ||2 ||l||2
and similarly
||vec(δWitp )||2 ≤ η̂||v||2 ||J −1 ||2 ||l||2 + ||o(η̂)||2 ||v||2 ,
where ||J T ||2 and ||J −1 ||2 are matrix Euclidean norms, i.e. the largest singular value
1
of (JfM . . . Jfi+1 )T , λmax , and the largest singular value of (JfM . . . Jfi+1 )−1 , λmin
T
(λmin is the smallest singular value of (JfM . . . Jfi+1 ) , because fk is invertable, so all
the smallest singular values of Jacobians are larger than 0). Finally, if η̂ is sufficiently
small, the angle α between vec(δWibp ) and vec(δWitp ) satisfies:
cos(α) =
≥
hvec(δWibp ), vec(δWitp )i
||vec(δWibp )||2 · ||vec(δWitp )||2
η̂||v||22 ||l||22 − hJ T l, o(η̂)i||v||22
1
)||l||2 + ||o(η̂)||2 ||v||2 )
(||v||2 λmax ||l||2 )(η̂||v||2 ( λmin
−hJ T l,o(η̂)i
η̂||l||22
λmax ||o(η̂)||2
+
η̂||l||2
1+
=
λmax
λmin
=
1 + ∆1 (η̂)
+ ∆2 (η̂)
λmax
λmin
where the last expression is positive if η̂ is sufficiently small and cos(α) ≤ 1 is trivial.
B
Proof of Theorem 2
Proof. Let e = ĥi − hi . Applying Taylor’s theorem twice, we get
ĥi − fi (ĥi−1 ) = ĥi − fi (hi−1 + gi (ĥi ) − gi (hi )) = ĥi − fi (hi−1 + Jgi e + o(||e||2 ))
= ĥi − fi (hi−1 ) − Jfi (Jgi e + o(||e||2 )) − o(||Jgi e + o(||e||2 )||2 )
= ĥi − hi − Jfi Jgi e − o(||e||2 ) = (I − Jfi Jgi )e − o(||e||2 )
where the vector o(||e||2 ) represents the remainder satisfying lime→0 o(||e||2 )/||e||2 =
0. Then for ||ĥi − fi (ĥi−1 )||22 we have
||ĥi − fi (ĥi−1 )||22 = ((I − Jfi Jgi )e − o(||e||2 ))T ((I − Jfi Jgi )e − o(||e||2 ))
= eT (I − Jfi Jgi )T (I − Jfi Jgi )e − o(||e||2 )T (I − Jfi Jgi )e
−eT (I − Jfi Jgi )T o(||e||2 ) + o(||e||2 )T o(||e||2 ))
= eT (I − Jfi Jgi )T (I − Jfi Jgi )e + o(||e||22 )
≤ λ||e||22 + |o(||e||22 )|
(A-1)
Difference Target Propagation
17
where o(||e||22 ) is the scalar value resulting from all terms depending on o(||e||2 ) and
λ is the largest eigenvalue of (I − Jfi Jgi )T (I − Jfi Jgi ). If e is sufficiently small to
guarantee |o(||e||22 )| < (1 − λ)||e||22 , then the left of Equation (A-1) is less than ||e||22
which is just ||ĥi − hi ||22 .
| 9cs.NE
|
1
The SCJ Small Parsimony Problem for Weighted
Gene Adjacencies
arXiv:1603.08819v2 [cs.DS] 26 Sep 2016
Nina Luhmann, Manuel Lafond, Annelyse Thévenin, Aı̈da Ouangraoua, Roland Wittler and Cedric Chauve
Abstract—Reconstructing ancestral gene orders in a given phylogeny is a classical problem in comparative genomics. Most existing
methods compare conserved features in extant genomes in the phylogeny to define potential ancestral gene adjacencies, and either try
to reconstruct all ancestral genomes under a global evolutionary parsimony criterion, or, focusing on a single ancestral genome, use a
scaffolding approach to select a subset of ancestral gene adjacencies, generally aiming at reducing the fragmentation of the
reconstructed ancestral genome.
In this paper, we describe an exact algorithm for the Small Parsimony Problem that combines both approaches. We consider that gene
adjacencies at internal nodes of the species phylogeny are weighted, and we introduce an objective function defined as a convex
combination of these weights and the evolutionary cost under the Single-Cut-or-Join (SCJ) model. The weights of ancestral gene
adjacencies can e. g. be obtained through the recent availability of ancient DNA sequencing data, which provide a direct hint at the
genome structure of the considered ancestor, or through probabilistic analysis of gene adjacencies evolution. We show the
NP-hardness of our problem variant and propose a Fixed-Parameter Tractable algorithm based on the Sankoff-Rousseau dynamic
programming algorithm that also allows to sample co-optimal solutions. We apply our approach to mammalian and bacterial data
providing different degrees of complexity. We show that including adjacency weights in the objective has a significant impact in
reducing the fragmentation of the reconstructed ancestral gene orders.
An implementation is available at http://github.com/nluhmann/PhySca.
F
1
I NTRODUCTION
Reconstructing ancestral gene orders is a long-standing
computational biology problem with important applications, as shown in several recent large-scale projects [15],
[28], [29]. Informally, the problem can be defined as follows:
Given a phylogenetic tree representing the speciation history leading to a set of extant genomes, we want to reconstruct the structure of the ancestral genomes corresponding
to the internal nodes of the tree.
Existing ancestral genome reconstruction methods concentrate on two main strategies. Local approaches consider
the reconstruction of one specific ancestor at a time independently from the other ancestors of the tree. Usually, they
do not consider an evolutionary model and proceed in two
stages: (1) comparing gene orders of ingroup and outgroup
species to define potential ancestral gene adjacencies, and (2)
selecting a conflict-free subset of ancestral gene adjacencies
– where a conflict is defined as an ancestral gene extremity
belonging to more than two potential adjacencies, e. g. due
to convergent evolution –, to obtain a set of Contiguous
Ancestral Regions (CARs) [7], [13], [24]. The second stage
of this approach is often defined as a combinatorial optimization problem aiming to minimize the number of dis•
•
•
•
•
N. Luhmann and R. Wittler are with the International Research Training
Group “Computational Methods for the Analysis of the Diversity and
Dynamics of Genomes”, Bielefeld University, Germany.
N. Luhmann, R. Wittler and A. Thévenin are with the Genome Informatics
group, Faculty of Technology and Center for Biotechnology, Bielefeld
University, Germany.
M. Lafond is with the Department of Computer Science and Operational
Research, Université de Montréal, Montréal, Canada.
A. Ouangraoua is with the Department of Computer Science, Université
de Sherbrooke, Sherbrooke, Canada.
C. Chauve is with the Department of Mathematics, Simon Fraser University, Burnaby (BC), Canada.
carded ancestral adjacencies, thus maximizing the number
of selected adjacencies [7], [24], [26]. This stage follows
principles common in scaffolding methods used to obtain
gene orders for extant genomes from sequencing data [10],
[25]. This approach was recently used to scaffold an ancestral pathogen genome for which ancient DNA (aDNA)
sequencing data could be obtained [30]. Global approaches
on the other hand simultaneously reconstruct ancestral gene
orders at all internal nodes of the considered phylogeny,
generally based on a parsimony criterion within an evolutionary model. This so called Small Parsimony Problem has
been studied with several underlying genome rearrangement models, such as the breakpoint distance or the DoubleCut-and-Join (DCJ) distance [2], [21], [37]. While rearrangement scenarios based on complex rearrangement models
can give insights into underlying evolutionary mechanisms,
from a computational point of view, the Small Parsimony
Problem is NP-hard for most rearrangement distances [34].
One exception is the Single-Cut-or-Join (SCJ) distance, for
which linear/circular ancestral gene orders can be found
in polynomial time [17], however constraints required to
ensure algorithmic tractability yield fragmented ancestral
gene orders.
The two approaches outlined above optimize somewhat
orthogonal criteria. For example, the underlying goal of
the local approach is to maximize the agreement between
the resulting ancestral gene order and the set of potential
ancestral adjacencies, independently of the other ancestral
gene orders. Would it be applied independently to all ancestral nodes, potential ancestral adjacencies exhibiting a
mixed profile of presence/absence in the extant genomes
might then lead to a set of non-parsimonious ancestral gene
orders. The global approach aims only at minimizing the
2
evolutionary cost in the phylogeny and can result in more
fragmented ancestral gene orders. Nevertheless, there is
little ground to claim that one approach or the other is more
accurate or to be preferred, and the work we present is an
attempt to reconcile both approaches.
We introduce a variant of the Small Parsimony Problem
based on an optimality criterion that accounts for both an
evolutionary distance and the difference between the initial
set of potential ancestral adjacencies and the final consistent
subset of adjacencies conserved at each ancestral node.
More precisely we consider that each potential ancestral
gene adjacency can be provided with a (prior) non-negative
weight at every internal node. The contribution of the
discarded adjacencies to the objective function is then the
sum of their weights. These adjacency weights can e. g. be
obtained as probabilities computed by sampling scenarios
for each potential adjacency independently [12] or can be
based on ancient DNA (aDNA) sequencing data providing
direct prior information assigned to certain ancestral nodes.
It follows that the phylogenetic framework we present can
then also assist in scaffolding fragmented assemblies of
aDNA sequencing data [22], [30].
We prove NP-hardness of the problem variant we introduce and describe an exact exponential time algorithm for
reconstructing consistent ancestral genomes under this optimality criterion, based on a mixed Dynamic Programming
/ Integer Linear Programming approach. We show that
this Small Parsimony Problem variant is Fixed-Parameter
Tractable (FPT), with a parameter linked to the amount of
conflict in the data. Moreover, this also allows us to provide
an FPT sampling algorithm for co-optimal solutions, – a
problem recently addressed in [27] using a MCMC approach. We evaluate our method on a simulated dataset and
compare our results to several other methods reconstructing
ancestral genomes. Further, we apply our method to two
real data sets: mammalian genomes spanning roughly one
million years of evolution, and bacterial genomes (pathogen
Yersinia) spanning 20, 000 years of evolution and for which
some aDNA sequencing data is available. We show that
we can reduce the fragmentation of ancestral gene orders
in both datasets by integrating adjacency weights while
reconstructing robust ancestral genomes.
This paper is an extended version of the work previously
presented in [23], particularly including new results on
simulated datasets and a hardness proof of the defined
problem.
2
2.1
BACKGROUND AND PROBLEM STATEMENT
Genomes and adjacencies
Genomes consist of chromosomes and plasmids. Each such
component can be represented as a linear or circular sequence of oriented markers over a marker alphabet. Markers
correspond to homologous sequences between genomes,
e. g. genes or synteny blocks. We assume that each marker
appears exactly once in each genome, so our model does
not consider duplications or deletions. To account for its
orientation, each marker x is encoded as a pair of marker
extremities (xh , xt ) or (xt , xh ).
An adjacency is an unordered pair of marker extremities,
e. g. {xt , yh }. The order of markers in a genome can be
encoded by a set of adjacencies. Two distinct adjacencies
are said to be conflicting if they share a common marker
extremity. If a set of adjacencies contains conflicting adjacencies, it is not consistent with a mixed linear/circular genome
model. We assume that the set of adjacencies for an extant
assembled genome is consistent. The set of adjacencies for
one genome naturally defines an adjacency graph, where
nodes represent marker extremities and edges represent
adjacencies. Conflicting adjacencies can be identified as
branching nodes in this graph.
2.2 The Small Parsimony Problem and rearrangement
distances
In a global phylogenetic approach, we are given a
phylogenetic tree with extant genomes at its leaves and
internal nodes representing ancestral genomes. We denote
by A the set of all adjacencies present in at least one extant
genome and assume that every ancestral adjacency belongs
to A. Then the goal is to find a labeling of the internal nodes
by consistent subsets of A minimizing a chosen genomic
distance over the tree. This is known as the Parsimonious
Labeling Problem.
Definition 1 (Parsimonious Labeling Problem). Let T =
(V, E) be a tree with each leaf l labeled with a consistent
set of adjacencies Al ⊆ A, and d a distance between
consistent sets of adjacencies. A labeling λ : V → P(A)
with λ(l) = Al for each leaf is parsimonious for d if
none of the internal nodes v ∈ V contains a conflict and
it minimizes the sum W (λ, T ) of the distances along the
branches of T :
X
W (λ, T ) =
d λ(u), λ(v) .
(u,v)∈E
This problem is NP-hard for most rearrangement distances taken as evolutionary models. The only known
exception is the set-theoretic Single-Cut-or-Join (SCJ) distance [17]. It defines a rearrangement distance by two operations: the cut and join of adjacencies. Given two genomes
defined by consistent sets of adjacencies A and B , the SCJ
distance between these genomes is
dSCJ (A, B) = | A − B | + | B − A | .
The Small Parsimony Problem under the SCJ model
can be solved by computing a parsimonious gain/loss
history for each adjacency separately with the dynamic
programming Fitch algorithm [18], [19] in polynomial time.
Consistent labelings can be ensured with the additional
constraint that in case of ambiguity at the root of the tree, the
absence of the adjacency is chosen [17]. As each adjacency is
treated independently, this constraint might automatically
exclude all adjacencies being part of a conflict to ensure
consistency. This results in an unnecessarily sparse reconstruction in terms of reconstructed adjacencies and thus
more fragmented genomes higher up in the tree.
2.3
Generalization by weighting adjacencies
When considering an internal node v , we define node u as
its parent node in T . We assume that a specific adjacency
3
graph is associated to each ancestral node v , whose edges
are annotated by a weight wv,a ∈ [0, 1] representing a confidence measure for the presence of adjacency a in species v .
Then in a global reconstruction, cutting an adjacency of a
higher weight has higher impact in terms of the optimization criterion than cutting an adjacency of lower weight.
Formally, we define two additional variables for each
adjacency a ∈ A at each internal node v ∈ V : The presence
(or absence) of a at node v is represented by pv,a ∈ {0, 1},
while cv,a ∈ {0, 1} indicates a change for the status of an
adjacency along an edge (u, v), i.e., pu,a 6= pv,a . We consider
the problem of optimizing the following objective function,
where α ∈ [0, 1] is a convex combination factor.
Definition 2 (Weighted SCJ Labeling Problem). Let T =
(V, E) be a tree with each leaf l labeled with a consistent
set of adjacencies Al ⊆ A and each adjacency a ∈ A
is assigned a given weight wv,a ∈ [0, 1] for each node
v ∈ V . A labeling λ of the internal nodes of T with
λ(l) = Al for each leaf is an optimal weighted SCJ labeling
if none of the internal nodes v ∈ V contains a conflict
and it minimizes the criterion
X
α(1 − pv,a )wv,a + (1 − α)cv,a
D(λ, T ) =
v,a
Further, we can state the corresponding co-optimal
sampling problem. A sampling method is important to
examine different co-optimal rearrangement scenarios
that can explain evolution toward the structure of extant
genomes.
Definition 3 (Weighted SCJ Sampling Problem). Given the
setting of the Weighted SCJ Labeling Problem, sample
uniformly from all labelings λ of the internal nodes of T
that are solutions to the Weighted SCJ Labeling Problem.
2.4
Problem complexity
Aside of the many heuristics for the Small Parsimony
Problem for non-SCJ rearrangement models (see for example [21], [36], [37] for the DCJ distance), there exist a few
positive results for the Weighted SCJ Labeling Problem with
specific values of α.
If α = 0, the objective function corresponds to the Small
Parsimony Problem under the SCJ distance and hence a
solution can be found in polynomial time [17]. A generalization towards multifurcating, edge-weighted trees including
prior information on adjacencies at exactly one internal
node of the tree is given in [22]. Recently, Miklós and
Smith [27] proposed a Gibbs sampler for sampling optimal
labelings under the SCJ model with equal branch lengths.
It starts from an optimal labeling obtained as in [17], and
then explores the space of co-optimal labelings through
repeated constrained parsimonious modifications of a single
adjacency evolutionary scenario. This method addresses the
issue of the high fragmentation of internal node labelings,
but convergence is not proven, and so there is no bound on
the computation time.
If α = 1, i.e., we do not take evolution in terms of
SCJ distance along the branches of the tree into account,
we can solve the problem by applying independently a
maximum-weight matching algorithm at each internal
node [26]. So the extreme cases of the problem are tractable,
and while we assume that the general problem is hard, we
will now prove it for a small range of α.
Theorem 1. The Weighted SCJ Labeling Problem is NP-hard
for any α > 33/34.
We show the hardness by reduction from the Maximum
Intersection Matching Problem, which is defined as follows.
Let G1 and G2 be two graphs on the same vertex set. Find
a perfect matching in G1 and G2 such that the number
of edges common to both matchings is maximized. We
prove NP-hardness of this problem by reduction from
3-Balanced-Max-2-SAT (see appendix for details).
Theorem 2. The Maximum Intersection Matching Problem
is NP-complete.
The relation of the Weighted SCJ Labeling Problem
and the Maximum Intersection Matching Problem can be
sketched as follows. For a given instance of the Maximum
Intersection Matching Problem, G1 and G2 , we construct
a tree that contains the edges of both graphs as potential
adjacencies. For α > 33/34, an optimal labeling of two
internal nodes then corresponds to perfect matchings in
G1 and G2 . Maximizing the number of common edges of
the matching then minimizes the SCJ distance between the
nodes. A detailed proof is given in the appendix.
3
M ETHODS
In order to find a solution to the Weighted SCJ Labeling
Problem, we first show that we can decompose the problem into smaller independent subproblems. Then, for each
subproblem containing conflicting adjacencies, we show
that, if it contains a moderate level of conflict, it can be
solved using the Sankoff-Rousseau algorithm [32] with a
complexity parameterized by the size of the subproblem.
For a highly conflicting subproblem, we show that it can be
solved by an Integer Linear Program (ILP).
3.1
Decomposition into independent subproblems
We first introduce a graph that encodes all adjacencies
present in at least one internal node of the considered
phylogeny (Def. 4 and Supplementary Fig. 7). As introduced
previously, we consider a tree T = (V, E) where each node
is augmented with an adjacency graph.
Definition 4 (Global adjacency graph). The set of vertices
VAG of the global adjacency graph AG consists of
all marker extremities present in at least one of the
adjacency graphs. There is an edge between two vertices
a, b ∈ VAG that are not extremities of a same marker, if
there is an internal node in the tree T whose adjacency
graph contains the adjacency {a, b}. The edge is labeled
with the list of all internal nodes that contain this
adjacency.
4
Each connected component C of the global adjacency
graph defines a subproblem composed of the species
phylogeny, the set of marker extremities equal to the
vertex set of C , and the set of adjacencies equal to the
edge set of C . According to the following lemma, whose
proof is straightforward, it is sufficient to solve each such
subproblem independently.
Lemma 1. The set of all optimal solutions of the Weighted
SCJ Labeling Problem is the set-theoretic Cartesian
product of the sets of optimal solutions of the instances
defined by the connected components of the global
adjacency graph.
To solve the problem defined by a connected component C of the global adjacency graph containing conflicts,
we rely on an adaptation of the Sankoff-Rousseau algorithm
with exponential time complexity, parameterized by the size
and nature of conflicts of C , and thus can solve subproblems
with moderate amount of conflict.
3.2
Overview of the Sankoff-Rousseau algorithm
The Sankoff-Rousseau dynamic programming algorithm [32] solves the general Small Parsimony Problem for
discrete characters. Let L be the set of all possible labels of
a node in the phylogeny. Then for each node u in the tree,
the cost c(a, u) of assigning a label a ∈ L to this node is
defined recursively as follows:
X
c(a, u) =
min c(b, v) + d(a, b)
v child of u
b∈L
where in our case d(a, b) is defined as in Definition 2. This
equation defines a dynamic programming algorithm whose
base case is when u is a leaf in which case c(a, u) = 0 if
λ(u) = a and c(a, u) = ∞ otherwise. Afterwards, choosing
a label with the minimum cost at the root node and backtracking in a top-down traversal of the tree results in a most
parsimonious labeling. We refer to [14] for a review on the
Sankoff-Rousseau algorithm.
3.3
Application to the Weighted SCJ Labeling Problem
In order to use the Sankoff-Rousseau algorithm to solve the
problem defined by a connected component C of the global
adjacency graph, we define a label of an internal node of
the phylogeny as the assignment of at most one adjacency
to each marker extremity. More precisely, let x be a marker
extremity in C , v an internal node of T , and e1 , . . . , edx be
all edges in the global adjacency graph that are incident to x
and whose label contains v (i.e., represent adjacencies in the
adjacency graph of node v ). We define the set of possible
labels of v as Lx,v = {∅, e1 , . . . , edx }. The set of potential
labels Lv of node v is then the Cartesian product of the label
sets Lx,v forQ
all x ∈ V (C) resulting in a set of discrete labels
for v of size x∈V (C) (1 + dx ). Note that not all of these joint
labelings are valid as they can assign an adjacency a = (x, y)
to x but not to y , or adjacency a = (x, y) to x and b = (x, z)
to z thus creating a conflict (see Supplementary Fig. 8 for an
example).
For an edge (u, v) in the tree, we can then define a
cost matrix that is indexed by pairs of labels of Lu and
Lv , respectively. The cost is infinite if one of the labels is
not valid, and defined by the objective function otherwise.
We can then apply the Sankoff-Rousseau approach to find
an optimal labeling of all internal nodes of the tree for
connected component C .
Note that, if C is a connected component with no conflict, it is composed of two vertices and a single edge, and
can be solved in space O(n) and time O(n).
3.4
Complexity analysis
The time and space complexity of the algorithm is obviously exponential in the size of C . Indeed, the time (resp.
space) complexity of the Sankoff-Rousseau algorithm for an
instance with a tree having n leaves and r possible labels
for each node is O(nr2 ) (resp. O(nr)) [14]. In our algorithm,
assuming n leaves in T (i.e. n extant species), mC vertices in
the global adjacency graph of C and a maximum degree dC
for vertices (marker extremities) in this graph, (1 + dC )mC is
an upper bound for the size of the label set Lv for a node v .
Moreover, computing the distance between two labels of
Lv and Lu , where (u, v) is an edge of T , can trivially be
done in time and space O(mC ): If both labels are valid, it
suffices to check how many common adjacencies are present
in both labels, while deciding if a label is not valid can be
done by a one-pass examination of the label. Combining
this with the Sankoff-Rousseau complexity yields a time
complexity in O(nmC (1 + dC )2mC ) and a space complexity
in O(nmC (1 + dC )mC ).
Given a general instance, i.e., an instance not limited to a
single connected component of the global adjacency graph,
we can consider each connected component independently
(Lemma 1). For a set of N markers and c connected
components in the global adjacency graph defining a
conflicting instance, we define D as the maximum degree
of a vertex and M as the maximum number of vertices in
all such components. Then, the complexity analysis above
shows that the problem is Fixed-Parameter Tractable (FPT).
Theorem 3. The Weighted SCJ Labeling Problem can be
solved in worst-case time O(nN (1 + D)2M ) and space
O(nN (1 + D)M ).
In practice, the exponential complexity of our algorithm depends on the structure of the conflicting connected
components of the global adjacency graph. The dynamic
programming algorithm will be effective on instances with
either small conflicting connected components or small degrees within such components, and will break down with
a single component with a large number of vertices of
high degree. For such components, the time complexity is
provably high and we propose an ILP to solve them.
3.5
An Integer Linear Program
We can formulate the optimization problem as a simple ILP.
We consider two variables for any adjacency a and node v ,
pv,a ∈ {0, 1} and cv,a ∈ {0, 1}, defined as in Section 2.
5
Minimize
X
α(1 − pv,a )wv,a + (1 − α)cv,a
v,a
subject to
pv,a + pu,a − pp,a ≥ 0 for (p, u), (p, v) ∈ E(T )
pv,a + pu,a − pp,a ≤ 1 for (p, u), (p, v) ∈ E(T )
pv,a + pu,a + cv,a ≤ 2 for (u, v) ∈ E(T )
pv,a + pu,a − cv,a ≥ 0 for (u, v) ∈ E(T )
pv,a − pu,a + cv,a ≥ 0 for (u, v) ∈ E(T )
− pv,a + pu,a + cv,a ≥ 0 for (u, v) ∈ E(T )
X
X
pv,a ≤ 1 and
pv,a ≤ 1
a=(xt ,y)
(c1 )
(c2 )
(c3 )
(c4 )
(c5 )
(c6 )
a=(xh ,y)
for any marker x and node v
(c7 )
The constraints ensure parsimony (c1 and c2 ), consistency of the solution (c7 ) and define the correct value for cv,a
dependent on the value of pa along an edge (u, v) (c3−6 ).
This ILP has obviously a size that is polynomial in the size
of the problem.
3.6
Sampling co-optimal labelings
The Sankoff-Rousseau DP algorithm can easily be modified
to sample uniformly from the space of all optimal solutions
to the Weighted SCJ labeling Problem in a forwardbackward fashion. The principle is to proceed in two stages:
first, for any pair (v, a) we compute the number of optimal
solutions under this label for the subtree rooted at v . Then,
when computing an optimal solution, if a DP equation has
several optimal choices, one is randomly picked according
to the distribution of optimal solutions induced by each
choice (see Appendix for more details). This classical
dynamic programming approach leads to the following
result.
Theorem 4. The Weighted SCJ Sampling Problem can be
solved in worst-case time O(nN (1 + D)2M ) and space
O(nN (1 + D)M ).
For subproblems that are too large for being handled by
the Sankoff-Rousseau algorithm, the SCJ Small Parsimony
Gibbs sampler recently introduced [27] can easily be modified to incorporate prior weights, although there is currently
no proven property regarding its convergence.
3.7
Weighting ancestral adjacencies
A first approach to assign weights to ancestral adjacencies
consist in considering evolutionary scenarios for an adjacency independently of the other adjacencies. An evolutionary scenario for an adjacency is a labeling of the internal
nodes of the species phylogeny T by the presence or absence
of the adjacency, and the parsimony score of a scenario is the
number of gains/losses of the adjacency along the branch of
T , i. e. the SCJ score for this single adjacency. For a scenario
σ , we denote by p(σ) its parsimony score. Its Boltzmann
p(σ)
score is then defined as B(σ) = e− kT , where kT is a given
constant. If we denote the set of all possible evolutionary
scenarios for the adjacency {x, y} by S(x, y), the partition
function of the adjacency and its Boltzmann probability are
defined as
X
B(σ)
Z(x, y) =
B(σ), P r(σ) =
.
Z(x, y)
σ∈S(x,y)
The weight of the adjacency at internal node v is then the
sum of the Boltzmann probabilities of all scenarios where
the adjacency is present at node v . All such quantities can
be computed in polynomial time [12].
Parameter kT is useful to skew the Boltzmann probability distribution: If kT tends to zero, parsimonious scenarios are heavily favored and the Boltzmann probability
distribution tends to the uniform distribution over optimal scenarios, while when kT tends to ∞, the Boltzmann
distribution tends toward the uniform distribution over
the whole solution space. In our experiments, we chose a
value of kT = 0.1 that favors parsimonious scenarios but
considers also slightly suboptimal scenarios.
When aDNA sequence data is available for one or several ancestral genomes, markers identified in extant species
can be related to assembled contigs of the ancestral genome,
as in [30] for example. For an ancestral adjacency in a species
for which aDNA reads are available, it is then possible
to associate a sequence-based weight to the adjacency –
either through gap filling methods (see Section 4, where we
use the probabilistic model of GAML [11]), or scaffolding
methods such as BESST [31] for example. In comparison to
the weighting approach described above, these weights are
then not directly based on the underlying phylogeny, but
provide an external signal for the confidence of adjacencies
at the respective internal node.
4
R ESULTS
We evaluated our algorithm on a simulated dataset and
compared its sensitivity and precision to several other reconstruction methods. Further, we applied our method to
two real datasets: mammalian and Yersinia genomes. The
mammalian dataset was used in the studies [13] and [27].
It contains six mammalian species and two outgroups,
spanning over 100 million years of evolution, and five
different marker sets of varying resolution (minimal marker
length). Our experimental results consider issues related
to the complexity of our algorithm, the use of a pure SCJ
reconstruction (obtained when the α parameter equals 0)
and the relative impact of the value of α on both the total
evolutionary cost and the ancestral gene orders fragmentation. Our second dataset contains eleven Yersinia genomes,
an important human pathogen. This dataset contains contigs
from the recently sequenced extinct agent of the Black Death
pandemic [9] that occurred roughly 650 years ago. We refer
to Supplementary Fig. 9 and 10 for the species phylogenies
of these two datasets.
4.1
Simulations
We created simulated datasets as described in [16]: with a
birth-rate of 0.001 and a death rate of 0, we simulated 20
binary trees with 6 leaves and scaled the branch lengths
such that the tree has a diameter D = 2n, where n is
the number of markers in each unichromosomal genome.
6
0.7
0.6
0.7
0.8
0.9
●
●
●
●
0.4
0.4
●
●
●
0.5
●
0.5
●
0.6
Sensitivity
0.6
●
0.5
Sensitivity
0.7
0.8
kT = 1
0.8
kT = 0.1
1.0
0.5
0.6
0.7
●
1.0
●
●●
0.9
1.0
●
0.6
F0.5 score
●
0.4
0.4
0.5
0.6
●
0.5
F0.5 score
0.9
0.7
●●
0.7
●
0.8
Precision
0.8
0.8
Precision
0.5
0.6
0.7
0.8
0.9
1.0
0.5
0.6
0.7
F1 score
MGRA
0.8
F1 score
RINGO
PHYSCA α
SCJ
0
0.3
ANGES
0.5
0.8
ues of α, the influence of the weighting becomes apparent:
for kT = 0.1, the precision only decreases for α = 1, while
for kT = 1, the precision decreases also for lower values
of α, however leading to more complete reconstructions. In
comparison, both DCJ-based methods RINGO and MGRA
produce less fragmented solutions by recovering more true
adjacencies under the jeopardy of also reconstructing more
false adjacencies. The sensitivity and precision of Fitch-SCJ,
ROCOCO and ANGES are comparable to our method for
low to medium values of α.
The F1 score assesses the relation of sensitivity and precision with equal importance. RINGO achieves a better F1
score than all other methods. The F0.5 score emphasizes the
precision of a method over its sensitivity. With this measure,
our method with kT = 1 and α = 0.5 outperforms the other
tools, while ROCOCO and ANGES also reach similarly good
scores.
In general, it can be seen that the equal contribution of
global evolution and local adjacency weights in the objective
function provides a reliable reconstruction and further a
useful tool to explore the solution space under different
values of α.
ROCOCO
1
Fig. 1. Average precision and sensitivity (top), and F1 and F0.5 (bottom)
of reconstructions on 20 simulated datasets. Adjacency weights have
been obtained with parameters kT = 0.1 (left) and kT = 1 (right).
The root genome with n = 500 markers is then evolved
along the branches of the tree by applying inversions and
translocations with a probability of 0.9 and 0.1 respectively.
The number of rearrangements at each branch corresponds
to the simulated branch length, the total number of rearrangements ranges from 1242 to 2296 in the simulated
trees. We compare results of our implementation PHYSCA
for different values of α ∈ {0, 0.3, 0.5, 0.8, 1} with the tools
RINGO [16], MGRA [4], Fitch-SCJ [8], ROCOCO [33], [35]
(dense approach for signed adjacencies) and ANGES [20]
(adjacencies only). We computed adjacency weights as described in subsection 3.7 with the software DeClone [12] and
parameter kT ∈ {0.1, 1}.
The methods RINGO and MGRA are global approaches
minimizing the DCJ-distance in the tree, while ANGES
reconstructs specific ancestors locally in the tree and is
applied for each node separately. For α = 0, our objective is
finding a consistent, most parsimonious solution and equals
the objectives of Fitch-SCJ and ROCOCO, where Fitch-SCJ
always finds the most fragmented solution whereas ROCOCO and our method aim at reporting least fragmented
reconstructions.
We measured sensitivity and precision of the reconstructions based on the comparison of simulated and reconstructed adjacencies by the different methods. A high
sensitivity indicates the ability to recover the true marker
order of ancestors in the phylogeny, while a high precision
denotes few wrongly reconstructed adjacencies. As shown
in Figure 1, our method reaches a high precision of 0.99
for all values of α ≥ 0.5, while increasing the sensitivity in
comparison to the pure Fitch-SCJ solution by reducing the
fragmentation of the reconstructed scaffolds. For higher val-
4.2
Mammalian dataset
We used the markers computed in [13] from whole-genome
alignments. The extant species contain a diverse number of
chromosomes ranging from 9 chromosomes in opossum to
39 chromosomes in pig. Unique and universal markers were
computed as synteny blocks with different resolution in
terms of minumum marker length. Note that all rearrangement breakpoints are therefore located outside of marker
coordinates. It results in five different datasets varying from
2, 185 markers for a resolution of 100 kb to 629 markers for
a resolution of 500 kb.
We considered all adjacencies present in at least one extant genome as potentially ancestral. To weight an adjacency
at all internal nodes of the tree, we relied on evolutionary
scenarios for each single adjacency, in terms of gain/loss,
independently of the other adjacencies (i. e. without considering consistency of ancestral marker orders). We obtain
these weights using the software DeClone [12], and we refer
to them as DeClone weights. We considered two values of the
DeClone parameter kT , 0.1 and 1, the former ensuring that
only adjacencies appearing in at least one optimal adjacency
scenario have a significant DeClone weight, while the latter
samples adjacencies outside of optimal scenarios. For the
analysis of the ancestral marker orders obtained with our
algorithm, we considered the data set at 500 kb resolution
and sampled 500 ancestral marker orders for all ancestral
species under different values of α.
Complexity
The complexity of our algorithm is dependent on the size
of the largest connected component of the global adjacency
graph. In order to restrict the complexity, we kept only
adjacencies whose weights are above a given threshold x.
Figure 2 shows the expected decrease in computational
complexity correlated to threshold x for the five different minimal marker lengths. In most cases, all connected
components are small enough to be handled by our exact
number of labels (log scale)
1e+03
1e+07
1e+11
7
3
3
1
1
marker length
1
1
3
3
2
2
100
200
300
400
500
1
2
2
4
4
4
4
5
5
5
5
0.4
0.5
threshold x
0.6
2
3
4
5
0.2
0.3
0.7
0.8
Fig. 2. Number of different labels for the largest connected component in
each of the mammalian datasets. This statistic provides an upper bound
for the actual complexity of our reconstruction algorithm.
100
number of CARs
90
●
●
●
●
●
●
80
●
●
70
●
●
60
50
40
Boreo.
Euarcho.
Ferun.
Primat.
ancestor
Rodent.
Theria
Fig. 3. Number of reconstructed CARs at each internal node in 500
samples for the mammalian dataset with 500 kb resolution, x = 0.2 and
α = 0.
algorithm in reasonable time except for very large components in the marker sets with higher resolution under a low
threshold x. For the 500 kb dataset with x = 0.2 and kT = 1,
the computation of one solution takes on average 200 s on a
2.6 GHz i5 with 8 GB of RAM. It can be reduced to 30 s when
DeClone weights are based on kT = 0.1. This illustrates
that our algorithm, despite an exponential worst-case time
complexity, can process realistic datasets in practice.
Optimal SCJ labelings
Next, we analyzed the 500 optimal SCJ labelings obtained
for α = 0, i. e. aiming only at minimizing the SCJ distance,
and considered the fragmentation of the ancestral gene
orders (number of CARs) and the total evolutionary distance. Note that, unlike the Fitch algorithm used in [17], our
algorithm does not favor fragmented assemblies by design
but rather considers all optimal labelings. Sampling of cooptimal solutions shows that the pure SCJ criterion leads
to some significant variation in terms of number of CARs
(Figure 3). In contrast, Table 1 shows that most observed
ancestral adjacencies are present in all sampled scenarios.
About 5% of adjacencies, mostly located at nodes higher
up in the phylogeny, are only present in a fraction of all
sampled scenarios, indicating that there is a small number
of conflicts between potential adjacencies that can be solved
ambiguously at the same parsimony cost.
TABLE 1
Frequency of adjacencies in 500 samples with α = 0 as percentage of
optimal labelings they appear in.
Ancestor
Boreoeutheria
Euarchontoglires
Ferungulates
Primates
Rodentia
Theria
root node
f = 100%
Frequency f
100% > f > 50%
f < 50%
94.66
95.42
96.53
98.82
99.49
97.67
92.23
1.07
0.88
0.55
0.34
0.34
0.89
1.23
4.27
3.79
2.92
0.84
0.17
1.43
6.53
The optimal SCJ distance in the tree for α = 0 is 1, 674,
while the related DCJ distance in the sampled reconstructions varies between 873 and 904 (Figure 4). In comparison,
we obtained a DCJ distance of 829 with GASTS [36], a
small parsimony solver directly aiming at minimizing the
DCJ distance. More precisely, over all ancestral nodes, 70
adjacencies found by GASTS do not belong to our predefined set of potential ancestral adjacencies and another 147
appear in the 500 samples with a frequency below 50%. This
illustrates both a lack of robustness of the pure SCJ optimal
labelings, and some significant difference between the SCJ
and DCJ distances.
Finally, we compared the Boltzmann probabilities of
ancestral adjacencies (DeClone weights) with the frequency
observed in the 500 samples. There is a very strong agreement for DeClone weights obtained with kT = 0.1 as
only 14 ancestral adjacency have a DeClone weight that
differs more than 10% from the observed frequency in the
samples. This shows that, despite the fact that the DeClone
approach disregards the notion of conflict, it provides a
good approximation of the optimal solutions of the SCJ
Small Parsimony Problem.
Ancestral reconstruction with DeClone weights and varying
values of α.
For α > 0, our method minimizes a combination of the
SCJ distance with the DeClone weights of the adjacencies
discarded to ensure valid ancestral gene orders. Again,
we sampled 500 solutions each for different values of α
with the 500 kb data set. We distinguish between DeClone
parameter kT = 0.1 and kT = 1. Figures 4 and 5 show the
respective observed results in terms of evolutionary distance
and fragmentation.
For kT = 0.1, the optimal SCJ and DCJ distance over
the whole tree hardly depends on α. Including the DeClone
weights in the objective actually results in the same solution,
independent of α > 0. In fact, while applying a low weight
threshold of x = 0.2, the set of potential adjacencies is
already consistent at all internal nodes except for a few
conflicts at the root that are solved unambiguously for all
values of α. This indicates that building DeClone weights
on the basis of mostly optimal adjacency scenarios (low kT )
results in a weighting scheme that agrees with the evolution
along the tree for this dataset. More importantly, Figures 4
and 5 show that the combination of DeClone weights
followed by our algorithm, leads to a robust set of ancestral
gene orders.
●
14
●
SCJ, kT 0.1
SCJ, kT 1
number of CARs
distance
1000
1700
1800
8
DCJ, kT 0.1
DCJ, kT 1
BD
node1
node5
node6
●
12
●
10
●
8
●
6
●
●
●
●
4
2
0.5
alpha
0.8
1
alpha
Fig. 4. SCJ distance (upper half) and DCJ (lower half) distance in the
whole tree for all samples and selected values of α in the mammalian
dataset.
●
●
●
●
70
●
●
60
50
Euarchonto.
Primates
Theria
0
0.3
0.5
0.8
1
0
0.3
0.5
0.8
1
40
0
0.3
0.5
0.8
1
number of CARs
80
alpha
Fig. 5. Number of CARs in the mammalian dataset in all samples at
selected internal nodes for different values of α reconstructed with
DeClone weights under kT = 0.1. While the number of CARs differs
in the case of α = 0 where the adjacency weights are not considered,
the fragmentation stays constant for the other values of α.
In comparison, for kT = 1, we see an increase in SCJ and
DCJ distance for higher α, while the number of CARs at internal nodes decreases, together with a loss of the robustness
of the sampled optimal results when α gets close to 1. It can
be explained by the observation that the weight distribution
of ancestral adjacencies obtained with DeClone and kT = 1
is more balanced than with kT = 0.1 as it considers suboptimal scenarios of adjacencies with a higher probability.
It further illustrates that, when the global evolutionary cost
of a solution has less weight in the objective function, the
algorithm favors the inclusion of an adjacency of moderate
weight that joins two CARs while implying a moderate
number of evolutionary events (for example an adjacency
shared by only a subset of extant genomes). From that point
of view, our algorithm – being efficient enough to be run
on several values of α – provides a useful tool to evaluate
the relation between global evolution and prior confidence
for adjacencies whose pattern of presence/absence in extant
genomes is mixed.
4.3
0
0.3
0.5
0.8
0.3
0
0.3
0.5
0.8
0
0
0.3
0.5
0.8
●
0
0.3
0.5
0.8
900
●
Yersinia pestis dataset
We started from fully assembled DNA sequences of seven
Yersinia pestis and four Yersinia pseudotuberculosis genomes.
In addition, we included aDNA single-end reads and 2 134
contigs of length > 500bp assembled from these reads for
Fig. 6. Reconstructed number of CARs in the yersinia dataset with aDNA
weights at the BD node and 0 otherwise, for four ancestral nodes.
the Black Death agent, considered as ancestral to several
extant strains [9]. We refer to this augmented ancestral node
as the Black Death (BD) node. The marker sequences for
all extant genomes were computed as described in [30],
restricting the set of markers to be unique and universal. We
obtained a total of 2, 207 markers in all extant genomes and
2, 232 different extant adjacencies, thus showing a relatively
low level of syntenic conflict compared to the number of
markers, although it implies a highly dynamic rearrangement history over the short period of evolution [30].
As for the mammalian dataset, we considered as potentially ancestral any adjacency that appears in at least
one extant genome. However for this dataset, reducing
the complexity by applying a weight threshold x was not
necessary. For the BD node, adjacency weights can be based
on the given aDNA reads for a given potential ancestral
adjacency as follows. First, we used FPSAC [30] to compute
DNA sequences filling the gaps between any two adjacent
marker extremities (obtained by aligning the gap sequences
of the corresponding conserved extant adjacencies and reconstructing a consensus ancestral sequence using the Fitch
algorithm). Then we computed the weights as a likelihood
of this putative gap sequence given the aDNA reads, using the GAML probabilistic model described in [11]. Each
adjacency together with its template gap sequence details a
proposition for an assembly A as a piece of the real ancestral
sequence, and given the aDNA read set
Q R, the model
then defines a probability P r(R|A) =
r∈R P r(r|A) for
observing the reads R given that A is the correct assembly.
The probability P r(r|A) can be computed by aligning r to
the assembly A while the alignment is evaluated under an
appropriate sequencing error model. We refer to [11] for
details.
Ancestral reconstruction with aDNA weights
Again we sampled 500 solutions for this dataset. We computed the weights at the BD node based on the aDNA data,
while adjacencies at all other nodes were given weight 0.
Hence we can investigate the influence of including the
aDNA sequencing data in the reconstruction while for the
rest of the tree, the weights do not impact the objective
function. Moreover, this weighting scheme addresses the
issue of potential BD adjacencies with a low weight due
to the difficulty of sequencing ancient DNA.
9
As shown in Figure 6, for selected internal nodes of
the phylogeny, the pure SCJ solutions at α = 0 result
in the highest fragmentation, while the number of CARs
decreases as we increase the importance of the adjacency
weights in the objective of our method. For the BD node,
when including the aDNA weights, the fragmentation is
decreasing while the reconstructions for each α > 0 are
robust. At the other nodes, the applied sequencing weights
also reduce the fragmentation except for node6 which is
located in the pseudotuberculosis subtree and hence more
distant to the BD node. This shows that the aDNA weights
not only influence the reconstructed adjacencies at the BD
node, but also other nodes of the tree.
5
C ONCLUSION
Our main contributions are the introduction of the Small
Parsimony Problem under the SCJ model with adjacency
weights, together with an exact parameterized algorithm
for the optimization and sampling versions of the problem.
The motivation for this problem is twofold: incorporating
sequence signal from aDNA data when it is available, and
recent works showing that the reconstruction of ancestral
genomes through the independent analysis of adjacencies is
an interesting approach [5], [12], [17], [27].
Regarding the latter motivation, we address a general
issue of these approaches that either ancestral gene orders
are not consistent or are quite fragmented if the methods are
constrained to ensure consistency. The main idea we introduce is to take advantage of sampling approaches recently
introduced in [12] to weight potential ancestral adjacencies
and thus direct, through an appropriate objective function,
the reconstruction of ancestral gene orders. Our results on
the mammalian dataset suggest that this approach leads
to a robust ancestral genome structure. However, we can
observe a significant difference with a DCJ-based ancestral
reconstruction, a phenomenon that deserves to be explored
further. Our algorithm, which is based on the SankoffRousseau algorithm similarly to several recent ancestral
reconstruction algorithms [5], [12], [27], is a parameterized
algorithm that can handle real instances containing a moderate level of syntenic conflict. Our experimental results on
both the mammalian and bacterial datasets suggest that
introducing prior weights on adjacencies in the objective
function has a significant impact in reducing the fragmentation of ancestral gene orders, even with an objective function with balanced contributions of the SCJ evolution and
adjacency weights. For highly conflicting instances, it can
be discussed if a reconstruction through small parsimony is
the right approach to solve these conflicts or if these should
be addressed differently.
Our sampling algorithm improves on the Gibbs sampler
introduced in [27] in terms of computational complexity and
provides a useful tool to study ancestral genome reconstruction from a Bayesian perspective. Moreover, our algorithm
is flexible regarding the potential ancestral gene adjacencies
provided as input and could easily be associated with other
ideas, such as intermediate genomes for example [16].
There are several research avenues opened by our work.
From a theoretical point of view, we know the problem
we introduced is tractable for α = 0 and α = 1, and
we show it is hard for α > 33/34, but it remains to
see whether it is hard otherwise. Further, given that the
considered objective is a combination of two objectives
to be optimized simultaneously, Pareto optimization is an
interesting aspect that should be considered. Our model
could also be extended towards other syntenic characters
than adjacencies, i. e. groups of more than two markers,
following the ancient gene clusters reconstruction approach
introduced in [33]. As ancestral gene orders are defined by
consistent sets of adjacencies, the principle of our dynamic
programming algorithm could be conserved and it would
only be a matter of integrating gene clusters into the objective function. From a more applied point of view, one
would like to incorporate duplicated and deleted markers
into our Small Parsimony Problem. There exist efficient
algorithms for the case of a single adjacency [5], [12] that
can provide adjacency weights, and natural extensions of
the SCJ model to incorporate duplicated genes. However
it remains to effectively combine these ideas. Finally, again
due to the flexibility and simplicity of the Sankoff-Rousseau
dynamic programming algorithm, one could easily extend
our method towards the inference of extant adjacencies if
some extant genomes are provided in partially assembled
form following the general approach described in [1], [3].
This would pave the way towards a fully integrated phylogenetic scaffolding method that combines evolution and
sequencing data for selected extant and ancestral genomes.
Acknowledgements
NL and RW are funded by the International DFG Research
Training Group GRK 1906/1. CC is funded by NSERC grant
RGPIN-249834.
R EFERENCES
[1]
Aganezov Jr., S., Sitdykova, N., Alekseyev, M.: Scaffold assembly
based on genome rearrangement analysis. Comput. Biol. and
Chemistry 57, 46–53 (2015)
[2] Alekseyev, M., Pevzner, P.A.: Breakpoint graphs and ancestral
genome reconstructions. Genome Res. 19, 943–957 (2009)
[3] Anselmetti, Y., Berry, V., pthers: Ancestral gene synteny reconstruction improves extant species scaffolding. BMC Genomics
16(Suppl 10), S11 (2015)
[4] Avdeyev, P., Jiang, S., Aganezov, S., Hu, F., Alekseyev, M.A.:
Reconstruction of ancestral genomes in presence of gene gain and
loss. Journal of Computational Biology 23(3), 150–164 (2016)
[5] Bérard, S., Gallien, C., et al.: Evolution of gene neighborhoods
within reconciled phylogenies. Bioinformatics 28, 382–388 (2012)
[6] Berman, P., Karpinski, M.: On some tighter inapproximability
results. In: International Colloquium on Automata, Languages,
and Programming. pp. 200–209. Springer (1999)
[7] Bertrand, D., Gagnon, Y., et al.: Reconstruction of ancestral genome
subject to whole genome duplication, speciation, rearrangement
and loss. In: WABI 2010, LNCS, vol. 6293, pp. 78–89. Springer
(2010)
[8] Biller, P., Feijão, P., Meidanis, J.: Rearrangement-based phylogeny
using the single-cut-or-join operation. IEEE/ACM TCBB 10(1),
122–134 (2013)
[9] Bos, K.I., Schuenemann, V., et al.: A draft genome of yersinia pestis
from victims of the black death. Nature 478, 506–510 (2011)
[10] Bosi, E., Donati, B., et al.: Medusa: a multi-draft based scaffolder.
Bioinformatics 31(15), 2443–2451 (2015)
[11] Boža, V., Brejová, B., Vinař, T.: GAML: Genome Assembly by
Maximum Likelihood. Algorithms Mol. Biol. 10, 18 (2015)
[12] Chauve, C., Ponty, Y., Zanetti, J.: Evolution of genes neighborhood within reconciled phylogenies: An ensemble approach. BMC
Bioinformatics 16(Suppl. 19), S6 (2015)
10
[13] Chauve, C., Tannier, E.: A methodological framework for the
reconstruction of contiguous regions of ancestral genomes and
its application to mammalian genomes. PLoS Comput. Biol. 4,
e1000234 (2008)
[14] Csűrös, M.: How to infer ancestral genome features by parsimony:
Dynamic programming over an evolutionary tree. In: Models and
Algorithms for Genome Evolution, pp. 29–45. Springer (2013)
[15] Denoeud, F., Carretero-Paulet, L., et al.: The coffee genome provides insight into the convergent evolution of caffeine biosynthesis. Science 345, 125527 (2013)
[16] Feijǎo, P.: Reconstruction of ancestral gene orders using intermediate genomes. BMC Bioinformatics 16(Suppl. 14), S3 (2015)
[17] Feijǎo, P., Meidanis, J.: SCJ: a breakpoint-like distance that simplifies several rearrangement problems. IEEE/ACM TCBB 8, 1318–
1329 (2011)
[18] Fitch, W.: Toward defining the course of evolution: minimum
change for a specific tree topology. Syst. Biol. 20, 406–416 (1971)
[19] Hartigan, J.A.: Minimum mutation fits to a given tree. Biometrics
29(1), 53–65 (1973)
[20] Jones, B.R., Rajaraman, A., Tannier, E., Chauve, C.: Anges: reconstructing ancestral genomes maps. Bioinformatics 28(18), 2388–
2390 (2012)
[21] Kovác, J., Brejová, B., Vinar, T.: A practical algorithm for ancestral
rearrangement reconstruction. In: WABI 2011, LNCS, vol. 6833, pp.
163–174. Springer (2011)
[22] Luhmann, N., Chauve, C., et al.: Scaffolding of ancient contigs
and ancestral reconstruction in a phylogenetic framework. In: BSB
2014, LNCS, vol. 8826, pp. 135–143. Springer (2014)
[23] Luhmann, N., Thévenin, A., Ouangraoua, A., Wittler, R., Chauve,
C.: The scj small parsimony problem for weighted gene adjacencies. In: International Symposium on Bioinformatics Research and
Applications. pp. 200–210. Springer (2016)
[24] Ma, J., Zhang, L., et al.: Reconstructing contiguous regions of an
ancestral genome. Genome Res. 16, 1557–1565 (2006)
[25] Mandric, I., Zelikovsky, A.: Scaffmatch: scaffolding algorithm
based on maximum weight matching. Bioinformatics 31(16), 2632–
2638 (2015)
[26] Maňuch, J., Patterson, M., et al.: Linearization of ancestral multichromosomal genomes. BMC Bioinformatics 13(Suppl 19), S11
(2012)
[27] Miklós, I., Smith, H.: Sampling and counting genome rearrangement scenarios. BMC Bioinformatics 16(Suppl 14), S6 (2015)
[28] Ming, R., Van Buren, R., et al.: The pineapple genome and the
evolution of CAM photosynthesis. Nature Genetics (2015), in press
[29] Neafsey, D.E., Waterhouse, R.M., et al.: Highly evolvable malaria
vectors: The genome of 16 Anopheles mosquitoes. Science 347,
1258522 (2015)
[30] Rajaraman, A., Tannier, E., Chauve, C.: FPSAC: fast phylogenetic
scaffolding of ancient contigs. Bioinformatics 29, 2987–2994 (2013)
[31] Sahlin, K., Vezzi, F., et al.: BESST - efficient scaffolding of large
fragmented assemblies. BMC Bioinformatics 15, 281 (2014)
[32] Sankoff, D., Rousseau, P.: Locating the vertices of a steiner tree in
an arbitrary metric space. Math. Programming 9, 240–246 (1975)
[33] Stoye, J., Wittler, R.: A unified approach for reconstructing ancient
gene clusters. IEEE/ACM TCBB 6(3), 387–400 (2009)
[34] Tannier, E., Zheng, C., Sankoff, D.: Multichromosomal median
and halving problems under different genomic distances. BMC
Bioinformatics 10(1), 120 (2009)
[35] Wittler, R.: Phylogeny-based Analysis of Gene Clusters. Ph.D.
Thesis, Faculty of Technology, Bielefeld University (2010)
[36] Xu, A., Moret, B.: GASTS: parsimony scoring under rearrangements. In: WABI 2011, LNCS, vol. 6833, pp. 351–363. Springer
(2011)
[37] Zheng, C., Sankoff, D.: On the pathgroups approach to rapid small
phylogeny. BMC Bioinformatics 12(Suppl 1), S4 (2011)
11
A PPENDIX
Results: supplementary material
Methods: supplementary material
Mouse
Rodentia
Rat
{xt, yh}
Euarchontoglires
v
xt
yh
v
Human
Primates
w
{xt, zt}
Boreoeutheria
zt
w
Theria
Macaca
Cow
Ferungulates
Pig
Opossum
Fig. 7. Simplified example of a connected component in the global
adjacency graph. All internal nodes of the tree are augmented with
adjacency graphs. At node v , the graph contains adjacency {xt , yh }, at
node u, the graph contains adjacency {xt , zt }. Along the edge {u, v},
one adjacency has been cut, the other has been joined. In the global
adjacency graph, we see a connected component that contains two
edges incident to xt . This component does not contain a conflict, just
a rearrangement. See Fig. 8 below for a C that contains a conflict.
Chicken
Fig. 9. Underlying phylogeny of the considered genomes for the mammalian dataset used in the evaluation.
Yersinia pestis Antiqua
2
Black Death
aDNA
1
4
3
Yersinia pestis C092
Yersinia pestis Z176003
C
v
bt
a3 w
dh
Cv
ch
v
a1
v,w
a2
a4 v,w
fh
v
bt
Yersinia pestis Nepal516
ch
Yersinia pestis KIM10
5
a1
a4 v
v
a2
a5
fh
v
et
Yersinia pestis biovar Microtus
Yersinia pestis Pestoides F
a5
et
Yersinia pseudotuberculosis IP31758
6
8
(bt, ), (ch,a4), (et,a4), (fh, )
invalid:
(bt,a1), (ch,a4), (et,a5), (fh,a5)
bt
ch
et
fh
a1
a1
a2
a5
a2
a4
a4
a5
Fig. 8. Example for a given connected component C . At each node v
of the tree, possible assignments are defined by the connected component containing all edges annotated with v . Possible assignments for a
marker extremity, e. g. bt , are defined by the incident adjacency edges,
hence in this example we can e. g. assign a1 ,a2 or ∅ to bt . Here two
valid joint labeling for v are shown, while the third one assigns different
adjacencies for bt and ch and is therefore invalid.
Yersinia pseudotuberculosis PB1
7
Yersinia pseudotuberculosis IP32953
Fig. 10. Underlying phylogeny of the considered genomes for the
Yersinia dataset used in the evaluation.
14
●
BD
node1
node5
node6
12
number of CARs
valid:
(bt,a1), (ch,a1), (et,a5), (fh,a5)
Yersinia pseudotuberculosis YPIII
10
●
8
6
4
0
0.3
0.5
0.8
1
0
0.3
0.5
0.8
1
0
0.3
0.5
0.8
1
BD
node1
node5
node6
number of CARs
14
12
10
8
6
4
0
0.3
0.5
0.8
1
2
0
0.3
0.5
0.8
1
gives the number of optimal solutions for the subproblem
rooted at v . At the root, we might have the choice between
different labels with minimum cost. Let Lroot be the set of
these labels, then the number of overall possible co-optimal
solutions is simply the sum of the number
of solutions
P
for all optimal root labels: C(root) =
l∈Lroot C(root, l).
Subsequently in the top-down traversal, choose a label
l ∈ Lroot with probability C(root,l)
. If at an internal node
C(root)
more than one label in a child node induced the minimum
value, choose one of these labels analogously.
Fig. 11. Yersinia dataset: Reconstructed number of CARs with DeClone
weights at kT = 1.
0
0.3
0.5
0.8
1
l∈Ly
alpha
0
0.3
0.5
0.8
1
l∈Lx
0
0.3
0.5
0.8
1
2
A DP sampling algorithm.
In the bottom-up traversal, in addition to the minimal cost
induced by labeling a node v with a specific label a ∈ L,
we can also store the number of optimal solutions under
this label for the subtree rooted at v . Let x and y be the
children of v , and Lx and Ly the sets of labels that induced
the minimum value for a at v (which means a label out of
these sets is assigned in the backtracking phase if v is labeled
with a). Then
X
X
C(v, a) =
C(x, l)
C(y, l)
alpha
Fig. 12. Yersinia dataset: Reconstructed number of CARs with DeClone
weights at kT = 0.1.
12
NP-Hardness for high values of α
0
X1
In this section we show that Weighted SCJ labeling is NPhard for 33/34 < α < 1.
17
+
0
1
17
3
16
15
X2
2
-
+
1
2
3
16
-
Z1
15
4
-
-
4
C1 = (x1 V x2)
The Maximum Intersection Matching Problem
Maximum Intersection Matching Problem (MIMP)
Input: two graphs G1 and G2 with V (G1 ) = V (G2 ), and
an integer r.
Question: does there exist a perfect matching M1
in G1 and a perfect matching M2 in G2 such that
|E(M1 ) ∩ E(M2 )| ≥ r?
We now introduce the problem used for showing the
hardness of MIMP. The 3-Balanced-Max-2-SAT problem is,
given n boolean variables x1 , . . . , xn and a set of m clauses
C = {C1 , . . . , Cm }, each with exactly two literals (e.g. (x ∨
y)), such that each variable appears in the clauses 3 times
positively and 3 times negatively, to find an assignment to
the n variables such that a maximum of clauses are satisfied.
Note that m = 3n since each variable appears in 6 clauses
and each clause has 2 variables. This problem is NP-hard
by a reduction from Max-Cut on 3-regular graphs [6]. We
reduce 3-Balanced-Max-2-SAT to MIMP, and we begin by
describing how we transform a given 3-Balanced-Max-2SAT instance into a MIMP instance. Figure 13 illustrates the
main ideas behind the reduction.
Let x1 , . . . , xn be the set of n variables and C =
{C1 , . . . , Cm } be the set of clauses of the given 3-BalancedMax-2-SAT instance. In our construction the graph G1 corresponds to the variables and G2 to the clauses.
The graph G1 is the disjoint union of n cycles X1 , . . . , Xn
each of length 18, one for each variable xi . For each
1 ≤ i ≤ n, the cycle Xi corresponding to xi has vertices
i
in cyclic order. Call the edges v0i v1i , v6i v7i and
v0i , v1i , . . . , v17
i i
i
i i
v16 negaand v15
v12 v13 positive, and the edges v3i v4i , v9i v10
tive. Note that Xi has exactly two perfect matchings: one
containing all the positive edges, and one containing all
the negative edges. Since G1 has n such cycles, G1 has 2n
perfect matchings, which are in bijection with the possible
xi assignments. Also note that no two positive or negative
edges share a vertex.
Some additional notations are needed before we can
construct G2 . Let C1+i , C2+i , C3+i be the 3 clauses containing
variable xi positively. Map each of the 3 positive edges of
Xi to a distinct clause arbitrarily (e.g. we map v0i v1i to C1+i ,
i i
v6i v7i to C2+i and v12
v13 to C3+i ). Similarly let C1−i , C2−i , C3−i
be the clauses in which xi appears negatively, and map each
negative edge of Xi to one of these clauses in a distinct
manner.
Now, G2 has the same vertex set as G1 , and is the union
of m cycles Z1 , . . . , Zm of length 6, one for each clause
in C . Let xj and xk be the two variables contained in a
j
clause Ci . Let vaj va+1 be the edge of the cycle Xj mapped
k k
to Ci , and vb vb+1 be the edge of the cycle Xk mapped
to Ci . Then the cycle Zi corresponding to Ci is, in cyclic
j
j
k
k
order, vaj va+1 va+5 vbk vb+1
vb+5
vaj , where addition is to be
taken modulo 18. Thus Zi contains exactly two edges in
j
k
common with G1 : vaj va+1 is from Xj and vbk vb+1
is from
Xk . Importantly, Zi has 2 possible perfect matchings: one
j
k
contains vaj va+1 and the other has vbk vb+1
.
14
13
5
6
+
+
7
12
11
10
-
8
9
14
13
5
6
+
+
7
12
11
10
-
8
9
Fig. 13. Illustration of a clause gadget used for the MIMP reduction. The
cycles X1 and X2 are the cycles of G1 corresponding to variables x1
and x2 . The ‘+’ edges are positive, and the ‘-’ edges are negative., The
cycle Z1 (dashed edges) corresponds to the clause C1 = (x1 ∨ x2 ),
where here the edges v01 v11 and v32 v42 are hypothetically mapped to C1 .
Other clause cycles are not shown.
Intuitively speaking, if an edge e is in both G1 and G2 ,
say on the cycles Xi of G1 and Zj of G2 , then e corresponds
to the fact that xi appears in clause Cj . If e is positive, setting
xi = T rue satisfies Cj , and if e is negative, setting xi =
F alse satisfies Cj .
We first prove some structural properties of G1 and G2
which will be of use later on. We denote by G the complement
of a graph G.
Lemma 2. The graphs G1 and G2 satisfy the following
conditions:
1)
2)
3)
4)
|V (G1 )| = |V (G2 )| are even;
G1 is the disjoint union of cycles of length at most
18;
G2 is the disjoint union of cycles of length at most
6;
G1 and G2 have a perfect matching in common, i.e.
there is a set E ⊆ E(G1 ) ∩ E(G2 ) such that E forms
a perfect matching in both G1 and G2 .
Proof: Conditions 1) and 2) are immediate from our
construction. For 3), we need to show that the m cycles of
G2 are disjoint. We argue that each vertex of G2 appears
in exactly one cycle. First, each vertex vaj gets included in
at least one cycle Zi : if vaj is an endpoint of a positive or a
negative edge e, then e is mapped to some clause Ci , and
so vaj is on the Zi cycle. Otherwise, vaj is not on a positive
or negative edge, and thus we have a ∈ {2, 5, 8, 11, 14, 17}.
Therefore a − 5 ∈ {15, 0, 3, 6, 9, 12} (modulo 18). In this
j
case, va−5 is on a positive or negative edge, and one can
j
check that vaj is in the same cycle Zi as va−5 . To see that no
vertex gets included in two cycles, observe that |V (G2 )| =
|V (G1 )| = 18n = 18(m/3) = 6m. Thus G2 has m cycles of
length 6 spanning 6m vertices, which is only possible if no
two cycles of G2 intersect. For condition 4), there are many
choices, and we exhibit one. Let 1 ≤ j ≤ n and 1 ≤ a ≤ 18.
One can check that in both G1 and G2 , vaj is not a neighbor
j
j
of va+2 . Thus the set E = {vaj va+2 : 1 ≤ j ≤ n, 0 ≤ a ≤ 17}
(addition modulo 18) is a perfect matching of G1 and G2 .
Observe that since G2 is a disjoint union of cycles, it
has 2m perfect matchings since each cycle has 2 possible
perfect matchings. We are now ready to prove the hardness
of MIMP.
13
Proof of Theorem 2: It is easy to see that MIMP is in
NP, since checking that two given matchings are perfect and
share r edges in the two graphs can be done in polynomial
time. Also, G1 and G2 can obviously be constructed in
polynomial time. To prove NP-hardness, we claim that,
using the construction described above, at least r clauses
of C are satisfiable if and only if G1 and G2 admit a perfect
matching with at least r edges in common.
(⇒) Let A be an assignment of x1 , . . . , xn satisfying r
clauses of C . In G1 , for each Xi cycle, take the positive
perfect matching of Xi if xi is positive in A, and the negative
perfect matching otherwise. The result is a perfect matching
M1 of G1 . To construct a perfect matching M2 of G2 , let
Ci be a clause satisfied by the value of a variable xj in A.
Then, take in M2 the perfect matching of Zi that contains
j
the edge vaj va+1 of Xj mapped to Ci . This edge is shared by
both our choices in G1 and G2 . We apply this choice of Zi
perfect matching for each clause Ci satisfied by A (and pick
the matching of Zi0 arbitrarily for an unsatisfied clause Ci0 ).
Since r clauses are satisfied, there must be r common edges
in M1 and M2 .
(⇐) Let M1 and M2 be the perfect matchings of G1 and
G2 with common edges F = E(M1 ) ∩ E(M2 ), |F | ≥ r.
Observe that each edge of F must be either positive or
negative in G1 . We claim that the following assignment A
satisfies at least r clauses: set xi = T rue if F has a positive
edge in Xi , set xi = F alse if F has a negative edge in Xi ,
and set xi arbitrarily if F has no positive or negative edge
in Xi . No xi can be assigned two values since F ⊆ M1
and M1 is in a perfect matching of G1 (and thus cannot
contain both positive and negative edges from the same
cycle Xi ). Let Z = {Zi1 , Zi2 , . . . , Zik } be the clause cycles
of G2 containing an edge of F . For each Zi in G2 , there can
only be one edge of Zi in F (recall that each possible perfect
matching of Zi has exactly one edge in common with G1 ).
Hence since |F | ≥ r, there must be at least r clause cycles in
Z , and thus at least r satisfied clauses.
Hardness of Weighted-SCJ-labeling
Given an MIMP instance G1 and G2 that satisfies the conditions of Lemma 2, we construct a Weighted-SCJ-labeling
instance consisting of a phylogeny T , consistent extant
adjacencies on the leaves of T and weighted ancestral adjacencies on its internal nodes. Let 2n = |V (G1 )| = |V (G2 )|.
The Weighted-SCJ-labeling instance has n markers, and the
phylogeny is a tree T with 4 leaves A, B, C, D and 3 internal
nodes X, Y and Z . The node X is the parent of A and B ,
the node Y is the parent of C and D and the node Z is the
parent of X and Y . Let E ⊆ E(G1 ) ∩ E(G2 ) such that E
forms a perfect matching in both G1 and G2 . The marker
extremities are V (G1 ) = V (G2 ) and the marker head-tail
pairs are given by E (i.e. if xy ∈ E , then x and y are the two
extremities of a marker).
Since G1 is a union of disjoint cycles of even length, its
edges can be partitioned into two disjoint perfect matchings M1,1 and M1,2 . More precisely, let M1,1 be a perfect
matching of G1 . By removing the set of edges M1,1 from
G1 , we obtain another perfect matching M1,2 . The extant
adjacencies for the leaf A are given by M1,1 , and those of B
by M1,2 . Similarly, G2 can be split into two disjoint perfect
matchings M2,1 and M2,2 . The extant adjacencies of leaf C
are given by M2,1 and those of leaf D by M2,2 .
The set of adjacencies under consideration is A :=
E(G1 ) ∪ E(G2 ). In X (respectively Y ), all adjacencies corresponding to an edge of G1 (respectively G2 ) have weight
1. All other adjacencies, including all adjacencies in Z ,
have weight 0. Note that in this construction, any ancestral
adjacency a with non-zero weight can be found in some leaf
of T , which corresponds to the initialization in our method.
Also note that no adjacency connects two extremities
from the same marker. Indeed, all adjacencies on T , extant
or ancestral, are taken from either G1 or G2 , whereas marker
extremities are defined by E , which is taken from the complement of G1 and G2 . Thus there is no marker/adjacency
inconsistency.
We show that if α > 33/34, Weighted SCJ labeling is
NP-hard. This is mainly because we are “forced” to assign
a perfect matching to X and Y , as we show in the next
Lemma. We then need to choose these matchings so as to
maximize their intersection, since this minimizes the SCJ
distance on the XZ and Y Z branches.
Lemma 3. Let α > 33/34. Let λ be any labeling for the
instance T constructed from G1 and G2 as above. Then
there is a labeling λ0 such that the edges corresponding
to λ0 (X) (respectively λ0 (Y )) form a perfect matching of
G1 (respectively G2 ), and D(λ0 , T ) ≤ D(λ, T ).
Proof: Call a cycle C of G1 (resp. G2 ) covered by λ
if λ(X) (resp. λ(Y )) has exactly |V (C)|/2 edges of C , i.e.
λ(X) (resp. λ(Y )) contains a perfect matching of C . If every
cycle of G1 and every cycle of G2 is covered by λ, then
the Lemma holds. Otherwise, we show that if there is an
uncovered cycle C , it is advantageous to modify λ in order
to make C covered.
Thus let C be a cycle of either G1 or G2 that is not covered
by λ, and denote c := |V (C)|. We have c ≤ 18. Let W = X
if C is in G1 , and W = Y if C is in G2 instead. Let M be the
edges of C that correspond to an adjacency labeled by λ(W ).
Let d be the number of vertices of C that do not belong to
an edge of M (i.e. vertices of C that do not belong to any
adjacency of λ(W ), or that belong to an adjacency that does
not correspond to an edge of C ). Since C is not covered we
assume that d ≥ 1. Now, M has (c − d)/2 edges. As c is
even, this implies d ≥ 2 as it must also be even. Obtain λ0
from λ by first removing all adjacencies of λ(W ) including
a vertex of C , then adding to λ(W ) the perfect matching of
C that has the most edges common to M , observing that we
retain at least |M |/2 edges of M (and hence remove at most
|M |/2). Then the number of adjacencies in λ but not in λ0 is
(c−d)/2
18−d
at most d + |M |/2 = d +
= d + c−d
2
4 ≤ d + 4 . On
the other hand at most c/2 − |M |/2 = c/2 − (c − d)/4 =
c+d
18+d
adjacencies are assigned by λ0 but not by λ. As
4 ≤
4
X and Y have 3 incident edges, the increase in the SCJ cost
incurred by changing from λ to λ0 is at most (1 − α) · 3(d +
18−d
+ 18+d
4
4 ) = (1 − α)(3d + 27). On the other hand, the
cost of adjacency weights is decreased by αd/2, since in C ,
|M | = (c − d)/2 edges of weight 1 are removed, and c/2
edges of weight 1 are added, and so d/2 new adjacencies of
weight 1 are gained. The total difference in cost from λ to
λ0 is at most (1 − α)(3d + 27) − αd/2. One can check that
for d >= 2 and α ≥ 33/34, this total difference is no more
14
than 0. Therefore, we can modify λ to cover C at an equal
or better cost. By applying this idea on every cycle of G1
and G2 , we get a labeling λ0 such that λ0 (X) is a perfect
matching of G1 and λ0 (Y ) is a perfect matching of G2 .
We can finally show the hardness of Weighted-SCJlabeling.
Proof of Theorem 1:
To prove NP-hardness of
Weighted-SCJ-labeling, we use the construction described
above, which can easily be seen to be feasible in polynomial
time. We show that G1 and G2 admit a perfect matching
with k common edges if and only if T as constructed above
admits a labeling total cost at most (1 − α)(6n − 2k) + α · 2n
if α > 33/34.
(⇒) Let M1 and M2 be perfect matchings of G1 and
G2 , respectively, that share edge set I = E(M1 ) ∩ E(M2 ),
with |I| ≥ k . We label the internal node X of T with the
adjacencies corresponding to M1 , the internal node Y with
the adjacencies corresponding to M2 , and internal node Z
with the adjacencies corresponding to I . Now, each of the n
adjacencies in λ(X) is found in either λ(A) or λ(B) (but not
both). Thus |λ(X) \ λ(A)| + |λ(X) \ λ(B)| = n. Conversely
|λ(A)\λ(X)|+|λ(B)\λ(X)| = n because half of λ(A)∪λ(B)
is in λ(X). Thus the AX and BX branches, together, induce
an SCJ cost of 2n. Similarly, the CY and DY branches
induce an SCJ cost of 2n. There are n − k adjacencies of
λ(X) not in λ(Z) and all adjacencies of λ(Z) are in λ(X),
and so the XZ branch has SCJ cost n − k . Likewise, Y Z has
SCJ cost n − k . The total SCJ cost is 6n − 2k .
As for the cost of adjacency weights, X has 2n adjacencies of weight 1, and n of them are used by λ(X). The same
holds for Y . Thus X and Y , together, induce an ancestral
adjacency cost of 4n − 2n = 2n. The total weighted cost is
(1 − α)(6n − 2k) + α · 2n as desired.
(⇐) Suppose that G1 and G2 is a “no” instance, i. e. no
pair of perfect matching have k edges in common. We show
that any labeling has cost strictly greater than (1 − α)(6n −
2k) + α · 2n. As per Lemma 3, there is an optimal labeling λ
for T such that λ(X) is a perfect matching of G1 and λ(Y )
is a perfect matching of G2 . Choose such an optimal λ that
labels Z with a minimum number of adjacencies. By the
same argument as above, the branches AX, BX, CY and
DY incur, together, an SCJ cost of 4n. As for XZ and Y Z ,
let a be an adjacency in λ(Z), and recall that a has weight
0 in Z . Then a must be in one of λ(X) or λ(Y ), since if a
is in neither of λ(X) nor λ(Y ), then it should be removed
as λ is suboptimal. If a is in λ(X) but not in λ(Y ), presence
or absence of a in λ(Z) has the same cost, and thus a 6∈
λ(Z) by our choice of λ. Thus λ(Z) ⊆ λ(X) ∩ λ(Y ). Since
G1 and G2 form a “no” instance and X and Y are labeled
by a perfect matching, we must have |λ(X) ∩ λ(Y )| < k ,
and therefore |λ(Z)| < k . The cost incurred by the XZ and
Y Z branches is 2n − 2|λ(Z)| > 2n − 2k . We get that the
total SCJ cost incurred by λ is strictly greater than 6n − 2k .
Exactly 2n of the 4n putative ancestral adjacencies for X
and Y are taken by λ, and so the ancestral adjacency cost is
2n. In total, the cost of the labeling λ is strictly higher than
(1 − α)(6n − 2k) + α(2n), and T is also a “no” instance.
The above analysis, in particular Lemma 3, can certainly
be improved in order to yield better bounds for α. Nevertheless, the question of whether Weighted SCJ labeling is
NP-hard for every 0 < α < 1 is wide open.
| 8cs.DS
|
Innocent Strategies are Sheaves over Plays
arXiv:1409.2764v1 [cs.PL] 9 Sep 2014
Deterministic, Non-deterministic and Probabilistic Innocence
Takeshi Tsukada
C.-H. Luke Ong
University of Oxford
JSPS Postdoctoral Fellow for Research Abroad
[email protected]
University of Oxford
[email protected]
Abstract
study of the game semantics of non-deterministic lambda calculus was motivated, in particular, by a desire to introduce abstraction refinement to higher-order model checking based on the nondeterministic λY-calculus.
Let us begin with a quick overview of the HO/N-style games.
Types are interpreted as arenas, and programs of a given type are
interpreted as P-strategies for playing in the arena that denotes the
type. Recall that an arena A is a set of moves MA equipped with
an enabling relation, (`A ) ⊆ (MA ∪ {?}) × MA , that gives A
the structure of a forest (whereby a move m is a root, called initial, just if ? `A m); furthermore, moves on levels 0, 2, 4, . . . of
the forest are O-moves, and those that are on levels 1, 3, 5, . . . are
P-moves. A justified sequence of A is a finite sequence of O/Palternating moves, m1 m2 m3 . . . mn , such that each non-initial
move mj has a pointer to an earlier move mi (called the justifier
of mi ) such that mi `A mj . A key notion of HO/N games is the
view of a justified sequence: the P-view of a justified sequence s
is a certain justified subsequence, written dse, consisting of moveoccurrences which P considers relevant for determining his next
move (similarly for the O-view bsc of s). A play then is a justified
sequence, m1 m2 m3 . . ., that satisfies Visibility: for every i, if mi
is non-initial then its justifier appears in dm1 m2 . . . mi e (respectively bm1 m2 . . . mi c) if mi is a P-move (respectively O-move). A
strategy σ over an arena A is just a prefix-closed set of even-length
P
plays s; σ is said to be deterministic if whenever s mP
1 , s m2 ∈ σ,
P
then mP
=
m
.
(We
use
superscript
P
to
indicate
a
P-move;
sim1
2
ilarly for O-move.) Recall that a strategy σ is said to be innocent if
it is view dependent i.e. for all s ∈ σ
Although the HO/N games are fully abstract for PCF, the traditional
notion of innocence (which underpins these games) is not satisfactory for such language features as non-determinism and probabilistic branching, in that there are stateless terms that are not innocent.
Based on a category of P-visible plays with a notion of embedding as morphisms, we propose a natural generalisation by viewing innocent strategies as sheaves over (a site of) plays, echoing a
slogan of Hirschowitz and Pous. Our approach gives rise to fully
complete game models in each of the three cases of deterministic,
nondeterministic and probabilistic branching. To our knowledge, in
the second and third cases, ours are the first such factorisation-free
constructions.
1.
Introduction
Game semantics is a powerful paradigm for giving semantics to
a variety of programming languages and logical systems. Both
HO/N games [10, 14] (based on arenas and innocent strategies) and
AJM games [2] (based on games equipped with a certain equivalence relation on plays, and history-free strategies) gave rise to
the first syntax-independent description of the fully abstract model
for the functional programming language PCF. The HO/N-style
games, based on arenas and history-sensitive strategies, have been
extended to give a fully abstract model for Idealised Algol (PCF
extended with locally-scoped references) [1]. Definability, a crucial step of the completeness argument, was established by showing that every compact history-sensitive strategy factorises through
an innocent strategy. Using the same factorisation technique, fully
abstract HO/N-style game models have been constructed for a spectrum of Algol-like languages, including Idealised Algol augmented
with language features such as non-determinism [8] and probabilistic branching [5].
Perhaps surprisingly, it is problematic to extend innocent strategies to model PCF extended with non-determinism [7]. A famous
game model by Harmer [7] is based on factorisation, decomposing a given non-deterministic strategy into a non-deterministic oracle and a deterministic innocent strategy. To our knowledge, the
problem of a factorisation-free fully complete game model for the
simply-typed non-deterministic lambda calculus is open; the same
problem is also open for lambda calculus augmented with probabilistic branching. This paper presents a new approach to innocent strategies, based on sheaves over a site of plays, that yields
fully complete game models for lambda calculi extended with these
branching constructs.
We are interested in the simply-typed lambda calculi because
they have good algorithmic properties, notably, the decidability
of compositional higher-order model checking [15, 18], which is
proved using HO/N-style effect arenas and innocent strategies. Our
P
O
P
(s ∈ σ ∧ ds mO
1 m2 e ∈ σ) ⇐⇒ s m1 m2 ∈ σ
(1)
It is an important property of innocence that—in the sets-of-plays
presentation of strategies—every deterministic innocent strategy
can be generated by the set of P-views contained in it. The category
of arenas and innocent strategies gives rise to a fully complete
model of the simply-typed lambda calculus [10].
However, as Harmer observed in his thesis [7], the notion of
innocence breaks down when one tries to use it to model (stateless)
non-deterministic functional computation.
Example 1. Take simply-typed λ-terms tt := λxy.x and ff :=
λxy.y of type B = o → o → o, and M1 := λf.f (tt + ff)
and M2 := (λf.f tt) + (λf.f ff) of type (B → o) → o, where
+ is the construct for non-deterministic branching. Assuming the
call-by-name evaluation strategy, these terms can be separated by
the term N := λg.g (g ⊥ z) ⊥, where ⊥ is the divergence term,
i.e. M1 N may converge but M2 N always diverges. In the HO/N
game model (see, for example, [7]), σi := [[Mi ]] are strategies over
the arena (({d} → {d0 } → {c}) → {b}) → {a}, for i = 1, 2.
Note that σ1 and σ2 are distinct as strategies: for example (we omit
pointers from the plays as they can be uniquely reconstructible)
1
2018/2/15
a b c d c d0 ∈ (σ1 \σ2 ). However σ1 and σ2 contain the same set of
non-empty, even-length P-views, namely, {a b, a b c d, a b c d0 }.
of the jointly surjective families of morphisms with codomain s,
then (Obj (PA ), ≤, .) is a preorder presentation, as discussed in
the preceding. In our sheaf-theoretic approach, an innocent strategy
of arena A, whether deterministic or not, is a sheaf σ over the site
(PA , J). The intuition is that a sheaf σ : Pop
that maps
A →
every s to either a singleton set or the emptyset (which is so if
the strategy σ is deterministic) corresponds to an ideal Iσ of the
associated preorder presentation whereby s ∈ Iσ if and only if
σ(s) 6= ∅.
The preceding example shows the sets-of-plays approach works
well for expressing, and even composing, non-deterministic strategies for stateless programs; the only problem is that, in general, the
set of P-views cannot be a good generator for these strategies.
The problematic term is M2 . It applies the argument f to tt
or ff, non-deterministically, but the branch has already been chosen
when M2 responds to the initial move. So a b c d c d0 is not playable
by M2 , although innocence requires it to.
Our approach is to admit that M2 has two possible responses
to the initial move: they give the same play a b but have different
internal states. Thus a strategy is formally a mapping from plays
to sets that represent the internal states. For example, [[M2 ]](a b) =
{tt, ff}, where tt means the left branch and ff the right branch. Now
the P-views for [[M2 ]] are, say, {a btt , a btt c d, a bff , a bff c d0 }.
Notice that a btt c d c d0 and a bff c d c d0 are no longer forced by innocence to be admissible plays. From this viewpoint, a deterministic strategy is a mapping from plays to empty or singleton sets.
In what follows, we discuss how to formalise this idea.
Set
Set
Our contributions Our thesis is that sheaves Pop
genA →
eralise innocent strategies of the arena A. (Indeed the sheaves approach seems more general than innocence, since it appears capable
of capturing the computation of single-threaded (history-sensitive)
strategies as well.)
Given arenas A, B and C, we define a category IA,B,C whose
objects are interaction sequences of the triple (A, B, C) in the
usual sense, and whose morphisms f : u → u0 are injective maps
that preserve moves, justification pointers, and basic blocks (which
are sequences of moves that begin with an O-move of A ⇒ C, and
end with a P-move of A ⇒ C, with all intermediate moves from
B). Let u ∈ IA,B,C , we write uA,B , uB,C and uA,C for the
standard projections of u to the component arenas. Given sheaves
σ1 : Pop
and σ2 : Pop
, there is a natural way to
A,B →
B,C →
compose them. (We write PA,B to mean PA⇒B .) Define a presheaf
σ1 ; σ2 : Pop
, which acts on objects as follows:
A,C →
a
(σ1 ; σ2 )(s) :=
σ1 (uA,B ) × σ1 (uA,B )
Ideal-based innocence Before we explain the main ideas behind
our sheaf-theoretic approach to innocence, it is helpful to consider
a category of plays PA , and an alternative view of deterministic
innocent strategies as ideals of a preorder presentation [11]. The
objects of the category PA are (even-length) justified sequences
of the arena A satisfying O/P-alternation and P-visibility (but not
necessarily O-visibility), which we shall henceforth call plays (by
abuse of language). The morphisms f : s → s0 are injective maps
that preserve moves, justification pointers, and pairs of consecutive
O-P moves. A morphism can permute such pairs, provided the
pointers are respected. For example, for each play s, there are
P
morphisms dse → s and s → s mO
1 m2 .
A preorder presentation is a triple (P, ≤, .) where (P, ≤) is a
preorder and . ⊆ P(P ) × P is called a covering relation (we read
U . s as “U covers s”). A subset I ⊆ P is called an ideal if (I1) I
is lower-closed i.e. if t ∈ I and s ≤ t then s ∈ I, and (I2) for every
covering U .s, if U ⊆ I then s ∈ I. A preorder presentation can be
extracted from the category PA , namely, (Obj (PA ), ≤, .) whereby
s ≤ s0 just if there is a morphism f : s → s0 , and U . s just if
U = {sξ }ξ∈Ξ for some family of morphisms, {fξ : sξ → s}ξ∈Ξ ,
which is jointly surjective, meaning that the union of the set of
move-occurrences that appear in the image of fξ , as ξ ranges over
Ξ, is the set of all move-ocurrences in the play s.
Then ideals of the preorder presentation (Obj (PA ), ≤, .) are
P
innocent strategies. Notice that, because s ≤ s mO
1 m2 and
P
O
P
ds mO
m
e
≤
s
m
m
,
condition
(I1)
of
ideal
gives
the
⇐1
2
1
2
P
direction of (1). Further since the set {s, ds mO
1 m2 e} covers
P
s mO
1 m2 , condition (I2) gives the other direction of (1).
C
Set
Set
u∈IA,B,C :uA,C =s
We show that the composite σ1 ; σ2 is well-defined:
(i) σ1 ; σ2 : Pop
A,C →
Set
(iii) composition is associative up to natural isomorphism:
(σ1 ; σ2 ); σ3 ∼
= σ1 ; (σ2 ; σ3 )
Furthermore, the category whose objects are arenas and whose
morphisms σ : A → B are (equivalence classes of isomorphic)
sheaves σ : Pop
is cartesian closed.
A,B →
Just as innocent strategies are view dependent, so there is a
compelling sense in which sheaves on plays, Pop
, depend
A,B →
on (indeed, are determined by) sheaves on views, Vop
,
A,B →
where VA,B is a full subcategory of PA,B . The subcategory VA,B ,
whose objects are nonempty P-views, is a preorder, and the induced
topology is trivial (every object has a unique covering sieve which
is maximal). Since every object in PA,B has a covering sieve by
objects of the subcategory VA,B , thanks to the Comparison Lemma
[3, 19], ι∗ : Sh(PA,B ) → Sh(VA,B ) gives an equivalence of the
respective categories (of sheaves), where ι : VA,B ,→ PA,B is the
embedding.
Sheaves on views are important because they are easier to understand and calculate with than sheaves on plays; conversely, composition of the latter is easier to describe than that of the former.
Let τM : Vop
be the denotation of a non-deterministic
A →
λ-term M . Then given p ∈ VA , τM (p) corresponds to the set of
all possible runs (qua plays) of M whose P-view is p. Returning to
Example 1:
Set
Set
Set
C
C
C
C
C
Set is a sheaf
(ii) σ1 ; σ2 is the left Kan extension of the functor F : Iop
A,B,C →
, whose action on objects is u 7→ σ1 (uA,B )×σ2 (uB,C ),
op
along the projection functor Iop
A,B,C → PA,C .
op
From ideals to sheaves A presheaf, F :
→
, is a
contravariant functor, assigning data (a set of “internal states”) to
each object s of . The definition of sheaf of a site is technical,
and a version is presented in the preliminaries subsection. Here
we can think of a sheaf over a site as an extension of the notion
of an ideal of a preorder presentation. A site is a pair ( , J)
where
is a category, and J, called a coverage, assigns to each
object s of
a collection of covering families, each of the form
{fξ : sξ → s}ξ∈Ξ . Intuitively a presheaf, F : op →
, is
a sheaf over the site ( , J) just if the data assigned to a given
object s (meaning the elements of F (s)) can be systematically
tracked by the data locally defined over the family {fξ : sξ →
s}ξ∈Ξ (meaning the elements of F (sξ ), as ξ ranges over Ξ), for
all covering families of s; further, every matching family of such
locally assigned data uniquely determines a datum assigned to s (an
element of F (s)). Thus, take the site (PA , J) where J(s) consists
C
Set
Set
Set
Set
Example 2. Using the notation in Example 1, let p0 = a b,
p1 = a b c d and p2 = a b c d0 . For i = 1, 2 define τi ∈
Sh(V({d}→{d0 }→{c})→{b},{a}) ) to be the sheaf-over-views de-
2
2018/2/15
notation of Mi . Then
τ1 (p0 ) = {x1 }
τ1 (p1 ) = {y1 }
τ1 (p2 ) = {z1 }
presheaf is a sheaf for a site if it is a sheaf for every covering family
of the site.
A sieve on an object s in a category is a family of maps with
codomain s that are closed under precomposition with maps in .
Given a family {fξ : sξ → s}ξ∈Ξ , the sieve it generates is the
family of all maps g : t → s with codomain s that factor through
some fξ . A presheaf is a sheaf for a family {fξ : sξ → s}ξ∈Ξ if,
and only if, it is a sheaf for the sieve it generates. If S is a sieve
on s and g : t → s is a map, we define g ∗ (S) to be the sieve on
t consisting of all maps h with codomain t such that g ◦ h factors
through some map in S.
A Grothendieck topology is a map J that assigns to each object
s of a collection J(s) of sieves on s, called covering sieves, that
satisfies the following:
C
τ2 (p0 ) = {x21 , x22 }
τ2 (p1 ) = {y21
}
τ2 (p2 ) = {
z22 }
Notice that in the set of plays [[M2 ]], there are two independent
plays (which have the P-view) p0 .
Our approach gives rise to fully complete game models in each
of the three cases of deterministic, nondeterministic and probabilistic branching. To our knowledge, in the second and third cases, ours
are the first such factorisation-free constructions.
C
Related work The standard notion of innocence does not work
well for certain language features, such as non-determinism. To
address the deficiency, Levy [13] proposed a category of P-visible
plays and viewing morphisms. This is essentially our category PA
of plays. However in op. cit. an innocent strategy σ is still defined
to be a certain set of plays, namely, a lower-closed set of objects
of the category: if t ∈ σ and s → t is a morphism, then s ∈ σ.
Because this definition captures only one of the two requirements
of innocence (i.e. ⇐ of (1)), Levy’s construction will likely not
yield accurate (fully complete) models of the non-deterministic λcalculus.
A related approach by Hirschowitz et al. [6, 9] does view strategies as presheaves (and sheaves) on a category of plays. However, in contrast to our focus on higher-type computation, they
are concerned with CCS-style concurrent computation which they
model as multi-player games. Strategies are presheaves on a category of plays EX over a position X, and a strategy is deemed
innocent if it is determined by its restriction to a subcategory of
views VX ,→ EX . A position is an undirected graph describing
the channel-based communication topology connecting the players, and plays are certain “glueings” of moves over a position, with
moves built-up using CCS constructs. Thus the connexions with
our work seem superficial.
Winskel et al. [16, 17] have worked extensively on causal games
as models of true concurrency, from the viewpoint of strategies as
event structures with symmetries. Recently Clairambault et al. [4]
built a conservative extension of HO/N games in a truly concurrent
framework. An extensional quotient of their model yields a fully
abstract model of PCF with parallel or.
Perhaps surprisingly, the question of what is the proper notion
of innocence in the presence of non-determinism is still open.
Harmer and McCusker [8] seem only concerned with stateful nondeterministic programs, namely non-deterministic Idealised Algol.
(i) The maximal sieve, {f | cod(f ) = s}, is in J(s).
(ii) (Stability) If S ∈ J(s) then h∗ (S) ∈ J(t) for every map
h : t → s.
(iii) (Transitivity) If S ∈ J(s) and R is a sieve on s such that
h∗ (R) ∈ J(t) for every h : t → s in S, then R ∈ J(s).
Lemma 3. For every coverage, there is a unique Grothendieck
topology that has the same sheaves.
N
Notation We write
for the set of all positive integers. For an
integer n, we define [n] := {k | 1 ≤ k ≤ n} and [n]0 := {k |
0 ≤ k ≤ n}. For a category C, we write x ∈ C to mean that x is
an object of C.
2.
Set
Sites of Plays
This section defines sites of plays over an arena. The innocent
strategies are just sheaves over those sites. The category of plays
has a subcategory of views. We prove that the sheaves over plays is
equivalent to sheaves over views: this generalises view dependency
to non-deterministic computation.
2.1
Plays
The definition of arenas is standard (as in [10]) except that all
moves are questions.
Definition 4 (Arena). An arena is a tuple A = (MA , λA , `A ),
where MA is a finite set of moves, λA : MA → {P, O} is
an ownership function and (`A ) ⊆ ({?} + MA ) × MA is an
enabling relation that satisfies the following conditions: (1) for
every m ∈ MA , there is a unique x ∈ {?} + MA such that
x `A m, and (2) if ? `A m, then λA (m) = O. If m `A m0 , then
λA (m) 6= λA (m0 ).
Technical preliminaries In the following we review the basic
definitions of coverage, Grothendieck topology and sheaves, and
refer the reader to the book [12] for an exposition.
A coverage on a category C is a map J assigning to each object
s of C a collection J(s) of families {fξ : sξ → s}ξ∈Ξ of maps
with codomain s, called covering families, such that the system of
families is “stable under pullback”, meaning: if {fξ : sξ → s}ξ∈Ξ
is a covering family and g : t → s is a map, then there is a covering
family, {hν : tν → t}ν∈N , such that each g ◦ hν factors through
some fξ . A number of saturation conditions are often imposed on
a coverage for convenience. A site is a category equipped with a
coverage J, written ( , J).
Given a family S = {fξ : sξ → s}ξ∈Ξ of maps with codomain
s, and a presheaf F : op →
, a family of elements {xξ ∈
F (sξ )}ξ∈Ξ is said to be matching for S if for all maps g : t → sξ
and h : t → sξ0 , if fξ ◦ g = fξ0 ◦ h then F (g)(xξ ) = F (h)(xξ0 ).
An amalgamation for the family {xξ ∈ F (sξ )}ξ∈Ξ is an x ∈ F (s)
such that F (fξ )(x) = xξ for every ξ ∈ Ξ. A presheaf F : op →
is a sheaf for a family S = {fξ : sξ → s}ξ∈Ξ of maps
just if every matching family for S has a unique amalgamation. A
C
C
C
For an arena A, the set MO
A of O-moves is defined as {m ∈
MA | λA (m) = O}. The set of P-moves is defined by MP
A :=
{m ∈ MA | λ(m) = P}. A move m is initial if ? `A m. An
arena is prime if it has exactly one initial move.
We write {m} for the arena that has one O-move m and no Pmoves. For a prime arena A and an arena B, B → A is the arena
whose moves are MA + MB where the initial B-move is enabled
by the unique initial A-move. For example, {m1 } → {m2 } →
{m3 } consists of an O-move m3 and P-moves m1 and m2 with
? ` m3 , m3 ` m1 and m3 ` m2 .
Unlike the standard formalisation, in which notions such as
justified sequences and plays are parametrised by arenas, we
parametrise them by a pair of arenas (A, B), corresponding to
the exponential arena A ⇒ B in the standard formalisation. This
change simplifies some definitions.
C
Set
C
Definition 5 (Arena pair). Let A = (MA , λA , `A ) and B =
(MB , λB , `B ) be arenas. The moves of (A, B) is the disjoint
union of moves, say MA,B := MA + MB . We define P-moves
3
2018/2/15
O
P
O
P
by MP
A,B := MA + MB and O-moves by MA,B := MA +
0
0
MO
.
For
m,
m
∈
M
,
we
write
m
`
m
just
if
either
A,B
A,B
B
(1) m, m0 ∈ MA and m `A m0 , or (2) m, m0 ∈ MB and
0
m `B m , or (3) ? `B m ∈ MB and ? `A m0 ∈ MA . We
write ? `A,B m just if ? `B m ∈ MB .
we introduce a richer structure to plays, organising them into a category. This is essentially the category introduced by Levy [13].
It is useful to view an even-length alternating justified sequence
is a sequence of pairs of O- and P-moves, which we shall call a
block (or an O-P block).
For a pair (A, B), an initial A-move is a move m ∈ MA ⊆
MA,B such that ? `A m: do not confuse it with ? `A,B m, which
is impossible. An initial B-move is defined similarly.
Definition 10 (Morphism between plays). Let m1 . . . mn and
m01 . . . m0n0 be plays of length n and n0 , respectively. A morphism
between plays is an injection f : [n] → [n0 ] s.t. for every k ∈ [n]
Definition 6 (Justified sequence). Let (A, B) be a pair of arenas. A
justified sequence of (A, B) is a finite sequence of moves equipped
with justification pointers. Formally it is a pair of functions s :
[n] → MA,B and ϕ : [n] → [n]0 (for some n) such that
(i) mk = m0f (k) (as moves),
(ii) mi x mk implies m0f (i) x m0f (k) (and similarly for ? x
mk ), and
(iii) if an O-move mk is followed by a P-move mk+1 , then m0f (k)
is followed by m0f (k+1) (i.e. f (2l − 1) + 1 = f (2l) for all l).
• ϕ(k) < k for every k ∈ [n], and
• ϕ respects the enabling relation: ϕ(k) 6= 0 implies s(ϕ(k)) `A,B
I.e. a morphism between plays is an injective map between O-P
blocks that preserves moves and justification pointers. We define
img(f ) := {f (k) | k ∈ [n]} ⊆ [n0 ].
s(k), and ϕ(k) = 0 implies ? `A,B s(k).
As usual, by abuse of notation, we often write m1 m2 . . . mn for
a justified sequence such that s(i) = mi for every i, leaving
the justification pointers implicit. Further we use m and mi as
metavariables of occurrences of moves in justified sequences. We
write mi x mj if ϕ(j) = i > 0 and ? x mj if ϕ(j) = 0. We
call mi the justifier of mj when mi x mj . We write x+ for the
transitive closure of x. We write |s| for the length of s.
Example 11. (i) Let s = m1 . . . mn be a play and s0 = m1 . . . ml
be its (even-length) prefix. Then f (i) = i (for i ≤ l) is a morphism
f : s0 → s. In other words, each prefix s0 ≤ s induces a
morphism. (But this may not be the unique morphism of s0 → s.)
(ii) Let s = m1 . . . mn−1 mn and assume that mk x mO
n−1 .
Then we have f : (m1 . . . mk mn−1 mn ) → s, where f (i) = i
(if i ≤ k), f (k + 1) = n − 1 and f (k + 2) = n. (iii) For
every play s, we have a unique morphism f : dse → s that
maps the last move of dse to the last move of s (though there may
exist another morphism that does not satisfy this condition). In this
sense, the morphisms of the category is an generalisation of the
notion of P-views. (iv) Let s = s0 mn−3 mn−2 mn−1 mn be a
play and assume that the justifier of mn−1 is not mn−2 . Let s0 be
the play s0 mn−1 mn mn−3 mn−2 obtained from s by commuting
O-P blocks mn−3 mn−2 and mn−1 mn . There is an isomorphism
f : s → s0 , given by f (i) = i (if i < n − 3), f (n − 3) = n − 1,
f (n − 2) = n, f (n − 1) = n − 3 and f (n) = n − 2.
It is convenient to relax the domain [n] of justified sequences
to arbitrary linearly-ordered finite sets such as a subset of [n]. For
example, given a justified sequence (s : [n] → MA,B , ϕ : [n] →
[n]0 ), consider a subset I ⊆ [n] that respects the justification
pointers, i.e. k ∈ I implies ϕ(k) ∈ I ∪ {0}. Then the restriction
(sI : I → MA,B , ϕI : I → {0} ∪ I) is a justified sequence
in the relaxed sense. Through the unique monotone bijection α :
I → [n0 ], we identify the restriction with the justified sequence in
the narrow sense.
A justified sequence is alternating if s(k) ∈ MO
A,B iff k is odd
(so s(k) ∈ MP
A,B iff k is even).
Example 12. Let (A, B) = (({d} → {c}) → {b}, {a}) be a pair
of arenas and the play s = a b c d c0 d0 (where c = c0 and d = d0 as
moves) over (A, B) in which c0 points to b and all other moves are
justified by their preceding move. Let s0 = a b c d be another play
of (A, B). Then s0 can be regarded as a prefix of s and as the P-view
dse of s. The first perspective induces the morphism f : s0 → s,
where f (i) = i (for i ∈ [4]), and the second perspective does
g : s0 → s, where g(1) = 1, g(2) = 2, g(3) = 5 and g(4) = 6.
Definition 7 (P-View/P-visibility). Let m1 . . . mn be an alternating justified sequence over (A, B). Its P-view dm1 . . . mn e (or
simply view) is a subsequence defined inductively by:
dm1 . . . mn e := dm1 . . . mn−1 e mn
dm1 . . . mn e := mn
dm1 . . . mn e := dm1 . . . mk e mn
(if mn ∈ MP
A,B )
(if ? x mn ∈ MO
A,B )
(if mk x mn ∈ MO
A,B ).
Definition 13 (Category of plays). Let A and B be arenas. The
category PA,B of plays has plays of (A, B) as objects and as
morphisms those defined above.
More formally, given an alternating justified sequence s of length
n, its view is a subset I ⊆ [n]. The above equation gives the
restriction of s to I. A view is, in general, not a justified sequence
since the justifier of a move may have been removed.
Let mk be a P-move in the sequence. Its justifier is said to be Pvisible if it is in dm1 . . . mk e. An alternating justified sequence is
P-visible if the justifier of each P-move occurrence in s is P-visible.
Lemma 14. PA,B has pullbacks.
Proof. Let f : s1 → t and g : s2 → t. They are injective maps
f : [|s1 |] → [|t|] and g : [|s2 |] → [|t|]. Let I = img(f ) ∩ img(g).
The restriction of t to I is the pullback s1 ×t s2 .
Definition 8 (Play). An alternating justified sequence over a pair
(A, B) of arenas is a play just if it is P-visible and its last move is
a P-move m ∈ MP
A,B .
We give another definition of morphisms via commutation.
Definition 15 (Commutation of non-interfering blocks). Let s
be an even-length alternating justified sequence over (A, B). Let
m1 m01 m2 m02 be an adjacent pair of O-P blocks in s, i.e. s =
t m1 m01 m2 m02 t0 , where m1 and m2 are O-moves. We say that
the pairs are non-interfering if the justifier of m2 is not m01 . The
commuted sequence s0 is defined by s0 := t m2 m02 m1 m01 t0 (in
which the justification pointers are modified accordingly).
Remark 9. In contrast to the standard definition of play in innocent
game semantics (as in [10]), we do not require O-visibility. This
is technically convenient because O-visibility is not preserved by
commutations (see Definition 15). Note also that a play may have
several initial moves, i.e. we do not assume well-openness.
2.2
Morphisms between plays that respects P-views
A commuted sequence is not always a justified sequence: if
m02 is justified by m1 , then m02 in the commuted sequence is not
well-justified. If the justified sequence is P-visible, the commuted
In the traditional HO/N game models, the set of plays are considered as a poset ordered by the prefix ordering. In this subsection,
4
2018/2/15
(ii) For a play s = m1 . . . mn , the family {fk : dm1 . . . mk e →
s}k∈{2,4,...,n} is a covering family. Here fk is the composite of the
P-view embedding and the prefix embedding, i.e.,
sequence is a justified sequence. Furthermore the converse also
holds.
Lemma 16. Let P be a set of even-length alternating justified sequences over (A, B). Suppose that P is closed under commutations, i.e. for every sequence s ∈ P and every non-interfering adjacent pairs of blocks in s, the commuted sequence is also in P .
Then all justified sequences in P are plays.
f
k
s = dm1 . . . mk e −→ (m1 . . . mk ) −→ s.
dm1 . . . mk e −→
The covering family generalises the set of P-views of the prefixes.
(iii) The covering family is finer than the set of P-views. Let
s = m1 m2 m01 m02 (the repetition of m1 m2 twice). Then {f :
m1 m2 → s}, where f (1) = 1 and f (2) = 2, is not a covering
family. However {f : m1 m2 → s, g : m1 m2 → s}, where
g(1) = 3 and g(2) = 4, is a covering family. Notice that those two
families have the same set of the domain, say {m1 m2 }, which is
the set of P-views of s.
Proof. Let s = m1 . . . mn ∈ P and mk be a P-move occurrence in
s. We prove that the justifier of mk is in the P-view dm1 . . . mk e.
By commuting pairs as much as required, we can reach a sequence,
say s0 = m01 . . . m0n , such that m0l is the move corresponding
to mk in s and m02i x m02i+1 for every 2i < l. This means
that dm01 . . . m0l e = m01 . . . m0l and hence the justifier of m0l is
in the view. Since P-visibility is preserved and reflected by the
commutation of non-interfering blocks, s is P-visible.
Definition 21. An innocent strategy is a sheaf over PA,B .
Set
Remark 22. Let σ be a functor Pop
. It is pre-deterministic
A,B →
if σ(s) is empty or singleton for every s. A pre-deterministic functor can be determined by the set Pσ = {s ∈ PA,B | σ(s) 6= ∅}.
Since σ is a functor, the set Pσ is lower closed, i.e. s ∈ Pσ and
f : s0 → s in PA,B implies s0 ∈ σ. A pre-deterministic functor σ
is a sheaf just if s = m1 . . . mn ∈ Pσ iff {dm1 . . . mk e | k =
2, 4, . . . , n} ⊆ Pσ . To see this, observe that {dm1 . . . mk e →
s | k = 2, 4, . . . , n} is a covering family and the family of
unique elements {xk ∈ σ(dm1 . . . mk e)}k is a matching family
and thus there is an amalgamation x ∈ σ(s). In this sense, for
pre-deterministic strategies, the innocence is equivalent to the sheaf
condition. However, if σ(s) may have more than one element, innocence based on the set of views differs from the sheaf condition.
Every morphism can be expressed as the prefix embedding followed by commutations. This is insightful and technically useful.
Lemma 17. Every f : s → t in PA,B can be decomposed as
f
≤
g
g
gn
1
2
s−
→t = s−
→ t0 −→
t1 −→
... −
−
→ tn ,
where n ≥ 0, tn = t and gi is a commutation of adjacent O-P
blocks in ti−1 for every i ∈ [n]. (This decomposition is not unique.)
Proof. Let f : s → t and t = m1 . . . mn . If f is induced by
the prefix, then we complete the proof. Otherwise, there is an odd
number k ≤ |s| such that either f (k) − 2 ∈
/ img(f ) or f (l) =
f (k)−2 for some l > k. Then we claim that mf (k)−2 mf (k)−1 and
mf (k) mf (k)+1 in t is a non-interfering pair. Suppose otherwise,
i.e. the justifier of mf (k) is mf (k)−1 . Then f (k) − 1 ∈ img(f )
since f preserves the justification pointer. Let l0 ≤ |s| be the index
such that f (l0 ) = f (k) − 1. Since s(l0 ) x s(k), we have l0 < k.
Because l0 is even, we have f (l0 − 1) = f (l0 ) − 1 = f (k) − 2. In
summary, we have l < k such that f (l) = f (k) − 2, that contradict
the assumption. So the adjacent O-P blocks mf (k)−2 mf (k)−1 and
mf (k) mf (k)+1 in t is non-interfering.
Consider the commutation h : t → t0 and the inverse h−1 :
0
t → t, which is also a commutation. By applying the same
argument to h ◦ f : s → t0 , h ◦ f can be decomposed as
gn ◦ · · · ◦ g1 ◦ g0 , where g0 is induced by the prefix and gi
(i > 1) is a commutation. This inductive argument is justified
by the same way as the termination of the bubble sort. Then f =
h−1 ◦ gn ◦ · · · ◦ g1 ◦ g0 .
2.4
Definition 23 (Subcategory of P-views). A play s ∈ PA,B is a
P-view if dse = s and s is not empty. We use p as a metavariable
ranging over P-views. The category of P-views VA,B is the full
subcategory of PA,B consisting of P-views. We write ι : VA,B ,→
PA,B for the embedding. Henceforth we fix the topology for VA,B
to be that induced1 from PA,B : it is the trivial topology, i.e. every
P-view has only one covering sieve, namely, the maximal sieve.
Remark 18. Let σ be an innocent strategy in the standard sense,
i.e. an even-prefix closed subset of plays with a certain condition.
Then s ∈ σ and f : s0 → s in PA,B implies s0 ∈ σ. To see this,
observe that a commutation of s ∈ σ is in σ and use Lemma 17.
2.3
Sheaves over PA,B and its restriction to P-views
In innocent game models for deterministic calculi (such as [10]),
one often considers the restriction of strategies to P-views. A remarkable property is that an innocent strategy (qua set of plays) is
completely determined by the subset of P-views it contains. After
all, innocence means view dependence.
In this subsection, we shall see that a similar property holds
for sheaves over plays PA,B . This property comes from the topological structure of plays: every play is covered by P-views (see
Example 20(ii)). This observation gives a justification of defining
innocent strategies as sheaves.
The category of P-views is a poset. We write (p0 ≤ p) and
(p ≥ p0 ) for the unique morphism f : p0 → p (if it exists).
Because the topology is trivial, a sheaf over VA,B is just a
functor Vop
. A sheaf σ ∈ Sh(PA,B ) induces a sheaf
A,B →
σ ◦ ι over VA,B . The strategy σ can be reconstructed from the
restriction to P-views σ ◦ ι (up to natural isomorphism).
Topology of PA,B
Set
As for the innocent strategies σ for deterministic calculi, which is
a set of plays, a play s = m1 . . . mk is in the strategy σ iff P-views
for (even-)prefixes are in σ, i.e. {dm1 . . . mk e | k = 2, 4, . . . n} ⊆
σ. We use the Grothendieck topology to capture this condition.
Lemma 24 (Comparison). The functor ι∗ : Sh(PA,B ) 3 σ 7→
σ ◦ ι ∈ Sh(VA,B ) induces an equivalence of categories.
Definition 19 (Covering family / sieve). A family of morphisms
{fξ : sξ → Ss}ξ∈Ξ is said to cover s when they are jointly
surjective, i.e. ξ∈Ξ img(fξ ) = [n], where n is the length of s.
A covering sieve is a sieve that is a covering family. By abuse of
notation, we write PA,B for the site associated with this topology.
Since every play has a covering by P-views, Lemma 24 follows
from a standard result, known as the Comparison Lemma [19]
(see, for example, [3, Prop. p. 721] which generalises the classical
Example 20. (i) For a play s = m1 . . . mn , the family {f :
(m1 . . . mn−2 ) → s, g : dse → s} is a covering family. Here
f is induced by the prefix and g by the P-view (see Example 11).
1 Given a site C and a full subcategory D ,→ C, the induced topology on
D is defined by: a sieve S on D is covering iff the sieve (S) := {f ◦ h |
f ∈ S, dom(f ) = codom(h)} on C generated from S is covering.
5
2018/2/15
result in SGA4). However an explicit description of the adjoint
ι∗ : Sh(VA,B ) → Sh(PA,B ) is insightful and worth clarifying.
Let τ ∈ Sh(VA,B ) be a sheaf over P-views. Let s = m1 . . . mn
be a non-empty play and pk := dm1 . . . mk e for every even number k. We define a set of τ -annotations for s: a τ -annotation is a
sequence e2 e4 . . . en , where ek ∈ τ (pk ) for every even number k,
subject to the following condition: for every even number k ≤ n,
O
if mP
l x mk−1 , then el = τ (pl ≤ pk )(ek ). For a non-empty play
s ∈ PA,B , we write (ι∗ τ )(s) for the set of all τ -annotations.
Given f : s → s0 , which is an injective map f : [|s|] → [|s0 |],
the morphism (ι∗ τ )(f ) : (ι∗ τ )(s0 ) → (ι∗ τ )(s) is defined by:
associative up to isomorphism, and the arenas and sheaves form
a CCC (where isomorphic sheaves are identified).
(ι∗ τ )(f ) : e2 e4 . . . e|s0 | 7→ ef (2) ef (4) . . . ef (|s|) .
A justified sequence of the triple is a sequence over MA + MB +
MC equipped with justification pointers that respect the enabling
relation `A,B,C .
3.1
•
•
•
•
We define (ι∗ τ )(ε) := {∗} for the empty sequence. Then ι∗ τ :
Pop
is a functor.
A,B →
Set
= {x1 }
= {y1 }
= {z1 }
= x1
= x1
τ2 (p0 )
τ2 (p1 )
τ2 (p2 )
τ2 (f )(y21 )
τ2 (g)(z22 )
= {x21 , x22 }
= {y21
}
={
z22 }
= x21
= x22 ,
Definition 29 (Interaction sequence). Let (A, B, C) be a triple
of arenas. A justified sequence s over (A, B, C) is an interaction
sequence if
where f : (a b) → (a b c d) and g : (a b) → (a b c d0 ). Then
(ι∗ τ1 )(a b c d c d0 ) = {x1 y1 z1 }
(ι∗ τ2 )(a b c d c d0 ) = { }.
We write Pσ := {s ∈ PA,B | σ(s) 6= ∅} and Vσ := {p ∈ VA,B |
σ(p) 6= ∅}. Then Vσ1 = Vσ2 but Pσ1 6= Pσ2 . The set-of-views
approach fails to distinguish σ1 from σ2 .
P
O
• The last move is in MP
A,C = MA + MC , and
• sA,B and sB,C are plays of (A, B) and (B, C), respectively.
Switching condition and basic blocks Before defining the morphisms between interaction sequences, we introduces a useful tool
to analyse the interaction sequences.
Proposition 26. ι∗ τ ∈ Sh(PA,B ) for every τ ∈ Sh(VA,B ).
Proof. Let S = {fξ : sξ → s}ξ∈Ξ be a covering sieve and {xξ ∈
(ι∗ τ )(sξ )}ξ∈Ξ be a matching family. Each xξ is a τ -annotation
eξ,2 eξ,4 . . . eξ,|sξ | . It suffices to give an annotation e2 e4 . . . en for
s (here n = |s|). Let k ≤ n be an even number. Since S is
a covering sieve, it must be jointly surjective, i.e. k ∈ img(fξ )
for some ξ. When fξ (lk ) = k, we define ek = eξ,lk . This does
not depend on the choice of ξ since xξ is a matching family. The
resulting sequence e2 . . . en satisfies the required conditions. The
uniqueness is trivial.
Definition 30 (Switching condition). Let (A, B, C) be a triple of
arenas. A sequence over MA + MB + MC is said to satisfy the
switching condition if it is accepted by the following automaton
with the initial state OOO of which all states are accepting.
MP
C
1 OOO
\
MO
C
u
OPP o
∗
Proposition 27. ι∗ and ι form an adjoint equivalence.
Proof. Let τ ∈ Sh(VA,B ). For a P-view p = m1 . . . mn , an
annotation a2 a4 . . . an ∈ (ι∗ τ )(p) is uniquely determined by an ,
since ak = τ (fk )(an ) for the unique fk : (m1 . . . mk ) →
(m1 . . . mn ). This gives a bijection ψp for each p from τ (p) to
(ι∗ τ )(p), and to (ι∗ ι∗ τ )(p) through (ι∗ ι∗ τ )(p) = (ι∗ τ )(p).
For the other direction, let σ ∈ Sh(PA,B ). Let s = m1 . . . mn
be a play. Then x ∈ (ι∗ ι∗ τ )(s) is a sequence e2 e4 . . . en such
that, for every even number k ≤ n, ek ∈ σ(dm1 . . . mk e) and
el = σ(fk )(ek ) if ml x mk−1 . This means that {ak }k∈{2,4,...,n}
is a matching family of {dm1 . . . mk e → s}k∈{2,4,...,n} . Since σ
is a sheaf, there exists a bijection ϕs from (ι∗ ι∗ τ )(s) to τ (s).
It is easy to see that (ι∗ , ι∗ , ψ, ϕ) is an adjanction.
3.
For X ∈ {A, B, C}, if m `X m0 , then m `A,B,C m0 .
If ? `C m ∈ MC , then ? `A,B,C m.
If ? `C m ∈ MC and ? `B m0 ∈ MB , then m `A,B,C m0 .
If ? `B m ∈ MB and ? `A m0 ∈ MA , then m `A,B,C m0 .
A justified sequence s of a triple (A, B, C) induces justified sequences of (A, B), (B, C) and (A, C), basically by the restriction
of moves. The projection to the component (B, C), written sB,C ,
is just the restriction. The projection to the component (A, B), written sA,B , is the restriction to moves in MA,B in which ? x m
for an initial B-move m (whereas m0 x m in s for an initial Cmove m0 ). The projection to the component (A, C), written sA,C ,
is the restriction to moves in MA,C in which an initial A-move m
is justified by the move m0 such that m0 x m00 x m (so m00 is an
initial B-move and m0 an initial C-move).
Example 25. Consider an arena pair (({d} → {d0 } → {c}) →
{b}, {a}) and let p0 = a b, p1 = a b c d and p2 = a b c d0 (in
which every move is justified by its predecessor). Define τ1 , τ2 ∈
Sh(V({d}→{d0 }→{c})→{b}, {a} ) as follows:
τ1 (p0 )
τ1 (p1 )
τ1 (p2 )
τ1 (f )(y1 )
τ1 (g)(z1 )
Interaction sequences
Definition 28 (Justified sequence). Let (A, B, C) be a triple of
arenas. The enabling relation `A,B,C for the triple is defined by:
MP
A
MO
A
MO
B
/ POP
MP
B
A state express the owners of the next moves for components
(A, B), (B, C) and (A, C) in this order.
The switching condition generalises the O/P-alternation of justified sequences for a pair (A, B).
Lemma 31. Interaction sequences satisfy the switching condition.
Proof. Observe that each state of the automaton is determined by
the first two component. Thus the O-P alternation for (A, B) and
(B, C) components suffice for the switching condition.
Recall that basic constituents of plays are pairs of consecutive
O-P move occurrences, called O-P blocks. Thanks to the switching
condition (Lemma 31), we know that interaction sequences consist
of what we shall call basic blocks: a basic block is a sequence of
consecutive move occurrences in the interaction sequence, starting
O
from a move in MP
A,C and ending with a move in MA,C , possibly
having moves in MB as intermediate moves.
Interaction and composition
This section introduces the notion of interaction sequences and
defines the composition (σ1 ; σ2 ) ∈ Sh(PA,C ) of sheaves σ1 ∈
Sh(PA,B ) and σ2 ∈ Sh(PB,C ), generalising the composition of
deterministic innocent strategies as in [10]. The composition is
6
2018/2/15
The category of interaction sequences Given a triple (A, B, C),
P
a generalised P-move is a move in MO
A + MB + MC . This can be
P
P
P
written as MA,C + MB and as MA,B + MB,C . An generalised
O
O-move is a move in MP
A + MB + MC .
The preceding move nl−2 is pointed by nl−1 , so by the induction
hypothesis, nl2 is not in t1 . If nl2 is in s0 , then all preceding moves
are in s0 . If nl−2 is in t2 , by iterating the same argument, we
conclude that n1 . . . nl does not contain moves in t1 . Since uB,C
is a play, its justifier is in its P-view. Hence not a move in t1 .
We prove that vB,C is a play, using the above claim. Notice that
vB,C is obtained by commuting adjacent O-P blocks in uB,C as
much as required. The above claim implies that every O-P block
in t1 B,C does not interfere to any O-P block in t2 B,C . Since
commutation of non-interfering O-P blocks preserves P-visibility,
vB,C is a play. Similarly vA,B is a play.
Definition 32. Let (A, B, C) be a triple of arenas and s, s0 be interaction sequences over (A, B, C). Suppose that s = m1 . . . mn
and s0 = m01 . . . m0n0 . A morphism between s and s0 is an injective
map f : [n] → [n0 ] which satisfies:
• mk = m0f (k) (as moves),
• mi x mk implies m0f (i) x m0f (k) (and similarly for ? x
mk ), and
Lemma 37. For every interaction sequence u of (A, B, C), the
projection uA,C is a play.
• if a generalised O-move mk is followed by mk+1 , m0f (k) is
followed by
m0f (k+1)
(i.e. f (k + 1) = f (k) + 1).
Proof. Let u be an interaction sequence of (A, B, C). We define
the set P of interaction sequences as the least set that satisfies
(1) u ∈ P , and (2) if v ∈ P and v 0 is obtained from v by
commuting a non-interfering basic blocks, then v 0 ∈ P . In (2),
v 0 is an interaction sequence by Lemma 36. Consider P A,C :=
{vA,C | v ∈ P }. This is a set of alternating justified sequences of
(A, C) that is closed under the commutations. By Lemma 16, each
element in P A,C is a play. So uA,C is a play.
In other words, a morphism between interaction sequences is an
injective map between the respective occurrence-sets that preserve
moves, justification pointers and basic blocks.
Definition 33. Given arenas A, B and C, the category of interaction sequences, written as IA,B,C , has interaction sequences as
objects and morphisms defined above.
Remark 34. One can introduce the topology to IA,B,C as follows,
though we shall not use them: A family of morphisms {fξ : sξ →
s}ξ∈Ξ
S in IA,B,C is said to cover s if they are jointly surjective,
i.e. ξ∈Ξ img(fξ ) = [n], where n is the length of s.
Projections as functors Given an interaction sequence u ∈
IA,B,C , the projections uA,B , uB,C and uA,C are plays of
(A, B), (B, C) and (A, C), respectively. Those projections are
naturally extended to functors: given interaction sequences u, v ∈
IA,B,C and a morphism f : u → v, the restriction f A,B of f is a
morphism f A,B : uA,B → uA,B .
Projection to (A, C) component The projections of an interaction sequence onto (A, B) and (B, C) components are plays by
definition. We show that the projection onto (A, C) component is
also a play.
Lemma 38. The projection A,B : IA,B,C → PA,B , B,C :
IA,B,C → PB,C and A,C : IA,B,C → PA,C are functors.
Definition 35 (Commuting an adjacent pair of non-interfering
blocks). Let u be an interaction sequence of (A, B, C). Let
m1 v1 m01 m2 v2 m02
u
:= {i ∈
Proof. Recall that uA,B is the restriction of u to IA,B
[|u|] | u(i) ∈ MA,B }. A morphism f : u → u0 in IA,B,C ,
which is an injection f : [|u|] → [|u0 |] on sets, is mapped to
u0
u
u
. It is easy to see that this is functorial.
→ IA,B
f IA,B
: IA,B
be an adjacent pair of basic blocks in u, where m1 and m2 are
0
0
P
moves in MO
A,C , m1 and m2 are moves in MA,C , and v1 and v2
are sequences of moves in MB ; i.e. u = u0 m1 v1 m01 m2 v2 m02 u1 .
We say that the pair of basic blocks are non-interfering if the
justifier of m2 is not m01 . The commuted sequence u0 is defined by
u0 := u0 m2 v2 m02 m1 v1 m01 u1 (in which the justification pointers
are modified accordingly).
Lemma 39. Let f : s → t in PA,C and v ∈ IA,B,C such that
vA,C = t. Then there exists unique f¯v : u → v in IA,B,C such
that f¯v A,C = f .
Lemma 36. Let u be an interaction sequence of (A, B, C) and
let v be obtained from u by commuting an adjacent pair of noninterfering blocks. Then v is an interaction sequence.
Proof. Observe that the O-P blocks in t bijectively correspond to
the basic blocks in v. Since a morphism f : s → t is an injective
map between O-P blocks, the bijection between O-P blocks and
basic blocks determines f¯v : u → v. So f¯v is unique if it exists.
We prove the existence. If f is a commutation, Lemma 36 suffices.
If f is an embedding induced by a prefix, existence of f¯v is trivial.
Lemma 17 says that these cases are enough to prove the claim.
Proof. Let u = s0 t1 t2 s00 and v = s0 t2 t1 s00 , where t1 and t2
are non-interfering basic blocks, i.e. the justifier of the first move
in t2 is not the last move in t1 . Let t2 = m1 . . . mk . We prove tho
following claim:
Let mi be a move in t2 . Then the justifier of mi is not in t1 .
In other words, A,C : IA,B,C → PA,C is a fibration of which
each fibre is a discrete category. We write f ∗ (v) for the object u in
the lemma and f¯v for the morphism.
We prove this by induction on i.
We prove the base case i = 1. Since m1 ∈ MP
A,C , by the
definition of the basic block, its justifier is in MO
A,C . Because t1 is
a basic block, the unique move in MO
A,C is the last move. By the
assumption the justifier of m1 differs from the last move of t1 , as
desired.
We prove the induction step. Let mi be a move in t2 (i > 1).
P
Then mi is either in MP
B,C or in MA,B . Suppose that mi ∈
P
MB,C . Since u is an interaction sequence, uB,C is a play. In
particular the justifier of mi is in d(s0 t1 m1 . . . mi )B,C e. Let
n1 . . . nl be the P-view. We show that no move in this sequence is
in t1 . First nl = mi and its immediate predecessor nl−1 are in t2 .
3.2
Composition
Let σ1 ∈ Sh(PA,B ) and σ2 ∈ Sh(PB,C ) be sheaves. We define
the composite (σ1 ; σ2 ) : Pop
, which shall be proved to
A,C →
be a sheaf. For a play s ∈ PA,C , the set (σ1 ; σ2 )(s) is defined by
a
(σ1 ; σ2 )(s) :=
σ1 (uA,B ) × σ2 (uB,C ).
Set
u∈IA,B,C : uA,C =s
So an element in (σ1 ; σ2 )(s) is represented by a triple (u, e1 , e2 ),
where u ∈ IA,B,C such that uA,C = s, e1 ∈ σ1 (uA,B ) and
7
2018/2/15
of f since {xf }f ∈S is a matching family. Now by the construction,
uf = f ∗ (u).
∗
¯
Then we have a family T :=
S {fu : f (u) → u}f ∈S . This
family is jointly surjective, i.e. f ∈S img(f¯u ) = [|u|], since S is
jointly surjective on O-P blocks of s, which bijectively correspond
to basic blocks of u. Hence T A,B := {f¯u A,B | f ∈ S}
and T B,C := {f¯u B,C | f ∈ S} are covering families and
{yf }f ∈S and {zf }f ∈S are matching families of them. Hence there
exist amalgamations x ∈ σ1 (uA,B ) and y ∈ σ2 (uB,C ). Then
(u, x, y) ∈ (σ1 ; σ2 )(s) is the amalgamation.
The uniqueness of u follows from the construction and the
amalgamations x and y are unique since σ1 and σ2 are sheaves.
e2 ∈ σ2 (uB,C ). For a morphism f : s → t in PA,C , (σ1 ; σ2 )(f )
is a function given by
(u, e1 , e2 ) 7→ (f ∗ (u), σ1 (f¯u A,B )(e1 ), σ2 (f¯u B,C )(e2 )).
In the preceding, we use the common notation x · f to mean
F (f )(x) where F : Cop →
, f : s → t is a morphism of
, and x ∈ F (t). By this notation, the second component can be
written as e1 · (f¯u A,B ) and the third component as e2 · (f¯u B,C ).
Categorically, the composite is the left Kan extension.
Set
C
Lemma 40. Assume σ1 ∈ Sh(PA,B ) and σ2 ∈ Sh(PB,C ). Let
F : Iop
be a functor defined by F (u) := σ1 (uA,B ) ×
A,B,C →
σ2 (uB,C ). Then the composite (σ1 ; σ2 ) is the left Kan extension
op
of F along the projection π : Iop
A,B,C → PA,C .
Set
3.3
O
σ1 ;σ2
π
Iop
A,B,C
op
/ Pop
A,B × PB,C
σ1 ×σ2
Associativity
The associativity of composition (up to natural isomorphism) is
proved by studying “generalised” interaction sequences IA,B,C,D
that have two internal components. This is a standard technique.
Pop
A,C
/
Set × Set
,/
Set
Definition 43. Given a quadruple (A, B, C, D) of arenas, the
enabling relation `A,B,C,D on MA,B,C,D := MA +MB +MC +
MD is defined by: (1) if m `X m0 for some X ∈ {A, B, C, D},
then m `A,B,C,D m0 , (2) if ? `D m, then ? `A,B,C,D m, (3) if
? `D m and ? `C m0 , then m `A,B,C,D m0 , (4) if ? `C m and
? `B m0 , then m `A,B,C,D m0 , and (5) if ? `B m and ? `A m0 ,
then m `A,B,C,D m0 . A justified sequence over (A, B, C, D)
is a sequence of MA,B,C,D equipped with pointers that respect
`A,B,C,D . Given a justified sequence w over (A, B, C, D), the
projections wA,B,C onto interaction sequences and wA,B onto
plays are defined in the obvious way. A justified sequence over
(A, B, C, D) is an interaction sequence if wA,B , wB,C and
P
O
wC,D are plays and its last move is in MP
A,D = MA + MD .
Proof. The universal natural transformation α : F → (σ1 ; σ2 ) ◦ π
is given by
αu : F (u) 3 (e1 , e2 ) 7→ (u, e1 , e2 ) ∈ (σ1 ; σ2 )(π(u)).
Set
Assume a functor H : Pop
and a natural transformation
A,C →
β : F → H ◦π. Thus for every u ∈ IA,C,B , we have βu : F (u) →
H(π(u)). Now γs : (σ1 ; σ2 )(s) → H(s) is defined by
γs (u, e1 , e2 ) := βu (e1 , e2 )
`
(recall that (σ1 ; σ2 )(s) = u: π(u)=s σ1 (uA,B ) × σ2 (uB,C )).
Then γ is natural and γπ(u) ◦ αu = βu for all u. Uniqueness of γ
comes from the universal property of coproducts.
Definition 44 (Switching condition). Let (A, B, C, D) be a quadruple of arenas and s be a sequence over MA,B,C,D . It satisfies the
switching condition if it is accepted by the following automaton
from the initial state OOO (all states are accepting).
Remark 41. In the traditional set-theoretic HO/N game semantics, the composite of strategies PA,B and PB,C (i.e. even-prefix
closed subsets of plays over (A, B) and over (B, C), respectively) is defined by (PA,B ; PB,C ) := {s ∈ PA,C | ∃u ∈
IA,B,C . uA,B ∈ PA,B and uB,C ∈ PB,C }. Our composition
satisfies (Pσ1 ); (Pσ2 ) = P(σ1 ;σ2 ) , where Pσ = {s | σ(s) 6= ∅}.
The composite of sheaves is again a sheaf.
o
OOO
O
MP
D
MO
D
OOP o
Theorem 42. Let σ1 ∈ Sh(PA,B ) and σ2 ∈ Sh(PB,C ) be
sheaves. Then σ1 ; σ2 is a sheaf over PA,C .
MP
A
MO
A
MO
C
/ POO
O
MO
B
MP
B
/ OPO
MP
C
The three components of states correspond to (A, B), (B, C) and
(C, D) in this order.
Proof. Let s = m1 . . . mn ∈ PA,C be a play, {f : sf →
s}f ∈S ∈ J(s) be a covering sieve and {xf ∈ (σ1 ; σ2 )(sf )}f ∈S
be a matching family. By the definition of σ1 ; σ2 , we have
a
xf = (uf , yf , zf ) ∈
σ1 (uA,B ) × σ2 (uB,C ).
Lemma 45. Every interaction sequence over (A, B, C, D) satisfies the switching condition.
Proof. This is because the automaton checks if each component is
O-P alternating.
u
P
A basic block consists of the start move in MO
A,D = MA +
P
the last move in MA,D and intermediate moves in MB +
MC . An morphism f : w → w0 between interaction sequences
over (A, B, C, D) is an injective map between move occurrences
that preserve moves, the justification pointers and basic blocks.
We write IA,B,C,D for the category of generalised interaction sequences.
We claim that there exists u such that:
MO
D,
• uA,C = s, and
• uf = f ∗ (u) for every f ∈ S.
If such u exists, there is a bijective correspondence between basic
blocks of u and O-P blocks of s. This correspondence tells us
the start and the last moves of each block. So it suffices to fill
the intermediate B-moves for each basic block. Consider the kth
basic block. Since S is a covering sieve, we have a morphism
f : sf → s ∈ S such that 2k ∈ img(f ) (recall that kth O-P block
is m2k−1 m2k ). Let l be the index such that f (l) = 2k. Recall that
xf = (uf , yf , zf ) with uf A,C = sf . Then the basic block of
uf corresponding to the O-P block m0l−1 m0l in sf = m01 . . . m0|sf |
tells us the kth basic block of u. This is independent of the choice
Lemma 46.
• Projections from IA,B,C,D (e.g. A,B,C and A,B ) are functors.
• Composition of projections is a projection, e.g.
A,B,C
B,C
B,C
IA,B,C,D −→ IA,B,C −→ PB,C = IA,B,C,D −→ PB,C .
• The projection A,D : IA,B,C,D → PA,D is a discrete fibration.
8
2018/2/15
Proof. The first two claims are easy to see. The third claim can be
proved by the same technique to Lemma 39
Lemma 48. The simultaneous composition is naturally isomorphic
to sequential compositions σ1 ; (σ2 ; σ3 ) and (σ1 ; σ2 ); σ3 .
Lemma 47. Let u ∈ IA,B,D and v ∈ IB,C,D . If (uB,D ) =
(vB,D ), there exists a unique w ∈ IA,B,C,D such that u =
wA,B,D and v = wB,C,D . A similar statement holds for every
u ∈ IA,C,D and v ∈ IA,B,C .
Proof. Given s ∈ PA,D , consider a function ψs that maps an
element (w, e1 , e2 , e3 ) of
a
σ1 (wA,B ) × σ2 (wB,C ) × σ3 (wC,D )
Proof. Let u = m1 . . . mM ∈ IA,B,D and v = n1 . . . nN ∈
IB,C,D and suppose that π B,D (u) = π B,D (v). We construct
w ∈ l1 . . . lL ∈ IA,B,C,D . By the switching condition, u and v
must be accepted by the left and right automata, respectively,
to ((wA,B,D ), e1 , ((wB,C,D ), e2 , e3 )) of
a
a
σ1 (uA,B ) ×
σ2 (vB,C ) × σ3 (vC,D ).
0 q1
X
MP
D
MO
D
MP
D
MP
A
MO
A
v
q2 o
MO
D
/ q 3
MO
B
0 p1
X
v
p2 o
MP
B
w: wA,D =s
u: uA,D =s
This is a bijection thanks to Lemma 47. It is easy to show the
naturality of ψ.
Let us write F for the simultaneous composition and G =
(σ1 ; (σ2 ; σ3 )). Assume f : s → t in PA,D . Then F (f ); ψt maps
(w, e1 , e2 , e3 ) to
MP
B
MO
B
MO
C
/ p 3
MP
C
(w, e1 , e2 , e3 )
and w must be accepted by the automaton
rO 1 o
MP
D
MO
D
r2 o
MP
A
MO
A
MO
C
F (f )
7−→(f ∗ (w), e1 · (f¯w A,B ), e2 · (f¯w B,C ), e3 · (f¯w C,D ))
∗
ψt (f (w)A,B,D , e1 · (f¯w A,B ),
7−→
(f ∗ (w)B,C,D ), e2 · (f¯w B,C ), e3 · (f¯w C,D ))
/r
O4
MO
B
/ r3
u: vB,D =uB,D
MP
B
and ψs ; G(f ) maps (w, e1 , e2 , e3 ) to
(w, e1 , e2 , e3 )
MP
C
ψs
7−→((wA,B,D ), e1 , ((wB,C,D ), e2 , e3 ))
We construct a sequence of moves w such that wA,B,D = u and
wB,C,D = v. An intermediate state is a tuple (i, j, k, p, q, r) such
that i ≤ M , j ≤ N such that mi . . . mM B,D = nj . . . nN B,D ,
k is the current index of l and p, q and r are states of the above
automata from which mi . . . mM , nj . . . nN and lk . . . lL are accepted, respectively.
(f ∗ (wA,B,D ), e1 · (f¯wA,B,D A,B ),
((f¯wA,B,D B,D )∗ (wB,C,D ),
G(f )
7−→
B,D )
B,C ),
e2 · ((f¯w
A,B,D
B,C,D )
O
P
• (i, j, k, q1 , p1 , r1 ): Then mi ∈ MO
D +MA . If mi ∈ MD , then
•
•
•
•
(wB,C,D )
e3 · ((f¯wA,B,D B,D )(w
C,D ))).
By Lemma 49, we have f ∗ (w)A,B,D = f ∗ (wA,B,D ), so the
first components coincide. As for the second components, again by
Lemma 49, we have
(f¯w
)A,B = (f¯w )A,B,D A,B = f¯w A,B .
let lk = mi = nj and proceed to (i+1, j +1, k +1, q2 , p2 , r2 ).
If mi ∈ MP
A , then let lk = mi and proceed to (i + 1, j, k +
1, q3 , p1 , r4 ).
O
P
(i, j, l, q2 , p2 , r2 ): Then nj ∈ MO
C + MD . If nj ∈ MC ,
then let lk = nj and proceed to (i, j + 1, k + 1, q2 , p3 , r3 ).
If nj ∈ MP
D , then let lk = nj = mi and proceed to
(i + 1, j + 1, k + 1, p1 , q1 , r1 ).
O
P
(i, j, l, q2 , p3 , r3 ): Then nj ∈ MO
B + MC . If nj ∈ MB , then
let lk = mi = nj and proceed to (i+1, j +1, k +1, q3 , p1 , r4 ).
If nj ∈ MP
C , then let lk = nj and proceed to (i, j + 1, k +
1, q2 , p2 , r2 ).
P
O
(i, j, l, q3 , p1 , r4 ): Then mi ∈ MO
A + MB . If mi ∈ MA ,
then let lk = mi and proceed to (i + 1, j, k + 1, q1 , p1 , r1 ).
If mi ∈ MP
B , then let lk = mi = nj and proceed to
(i + 1, j + 1, k + 1, q2 , p3 , r3 ).
Other cases are never reached.
A,B,D
For the third components, recall that
(f¯w
)B,D = (f¯w )B,D .
A,B,D
and (f¯w )B,D : (f ∗ (w)B,D ) → (wB,D ). Since
(f¯w )B,C,D : (f ∗ (w)B,C,D ) → (wB,C,D )
is projected onto (f¯w )B,D , we have
((f¯w )B,D )∗ (wB,C,D ) = f ∗ (w)B,C,D .
For the fourth components, by using Lemma 49, we have
(f¯wA,B,D B,D )w
B,C,D
= (f¯w B,D )wB,C,D
= (f¯w B,C,D B,D )wB,C,D
The justification pointer for A-moves are determined by u and
others by v.
= (f¯w )B,C,D
(in general, for h : u → v in IB,C,D , we have (hB,D )v = h) and
(f¯w )B,C,D B,C = f¯w B,C
Given innocent strategies σ1 ∈ Sh(PA,B ), σ2 ∈ Sh(PB,C )
and σ3 ∈ Sh(PC,D ), their simultaneous composition F : Pop
A,D →
is defined by: for objects, F (s) is
a
σ1 (wA,B ) × σ2 (wB,C ) × σ3 (wC,D )
Set
as desired. The fifth component is the same.
Lemma 49. Let w ∈ IA,B,C,D and f : s → (wA,D ) in PA,D .
Then
f ∗ (w)A,B,D = f ∗ (wA,B,D )
and
f¯w A,B,D = f¯wA,B,D .
w∈IA,B,C,D : wA,D =s
and, given f : s → t in PA,D , the function F (f ) : F (t) → F (s)
maps (w, e1 , e2 , e3 ) ∈ F (t) to
(f ∗ (w), e1 · (f¯w A,B ), e2 · (f¯w B,C ), e3 · (f¯w C,D )).
9
2018/2/15
Proof. By definition, f¯w : f ∗ (w) → w in IA,B,C,D . Thus
3.5
∗
f¯w A,B,D : (f (w)A,B,D ) → (wA,B,D ).
Basically the full completeness is achieved by establishing the correspondence between the paths of terms in normal form and Pviews. This subsection describes the key lemma for full completeness, adapting the standard technique for HO/N game models.
An arena A is prime if it has a unique initial move. Then
A = B ⇒ {m} for some arena B and the initial A-move m.
Let A = A1 × · · · × An be an arena, where Ai is prime for
each i, and i ∈ [n]. Writing m2 for the unique initial Ai -move,
(m1 m2 ) ∈ VA,{m1 } . We define (m1 m2 )/VA,{m1 } as the full
subcategory consisting of P-views p > (m1 m2 ). (Since VA,{m1 }
is a poset, this coincides with the standard definition of the under
category.) Suppose Ai = B ⇒ {m2 }. There is an isomorphism
Both claims follow from f¯w A,B,D A,D = f¯w A,D = f .
Corollary 50. Composition is associative up to isomorphism.
3.4
Key lemma for full completeness
CCC of arenas and strategies
Definition 51. The category of arenas and strategies G has arenas
as objects and a sheaf σ ∈ Sh(PA,B ) as a morphism from A to B.
We regard that isomorphic sheaves define the same morphism. The
composition is defined in Section 3.2.
As usual, the identity morphisms are copycat strategies.
∼
=
χ(m1 m2 ) : (m1 m2 )/VA,{m1 } → VA,B ,
Definition 52. Let A be an arena. Let us write a move in MA,A =
MA +MA as l(m) and r(m) for m ∈ MA , in order to distinguish
the component. The relation ∼ is given by l(m) ∼ r(m) and
r(m) ∼ l(m) (i.e. ∼ relates the same move in the different
component). A play s = m1 m2 . . . mn ∈ PA,A is copycat if,
for every even number k ≤ n, (1) mk−1 ∼ mk , (2) ? x mk−1
implies mk−1 x mk , and (3) mj x mk−1 implies mj−1 x mk .
The copycat strategy idA ∈ Sh(PA,A ) is defined by: idA (s) =
{∗} if s is copycat and idA (s) = ∅ otherwise.
Proposition 53. (idA ; σ) ∼
=σ∼
= (σ; idB ) for σ ∈ Sh(PA,B ).
given by m1 m2 m3 . . . ml 7→ m3 . . . ml . Here we need to modify
the justification pointer as follows:
• If m2 x mk in LHS (then k = 3), then ? x mk in RHS.
• If m1 x mk in LHS, then m3 x mk in RHS.
• If mj x mk in LHS (j 6= 1, 2), then mj x mk in RHS.
This isomorphism is the key to prove full completeness.
Let τ ∈ Sh(VA,B ). Suppose that A = A1 × . . . An , where Ai
is prime for each i. Let i ∈ [n] and Ai = B ⇒ {m2 }. We define
the operation (m1 m2 ) B τ that “inserts” m1 m2 before the P-views
in τ , defined by:
In the rest of this subsection, we show that G is a CCC. It is an
adaptation of the standard arguments for HO/N game models.
((m1 m2 ) B τ )(m1 m2 )
((m1 m2 ) B τ )(m1 m2 p)
((m1 m2 ) B τ )(p)
Products and terminal object Given arenas A and B, the arena
A × B is defined by: MA×B := MA + MB , λA×B := [λA , λB ]
and (`A×B ) := (`A ) ∪ (`B ). We say a play s ∈ PA×B,A
is copycat if s does not contain B-moves and it is copycat as a
play of PA,A . The projection π1 ∈ Sh(PA×B,A ) is defined by:
π1 (s) = {∗} if s is copycat and π1 (s) = ∅ otherwise. The
projection π2 ∈ Sh(PA×B,B ) is defined similarly.
For a play s ∈ PA,B×C , we write sA,B for the restriction of
s to {i | mj x∗ mi for some mj ∈ MB }, where x∗ is the
reflexive and transitive closure of x. The restriction is a functor.
The terminal object is the empty arena having no moves.
:= {∗}
:= τ (p)
:= ∅
(otherwise).
To be precise, the second equation should be written as ((m1 m2 )B
τ )(m1 m2 p) := τ (χ(m1 m2 ) (m1 m2 p)). Then ((m1 m2 ) B τ ) ∈
Sh(VA,{m1 } ).
Lemma 55. Let τ ∈ Sh(VA,B ) and suppose that A = A1 × · · · ×
An , Ai is prime for all i, k ∈ [n] and Ak = B ⇒ {m2 }. Then
ι∗ ((m1 m2 ) B τ ) ∼
= hπi , ι∗ (τ )i; ev
where πi ∈ Sh(PA,Ai ) is the projection of the product and ev =
Λ(idAi ) ∈ Sh(P(B⇒{m2 })×B, {m2 } ) is the evaluation map.
Exponentials Let A and B be arenas. The exponential arena
A ⇒ B is defined by: (1) MA⇒B := {m ∈ MB | ? `B
m} × MA + MB , (2) λA⇒B (m) := ¬λA (mA ) (if m =
(mB , mA )) and λA⇒B (m) := λB (m) (if m ∈ MB ), where
¬O = P and ¬P = O. The enabling relation is defined by (a) if
? `A mA , then mB `A⇒B (mB , mA ), (b) if mA `A m0A ,
then (mB , mA ) `A⇒B (mB , m0A ), and (c) if mB `B m0B , then
mB `A⇒B m0B .
Given a play s ∈ PA,B⇒C , let us write θ(s) for the justified
sequence in which (mC , mB ) ∈ MB⇒C ⊆ MA,B⇒C is replaced
with mB ∈ MA×B,C . Then θ(s) is a play over (A × B, C).
Conversely, given a play s ∈ PA×B,C , let us write θ−1 (s) for
the justified sequence in which every B-move mB ∈ MB ⊆
MA×B,C is replaced with (mC , mB ) ∈ MB⇒C ⊆ MA,B⇒C
where mC is the initial C-move s.t. mC x+ mB . Since θ and
θ−1 do not change the order of move occurrences nor justification
pointers, they are functors. Furthermore θ−1 is the inverse of θ. So
PA,B⇒C is isomorphic to PA×B,C . Since θ maps views to views,
we have an isomorphism between VA,B⇒C and VA×B,C as well.
The isomorphism θ : PA,B⇒C → PA×B,C gives an isomorphism Λ : Sh(PA×B,C ) → Sh(PA,B⇒C ) : σ 7→ σ ◦ θ. This is a
natural bijection on hom-sets Λ : G(A × B, C) ∼
= G(A, B ⇒ C).
In summary, we have the following result.
4.
Sheaves model for deterministic λ→
This section develops the sheaves model for simply-typed λcalculus, the simplest functional programming language.
4.1
The target language
The standard simply-typed call-by-name λ-calculus extended to
have divergence ⊥. The syntax of terms is given by:
M ::= x | λx.M | M M | ⊥.
We consider simply-typed terms possibly having free variables.
Types are type environments are given by the grammar:
κ ::= o | κ → κ
Γ ::= · | Γ, x : κ.
The typing rules are standard, expect that ⊥ is considered as a
constant of the ground type o.
We study the equational theory of terms, precisely βη-theory.
The relation = is the least equivalence relation that satisfies
(λx.M ) N = M [N/x]
λx.M x = M
(x fresh)
0
and the congruence rules: if M = M , then M N = M 0 N and
N M = N M 0 . The normal form is defined by:
Lemma 54. G is a cartesian closed category.
Q ::= λx1 . . . xk .y Q1 . . . Qn | λx1 . . . xk .⊥
10
2018/2/15
4.2
This section studies an extension of λ→ having the non-deterministic
branch and interprets the calculus using G. We shall prove the
soundness of interpretation and the full completeness.
Deterministic strategies
Definition 56. An odd-length play is an odd-length alternating Pvisible justified sequence. (It is not a play because a play is of
even-length.) For an odd-length play s over (A, B), the immediate
extension ie(s) is a set of plays {sm | sm ∈ PA,C }.
5.1
(M1 + M2 ) N = (M1 N ) + (M2 N )
Definition 57. An innocent strategy`σ ∈ Sh(PA,B ) is deterministic if, for every odd-length play s, t∈ie(s) σ(t) is empty or singleton. It is finite if {p ∈ VA,B | σ(p) 6= ∅} is a finite set.
λx.(M1 + M2 ) = (λx.M1 ) + (λx.M2 )
M + (λx1 . . . xn .⊥) = M
and the associativity and commutativity of +. These equations
are sound with respect to the observational equivalence in the
call-by-name evaluation strategy, where the observable is mayconvergence. (They are not sound for must-convergence because
of the right equation.)
We define normal forms where n, k ≥ 0:
Remark 58. If σ is deterministic, then σ(s) is empty or singleton
for every s ∈ PA,B . So it is completely determined by a set
{s ∈ PA,B | σ(s) 6= ∅}. Through this translation, the sheaf-based
definition of innocent strategies coincides with the standard one.
Definition 59. A category of deterministic strategies Gdet is a
subcategory consisting of deterministic strategies.
R := Q1 + · · · + Qn
Gdet is well-defined since the identity idA deterministic and the
composition preserves determinacy.
5.2
Proof. Let σ1 ∈ Sh(PA,B ) and σ2 ∈ Sh(PB,C ) be deterministic strategies. Then for every odd-length play s of (A, C), there
exists at most one u such that uA,C = sm, σ1 (uA,B ) 6= ∅
and
` σ2 (uB,C ) 6= ∅ (see uncovering construction in [10]). Thus
u: uA,C ∈ie(s) σ1 (uA,B ) × σ2 (uB,C ) is empty or singleton.
Interpretation and soundness
The term Γ ` M +N : κ is interpreted as the coproduct [[M ]]+[[N ]]
in Sh(P[[Γ]],[[κ]] ). A simple way to describe the coproduct is to
use sheaves over views: since the sheaves over views are just
presheaves, the coproduct can be computed pointwise. So, given
τ1 , τ2 ∈ Sh(VA,B ), we have (τ1 + τ2 )(p) = τ1 (p) + τ2 (p). For
sheaves σ1 , σ2 ∈ Sh(PA,B ) over plays, we define σ1 + σ2 :=
ι∗ ((ι∗ σ1 ) + (ι∗ σ2 )) using the Comparison Lemma (Lemma 24).
Coproducts on the function position commutes with application.
Lemma 63. (hσ0 , σ1 +σ2 i; ev) ∼
= (hσ0 , σ1 i; ev)+(hσ0 , σ2 i; ev).
Since projections A×B → A and A×B → B are deterministic
and the isomorphism Sh(PA×B,C ) ∼
= Sh(PA,B⇒C ) preserves
determinacy, Gdet is a CCC.
Interpretation
Proof. (Sketch) By Lemma 24, it suffices to consider their restrictions on P-views. Let σ0 ∈ PA,B and σ1 , σ2 ∈ PA,B⇒C . Let us
write D := B × (B ⇒ C) for simplicity. Then the right-hand-side
on P-view p ∈ VA,C is given by
Simple types are interpreted as objects by
[[κ → κ0 ]] := [[κ]] ⇒ [[κ0 ]]
as well as type environemnts
((hσ0 , σ1 i; ev) + (hσ0 , σ2 i; ev))(p)
a
∼
σ0 (uA,B ) × σ1 (uA,B⇒C ) × ev(uD,C )
=
[[x1 : κ1 , . . . , xn : κn ]] := [[κ1 ]] × · · · × [[κn ]].
The interpretation of terms is fairly standard:
u:uA,C =p
[[x1 : κ1 , . . . , xn : κn ` xi : κi ]] := πi
0
Q := λx1 . . . xn .y R1 . . . Rk ,
where y R1 . . . Rk is fully applied. Every term has a unique normal
form (modulo the commutation of non-deterministic branches), or
is equivalent to λx1 . . . xn .⊥. Note that M + M 6= M in general.
Lemma 60. Composition preserves determinacy.
[[o]] := {mo }
The target language
Consider the simply-typed lambda calculus with ⊥ extended to
have the non-deterministic branch: M1 + M2 . The additional axioms are:
An odd-length play s ends with an O-move and the immediate
extension ie(s) is the set of all possible Proponent’s responses.
4.3
Sheaves model for nondeterministic λ→
5.
where y Q1 . . . Qn is fully applied, i.e. y Q1 . . . Qn : o. Every
term has a unique normal form.
a
+
0
[[Γ ` λx.M : κ → κ ]] := Λ([[Γ, x : κ ` M : κ ]])
σ0 (uA,B ) × σ2 (uA,B⇒C ) × ev(uD,C )
u:uA,C =p
[[Γ ` M N : κ]] := h[[M ]], [[N ]]i; ev
[[Γ ` ⊥ : o]] := ι∗ τ∅ ,
∼
=
a
σ0 (uA,B )×(σ1 (uA,B⇒C )+σ2 (uA,B⇒C ))×ev(uD,C ).
u:uA,C =p
where τ∅ ∈ Sh(V[[Γ]],[[o]] ) is the constant functor mapping to ∅.
A play is well-opened if it has exactly one move pointing to ?. If
uA,C has a unique initial C-move and ev(uB×(B⇒C),C ) 6= ∅,
then uB⇒C has a unique initial (B ⇒ C)-move and hence
uA,B⇒C is well-opened. So we can assume without loss of generality that uA,B⇒C ranges over well-opened plays. We claim that
for a well-opened play uA,B⇒C , we have a bijection on sets
∼ (σ1 + σ2 )(uA,B⇒C ).
σ1 (uA,B⇒C ) + σ2 (uA,B⇒C ) =
Theorem 61 (Soundness). M = N iff [[M ]] ∼
= [[N ]].
Proof. This is a special case of Theorem 66 below.
Theorem 62 (Full completeness). Let Γ be a type environment,
κ be a simple type and σ ∈ Sh(P[[Γ]],[[κ]] ). If σ is finite and
deterministic, there exists a term Γ ` M : κ such that σ ∼
= [[M ]].
The required natural isomorphism is the consequence of the claim.
Assume uA,B⇒C = s = m1 . . . mn and let pk := dm1 . . . mk e
(for k ∈ {2, 4, . . . , n}) and {fk : pk → s}k∈{2,4,...,n} be
a covering family. Then (ι∗ ((ι∗ σ1 ) + (ι∗ σ2 )))(s) is the set of
sequences of the form e2 . . . en , where ek ∈ σ1 (pk ) + σ2 (pk ).
Since s is well-opened, f2 : p2 → s is factor through fk :
Proof. The set {p ∈ V[[Γ]],[[κ]] | σ(p) 6= ∅}, which is finite and
prefix-closed, gives a finite view function in the sense of [10]. A
term M that denotes σ can then be constructed by induction on
the size of the view function, following the proof of Prop. 7.4 in
op. cit.. One can directly construct a term M using Lemma 55.
11
2018/2/15
pk → s for every k. This means that ek ’s come from the same
component as e2 . So ek ∈ σ1 (pk ) for all k or ek ∈ σ2 (pk ) for
all k. Hence (ι∗ ((ι∗ σ1 ) + (ι∗ σ2 )))(uA,B⇒C ) has a bijection to
(ι∗ ι∗ σ1 )(uA,B⇒C ) + (ι∗ ι∗ σ2 )(uA,B⇒C ) as desired.
5.3
Full completeness
A sheaf σ ∈ Sh(PA,B ) is finite if
`
p∈VA,B
σ(ι(p)) is finite.
Lemma 67. Every finite sheaf σ ∈ Sh(PA,{m1 } ) can be decomposed as σ ∼
= σ1 + · · · + σn , where σi is deterministic on initial
response for all i.
Let us write [[M ]]V for its the restriction on views, i.e. ι∗ [[M ]].
For a term in normal form, its view restriction can be computed by
the induction on the structure. By definition,
[[Γ ` λx.Q : κ → κ0 ]]V ∼
= [[M ]]V ◦ θ,
∗
Proof. Let
` τ = ι σ be the restriction of σ to views. Consider the finite set p∈ie(m1 ) τ (p), which we write as {(p1 , a1 ), . . . , (pn , an )}
(ai ∈ σ(pi ) for each i ∈ [n]). We define τi ∈ Sh(Vop
A,B ). On objects,
where θ : V[[Γ]],[[κ→κ0 ]] → V[[Γ×κ]],[[κ0 ]] is the isomorphism, and
[[Γ ` Q1 + · · · + Qn : κ]]V ∼
= [[Q1 ]]V + · · · + [[Qn ]]V . The next
lemma gives the interpretation of head variable, which is a consequence of Lemma 55.
τi (p) := {a ∈ τ (p) | pi ≤ p and ai = a · f where f : pi → p }.
Then τi (p) ⊆ σ(p) for every i and p. For f : p → p0 , we define
τi (f ) as the restriction of τ (f ) : τ (t0 ) → τ (t) to τi (t0 ) ⊆ τ (t0 ). It
is easy to see that τi is a functor. Then we have
τ ∼
= τ1 + · · · + τn .
Lemma 64. Assume a term Γ ` xi R1 . . . Rn : o where xi : κi ∈
Γ. Let m1 be the unique initial move of [[o]] and m2 be the unique
initial move of [[κi ]]. Then
[[Γ ` xi R1 . . . Rn : o]]V ∼
= (m1 m2 ) B h[[R1 ]]V , . . . , [[Rn ]]V i.
To see this, consider a ∈ τ (p) for some p. Let p0 be the first two
moves of p and let a0 = a · f , where f : p0 → p (unique). Then
(p0 , a0 ) is (pi , ai ) for some i ≤ n. Hence a ∈ τi (p). Furthermore
such i is unique by the construction. So we have the claimed natural
isomorphism. Letting σi := ι∗ τi , we obtain the statement.
Let B be a prime arena and m1 be the unique initial move.
A
` sheaf σ ∈ Sh(PA,B ) is deterministic on initial response if
s∈ie(m1 ) σ(s) is singleton.
Lemma 65. [[Q]] is deterministic on initial response.
Theorem 68 (Full completeness). Let Γ be a type environment, κ
be a type and σ ∈ Sh(P[[Γ]],[[κ]] ). If σ is finite, there exists a term
Γ ` M : κ such that σ ∼
= [[M ]].
Proof. By induction on the structure of Q. If Q = xi R1 . . . Rn ,
this follows from Lemma 64. If Q = λx.Q0 , then [[Q0 ]] is deterministic on initial response and Λ : Sh(PA×B,C ) → Sh(PA,B⇒C )
preserves this property. Hence [[Q]] = Λ([[Q0 ]]) is deterministic on
the initial response.
Proof. By induction on the number of elements in
`
p∈V[[Γ]],[[κ]]
σ(p)
and the structure of κ. If κ = κ1 → κ2 , consider Λ−1 (σ) ∈
Sh(P[[Γ]]×[[κ1 ]],[[κ2 ]] ) and apply the induction hypothesis. Suppose
that κ = o. If σ has several initial responses, then by applying
Lemma 67, we have σ = σ1 + · · · + σn (n ≥ 2). By the induction hypothesis, we have σi ∼
= [[Mi ]] for every i and thus
M1 +· · ·+Mn is the required term. Suppose that σ is deterministic
on initial response. Let (m1 m2 , a) be the unique response. Since
σ ∈ Sh(P[[Γ]],[[o]] ), m1 is the unique initial move of [[o]] and m2 be
the unique initial move of [[κk ]], where xk : κk ∈ Γ for some xk .
Suppose that κk = κ01 → · · · → κ0l → o. We define the sheaf
τ 0 ∈ Sh([[Γ]], [[κ01 ]] × · · · × [[κ0l ]]) by:
Theorem 66 (Soundness). M = N iff [[M ]] ∼
= [[N ]].
Proof. To prove the left-to-right direction, it suffices to show
the all equations are valid. The equation [[(M1 + M2 ) N ]] ∼
=
[[(M1 N ) + (M2 N )]] follows from Lemma 63. Because + is the
coproduct, it is commutative and associative. Because ι∗ [[⊥]] is the
constant functor to ∅, we have σ + [[⊥]] ∼
= σ for every σ.
To prove the converse, assume that [[M ]] ∼
= [[N ]] for normal
terms Γ ` M : κ and Γ ` N : κ. Let m1 be the unique initial
∼
move
` of [[κ]]. Then, since`[[M ]] = [[N ]], we have a bijection between
[[M
]](s)
and
s∈ie(m1 ) [[N ]](s). Let n be the number of
s∈ie(m1 )
elements of those sets. Then M ≡ Q1 + · · · + Qn since [[Qi ]] is
deterministic on initial response for every i ∈ [n] by Lemma 65.
Similarly N ≡ Q01 + · · · + Q0n . Since [[M ]] ∼
= [[N ]], there is a
bijection ϕ : [n] → [n] such that [[Qi ]] ∼
= [[Qϕ(i) ]]. By the induction
hypothesis, Qi = Qϕ(i) . So M = N .
Suppose that
τ 0 (p) := σ(χ−1
(m1 m2 ) (p)),
0
where χ−1
(m1 m2 ) : V[[Γ]],[[κ1 ]]×···×[[κ0n ]] → V[[Γ]],[[o]] : p 7→ m1 m2 p
(see Section 3.5). Then ι∗ σ ∼
= ((m1 m2 ) B τ 0 ) because σ is
deterministic on initial response. So by Lemma 55, we have
0
σ∼
= hπk , τ i; ev
where πk : [[Γ]] → [[κk ]] is the projection. By the induction hypothesis, we have Mi for each i ∈ [l] such that τ 0 ∼
= h[[M1 ]], . . . , [[Ml ]]i.
Recall that [[xk ]] ∼
= πk . Since G is a CCC, the application of
the product can be rewritten by the series of applications. Hence
σ∼
= [[xk M1 . . . Ml ]] as desired.
M ≡ λx1 . . . xk .y R1 . . . Rn
0
0
N ≡ λx1 . . . xk .y 0 R10 . . . Rn
We can assume without loss of generality that k = 0. Then by
Lemma 64, we have
∼ (m1 m2 ) B h[[R1 ]]V , . . . , [[Rn ]]V i
[[M ]]V =
Example 69. Let tt = λxy.x and ff = λxy.y. Recall the example
in Introduction, M1 = λf.(f tt) + (f ff) and M2 = (λf.f tt) +
(λf.f ff). Then ι∗ [[M1 ]] = τ1 and ι∗ [[M2 ]] = τ2 , where sheaves τ1
and τ2 over P-views can be found in Example 25.
and
0
0
0
[[N ]]V ∼
= (m1 m2 ) B h[[R1 ]]V , . . . , [[Rn ]]V i
where m2 is the initial move for y : κ0 ∈ Γ and m02 is the initial
move of y 0 : κ00 ∈ Γ. Since [[M ]]V ∼
= [[N ]]V , we have m2 = m02 ,
0
0
which implies y = y and n = n . Furthermore [[M ]]V ∼
= [[N ]]V
implies [[Ri ]]V ∼
= [[Ri0 ]]V for all i ∈ [n] and thus [[Ri ]] ∼
= [[Ri0 ]]. By
the induction hypothesis, Ri = Ri0 and hence M = N .
6.
Sheaves model for probabilistic λ→
We have seen that a term of the non-deterministic λ→ is modelled
by a sheaf σ which maps a play s to a (finite) set σ(s). An element
of σ(s) represents a particular choice of branches by which the
term behaves like s.
12
2018/2/15
In this section, we shall study a non-deterministic sheaf σ
equipped with a weight map µ which assigns each choice (s, a)
(where s is a play and a ∈ σ(s)) with a positive real number
µ(s, a).
Definition 73 (Weighted innocent strategy). A weighted innocent
strategy over pairs (A, B) of arenas is a pair (σ, µ) of an innocent
non-deterministic strategy σ ∈ Sh(PA,B ) and an innocent weight
map µ for σ.
The target calculus: weighted and probabilistic λ→
Similar to the deterministic / non-deterministic cases, a weighted
innocent strategy is determined by its restriction on views.
6.1
The target language is an extension of the nondeterministic λ→
studied in the previous section. The new feature is the term constructor c · M , where c is a positive real number. The additional
equations are:
λx.(c · M ) = c · (λx.M )
Lemma 74. Assume σ, σ 0 ∈ Sh(PA,B ) and a natural isomor∼
=
phism ϕ : σ → σ 0 . Let µ and µ0 are innocent weight maps for
0
σ and σ , respectively. If µ(p, a) = µ0 (p, ϕ(a)) for every P-view
p ∈ VA,B , then µ(s, a) = µ0 (s, ϕ(a)) for every play s ∈ PA,B .
c · (M + N ) = (c · M ) + (c · N )
c1 · (c2 · M ) = (c1 c2 ) · M
(c · M ) N = c · (M N )
Proof. By induction on the length of s. Let s = s0 m1 m2 be a play
and e ∈ σ(s). If s is a P-view, the claim is just assumed. Suppose
that s is not a P-view. We have a covering family {f : s0 → s, g :
dse → s}. Since the pullback g ∗ (f ) : p0 → dse is in VA,B ,
and c · ⊥ = ⊥. These equations are admissible in the sense that
M = N implies M and N are observably equivalent in the standard call-by-name operational semantics (where the observable is
the probability of convergence). The probabilistic λ→ is a fragment
of this calculus in which nondeterministic branch and the weight
construct
P are restricted to the form (c1 · M1 ) + · · · + (cn · Mn ),
where n
i=1 ci ≤ 1.
Remark 70. The rule M (c · N ) = c · (M N ) is unsound, because
the application is not linear on the argument. For instance, if the
argument is called twice as in (λf.f (f (z)))(c · λx.x), the resulting
coefficient is c2 :
µ(f, a) = µ(g ∗ (f ), e · g) = µ0 (g ∗ (f ), ϕ(e · g))
= µ0 (g ∗ (f ), ϕ(e) · g) = µ0 (f, ϕ(e)).
By the induction hypothesis, we have
µ(s0 , e · f ) = µ0 (s0 , ϕ(e · f )) = µ0 (s0 , ϕ(e) · f ).
So we conclude
µ(s, e) = µ(f, e)µ(s0 , e · f )
(λf.f (f (z)))(c · λx.x) = (c · λx.x)((c · λx.x)z)
= c · ((λx.x)(c · ((λx.x)z))
= µ0 (f, ϕ(e))µ0 (s0 , ϕ(e) · f ) = µ0 (s, ϕ(e))
as desired.
= c · c · ((λx.x)z) = c2 z.
Similarly, if the argument never be called as in (λx.z)(c · N ), the
coefficient c does not affect, e.g. (λx.z)(c · N ) = z = (λx.z) N .
A normal form is defined by:
Lemma 75. Let τ ∈ Sh(VA,B ). Every weight map µ0 for τ can
be extended to an innocent weight map for ι∗ τ .
R := c1 · Q1 + · · · + cn · Qn
Proof. Given a non-empty P-view p = p0 m1 m2 ∈ VA,B and
e ∈ τ (p), we define δ(e) := µ0 (p0 → p, e) (if p0 6= ε) and
δ(e) := µ0 (p, e) (if p0 = ε). We give a weight map µ for ι∗ σ. Let
s = m1 m2 . . . mn ∈ PA,B and x ∈ ι∗ τ . Then x is of the form
e2 e4 . . . en , where ek ∈ τ (dm1 . . . mk e) for every even number
k ≤ n. The weight for x = e2 e4 . . . en is defined by:
Q := λx1 . . . xk .y R1 . . . Rn ,
where y R1 . . . Rn is fully applied. Every term has a unique normal
form (modulo commutation of the non-deterministic branches), or
is equivalent to λx1 . . . xk .⊥. Note that 2 · M + 2 · M 6= 4 · M .
6.2
Sheaves with weight
Set
µ(s, e2 e4 . . . en ) := δ(e2 )δ(e4 ) . . . δ(en ).
Definition 71 (Weight). Let F be a functor Dop →
. A weight
map µ assigns, for each s ∈ D and a ∈ F (s), a positive real
number µ(s, a) ∈ R+ .
It is easy to see that µ is innocent.
So one can define a weighted innocent strategy as a pair of a
sheaf over P-views and a weight function for it.
Let σ ∈ Sh(PA,B ) and µ be a weight map. Given a morphism
f : s → t in PA,B and an element a ∈ σ(t), we define µ(f, a) :=
µ(t, a)/µ(s, a · f ). Notice that µ(g ◦ f, a) = µ(g, a)µ(f, a · g).
Definition 76. The category of weighted innocent strategies Gw
has arenas as objects and weighted innocent strategies as morphisms. Here (σ1 , µ1 ) and (σ2 , µ2 ) are identifies if there exists a natural isomorphism preserving weights. A composition of
weighted innocent strategies (σ, µ) and (σ 0 , µ`0 ) is ((σ; σ 0 ), µ00 ),
where for each s and (u, e, e0 ) ∈ (σ; σ 0 )(s) = u:π(u)=s σ(uA,B )×
σ 0 (uB,C ), where e ∈ σ(uA,B ) and e0 ∈ σ 0 (u0 B,C ), we define
Definition 72 (Innocence on weight). Let σ ∈ Sh(PA,B ) be a
sheaf and µ be a weight map. The weight map µ is innocent if it
satisfies the following conditions: (1) µ(ε, ∗) = 1, and (2) given a
covering family {f : s → u, g : t → u} and a ∈ σ(u), consider
the pullback diagram
/s
s ×u t
g ∗ (f )
t
g
/ u
µ00 (s, (u, e, e0 )) = µ(uA,B , e) µ0 (uB,C , e0 ).
f
Associativity of the composition can be easily shown.
Lemma 77. Gw is a cartesian closed category.
then µ(f, a) = µ(g ∗ (f ), a · g).
Proof. Given a deterministic innocent strategy σ ∈ Sh(PA,B ),
the trivial weight map µ is defined by µ(s, e) = 1 for every s
and e. Then idA with the trivial weight map is the identity and
π1 ∈ Sh(PA×B,A ) and π2 ∈ Sh(PA×B,B ) with the trivial weight
maps are projections. The natural isomorphism G(A × B, C) =
Sh(PA×B,C ) ∼
= Sh(PA,B⇒C ) = G(A, B ⇒ C) has obvious
extension to weighted innocent strategies. Hence Gw is a CCC.
The typical case is that u = v0 v1 v2 , s = v0 v1 , t = v0 v2 and
s ×u t = v0 . Intuitively µ(f, a) is the weight of playing v2 from
s = v0 v1 (that reaches to the state a ∈ σ(s)) and µ(g ∗ (f ), a · g)
is the weight of playing v2 from v0 (that reaches to the state
a · g ∈ σ(t), the restriction of a to t). The innocence of the weight
map requires that the weight for playing v2 is independent of the
situation.
13
2018/2/15
6.3
Since the P-view of t ∈ ie(s) is given by dte = ds0 mm0 e =
p0 mm0 (for some m0 ), we have a bijection from ie(s) to ie(p0 m).
Since {ft , gt } is a covering family, a pair (bt , dt ) ∈ σ(s0 )×σ(dte)
such that bt · ft∗ (gt ) = dt · gt∗ (ft ) bijectively corresponds to
e ∈ σ(t). So there exists a bijection between {e ∈ σ(t) |
e · ft = e0 } and {e ∈ σ(dte) | e · gt∗ (ft ) = e0 · ft∗ (gt )}. Since
gt∗ (ft ) : p0 → dte is the prefix embedding and ft∗ (gt ) : ds0 e → s0
is the P-view embedding that is independent of t, we conclude
X X
µ(gt∗ (ft ), e · gt )
Semantics of weighted λ→
0
µ00
Let τ and τ be sheaves over P-views of (A, B) and µ0 and
be
weight maps for τ and τ 0 , respectively. The weight map [µ0 , µ00 ]
for τ + τ 0 is defined by [µ0 , µ00 ](p, e) := µ0 (p, e) (if e ∈ τ (p))
and [µ0 , µ00 ](p, e) := µ00 (p, e) (if e ∈ τ 0 (p)). We define c ⊗ µ0 by
(c ⊗ µ0 )(p, e) := cµ0 (p, e).
The same operations can be defined for weighted innocent
strategies through Lemma 75. Given a weighted innocent strategy
(σ, µ), we define c⊗µ the unique extension of c⊗µ0 to σ, where µ0
is the restriction of µ to P-views. Then (c ⊗ µ)(s, e) = ck µ(s, e),
where k is the number of the moves in s that point to ?. It is easy
to check that the equations about weights are sound for this interpretation, by using the next lemma.
t∈ie(s) e∈σ(t)
e·ft =e0
X
p∈ie(p0 m)
e∈σ(p)
e·(p≥p0 )=e0 ·h
=
Lemma 78. Let s be a well-opened play and e ∈ σ(s). Then
(c ⊗ µ)(s, e) = c(µ(s, e)).
µ((p ≥ p0 ), e) ≤ 1
where h : ds0 e → s0 is the P-view embedding.
Lemma 79. M = N iff [[M ]] = [[N ]].
Because the probabilistic λ→ is a fragment of the weighted
calculus, all the properties including soundness and adequacy are
applicable fro the probabilistic calculus. Full completeness can be
proved by the same way as the weighted case.
Let B be a prime arena. A weighted innocent`strategy (σ, µ)
of (A, B) is deterministic on initial response if s∈ie(o) σ(s) is
singleton and µ(s, e) = 1 for its unique element (s, e). The next
lemma can be proved by the same way as Lemma 67.
Theorem 84 (Full completeness). Let κ be a simple type, σ ∈
Sh(P[[Γ]],[[κ]] ) be finite and µ be a probabilistic weight. Then
(σ, µ) = [[M ]] for some probabilistic term Γ ` M : κ.
Lemma 80. Every finite weighted innocent (σ, µ) strategy can be
decomposed as c1 ⊗ (σ1 , µ1 ) + · · · + cn ⊗ (σn , µn ), where (σi , µi )
is deterministic on initial response.
Concluding remarks As presented, our model treats neither recursion nor primitive data types such as boolean. Further the target
languages are restricted to simply-typed calculi. However we believe that these restrictions can be relaxed.
We will apply the sheaf-theoretic approach in the paper to study
the model checking of non-determinsitic calculi, such as nondeterministic PCF and its call-by-value version, and to develop
a semantics of refinement dependent types.
The full completeness for the weighted calculus is proved by the
same technique as in the proof of Theorem 68, using Lemma 80.
Theorem 81 (Full completeness). Let (σ, µ) be a weighted innocent strategy for ([[Γ]], [[κ]]) and suppose that σ is finite. Then there
exists a term Γ ` M : κ such that (σ, µ) ∼
= [[M ]].
6.4
X
Semantics of probabilistic λ→
A weighted innocent strategy (σ, µ) is probabilistic if, for every
odd-length play s = s0 m and e0 ∈ σ(s0 ), the sum of weights of
possible responses that extends (s, e0 ) is less than 1.
References
[1] S. Abramsky and G. McCusker. Linearity, sharing and state: a fully
abstract game semantics for Idealized Algol with active expressions.
In Algol-like Languages, pages 297–329. Birkhaüser, 1997.
Definition 82. A weighted innocent strategy (σ, µ) over (A, B)
is probabilistic if, for every odd-length play s = s0 m and e0 ∈
σ(s0 ), we have
X
X
µ(ft , e) ≤ 1
[2] S. Abramsky, R. Jagadeesan, and P. Malacaria. Full abstraction for
pcf. Inf. Comput., 163(2):409–470, 2000.
[3] A. Beilinson. P-adic periods and derived de Rham cohomology.
J. AMS, 25(3):715–738, 2012.
t∈ie(s) e∈σ(t): e·ft =e0
where ft : s0 → t is the prefix embedding. It can be strictly less
than 1; the difference is the probability of divergence. A sheaf τ
over views with a weight map µ0 is probabilistic when the same
condition holds (but s is restricted to P-views).
[4] S. Castellan, P. Clairambault, and G. Winskel. Concurrent HylandOng games. Lecture slides, IHP Workshop on Semantics of Proofs
and Programs, 2014.
Lemma 83. (σ, µ) is probabilistic iff its restriction to views is.
[6] C. Eberhart, T. Hirschowitz, and T. Seiller. Fully abstract concurrent
games for pi. CoRR, abs/1310.4306, 2013.
[5] V. Danos and R. Harmer. Probabilistic game semantics. ACM Trans.
Comput. Log., 3(3):359–382, 2002.
Proof. Let σ ∈ Sh(PA,B ) and τ = ι∗ σ ∈ VA,B . Let s = s0 m be
an odd-length play and e0 ∈ σ(s0 ). We prove
X
X
µ(ft , e) ≤ 1
[7] R. Harmer. Games and Full Abstraction for Nondeterministic Languages. PhD thesis, Imperial College, 1999.
[8] R. Harmer and G. McCusker. A fully abstract game semantics for
finite nondeterminism. In LICS, pages 422–430, 1999.
t∈ie(s) e∈σ(t): e·ft =e0
[9] T. Hirschowitz and D. Pous. Innocent strategies as presheaves and
interactive equivalences for ccs. Sci. Ann. Comp. Sci., 22(1):147–199,
2012.
by induction on the length s, where ft : s0 → t is the prefix
embedding. If dse = s, then every t ∈ ie(s) is a P-view. Hence the
claim follows from the assumption.
Assume that dse 6= s. Let s0 = m1 . . . mn , mj be the justifier
of m and p0 = dm1 . . . mj e. Consider the covering family {ft :
s0 → t, gt : dte → t} for every t. Then we have µ(ft , e) =
µ(gt∗ (ft ), e · gt ) for every t ∈ ie(s). So it suffices to prove that
X
X
µ(gt∗ (ft ), e · gt ) ≤ 1
[10] J. M. E. Hyland and C.-H. L. Ong. On full abstraction for PCF: I, II,
and III. Inf. Comput., 163(2):285–408, 2000.
[11] A. Jung, M. A. Moshier, and S. J. Vickers. Presenting dcpos and dcpo
algebras. Electr. Notes Theor. Comput. Sci., 218:209–229, 2008.
[12] S. M. Lane and I. Moerdijk. Sheaves in Geometry and Logic. SpringerVerlag, 1992.
[13] P. Levy. Morphisms between plays. Lecture Slides, GaLoP, 2013.
t∈ie(s) e∈σ(t): e·ft =e0
14
2018/2/15
[14] H. Nickau. Hereditarily sequential functionals. In LFCS, pages 253–
264, 1994.
[15] C.-H. L. Ong. On model-checking trees generated by higher-order
recursion schemes. In LICS, pages 81–90, 2006.
[16] S. Rideau and G. Winskel. Concurrent strategies. In LICS, pages 409–
418, 2011.
[17] S. Staton and G. Winskel. On the expressivity of symmetry in event
structures. In LICS, pages 392–401, 2010.
[18] T. Tsukada and C.-H. L. Ong. Compositional higher-order model
checking via ω-regular games over Böhm trees”. In CSL/LICS, 2014.
[19] J.-L. Verdier. Fonctorialité de catégories de faisceaux. In Théorie des
topos et cohomologie étale de schémas (SGA 4), Tome 1, pages 265–
298. Springer-Verlag, 1972. Lect. Notes in Math. 269.
15
2018/2/15
| 6cs.PL
|
Spacetimes with Semantics (III)
The Structure of Functional Knowledge Representation
and Artificial Reasoning∗
Mark Burgess
arXiv:1608.02193v4 [cs.AI] 1 Aug 2017
6th August 2016
Minor improvements 1 August 2017
Abstract
Using the previously developed concepts of semantic spacetime, I explore the interpretation of knowledge
representations, and their structure, as a semantic system, within the framework of promise theory. By assigning
interpretations to phenomena, from observers to observed, we may approach a simple description of knowledgebased functional systems, with direct practical utility. The focus is especially on the interpretation of concepts,
associative knowledge, and context awareness. The inference seems to be that most if not all of these concepts
emerge from purely semantic spacetime properties, which opens the possibility for a more generalized understanding of what constitutes a learning, or even ‘intelligent’ system.
Some key principles emerge for effective knowledge representation: 1) separation of spacetime scales, 2) the
recurrence of four irreducible types of association, by which intent propagates: aggregation, causation, cooperation, and similarity, 3) the need for discrimination of identities (discrete), which is assisted by distinguishing
timeline simultaneity from sequential events, and 4) the ability to learn (memory). It is at least plausible that
emergent knowledge abstraction capabilities have their origin in basic spacetime structures.
These notes present a unified view of mostly well-known results; they allow us to see information models,
knowledge representations, machine learning, and semantic networking (transport and information base) in a
common framework. The notion of ‘smart spaces’ thus encompasses artificial systems as well as living systems,
across many different scales, e.g. smart cities and organizations.
Contents
1
Introduction
4
2
The structure and semantics of spacetime
2.1 Semantic spacetime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 The concept of knowledge as an iterated equilibrium . . . . . . . . . . . . . . .
2.3 The concept of knowledge spaces: spaces that can learn and represent knowledge
2.4 Spatial semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 Distance, similarity, and metrics . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Locatability of assets (sorting and discriminating identities) . . . . . . .
2.4.3 Physical representation of semantic differences . . . . . . . . . . . . . .
2.4.4 Metric and semantic coordinate bases, spanning trees . . . . . . . . . . .
2.4.5 Symmetry breaking and deterministic descriptions . . . . . . . . . . . .
2.5 Temporal semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 The semantics of spacetimes and dynamical systems . . . . . . . . . . . . . . .
2.6.1 Assigning coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.2 Time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
5
5
6
7
7
9
10
10
11
12
14
14
14
∗ This series of documents is aimed at bridging the world of natural science (especially physics) and computer science, with a compromise
on the level of rigour that I have improvised with an eye on practical applications. It is probably too ambitious in scope and detail, but bridges
may serve a purpose even with gaps.
1
3
4
5
The structure and semantics of data patterns and representations
3.1 Information content of space . . . . . . . . . . . . . . . . . . . . . .
3.2 Data and data types . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Schemas, types, and the chemistry of composition . . . . . . . . . . .
3.4 Type calibration, bases, and matroids (recap) . . . . . . . . . . . . . .
3.5 Trust and certainty in data calibration . . . . . . . . . . . . . . . . .
3.6 Data dimensionality: aggregates . . . . . . . . . . . . . . . . . . . .
3.7 The role of agent phase in data representation . . . . . . . . . . . . .
3.8 Data encoding relative to fixed markers (protocols) . . . . . . . . . .
3.9 Coordinates: spanning trees and tuples over data . . . . . . . . . . . .
3.10 Streams, state, and memory . . . . . . . . . . . . . . . . . . . . . . .
3.10.1 Assessment and timescales . . . . . . . . . . . . . . . . . . .
3.10.2 Mapping input state to concurrent context . . . . . . . . . . .
3.10.3 Transmuting temporal association into spatial representation .
3.11 Data classification and summarization . . . . . . . . . . . . . . . . .
3.11.1 Stable summarizations: attractors . . . . . . . . . . . . . . .
3.11.2 Repeated sampling, significance of data and attached meaning
3.12 Distributions, patterns, and their information content . . . . . . . . .
3.12.1 Ontological structure . . . . . . . . . . . . . . . . . . . . . .
3.12.2 Pattern acquisition: generalization and discrimination . . . . .
3.13 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
16
18
19
20
20
21
22
24
26
27
27
28
30
31
32
32
33
33
34
Memory, addressing, and spacetime patterns
4.1 Relationship between switching and memory . . . . . . . . . . . . . . . . . . . .
4.1.1 Conditional promises . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Timescales for deterministic switching . . . . . . . . . . . . . . . . . . . .
4.2 Spatial representation of encoded memory . . . . . . . . . . . . . . . . . . . . . .
4.3 Spatial addressability of memory . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Address discrimination by constrained flooding or broadcast propagation (-)
4.3.2 Address discrimination by routed propagation (+) . . . . . . . . . . . . . .
4.3.3 Criteria for addressability of data in memory . . . . . . . . . . . . . . . .
4.3.4 Non-local link memory . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.5 Teaching spacetime to tokenize patterns . . . . . . . . . . . . . . . . . . .
4.4 Temporal addressability of memory . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Sequences and oriented paths . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 Knowledge as equilibrium, with constant semantics . . . . . . . . . . . . .
4.4.3 Advanced-retarded spacetime propagation and addressability . . . . . . . .
4.4.4 Splitters and oriented discriminators, encoding decisions . . . . . . . . . .
4.4.5 Role of separated timescales in tokenizing . . . . . . . . . . . . . . . . . .
4.5 Nominally addressed memory . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Long range ordered spacetimes: cellular automata . . . . . . . . . . . . . . . . . .
4.7 Graph dynamics, thinking, and reasoning . . . . . . . . . . . . . . . . . . . . . .
4.8 Agent phase and conditions for memory, smart behaviour . . . . . . . . . . . . . .
4.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
34
35
35
37
38
38
39
39
41
43
43
44
44
45
45
46
47
47
49
50
52
52
The structure and semantics of knowledge
5.1 Learning, familiarity, and linguistic tokens . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Can familiarity be passed on? . . . . . . . . . . . . . . . . . . . . . . . .
5.1.2 Tokenized knowledge: topic maps and semantic indices . . . . . . . . . .
5.1.3 A minimal observational ontology for knowledge, based on promise theory
5.2 Labelling and naming conventions: the observer’s view of the world . . . . . . . .
5.2.1 Interior and exterior spaces of an observer . . . . . . . . . . . . . . . . . .
5.2.2 Interior and exterior agents . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.3 The named roles in knowledge representation . . . . . . . . . . . . . . . .
5.2.4 The observer’s clock, and sense of time . . . . . . . . . . . . . . . . . . .
5.3 Concepts, interior structure, and semantic superagents . . . . . . . . . . . . . . . .
5.3.1 Encoding or realization of concepts (reification) . . . . . . . . . . . . . . .
5.3.2 Discreteness of concepts . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.3 Role of context in discretization . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
53
53
53
54
54
55
55
56
56
57
57
58
59
60
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5.4
5.5
5.6
6
7
8
5.3.4 How are concepts acquired? . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.5 Concept feedback, or introspection about THINGS through things . . . . . .
Associations defined . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1 Semantic addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.2 Concurrently active concepts . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.3 Tensor structure of association . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.4 Associations may connect any agent or scale . . . . . . . . . . . . . . . . .
5.4.5 How are associations identified? . . . . . . . . . . . . . . . . . . . . . . . .
5.4.6 Flooding, routing, and association . . . . . . . . . . . . . . . . . . . . . . .
5.4.7 Autonomy and association (a notational caution) . . . . . . . . . . . . . . .
5.4.8 Quasi-transitive or propagating associations . . . . . . . . . . . . . . . . . .
5.4.9 The irreducible positive associations . . . . . . . . . . . . . . . . . . . . . .
5.4.10 Non-transitive semantics in associations . . . . . . . . . . . . . . . . . . . .
5.4.11 Embellishing associations for deeper inference . . . . . . . . . . . . . . . .
Context as a conditional switch for associative memory . . . . . . . . . . . . . . . .
5.5.1 Conditional association and structure of context . . . . . . . . . . . . . . . .
5.5.2 Constructing embellished or contextualized associations . . . . . . . . . . .
5.5.3 Causation and aggregation are not completely orthogonal . . . . . . . . . . .
5.5.4 Information argument for separation of context . . . . . . . . . . . . . . . .
5.5.5 Context atoms: coarse grained compression to awareness of ‘now’ . . . . . .
5.5.6 Spacetime template patterns for efficient recognition and recall . . . . . . . .
5.5.7 Stability of contextual knowledge and weight maintenance (sleep functions?)
5.5.8 Sensory perception: from context to reasoning . . . . . . . . . . . . . . . .
Spacetime representation of scaled concepts . . . . . . . . . . . . . . . . . . . . . .
5.6.1 Examples of irreducible positves . . . . . . . . . . . . . . . . . . . . . . . .
5.6.2 Noun and verb concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.3 Irreducible positves and negatives (encoding NOT) . . . . . . . . . . . . . .
5.6.4 Emotional origin to logical negation? . . . . . . . . . . . . . . . . . . . . .
5.6.5 Invariant representations and its relationship to tokenization . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
60
61
62
62
62
65
65
66
67
67
68
69
71
72
73
73
73
74
75
78
79
79
81
82
82
84
84
85
86
Inference, logic, and reasoning in semantic spaces
6.1 Reasoning from learning . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Associative model of reasoning . . . . . . . . . . . . . . . . . . . . . .
6.3 Couch’s rules for inference . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Stories and narrative reasoning . . . . . . . . . . . . . . . . . . . . . .
6.5 Story types, classified by their promises and spacetime associations . .
6.6 Inferences from algebra of association: type symmetry transformations .
6.6.1 Questions and answers: contextual stories . . . . . . . . . . . .
6.6.2 The storytelling hypothesis . . . . . . . . . . . . . . . . . . . .
6.7 Semantic commentary of structures . . . . . . . . . . . . . . . . . . . .
6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
87
87
88
91
91
92
94
96
97
97
98
Smart spaces
7.1 Interactive spaces . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 What does smart mean? . . . . . . . . . . . . . . . . . . . . . .
7.3 Promise view . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4 Benefits of a knowledge representation in real space . . . . . . .
7.4.1 Memory . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.2 Sensor types and dimensionality of context . . . . . . .
7.4.3 Spatial reasoning with concept and association structures
7.4.4 Temporal reasoning with pathways and stories . . . . .
7.4.5 Locality, centralization, and economies of scale . . . . .
7.4.6 Central controller versus brains . . . . . . . . . . . . .
7.4.7 Innovation in smart spaces . . . . . . . . . . . . . . . .
7.4.8 Auto-reasoning and searching for answers . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
98
99
99
100
101
101
102
103
105
107
108
109
111
Summary
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
111
A Coordinatization of data streams
113
3
B Invariant representation by spacetime aggregation example
1
116
Introduction
A spacetime is an extended network of interconnected agencies that collectively support the notions of location,
time, direction, and dimension. In a semantic spacetime, special meaning is attributed to locations and times; a
semantic spacetime adds interpretation, and even function, to the experience of space by an observer. Semantic
spacetimes may be models for everything from real structures to virtual collections of objects, in either a solid
or a gaseous state of organization: buildings, warehouses, databases, computers, machines, organisms. Semantic
spacetimes are associated with knowledge systems; however, we should not jump to conclusions about what this
means: a Turkish bathhouse, or a supermarket, could be representable as a knowledge space, not only the library
in Alexandria. As such, semantic spacetimes are intended as a universal theory of functional interpretation within
any extended system, that behaves as a natural generalization of the physics of material scaling to more generally
immaterial worlds.
In paper I of this series [1], it was shown that, if we take a collection of autonomous agents that promise to
collaborate in the construction of a spacetime, then it becomes possible to construct semantics in a simple way,
with clearly defined laws. Paper II [2] showed that, by separating promises that refer to local (scalar) properties
of agents from those relating to relationships between agents (tensors), we identify that a promise spacetime has
both scalar (material) properties and vector (adjacency, cooperative, non-self) properties, and hence we have a
description that scales as a natural generalization of the physics of coarse graining: we can combine agents into
super-agents, or coarse grains, without loss of information as long as we account for both reducible and irreducible
promises. A complete definition of scaling opened the way to a natural definition of tenancy, i.e. the hosting
of functional agencies within a space, with properties or attributes implemented over a region bounded by any
criterion (not necessarily a spatial boundary). Limited range promises imply a horizon for context, in which
dominant ‘rules’ or behavioural patterns may differ across cells. This is also a basis for defining and understanding
biological organisms as smart agencies. It forms a simple basis of a spacetime interpretation of shared context, and
categorization.
In this final part of the promise-spacetime manifesto, I want to build on all these notions to explore how
extended, differentiated structures can interact with one another to encode and encode knowledge. Such structures
would find use both as repositories of information, and as functional machinery, and as ‘smart spaces’1 . An obvious
application of this work is to understand the coming age of embedded information technology: the so-called
Internet of Things, Smart Cities, etc, which are non-simply connected spacetimes of significant scale, complexity,
and inhomogeneity. The outline for the topics is as follows:
1. A review of spacetime and its semantics.
2. A description of data as semantic elements, in the language of promise theory.
3. How memory can be constructed, addressed, and used, in order to equip spaces with a capability for data
representation and learning.
4. How knowledge is acquired and represented, by summarization.
5. How reasoning occurs in the shape of constrained spacetime trajectories.
6. The application of the foregoing ideas to functional spaces.
2
The structure and semantics of spacetime
The artificial spaces we create are filled with meaning. Parking spaces are different from living spaces; a beach
is different from a cliff. Meaning is partially engineered, by repeated trial and error, and is partially serendipitous
in its origin. To evolve something to the point where it becomes ‘fit for purpose’, we accept that there must be
selection criteria that prefer one functional form over another; these criteria take on the role of goals, some of
which are engineered and some are emergent, based purely on observed success in a functional role.
This repetitive process, which leads to an island of possible temporary stability, is called learning, and it take
the form of an interactive relationship between and agent and its subject, whose equilibrium outcome is what we
1 This includes but goes beyond the current interest in so-called artificial intelligence, e.g. using Artificial Neural Networks (ANN) to recognize complex patterns. Such a neural view of reasoning is too limited and too specific to represent a comprehensive approach to engineering.
4
call knowledge [3]2 . Knowledge may decay as quickly as it has been won: it requires maintenance, and we can
talk about its stability, i.e. its usefulness or applicability to a context. In this section, we explore how this process
occurs, with or without the intentional direction of human involvement.
2.1
Semantic spacetime
A semantic spacetime is a structured collection of semantic elements (see section 6.6. of paper I) [1]. To make a
spacetime, with both scalar (material) attributes and vector (adjacency) bindings, from the atomic building blocks
of autonomous agents, one defines the basic elements in a semantic space:
Definition 1 (Semantic element) A semantic element is a tuple hAi , {πscalar j , . . .}i consisting of a single autonomous agent, and an optional number of scalar material promises.
A semantic element is an autonomous agent (in the sense of promise theory), surrounded by a halo of promises
that imbue it with semantics (see fig. 4). Examples include any objects that could be relevant within a functional
context: people, machinery, and passive structures like buildings or trees.
Definition 2 (Semantic spacetime) A collection of semantic elements, in any phase (gas or solid) [1], for which
a local change in state, promises or configuration represents a local unit of time.
There are multiple reasons to want to define semantic spacetimes. The definition is a direct match for the day to
day workings of materials, chemistry, biology, and information technology. More generally, the reasons relate to
the communicability of knowledge and functionality between interacting agents:
• To find the requirements of a low level representation for concepts as spacetime patterns.
• To understand arbitrary spacetimes as functional machinery, with conceptual utility.
The field of linguistics provides the essential clue that this agenda is possible, since languages fit both of the
requirements above. Indeed, what we are looking for is a language, but a formal language that can be used to
understand the design of machinery with purpose, or discern the relative purposes of systematic mechanisms.
Semantics are one part of successful functioning, but they are static relationships. To utilize semantics for
purpose we have to place them into a dynamically stable context. This is where the concept of knowledge becomes
important.
2.2
The concept of knowledge as an iterated equilibrium
The concept of knowledge has special status. It is defined, in these notes, by analogy with the knowing of a person
or a friend [3]. Knowledge is not something we acquire by hearsay, not something acquired by reading once in
a book. It is the result of an ongoing relationship of repeated assessment, made by an agent, over time. During
repeated assessment, an agent gauges the relative importance of outcomes and their stability to change. Knowledge
is thus quite different from data: it is a summarized aggregate representation of data, to which we have attached
an interpretation of significance. It has been confirmed iteratively to the point of trust [5], by repeated sampling.
Knowledge is not ‘big data’ or bulk information in raw form: to sieve out knowledge from an extensive stream of
samples, a reduction over equivalences of quantitative and qualitative measures has to take place [3].
The repeated sampling of semantically equivalent data by an agent is called learning. In order for there to be
learning, there must be at least a promise of offered observability (+), and the receiver must promise to accept or
receive (-) data faithfully. The integrity of a measurement cannot exceed the overlap of these two promises [6].
observed
+attributes
−−−−−−→
observer
(1)
observed ←−−−−−− observer
(2)
−attributes
The stability of every sampled observation made is uncertain, because an agent does not know how fast the world is
changing without repeating its observations. Agents choose a sampling rate Tsample which (by the Nyquist theorem)
should be at least twice as fast as the rate of change of the promise it is assessing, and hence also the rate at which
it is able to form a stable assessment of what it observes [3, 7, 8].
2 According the the UK government IT infrastructure library [4], the sequence DIKW (data, information, knowledge, wisdom) is the ladder
of increasing meaning that humans go through in the course of their working interactions.
5
Definition 3 (Learning about π) The sampling, equilibration, and summarization of observational assessments
concerning a promise π made by another agent, repeated over a timescale Tlearn > 2Tsample .
Let α(π) be an assessment of promise π, made by some agent. On each sample time t + 1, an observer agent
applies a learning function L to combine the previous average assessment E(α(π)t ) with a new sample α(π)t+1
to form:
E (α(π)t+1 ) = L (α(π)t , E (α(π)t ))
(3)
Learning defines a clock that ticks at a rate Tsample .
Definition 4 (Knowledge of π) A stable summary of the iterated assessment α(π)Tknow , of one or more promises
π, is formed by equilibration of the samples over a timescale Tknow 2Tsample .
Because knowledge defines a process with a timescale, the failure to confirm it relative to other changes leads to
its decay. Trust in knowledge needs to be replenished. This is the main difference between knowledge and data.
Lemma 1 (Knowledge decay) The uncertainty of knowledge is a geometric function of the time since last learning
took place. If L is a convex or linear function of its arguments, which weights old assessments less favourably than
new by a factor ` < 1, then an assessment α(π)τ from arbitrary time τ has a geometric rate of attenuation by `r ,
where r ' Tlearn /Tsample .
The proof follows from the iteration r times of a convex form, such as Lt = (1 − `)si + `Lt−1 . Let α(π)τ → x,
and each new sample be si , for i = 1, 2, . . . r. Then the contribution to Lr is `r x x. Thus knowledge must
be reassessed, and learning must continue as long as new assessments are available, else the uncertainty of the
knowledge If sampling stops, the value of L fails to track its value. For each sample missed, there is a cumulative
error which grows like (1 − `)r . A corollary is that the significance of data to an assessment decays at the same
rate with its relative age. By virtue of Nyquist’s theorem [7, 9], we can also say that:
Lemma 2 (Fidelity and learning rate) Learning can only represent source values faithfully if the rate of sampling is greater than twice that of the fastest rate of change in the data, i.e. 2/Tsample < ∂π/∂t.
From the definition of knowledge:
Lemma 3 (Knowability of information) A fact is knowable by agent A if the promise can be assessed by an
agent A as kept over multiple assessments.
where:
Definition 5 (Fact) A promise of information to be taken on trust.
2.3
The concept of knowledge spaces: spaces that can learn and represent knowledge
What do space and time mean to us? Let’s explore the semantics of spacetime further, to better understand how we
experience it. The scaling characteristics of promises were investigated in paper II [2]. One of the challenges facing
us, with respect to understanding knowledge systems as spacetimes and vice versa, is the separation of dynamical
timescales and spatial regions during observation and measurement (see figure 1). This is a stability criterion. In
dynamics, weakly coupled dynamical phenomena separate along distinct timescales, and thus the semantics we
build on top of them must do the same. A natural hypothesis is therefore that promises, which encode information,
separate into at least two timescales: I call these context (fast) and knowledge (slow), such that quickly assessed
context modulates more slowly accumulated knowledge, in order to adjust its relevance. From this, we see that
there is a set of timescales for knowledge to emerge, including:
1. Sample or observation rate.
2. The time to summarize observations (form sense of context).
3. Learning rate (accumulation of experience).
4. Knowledge lifetime (decay rate in the absence of observation).
5. Reasoning about knowledge (associative narratives).
The extraction of what we call reasoning, or stepwise narratives, built from the semantic relationships between
observations have to be identified from the resulting network of associations. This is the agenda for paper III.
What kind of data lead to the kind of functional knowledge about which we can reason? An ability to reason
takes more than single samples; it requires the learning of relationships between the things too. I’ll return to this
in section 4. For now, we may observe that there is a clear role for space and time in data representation, and the
hypothesis we shall explore in these notes is the idea that these spacetime aspects are, in fact, the only basis for
knowledge representation.
6
THE MIND
System I
Agents
System II
Agencies (semantics)
Quick
States
Slow
CONCEPTS &
TOPICS
CONTEXT
classes
Relationships
LEARNING
associations
SAMPLES &
MEASURES
Ephemeral
Persistant
THE WORLD
timescale (dynamics)
Figure 1: To find a dynamical basis for knowledge semantics, we need to identify separable timescales, from which we can
identify stable ‘knowable’ phenomena. Learning occurs at two timescales: short-lived convergent states we call ‘context’ about
‘now’, and long-lived associative bindings based on accumulated experience, that form the stories on which we base reasoning.
2.4
Spatial semantics
What things do we want to say about space? As described in paper I, a space may be a collection on unrelated
points (like a gas of disconnected bubbles); but, usually, we expect points to be interconnected allowing motion
from one point to another [1]. Adjacency is the most primitive irreducible promise that characterizes space. In
promise theory, adjacency is a directional concept: A may be adjacent to B without B being adjacent to A. This is
a deliberate feature that models functional behaviours like semi-permeable membranes, diodes, and authorizationaccess controls.
The semantics of location take on a variety of forms, representing four main concepts to do with space: topology, location, distance, and direction. Many of these concepts are intertwined. Most of our familiar conventions
for space and time are based on a history of thinking of a Cartesian theatre model, in which spacetime is a backdrop
on which interactions happen between agents, independently of absolute location, only relative to other agents3 .
This separation of spacetime and matter is rooted both in convenience, as well as in the agenda of making science
impartial. Einstein showed, however, that this separation is both misplaced and inconsistent (impartial does not
mean objective) [11]: the presence of material agents within spacetime is closely related, if not actually equivalent
in some sense, to the local promises made by spacetime itself. This hints that a unified description for spacetime
and matter would make more sense [1].
In promise theory, the alphabet of symbols, in a coded structure, is composed of the atoms of space:
Definition 6 (Location agents) These are irreducible sites that take up space and can emit and absorb signal
agents. They may not overlap.
Agents in the role of signal agent form the alphabet of scaled inter-agent language.
Definition 7 (Signal agents) They may be created and destroyed, subsequently emitted and absorbed, by location
agents. They can occupy the same space, since they end up and accumulate at end points.
2.4.1
Distance, similarity, and metrics
The notion of distance is related to the ability to establish a coordinate system on space. The challenge of any
coordinate system is to make locations easy to find, and give meaning to the distance between locations. Two
kinds of distance may be defined in a semantic space: metric and semantic.
3 For
a background on the concept of this philosophy, see for example [10].
7
Definition 8 (Metric (quantitative) distance) A measure of coordinate-similarity in position.
Example 1 (Metric distances) Various aspects of topology and metric distance play into how we understand
space.
1. Adjacency, connectivity, continuity, and dimension - characterizing allowed transitions, doorways, the
existence of dimensions or degrees of freedom, discrete or continuous. Which channels make it possible to
move? What is the smallest distance an agent can move? Are there disallowed transitions? Adjacency is a
cooperative concept (see paper I).
2. Location, position: (nominal) identifying and promising position requires distinguishable coordinates. Location does not specifically require numerical coordinatization of a space, as long as the naming of locations
is unique, e.g. using signposts. The villages of our rural past were distinguishable in spite of a lack of maps.
A description of relative locations depends on direction and metric distance.
3. Metric distance between locations: (cooperative) given nominal positions of agents, by cooperation the
agents can form a coordinate system with mutually agreed understanding of the meaning of distance. This
then allows comparative measures.
The concept of distance is related to dimension, because a projection of a spatial relationship onto a ‘screen’
of lesser dimension can distort the ability to measure or observe the distance. What seems to be close, in
one interpretation of the observed world need not be close in the representational facsimile world.
4. Semantics of metric distance: What we do mean by ‘near’, ‘far’, ‘local’, etc (qualitative approximate
discriminants).
5. Direction: favoured adjacency amongst the possible degrees of freedom at a location, often associated with
a coordinate representation, or an external relative marker that breaks a rotational symmetry, like fixed stars
(see paper I).
6. Containment: the constraining of locations within a perimeter or other boundary.
7. State or pattern: configuration patterns of internal degrees of freedom, whose effect may or may not be
visible in the exterior. Hidden variables are like embedded dimensions projected into a subspace of the
observer (e.g. Kaluza Klein theory). This is familiar in physics where properties like charge and spin
are represented as hidden, non-spatial dimensions degrees of freedom, generally characterized by their
symmetries4 .
Example 2 (Effectively fixed markers) The fixed stars move so slowly and insignificantly relative to the Earth,
that we can use them as effective boundary conditions, even though they break no symmetries. By pegging a slow
moving phenomenon (e.g. the perceived motions of the fixed stars) to something changing more quickly, one can
create an illusion of symmetry breaking, in the form of an effectively static boundary condition. The propagation
itself may be complicit in the assessments of promises (see figure 2).
There are no cooperative promises between the fixed stars, and yet they hold their formation in such a way that
can be used. Fixed stars may promise something resembling an amorphous solid, that provides a reference without
actual symmetry breaking, by virtue of an almost complete decoupling of timescales.
The fixed stars are a form of gaseous memory, joined together by our observation on Earth. They are (for most
intents and purposes) causally and functionally independent. The promises they are (or were) able to make to us
are rather simple and non-interactive. If this situation happened on a closer scale, we would not easily trust such
unidirectionality. Science places much trust in the constancy of the heavens5 .
When we use the term ‘quantitative’, most of us are conditioned to thinking of distances being real-valued continuum measurements. This may or may not be the case, depending on the semantics of the measuring scale.
Definition 9 (Semantic (or qualitative) distance) A measure of similarity in interpretation.
Example 3 (Semantic distances) Semantic distances are based on comparison of symbolic sequences:
• Hamming distance:
The number of symbols that differ between two sequences.
4 Kaluza-Klein
theory is not capable of representing the known symmetries as spacetime translations on a continuous differentiable spacetime, but continuum translations would not be the correct model in a discrete network.
5 ”I am constant as the northern star, of whose true-fix’d and resting quality there is no fellow in the firmament.” Julius Caesar (III, i, 60
62) Just before Caesar is stabbed in the back for his intransigence.
8
Figure 2: Propagation of signals (latency) can effectively break symmetries too by providing such a separation of scales as to
be indistinguishable from a fixed boundary condition.
• Hop count in an associative network:
An obvious measure of distance in a graphical structure in the number of hops between nodes; i.e. the
adjacency count. The concept of distance in a network of associations may simply be used to define a notion
of similarity of the concepts held at the nodes.
• Semantic hashing:
Semantic hashing [?, 13] has a specific meaning in the literature. It refers to an attempt to reduce text
in a document to a small feature space, based on importance of keywords and tags, in such a way that
similar documents have similar hashes. A hash promises a semantic summary of a superagent collection of
individual words. The metric properties of the feature space are preserved by the hashing, but the choice
of keywords is manual and ad hoc (see example 9). Similarity does not promise to represent semantic
relationships that are not derivable from the tags.
• Sparse distributed representations:
Mapping similar keys to similar locations, in a sparse encoding. Similar to semantic hashes.
• Semantic routing:
The goal we seek in classifying meanings is achieve a consistent semantic routing of observed phenomena.
This is partly the goal of ontology (see section 2.4.4). The ability to bring universal meaning to phenomena
is fraught with misunderstandings, however [14]. Methods fail to be universal because they impose an
ontology from the top down [3], rather than adapting to context.
Example 4 (Nominal addressing) By using semantic labels for locations, instead of metric coordinate, we
can develop different approaches to locating attributes. This idea has been exploited in Name Data Networking and Content Centric Networking [15, 16].
Example 5 (Semantic routing and ontology) Before the successes of machine learning and pattern recognition, ontologies were designed by hand, using logical languages like OWL [14]. This was unsuccessful,
as ontologies are not universal; they are spanning sets (see paper I). Today, ontology is assumed to be
discovered in context by machine learning methods like PCA [17].
Unlike metric distances, qualitative aspects are not continuously distributed, but they can still be measured (quantitatively) using the notion of sets and basis sets.
2.4.2
Locatability of assets (sorting and discriminating identities)
The ease of learning, representing, and then finding assets stored in a semantic space (e.g. storing and retrieving
concepts, from some starting key), depends on being able to reliably route keys to the same (or at least) similar
locations in a memory every time we detect them. If one could further arrange for similar concepts to map to similar
spatial locations, then an approximate sensation or idea might even trigger a more complete or specific memory,
and vice versa. This would be the beginning of what we mean by representing associative ‘lateral thinking’.
9
Example 6 (Closeness of related concepts) Must similar concepts be close together in a knowledge space? This
must depend on the encoding of the space and the mechanism for coordinatizing it. Without this mapping, there is
no a priori link between the two.
Recent work in Nature using MRI scans of English language readings showed how brain activations for words
were hashed to broadly cohesive cortex regions [18], with similarities even across different subjects. The authors
posited that English words fell into clustered meanings that they labelled broadly as the categories: visual, tactile,
outdoors, place, time, social, mental, person, violence, body-part, number. In current parlance, these could be also
called spacetime, state, self, senses, relationships, and danger. These results are based on single words, however
in Hawkins [19], memory is assumed to be formed as temporal fragments, so such individual words may not have
the significance of independent semantic atoms. Nevertheless, the results are intriguing6 .
2.4.3
Physical representation of semantic differences
Should semantic distances be proportional to metric distances when representing knowledge as space? Spacetime
accounting principles, like the conservation of energy, and principle of least action suggest that it would be natural
for this to be the case; however, that presumes a continuum model of space. In network structures, small distances
can be quickly amplified into large ones by branching. If we encode processes for structural discrimination, which
sort encoded inputs, these must lead to bifurcations, which in turn can lead to wide (exponentially diverging)
separations at low cost across several spacetime dimensions, so what is ‘close’, in one measure, might actually end
up appearing far apart by another.
Finding an object may or may not imply physical visitation, so the constraints on what ‘easy to find’ means are
relative to the agent concerned. In paper II, the role of directories (content indices), for locating promises within
agent boundaries, was described. A promise directory or index is a structure that acts as a transducer between
the semantics of promises and a metric map of resources. It plays the role of a lookup table or hashing function,
converting roles to locations for interior agents. Directories are intrinsically linked to dynamic and semantic
transparency.
Example 7 (Searching versus coordinatization) An item in a warehouse has to be retrieved by a physical visitation (e.g. person or robot), but retrieval of information about its presence may be looked up in a directory or
inventory.
To map out a space, we need markers and/or graduations to measure distance. Markers that select particular
locations, breaking translational symmetry, are used to locate absolute points. Inequivalent semantics thus break
translation symmetries. To map out knowledge in spacetime, we have only spatial degrees of freedom with which
to encode persistent memory. The distance between locations is significant to the ability to retrieve encoded
information, as it leads to notions like association and locatability.
Example 8 (Map of spatial semantics) How do we find a shop in a city, or an item in a supermarket? Is there
some directory or hashing function, or a funnel that routes messengers to directed locations?
Example 9 (Discriminating and sorting structures) Coin sorting machines sorts coins and eject them into different piles. Passport queues discriminate people by their countries of origin. Baggage handling systems discriminate luggage by origin flight number.
An observer examining another observer’s semantic space might see a projection of points that appears to
bear no resemblance to the relative separation of the original spatial relationships. A metric preserving map M ()
exhibits isometry, i.e.:
D0 (M (x), M (x0 )) = D(x, x0 )
(4)
Conformal maps preserves angles, shear map preserve areas, and so on. We have to understand these spacetime
concepts and translate them into knowledge representations to appreciate the effects they transmit to our comprehension of what we believe to be recorded knowledge.
2.4.4
Metric and semantic coordinate bases, spanning trees
The spanning of graphs and vector spaces with metric coordinate systems was discussed in paper I [1]. Spanning
trees are one of the main approaches to the addressability and navigation of discrete spacetime structures, as they
form natural hierarchies for tuple-based addressing. The same approach may be used for semantic decompositions,
with nominal rather than ordinal values.
6 At
the time of writing, some of these MRI results are uncertain due to the discovery of a longstanding bug in the software.
10
Example 10 (Nominal coordinates) Computer file systems, web documents, markup languages, or document
data formats, like XML and JSON, all have a regular hierarchical structure from the root of each document.
This may be used to define a coordinate systems, based on the names and ordered numberings of identifiable objects within the document. Locations are specified by URI, URN, URL path identifiers. These path locators provide
nominal semantic coordinates based on a spanning tree formed from the markup elements. This is a generalization
of hierarchical file naming, e.g. C:\parent\children\sarah.
In order for similar things (semantics or meanings) to map to similar locations, the encoding of properties that form
the key must promise to be similar too. Thus ‘car’ and ‘lorry’, as words, would not be likely to occupy a similar
neighbourhood, but if both are associated with wheels and transport, then they can be. Hierarchical classification
can thus play a role in ‘namespacing’ concepts, as we shall see in section 5.1.3. This is the basis for taxonomy by
the idea of ‘prototypes’ [20].
Definition 10 (Taxonomy) A directed acyclic graph (tree) of name agents forming an ordered hierarchy of concepts and exemplars, with the level of generalization increasing towards the root of the tree, and sub-categories
diverging towards the leaves.
Tree structures apply a discrimination criterion to separate and resolve mutually exclusive branches, and separate
conceptual branches into ‘namespaces’ (see paper II). This forms a path-like coordinate system, as is common
in computer file systems, with directories for each nested superagent. A generalization of taxonomy, as used in
computer science, is found in attempts to apply first order logic to relationships between named concepts. This is
called ontology. The specific meaning of ontology has evolved since the widespread use of machine learning.
Definition 11 (Ontology) A network of agents, representing concepts, contexts, and semantic associations, whose
spanning tree forms a taxonomic hierarchy.
It follows from basic promise theory axioms that each agent is free to decompose the world, according to its
perception, into (super)agents by its own criteria and resources [2, 6]. Similarly it is free to associate meanings
to their exterior promises. As one forms spanning trees of such named agencies, physically or conceptually, by
annotating the representation of the physical domains with names and labels (see section 6.7). As an approach
to addressing, taxonomies may be inefficient, if they become too deep because they must be parsed serially and
recursively. Fast hashing aims to reduce the dimensionality of a space for approximate match. Shallow, broad
trees, or balanced trees are often preferred as data storage structures for their predictability and cost during parsing
[21, 22].
2.4.5
Symmetry breaking and deterministic descriptions
Purely relative, differential characterizations of a system (such as the propagation rules we refer to as laws of
motion), are essentially empty of content, in the sense that they make vacuous predictions based on known symmetries and constraints. Without fixed boundary information, e.g. about initial or final states, no predictions can
be made, because there is no content, only relative constraints. A propagation rule can thus be made to predict
almost anything; only sources and sinks of information are non-invariant, symmetry breaking, singular states and
transitions [23] that can anchor these predictions to some fixed information. Boundary markers inject information
into a spacetime description (see section 3.11.2).
The kind of deterministic prediction, which follows from a propagation of boundary conditions, tempts the idea
of a rigid universe, with mechanical predictability. The classical notion of Newtonian spacetime is effectively such
a solid-state lattice of points, with rigid adjacencies. It brings a mathematical predictability to spacetime paths: if
we traverse a link in a known direction, we can count where we are in this predictable world. We are so used to
thinking in this way that it is tempting to believe that we cannot represent or encode much information without
such a rigid spacetime. However, this is not the case. Even agents in a fluid phase may imitate the information
in a solid phase, if there is a sufficient separation of timescales. Example 2, about the fixed stars, indicates how
this comes about in a well known astronomical case. Other forms of dynamical symmetry breaking are found in
biology (where membranes and differentiated regions of cells combine from differentiated scale interactions), and
in weather patterns (where clouds, cells, and fronts lead to the formation of spatial cells of high and low pressure
or moisture).
The separation of timescales is how effective local boundary conditions emerge in a dynamical system, allowing symmetry to be effectively locally (dynamically) broken for all intents and purposes, at one timescale, while
actually remaining globally invariant in the greater reckoning. However, there is a conundrum here when we think
of systems as discrete information: how can we understand timescales, when the very clocks that define them are
made of the very stuff they measure.
11
2.5
Temporal semantics
Time is the rate at which an observer experiences change, i.e. the rate at which configurations in space are perceived
to change. We don’t need to know why spacetime elements change, but when they do, an observer’s clock ticks.
How an observer perceives time may depend on what context or state of mind the observer is in (e.g. awake,
asleep, busy, idle, and equivalent non-human states). This contextual aspect to temporal interpretation affects an
observer’s ability to assess and characterize existing promises and changes; it affects the granularity of time’s
agency, and its sequential ordering.
Time cannot be defined without a semantic space (i.e. some collection of agents that can represent state),
because time does not happen unless there is a clock to measure it. Thus, without spacetime semantics, there is no
natural justification for a relationship between space and time, and time could not exist in the way we understand
it. To define the smallest clock we need sufficient number degrees of freedom to represent uniquely distinguishable
states. Intervals of time are often defined ad hoc, based on a standard scale for speed, in the mathematical tradition
of imagining time to be an independent theatre, but Einstein showed us that this is not a consistent viewpoint: we
must understand that clocks are made of space, and thus space (agency) and time are inseparable.
Example 11 (Stories, narratives, world line knowledge) Spacetime structures that relate to knowledge modelling
include processes, concepts, associations, contexts, etc. These form storylines or narratives, which in turn describe
an experience of ‘proper time’ for an observer7 . For example: a person comes into the entrance, goes through a
changing room that promises lockers for storage, then the bathing rooms, and so on.
Example 12 (Temporal semantics) The many interpretational semantics of time include the following examples.
1. Local metric/sequential time: (e.g. a stopwatch counting). In practice, all clocks have finite number of
configurations to use for memory and thus can count only for a limited interval After this, time must repeat,
though not necessarily in the same order on each covering.
e.g. we may construct a precise equational specification of metric time TC as a simple counting mechanism
that satisfies the algebra an image Alg(ΣΦ , EΦ ) of a clock C. This is a simple clock that wraps around to
zero when it reaches a maximum counting value. The value ‘1’ represents a tick of the clock. Let’s give a
detailed explication of this one case to illustrate.
signature ΣC
:
TC ∈ (Z)
0
:
→ TC
1
:
→ TC
→ TC
sorts
constants
Tmax > 1
:
operations
:
·
:
TC × TC → TC .
T +0
=
T,
0+T
= T,
equations ΦC
(∀T ∈ TC )
(∀T, ∃ − T ∈ TC )
T + Tmax
= T
Tmax + T
T1 + T2
= T
= T2 + T1 ,
T1 + (T2 + T3 )
=
(T1 + T2 ) + T3 ,
T + (−T )
=
0
(−T ) + T
=
0
end
(5)
This is a discrete representation of ZN where N = Tmax . Quasi-continuous versions of a clock can be
constructed by increasing the size of Tmax . We can extend this approach to specify formal semantics for any
measure on a collection of agents.
7 The
concept of proper time is well known in general theory of relativity as the time experienced by an observer along its path [24].
12
2. Clock/calendar time:
A tuple of metric time clocks that accumulate state in a cylindrical coordinate system e.g. (seconds, minutes,
hours, days, . . .), or ‘12:06:22’. Each dimension is a different granular decompositions of accumulated
state, representable as a metric clock. As the second hand of the clock wraps around, the next counter
advances, in a cylindrical wrapping.
Example 13 (Coordination by fixed clock) The temporal analogy to measuring positions relative to the
fixed stars, or fixed markers, is measuring time relative to a common clock. Agents can coordinate their
activities relative to a fixed clock, or (if they have sufficient internal time-keeping capabilities) they can
synchronize their own internal clocks by extensive peer to peer coordination.
3. Time as a state and as an agent:
A map Tstate of string names SN to clock or calendar times TC that characterize all of the times on a calendar
or clock times, e.g. ‘now’, ‘Monday morning’, ‘Friday afternoon’, ‘then’, ‘recently’, ‘last week’,... It could
be distance travelled from an origin, or amount of sand fallen.
4. Global metric/sequential time:
In which an entire system is its own clock, i.e. any change represents a tick of a global clock. This behaves
exactly like a transaction log or software version control system, in which a tick corresponds to an ordered
transactional ‘commit’ to a database i.e. each tick is the record of a change between adjacent spacelike
hypersurface.
5. Rate of change:
Distinguishability of temporal state, at fixed location, relative to an observer’s separate clock.
6. Time as a resource:
Do you have time?
7. Trajectories:
Open spacetime patterns: non-repeating spacetime path.
8. Orbits:
Closed spacetime patterns: circling, waiting, eating, sleeping.
Absolute and relative notions of time do exist (see [23] for a discussion in the framework of equational specifications and state machines). A discrete aggregate concept of time is essential to creating spacetime associations,
because if time had infinite resolution, it would be infinitely improbable that coordinate pairs would identify the
intended points.
Definition 12 (Concurrent interval) A collection of events, signals, or ticks, arriving either serially or in parallel, that are considered to occur ‘simultaneously’, during a single tick of a clock.
Clocks cannot have infinite resolution, as this would lead to the need for infinite information.
Example 14 (Concurrent data) Any discrete aggregation process leads to a concurrent time approximation.
Each time we bake a cake, we assemble inputs into a unit output. There is no discernable order to the mixture
at the granularity of the cake. Each time a population census is taken, the counting result stands as a single moment in time until the next one it performed, and there is no memory of the process by which the counting took
place. The entire process is a concurrent interval, coarse grained into a single output. The uncertainty about
associating concurrent events grows with the coarseness of the concurrent interval.
Concurrent intervals include discretizations of streams: network packets, single promises, packages, enveloped
messages, containers, words, etc. Any natural, discrete unit of aggregation may be coarse grained into an effective
unit of time.
The concept of concurrent or simultaneous events is only possible in a discrete spacetime. The discretization of
spacetime, whether one chooses to think of it as fundamental or a limitation of an observer’s faculties, is essential
to the ability to limit the resolution of measurement information, and generate repeatable sensory experiences that
can become knowledge. In this sense, approximation is the key to be able to represent stable knowledge.
13
2.6
The semantics of spacetimes and dynamical systems
For dynamical and functional interaction, space and time interact in inseparable ways, so it is convenient to simply
talk about spacetime. The temporal aspect of ‘world lines’, or paths that trace the linear experience of an observer,
play a significant role in staging experiences into storylines that we associate with meaning. The selection of a
preferred path is a form of directional symmetry breaking in spacetime. Direction is one dimensional, thus time is
one-dimensional.
In any extended structure, which links agents together in chains by adjacency, sequential partial ordering is
a necessary phenomenon. Thus, timelines, world lines, and storylines are a necessary outcome of understanding
interaction trajectories for localized observers8 . This unidirectional sequencing manifests in a number of ways:
• Bottom up representation of communicated language, as symbolic streams.
• Top down reduction of patterns to find atomic symbol alphabet (reductionism).
Processes that are bounded at a spacetime location lead also to a directionality:
1. Converging towards a fixed end state.
2. Diverging from a fixed start state.
2.6.1
Assigning coordinates
Ordering of agents and signals is a cooperative spacetime phenomenon (see paper I) that requires there to be symmetry breaking, i.e. the definition of a preferred direction in the degrees of freedom of a collection of agents. How
one assigns ordered labels to a collection of agents goes beyond the scope of this work, but has been considered
elsewhere9 . The use of a central coordinator or observer, connected directly like a routing hub, with the ability to
discriminate paths has many advantages, because it can be the discriminator of order as well as the agent by which
messages are routed. This obviates the need for the end node agents to promise a specific identity. A hub agent
would naturally double as a router for passing signals10 .
Example 15 (Cooperative waves) Waves that pass through agents by local or global cooperation are spacetime
relative phenomena based on symmetry breaking. For example, green waves that move through traffic lights to
keep traffic moving are timed relative to a fixed wall clock, or controlled by a single controller. All traffic lights
promise to turn green at a coordinated time relative to a fixed marker (like the fixed stars/fixed wall clock examples).
Similarly, fan waves at sports games propagate by observation of nearest neighbours.
2.6.2
Time series
Time series are a form of data that expresses a trend or a pattern over an interval of sequential times. The meaning
of time series depends on the semantics of the time-base. Time series may represent open ended trajectories or
closed circuit periodograms to respect and model the underlying freedoms of a spacetime trajectory [28–30].
Example 16 (Periodic time-series) Behavioural data form time series with distinct patterns (see figure 3). In
the figure, we can imagine measuring resource consumption in a workplace, or the rate of information flow in a
computer network. The time base covers a single working week from Monday to Sunday. When we reach the end
of one week of samples, we overlay new data at the corresponding time of weekday by modulo arithmetic. The
aggregations occur with the same temporal semantics for each interval, assuming that there is a repeated pattern of
behaviour each working week. Unbiased data measurements reveal strongly auto-correlated patterns that confirm
this pattern in many systems driven by human behaviour. By representing the pattern in the semantics of the timebase, we can effectively learn the temporal pattern efficiently, and compare only semantically equivalent times.
This temporal cyclicity traces back to the rotation of the planets.
Example 17 (Linear time-series) Discrete time series are found in symbolic strings, whose time series form syntax. In computational and linguistic literature, a grammar is an alphabet with a number of generative rules,
leading to computable patterns. In bio-informatics, grammars may be seen in DNA [31]. Grammars are classified
8 Even if the goal is to manufacture an artificial representation of a problem, such as the layout of a department store, such a structure has to
be able to learn and adapt by evolving in relation to interactions with the agents that parse its structure.
9 Unique addresses can be assigned to a gas of agents too in a self-organizing manner in ad hoc networks [25, 26]. If addressing comes with
crystallization of a addressable lattice, e.g. a collection of cells that self-organize into a regular pattern...then we need a notion of di-polarity or
orientation for the joining of independent fragments to bring about monotonicity in the numbering. In an already solid phase, the problem is
simpler, and the symmetry is broken by choosing an end point, e.g see the self-organizing spanning trees [27].
10 This is true in the design of IPv6, for example.
14
graduated data
Monday morning
Monday afternoon
Monday evening
Monday night
Tuesday morning
Figure 3: A time series measures sampled data against a semantically significant ordinal timescale. The ordering might be
open, bifurcating or periodic. The vertical scale resolution is the self-calibrated perception of sensors, and the horizontal scale
resolution is based on sampling frequency relative to some larger model that is identified as meaningful in this context (e.g. the
working week).
by the kind of computer needed to parse the pattern faithfully [32,33]. In promise theory we phrase it as a voluntary
cooperation construction: a grammar is a collection of agents and superagents that make promises, and whose top
level adjacency has the form of a one dimensional chain11 . Languages are related to algorithmic state machines,
according to Chomsky’s theorem [34]: they are spacetime phenomena, with concurrent words/symbols/statements
and temporal progression.
Lemma 4 (Inseparable space and discrete time (including time pattern)) The representation of a timeseries,
as a concurrent phenomenon, involves spatial pattern representation within a concurrent time slice.
Any repeated pattern marks out a discrete interval whose semantics represent a scaled concurrent phenomenon.
3
The structure and semantics of data patterns and representations
Underlying all communication is the ability to attach meaning to patterns. It is the basis of information transfer
in communications theory, linguistics, and computation. Grammars are scaling patterns of the agencies by which
linguistic symbols are carried, and projected into sequential (one dimensional) streams. Sequential streams are,
by their nature, temporal when communicated, but may be stored in a spatial memory. Chomsky showed that
grammars are associated with transition systems operating on discrete linear patterns [34]; from this we have a
theory of signalling and data encoding [32] that goes beyond Shannon’s basic information theory [9].
The encoding of information signals may take a variety of representations, of which natural languages are
the most complex forms. In information technology, it has become the norm to speak simply of ‘data’, as a
stream of symbolism for encoding by a plethora of underlying binary formats (ASCII, EBCDIC, UNICODE, least
significant byte first, binary-coded decimal, big-endian, little-endian, etc) is taken for granted, precisely because it
is of fundamental importance.
3.1
Information content of space
Consider a set of agents Ai , forming a semantic space. At each point, an agent may have an alphabet of promisable
characteristics Ci , with index αi = 1, 2, . . . Ci . We can define the distribution of promises by
pαi = p
0
+α
i
if αi is promised, i.e.Ai −−→
A?
otherwise
(6)
11 Human language is linear too, but thinking is not: we experience flashes of insight, in which it is not necessary to recount a story in our
minds, step by step (though we may do so later to rationalize it). This is evidence that thinking is not performed in the manner of a Turing
machine.
15
and
Ci
X
pαi = 1.
(7)
αi =1
Then the intrinsic information at point Ai is the Shannon entropy [9]
Si = −
Ci
X
pαi log pαi .
(8)
αi =1
We note that the information at each point is potentially based on a different alphabet Ci 6= Cj , and is therefore a
priori uncalibrated. In order to compare the information over a path or region of spacetime, a common alphabet is
needed (see also the discussion in paper II, section 2).
In order to information to propagate to neighbouring agents, e.g. to act as a fixed marker signpost, agents need
to have a non-empty overlap in their promisable states, i.e. a common sub-alphabet
Ai
Aj
+C
−−→
−C
−−→
Aj
(9)
Ai .
(10)
Thus, over an extended region or path, spacetime agents must have a minimum common alphabet of symbolic
promises (see also paper II, section 2). This issue is related to the notion of schemas in data.
Definition 13 (Translation invariance of information alphabet) A region of space may be said to be translationally invariant over its information alphabet if the alphabet of each agent is everywhere constant, i.e.
d+ C
i
= Ci+ − Ci = 0.
(11)
This is a non-local coordination, or long range ordering. See paper I about derivative d+ .
The injection of new information, at a particular location i, is an insertion of information that can be promised to
agents capable of recognizing the same alphabet.
Definition 14 (Non-local informational entropy) The distribution information of alphabetic type α is defined by:
X
SR (α) = −
pαi log pαi .
(12)
i∈R
This represents the intrinsic information in the variation of property α over a region of space.
Lemma 5 (Breaking translation invariance reduces information) If the space is translationally invariant for
all α, then it has maximum entropy log |R|. If the α invariance is broken, then SR (α) < log |R|, for any or all α.
This is proof is self-explanatory. It refers to changes in the promises at points within a region, for fixed alphabet.
A change in alphabet can also break translational invariance, as observed by an independent observer who can
recognize the alphabets of all the agents, however this change is not generally recognizable to neighbouring agents
whose alphabets are not equivalent, because they may only be able to recognize the alphabetic properties they have
in common.
3.2
Data and data types
Any encoded pattern, in any medium, represents data, and may be measured for its information content. We
should distinguish data as a promise from the promise concept of data containers or agency. Data are represented
by promised values, while the containers of the representation are agents. Thus, even non-scalar data type may be
represented by scalar promises (scalar refers to two different things here).
Definition 15 (Data) A stable pattern promised by any kind of agent or superagent structure.
The significance of the pattern depends on the overlap of promiser and promisee’s interpretation [6]. In promise
theory, agents are the only medium in which patterns can be represented, by design. Data are thus the pattern
values promised by agents. Data values fall into two major interpretations:
16
Definition 16 (Nominal data values) The mapping of patterns indicating into a set of symbols with distinct identity.
Definition 17 (Numerical (ordinal/cardinal) data values) Quantitative representations indicating relative magnitude on some tuple of generally real-valued scales.
Example 18 (Tuple spanning of values) The representation of numbers in the Arabic system uses tuples based on
powers of 10, using symbols 0,1,2,3,4,5,6,7,8,9. For instance, one thousand and twenty three is written (1,0,2,3)
or simply 1023, where the coordinate tuples map powers of ten (103 , 102 , 101 , 100 ).
Example 19 (Data containers) In computer programming languages, the agents that promise data are called
variables, constants, buffers, etc; in communications protocols, and documents, the agents are called fields, etc.
When communicating data, a sender S agent might copy data from one representation, in its internal memory, into
another, within signal agent, which is then emitted and absorbed by the receiver R.
Definition 18 (Scalar data) The promise of scalar data, e.g. X = 8 has the form
+8
AX −−→ R
(13)
In general data must be ‘typed’, i.e. must promise an intended interpretation in order to make sense to a recipient:
Definition 19 (Typed scalar data) A typed promise of scalar data, is a body doublet including an intended type
and a value, e.g.
+(integer,3)
−−−−−−−−→
AX
+(string,3)
−−−−−−−→
AY
R
(14)
R
(15)
Aggregate data are promised in the same way.
Definition 20 (Aggregate typed data) The promise of non-scalar typed data, e.g.
X = (integer1, integer 3, string ‘nine0 ) has the form
+(integer1,integer 3,string ‘nine0 )
AX −−−−−−−−−−−−−−−−−−−−−−→ R
(16)
Lemma 6 (Scalar representation) All data promised by a single agent may be represented by scalar promises.
The proof of this is trivial: promise body can contain any bundle of promises:
+b
A −−→ R
(17)
where b can represent any tuple: a scalar value, an array, a bundle of values, and even a number of internal agents.
In other words, scalar promises are not the same as scalar values. Scalar data refer to single named values. A scalar
promise refers to an internal property of a single agent [1]. We may also remark that agents that promise data form
Scalar
Scalar
Scalar
Scalar
Vector
A
Scalar
A
Vector
Scalar
Vector
A
Thing /element
Thing/element
Thing/element
Figure 4: Elements of semantic space may be viewed as autonomous agents surrounded by a cloud of scalar promises, joined
into a semantic web by vector promise bindings.
a semantic space.
17
Lemma 7 (Data form a semantic space) Agents promising data are the elements of a semantic space.
The proof is straightforward: a semantic space is a collection of agents that make scalar promises. They make
vector promises to coordinate in groups. Data values are spacetime scalars, whether single valued or aggregate
types12 . Objects promise non-ordered adjacency (superagency) in clusters of non-similar agents. Arrays promise
ordered adjacency in clusters of similar agents. The phase of the space (i.e. its adjacency structure) need not be
defined.
3.3
Schemas, types, and the chemistry of composition
If combinations of atomic data are received on a regular basis, it makes sense to bind them into homogeneous
‘molecules’ of fixed types, which can be recognized and stacked efficiently in space. Aggregated data thus form
superagents that promise multiple internal data items under a single identity. The members inside form a ‘schema’
which is given by the superagent directory [2]. This leads to a simple form of hierarchical containment. Heuristically, schemas are molecular scale superagents, formed from atomic agents.
Definition 21 (Schema, data class, or compound type) A schema S is a bundle of sub-agents and concurrent
promises within a superagent. A schema is thus closely linked with the superagent directory, revealing its specific
internal structure.
A schema acts as a declaration of constraint on the form of a superagent’s composition, without fixing the values
of any instance of the pattern.
Example 20 (Data schema) An object X may contain internal members, whose directory is a set of promises
about structure, but not value.
X : {
integer subagent1;
real
subagent2;
string subagent3;
}
The schema promises three sub-agents, each of which makes scalar promises about their type, from an alphabet,
which contains at least integer, real, and string.
Definition 22 (Schema instance) An manifest superagent whose directory is assessed to be an instance of schema
S.
Schemas are well known in database theory, where tabular structures are ubiquitous [35].
Lemma 8 (A schema has constant semantics) A schema, and all instances of a schema imply constant semantics
for each distinct promise.
As key-value pairs, the semantics of keys are constant, but the values of data are not. The purpose of a schema is to
promise constant semantics across a collection of data, whose values may vary. This promises semantic calibration,
ensuring that data can be compared meaningfully.
Example 21 (Schemaless data) Data that do not have a homogeneous spatial template are schemaless. Their
representation does not have a bounded alphabet. A JSON file used as a specification with an ad hoc schema
(sometimes called schemaless) has only a single data point, so we have no statistically significant basis for building
a relationship with it, unless it is re-iterated intentionally to underline its authority. This kind of data may crave
greater trust, based on an external authority, to reliably interpret the schema.
Example 22 (Fixed or homogeneous schemas) A repeated schema, such as sensor data in the same format, but
which never repeat the same value are uncertain in value, but one has increasing confidence in the meaning of the
schema.
The formation of a schema is a form of spacetime symmetry reduction: agents take on a restricted pattern, instead
of one exploring all possible freedoms, which may be then repeated like a template. This is an alphabet reduction
in the informational sense of section 3.1.
In a space whose agents promise a homogeneous schema, locations may not be equivalent in their data values,
but locations are semantically equivalent in their alphabets. There is still residual symmetry between agents with
12 The
term scalar is also used about single value data types, but that is a different meaning. Here we mean the tensor meaning of scalar.
18
the same schema. If agents are schemaless, there is complete symmetry breaking, with neither a common alphabet
nor a common value. The breaking of data symmetry establishes identifiable anchors or signposts by which data
can be compared and aligned, but the ability to use these as anchors assumes that neighbouring agents can recognize
the inhomogeneity, and position themselves relative to it.
Lemma 9 (Schema translation) If a collection of agents has different schemas, they may interact with external
agents via any agent whose schema alphabet overlaps with all of the different schemas.
The proof of this follows from the propagation criterion in (10). An intermediate agent can select and expose only
those parts of a schema that are common, and present a calibrated interface to the irregular schemas.
A schema alphabet is a simple grammar. The information contained in its labels may be used to separate data
into different contexts, as it is parsed sequentially. If a schema changes over time, that too could indicate a change
of context in the data, but it could be a challenge for the semantics of the data. In learning, we normally wish to
compare to similar data at several times, with a calibrated schema, so that we may assume the data have the same
interpretation on each sampling (this is the cornerstone of the scientific method). If a schema changes, there is
additional uncertainty about data interpretation, which cannot be easily quantified. The need for a stable schema
thus depends on the extent to which one expects to form a stable iterative knowledge relationship with a data
source.
Rare events
Ad hoc search/detection
No representative
Variable schema
Regular events
Specialized sensor/cached assumptions
Represented by an average
Fixed schema
The usefulness of cooperation over schema calibration may change the criterion for spacetime organization.
Different types of agent (i.e. agents with different schemas) lead to compound that express a functional chemistry
with different functional properties.
Example 23 (Components and compounds) Electronic components fall into schemas we could call resistors,
capacitors, inductors, transistors, etc. For the purpose of searching it is natural to store instances of the same
schema close together in a semantic space, e.g. in a warehouse. When exploiting the chemistry of interaction, e.g.
to build a television, the chief concern becomes to mix these components efficiently, so that their unique functions
are near to where they are needed. Then keeping all resistors together is no longer supports the intended semantics.
3.4
Type calibration, bases, and matroids (recap)
In a promise graph, scalar data type promises may be represented as vector promises, connecting to an authoritative calibration agent (matroid basis), which acts as a ‘standard candle’. Or, they may be calibrated by mutual
equilibrium over a collection of agents that claim similar promises (see figure 5, and the discussion in paper I).
For the former case, each such singularity represents one unique scalar property (see figure 5): this is how one
forms a matroid or spanning set for nominal properties (see [1]). Thus conceptual linkage becomes associated with
spacetime singular agents, or anchors, that ensure mutual calibration, and semantic stability. The alternative of
mutual linkage into a complete graph involves labelling not just a single agent, but pairs of links between all the
agents, with O(N 2 ).
R
G
B
3
3
2
2
4
Blue
(1,0,0)
1
(4,1,0)
(2,0,0)
4
(5,2,0)
+You are Blue
− Yes, I’ll buy that
(3,0,1)
1
Figure 5: Calibrating a type property, either by equilibrating with a single reference source (singularity), or by normalizing a
clique.. Right, an explicit matroid basis for global properties with three property hubs adding three components to the coordinate
tuples.
19
The desire for uniqueness and stability often leads us to construct the single source model for property/type
attribution. This makes concepts rigid, localized, and standardized rather than plastic, distributed and subject to
shifting equilibria. Equilibration with a central source has a cost advantage (O(N ) versus O(N 2 )), and turns
the scalar property into its own dimensional coordinate index, as there is now a direct route from this to every
instance of it (see figure 5). This choice of basis is related to the centralization or decentralization of knowledge
representations.
Definition 23 (Symmetry breaking) The loss of locational equivalence under arbitrary spatial translations. Spatially distributed promises take on only a restricted set of values or patterns, compared to what is otherwise possible
at only a subset of particular locations.
In this sense, symmetry breaking is involved in the crystallization of a gas with arbirary positions to a solid state, as
well as any introduction of edges, boundaries or markers, whether by equilibration or singular agent representation.
3.5
Trust and certainty in data calibration
Accepting an authoritative matroid basis-agent as a calibrated standard for a spanning property assumes a level of
trust. Trust is acquired by a learning process, i.e. it is based on extended relationships [5]. A promise theoretical
subtlety here is that the property represented by a matroid basis agent (e.g. blue in figure 5) is promised by the basis
agent, and is accepted by the associated clients. Any agent could set itself up as the authoritative source for ‘blue’,
and stand in competition, e.g. namespaces and scoped blocks limit the reach of definitions. This becomes particular
relevant as we consider representations of authoritative identity in humans spaces , such as homes, communities,
and cities.
Example 24 (Types in natural science) What agency is the authoritative source for the scalar property of being
an electron?
3.6
Data dimensionality: aggregates
Data are processed serially (each transaction generating the notion of system time); however, concurrent promises
or parallelism may be used to represent data in parallel threads, and nested superagency may be used to bundle
orthogonal degrees of freedom as concurrent promises. Concurrent promises are thus the basis for aggregate data
representations.
We can use the scaling arguments of paper II [2] to express data representations in terms of agents and superagents. By wrapping data inside layers of containers (agents within agents), we combine schemas, promise
independent names, and form namespaces to limit scope. Internal structure is promised by a superagent directory
or index [2].
Two kinds of aggregation may be formed from scalar values:
Definition 24 (Data structure or data object) A uniquely named superagent S, formed from uniquely named
subagents Ai , each of which promises values. Let S = {A1 , . . . , An }, where each subagent promises typed
scalar values
S
Ai
Ai
+name
S
−−−−−→
+namei
−−−−→
+(typei ,valuei )
−−−−−−−−−−→
R
(18)
S
(19)
R
(20)
Some structures may be transparent, while others maybe opaque to external agents.
Definition 25 (Transparent, typed structure or object) An object Sτ , of aggregate type τ is transparent, iff it
promises directory of members to R:
+directory
τ
Sτ −−−−−−→
R.
(21)
The directory of type τ contains the types and data member names of the superagent S. See section 3.7 of paper
II [2] on superagency and directories.
See paper II for a description of directories.
Example 25 (Data object) An object X may contain internal members:
20
X : {
subagent1 : 3,
subagent2 : 4,
subagent3 : 5
}
If the object is opaque, it promises only its exterior name X, and the interior structure is not promised, i.e. it is
unknown to outside agents. We might then call X a ‘blob’. If it is transparent, there is a directory of its internal
agents and their promises:
X.subagent1 = 3
X.subagent2 = 4
X.subagent3 = 5
Objects are inhomogeneous collections of subagents that promise any kind or type. Arrays are totally ordered
collections of agents promising symmetrical types (see section 5.4 of paper I [1] concerning ordering).
Definition 26 (Array) A totally ordered collection of semantically equivalent, symmetrical objects, under a single
common name. Each member of an array makes all the same type promises, as the others, but may have different
values. A uniquely named superagent S, formed from uniquely ordered subagents Ai , each of which promises
different values of the same type. Let S = {A1 , . . . , An }, where each subagent promises typed scalar values
Sτ
Ai
+(+τ,nameS )
−−−−−−−−→ R
+(Sτ ,namei )
−−−−−−−→
R
(22)
(23)
The type and value may themselves be uniformly compound objects.
Each element in an array thus promises:
• Membership to the compound ordered structure.
• A value promise, indicating its primary function.
• An index number or name promise (elements may be associative, i.e. nominal rather than ordinal in general),
indicating its intended position within the collective.
Example 26 (Data array) An array X may contain internal members of any type, including composite objects,
as long as they all have the same schema. Suppose the type τ is a doublet object (integer, real)
X[0] = (integer, 3) (real, 2.72)
X[1] = (integer, 4) (real, 3.14)
X[2] = (integer, 5) (real, 1.0)
The dimension of an object class is the number of scalar degrees of freedom, promised by it. The dimension of a
representation is the number of promised scalar values, or occupied slots. In an instantiated object, all degrees of
freedom are filled, thus the dimension of an instance could be considered zero, but it is more useful to refer to the
dimension of its schema.
3.7
The role of agent phase in data representation
If we follow the axiomatic approach, and represent each independently variable property of data by a separate
agent, perhaps within the collected boundary of a superagent, then the three levels of data aggregation are represented by the ordered and disordered material phases of semantic space:
• Scalars are atomic, ad hoc agents with only external promises.
• Objects are composite scalars, i.e. connected molecular superagents, with both internal and external promises,
and inhomogeneous composition.
• Arrays are homogeneous lattices of scalar promises, i.e. agents in crystalline lattice, with homogeneous
long range order. Arrays offer limited translational symmetry, with a homogeneous schema, by chaining
together agents into a regular ordered structure.
21
Example 27 (Promise graph vs JSON) Consider a typed data stream (e.g. JSON encoded data). We can extract
values and model their agency in representing functional behaviour in a system.
{
"kind": "Service",
"metadata": {
"name": "myagent"
},
"spec":
{
"ports": [ 8765, 9376 ],
"selector": {
"agent": "example"
},
"type": "LoadBalancer"
}
}
This stream is composed of a signal agent named ‘kind’, and two signal superagents called ‘metadata’ and ‘spec’.
The signal agents represent concurrent promise bundles within their interiors.
+service
kind
selector
agent
example
0
8766
9376
metadata
name
ports
LoadBalancer
1
type
Figure 6: Promise graph for the JSON data representation in the example 27. Scalar promises are simple aliases for the agent
names, while the aggregate superagent ports vector requires an additional solid-state vector ordering agreement between the
agents.
3.8
Data encoding relative to fixed markers (protocols)
Data may be transmitted as an array of raw values, or pre-annotated with labels and significant symbols, in the
form of a protocol. Smart sensory apparatus may perform preprocessing that annotates data values in a predictable
protocol, while raw sensors pass only simple signal sequences. Decoding the significance of data adds to its
semantic value and assists in building knowledge representations, by reducing the amount of recognition that is
needed relative to the behaviour of the sensors. How we utilize markers (signposts) to ground and punctuate
patterns is therefore of the utmost importance to understanding and deriving value from the structure of data. This
encoding and decoding is the realm of formal grammars, and grammatical protocol is the basis for language and
communication. Markers and delimiters break the equivalence of array positions for raw data, establishing anchor
points for aligning expectations between a sender and a receiver, i.e. a promiser and a promisee.
Definition 27 (Encoding) The establishment of boundary conditions in a semantic space that break free symmetries for the purpose of anchoring an interpretation of a collection of data promises.
22
In promise theory, scalar promises at fixed agent locations represent boundaries and boundary values (see paper I).
Data encoding involves an alphabet Σ. Some symbols are reserved with special meanings, e.g.
• Parentheses signify containment. Spatially, this plays the role of an embedded dimension.
• Arrays signify collections of agency that are equivalent under the membership, i.e. they have the same
semantics.
Definition 28 (Data stream or sequence) A data stream is a totally ordered set of signal agents emitted from a
source agent, and absorbed by a receiver agent.
A sequence transmitted from one agent to another involves a number of promises. The transmitted signal is
composed of signal agents that may or may not promise a total ordering. These promises are made both to the
sender and the receiver, else the sequence order cannot be transmitted faithfully. All symbolic language (including
phonetic language) depends on this construction.
1
2
S
3
R
4
5
6
Figure 7: A sequence is a spacetime coordinated promises. this is identical to figure 8 except that the source of the data
encoding is absent.
Example 28 (Parsing) In the traversal or parsing of streams, an agent13 assesses a stream of samples to identify
agents within the stream that have assumed meanings. These are called tokens in parsing. It searches a partially
ordered list of patterns (normally sorted by largest pattern first) and guesses the interpretation of these tokens.
Context free languages may have specific token orders, but do not attach different semantics to different spacetime
locations. Context specific languages (e.g. markup languages) may attach different semantics to different contexts,
or embedded regions within a stream.
Encoding may be serial (timelike) or parallel (spacelike):
Definition 29 (Serial, temporal encoding) In a temporal encoding there is an alphabet of symbols, some of which
are special symbols to mark out space and time: parentheses for spatial digressions. Item separators, like whitespace and punctuation, indicate the boundaries of agency. This is a stream or a protocol.
.
Example 29 (Non atomic symbols) Not all alphabets are atomic. In written Chinese, the symbols are formed
from a deeper level of atomic motifs, combined into a collections of overlapping symbols. Words too are symbols
that function collectively as concurrent data representations.
Definition 30 (Parallel, spatial encoding) In a parallel encoding, spatial dimensions are encoded either by direct
mapping in representation space, or by indexing with a tuple basis set. This is a memory array.
.
23
1
2
3
R
4
5
6
Figure 8: A possibly ordered collection of agents may represent aggregate data in a distributed manner. The interpretation of
the collection is entirely up to the receiving agent R.
Example 30 (Parallel distributed data encoding) Agents may cooperate with each other to represent structures
in a variety of ways (see figure 8). Parallel encoding of concurrent data, rather than sequential streams, are
common in disk striping, RAID configurations for disks, etc.
Definition 31 (Boundary markers and anchors (temporal punctuation)) Special points along a stream trajectory may imbue locations with a particular significance. Such anchor points include boundaries, signal delimiters,
parentheses, initial and final points, etc.
Example 31 (Anchoring by agent type (spatial signposts)) Any exterior promise (name, charge, spin, service
type, identity, etc) may be used to label an interstitial location within a homogeneous space. This is an explicit
symmetry breaking that injects information about location, by a reduction in non-local information.
Example 32 (Recursion markers) Parentheses may or may not be labelled to indicate different semantics. Inside
parentheses, special context dependent meanings may also apply, e.g. separators:
( , , ) [ , , ] {
3.9
; ; }
(* ; ;
*)
($ ; ; $)
Coordinates: spanning trees and tuples over data
Data may be coordinatized in two principal ways: by spanning tree, as a graph using nominal tuples, and as a
dimensional lattice or grid, using tuples (nominal or ordinal). Nominal tuples might look like this:
(grandmother,mother,child) -> scalar value
(grandfather,mother,child) -> scalar value
This may or may not assume ordering. Ordinal tuples have to assume an ordering:
(0,0,0) -> scalar value
(1,0,1) -> scalar values
An association with hierarchy is common, especially when the dimensionality is used to reduce spaces into subdivisions (see figure 9).
Spanning trees may be recursive by reference only, e.g. in linked type structures, objects can point to an object
of the same type, i.e. with symmetrical semantics:
type data_chain
{
integer value;
data_chain *pointer_to_next;
}
Example 33 (JSON and YAML data formats) The JSON data format is intended as a generic interchange format. JSON data structures are composed of three main kinds of value:
13 Such
an agent is called a parser or lexer.
24
document (root)
chapters
sections
left
right
paragraphs
sentences
parentheses
Figure 9: Spanning tree and data hierarchy. Ordering may be preserved by promising a policy of depth-first, left-first recursion.
• Named values (scalars), represented as assignments lval : rval.
• Named lists of values (arrays), delimited by [ and ], with comma separator.
• Lists of named values and named lists of values (objects or structs), delimited by { and }.
Order is mainly important inside arrays, where order is the only semantic distinction as the elements are homogeneous. These correspond to the agent structures above, identifying data as agents that promise both identities
(names) and values., e.g.
{
"kind": "Service",
"metadata": {
"name": "myagent"
},
"spec":
{
"ports": [ 8765, 9376 ],
"selector": {
"agent": "example"
},
"type": "LoadBalancer"
}
}
This structure can be coordinatized in a number of ways: e.g.
/kind -> Service
/metadata/name -> myagent
/spec/ports[0] -> 8765
/spec/ports[1] -> 9376
/spec/selector/agent -> example
/spec/type -> LoadBalancer
In this format, we understand the relational semantics as a spanning tree of absolute paths from a root node, and
the order is preserved as a ‘first in first out’ queue structure.
A less opinionated coordinatization may be assigned by ordinals. Here we cannot easily see the intended
naming of regions:
25
(
(
(
(
(
1,
1,
1,
1,
1,
1{,
1{,
1{,
1{,
1{,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
)[0]
)[1]
)[2]
)[3]
)[4]
->
->
->
->
->
"kind")
:
"Service"
"metadata"
:
( 1,
( 1,
( 1,
2{,
2{,
2{,
1{,
1{,
1{,
( 1,
( 1,
3{,
3{,
0,
0,
( 1,
( 1,
4{,
4{,
1{,
1{,
0, )[0] -> "ports"
0, )[1] -> :
( 1,
( 1,
4{,
4{,
2{,
2{,
1[, )[0] -> 8765
1[, )[1] -> 9376
( 1,
( 1,
4{,
4{,
3{,
3{,
0, )[0] -> "selector"
0, )[1] -> :
( 1,
( 1,
( 1,
4{,
4{,
4{,
4{,
4{,
4{,
1{, )[0] -> "agent"
1{, )[1] -> :
1{, )[2] -> "example"
( 1,
( 1,
( 1,
4{,
4{,
4{,
5{,
5{,
5{,
0, )[0] -> "type"
0, )[1] -> :
0, )[2] -> "LoadBalancer"
( 1,
( 2,
5{, 0, 0, ) ->
0, 0, 0, ) ->
0, )[0] -> "name"
0, )[1] -> :
0, )[2] -> "myagent"
0, )[0] -> "spec"
0, )[1] -> :
See appendix A. In YAML format, different delimiters are used:
kind: Service
metadata:
name: myagent
spec:
ports:
- 8765
- 9376
selector:
agent: example
type: LoadBalancer
YAML emphasizes temporal structure and relative state, Now the structure can only be understood in relative
terms, as parentheses are based on indentation level, i.e. ordinal coordinates within a two dimensional embedding
of the data within a page. Section markers are denoted by ‘- ’ and association is denoted by ‘: ’. These assume a
state machine rather than a stack for parentheses. A JSON or YAML structure is a Directed Acyclic Graph (DAG).
The coordinatization of a JSON structure is by name and container path (a spanning tree).
Example 34 (CSV data formats) Comma Separated Values (CSV) are an array format. The names of data are
represented by their position in a regular spacial array. Names are not provided as table headers, but are assumed
to have been agreed by the implicit schema. This is a tuple coordinatization.
3.10
Streams, state, and memory
In any system that observes input through some sensory stream, assessments are converted into state, and memory
is involved in storing that state. The timescales of real world events and occurrences span all possible scales.
Events might be ephemeral, relative to an observer, in which case they could only contribute to the observer’s
context, or they might persist for repeated sampling as statistical ensemble averages.
26
3.10.1
Assessment and timescales
There are three stages to assessing data, which are explored as a simple model for sensory knowledge in the
remainder of these notes:
1. Data sampling, with timescale Tsample .
2. Symbolization/tokenization/digitization, with timescale Ttoken .
3. Contextualization (combination of multiple sources), with timescale Tcontext .
Contextualization uses a short term memory to accumulate state, and keeps it as an short term assessment of the
state of the world. It is the ‘awareness’ of the agent’s situation.
Example 35 (Short term memory, modulation, and timescales) In later sections, we shall see how the short
term memory modulates the long term memory, allowing large amounts of stable information (slow) to be switched
on or off, by small amounts of context information (fast) (see figure 10). The modulation of long term learning by
situation awareness is a form of conditional switching.
Context contributes to defining the boundaries of applicability for memories, by allowing more permanent memories to be switched on and off in order to adapt to situations.
promises
+policy
+data visibility
sensor/parser
serialization
Input stream
e.g. json
context / meta
adaptation
Figure 10: Input stream is combined with a context according to promised policy, which is based on what has previously been
learned. The initial promise to learn and adapt are the way we bootstrap a system without pre-supervised training.
Example 36 (Feeding back learned patterns) The working week is a pattern that humans can learn, to bring
clarity to many processes that we sample and observe. However, this model may not be applicable to a train
journey, or a light switch, or a satellite’s environment. The major patterns of influence, experienced by an agent,
need to be learned over time, and calibrated for each kind of environment individually. Feeding back these patterns
to condition the way assessments are made leads to a form of anomaly detection, or immediate abnormality alert
[30].
Patterns are crucial to the identification and learning of semantics. If a stable pattern is already known, it can be
assumed, and learnt permanently by hardcoding to save the cost of repeated recognition. This is what happens in
the separation between DNA and protein networks, which are stable over very different timescales.
3.10.2
Mapping input state to concurrent context
A sensor is an input channel that is monitored by an agent. Sensors are designed to select specific signals from
what is promised by the exterior environmental source.
Definition 32 (Sensor) A promise by an agent R to accept specific data d through a filter F , from a source S:
S
R
+d
−−→
R
(24)
−−−→ S,
(25)
−F |d
combined with the promise to relate an instantaneous assessment to a subsequent agent A:
−d
α R−
−→S
R −−−−−−−−→ A.
(26)
Assessments form a data stream, which obey the Markov property [36] in the absence of memory, i.e. in the absence
of memory, every sample is assessed independently of what came before.
27
Sensors contribute to the state of the agent A, i.e. to the snapshot of an agent’s promisable characteristics. Memory
may be added to this later to enhance discrimination.
Definition 33 (State) A collection of concurrent assessments over an agent’s acceptance (-) promises.
The ability to assess concurrently assumes memory sufficient to promise the persistence of the assessments over
the interval, and perhaps beyond. The process by which assessments are made is familiar from information theory.
Definition 34 (Sampling) The projection of a sensor’s assessment into the alphabet of states of the receiver system.
As an autonomous agent, a sensor may filter input as it likes.
Lemma 10 (Sensor resolution) The information resolution of the sensor is less than or equal to the resolution of
the sample.
This is assumed obvious, and is stated without proof.
Definition 35 (Parsing) The repeated sampling of a sequential stream to decode its concurrent symbols.
Closely related to state is the notion of context:
Definition 36 (Context (version 1)) A cumulative assessment of the state of promises, given and received by an
agent, over a concurrent interval of time.
An equivalent restatement of this definition is used in section 5.3.2.
Context is a short term model of a system; it is a self-assessment of an agent’s place within its surroundings.
Because it needs to change quickly, in step with its environment, its complexity has to be limited relative to what
could potentially be stored in a long term knowledge model. The status of context is thus a bit like ‘hearsay’, as
contrasted with confirmed evidential knowledge.
Definition 37 (Situation awareness) The assessment of context, from a set of interior and exterior sensory inputs,
relative to an agent, in comparison to an on-going longer term path of intent.
To maintain context in an operational system, there has to be a process within an agent that can work fast enough to
make assessments. This defines a timescale, which we call Tsample . The process of digital sampling (symbolization)
and contextualization must be of a similar order of magnitude in their timescales, and could well be closely related.
We may assume that these timescales are of the same order of magnitude, since sampling and symbolization are
interdependent in any representation of assessment14 .
• The process of parsing structural information has state itself (Chomsky hierarchy). This can be referenced
in the interpretation of the data.
• The particular path taken in reading sensory input, i.e. a story or temporal structure can form a basis for an
evolving context.
• The state may be persistent or ephemeral.
3.10.3
Transmuting temporal association into spatial representation
A system’s spacetime state is its assessment of the present. It provides a notion of ‘here and now’, which, in
turn, provides a primary original foundation for associative linkage. Each frame of ‘here and now’ is like a
film reel that overlays new experiences and previously experienced concepts. It can be characterized only by
the combination of exterior sampled states with what has previously been learned and replayed from interior
knowledge representations. Interior representations can, moreover, only be based on interior spatial degrees of
freedom, i.e. what memory agents can promise. Labelled states could be assigned semantics directly, by linking
context tokens immediately into concepts, and they may take on semantics emergently over many iterations. Both
processes must be possible for non-trivial knowledge representation.
Example 37 (From sensor input to concept) Suppose we have sensors that sample and assess the exterior environment in an aircraft. How could we reach assessments that have the semantics of high level concepts, e.g.
14 Even if we imagine that the resolution of a sensor is real valued, with infinite resolution, it must be represented by internal states that can
be read in a finite amount of time, and hence must be finite. Thus all observations have finite resolution.
28
Agent
Plane
Wheel
Context State
flying, crashing, standing still in a hangar.
revolving, stationary, braking.
from the sensory measurements of speed, temperature, acceleration, etc.
A stream of assessments, by an observer agent, leads to a spacetime structure in which assessments become
connected. Shared semantics in a cluster could be used to define concepts from these.
Definition 38 (Temporally associated concepts (simultaneous activation)) Concepts may be called temporally
associated if they occur simultaneously within a concurrent interval, such as a context interval, or within the body
of a single promise.
The perceived simultaneity of context is where the association of long term concepts must begin (see figure 11).
This is a strategy used not only in cognition, but also in sensory systems like the immune response where lymphocytes require co-stimulation from antigen presenting cells to provide contextual confirmation of pathogen recognition. Initially, during the bootstrapping of a knowledge system, there is no longer term memory of concepts or
their associations to rely on for self-calibration of context, but sensory context is continuously available.
transitive
t1
simultaneous
context
t2
simultaneous
context
t3
(( ,
) ( , )) => (
, )
Figure 11: As concurrent units of data are received by an observer, associations happen in each interval by the co-stimulation
or co-activation of conceptual agents. This becomes transitive, by mutual association with intermediary. at a timescale greater
than the simultaneous/coincident signal agents.
key1
value1
additional interpretation by clustering
key2
value2
sensor group
context
Figure 12: An observer can annotate what is observed, in context, by adding associations between tokens (names and values)
within a context. Depending on the memory of the observer, the annotation may include short relative times, or be limited
to simultaneity: e.g. key 1 preceded key 2, or when key1 was used value 2 was also in play, where key1 and key2 represent
different but related aspects of a common sensor interpretation, in (partial) pattern.
Context is cumulative, over successive concurrent intervals, hence it is a spatiotemporal state. It can thus
represent regions and paths. A context may represent spacetime locale. It might be what just happened, or where
we are currently located.
29
Example 38 (Parsing a sentence (linguistics as cognition)) The parsing of a sentence involves short term state
memory (identified by the classifications of Chomsky theorem [34], and stochastic generalizations). An agent
parsing a stream must be able to hold all the input tokens that belong within a concurrent interval, and are
assigned semantics collectively, in order to represent the grammar faithfully. This suggests that the transmutation,
from individually sensed input words to semantics, divides into three stages: sensor recognition (words), short
term context (clauses, sentences, etc), and long term memory (meaning):
feedback
←−−−−
feed forward
input
−
−
→ Associations .
−
−
→
State
→
Context
Word
|
{z
} −−−−→ Concepts
| {z }
|
{z
}
sensor
awareness
(27)
meaning
As words accumulate in context, the scaled ‘simultaneity’ of the complete sentence fragment triggers associations
and concepts in long term memory. The long term memory may conceivably feed back into short term context
to further discriminate based on learning. From this perspective, language representations are indistinguishable
from sensor arrays, in which words represent different classes of virtual sensor stimuli. Next consider the sentence:
“I’m walking the dog.”. Taken at units, each word has meaning, and the sentence also has meaning. Additionally,
in the city of Newcastle Upon Tyne, in the UK, the phrase ‘walking the dog’ is a code ‘euphemism’ for ‘having
a beer at the pub’; thus, the sentence fragment has meaning at a larger scale. Meaning may be perceived at any
scale, and patterns may trigger quite different associations in different contexts.
Example 39 (Parsing a document) The parsing of a document is a scaled version of the parsing of a sentence.
In order to maintain longer term context, there needs to be continuity through short term memory. Thus, if an
agent puts down a book and picks it up later, it needs to reconstruct some context from long term memory before
continuing. The process of parsing a linear document, or other sequence of events, leads to an evolving historical
timeline which is cached as context. Context must expire and be replaced by a new context at approximately the
same timescale as sensing takes place. Sometimes new context might replace, other times it might accumulate. The
running context is built on spatial structure (titles and section containers), as well as associations and concepts
triggered from memory.
The low level semantic units are parsable (i.e. observable) senses, based on words or sounds, depending
on what sensory input channel is active. Then clusters of words attain independent meaning. Thus, spacetime
considerations alone suggest that, externally, linguistic words and phrases are just sensory fragments (as imagined in generative linguistics [37]), while internally they are transduced into concepts (as imagined in cognitive
linguistics [20]).
The examples above illustrate how context can be both a container for perceived short term correlation or a
dependency for switching on and off long-term promises. Thus context is about scaling of agent roles, by recursion
and by association.
3.11
Data classification and summarization
Looping back memory along side a fresh input stream, iteratively, in order to calibrate samples against expectations,
is how a knowledge equilibrium can be formed from assessments. Repeated measurement is a basic element of
experimental scientific method [3], and it is easy to see how this leads to a subjective notion of certainty in the
observer. Convergence to an attractor is one way to achieve stability; pre-existing memory acts as a seed for such
an attractor.
The reliance on repetition is a robust strategy, but it is also data intensive. It would be impractical and inefficient
to remember every individual assessment (i.e. every data point acquired) forever, so one needs to be able to
summarize data into representative characterizations (tokens) that are cheaper to store in memory15 . Several forms
of summarization of data may be used to form the basis of knowledge representations over different timescales:
• Context awareness (short term passive summarization): Tsample , Tconcurrent .
• Learning (long term reinforced or stabilized passive summarization): Tknow ' Tlearn , Ttoken , Tpattern .
• Adaptation (short or long term feedback based on active summarization): Tadaptation , Tattractor .
The principal distinctions in this list are about timescales; indeed, this is the spacetime lesson that allows stability
to converge. The cyclic repetition of the learning process, over a noisy inputs, supplies a constant stream of such
perturbations, so persistence is a form of noise reduction that allows stability of semantics. Stable semantics reduce
the cost of reevaluating meaning at every turn.
15 There is some empirical evidence for single neurons in the brain being able to represent complex tokenized concepts, in so-called ‘invariant
representations’ [38].
30
3.11.1
Stable summarizations: attractors
In order for random inputs to converge into clusters, at a stable location, there must be a mapping whose properties
are both flexible enough to adapt to unknown inputs and robust enough to recognize the same patterns repeatedly.
Small variations in a pattern would preferably map to locations that were associated with attractors, else every
variation of input would lead to an unstable and random new conceptualization. The implication is that even
associative memory would benefit from a metric quality (see section 2.4.1).
Iterative learning avoids data explosion by integrating samples into a form that is more stable than raw data.
Aggregate data measures, with a metric nature offer such a representative value for a collection of samples. In
promise theoretic terms, a data aggregation is formally an external promise made by the superagent over all semantically equivalent samples16 . Summarization of data leads to concepts like averages. Dynamical or quantitative
averages are well known from statistics. A semantic average, on the other hand is a qualitative value that represents
the common promises of a cluster of agents.
Definition 39 (Dynamical summary) Any single representative measure that assesses an ensemble of observations: Tknow ' Tlearn .
Definition 40 (Semantic summary) The formation of a semantic average over a bundle of related promises, from
any number of agents, leading to i) a superagent, and ii) and exterior promise for the superagent ensemble:
Tknow ' Tlearn .
Both dynamical and semantic summaries are formed using the available characteristics of data, described in section
3.11. For numerical data, the available characteristics are the modalities and moments of data distributions, while
for symbolic or qualitative interpretations, the averaging takes the form of semantic generalizations and collective
interpretations.
Definition 41 (Metric (dynamical) average) A standard ordinal scalar value that promises to represents a body
of dissimilar ordinal scalar values: Tknow ' Tlearn .
Definition 42 (Semantic average) A single nominal interpretative assessment that promises to represents a body
of dissimilar promises or assessments: Tknow ' Tlearn .
Example 40 (Averages) Averages are aggregations yielding a representative result. Dynamic or metric averaging
takes a number of quantitative scalar values and combines them to result in a value that promises a scalar result
to represent the bulk data: Tknow ' Tlearn .
summary
Figure 13: Averages and representative values are exterior promises belonging to a superagent. Averaging, or summarization,
is thus a form of spacetime coarse graining, i.e. a scale transformation.
Example 41 (Semantic averages as scaling generalizations) Semantic averages take a number of dissimilar promises
and represent them by a common standard, e.g. extracting what is meant by ‘a vehicle’ from ‘a car’, ‘a bus’, ‘a
train’, etc.
A trivial lemma may be noted, relating to these:
Lemma 11 (Averages express exterior promises for an ensemble superagent) Any collection of agents making
promises or assessments may collectively promise a scaled average to represent the samples; such a promise is
exterior to the superagent formed by the ensemble (see figure 13).
16 So-called
‘big data’ (unstructured data) do not constitute knowledge, but patterns revealed within them might.
31
The proof is simple (see paper II): let each assessment by observer agent A be represented by a signal agent ai .
Let each agent ai promise to be members of an ensemble (sample set)so that they form a superagent. Now, at
super-signal-agent scale there must be new external promises. An average is a external promise of the ensemble
superagent.
3.11.2
Repeated sampling, significance of data and attached meaning
We attribute meaning to distinguishable discrete measures, i.e. symbols, that stand out against background noise.
How we transform signal information into meaning is thus related to a signal’s entropy.
Example 42 (Significance and entropy) Culturally significant markers, signposts, and directional beacons, are
towers, statues, monuments, obelisks in a landscape that is relatively flat. Low entropy, low information structures
stand out symbolically, and we attribute them with special semantics.
Hypothesis 1 (Significance is inverse information) At some information scale σ, we can identify a significance,
of the general form:
Significance(σ) ∝ MaxEntropy − Entropy of representation .
(28)
Significance gives an anchor point to which we attach meaning. For graphs, network centrality can be used as the
source of the entropy distribution [39].
The more times data are collected in semantically equivalent contexts the more robust an average summarization of the data can be to new variation. By repeating a measurement, we can make one interpretation stand out
against all the others. This brings not only increasing certainty, but increasing significance. This is the idea behind
histograms. Regardless of the data classification alphabet used for the sampling procedure, repeated sampling will
lead to a histogram whose scales of variation, measured subject to Nyquist’s sampling limit, that represent the true
alphabet of variation in data.
Assumption 1 (Knowledge is an equilibrium state) Learning happens by iteration, and knowledge is the summary of iterative learning, over a timescale Tlearn . Without continually revisiting and re-testing observations, both
the certainty and significance of the data lose meaning. Stable knowledge is therefore the result of an equilibrium
cycle, fed by observation, pattern discrimination, and repeated resampling.
The remainder of this work rests on this assumption.
3.12
Distributions, patterns, and their information content
From the summarization scaling argument, we conclude that knowledge begins effectively with stable histograms.
Memory of multiple episodes builds up a summary over data episodes, from which we derive information, and
thence contextualized knowledge. Summary representation takes the form of a frequency distribution, relative to
a set of sampling classes, which, in turn, may be tokenized into an alphabet of effective recognizable tokens or
symbols. These may be found within the spatial characteristics of the frequency distributions:
• Modality (how many peaks does the distribution have?)
• Moments of the distribution (is it localized, symmetrical, skew, long tailed, etc).
• Smooth, discrete (bar chart histogram or smooth bell curve, etc).
• Wavy, frequency amplitude (what is its Fourier decomposition?)
• Continuous, broken, discontinuous (is it a differentiable total function, or a partial function?).
To transmute statistical weights into symbolic representations, one may re-digitize these different characterizations
into manageable, discrete characteristics:
Definition 43 (Tokenization) The process of mapping a set of assessed patterns and characteristics P into names
N:
C:P →N
(29)
over a timescale Ttoken .
In so doing, we must be clear that the mapping of a symbolic context C, based on observed data, is a function of the
promisable capabilities of the agent’s sensory apparatus. Thus, there is no way to separate concept or interpretation
from cognition. The ability to identify patterns is critical to forming certain kinds of associations (see figure 30
(b),(c),(d)).
32
3.12.1
Ontological structure
In a world model, there may be both observed phenomena and imagined phenomena. We have names of things,
and relationships between them. Both of these can be imagined or real. Thus a knowledge model can grow to be
richer than the sensory model, as it feeds back on itself.
THE MIND
Long−term
Memory
Short−term
Memory
CONCEPTS
classes/tags
BASIS/index
Exemplars
Things
situational metadata
numbers
sensory horizon
values
sensory horizon
SAMPLES
Docs
THINGS
MEASUREMENTS
OCCURRENCES
THE WORLD
Figure 14: Occurrences/exemplars live in their own world, with their own adjacencies. Conceptual models form aggregate
representations that are smaller than the sets of exemplars, and act as indices (coordinate bases) for them, with aggregate
meaning.
To learn an ontology, rather than assert one [14], we can assume that samples promise consistent semantics up
front [40]. It is then possible to learn structural relationships, i.e. clustering by divining a definition of distance,
using some arbitrary dimensional spanning set, e.g. using clustering methods like Principal Component Analysis
(PCA) [17], assuming that the spanning set is known. There is a bootstrap issue, by which observational data have
to stabilize into ‘knowledge’.
Hypothesis 2 (Stability hypothesis) Knowledge consists only of those phenomena that are dynamically stable,
hence from which we may assign consistent semantics.
Before the widespread use of dynamical pattern recognition, it was a long standing assumption that a model
of knowledge had to begin with an ontology, or a pre-imposed arrangement of ‘things’, subjects, issues, and
relationships into a tree. That approach failed to explain how understandings of relationships could come about
by learning, or be adapted to a changing world. Emergent models of an environment, on the other hand, with
identifiably separable clusters could be the basis of a time-varying emergent ontology. The concept of strongly
connected clusters (SCC) from graph theory [41] may serve as an emergent definition of ontological structure;
one could imagine using different semantic filters to select directed subgraphs to form overlapping ontological
islands17 .
The prediction of promise theory would be that concepts form from as a result of two processes, based on
observer independence:
• Intrinsic properties posited by a source (+ promises).
• Perceived and learned properties perceived by a receiver (- promises).
This basic promise theoretic theme follows from the ± propagation law, and runs throughout these notes, and
recurs on in many scenarios to explain subtleties of information propagation.
3.12.2
Pattern acquisition: generalization and discrimination
Concepts and their associations emerge through encodings that cluster related concurrent things. This can be
immediate, like a Markov process:
17 Indeed,
this observation was the beginning of successful web search with PageRank [42].
33
• Classification of concepts by senses alone.
• Simultaneous activation of concepts across parallel inputs.
and it can be enhanced by long term memory, in a Non-Markov process:
• Approximately simultaneous activation (coarse persistence in concurrent chunks).
• Feedback from previous learning into sensor stream.
We search for the most significant patterns in order to attribute meaning to them, by:
• Identification by aggregation: The precise accumulation of attributes by pattern activation. If every pattern
can only be activated once without context dependency, this leads to a limited number of traits, which are
stable by aggregation.
• Identification by division: A precise branching of attributes, leads to the discrimination of patterns into
modules, which are fragile by branching. Within any branch, any sub-branchings can bring further discrimination. Branching is not an idempotent operation.
The formation of a branching tree of concepts is called taxonomy or ontology span. Since decomposition is
arbitrary (we can branch or group arbitrarily and as many times as we like), the implication is that there cannot be
a single taxonomy or ontology based on a universal set of mutually exclusive properties.
3.13
Summary
A few main themes are behind the construction of knowledge from data:
1. The separation of scales, according to a principle of weak coupling. Each scale has its own semantics and
cohesion, structurally connected, but independently identifiable, with independent agency.
2. The identification of types, schemas, from the separation of spatial scales, i.e. the formation of regularized
superagents, and the identification of context from the separation of time into concurrent chunks (simultaneous and sequential).
3. The breaking of ‘translational’ symmetries is the key to encoding data reliably using anchors and signposts
to align semantics relative to distinguishable markers (fixed stars to steer by).
4. Four mechanisms occur, by which intent is passed along from one agent to another:
• Nested containment (scaling).
• Causal history (dependency).
• Composition, or cooperative refinement.
• Similarity (proximity).
Data are encoded in the form of patterns, which are subjectively identifiable as discrete symbols. The patterns are
transmitted, or written, in any semantic medium, by variations in the scalar promises of spatial agents over spatial
trajectories. Symbols may be combined into larger patterns, e.g. words and sequences, by tracing out spacetime
trajectories in the representation medium. Data are parsed in scaled, sequential streams of quasi-concurrent symbols. The order of the stream has meaning, but may not be deterministically related to the interpretation. To alter
and extend the interpretation of sequential orderings, one may define language semantics for the encoded patterns,
that are partially ordered at a larger scale (superagency). The representation of data as scalar promises is a clue
about how we may form memory out of any semantic space.
4
Memory, addressing, and spacetime patterns
A central part of what we consider knowledge and ‘smartness’ to be about is the ability to learn and recall concepts
and patterns. This requires memory, which, in turn, is naturally built on spacetime concepts. By choosing a
language of semantic spaces for memory, we acknowledge that almost any agent medium may be used to store
information: whether it be atoms, molecules, transistors, brains, or cities.
This section concerns how semantic spacetime can be used as a representation of addressable memory, one of
the key requirements for learning and knowledge representation.
34
Definition 44 (Memory agent) An agent M may be said to promise the role of memory (like a service) if it
promises to forward data d accepted from input source S, previously, to another agent R, upon receipt of an
appropriate request r:
S
M
R
M
R
d
−d
−−→
r
+d|r
−−−→
−d
−−→
M
(30)
S
(31)
M
(32)
R
(33)
M
(34)
(see figure 15). The replacement of impositions with promises, in these interactions, is also acceptable for voluntary
R
M
R
−r
S
M
R
S
Figure 15: Memory acts as a switch, with the promises here likened to a transistor. Data are pushed to memory, and accepted
on trust. Requests for data are pushed to memory and a conditional response is returned, as a service transaction.
cooperation lookup, such as in scheduled access. In that case, the request r → ∅ may be empty, for spontaneous
delivery, like a repeating beacon.
A memory agent is thus a kind of delayed forwarding device for information. This is explored in more detail below.
4.1
Relationship between switching and memory
In order to write to memory, by changing the patterns represented by a spacetime configuration, agents must be
capable of keeping conditional promises, which are the representation of switching interactions [6].
4.1.1
Conditional promises
Conditional promises are the basic atomic representation for a variable memory representation, because properties
are promised subject to conditional biases, e.g. an agent may promise to record sensor data subject to prevailing
conditions, or may promise to return a data value, subject to providing the agent address.
Definition 45 (Switch) An agent, which promises a distinct outcome, for each distinct assessment it makes of a
promised input.
output(condition)|condition
switch −−−−−−−−−−−−−−→ A
(35)
In other words, the output is a function of the condition. The condition modulates the promised role.
Lemma 12 (Dependencies are switches) Any functional dependency represents the modulation of a promise, and
hence acts as a memory element.
The proof follows from the structure of a dependency (see figure 15):
+b|dependency
S −−−−−−−−→ R,
(36)
where b is effectively a trivial function of the dependency, as the promise body is +b if the dependency is given,
and ∅ is not. Switches may be more complex than a single input, and may give up their stored data by different
channels (see figure 16). More complex switching structures can be used for forwarding or routing of information
that can perform sorting functionality, which is important for addressing specific elements.
35
S1
R1
S2
M
R2
S3
S4
R3
Figure 16: The flow of data through discriminator switches: with this generic form, we have the basis for predictable addressing. The ability to uniquely address locations in space, and to promising routing of data to that single destination is what allows
spacetime to encode data reliably, and for us humans to attach significance to the values.
Definition 46 (Routing (multi)switch slave) An agent with several imposed input channels Si and several output
channels Rj (see figure 16). For example, a simple configuration might impose:
Si
Mij
Rj
Mij
di
−di
−−→
rj
Mij
(37)
Si
(38)
Mij
(39)
+d|rj
−−−→ Rj
(40)
The switching agent M is now a matrix whose promise to forward depends on the request issued by a receiver.
Again, the imposed request may be null to represent a pull request for data by the receiver.
By extending the conditionals to a matrix, a policy for forwarding can be hardwired into the intermediate agent, as
an explicitly programmable relay.
Definition 47 (Forwarding or relay agent) An agent F which promises to accept data from an input channel Si
and relay it to an output channel Rj unaltered, subject to a promised condition rij :
Si
F
F
A?
di
F
(41)
Si
(42)
−−−−−→ Rj
(43)
−di
−−→
+di |rij
+rij
−−−→
F
(44)
The condition rij is called the routing or forwarding table and may be promised by any agent A? , including F
itself.
If we allow the intermediate agent to promise with full generality, the implication of this matrix of promises is that
the router acts as memory in the form of a routing table.
Lemma 13 (A routing switch can act as memory agents) Mij is equivalent to a collection of memory agents
with dim j retrievable values.
The proof is trivial from definition 44 by identifying M → Mij , S → Si and R → Rj , with data d → dj for
requests r → rj . Similarly:
Lemma 14 (A routing switch can act as a data forwarder) The multi-switch Mij contains a forwarding agent
F.
The proof follows by identification Mij → F , Rj → A? .
Example 43 (Network routers and forwarders) It might seem that a routing agent does not store data for very
long, but this is only a design choice of most routing switches. The definition above states that data values can be
retrieved from memory at some later time, which is a sufficient condition to act as memory. It does not specify how
many times values can be retrieved, or even how quickly. Routing switches could, in principle, keep all the data
they receive permanently, as long as they have sufficient internal capacity, and a sufficient number of addressable
channels, but that is not usually of interest. Some firewalls retain more detailed information for forensic analysis.
36
The configurations and examples above illustrate how memory retrieval by address is directly analogous to the
routing of information by request, and that the requests may be pre-determined in the form of a memory policy.
We can apply these structures to the problem of learning systems, and knowledge representation, by thinking of
memory addresses or data requests as being conditional biases which come via a sensor or input channel, arriving
as a queue of changing impulses. What this shows is that almost any spatial structure capable of being modulated
through its adjacencies to neighbours can store and retrieve state under the right circumstances.
Example 44 (Artificial Neural Networks) Artificial Neural Networks (ANN) are sequential fabrics made of nonlinear aggregating switches, whose agent connections promise weighted threshold values to other agents. By
training, i.e. adjusting, the weights of the whole layered fabric to match a desired state output, for given input,
they approximate something like N : 1 attractors. ANNs are effective on complex patterns, e.g. facial recognition.
They are used for sensor pattern recognition, in the language of this work.
4.1.2
Timescales for deterministic switching
To be able to construct stable knowledge systems from switching agents, their promises need to be kept over
timescales that support predictability even with non-deterministic inputs. The timescales for change and operation
of a conditional promise tell us about the dynamic and semantic stability of switching. We may define the stability
of this promise. Let us denote the timescale for change of promise components by T (·).
Definition 48 (Semantic stability of a switch or conditional promise) Consider a promise
+b|c
S −−−→ R
(45)
with timescale T (∆b) meaning the timescale over which one expects a change in the promise body, and T (∆c)
meaning the timescale over which one expects a change in the dependency modulation. The stability of the promise
means
T (∆b) T (∆c),
(46)
i.e. the body of the promise b is constant over the switching on and off by modulation. If b, c are continuous
functions, then we may express this as
∂c
∂b
.
∂t
∂t
(47)
If these stability criteria are not met, subsequent switching operations will lead to inconsistent results, i.e. the
promised behaviour will not be a total function of c.
Definition 49 (Dynamic stability of a switch or conditional promise) Consider a promise
+b|c
S −−−→ R
(48)
with timescale T (b) meaning the timescale it takes to keep a promise b from any starting state, and T (∆c) means
the timescale over which one expects a change in the dependency modulation.
T (b|c) T (∆c),
(49)
i.e. the promise b must be keepable quicker than the conditionals vary by modulation.
Lemma 15 (Non-deterministic switching) If the dynamical stability condition in (49) is not met, there must be
either a queueing instability or an overlapping and inconsistent state in the switch.
The proof is straightforward. Let the arrival rate be λ ' 1/T (c) and the service rate be µ ' 1/T (b|c). If
T (b|c) T (∆c) then λ µ, and there is no queue, else there must be a finite queue length. If the agent S does
not support queueing, then its state must be a superposition of all the queue states, i.e. c must be in an indeterminate
state.
Example 45 (Memory and learning network stability) Memory networks, such as computer memory, Bayesian
or neural networks, etc., are dynamically and semantically stable if their training b does not change significantly
for the duration of their operational usage c.
37
4.2
Spatial representation of encoded memory
Data are encoded and represented as strings of symbols. Symbols are represented as discrete (but possibly overlapping) patterns in a spatial medium, i.e. they are patterns in the properties of spacetime. Symbols are thus promise
theoretic agents (or superagents).
Definition 50 (Alphabet) A collection of unique symbols, represented as discrete spacetime patterns over any
configurable spacetime property.
Alphabetic symbols form serial structures called strings. Some strings may promise to conform to languages [32]
(see also the discussion in paper II).
Definition 51 (Language) Aggregate superagents of symbols that promise constraints L on their interior composition, classified by Chomsky rules and their stochastic generalizations, etc.
Natural languages comprise strings of words and phrases (such as in this document or in spoken words). Formal
languages, used in control systems, typesetting, protocols, etc, make more rigid promises about internal structure
of language objects. For the purpose of this work, it is not necessary to go into more detail than this.
Data may be transmuted through many different alphabets during the course of reading, copying, transmitting,
and storing sampled data in different memory representations.
Example 46 (Multiple language representations) Imagine reading aloud from a book, to an illustrator, who depicts a story in pictures, with captions written in another language. Each version of the story represents the story
in a different semantic (spacetime) medium.
Representations may not be fully equivalent.
Example 47 (Binary representation of languages) The alphabet of symbols for contemporary computers comprises the binary symbols 0 and 1. These are combined in fixed size strings (superagents) or tuples to represent
arbitrary pattern. At a higher level, these are grouped into 8 bit words over ASCII symbols, or 16 bit Unicode
symbols. These symbols form strings in different human and computer languages.
Example 48 (Codons) In biology, two alphabets of the genetic language are composed of the four almost identical
sets of symbols for DNA/RNA: G,A,T/U,C. The language of proteins is encoded by codons (triplets of DNA/RNA
symbols). These cluster into words formed of three letters, called codons, which represent amino acids.
The symbols in an alphabet represent a collection of states which can be promised by spacetime agents. Signal
agents, promising messages in these languages, can be emitted and absorbed in streams, as promise bodies, and
act as symbols themselves in order to communicate between location agents [2]. These few points about language
should be enough to describe serial communication in a general way.
4.3
Spatial addressability of memory
Storing information as spatial structures is, by far, the dominant form of memory representation at every scale. The
interior nature of available spacetime agents is central in determining the sophistication of the patterns that can be
supported by a memory representation. Simple agents, technological or biological, promise limited interior capabilities but can make up for information density through numbers. On the hand, a medium like the biomolecular
language of proteins has immeasurable complexity.
A minimum requirement for agents is that they must promise labels to be individually addressable, in such a
way that information stored may be retrieved from the either the same location, or one that is deterministically
equivalent. For this to be true, there must be cooperation between the agents, and each must promise a unique
identity, as either a name or a number, within their localized namespace (see paper II).
Example 49 (Pages of a book) A book is a collection of pages. Each page agent is numbered allowing addressability of its contents.
Two approaches may be used for addressability.
38
4.3.1
Address discrimination by constrained flooding or broadcast propagation (-)
All agents promise to recognize their own address or unique name. A sender imposes (floods) addressed information to all the agents, and every agent promises to only accept information labelled with its own address.
Let Mi be a set of memory agents, and A be any agent that can read of write to the memory. Each agent Mi
promises a value vi , and a unique address addri , sometimes referred to as its media access control address, to all
other agents:
+(vi ,addri )
Mi −−−−−−−→ A, Mj
∀i, j.
(50)
In general, address uniqueness cannot be promised without coordination. Such coordination could be expensive,
requiring 4N 2 promises amongst N agents, or 2N promises with a centralized source authority. In practice,
memory addresses are imposed on memory agents, by a single source authority, which we may call the factory F :
F
Mi
addri
−addri
−−−−→
Mi
(51)
F.
(52)
To write a value to a particular Mp , and agent A imposes a data value v, along with the intended address onto all
agents (this is called broadcasting or flooding=:
A
+(v,addrp )
{Mi }
(53)
All agents promise to accept such messages, conditionally if addri = addrp :
−v|addri =addrp
Mi −−−−−−−−−−→ A, Mj ,
∀i, j.
(54)
These promises are hardwired into the agents in many cases, however, unreliable agents may or may not promise
not to accept messages not addressed to them.
Example 50 (Unreliable agent addressing) If we chose the Mi at the scale of representing students in a classroom, then we could not assume that they would all promise to behave in such a simple mechanistic manner.
Several students could have the same name, as they originate from different ‘factories’, and so on. In other words,
depending in the scale and nature of the agents, we should be careful about what assumptions are made.
As the number of addressable agents grows, a flooding scheme imposes an increasing burden on the capabilities
of the receivers, to i) recognize longer and longer addresses, and ii) to respond quickly and unblock the channel
for future messages.
Example 51 The biological immune system uses a flooding mechanism to signal cells. It is impractical for cells
to have unique addresses, but cells of a given type are considered redundant, and have the same function. Bioaddressability targets classes of cells rather than individual cells. Cells have antigen receptors (Major Histocompatability Complices (MHC), etc.), which act as addresses for classes of equivalent cells. These sites bind only to
particular messenger cells like antibodies, and thus receive only the signals intended for them.
4.3.2
Address discrimination by routed propagation (+)
A way around the scaling limitations of flooding is to hardwire spacetime adjacencies so that there is a unique
and direct path to each agent from a source (see figure 17). An intermediate agent R, which we call a router, is
introduced between the sender of a message A and a finite number of receivers Mi , say in the range 0 < i < N .
Each leaf agent now has a unique identity to the intermediate routing agent, by virtue of its unique adjacency
relationship, and being isolated in its own cul de sac of the hub R. The router R accepts data from an agent A, if
the address lies in its range:
A
R
+(v,addrp )
R
(55)
−−−−−−−−−−−−→ A
(56)
−(v,addrp )|0<p<N
Further, it promises (or imposes) to forward the message to the Mi , without broadcasting it to all its children:
+(v,addrp )|p=i
R −−−−−−−−−−→ Mi
39
(57)
+1
−accept[1−3]
+2
router
+3
Figure 17: Individual identity promise, and voluntary identity recognition is a part of both flooding and non-flooding addressing schemes, if one considers the scaling of routed addresses. A router accepts a certain range of addresses and routes precisely.
If the paths (or ports in the language of [43]) to 1,2,3 were not unique, the end nodes would also have to promise to know their
identities and discriminate. Routing allows end nodes to operate with few promises. The principle can be applied recursively
to address arbitrarily large numbers of agents, at the cost of an increasing number of trusted intermediaries.
and Mi accepts:
−(v,addrp )
Mi −−−−−−−→ R.
(58)
It is no longer strictly necessary for Mi to make acceptance conditional on the address, if one assumes that the
configuration of agents is fixed and that R is a trusted intermediary and delivers the data correctly. Promise theory
makes it clear that one cannot place complete trust in intermediaries: they may be unable or unwilling to keep their
promises to relay information.
An alternative (which is more in keeping with the agent autonomy) is that the relationship between R and Mi
is reversed. Mi can promise its identity or leaf address to R, to be placed in a routing table (and index or directory)
for forwarding:
Mi
R
+addr
i
−−−−→
−addri
−−−−→
R.
(59)
Mi .
(60)
The difficulty with this is that the Mi have no way to make their addresses unique, without exterior coordination18 .
This conundrum is a namespace management issue, which I won’t explore further in these notes.
In summary, the intermediary R has a topology that gives it a unique path to each memory address. The router
promises to keep an index overview, or directory map, enabling it to switch on connected routes (see paper I).
It selects the path corresponding to the promised address, and forwards information accordingly, i.e. it acts as a
switch that discriminates direction, based on the promise of an address.
To scale this further, addresses may be made hierarchical, be splitting them into a routing prefix address and
a leaf address. Then router would modify its promise to accept data only if the routing prefix were addressed to
it specifically. This principle can be applied hierarchically to provide tuple addressing with direct memory access
along dedicated connections.
Example 52 Computer networks use both kinds of address discrimination. Bus architectures, like Ethernet or
WiFi, assign a factory unique machine (MAC) address to each network agent. Messages containing serially encoded addresses, are flooded to all agents in a region and receivers promise to accept messages addressed to them.
No ordering of addresses is needed: the agents are as a gas. For routed architectures, like the Internet protocol
(IP), addresses are routed by selecting the known pathway to the destination. This presumes a cooperative ordered
arrangement of IP addresses, in a solid state.
Example 53 Computer memory chips assign addresses to memory locations by using a discriminatory switch,
attached to an address bus. The address bus acts as the envelope for messages on the data bus, imposing addresses
on to the leaves. Collaborating memory chips are arranged in an order with hardwired signals to tell them their
address ranges. A hierarchical system of relative addressing, with bit discrimination, allows memory chips to
accept messages only for the address ranges they represent, and single out the addressable memory registers
uniquely.
18 In Internet addressing, which uses this scheme, the uniqueness of addresses must be handled manually, out of band, with an additional
network of promises.
40
4.3.3
Criteria for addressability of data in memory
By either of the foregoing methods, absolute addressability of memory implies the ability to uniquely signal or
retrieve information from every agent. Some basic promises enable this.
Names or addresses19 must be unique for all agents within a hierarchy of namespaces that connect senders
and receivers of data, i.e. every memory agent must promise an identity that is uniquely distinguishable by its
neighbours within its scope. Every agent must be distinguishable from every other locally, else it would be impossible to intentionally retrieve data stored in a memory agent, or for a memory agent to forward to another agent
deterministically.
Lemma 16 (Deterministic addresses) Let Mi be a collection of memory agents. Mi may be deterministically
addressed by any other agent Mj6=i , iff Mi promises a unique distinguishable identity.
The proof may be found by considering the properties of memory agents. If Mi and Mj cannot be distinguished,
then
Sk
Sk
dk
dk
Mi
(61)
Mj
(62)
cannot be distinguished, and data may be stored in unknowable locations, and may overwrite existing values.
Moreover,
Mi
Mj
+d|r
−−−→
+d|r
−−−→
R
(63)
R
(64)
cannot be distinguished, thus an agent would retrieve an arbitrary value from an unknowable location. If the agents
are distinguishable, the agents act independently and deterministically according to definition 44.
The ability to coordinatize a space, to assign unique tuple addresses, depends on its topology, i.e. on agent
adjacency. Non-trivial topologies may have loops and multi-valued coordinate, causing significant difficulties. This
is why most artificial structures are based on trees (directed acyclic graphs) or partially ordered lattice structures.
The acceptance of address labelled data cannot be avoided, even in a direct tree routing approach, however. Agents
need to recognize their own messages (see figure 17).
Further to this, we may say that in networks, where there is not direct adjacency with memory locations, the
existence of a complete name or address has to propagate from every memory location to all remote agents in order
to become addressable in practice, i.e. readable or writable20 .
Lemma 17 (Name/address propagation) An agent R may be addressed or signalled directly by an agent S iff
the full name of R propagates through all intermediary agents I1 . . . In to S, and I1 . . . In commit the route by
which the information propagated to memory (see figure 18).
level n
level 1
level 0
1
S
3
In
I1
5
start
send to R
2
4
R
end
R is on interface/port 3
who is R?
Figure 18: Agent names or addresses have to propagate out from their location R throughout a semantic space, before other
agents S can i) know that they exist, and ii) know how to reach them as a destination. Messages could be given a maximum
range, relative to a current position, iff the geometry of the space is fully understood by all agents.
19 Whether
we call this a name or an address is not really important. An address is just a long name, usually given relative to a namespace.
that the ability to route to R should not be confused with the ability to locate named data objects stored in the memory network
without pointing to their specific location, as in [15]. The latter is an independent problem. Only the forwarders that connect the address space
need full unique addresses.
20 Note
41
The implications of this lemma are that deep hierarchical naming or addressing of agents requires a routing infrastructure based itself on memory agents, thus scaling of memory may require recursively nested memory. To prove
this, we may assume that R promises its own identity:
+R
R ≡ AR −−→ .
(65)
In order for any other agent Ai to learn this identity, we must have
−R
Ai −−→ R.
(66)
Knowledge of R0 s identity can propagate iff
Ai
Ai
−R
−−→ Aj
+R
−−→
Aj6=R , ∀i, j 6= R.
(67)
(68)
Since S ∈ Ai6=R , S can learn of R0 s identity, and may try to promise or signal to R. The name of R may not
be private. All agents must be in scope of its name promise. Now, in order for S to signal R, there must be a
connected path along intermediaries I1 . . . In , all of which promise to forward data to an output interface along the
path. This information can only come from back-propagation from R to S, I1 . . . In , and these agents remember
R and the route ‘rij ’ to select the appropriate interface direction. The intermediaries must therefore have memory
of their own.
The costs of routing or flooding a network are quite different. In a pure flooding approach, with no optimized
routing, all agents are blocked by a single agent’s use of a shared network, locked into a single concurrent timeslice.
In a routed network, there is a setup cost of propagating addresses to the forwarding tables, and then a latency
proportional to rN for path length N .
Lemma 18 (Forwarding, name/address scope, and embedded memory) If intermediate agents I1 . . . In forge
a path from S to R, in a hierarchical spanning tree of non-overlapping nested regions (‘subnets’) with N tiers,
then N -tuple addresses may be used to bound (but not eliminate) the memory required for forwarding messages,
at each level of agency:
• R needs only R
• In needs only R, Im<n
• S needs the full path.
We can see this as follows. In order to route at each stage of the path, the intermediaries must have a routing table
‘rij ’. Without embedded memory, this cannot be achieved. In order to discriminate its own level, as well as pass
on data to all sub-levels, unidirectionally, this information is used. Other parts of a tuple address can be discarded.
Example 54 (IP addressing) The IP address subnet mask divides an encoded IP address into a prefix and a local
address part: (prefix,local). This is a two level spanning hierarchy, where local addresses are used at the end node,
and prefix is used for forwarding. With only two levels, the approach is simple but not very efficient, leading to high
memory usage in forwarding agents. If a deeper hierarchy, with longer tuples were used, the memory required to
forward could be reduced to enable local uniqueness.
Example 55 (Ethernet MAC addressing) Ethernet addresses have no namespaces (ignoring VLAN), and so delivery has to be by public flooding and self recognition. No memory is required for this, but each data transmission
occupies the entire channel for all agents, blocking their ability to promise simultaneous availability.
The foregoing results show how hierarchy plays a key role in scaling the local uniqueness of addresses. By
exploiting superagent structures as ‘fixed star’ markers, the addresses of individual agents can be made relative to
the superagent containers. This includes the addresses of sub-agents within them, and so on, recursively down the
levels.
Example 56 (Tuples, namespaces, and the Arabic numerals) Hierarchical namespaces are what allows us to
represent arbitrary numbers with only the Arabic symbols 0-9. By forming tuples of numbers we allow multiple
digits to represent different blocks of the final digit, e.g. 13 is address 3 in bank 1, and 23 is address 3 in bank
2. Without such tuple addressing, ordinal values would have to be represented by nominals, i.e. we would need
a different symbol for every number. For example, singlets, using nominal alphabet [A-Z], can encode only the
addresses [1..26], but doublets (a,b) using relative addresses [0-9][0-9] can encode numbers from 0..99, as they
two values vary independently, with tuple elements counting as logarithmic partitions.
42
4.3.4
Non-local link memory
In spatial memory, key-value networks may be approached in two ways: by storing data locally at fixed agent
locations, and non-locally by storing data as processes that span links. The memory models described thus far
may be called deterministic, local models of storage: they map a fixed address to a fixed location. A different
kind of memory is possible, based on network linkage, by observing that, between any number N of agents, there
may be as many as N 2 − N links. Thus, for large memory spaces, storing memory in the links rather than the
locations could be very efficient. The challenge lies in addressing and encoding the memories. This is the idea
behind Bayesian networks, and the related Neural Networks (for a review, see for example [17, 44, 45]).
Figure 19: Mapping an observed exterior space into an interior representation, through a sensory apparatus, may imply an
essentially isomorphic representation.
4.3.5
Teaching spacetime to tokenize patterns
Sensors are often designed, shaped, or honed, over their evolution, specifically to register particular spacetime
phenomena21 . Such sensor ‘designs’ favour the recognition of patterns by virtue of their structural properties, and
minimize the need for data analysis and post-processing of interpretation. For example, the ability to judge distance
might be enhanced by a sensor array with constantly spaced pattern between locations, acting as a graduated
measuring stick. Through this specific adaptation, sensors are made efficient as pattern discriminators, by learning
that is cached by evolutionary selection, over longer timescales. There is evidence that this is true in the human
senses: eyes and ear do not merely pass data to the brain but perform staged preprocessing and limited ‘type’
recognition, through the existence of specific cells for the measurement of motion. Research has further identified
fabrics of place cells, grid cells, head direction cells, border cells for seeing navigation of spacetime, and even
facial recognition regions, and so on [47–51].
Software techniques, using machine-learning, can simulate this preprocessing, by long-term sensor adaptation,
to train sensors with particular semantics; this is typically done by training artificial neural networks (ANN) to
identify patterns [52]22 . This ability to optimize the recognition by evolutionary learning should be considered
highly significant from a spacetime viewpoint. If one can ask a specific question about external phenomena,
with a specific answer, it is possible to build a discriminator based on the scaling of a fabric that uses semantic
addressing to yield a direct symbol token. This amounts to a large compression of information being forwarded into
a knowledge representation, and suggests that tokenization is a powerful strategy for making cognition efficient.
In sensor training, implicit associations become encoded, over multiple link patterns, by traversing spatial
memory lattices, whose geometry is specialized to play the role of parsers that trigger a specific tokenization
of the inputs. This is somewhat analogous to the creation of a directory or index for fast lookup. A particular
symbolic token may be recalled because it is adjacent, in the index coordinatization, for the context we are currently
exploring. With non-local link encoding, this process is still somewhat mysterious, but it amounts to a form of
semantic hashing.
This training of specialized sensors seems almost fantastic, however, it makes considerable sense from a processing viewpoint. Without a pre-seeded similarity of structure to lay out boundary markers, there is no obvious
way in which an unbiased associative structure could spontaneously break a symmetry to recognize spacetime characteristics like distance, direction, dimension, location, etc; yet we know that they do. What long term learning
enables is the caching of environmental boundary conditions that train discrimination structures by selection.
21 This
22 Thus
seems to be what Ashby called homomorphic machines [46] in his work on cybernetics.
ANNs are more akin to the preprocessing of a visual cortex than perhaps the brain that receives the interpreted signals.
43
Example 57 (Head direction cells) These are cells that perform a form of context assessment, approximately
tracking a compass direction of the head by relative motion. Grid cells and place cells in the brain have a regular
lattice structure that allow an image of space to be projected into an array of agents that mimic the same kind of
connectivity.
This partial mimicry of patterns by spatial structure goes beyond the recognition of exterior sensory patterns.
It applies internally too, in the representations of concepts. In a knowledge system, we have to answer how similar
patterns may be mapped to similar regions, so that uniqueness is assured over an appropriate region, (e.g. in a
brain network, how does the same concept avoid being senselessly repeated at different locations, like a plague?).
A notion of idempotent convergence seems important to maintain stability (i.e. concepts as singular desired endstates)
Example 58 (Deep learning applied to reasoning) Artificial Neural Networks (ANN) or deep learning networks
have proven to be effective at forming stable attractors in the recognition of discernable patterns, even though they
may be inefficient in terms of computational resources. Foreknowledge of a pattern, combined with system design
(where possible), can sometimes be used to simplify the recognition of patterns with much lower cost.
In [52–54], a multi-stage, highly trained series of ANNs is used as an advanced discriminator for sentence
parsing. This approach is interesting, as it does not attempt to preserve grammatical structure, except in word
compound groupings. Moreover, using multiply, individually trained stages, it approximates intentional promises
through a mixture of emergent behaviours and hand-crafted training design. The naming of concepts is emergent
(non-intentional), buried in the internal layers of of the neural network. Thus intermediate concepts may be
presumed to exist, but are only accessible from the encoding at the input, and the associative weights that lead
to particular network activations: their internal names are non human-readable. The use of specific associative
forms, belonging to the four irreducible types features in the authors’ solution.
The model in this work leads to the kind of representation one expects for the processing of an intentional
process: information is promised at the inputs and outputs, and training shapes the network weights to propagate
input to output conditions, assuming the existence of sufficient internal degrees of freedom to compute stable
attractors. We can thus understand ANN training as an approximate, non-linear, best-fit propagator matrix for
input/output, with hand-crafted spacetime symmetry breaking.
A working hypothesis, based on a semantic spacetime model, is that one would expect a successful sensor
apparatus to align its operation with causal structure and precedence, and operate like a ‘prism’ which transmutes
spatial patterns into separated sequential frames of input to be parsed in chunks23 . This hypothesis further seems
to be a necessary condition for remembering location, and navigating in artificial memory, e.g. chips, cities.
4.4
Temporal addressability of memory
Temporal memory includes the encoding of data like transaction logs, melodies, video, and any sensory stream.
How do we encode and recall such sequences in a reliable way?
4.4.1
Sequences and oriented paths
To encode an arrow of time, i.e. a directed sequence, it’s necessary to break the symmetry of a memory space in
some way, by establishing a precedence relation. One way is to distinguish locations and order their consecutive
addresses, either by labelled addressing or by using a hardwired discriminator structure. These may be formed by
any of the causal, irreducible promise types:
Lemma 19 (Partially ordered promises) Partially ordered addressable structures are representable as partially
ordered promise types:
• Precedence adjacency gradient (queues):
precedes
precedes
A1 −−−−→ A2 −−−−→ A3 .
(69)
• Conditional dependence or prerequisite:
later|earlier
earlier
A3 −−−−−−→ A2 −−−→ A1 .
(70)
23 Might we even show that the brain cannot simply be an aggregation of homogeneous autonomous agents, as the cortex seems to be, without
some pre-seeded symmetry breaking from sensory arrangements? There must be symmetry reduction in order to support the different tenant
modules as well as to distinguish different datatypes?
44
• Superagency and encapsulation:
inside
inside
A −−−→ S −−−→ S 0 .
(71)
By breaking the symmetry with an arrow of time, one finds a representation for sequential memory. Since learning
is not sequential but cyclic, long term memory must be able to reliably revisit such sequences to reach an equilibrium along the time vector to end in a summary state. It is plausible that each agent, as part of a sequence, could
reach some kind of independent averaged equilibrium, with multiple sequences representing the same semantics.
This begs the question: how does an observer know whether two sequences have the same intended semantics or
not?
Example 59 (Sequences with constant semantics) When does it make sense to compare two sequences (e.g.
comparing a shopping list with wedding invitations does not make sense)? The simple answer to this question is
that such comparisons must be baked into the immutable aspects of the sensory stream. The boundary conditions of
sensory experience are the only source of information on which to base such semantics. Thus we hypothesize that
this is where long term evolutionary adaptation of sensor recognition play a role. A separation of timescales allows
long term adaptation to similar spacetime processes to feedback and shape the short term sensory recognition of
similar patterns.
4.4.2
Knowledge as equilibrium, with constant semantics
I have proposed that knowledge is to be understood not as information but as an the stable equilibrium of an active
relationship [3, 14]. We may state this hypothesis as a less ambitious lemma:
Lemma 20 (Learning with immutable semantics) Learning by repeated observation over semantic sequences,
is possible iff all samples promise to be piecewise graph homomorphic to one another.
Without a consistent promise graph, data cannot be combined or mapped into a coordinate system reliably, with
preservation of structure. The simplest way to make measurements promise similar graphs is to use the same
sampling apparatus or sensor for all samples. This fixes the boundary conditions, and suggests that, even if sensors share certain infrastructure, it is natural for specific intermediary stages to be routed to specific spacetime
discriminating structures that encode graph-isomorphic structures with the appropriate ordering and interpretation.
Example 60 The eye shares a lens and optical infrastructure. The captured light may be processed by different
specialized memory cells as part of decoding or imposing semantics on the data:
Colour
Grid shape
→ . . . Concepts
(72)
Lens → Retina →
Orientation
...
Even with substitution of an element in the process, these have shared structure, and can share semantics.
4.4.3
Advanced-retarded spacetime propagation and addressability
The promises needed for propagating addressability of space, by comparing agent identities, are related to the
concepts of so-called advanced and retarded boundary conditions in the physics of field propagation and quantum
mechanics [55]. Referring to the discussion in section 4.3.3, we may define.
Definition 52 (Advanced propagation) Information originates from an end point and expands towards possible
starting points.
−d
R −−→ {I1 . . . In , S}.
(73)
Advanced propagation is like ‘back propagation’, as referred to in the training of learning networks.
Definition 53 (Retarded propagation) Information originates from a starting point and expands towards possible end points.
+d
S −−→ {I1 . . . In , R}.
45
(74)
These two propagation methods are conjugates of one another, and both processes are involved in determining a
route from start to end in a complete path. (see figure 18).
Only the end points represent information sources, assuming that intermediaries admit zero distortion24 . The
propagation of a message d from one to the other is possibly universal and carries no significance. However, the
joining of these defines an overlap of information transfer (what is sometimes called a transaction).
Lemma 21 (Path overlap of advanced and retarded propagation) The set of oriented transitions from S to R,
or vice versa is given by the overlap of advanced and retarded propagations for data ±d originating from R and
S respectively.
If there exists one or more connected paths from S to R, involving an arbitrary number of intermediate agents,
then it make be possible to propagate information in one or both directions to connect S with R. These mutual
flooding processes can be joined at any intermediate location along a viable path, provided the message types are
compatible (±d as a lock and key).
Example 61 (Quantum mechanics as an address routing protocol) This construction is very like Feynman’s notion of a spacetime path formulation of quantum mechanics [56], and the Feynman Wheeler absorber theory [57],
as well as Schwinger’s source theory [58]. This is discussed in the Transactional Interpretation of Quantum Mechanics [59]. Between an interacting system and observer, there is an ‘offer wave’ ψ(x, t) described by the wave
function of the system, and a ‘confirmation wave’ ψ † (x, t) travelling from the observer back to the system. The
overlap between these ψ † ψ is the probability of a connected path between them, as hψ|T |ψi are matrix elements
for specific path transitions. In Feynman’s construction is represents a sum over all possible connected paths that
satisfy the energy conservation payment conditions (see figure 20).
ψ
past
future
ψ
Figure 20: A handshake between two propagating waves represents a set of overlap path transitions connecting them.
In this view, the wavefunction of a system plays the role of a pre-arranged address structure (like a directory,
or name service) to distinguish spacetime locations. Indistinguishable locations cannot have different connection
probabilities, else they would be distinguishable. The wavefunction is thus an equilibrium information map of a
system at the snapshot instant of an interaction, and the solutions for ψ(x, t) have the status of an on-going knowledge relationship between observer and observed. Note that, although the time t appears in as an evolutionary
learning parameter in ψ(x, t), this is not the same time as is represented by the propagation from boundary points
S to R. Moreover, the time-reversal invariance represents path-reversal (route) invariance rather than absolute
time.
4.4.4
Splitters and oriented discriminators, encoding decisions
How is a preferred direction realized (reified) in a representational spacetime? There any may processes that
routinely break directional symmetries in the exterior world: the position of the sun, rivers, etc. On the interior of
a knowledge representation, there are no such landmarks, so associations are always relative to something in the
interior. There are two ways around this:
• Encode absolute boundary conditions spatially in memory as homomorphic ‘concepts’, and then associate
directions relative to those concepts.
• Form a Cartesian theatre [10] facsimile of the outer world and use interior properties to represent directional
processes: electrical currents (e.g. in a brain, or an electronic circuit) select a direction of flow; the act of
pushing a memory by imposition or force breaks its directional symmetry; start and end boundary conditions
select a direction, e.g. , like a contact potential, or being inside the plates of a capacitor, or a temperature
field, where some field gradient sets a bias.
Certain spacetime processes break symmetries, by virtue of their nature, e.g. prisms, gradient fields, etc (see figure
21). Sensors can make use of these devices in adapting to pattern encoding. Directional memory exhibit any axial
24 This is a common assumption, but also one that is wholly unjustified. The presence of intermediate agents is likely to distort the propagation
of information, without a considerable infrastructure of promises (see the discussion of common knowledge and the end-to-end problem in [6]).
46
+
+
+
+
+
+
potential gradient
+
direction
discriminator
sequence
_
_
_
_
grid
_
_
_
Figure 21: Ordered spacetime structures form directed graphs, either by autonomous cooperation of memory locations, or by
external calibration through discriminators, routers, prisms, etc.
symmetry. Branching treelike (conical) structures are sequential and directional, with discrimination built in. In a
treelike bifurcation, semantically close concepts could end up close or very far apart in memory space25 .
Example 62 (Optical interferometry and correlation) The optical versions of neural networks, which have much
the same kind of spacetime structure, are diffraction gratings, optical correlators, holographic systems, and other
interferoemeters. These encode information in frequency and phase and yield fringe patterns that can be used for
optical processing. This underlines the point that massively parallel ‘spatial computation’ may be performed in
many ways [62].
4.4.5
Role of separated timescales in tokenizing
As usual, the separation of scales informs the dynamical basis for the semantic machinery. Three timescales for
memory are generally accepted in human learning, and translate naturally for other learning systems by spacetime
considerations [63]:
1. Immediate memory: for data collection, i.e. the ability to hold ongoing experiences from fractions of a
second to seconds. The capacity of this register is very large, involves all modalities (visual, verbal, tactile,
and so on), and provides the ongoing sense of awareness or context.
2. Short-term memory: allows humans to hold information in mind for seconds to minutes. This is associated
with gathering of context.
A special sort of (procedural) short-term memory is called working memory, which refers to the ability to
hold information in mind long enough to carry out sequential actions. An example is searching for a lost
object; working memory allows the hunt to proceed efficiently, avoiding places already inspected.
3. Long-term memory entails the retention of information in a more permanent form of storage for days,
weeks, or even a lifetime. This is the equilibrated memory of a learning process.
4.5
Nominally addressed memory
Numerical addresses, based on metric coordinates are well suited to the labelling of ordered structures [1]. Nominal
addresses (semantic addresses) may also be used to represent ordered structures, if an intermediary supplies the
partial ordering of names by transmuting names into partial ordering, through its spacetime structure.
Address discrimination, using names and locations, is made using the two approaches in section 4.3, i.e. intermediary forwarding switch or broadcast flooding (called a bus architecture)26 . Flooding is an inefficient approach
since it occupies an entire region of space for the duration of a transaction.
Lemma 22 (Address routing) To transport a signal or agent from one location to another, it is necessary and
sufficient for agents to promise the ability to distinguish direction, and a map between a unique absolute identity
and the desired spacetime location.
25 Hawkins has presented ideas about Hierarchical Temporal Memory, in which sparse distributed representations are used to store sequences,
however too few details are in the public domain to evaluate the claims [19, 47, 60, 61].
26 If the brain were a bus architecture, memory would be routed by address. We know of no way in which individual neurons or clusters could
know their own addresses in a coordinated way, nor of a unique routing structure that discriminates on conceptual lines. It is plausible that
sensory bundles from major sensory regions: colour cells, place cells, frequency range cells, etc, could provide some of the labelling needed
for form an approximate regional (digital) address.
47
The proof is as follows. If an agent does not promise a unique name, i.e. a unique public scalar promise, it cannot
be distinguished from other agents. If the location has a unique identifier, and other agents promise to recognize
this unique name, the location will self identify once it has been located. It can now be searched for by covering
the entire space. Once located, sign posts or complete paths can be left to mark routes. A signpost is a lookup
table (map) that associates a class of locations with an identity. If agents do not promise a unique identity, we have
to be able to measure metric distance from a signpost or origin. Then we can assign numerical coordinates based
in distance and direction. Each coordinate system breaks translational invariance by choosing an origin. Either
the map is fixed in the name (coordinates) or paths are encoded into spacetime itself The topology of tuples has to
match the promised directional topology of the spacetime. If neither direction nor some encoded name are given,
unique locatability is impossible. We can now define semantic routing as routing without a metric coordinate
system:
Definition 54 (Semantic routing) The use of conditional forwarding based on persistently promised directions
and the use of lookup tables (directories) to match names and forwarding directions to the next signpost. Signposts
form a vector field that converge onto an externally stable set of destinations [64].
Example 58 illustrates the use of multi-stage semantic routing without explicitly known names. Another example
that uses naming based on weighted paths is the following.
Example 63 (Breadcrumbs, ants, and road-signs) Road signs use nominal addressing, with arrows pointing to
the direction to follow to find the next clue about a named object. The formation of a trail for nominal addressing
is basically a stigmergic process of discovery and breadcrumbs or pheromone trails [65].
Names (and thus nominal addresses) need not only be based on written language. To our senses, the name of a
door handle is given by its shape and relative position, just as the name of a protein is an attached protein with a
shape that gives it the ability to bind to a receptor, like a lock and key.
Example 64 (Machine addressing) Memory addressing in integrated circuitry and layer 2 (e.g. Ethernet) networks attached to buses makes use of self identification addresses. Data are addressed to a destination and flooded
to all agents in their region of space. The owner of the intended address accepts the data.
Example 65 (Named Data and Content Centric Networking) Named Data Networking [15], and other namebased location addressing schemes, like Content Centric Networking [16], Content Delivery Networks [66] use
names as the destination addresses of locations on the Internet.
n
Source
Source
+D(n)
+cache
+F(n)|D(n),I(n)
Receiver
Figure 22: In Named Data Networking (NDN), a trusted intermediary caches and passes along data by name.
Unlike the impositional approaches to data transmission, NDN is based on the idea of voluntary cooperation
[67], and is a form of conditional forwarding based on identity [6, 68–70]. Sources of data promises promise their
data +D(n) by name n, with a signature of the originator to specify provenance (see figure 22). Willing receivers
promise that they want data by name +I(n) using ‘interest promises’ also using the name. Name discovery requires
some kind of directory. Forwarding agents (routers) promise to accept both kinds of promise and forward data them
48
by name, if they have received both voluntary declarations, with a conditional forwarding promise F |D, I:
+D(n)
−−−−→
Router
(75)
Source
(76)
Source
(77)
Router
(78)
Receiver
(79)
Router −−−−−−−−−−−−−→ Receiver.
(80)
Source
Router
Router
Receiver
Router
−D(n)
−−−−→
+cache(n)
−−−−−−→
−I(n)
−−−−→
−I(n)
−−−−→
+F (n)|+D(n),−I(n)
The intermediate router must be completely trustworthy to both parties, as it has the ability to distort the data
and even replace it [6, 68]. The authenticity is presumed by virtue of signing. Although neither the publishing of
the data, nor the interest request specify where to find the data (i.e. they need not have any concept of location
address), the router and the end points do need to understand one anothers’ locations or addresses. Identity
might be encoded by direct and unique adjacency (signpost or interface direction), if the routes are unique. There
is always the possibility that the source and recipient may have moved relative to the router’s understanding of
topology (e.g. in a wireless of gaseous network, where broadcast flooding is the only approach). Thus the router
has to make further promises to give spacetime structure, but which do not involve data:
±addr
Source −−−→
±addr
Router −−−→
±addr
Receiver −−−→
±addr
Router −−−→
Router
(81)
Source
(82)
Router
(83)
Receiver.
(84)
Example 66 (Sparse Distributed Representations) Sparse Distributed Representations [71, 72] use a kind of
semantic hashing to map similar things to similar locations. Some distance measure is used for this, e.g. the
Hamming distance of keywords. An interesting property of Sparse Distributed Memory is that each keyword is its
own memory address, i.e. data and index hash are the same.
Example 67 (Conceptual memory) An intelligent agent has to be able to process multiple concepts and draw an
imaginary boundary around them in order to deem them related, like in a namespace. A simple machine needs to
be able to do this based on a mechanical principle, without some brain-like reasoning skills presumed27 .
4.6
Long range ordered spacetimes: cellular automata
A priori oriented structures are not the only way of breaking spacetime symmetries to distinguish memory locations. In discrete spacetimes, with sufficient symmetry, edge effects and interactions can also lead to dynamical
effects that form memory representations. Regular edge geometries encode directional information.
Many of the spacetime notions that we take for granted in the classical formalization of dynamical systems do
not apply to networks or arbitrary graphs (see paper I). For example, translation, velocity, momentum, derivatives,
conjugate pairs, phase space (position and momentum), waves, Fourier transforms, etc—all these concepts derive
from the existence of a smooth manifold, with a clearly definable set of coordinates28 . In generalized graphs,
where the topology of space changes at every hop, the idea of continuity and conservation are simply nonsense.
If so many of the classical concepts are absent or non-trivial to define, then how can we relate the concepts of
a discrete spacetime to classical notions? Finite state machines are the dynamical equivalent to canonical phase
space for discrete graphs. The correspondence is
(~x, p~) ↔ (Ai , πij )
(85)
where ~x is a position coordinate, p~ is a momentum vector, and Ai is an agent, πij is the promise graph over Ai 29 .
Although analogous, these constructions are in no way comparable structures, unless there is a basic homogeneity
or long range order in the spacetime adjacencies (the long range ordering of the scalar promises on top of the graph
is a separate issue). The study of such graphs as dynamical systems (rather than as approximations to Euclidean
27 We
may ask: how does the brain itself associate concepts? Simultaneous activation is one of few possibilities, as explained later.
don’t want to get into a discussion of differential or algebraic geometric ideas here, so I am being intentionally simplistic.
29 A half-hearted attempt to relate graphs to a Hamiltonian formulation has been posed in [73], leaving many questions unanswered.
28 I
49
or Riemannian space) has only received attention very recently [43, 74, 75], and turns out to be a quite non-trivial
problem. To make progress, one needs to fix a coordinate system (naming convention) for the links emanating
from each node agent (referred to as ‘ports’ in [43]). Graphs with this property are related to Cellular Automata
(CA) [76–78].
The hypothesis explored in these notes is that any stable spacetime phenomenon can be used as a form of
memory, and be attributed semantics by context. The behaviours of cellular automata models are essentially those
of reaction-diffusion models, and their study goes far beyond the scope of this work, but they are related to the
propagation of intent discussed in [79]. What is important to note here is that only special cases of CA may be
used to act as memory that does not simply leak away.
Example 68 Cellular automata (CA) have been used, in combination with learning methods, to model the evolution of land use in Geographical Information Systems (GIS) [80–84]. By modelling each point in space as an agent
(cell), with promises based on on iterative dependency on nearest neighbours, one achieves a local CA model. The
challenge with CA models is in matching boundary conditions around the edges, in order to break the symmetries of the lattice. This is how one injects information about actual models. To accomplish this, neural network
techniques of back-propagation have been combined with the CA models.
Even with a notion of propagation, it is not trivial to define something like a momentum, related to a graph
transition function (see section 4.7), and the usual formulations of Poisson brackets and Hamiltonian or action
principles make little sense without a scaling limit to a Euclidean space.
Example 69 (Self-healing communication trees) The cellular connection network infrastructure, proposed by
Borrill [27], to self-organize a self-stabilizing spanning tree on top of a homogeneous trusted lattice of agents with
a fixed number of ports, is a cellular automation whose function is to route messages using an optimized flooding
scheme and nominal addressing (flooding is optimized by learning of routing tables by the memory agents). The
operation is similar in concept to NDN, and makes use of purely local agent adaptation, with only emergent
organization. The aim is not to stabilize a rigidly addressed memory, but rather to stabilize an adaptive spanning
network across fixed nodes, each of which can act as memory. This is best suited for retrieving encoded memories,
already placed at the nodes, but does not immediately support a natural ordering of locations for storing distributed
sequences from an external source. The latter can be arranged by adding a globally ordered naming convention,
in accordance with the reduction law [2].
4.7
Graph dynamics, thinking, and reasoning
Knowledge spaces of concepts, adjacent by association, go beyond memory systems. The fact that they can
propagate intent, means that can they also compute results, in the sense of executing algorithms. As feedback
networks, the stability of concepts and outcomes is thus a matter for serious investigation. Our understanding of
the dynamics of generalized graphs is still in its infancy, but a number of special cases submit to analysis, and
challenge us to think about spacetime concepts more carefully.
Example 70 (Causal Graph Dynamics) The dynamical evolution of labelled graph systems is a theme that can
be related to a vast number of fields of study practical examples, from cellular automata models of land use, to
biological tissue, to ecological systems, and so on. Recent work, which tries to address the dynamics at a formal
level, builds on notions of dynamical systems from physics [43, 74, 75], by using labelled graphs, without explicitly
modelled semantic constraints. Graph vertices (analogous to promise theory agents) make connections to other
agents by named ‘ports’, i.e. links that are promise labels from a fixed alphabet of types. The semantics of the
types are unspecified, and agents do not have to have all kinds of ports. The type alphabet thus acts as a matroid
basis for vector promises. The model of causation is based on locality of interactions, i.e. neighbours, within some
adjacency radius. The locality is provably related to causality.
Causal Graph Dynamics shows that it can be a model of computation, thus (as a constrained subset) a fuller
spacetime dynamics can also represent computation. This ought to be obvious, since computers are made from
labelled spacetime, but it is still a source of controversy in some fields. There is also a tangible similarity with
quantum field theory.
Example 71 (Differentiated spaces) Spaces like rooms, buildings, and cities are strongly differentiated spaces,
where the forces of morphology have already reached an equilibrium that decouples their function from the dynamics within. However, the origin of that morphology presumably begins from a less differentiated state somehow.
Modelling the emergence of structure by exploring boundary conditions and symmetry breaking mechanisms is
also a major industry in science.
50
Example 72 (Cellular automata) Cellular automata have been studied since the pioneering work of Baricelli
and von Neumann and others [76–78, 85, 86]. A cellular automaton is a graph that normally forms a regular
degree graph of neighbouring cells, forming a lattice, each of which is an autonomous agent. Each agent has state
and that state evolves according to a promise dependent on its neighbouring cells. The dynamics are local, but
the relative times of transitions of each neighbour are not completely defined. A normal approach is to consider
a complete cellular automaton as a rigid spacelike hypersurface that evolves according to a global clock, with a
transition function mapping all cell states ci (t) synchronously:
T : ci (t + 1) → ci (t),
∀i.
(86)
The state at t + 1 for all agents is thus based on the frozen snapshot of states at time t, and all update syn-
c4
c2
c5
c3
c1
Figure 23: Cellular automata are regular degree graphs whose cellular agents promise to modify their state based on the states
of their nearest neighbours.
chronously. This does not imply infinite speed of propagation, because influence is restricted to nearest neighbour
interactions (see figure 23): the velocity of influence is one hop per time tick. However, it implies infinite speed of
coordination: i.e. long range order in spacetime is assumed, which is phenomenonologically plausible for small
scale living systems observed by light and sound, etc. Monte Carlo updating is also possible to simulate quantum physical systems like the XY-spin model [87]. Due to the multi-dimensional dependency of cellular automata
state transitions, the evolutionary outcome of the system is strongly dependent on the updating semantics of the
transition function.
Cellular automata act as computers (e.g. see [3, 78]), and can perform routing functions as addressable
networks (e.g. see [27]).
Example 73 (Directed eigenvector centrality) Eigenvector centrality is a measure of a graph based on the observation that the graph adjacency matrix is a transition matrix, whose self-stable flows result in a shift of link
weights onto nodes, through the principle eigenvector (a consequence of the Frobenius-Perron theorem). This was
well understood for undirected graphs. A variant of this technique was used by Google search for its PageRank
method of search ranking [42]. The result was extended for directed graph flows in [41]. For a directed graph, the
result is somewhat different to that for a non-directed graph. Only absorbing sinks nodes end up with any weight,
because transitive adjacencies funnel all the flows into sinks. If we apply this to a reasoning process, it would
imply that directed learning networks would not be able to distinguish between reasoned outcomes by relative
weight. However, a number of remedies were noted in [41], including the equivalent of back-propagation, and
feedback by pumping results back into the sensory input end of a network. This link between search ranking and
reasoning illustrates how search ranking acts as a simple kind of Bayesian network. The technique is particularly
interesting as it leads to stable regions [39] that could naturally be associated with concepts, based purely on
spacetime connectivity, and thus it offers insight into the way concepts may emerge without deliberate supervised
training in a true artificial intelligence.
Much remains to be understood about the behaviours of systems on which we are relying for artificial knowledge representation and reasoning. It is tempting to think that, because certain applications approximate deterministic behaviour, when driven by a dominant directional process, like input sensing, all behaviours must be therefore
be deterministic. That is far from the case. Indeed, even in the case of humans, we know that when external input
is switched off, the brain enters states of hallucination and dreaming, where reason seems to follow quite different
rules. Several questions stand out:
51
• How do graph dynamics play a role in knowledge representation?
• Is knowledge storage and recovery stable to the dynamical states of a knowledge system?
• Can all concepts be observed (read or written) simultaneously, or are is it possible to excite partial eigenvectors that are not simultaneously compatible, as in the case quantum mechanical canonical pairs [88]?
• What role do pointed graphs [75] and spanning trees play in bounding the determinism of reasoning?
• How does parallelism affect the linear nature of sequential reasoning?
We return to the matter of reasoning in section 6.
4.8
Agent phase and conditions for memory, smart behaviour
Agents may be in a free gaseous state, as atomic or molecular agents, or bound together into a solid lattice with
limited long range order (see paper I [1]).
• In a gaseous phase, agents mix freely and encounter one another by random walk, with short-lived adjacencies. This is a useful way of mixing and discovering new information. Coordinates in a gas can only
be through arbitrary numbering and message broadcast. Think of people milling around the lobby of a
hotel, meeting by chance, or calling out “Come in number 9, your time is up!”. Extended structures like
strings, stories, paths, cannot be remembered in a gaseous phase, because they have no stable or persistent
relationships.
• In a solid phase, information is located in a very regular, ordered way. Adjacency is a lasting knowable relationship. One can form a regular coordinate system with purely local predictability. Search is not required.
Think of visiting a specific tenant in a numbered hotel room.
The transfer of intent (including structure) requires both (+) and (-) promises to propagate, mutual bindings or
tensor promises are involved in structuring of data.
Lemma 23 (Sequential memory implies vector promises) To promise stable sequences, agents must bind with
vector promises.
This was discussed under temporal memory. To the extent that a solid implies a stable ordered sequence of vector
bonded agents, without necessarily having significant long range order, then memory requires a kind of permanent
solid state. Given this, there may be something to learn from cellular automata about how a smart knowledge
representation (i.e. ‘mind’) can remain sufficiently plastic to change.
Example 74 (Immunology and neurology) Memory structures and pattern recognition discriminators are often
constructed from solid state regular lattices, that provide regular calibration grids for observation. The brain
cortex, and other related memory structures are form a largely solid state network, albeit irregular on a small
scale. The immune system, by contrast, is capable of learning and rememberings previous infections based on
populations of antigen carrying lymphocytes (B and T cells). In order to accomplish this, the individual agents
must be able to promise greater capabilities as individuals. Thus, in a gaseous phase, they substitute scalar
promises for vector and tensor linkage. For a review, see [89, 90].
4.9
Summary
Information may be encoded as memory by using the familiar spacetime properties of spacetime:
• If a spacetime is deterministically addressable, it can store patterns of information in a retrievable manner.
• Separation of scales enables scale-free addressability.
• Consistent addressability is possible if memory agents are distinguishable, and there is end-to-end propagation of data between the agents in two directions between readers/writers, and memory agents.
• Four encoding methods may be matched with mechanisms for propagation of intent:
– Containment and aggregation.
– Causal ordering.
– Composition or cooperation of parts.
52
– Proximity or topological nearness.
These elementary considerations show that any semantic spacetime can perform a learning function, and can
therefore acquire information, and store it, leading to a knowledge representation. We can now study the structures
of tokenized knowledge in detail.
5
The structure and semantics of knowledge
The hypothesis, which forms the basis for this work, is that knowledge is the sum of an observer’s tokenized
sensory inputs, iteratively confirmed, over multiple observations, and laced with a feedback that is integrated into
a running assessment of ‘here and now’. This process leads to a unique world-view, for each observer, resulting
from each agent’s unique world line history. The non-linear feedback amplifies the fact that experience is nondeterministic, but if knowledge emerges around stable attractors, it may be cached in a compressed, tokenized
form. This is essentially like a linguistic representation of experience30 .
Constructing a plausible model for this is the subject of this section. Given the network complexity of associations formed during full-sensory learning, it is unrealistic to expect tokenized concepts and associations, committed
to memory, to be complete and faithful representations of the input. It would be a misunderstanding of knowledge
systems to expect any such thing. A symbolic knowledge representation can only be a tokenized approximation
of the fruits of learning31 . The tokenized form is not a substitute for the actual experience, only a summary of
its interpretation. The implication is that we cannot pass on complete experience in a tokenized (linguistic) form,
because it enters a system through different sensors and touches fewer concurrent sensations.
5.1
Learning, familiarity, and linguistic tokens
The implications of learning are profound. They suggest that deep knowledge cannot easily be imposed onto an
agent from the ‘top-down’, in symbolic form. In other words, one cannot induce knowledge by a short-cut, like
reading or listening to a lecture. One may try to simulate a complex experience, in a linearized stream of communication, just as a Turing machine can simulate any information based process eventually, assuming perfect fidelity,
but it cannot be done efficiently or with a predictable outcome, especially under the shadow of approximation,
because the process of learning is iterative and non-linear32 . The implications of this are not well understood, and
yet we work hard to represent knowledge in a variety of media, and design ‘artificial intelligence’ that has its own
approaches to learning.
Example 75 (Is a book knowledge?) Is knowledge the same as a representation or facsimile of knowledge? Do
we possess knowledge simply by owning a book? Is the process of knowing the content of a book about revisiting
it to gain familiarity? There is a clear temporal aspect to knowledge that we need to address. The taking of a
photographic snapshot of a space (a time-slice) may capture information, but we do not ‘know’ that information
by simply capturing the image: knowledge comes from revisiting the snapshot, relating it to other familiar concepts,
to the point of familiarity. Thus even an approximate representation does not become knowledge simply by being
disentangled from its network and encoded into a form that can be passed on.
To understand the implications of these remarks, in as broad a range of scenarios as possible, we need to get to
grips with how knowledge is valuable to different kinds of processes, without preconceived notions! Knowledge
is not just that which is familiar, but also descriptive of context.
5.1.1
Can familiarity be passed on?
Familiarity is a major part of our understanding of knowledge.
Definition 55 (Familiarity and significance of knowledge) The attribution of a weight or score that grows with
frequency of use or accumulated usage, in a memory network.
The non-linearity of learning and memory networks leads to a simple point:
30 This might help to explain why we experience such a close connection between thinking (in the sense of storytelling and reasoning) and
language.
31 In principle, the same level of approximation need not be a limitation of an artificial knowledge systems, with simple sensors, but the
economics of learning point to a diverging need for resources, and hence the eventual importance of approximation.
32 Promises of knowledge, made by other agents, can be assessed by a receiver and either accepted as trusted facts or merely as data. When
concepts and semantics are taken as authoritative standards, they change from being voluntary emergent promises to impositions, and hence
they lose their ability to avoid inconsistency and be represent local circumstances. Standards may be useful as scoped definitions, but their
validity is fragile to variation in the environment.
53
Lemma 24 (Familiarity cannot be propagated) Familiarity of a learning network cannot be transmitted faithfully through a serial (linguistic) representation.
This observation follows from the sensitivity to boundary conditions in non-linear systems, and the relativity of
familiarity with each concept to neighbouring concepts, in spite of intentionally deterministic procedures [91].
No two agents can be prepared and maintained in exactly the same state for an extended process of simulated
learning. They would have need to have to begin with identical initial conditions, and maintain identical external
conditions throughout the process of information acquisition. It could not be identical to the knowledge acquired
by experience, because a serial stream cannot be processed in the same way as a parallel sensory experience,
simultaneously, this is impossible (on many levels).
The network aspect of knowledge is crucial. Knowledge is not isolated: it has lasting bonds to notions of context. If lasting connection were enough for represent knowledge, any solid object (a building, a vase, a book) could
be used as a representation of knowledge. Any specific arrangement of space, persistent in time with sufficient
stability, or thing rendered timeless by isolation from change, could encode what has been accumulated, and could
be transmuted into familiar knowledge by repeated interaction.
5.1.2
Tokenized knowledge: topic maps and semantic indices
Knowledge management has been studied for many years in the context of library systems and of the semantic
web. One of the early technologies for representing concepts formally was the idea of Topic Maps (for a review,
see [92]). This emerged from the need for document classification and indexing. An approximate rendition of the
models is as follows:
• The world consists of subjects (things, people, ideas, etc) about which we can say something.
• These subjects map to ‘topics’ in a software data model.
• Our linguistic understanding of subjects is glued together with associations we make. These are descriptive
relationships between linguistic (nominal) data.
• Associations are made by individuals, making topic maps subjective viewpoints. They can be merged to
mimic learning from one individual to another, or a blending of experiences.
The net result of a topic map is a graph structure whose nodes are topics and whose links are associations. Additional structure is given to type concepts, which are meant to model context. The information model is simple to
parse and to search, in the manner of an index.
The strength of Topic Maps lies in their autonomous, subjective viewpoints, based on keywords. This further
allows topic maps to be merged without conflict (the language is promise oriented). However, their weakness
stems from a model that is deliberately abstract. Nonetheless, it supplied much of the inspiration for the present
work33 . By re-examining the ideas without those constraints, using an independent framework of semantic spaces,
we end up with a more detailed model that retains the simplicity of the old at a particular scale. A critique of topic
map knowledge representation was given in [14].
The idea that we can take a complex set of observed attributes, from a mixture of sensory inputs and pre-learned
concepts, and project them into a simplified ‘symbolic’ representation is profound; it is the key to ‘approximation’,
without which it would be likely impossible to bound the resources associated with reasoning (see section 6.4).
Tokenization solves three issues:
• Reduction of the dimensionality of sensory data.
• Compression of sensory information to a singular alphabetic form.
• Mapping variable inputs to an invariant representation.
5.1.3
A minimal observational ontology for knowledge, based on promise theory
Starting with agents and promises as a simple axiomatic framework, we can try to formulate a minimal set of
agencies¡ and promises that would be needed to represent knowledge. In order for concepts to be encoded, there
must exist some form of semantic space. The precise nature of the space may limit what knowledge it can learn
and encode. Based on a scaled view of spatial agency, and the axiom that knowledge is acquired independently
33 Topic maps probably suffered from premature standardization, an artificially taxonomic approach to classification ontology (to fit into a
data model technology based on typing), and closeness to a particular imagined use-case; thus the scaling of the model was never considered
at all.
54
by each observer, based on a context of its temporal interactions with exterior agents, we can now develop some
minimal number of structures for knowledge representation, by mapping the parts systematically to agents and
promises.
spacetime
(−)
(+)
senses
(+)
world
memory
(−)
mind
interior things
exterior THINGS
observing agent
horizon
Figure 24: Knowledge representation is based on the viewpoint of an observer agent, assessing the sensory promises made by
the agents in its exterior, and encoding its understanding in a system of interior agents making ‘knowledge promises’. These
knowledge promises map to ‘associations’ in semantic memory.
Example 76 (What happens in space stays in space) Is every familiar solid object something that can represent
knowledge? Certainly, why not. However, knowledge tokens are only isolated characterizations: a kind of knowledge, but not a faithful copy of the process that made them. They are discretized tokens that have been isolated
from their origins, i.e. abstracted representations. The fidelity of knowledge is not guaranteed.
5.2
Labelling and naming conventions: the observer’s view of the world
With reference to figures 24 and the original figure 1, we begin by with a model of a single agent observing
the exterior by assessing exterior agents through a number of sensors. Inside the observer (within its superagent
interior) are conceptual agents, which we may refer to as its ‘mind’, of unspecified nature, used to equip the
superagent observer with memory and knowledge representations. The primary distinction is thus the sensory
horizon between interior and exterior agencies.
5.2.1
Interior and exterior spaces of an observer
Distinguishing the exterior or physical world of an agent (that which can be sensed) from its internal representations34 is the first step in discussion knowledge.
• Exterior ‘world’ agents promise varying levels of transparency and predictability. The host agent’s understanding of the world comes from interacting with and assessing these, if it can.
• Interior mind agents: memory agents that record a semantic imprint in the form of concepts. Each interior
represents a single observer viewpoint, and an independent knowledge representation.
• Interface sensors promise to observe and assess that which has been promised observable.
We need not be more specific about how the representations work until a particular realization is needed.
34 Art
is the lie that enables us to realize the truth, said Picasso.
55
5.2.2
Interior and exterior agents
I use caps for exterior agents and lower case for interior representational agents, thus we may define:
Definition 56 (Exterior THING) An agent that makes promises in the exterior world of a specific agent, which
observes. Whatever the nature of the THING, it is the promises it makes that make it observable.
This maps to an observer’s characterization and concept:
Definition 57 (Interior thing) The interior representation of something promised on the exterior of an observing
agent, i.e. a full or partial concept.
The agent structures we need include the following35 .
Definition 58 (Exterior (real) agents) Agents that can be distinguished by any observer in the exterior world.
These are not interpretations, but structural spacetime identifications.
exterior
T HING
(+) P ROMISE
C ONNECTION
PATTERN
PATH
physical source agent
Exterior agent
Information supplied by a thing agent.
a physical adjacency between things.
A geometric configuration of agents.
A one-dimensional ordered pattern of THINGS .
These basic exterior agencies require a memory representation within the observer. The interior representations
are one to one, but the representations have the ability to make different kinds of promises:
Definition 59 (Interior (representational) agents) Internal structural agencies needed to map THINGS in the exterior world to interior ‘thing’ representations.
interior
Thing
Concept
(-) Promise
Connection (association)
Pattern
Path
5.2.3
mental representation receiving agent
An internal representation of a THING .
A scaled aggregate ‘thing’.
Information accepted from the THING to condition the internal thing.
An adjacency between interior things, with interpretation added.
A geometric configuration of agents.
A one-dimensional ordered pattern of THINGS .
The named roles in knowledge representation
The basic roles for categorization may be formed from the irreducible association types, by alias.
Definition 60 (Role promises by interior agents) Roles are promise patterns, common to a variety of agents.
These translate into concepts in an interior representation. A subset of knowledge structural relationships promised
by interior agents. Some examples:
interior agent promise
image / representation (default)
location
event
concept
pattern
association
story
topic / theme
subject /theme
model
categories
class
idea
35 I
mental promises
internal representation of a thing
a thing that represents a spatial position
a thing that represents a spacetime position
a generalizing label representing common promises
a geometric configuration of agents/promises
a promise relating subjects
a linear collection of agents ordered by associations
image of a Thing or Object
image of a Thing or Object
a superagent promise of complex spacetime representation
a generalization of several concepts
”
a selection of things taken together
hope the following definitions suffice as plausible to most readers (without trying to appease other nomenclature in the literature).
56
Note that, by contrast with Topic Maps, I treat the term ‘subject’ as an interior ‘mental’ agent, rather than a physical
agent. Associations between the interior concepts are explicit vector promises. Any scalar attribute can be made
into a generic vector promise by introducing a central matroid agent which represents the named attribute and
serves as a membership hub for all agents that possess the attribute.
These conventions follow from the basic promise theory representation of types (see paper II, section 2.6) as
agents that make specific behavioural promises, i.e. what we call roles. Thus types are simply aliases for roles. A
basis agent ‘thing’ can thus play many roles in knowledge representation without being a fundamentally different
type of object: roles, concepts, exemplars, generalizations, classes, categories, models, ideas, etc are all roles
played by thing representations rather than different fundamental types of thing/THING. There are relatively few
agent roles in the exterior world, because there are fewer ways to make feedbacks and links to form more complex
representations. In the interior (what we can refer to as ‘mind’), there are many more possibilities, unconstrained by
reality. Imaginations and analogies that result from feedback and mixing of patterns provide the richer menagerie
of forms that we use to describe knowledge.
5.2.4
The observer’s clock, and sense of time
When an agent samples its sensory stream, the formation of a context from the samples sets a timescale associated
with the input Tsensor . This is the clock by which observations are accepted (not the rate at which impulses are
THINGS
input
state
memory model
sensors
context
concepts
associations
introspection
sensing
behaviour
Figure 25: The functional roles involved in continual transmutation of observation into state and then concepts (memory
model), from left to right. Notice that agents’ promises (not the agents per se), on the left, determine what later agents, on the
right, can learn and know. Memory feedback, from the right, can also cycle back around, and be placed on a par with direct
observation. We could speculatively refer to memory feedback, without new input, as dreaming.
necessarily given). It determines an agent’s ability to receive information and pattern it into a sense of ‘now’.
Lemma 25 (An agent’s internal observational clock completely determines its sense of time) Let A be an agent
promised a steam of data that arrive with an average interval of T+ units, and are accepted every T− units. Thus
the expectation time to receive a sensory signal is greater than or equal to T− .
The proof follows from agent autonomy: no change can occur in the agent unless it accepts the external data or
modifies its own internal data by its own mechanisms. This is what is meant by its internal clock.
5.3
Concepts, interior structure, and semantic superagents
The obvious function of concepts is to aggregate observed THINGS into summarizable tokens, labelled by the
context in which they were observed. Concepts begin as representations of external spacetime phenomena, in the
internal spacetime of an observer. However, one there is a sufficient library of concepts, new ones can also be
invented ad hoc from within, by mixing and distorting existing ones. If an agent forms a ‘new’ idea, this can be
learnt and remembered, as a separate internal concept, with the iterative learning replaced by repeatedly rehearsing
the idea, simply by thinking a lot about it.
Definition 61 (Concept) An agent, or superagent, within the internal semantic memory space of an observer,
which promises a representation of some identifying characteristics, associated with imaginable or externally
observed phenomena, and addressable by a locally unique name, for each context in which it arises.
57
This slightly involved definition allows for concepts to be observed directly, modified internally, or even invented
randomly. Concepts are distinguished from one another, by their agent boundaries. Being discrete, they can be
thought about separately, thus they are discrete semantic units. However, the fact that they can be thought about
independently of sensory input means they can also be invented without observation. Their physical encoding
in spacetime may overlap with other concepts, depending on the scale and composition of their representation.
Concepts need not be mutually exclusive: they may overlap with one another. Their contextual (conditional)
nature means that they are unlikely to be completely reducible to unique, mutually exclusive semantic sets (like a
periodic table of atomic elements); rather, they will be like a rich and varied chemistry of nuanced compounds.
New concepts are created all the time in our human minds, and it makes sense to extend this process, by
analogy, to recognize the possibility for ideation in any kind of learning system. The graphical representation of
concepts then takes the form of patchworks, i.e. non-orthogonal networks of connections, linking to matroidal
basis members. This structure reflects how the concepts coalesce from rough, noisy environmental interactions
over a range of contexts. Concepts (including synonyms, aliases, etc) might well reach into a postulated ‘space of
all notions’ pervasively, in some regions, and barely cover at all in others. Concepts may also overlap by sharing
crucial sensory ingredients.
thing
thing
+concept
span
thing
"exemplifies"
"generalizes"
thing
Figure 26: Concepts promise a basic set of characteristics that generalize the things that play the role of its exemplars. Note
the spacetime structure is a very familiar one of aggregation and scaled representation (a scaling transformation).
5.3.1
Encoding or realization of concepts (reification)
To represent, encode, or ‘reify’ (in knowledge jargon) concepts in a knowledge representation, we identify the
available agents, belonging to the representational medium, at the representational scale (see figure 27). It seems
natural to assume that any concept may be a superagent, formed from collaborative subagents, in the general case.
This means that concepts may overlap in two ways: through the presence of shared interior subagents that belong
to more than one concept, and through the repetition of exterior promises made between concepts (see figure 27).
Example 77 (Concept representation medium) Concepts may be attributed to memory configurations written
on any semantic space or medium, e.g. a crystal lattice, a collection of proteins, an agricultural planting pattern,
the layout of a town or city, etc.
Figure 27: The overlap of two conceptual regions might involve agents who represent common concepts, but might only
include their promises.
It may be assumed that there is a ground state for the conceptual fabric, i.e. that there exist basic building blocks
for concept agents, in the form of discrete semantic units. All concepts would be broken down into these, but there
may be no upper limit to their combination into superagents, allowing concepts of any aggregate size. This ground
58
state set must be finite, but might be infinitely extensible. Limits on the size of conceptual agent clusters would
likely be economic in nature.
Definition 62 (Continuity of concepts) An overlapping collection of clusters, composed from base agents, and
linked by association, bridges clusters into a single continuous cluster.
From these simple speculative proposals, grounded in graph theoretic parsimony, we arrive at a natural picture of
concepts. Concept space is continuous, in the sense that it contains the overlap of open collections of clustered
basis agents, but concepts themselves are more or less discretely identifiable, with the caveat that it might be
difficult to decide exactly where the boundaries between them lie in general.
Example 78 (Cognitive linguistics) In the field of cognitive linguistics, concepts are treated as ‘semantic units’,
(i.e. signal agents in PT parlance) that exist independently of language [20]. Although the concepts of cognitive
linguistics are not well developed in a formal sense (even by their own criteria), they seem to be broadly compatible
with a promise theory view in which concepts act as spanning sets for words in contexts. This is the view I take here,
based on experiences with systems of agents in CFEngine [93], etc. Langacker is vague on what constitutes the
spanning sets, but indicates that the following general containers are involved in classifying concepts: space, size,
shape, orientation, function (container, force, motion, role), material nature or composition, and also economic
issues like cost. These regions are called ‘spaces’, and have a close resemblance to the notion of namespaces [2]
as containers of concepts, analogous to directories of files on a computer. In Langacker’s view. metaphors like in
the associative overlap between conceptual classifications36 .
There is a superficial similarity in structure between Cognitive Grammar [20] and the spacetime view taken in
this note series. It is plausible that this spacetime view could provide a more rigorous notion of conceptualization
that is compatible with it.
5.3.2
Discreteness of concepts
Some properties of concept agent representations:
• The discreteness of concepts does not imply discreteness of their semantics, and vice versa.
• Concepts need not be mutually exclusive in their degrees of freedom (orthogonal with respect to a basis
set, as described in paper I): just as vectors may be partially independent without actually being totally
orthogonal (e.g. East and North-East are linearly independent, but not orthogonal). So we can think of
concepts as tensors in the promise matroids of knowledge representation space.
Example 79 (Non-orthogonal concepts) The concept of blue is independent of the concepts of colour or turquoise,
but it is not completely orthogonal or unrelated to them.
Example 80 (Artificial learning networks) Artificial Neural Networks (ANN) can be used to learn concepts from
a curated input stream. Due to the nature of ANNs, concepts are not represented by any single node, but rather
as virtual knowledge space agents that are distributed throughout links between aggregated basis agents in a
distributed and overlapping way. This contrasts with a semantic network, e.g. Topic Maps (TM), where concepts
are pre-curated without learning. In ANN, concepts are decentralized in their representation in order to promote
fuzzy matching (as a non-linear function of a sum over weighted bases from multiple network layers, with i, j
running schematically over layers and members); in TM concepts are completely centralized in single graph nodes
(with matroid basis vector êi and i running over the graph nodes) and have no representation that can be matched
to data:
X
ANN
concept ' V
wij êij
(87)
ij
concept
TM
'
êi .
(88)
Thus TM has simplicity and efficiency of representation, while ANN has generality and a basis in sensory learning.
36 Langacker riles against the notion that concepts must be discrete structures, presumably in protest against the discrete mutually exclusive
treelike structures of generative linguistics [37], suggesting instead that concepts might overlap; however, there is no contradiction between
discreteness of superagency and the possibility of overlap.
59
Real world objects are composed of discrete elementary properties combined into compounds (like the table
of atomic elements, or the 19th century preoccupation with taxonomies), while the conceptual models need not
be limited to that. An appeal to topology allows us to formulate concepts as patches, which (together) form a
covering of a space of basis parts. This shows that, even without such a fundamental ‘table of elemental concepts’,
it is possible for a finite number of agencies to encode properties to span an effectively infinite realm of possible
multiset combinations.
5.3.3
Role of context in discretization
Contextualization of concepts further allow single concepts to be represented as a set of agents, distributed over the
combinatoric contexts. The encoding of concepts in the neuronal space of the human brain is one of the mysterious
but highly relevant examples of encoding of concepts.
Example 81 Recent studies of brain activity using MRI [18] to scan the brains of listeners during English language readings showed how brain activations were grouped into to broadly cohesive cortical regions [18], with
similarities even across different subjects. The authors posited that English words were mapped by brains into
broad clustered meanings that they labelled with the following categories: visual, tactile, outdoors, place, time,
social, mental, person, violence, body part, number. In current parlance, these could be also called spacetime,
state, self, senses, relationships, and danger. The status of these studies is currently under scrutiny, at the time of
writing, due to the discovery of software bugs going back several years, that may be showing false positive results.
The aggregation of observables in this example suggests how apparent ‘types’ or categories of knowledge
emerge, which could, in turn, be identified with tokenized context states (see later discussion around figure 39).
Context-delineated aspects of a concept representation are ‘roles’ in the promise theoretic sense. Since these are
stable under repetition, by the assumption that context is stable, they must be shaped, relative to a network spanning
set, by the same semantic discriminator that fashions context, working in a predictable manner.
A closer look at context is thus warranted, based on the foregoing remarks, and a rephrasing of the definition
of context may also be posited:
Definition 63 (Context (version 2)) The accumulation of state, over the recent history of an observer, such as to
form the basis for the semantics of long term observations.
This rough description will be the model of context for explaining how tokenized names themselves become
concepts. From this starting point, we can obtain trains of thought, patterns of usage, and so on, just by linking
agents together. Because context linkage plays the same role as matroid representation, context effectively becomes
a part of the semantic coordinate system for observational memory.
5.3.4
How are concepts acquired?
Concepts are memory agents (and clusters of agents, linked by association) in knowledge space; that is, they are
interior agents, within the interior of an observer horizon. The promises they can make are about being associated
with one another in different ways, their strength of association, freshness, etc (note the difference between a
promise and an association). A concept may be promised internally (i.e. constructed) as a mirror of a promised
relationship in the exterior ‘objective’ world of the observer; alternatively, it may be promised entirely on the basis
of interior patterns, by introspective thinking about tokenized patterns and their perceived similarities.
For an imprint to be memorized, a THING Oi , in the objective exterior, promises its visible ‘essence’ ei to an
observer’s eyes or sensors Si , allowing the observer to sample it (see the left hand view in figure 28):
O1
S
+e
1
−−→
−e1
−−→
S
(89)
O1
(90)
If the promises are kept, this sampling occurs on a timescale Texterior . The agent E imposes a representation into
the memory container whose image is some agency C1 :
S
C1
+R1 |e1
−R1
−−−→
C1
(91)
S
(92)
If kept, the latter exchange happens a timescale Tinterior . For this to work, one must have
Tinterior ≤ Texterior ,
60
(93)
L
O1
O1
O2
+e 1
O2
−e2
+e1
−e1
+e 2
S
+e 2
S
+R 2
+R 1
+R 1
C1
−e2
−e1
directed association
Observer
A
C1
C2
+R 2
C2
directed association
Figure 28: Associations are formed in the mind of an observer. on the left, the association is formed without external impetus.
On the right, the observation of a link (perhaps through an intermediary L) is observed and represented as an association.
Associations may or may not reflect actual structure of the real world: some we call abstract, imaginary, or conceptual.
i.e. a mind must work quickly relative to what it hopes to observe.
The importance of these timescales is particularly significant when sensor discriminators are chained together,
in order to modulate one anothers’ capabilities. In that case, the timescales modulate and throttle one another.
Chaining is a technique attempted as a way of using neural networks for non-trivial reasoning (see example 111).
Example 82 (Social networks as concepts) The formation of social networks has the same basic structure as the
knowledge networks alluded to above. A social network may be considered a smart space, and indeed an intrinsic
knowledge representation. Links between members of a social network lead to clustering of cliques and regions.
Recommendations by one member of a group get passed on to others until a group also represents an interest in
that topic, by epidemic transmission [94, 95].
Members who are not interested in the mentioned topic will fall away or start new groups. Thus overlapping
memory locations cluster around topics to which they feel an affinity, forming concepts. Their affinity depends on
a context in the wider world that brings about the topics in question. Parts of the concept cluster may come at the
topic from different angles, triggered by different atoms of context during a concurrent activation.
5.3.5
Concept feedback, or introspection about THINGS through things
Once learnt, the stability of concepts and their address locations can be improved by feedback from the inner
representation into the context assessment stage of the sampling sensory apparatus S, which allows recognition
and discrimination of internal memories along side new inputs. This serves both a calibrating, self-stabilizing role,
and it also suggests the possibility of deliberate calibration extension, by matching and linking concepts relative to
similarly patterned concepts, when inputs are subdued. This introspective processing of things, without the anchor
of sensory reality, seems like the likely origin of what we consider to be new ideas. If existing memories shape
new versions of concepts, then conceptual reinforcement may be cyclic:
O1
S
S
C1
C1
+e
−−→
−e
−−→
+Rnew |e1 ,Rprevious
−Rnew
−−−−→
−Rprevious
−−−−−→
S
(94)
O
(95)
C1
(96)
S
(97)
S
(98)
This has precisely the form of a Bayesian learning update process for non-empty ei . For empty ei , it represents
thinking or introspection, while input is suppressed37 . For now, I will leave this possibility open ended. However,
there is some evidence that this mechanism is active in human brains [19], as in figure 25.
Definition 64 (Introspection) When an observer agent feeds back concepts from long term learning, back into its
sensor and context stream as though they were new input.
37 Dreaming and introspection are not exactly the same, as we know from brain research. During dreaming, not only are sensory inputs
suppressed, but long wavelength waves stimulate the brain at a coarse scale, in some kind of annealing fashion.
61
O1
O2
+e 1
−e2
−e1
+e 2
Observer
S
2
+R 1
C1
+R
directed association
C2
Figure 29: Once learnt, the stability of concepts and their address locations can be improved by feedback from the inner
representation into the sampler E, allowing recognition and discrimination.
5.4
Associations defined
Associations are the links between the material agents of conceptual knowledge space. Let’s consider how a
promise viewpoint points towards the origins of associations as a subject of the promises made by these agents.
5.4.1
Semantic addressing
How one enters or stimulates activity in a knowledge network at a suitable location, for thinking about a particular
concept, is a problem where context and association must overlap. Knowing both the symbols and the exact
coordinate address of every concept in memory would be potentially cumbersome. Indices and directories can
be used to short-cut the linking of semantics to locations, but there is also the possibility that memory could be
addressed by its own tokenized content. This is the idea behind semantic addressing, as used in databases and
other key-value stores. The underlying mechanisms of databases make use of a indices and directories to simulate
semantic addressing; however, other technologies, based on hashing, or functional symbolic discrimination, do not
need to use the indirections of look-up tables.
5.4.2
Concurrently active concepts
Semantic addressing is related to conceptual association. An association is a pairing of concepts or data, such
that one is a token representation for the other, such as in a key-value pair. It involves a correlation of some kind.
From promise theory axioms, we know that a promise of correlation can only be made locally by a single agent,
as an independent arbiter [6]. This can take several forms (see figure 30) and explanations below. Associations are
O2
+k
O2
+k
+v
+v
R
S
R
O1
R
O1
(a)
(b)
(c)
O2
L
R
O1
R
S
(d)
(e)
Figure 30: The tensor forms of correlation, inferable as associations (see definitions in text).
semantic links between concepts. They originate from clusters of promises.
62
Definition 65 (Association) An association A is a labelled map between concepts. A simple diadic vector association C1 has a relationship with C2 involves a link between two concepts, and a third concept which is name of
link type/role itself.
A : concept × concept → concept.
(99)
We use the lower-half-arrow notation represent a directed association:
A
named association
C −−−−−−−−−+ C 0 ≡ C −→ C 0 .
(100)
Unlike a simple promise, an association a always has a conjugate or reciprocal interpretation a. An association
implies the existence of a directed path between C and C 0 .
An association may be the subject of a cooperative promise between agents.
Lemma 26 (Unlabelled concept-association graphs are non-directed) The assumption of auto-reciprocal associations means that each association provides bidirectional labelled connectivity in the knowledge graph with a
conjugate label.
The proof follows from the assumption that each association has a mandatory conjugate or reciprocal meaning,
and the existence of a path between the concepts. This means there is a bidirectional association between any two
concepts, but with conjugate labels.
0 a
+C
−adj
C −
C −−−→ C 0
a
+adj
0
=⇒
.
(101)
C−
+ C 0 =⇒
C −−−→ C
+adj
0 −adj
C 0 −−−→ C
C −−−→ C
Hence there is a non-directed adjacency between C and C 0 . The relationship between associations and promises
is complicated. An association is a promise of virtual adjacency between concepts, which are superagent cluster;
however, the meanings of the concepts and associations depend on the promises of the entire knowledge system.
I’ll leave it as an exercise to the reader to imagine this this entire construction (see figure 30).
Example 83 (Associative notation)
inside
T −−−+ T 0 ,
(102)
meaning T is associated with T 0 , with the interpretation T ‘is inside’ T 0 .
Definition 66 (Association structure) A collection of assessments, which may be interpreted by an observer R
as a relative adjacency between two things or concepts, taking the form of a total function A : K → V , (written
A
k −+ v). It maps keys K to values V , and is defined over some interior representation belonging to R.
(a) Auto-calibrated association. Scalar association by coincident promiser. A source agent S promises both a
key +k and a value +v to an unspecified agent. Agent R is in scope of these promises and assesses them
simultaneously:
+k
S
−−→
A?
(103)
S
−−→
A?
(104)
A:k −
+v
−−−−−−→ A?
(105)
R
R
+v
R
Example 84 (of definition 66 (a)) These are dual or multiple roles, i.e. scalar promises that are associated
by virtue of originating from the same agent: e.g. linguistic compounds like police-woman, bank-teller,
food-truck, leather-jacket, table-leg, car-factory, apartment-building, living-room etc. These are not merely
linguistic compounds but conceptual associations too.
(b) Observed correlation. Bi-scalar association, with key and value observed in separate agents.
+v
−−→
O1
+k
−−→
O2
−v
−−→
R
R
R
−k
−−→
+ v|k,v
−
−−−−−−−−→
A:k
63
R
(106)
R
(107)
O1
(108)
O2
(109)
A?
(110)
R assesses that k and v are simultaneously kept from different dependent sources and associates them with
the name A.
Example 85 (of definition 66 (b)) The observation of correlations between qualities that have not been
promised. Agents are authors, or police, seeing shapes in clouds, or faces on cars, i.e. the association
by the observer, without reference to the outer reality.
(c) Reported association. Vector association, inferred by scalar promise from separate agents.
+A
−−→
O1
O2
(111)
O1
(112)
R
(113)
R
(114)
O1
(115)
O2
(116)
A:O1 −
+ O2 |(O1 −−→O2 ),(O2 −−→O1 )
−−−−−−−−−−−−−−−−−−−−−−−−−→ A?
(117)
−A
−−→
O2
+A
+(O1 −−→O2 )
−−−−−−−−−→
O1
−A
+(O2 −−→O1 )
−−−−−−−−−→
O2
+A
−(O1 −−→O2 )
−−−−−−−−−→
R
−A
−(O2 −−→O1 )
−−−−−−−−−→
R
+A
R
−A
R assesses the promises made by O1 and O2 as simultaneously kept, and infers the association between
these two agents under the name A.
Example 86 (of definition 66 (c)) Observation of a connection, which the agents both promise to have, may
be based on adjacency, causation, generalization, etc: e.g. a person on bike, device plugged in, Mark’s
dog, A supplies B, A depends on B, A and B are partners, the agents ‘know’ each other in the real world,
the agents agree about subject X.. This is a formation of cluster of collaborative molecular roles from
independent atomic roles.
(d) Reported calibration. Bi-vector association through intermediate agent. Generalization of (c).
+b
1
−−→
O1
L
(118)
L
(119)
R
(120)
R
(121)
O1
(122)
O2
(123)
A:O1 −
+ O2 |(O1 −−→O2 ),(O2 −−→O1 )
−−−−−−−−−−−−−−−−−−−−−−−−−→ A?
(124)
−b2
−−→
O2
b1
+(O1 −
→L)
−−−−−−−→
O1
b2
+(O2 −
→L)
−−−−−−−→
O2
b1
−(O1 −
→L)
−−−−−−−→
R
b2
−(O2 −
→L)
−−−−−−−→
R
+A
L
R
−A
Example 87 (of definition 66 (d)) The calibration of roles, and innovative mixing. These are promises that
unite agents by a common (matroid) basis element: partners brought together by an interloper or matchmaker, the members of a club, agents authorized to act (e.g. keepers of the password), members of a country,
all agents facing North.
(e) Reported encapsulation. Scaled superagent informs as a function via index/directory lookup table. Generalization of (a).
asserted
S
+(k −−−−+ v)
−−−−−−−−−−→
R
(125)
S
(126)
A?
(127)
asserted
R
−(k −−−−+ v)
−−−−−−−−−−→
A
R
k−
+v
−−−−−→
R parrots the trusted association promised by superagent S, about its interior state.
64
Example 88 (of definition 66 (e)) An association promised by a single encapsulating agent: e,g. I am a
shopping mall, a city, a community, and these agents are part of my makeup (here is is my internal directory).
The type of exterior promise involved in the simultaneous assessment by R’ need not be mentioned, as it plays no
real role, as long as the promise type matches across the concurrent assessment. Also, as usual, it is understood
that the promises, on which R bases its assessments, may not always be kept. Associations are, consequently, at
least as uncertain as the promises that lead to them.
In terms of physical representation, spacetime proximity (e.g. adjacency, such as in a cellular automaton or
an ionic chemical bond) is a more robust or less uncertain promise of association than say remote correlation
by message exchange (e.g. pen pals or a covalent chemical bond), because it does not rely on the promises of
intermediate agents to glue together the association. In either case, the associative bindings are reinforced by
frequent revisitation, i.e. learned familiarity, increasing trust [5].
5.4.3
Tensor structure of association
Associations are constraints on the spacetime structure of a semantic space. Associations are made of promises, but
not vice versa. As a constraint, an association involves spacetime symmetry reduction, such as a regular ordering
or clustering around anchor points. Concepts-association graphs are a subset of promise graphs.
Lemma 27 (Associations are tensors) All tensor forms may be constructed by involving clusters of N agents.
Referring to paper I, the proof follows trivially; the rank of the tensor is less than or equal to the number of agents
involved in an association. In a mathematical tensor, each collection of indices represents a subjective agent viewpoint, which could potentially be transformed into another agent’s perspective, by redefining agents. This is known
as covariance [55]. It is intriguing to wonder if the spacetime properties of memory allow covariance of tensor
representations to play a role in cognitive perspective too. Context may play the role of such a transformation, as
it alters the clustering structures of agents in a knowledge representation, and the promised associations made by
the superagent clusters.
Example 89 (Associative arrays) In a computer, associative structures form directories or index maps (see paper
II). In order to link concepts semantically, there has to be an underlying dynamical addressing of spacetime, which
acts as intermediary. Transforming the addresses, in a way that respects spacetime structure, leads to a different
perspective.
Example 90 (Superagent directory lookup table) A superagent directory is an associative array in which keys
are mapped to promises. Agent directories map interior agent names/addresses to their promises.
Example 91 (Entity-Relation models) In Entity Relation (SQL) modelling, the graphical structure is built from
lookup tables, and addresses are hidden from users, replaced with primary keys. Thus spacetime locations may
be found only by searching, or using an index to map primary keys (see figure 31). Underlying this user interface
legerdemain is a ordinal addressing scheme for referring to the locations: just as a book index maps from keywords
to page numbers, the ordinal page numbering is complicit in making the scheme predictable and efficient. An index
exploits the solid state structure of a space make location predictable. There are then all kinds of structures with
properties that optimize the time to retrieve the item.
Example 92 (Lookup tables) In data science, associative arrays are simple data maps, mapping a domain into
a codomain. As associative data array is a simple mapping between sets. Consider a set of three colour concept
classes: C = {red, green, blue} and a set of data values D = {000000, . . . , F F F F F F } corresponding to these
classes. We may form a mapping M between these concepts and ranges of discrete colour codes in RGB space.
A[red]
A[green]
A[blue]
5.4.4
= {010000, . . . , F F 0000}
(128)
{000100, . . . , 00F F 00}
(129)
= {000001, . . . , 0000F F }
(130)
=
Associations may connect any agent or scale
An association is not limited to a particular kind of agent, as illustrated by these examples.
Example 93 (Functions) Functions association inputs with outputs, e.g. looking up RGB codes for colours:
Colour(c) ↔ Codes(d)
(131)
Or as an associative functional map function A, with c ∈ C and d ∈ D:
c = A(d).
65
(132)
Timeseries
Alert type
link
(12,34,6)
ALARM1
ALARM2
23:56:01 Monday
Figure 31: Entity relation models in databases allow different types of data to be standardized as tabular templates and then
linked together through data intermediaries forming a map, as long as each instance of a table they has a primary index, i.e. a
coordinate system that can be indexed against.
Example 94 (Associating different schemas) What if we want to keep the different memory types distinct, in
special kinds of memory, and still weave them together into an associative web? i.e. we don’t try to transduce them
all into words, or data values, but keep timeseries, words, image fragments, etc all separate but Entity Relation
theory introduced for database modelling attempted to answer this question [35].
5.4.5
How are associations identified?
Consider how two concepts can become associated, in a memory representation. Referring to equations (90) and
(92), we may add a single promise pair from S to itself, which loops back, to form the association. By continuously
keeping and assessing this promise, the agent can learn the association, as a stable representation in its knowledge
space. An association, intended by the sensor observer S is a promised scalar relationship about two sub-agents,
inside its horizon:
some association
±C −−−−−−−−+ C2
S −−−1−−−−−−−−−−−−→
S
(133)
If the association builds on an actual observation of objects O1 and O2 , then that relationship would be communicated by the promises of e1 , e2 , and then one could consider the association to be a conditional promise:
some association
±C1 −−−−−−−−+ C2 |e1 ,e2
S −−−−−−−−−−−−−−−−−−−→ S
(134)
If the association is based on reasoning by the observer then
some association
±C1 −−−−−−−−+ C2 |e1 ,e2
Observer −−−−−−−−−−−−−−−−−−−→ Observer.
(135)
To associate concepts, we must have simultaneous triggering of concept clusters. No other information exists to
join two concepts together. Since the concepts themselves are bootstrapped by the equation (134)
Lemma 28 (Simultaneous coincident context) The semantics of association are those of simultaneous activation
of context.
From equations (134) and (135), there are two possible forms of concurrent activation, which are the only concurrent events that contain both concepts as information. Thus the only causal possible relationships are these
two routes. Since exact simultaneity is impossible, coarse graining of simultaneity is an essential requirement to
coarse graining together concepts into associated clusters. This means there is a timescale for association formation
Tassociation ≥ Tinterior .
66
5.4.6
Flooding, routing, and association
It is worth remarking that, in order to form concepts by association, concepts may be addressed from inputs according to either the flooding or routing addressing methods. Flooding (even within a restricted region) is a divergent
process, which is therefore expansive and leads to possible generalization of concepts. Routing is convergent and
specific, allowing precision. These complementary ‘advanced’ and ‘retarded’ propagation mechanisms are at work
in spacetime networks at all levels.
5.4.7
Autonomy and association (a notational caution)
Consider figure 32. If we represent composition, within a knowledge space, as a promise (case (a)), a literal reading
of the notation
+b
S −−→ R;
(136)
appears to claim that the associative label comes between the two concepts it links (case (b)). If this were the case,
the intermediate agent would need complex semantics to observe both S and R and act as a switch to relay the
intention between the two rather than impeding them. As an intermediary, it would be more likely to impede or
distort transmission than represent the other’s intent by proxy [6]. Thus a connection would require the creation of
a new agent and a conditional promise to activate the connection. Associations promised by autonomous agents,
like cells or dumb devices, do not have the internal capabilities to keep advanced promises. If b is realized as an
+b
S
R
S
+b
(a)
R
(b)
+b
S
R
(c)
Figure 32: A promise of a body +b (a) can impose signal agents, or route the third party without it being emitted/absorbed.
It is not correct to imagine the body provider as an intermediary (b), rather it is a source that must be relayed by the agent who
makes the promise (c), because b cannot promise on S’s behalf.
intermediate agent, then it blocks S’s direct promise, rather than enabling it, and makes no guarantee of a promise
of its own to R. This is both a complication, expensive, and uncertain. A simpler representation, which observes
the principle of autonomy [6], is shown in case (c), in which b serves as the source of the promise body, and S
promises to route it directly to R. In this form, the source of the intention acts as a basis (matroid) agent, which
may be constant, and S has a fixed promise for all b rather than requiring a specialized promise for every b. Thus,
this approach is economically more scalable. All the dynamics are directly between the concept agents. By the
conditional promise law:
+data
b
−−−→
S
+data
−−−→
S
−data
b
S
−
−
−
→
b
+b
+data(b)
(137)
≡>
S −−→ R
≡>
−−−→ R
+data|data(b)
S −−−data
S −
−−−−−−→ R
R −−−→
S
−data
R
−−−→
S
Theorem 1 (No go for bipartite lattice) A programmable promise graph can not be implemented as a bipartite
y
+b
x
graph, by proxy (see figure 32(b)), i.e. S −−→ R cannot be represented in the form S −
→B −
→ R, for b, for any
x, y.
From the discussion above, the proof follows from the fact that a bipartite lattice would not propagate the intended
promises without additional promises to switch on the intermediate node, leading to a recursive deficit in switching.
This suggests that a bipartite lattice is not a viable way to implement associations. In order to make a proxy
representation, we would need
S
B
R
+x
−−→
+b
−−→
−x
−−→
67
R, B
(138)
B, S, R
(139)
S, B,
(140)
where x is to be determined. In order to transmit intent +b to intermediate agent B, +x would need to be of the
form
E
B
B
S
def (b)
−−−−→
−def (b)
−−−−−→
B
(141)
S, R
(142)
+b|def (b)
−−−−−−→ R
?
−
→
(143)
R
(144)
for some external control source agent E. The latter is needed to promise intent from S to R, but there is no ?
that can promise the intent to follow B from S. If S = E, there would need to be a direct channel from S to R,
which violates the bipartite structure. Section 5.4.11 illustrates a workaround for this, however, using intermediate
switching agents for more complex linkage; the result is not a direct association, but an implicit association by
reasoning. The routing construction does not violate the principles of autonomy, and can lead to a simple direct
association:
Lemma 29 (Promises may be routed) A promise graph in space can implemented as a conditional routing switch
(figure 32(c)).
This follows from the conditional promise law [6], in which a promise from the basis agent for the association
label acts as the switch.
B
S
S
5.4.8
+def (b)
−−−−−→
+b|def (b)
−−−−−−→
−def (b)
−−−−−→
S
(145)
R
(146)
B, R.
(147)
Quasi-transitive or propagating associations
Intent propagates between agents if there is what we might call a quasi-transitivity as a result of promises made
and kept [79]. In a knowledge network, this may translate into concept associations which propagate if there are
associations that have certain spacetime transformation properties. Associations can correlate causal influence over
two dimensions (scale and location). Let C (n) be conceptual (super)agents at an aggregate scale, labelled by n:
• Vertical scaling associations (North-South) which aggregate or decompose concepts by membership, i.e.
scale resolution:
(n+1) +generalizes
Ci
(n)
−−−−−−−−+ Cj
(148)
These associations break scaling symmetry by pinning concepts (see below), or selecting a direction for
scale.
is bigger than
CMy dad −−−−−−−+ Cyour dad .
(149)
• Horizontal associations (East-West): within each superagent (see figure 33):
similarity
Ci −−−−−+ Cj
(150)
or between superagents, e.g.
depends on
Si −−−−−−+ Sj .
(151)
• Vector associations that may result in a quasi-transitive inference, i.e suggest a propagation of intent over a
potentially long range. These are related to spacetime properties (see table 1).
• Vector associations that result in no transitive inference, i.e. suggest short range finite-dimension transitions
rather than spatially extensive properties. These are related to interior state machines (like hidden variables).
68
superagents
N
W
E
subagents
S
Figure 33: The common use of compass directions to refer to a scaling hierarchy of increasing generalization, i.e. increased
coarse graining, while other associative properties happen in the planar scaling layers. This language of N,S,E,W is common
in the networking industry. All the irreducible associations refer to constraints on the scaling and the topology of the space of
agents.
5.4.9
The irreducible positive associations
The irreducible quasi-transitive structural associations are documented in table 1. The generally transitive or
propagating properties fall into the familiar irreducible classes (see table 1). Let A(n) be agents in either a real or
a knowledge space, which may include C (n) :
1. Proximity/Adjacency: (space) mutual, classically non-directed relationships represent non-directional role
symmetries that have not been broken. e.g. overlaps with, shares with, associates with, joins, connects with,
relates to, etc.
(n)
Asimilarity : Ai
(n)
For every Asimilarity : Ai
(n)
→ Aj .
(n)
(n)
→ Aj , there is also a Asimilarity : Aj
(152)
(n)
→ Ai .
Lemma 30 (Similarity associations are symmetry preserving) The class of similarity associations represents symmetry preserving transformations, mapping one concept into an equivalent one that may be
reversed by inverting the direction
s
s
C −
+ C2
C −
+ C1
s
s
If s is a similarity association C1 −
+ C2 , then C2 −
+ C1 , because A −−1−−−−→
A =⇒ A −−2−−−−→
A.
2. Sequence/Dependency: (time) associations that describe causal pathways in a system, whereby one agent
depends on the prior keeping of a promise by another, or builds on the existence of a prior concept. e.g. depends on, reinforces, contributes to, asserts, documents, annotates, expresses, elaborates, extends, involves.
(n)
Acausal : Ai
(n)
→ Aj .
(153)
Lemma 31 (Causal associations break equivalence symmetry) Causal associations break the directional
symmetry of mutual semantics.
Selecting a precedence between two concepts imposes a local direction on the map, thus selecting a preferred
direction or concept, and breaking the ‘reflection’ symmetry or equilibrium between concepts.
Causal association may be chained together to represent world lines, because any change in an association
measures time, and selects a preferred direction.
3. Scaling/containment: (boundary enclosure) describes a transformation from one scale to another, in which
multiple concepts at the smaller scale contribute to the larger scale. e.g. the relationship between a collective
agent and a member agent: generalizes, exemplifies, abstracts, models, etc.
Let (super)agents at scale n be represented by A(n) (see paper II), then
Acoarse : A(n) → A(n+1) .
69
(154)
ST T YPE
1
2
3
4
F ORWARD
is close to
approximates
is connected to
is adjacent to
is correlated with
F ORWARD
depends on
is caused by
follows
F ORWARD
contains
surrounds
generalizes
F ORWARD
has name or value
characterizes
represents/expresses
promises
R ECIPROCAL
is close to
is equivalent to
is connected to
is adjacent to
is correlated with
R ECIPROCAL
enables
causes
precedes
R ECIPROCAL
is a part of / occupies
inside
is an aspect of / exemplifies
R ECIPROCAL
is the value of property
is a property of
is represented/expressed by
S PACETIME STRUCTURE
PROXIMITY
“near”
Symmetrizer
S PACETIME STRUCTURE
GRADIENT/DIRECTION
“follows”
S PACETIME STRUCTURE
AGGREGATE / MEMBERSHIP
“contains”
S PACETIME STRUCTURE
DISTINGUISHABILITY
“expresses”
Asymmetrizer
Table 1: The four irreducible association types are characterized by their spacetime coincidence or adjacency. Note that, in
promise theory, even relationships like ‘is correlated with’ are directed relationships: one may not assume that the assessment
of a mutual property is mutually assessed. Similarly the expression of a property is a cooperative relationship.
Lemma 32 (Coarse grain associations break universal scaling) Coarse graining associations pin semantics at a particular semantic scale, breaking renormalization scaling symmetry.
By naming a specific agent or superagent, e.g.
generalizes
Canimal −−−−−−+ Cdog .
(155)
we break a generic scale-free relationship.
The formation of superstructure by associative clustering leads to generalizations and new types. Generalization of instances into conceptual hubs allows self-organized model learning, and increases the scope for
fuzzy matching.
Lemma 33 (Smoothing) If we take a vector field and coarse grain the space so that the points defining the
vector are effectively coincident at the new scale, then we average away the gradients. This is the process
of smoothing and noise reduction. It transforms similarity states into a single thing with representative
semantics, i.e. a common concept.
4. Promise/Expression: (property) if we associate agents, while retaining the special names and characters, it
is like the formation of molecules. This breaks the symmetry of roles between the component agents in a
cooperative association, e.g. represents, is a name for, is the doorway to, etc.
(n)
Arepresentation : Ai
(m)
→ Aj
.
(156)
Lemma 34 (Representation breaks role symmetry) Representational associations break the symmetry of
mutual semantics, selecting one part to act as a proxy labelling the other.
Lemma 35 (Transitivity no go) Transitivity of associations cannot be promised for arbitrary associations.
The proof follows from the basic axiom that no agent may make a promise on behalf of another. Concepts are
(super)agents, and no agent can promise a property on behalf of another, i.e. that an association will be enforced
beyond itself.
70
Example 95 (Transitivity is not assumed) If A promises it is connected to B, and B that it is connected to C,
we simply don’t know whether A is connected to C, either directly or through B. it might be true that there is a
connected path from A to C. This is not a certainty, because B might not allow transmission of anything between
the two. The fact that agents are chained together in relationships is no guarantee of their.
Example 96 (Vertical association limits (hierarchy of generalization)) Vertical associations, which take us from
exemplary THINGs to a generalized class or concept, could be stacked to any depth, in principle. In practice, however, human faculty for stacking concepts is quite limited: we can typically keep only a handful of ideas in mind
at any one time [3, 96, 97]. Moreover, in any practical system, the apparatus by which associations are made and
classified is finite, and will have to be understood by a human with limited faculties, thus it is not implausible
that these vertical links would be finite in number, and perhaps specialized with distinct semantics that reflect the
transformation staging in learning process (recall the stages in the parsing example 38). For example, a limited
set of stages could lead to the following interpretations:
1. Species in a taxonomy.
2. Things in a taxonomy.
3. Concept hierarchy.
4. Sensor data hierarchy.
Going up the list corresponds to ‘is generalized by’, and going down the list corresponds to either ‘is exemplified
by’ or ‘is composed of’38 .
5.4.10
Non-transitive semantics in associations
Not all associations make even quasi-transitive sense. Consider human relationships, like father, daughter, etc. If
we write
is father to
is father to
Fabian −−−−−+ Simon −−−−−+ Dawn,
(157)
it is by no means true that Fabian is father to Dawn. The father relationship is pinned at a specific short range: it
cannot form a state of long range order. It is, in fact, is a short-range form of association, and a causal precedence
association. Interestingly, it does not propagate, because the concept of father is limited. This now has the direct
implication of that we can state as a general law: that loss of long range symmetry is directly associated with
knowledge encoding, as stated formally in the final conclusions.
Simon
is
fa
er
th
to
precedes
son
father
Fabian
represents
Figure 34: Constrained association by something like vector addition of irreducible types.
Imbuing relationships with labels that constrain the semantics locally also constrains the propagation of intent.
This is significant: it means that a plain graph theory cannot represent hierarchical semantics, and we need to
understand specifically how concepts, embedded in the associations, further constrain the spacelike clustering of
38 It is interesting to note that the brain cortex has neuron column structures that can form hierarchy up to depth 6, and to speculate about
whether this is related [19].
71
concepts. If we decompose ‘is father to’ into a causal/precedence association and a representation association (see
figure 34, and recall section 5.4.7):
represents
precedes
father of
X −−−−+ Y = father )−−−−− X −−−−+ Y
(158)
represents
father of
precedes
−−−−+ ' )−−−−− ⊕ −−−−+ .
(159)
What this shows is that our understanding of concepts is more complicated than our understanding of spacetime; it involves higher tensor forms. If the father concept were just any old agent, and ‘is father to’ were just a
generic ‘source of’, then the sequence
is source of
is source of
Fabian −−−−−−+ Simon −−−−−−+ Dawn,
(160)
now makes quasi-transitive sense. Why the father concept does not propagate does not have an obvious spacetime
explanation, but may be addressed from the viewpoint of reasoning (see section 6). Semantics are sometimes
pinned to particular spacetime scales, and cannot be generalized beyond them. Thus, the emergence of semantics
may be associated with spacetime symmetry breaking.
5.4.11
Embellishing associations for deeper inference
To represent higher tensor forms, where associations are more complex than a simple link between two concepts,
we may consider embellished associations. These are constructed by introducing an intermediate concept and
treating the association as an inference (see section 6). Although the direction of a narrative still follows a direction
from a promiser concept to a promisee, other concepts may embellish the descriptions of the associations, as labels.
Thus vectors associations generalize to tensors, with multiple concepts clustered into the association itself.
Example 97 (Rolling context) Context plays a key role in the iterative association of concepts (see figure 35):
murder
library
cause | context
Professor plum
Mrs Peacock
Figure 35: Embellished association is a general tensor.
From-Concept
Professor plum
Association
murders with a knife
Context
in the library
To-Concept
Mrs Peacock
In the next iteration, the context would accumulate the foregoing concepts until they age beyond relevance:
From-Concept
Police
Association
arrive
Context
the murder of professor plumb
in the library by Mrs Peacock.
To-Concept
in the library
What kind of spacetime structure accounts for this graph (see figure 35)? It is a tensor of rank less than or
equal to four:
Ac1 c3 c4 c2
(161)
The only natural object with this structure, which does would not violate the ‘no go’ theorem for bi-partite
causaton, is the context aggregator itself, which would build the embellishments as context. This is not yet well
understood, at the time of writing, but should become clearer with some model building39 . The conditional activation decided by the hub could be based on a summation threshold value or it could be a binary logic. In a computer
system, binary decisions are assumed. In biochemical processes, they seem unlikely40 .
39 Experiments
are underway in the ‘percolibrium’ project.
we cannot extended the number of these simultaneous connections indefinitely, it suggests that there is a limit to how many simultaneous concepts can be processed in one go, unless they are parallelized, meaning that a Turing machine trying to simulate the knowledge
process would take longer to understand complex associations than simple ones, proportional to the number of concepts involved.
40 Since
72
murder
Professor plum
library
?
? ?
Mrs Peacock
Figure 36: Embellished association is an intermediate routing hub or switch that allows simultaneous activation of the concepts. The hub plays a key role in knowing such complex associations.
5.5
Context as a conditional switch for associative memory
A knowledge system needs to be able to map associations to contexts, in order to differentiate and nuance usage, so
a semantic space must be able to handle context-specific adjacencies, both to distinguish and coordinatize multiple
meanings. To realize this, associative memory must be switchable. A switch is a device that exploits boundary
conditions in order to link or unlink spatial regions. We can imagine context as acting analogously to the bias
current of a transistor:
Collector
Concept
+current|bias
−−−−−−−−→
Emitter
+related|context
−−−−−−−−−−→ Exemplar
(162)
(163)
Context thus becomes a modulator for permanent conceptual and associative memory, in accordance with the
conditional assistance law [6]. From (163), and the rules of promise theory, we see that the promise to deliver
context information must be kept by a distinct agent in order to assist in the keeping of a conditional promise. As
a model for this, we return to figure 25, consider how this makes sense41 .
5.5.1
Conditional association and structure of context
Conditionality breaks translation symmetry explicitly, by introducing boundary conditions which modulate the
concept space creating inhomogeneity, and it is associated with switching; thus, now we see a direct link between
switching, memory, and knowledge representation at a fundamental spacetime level. Consider associations of the
form:
association|conditional context
C −−−−−−−−−−−−−−−−+ C 0 .
(164)
The associational link is now considered only to be a link if the condition is part of the context of the observer.
Example 98 In tabular form:
From-Concept
Hot
Hot
Hot
Association
characterizes
characterizes
characterizes
To-Concept
Spain
Chilli
Sarah
Context
weather
food
sexual attraction
We see that concepts may have many interpretations that are labelled by context. In ontology, e.g. in topic maps,
associations and concepts are classified into types; this approach was criticized, as it leads to static, impositional
data models that easily form contradictions to invalidate themselves [14, 92]. An approach using conditionals,
where the current context belongs to a set of token characteristics, addresses the deficiencies of a type system,
though it leads to more complex graphs42 .
5.5.2
Constructing embellished or contextualized associations
The need to represent the kind of contextualized associations as described in the previous section, has structural
implications for concepts. If we assume the existence of aggregating agents, such as the one shown in figure 37,
then we see that the inputs to an agent from different contextual triggers begin to dominate the links to the concept
41 Any superagent capable of assembling a working context can use it as a coded memory address in a routing fabric analogous to a Banyan
switch [98].
42 The ideas is similar to the demotion of static schemas in database technology and their replacement with schema-less data formats.
73
(n+1)
A
S1
(−c)
S2
(+ C)
(−c)
S3
(−c)
S4
(−c)
M
R1
θ
A (n)
Figure 37: An aggregating agent, whose promise to R1 depends on the promises from a number of source agents is an
ambiguous semantic situation. An aggregation of equivalent sources is essentially a membership relationship in which the
agent M plays the same role as a generalization, mapping from level A(n) → A(n) or from level A(n) → A(n+1) .
superagent, shown by the dotted line in figure 38. For each outgoing association from the single concept to other
concepts, conditionalized by context, there must be an incoming bundle of promises to accept context inputs. This
suggests a potentially large asymmetry between inputs and outputs that places a burden on a single agent. A
hypotheical resolution would be to build a concept bundle from interior subagents, say one for each association.
A concept with, say, two outgoing associations (shown by the dotted line in figure 38), would then have two such
basic aggregating elements. This generalizes to more. This, in turn, suggests that the concept forms as a cluster of
several such aggregating switches, linked through a binding basis agent (see figure 38), which demarks the single
concept as a discrete entity.
For each association, promised by the concept agent, there are likely to be several context triggers, arising from
different stages of the sensory inputs chain. Given that agents have limited internal capabilities, there is also likely
to be a limit to the number of contextual inputs that can condition an associative memory link.
Concept superagent
member2
Concept
binding element
member1
Outgoing contextualized associations
Incoming context
Figure 38: A concept forms as a cluster of several aggregating switches, joined through a binding basis agent.
5.5.3
Causation and aggregation are not completely orthogonal
The irreducible associations are not without some overlap or ambiguity. Consider the case of an aggregating agent,
like the one discussed in figure 37. This is a structure that is key to representing conditional promises, and thus a
contextualized associations of the form:
M
M
+C|c1 ,c2 ,c3 ,c4
−−−−−−−−−→
−c1 ,c2 ,c3 ,c4
−−−−−−−−→
74
R1
(165)
R1
(166)
The triggering context −{c1 , c2 , c3 , c4 } is aggregated by M in order to activate the promise +C of an association between M and R1 . Another interpretation might place M in the role of a generalization of the concepts
S1 , S2 , S3 , S4 . The ambiguity arises because a generalizing aggregate concept may be considered to be caused
(in some sense) by the members that belong to it. However, aggregation usually presupposes that the Si are
symmetrical in their semantics.
• (Symmetry of contexts) Each instance leads to (has a causal influence on) the existence of a generalization.
• (Asymmetry of contexts) A dependency or contributing ingredient is not generalized by the combination of
contributors.
Thus causation is a more primitive association, in some sense. The main difference is that generalization requires
the ingredients to be symmetrical in their semantics. The members of a group need some unifying property (such
as a basis set or matroid element). In practice, a generalization is a membership in a superagent M whose representative basis is the superagent itself. For source agents that are symmetrical in their semantics, one could imagine
a transformation from one type of association to another, formed as a superposition of semantics, represented by
the angle θ in figure 37. This transformation does not invalidate the irreducibility of the coarse graining association. Indeed, it is natural that causal influence would be able to traverse concepts belonging to different levels of
aggregation. Introspective association would tend to level the hierarchy of concepts.
Finally, it is worth noting that the law of dual or complementary promises [6] indicates how ± can be exchanged, turning what is context into what is concept, or vice versa. Thus, in a reasoning situation, the roles of
context and representative concepts can easily be thrown into reverse.
5.5.4
Information argument for separation of context
To be more precise about the multiple stages of conceptualization, and their timescales, from quick the identification context tokens to long-term associative clusters (in different kinds of memory representation), we consider an
argument based on intrinsic information, as measured by the Shannon entropy. The Shannon entropy is a measure
of information efficiency [9, 64]. Both short term context and long term memory contain information, in the form
of internal links and concepts. We don’t need to now specifically how information is encoded, but we can assume
that it is quantifiable, through a symbolic alphabet, as an informational entropy.
Definition 67 (Shannon entropy of a memory configuration) Let A be some associative configuration, formed
from an alphabet of links between concepts, in a graph Aij of all possible concept agents C and superagents S,
where i = 1, . . . , C + S, and let p(A) be the probability of links being active. We may define the information of
this configuration to be:
C+S
X
S(A) = −
pij log pij .
(167)
i<j=1
This intrinsic information may be used to compare the amount of information in short-term context memory with
the amount of information in long term memory.43 An agent or collection of agents responsible for context classification takes observations about the state of the world, at each clock tick of the knowledge system, and labels a
set of states, which are used to predicate new associations between pairs of concepts. Let’s define these collections
of agents using superagents at a large scale:
Definition 68 (Adaptive knowledge system components)
S
= Sensor input agent graph
(168)
C
= Context agent graph
(169)
A
= Knowledge/Concept association agent graph (sum of all persistent memories)
(170)
We can use information to make a strong statement:
43 Physicists and computer scientists define information differently, in a subtle way. Information in physics is about uniqueness of an
experiment within an ensemble of experiments, so it measures the uniformity of statistical patterns across many episodes (external time).
Information in Shannon’s theory of communication is about uniformity of states within a single episode or sample, thus it measures the
uniformity of patterns within a single experiment (over internal time). This is why information scientists would say that information is entropy,
and why physicists say that is it the opposite of entropy!
75
Hypothesis 3 (The information resolution of context must be less than that for knowledge) A context-switched
knowledge system, capable of distinguishing non-trivial contexts, with introspective feedback, divides into two independent agencies C and A, representing short-term context and long-term associative memory, where S(C)
S(A) and the timescales for acquisition T (C) T (A). This is necessary to have differentiated classes of associated knowledge determined by context.
The proof of this lies in the following considerations:
• Suppose that the context memory were the same agency as the long term memory plus the new sensory
input, i.e. C = A. The feedback from A onto itself is now indistinct from the ordinary associations, and thus
the use of total knowledge as a switch for knowledge context is an automorphism with no class distinctions.
This memory must be homogeneous, and therefore does not allow for persistent context distinctions. In
other words, it cannot remember distinct associations as conceptual entities, because C = A. This means we
can rule out C = A in a contextual associative system, and there must be (at least) two agencies. A further
consequence of this case is that the timescale for context evaluation must be the same as the timescale for
knowledge acquisition and retrieval. Since long term knowledge requires several iterations of sampling over
a context to stabilize, with introspection involving both C and A, nT (C) < T (A), for some number of
iterations n > 1.
• Next, suppose instead that S(C) 6= S(A). The average amount of independent information in each context is
then of the order S(A)/S(C) in each context, allowing contexts to be inhomogeneously encoded scenarios
within the total amount of long term knowledge. In order to contain a non-trivial amount of context specific
knowledge, we therefore need S(C) S(A), implying that whatever agency keeps context, its capacity for
concept resolution must be low compared to long term memory. If this condition were not met, the contexts
would merge into one another, and lose distinction, eliminating smart behaviour44 . With a lower information
content, context can be addressed and operated at a separate and faster timescale than long term memory
processes, leading to fast short term memory, and slower learning of knowledge T (C) T (A).
Corollary 1 (Short term context is faster than long term association) If C is a distinct agency that is a conditional prerequisite for switching A, then the timescale T (C) for a transition is C must be less than or equal to the
timescale T (A) for a transition of A’s configuration.
Context is an agent’s assessment of the present, or its ‘situation awareness’; this may include an interpreted ‘emotional state’ for the agent; these structural entities can all be derived from basic spacetime considerations, in a
representation independent way45 .
Example 99 (CFEngine’s model) The CFEngine software [93] makes use of a strongly similar system, defining
classes based on environment discovery, and using hints about what the user is thinking, e.g. -Dclass. This
replaces a more familiar form of reasoning by first order branching logic. Sensory inputs may be extended indefinitely to discriminate very specific states. Thus CC represents a kind of short term memory for what the
operational system of machines and human operators is collectively thinking about, or focused on.
Example 100 (Immunology model) Context is provided by antigen presenting cells. Long term memory is provided by the populations of epitopes on antibodies in lymph nodes and in circulation as memory cells [89, 90].
Example 101 (Genetics and epigenetics) Genetic encoding operates at two levels: there is long term memory
encoded in gene sequences on chromosomes. Then there is an environment of protein networks that activate genes
at specific moments. These encode short term context, within an organism as a learning system.
Example 102 (Do neural networks have fast and slow memory?) In an artificial neural network, layers of neurons are connected into a high density feed-forward network, in which propagation occurs at a single speed. This
fast propagation acts as a context encoding for conditioning the long term memory tokenization. Once trained,
this means that a neural network has only a single timescale. However, when the network is learning, it feeds
sensory data forward as context, to be adjusted by back-propagation over a longer training timescale. This is
the separation of timescales for ANN training. Once the network is trained, it operates purely as an advanced
multi-dimensional sensor at high speed, similar to the way an eye can process visual signals quickly because of
millions of years of evolutionary learning. Similar to the eye, once training has stopped, the ANN cannot learn
anything new, without a separate slow process of adjustment, thus ANNs do not contradict these considerations.
44 This may well be an effect that happens as the memory of a smart observer reaches its full capacity, or a critical percolation threshold
between contexts is reached.
45 There has been some speculation about the nature of brain consciousness as a state of matter, that makes some connection with spacetime
notions too [99, 100]. That work goes far beyond the scope of the present work however. The extent to which the present structural concerns
are represented in brains remains to be seen, though there does not seem to be any obvious contradiction as far as I can tell.
76
The implication of conditional promises in context switching reaffirms the basic separation of learning into
two timescales: a short-term evaluation or classification of the current ‘state of mind’ in a thinking process, and a
long-term linkage between concepts, which can be paged in by the short term modulation. The relative information
also supports the notion that knowledge would converge, by feedback, to a tokenized representation, separate from
detailed sensory input. The question remains, however, as to whether context has to be a completely separate memory, or simply a distinct agent that transforms long term memory into context? From the foregoing considerations,
it seems that C would be a coarse graining of both A and sensory input S?
In an introspective system, a component of context could derive from fed-back long-term associative knowledge, along side the interpretation of direct observations (see figure 39). However, it also seems likely that context
has a role closer to observation than to accumulated knowledge. The question of how knowledge is bootstrapped
provides some direction here. A memory that has no initial concepts still has to be able to generate context according to this architecture. This suggests that context is closer to sensory interpretation than to long term memory.
That hypothesis is further backed up by the argument that contextualization has to happen at approximately the
same timescale as observation; thus, even with feedback from long term concepts, context is the imprint of the
sensory stream onto specialized knowledge. Thus, promise theory makes a simple prediction:
Hypothesis 4 (Separate agent for context adaptation) Context needs a distinct and internally coordinated agent,
of unspecified structure, that promises to quickly summarize the current state of the the external sensory input and
recent history, for use in selecting the relevance of context-dependent relationships between the representations
above. The agent may promise only limited capabilities, meaning that context has to focus on particular parts of
the sensory input in practice.
Because the only mechanism for coupling of concepts is their simultaneous activation, and activation is assumed
to be modulated by context, the prediction is that conceptual memory is built on a hierarchy of timescales, not on
a hierarchy of a priori semantics. In a knowledge space, associational adjacency is conditionality on a context. It
is thus entirely reasonable to expect to find different kinds of memory for different purposes, supporting different
recognition mechanisms and timescales just as we make different kinds of data store in computing.
2
1
3
Long term
memory
4
coarse feedback
Associative
concepts
Context
Short term
memory
inputs
Figure 39: Context discrimination leads to different versions of a concept, according to sensory input and coarse feedback.
Context assessment needs to be fast and disposable, i.e. short term. The formation of associations, or contextual relationships
to other memories and concepts, can be slower, to achieve higher certainty. The integration of experiences into a network of
associated concepts could perhaps be interpreted as a form of quality control. Once concepts have been learnt, over long times,
they too can be used to build more robust contexts, feeding back into the perception of new states.
Example 103 (Scientific method as a learning system) The scientific method is a learning system that acts at
the level of a society.
• S: the sensors are the scientists who study the world.
• C: the context is the papers and the theoretical and experimental results.
• A: the long term knowledge is the cultural memory and literature.
At the scale of a society, the observation horizon is formed by a changing population of scientists, who produce
noisy inputs, because they are constantly in flux, like the disposable cells in a single organism. Since scientists
come and go much more quickly than the literature A, the results could appear unstable. However, there is a
stability mechanism for S arising by social interaction. This leads to norms and cultural inertia, which may drift
slowly over time, but offers a more evolutionary timescale that allows A to stabilize.
77
5.5.5
Context atoms: coarse grained compression to awareness of ‘now’
To remain viable, as a fast switching discriminator, a context classifying agent must assemble from input and long
term memory and compress or tokenize state in ‘real time’46 . The coding of irreducible associations, discussed
in section 5.4.9, enable long term concepts to aggregate hierarchically (though not mutually exclusively) into
superagent generalizations. The names or identities of these generalizations would be suitable compressible coarse
grains for use in context too. In a similar way, input sensors could perform a partial coarse graining automatically
as part of their function. This is how one may dissect deeper specialized knowledge into multiple categories, in
real time, without them becoming so many as to merge into one another, i.e. S(C) S(A), from section 5.5.4.
We know, from spacetime physics, that structural stability relies on the ability to separate scales by assuring weak
coupling or full decoupling between contexts47 .
Samples are themselves signal agents, which encode data, and are associated with the spacetime locations of
the sampling points. Thus, given the assumption that the variety and size of the external environment greatly
exceeds that of the observer, samples need a notion of context in order to support adaptation. They cannot easily
be attributed meaning without the annotation of context.
What might context atoms look like in practice? The table below shows some examples that lend themselves
to specific kinds of sensors:
Coarse grained sensory context atom
day, time, location
light, dark, red, green, blue
hot, cold, free, constrained
salt,sweet, sour, savory, etc
open, closed
blocked, open, snowed under
available, unavailable
parking, parked, shunt, cruising
inside, outside
sensor type
spacetime
sight
touch
taste
bridge
road
shop supplies
transport
sense of self “me”
In spacetime terms, context is associated with coarse boundary information about an observed scenario. Knowledge is accumulated gradually by observation with continuous introspection. This is integrated relative to an
already accumulated notion of ‘self’. Similarly, in the sensory data collection, one always associates the concept
of “here and now” with a sensed context, calibrated by the long-term accumulated experience. The evolving semantics of “here and now”, developing over time, are what leads to higher semantics of associations, i.e. that
which goes beyond ‘concurrent activation’.
It seems likely, given the sheer volume of sensory inputs, and possible feedbacks, that only a partial prioritized
context would be used to encode and retreive memories48 . Thus, one would expect a reasoning engine with limited
processing capacity to ‘miss’ in comprehending certain concepts, if by chance a partial context led reasoning
astray. Concepts associated with coarse grained context may well only converge into coherent clusters over many
averaged experiences and long times. Thus the alphabet of context one could access, at any given moment, might
be much smaller than the actual potential resolution of a sensory input system.
Example 104 (Some human context concepts) The separation of human senses lead us to distinguish spatial
concepts from other variations in the environment. The elements of context may include measures that we do not
consider to be spatial, but which fit easily into a concept of semantic spacetime:
Sense
Sight
Sound
Touch
Smell
Taste
Dimension
time, space, colour.
time, space, pitch, timbre.
time, space, composition, temperature, pressure, humidity.
time, composition.
time, composition.
46 Realtime is self-defining here, as the perception of the observer is only as fast as it can evaluate context. Evolutionarily, it can only adapt
to its environment and be ‘smart’, if its internal clock is at least as fast as its environment.
47 Some contexts might be triggered by all concepts, affecting all parts of memory equally. This might happen when a concept is not clearly
related to an existing concept in the current context. Then context together with feedback from long term concepts cannot narrow a focus
to a specialized area, and thus might tend to map to all or no areas. Such unspecific stimuli might be what animals experience as emotional
sensations.
48 We humans tend to focus one particular sense at a time, when we are reasoning; e.g. we concentate on sounds instead of sight, or touch
instead of looks, taking these one at a time in order to ‘concentrate’.
78
Example 105 A sequence of sampled music can be labelled at a variety of scales. The sampling of a piece of
music is digitized (or packetized) into a sequence of standardized, temporally ordered samples that capture the
dynamics of the sound. The semantics of these samples are that they promise a representation of music according
to an agreed encoding convention that may be used by any promisee to reproduce the intended meaning of the
performance. Each packet sample may be labelled with:
• The time and location of recording.
• The relative times (i.e. order) of the discrete sound samples, from the perspective of the observing agent.
• The data values of the sampled sound representation.
To reproduce the music, we don’t need to know the spacetime location of the original sampling process, but we do
need to know the relative spacetime arrangement of the samples, so as to retransmit them faithfully over the scale
of the total sample, during playback. This suggests a minimal data representation of the form
(ordinal sequence label, data value)
i.e. (time stamp, digitized sound frequency, amplitude)
5.5.6
Spacetime template patterns for efficient recognition and recall
Optimizing the use of space and time to to encode specialized patterns, suggests that certain patterns or ‘fabric’
topologies for memory might assist in the identification and discrimination of sensory forms. Directional patterns
could benefit from directional sensors; visual sensors that measure relative positions would benefit from a coordinate grid representation, etc. This is one more way in which the structural properties of knowledge are affected
by the structural properties of spacetime: they are constrained by them. The earlier discussion of grid cells, place
cells, directional cells in biology are an example of this (see section 4.3.5).
Example 106 (Efficient fabrics for mapping semantic context into long term memory) The Sparse Distributed
Representations (SDR) [19, 61, 71, 72] provide a plausible (even likely) mechanism that could evolve in biological systems. The connection between running timelines and context are best developed in studies of linguistics [20, 101], and the notion of a state of mind in contemporary artificial knowledge systems need not be extrapolated too far to explain the whole of free will and consciousness [10], but it is clear that there are common
characteristics that open an avenue of plausible investigation.
Directed networks, including Bayesian networks and neural networks [17, 44] are a natural area of study, as they
form generic discriminating structures. As branching, rather than converging processes, they may not be efficient,
but they can be generic, and they automatically route connections. Routing structures are an attractive model for
knowledge addressing, because a flooding of a bus would be inefficient and would block the system with a strong
coupling to a single timescale T (A) (see section 4), with inputs constrained to the width of the context bus.
5.5.7
Stability of contextual knowledge and weight maintenance (sleep functions?)
It seems plausible that ‘state of mind’ may be considered a fair synonym for context: i.e. a collection of contemporaneous or coincidental signals and concepts associated by simultaneous or concurrent inputs. The number of
sensory inputs feeding into that state thus plays a role in determining whether a rich enough context can be found
to refining and tokenize concepts, and provide robust conceptual representation in long term memory; similarly
in determining the number of independent degrees of freedom by which sensations can be felt or imagined by the
observer49 .
Lemma 36 (Concepts are distributed) Different aspects of a ‘thing’ may end up being encoded in quite different
parts of a semantic space (e.g. a brain). Localization of concepts is not guaranteed.
If multiple sensory inputs bring simultaneous activity into data summarization, different impressions may not all
converge to a single localized region to form a coherent representation as a ‘thing’, and an observer equipped with
a knowledge bank will be thrown into confusion. Thus, a knowledge representation needs to be able to handle
multi-dimensional parallel assessments of context as an addressing scheme for concept storage.
Because context acts either like a multi-switch, or as a network of signal concentrations that form a basis
for routing like in ANN, the routing mesh needed to coordinatize concepts has the structure of a directed graph,
49 It seems plausible that emotional sensation is no more than a particular state of mind. In an organism with myriad sensors, an emotional
sensation could be complex, an approximate facsimile of that entire complex sensory apparatus. For a system with few sensors, an emotion
would be little more than a feeling of hot and cold. In humans, emotions can feel intense, presumably for this reason.
79
evaluate
(emotion)
SHORT
+value assessment
+graduation of state
LONG
stabilized semantic
calibration
Figure 40: Self-calibration, where previously learned patterns are used as a measuring stick for new observations, leading to
a convergent result. If there is no convergence, we remain confused about measurements.
from the simplest signal wire to a very complicated routing fabric. Two dynamical processes may be at work in
associative knowledge representation: iterative learning (cyclic repetition, as in figure 40) [30, 102], as in section
2.2, and directed flow through discriminating fabrics. The latter have been studied in [41]. What is known, from
such studies of the underlying graph dynamics, is that all of these structures and processes may be unstable to
perturbations.
Lemma 37 (Instability of associations) Knowledge may be unstable to perturbations.
Annealing, smoothing
old concept
Recent concept
emphasised
Figure 41: De-emphasizing recent memory to stabilize the associative graph, by locally diffusing importance scores with
post-processing (like stimulated coherence in physics).
Instability is combatted by regular maintenance, at a spacetime scale that matches potential perturbations. This
is the essence of Shannon’s error correction theorem, and the maintenance theorem [7]. In sensory learning and
context awareness, it makes sense for a knowledge system to bias inputs in favour of what is happening right now,
rather than paying too much attention to the past. However, if taken to the extreme, focus on too short a timescale
would lead to ignoring long term experience in favour of ad hoc spikes of activity in disparate locations. Agents
that experienced many different contexts over long times, but only a few on short times, would be constantly
destabilized by a fickle accumulation of new unbalanced inputs. In order for memories not to become too biased
in favour of the present, some form of annealing or smoothing of the memories may be needed to compensate for
these biases (see figure 41)50 .
Hypothesis 5 (Annealing of knowledge) A form of low frequency (long wavelength) process that is de-focused
relative to the spatial scale of a concept representation (so as not to cause damage to individual memories), is
needed to smooth out associative link strengths in a dynamically learning knowledge representation, to apply a
fair weighting policy on memories.
The need for such a process makes complete sense from a spacetime viewpoint, but semantically it seems speculative. It is only by unifying both viewpoints that it becomes a very plausible stabilizing maintenance mechanism
for a learning system, analogous to time series smoothing51 .
50 It
51 In
is fascinating to speculate as to whether this is the (or one of the) function(s) of dreaming in animals.
the field of cybernetics, the idea of building a brain as a controller has been discussed for nearly a century, in different forms. Ashby,
80
5.5.8
Sensory perception: from context to reasoning
We associate intelligence with adaptation based on reasoning. It is therefore instructive to think about how tokenized context can lead to knowledge representations that support reasoning. These forms should emerge, rather
than being imprinted by design. This topic is a bridge to the next section about automated reasoning and inference
from knowledge representations.
There are two obvious reasons for wanting to study the emergence of reasoning structures. The first is pragmatic: the less supervision a learning process requires to think about context, the more impartial and valuable it
is. The second is more related to understanding how our arbitrary semantic spacetimes (including brains) may
actually work: no one is obviously born with a large hardwired conceptual ontology, so we have to explain how
ontologies emerge through learning [14].
The semantics of spacetime meet the semantics of observational input in the concept of streams. Streams are
ordered samples of data that form the basis for data acquisition, repetition, and thence stable knowledge (figure 45).
A key question, in the formation of practical knowledge representations, is: how can we automatically annotate the
a collection of data, to bring about an automatic emergent semantic model, i.e. not design an information model,
but a model of associative meaning? We return to this practical subject more fully in section 6.7. The process has
to involve a series of dynamical transformations, which correspond to spacetime structures. Schematically:
Sensing → Agentization(basis) → State → Association graph,
(171)
i.e.
semantic coordinatization
Arrival process −−−−−−−−−−−−−→ Matroid basis → High level graphs.
(172)
New impulses are recorded and recognized, then they are consistently routed to certain ‘empty’ memory agent
locations. These take on the role of proxies, in order to represent the incoming signals and eventual join them up
into concepts52 . Encoding of concepts and associations into a physical medium is a spacetime process: it involves
coordinatized spatial location of agents for pattern encoding, and sequential as well as cyclic time patterns for
coding and decoding with repetition for iterative confirmation. It is a process by which scaled knowledge ‘symbols’
are constructed from pre-adapted sensors53 , attending to a separation of timescales. Without such a separation, a
confusion of semantics would result.
Through the separation into context and associative concepts, we can now explain how a set of coarse characteristics, named or labelled quickly, may lead to the formation of a switched network, connecting or isolating
concepts only under specific circumstances, stimulated from the environment and from introspective feedback. A
bus of context reaches into long term memory, acting much like a transistor bias, or network routing fabric, but
with semantic addressing forming the interface to coordinate tuple addresses. Any spatial structure with this quick
approximate response to its surroundings can thus store contextual memories.
In the next section, we consider how this leads to associatively connected routes through a knowledge space.
If one starts at some context relevant location, and traces neighbouring concepts by association, focused pathways
and divergent trees (explosions from a starting point) can spread out, activating related concepts in order. This may
be perceived as reasoning, explanation, or brainstorming, depending on how focused the result seems to be. How
we get from associations to storylines and what we consider to be rational logic is the subject of the next section.
Example 107 (Named Data Networking (again)) The Named Data Networking example, used earlier [15], has
the structure of a routing network, based on semantic addressing rather than on explicit spacetime coordinates, but
does this make it also a representation of concepts, i.e. a knowledge representation (see figure 42)? A knowledge
representation lies inside a single observer. Based on the figure from example 65, we may draw two horizons that
represent where we consider the observer to be, in alternative interpretations of NDN as a knowledge space. We
can say immediately that NDN forms a fabric that switches data based on specialized sensors.
The observer sensors are the sources of data that transmit data by name. It is most natural to choose the outer
horizon as the boundary of the observer. Thus the sum of all receivers is the ‘mind of the observer’, in this proposed
mapping, i.e. ‘interest space’ is where concepts could be encoded.
in particular, discussed the notion of designing a brain based on the continuum feedback principles of his day [46, 103]. In his work, he notes
the separation of reflex responses and centralized control feedback. He also identifies the reliance of learning on a selection criterion that leads
principally to stability of knowledge. This suggests a separation of scales balancing dynamical convergence with modulation, very like the
discrete version here.
52 This process is sometimes called reification in computer science, but this terminology is closely associated with the Resource Description
Framework [104] for the semantic web [105], and related technologies [106], so I shall avoid using it.
53 A semantic ‘symbol’ may be represented by any convenient clusters of agents, they needn’t be represented as a non-overlapping alphabet
of agents, though there will always be a transformation to such a picture, in principle.
81
n
Source
Source
+D(n)
+cache
outer horizon
+F(n)|D(n),I(n)
inner horizon
Receiver
Figure 42: Is Named Data Networking (NDN) a knowledge representation supporting concepts and associations?
If the router keeps transaction logs, then it has a kind of memory of simultaneous activation of receivers that
showed common interest in a piece of data at the same moment. However, there is no exterior correlation between
data files, as there is no spacetime model relating files and data streams to a single model, so the a meaning of
such a correlation would be dubious. Because all data have to have a unique name in NDN, the ‘sensors’ (or
data sources) are unlike the sensors that provide data streams in organisms or knowledge systems, as the named
data never change: they are frozen in time. However, we could instead define sensors that point to aggregations of
similar things, e.g. different versions of a document, or chunks of a video stream. The aggregation of these can be
understood as the naming of superagents in the observer’s exterior.
Can we map concepts and associations to NDN structures? A concept would have to be mapped to a resting
place for related data in the space of receivers. The clustering of receivers with similar receipts could be interpreted
as a concept, however these are not linked except by common routing (like coins sorted into categories by a
machine). An association would be a link between two concepts. There are no structures that join the receivers
together except the routers. Correlations between the interests of autonomous receivers are the most tangible form
of semantic link. Unlike a true knowledge space, there is no way to promise persistent associations. Data interest
may be indirectly associated through the routing hub’s ephemeral promise to forward inputs, but these links are
neither persistent nor.
In conclusion, NDN does not form a knowledge representation, even though it is a semantic routing fabric.
It’s function is not be form a memory of the world, but to disseminate data in a fluid gaseous state. Chance
correlations could be observed between receivers, analogous to concepts, but the lack of context or permanent
association prevents them from iterative associative learning.
5.6
Spacetime representation of scaled concepts
Let’s consider how invariant and tokenized representations emerge from scaled agent clusters and irreducible association.
5.6.1
Examples of irreducible positves
To see how the irreducible types can lead to plausible encodings of concepts, through spacetime associations, it is
helpful to look at examples. By reducing to the most elementary forms, we end up with a kind of pidgin language,
but this seems like a good start that mirrors how language begins. Let’s characterize the irreducible associations
generically, without trying to nuance the expressions, to see if we can make ideas comprehensible by reduction:
Type
1.
2.
3.
4.
Verb
approaches
follows
contains
expresses
Prepositional association
near
before/after
inside/around
instead of
Spacetime interpretation
similarity/distance
causal/ordered
scale/aggregation
boundary structure (name)
Let’s try to combine these into more complex concepts, br reducing them to triplets of the form:
(concept, irreducible association, concept)
(173)
These examples may not be the only possible representations (if they were, they might be fragile), but they are
hopefully plausible and maximally simple, avoiding higher tensor forms in the interest of low cost simplicity.
Example 108 (Scaled conceptual cluster reductions for inference) Consider the following linguistic representations of common concepts:
82
1. “She is sitting on the roof”
This is a description of a static atemporal state. We could argue that these capture the situation:
(she, near, roof)
(she, is contained by, roof)
However, they cannot easily be distinguished from other similar events. Two different versions irreducible
concepts may apply to this situation. Either is approximately true, and the superposition of them, in provides
more specificity. Being inside a roof might sound odd, but we have no trouble thinking of someone being
inside the boundary of a roof, as a two dimensional space that marks a three dimensional volume. Our visual
apparatus is used to projecting images and concepts into two dimensional frames, which has the effect of
blurring these distinctions ‘metaphorically’.
A recursive approach offers a more precise representation than this simple direct representation, and illustrates how grammatical recursion helps to sort out the represenation of events.
(she, expresses, sitting on roof)
(sitting on roof, expresses the role, sitting)
(sitting on roof, expresses the attribute, on roof)
2. “She climbs onto the roof”
This is a description of a temporal process, leading to a future state.
(she, expresses, on roof)
(she, expresses, climbing onto) (climbing onto, expresses role, movement towards)
The superposition of these two association paths captures the idea. Alternatively, a interpretation that
includes causal transition may be included:
(she, leads to, on roof) (she, causes, on roof)
3. “Acid burns stuff”
This is a description of a temporal process.
(acid, leads to, burning)
(acid, expresses, burning)
(acid, leads to, burned stuff)
The challenge here is in linking the concept of burning (a verb) to a pair of things: acid and stuff. The use of
adjective clusters ‘burned stuff’ as an independent agent is one solution. This has the form of a superagent,
linked by context. One solution is the use of aggregate superagent clusters to employ recursion:
((acid, near ,stuff), expresses, burning)
((acid, near ,stuff), leads to, burning)
4. “Alice looks good”
This is description of state.
(Alice, near, good)
(Alice, expresses,good looks)
5. “The dog who loves adventure”
This is an explication of state, with an additional twist.
(Dog, near, adventure)
(Dog, expresses,love of adventure)
(Love of adventure, has role, love)
(Love of adventure, has attribute, adventure)
What is new here is that it introduces emotion into the expression. The use of ‘love’ may add weight or
priority to this property above others. It suggests a dimension of representation that is otherwise missing in
the simple spacetime focused discussion here.
83
6. “People must obey the law”
This is an explication of desired state, which introduces a temporal aspect pointing to the future.
(People, near, law)
(People, express, obey the law)
(Obey the law, has role, obey)
(Obey the law, has atribute, the law)
(Obey the law, depends on, the law)
The notion of force, expressed by ‘must’ is not a spacetime concept, but could be captured by the final causal
annotation. It refers to a desire, which has an emotional component, and also a future (temporal) component.
This is an intention, a goal, or a desired state.
These formulations may not seem linguistically familiar, but they are logically fair approximations to what we
mean by more habitual expressions. It is possible that there may be value in distinguishing refinements of these
wordings as independent associative link types, e.g. refinements of the form
X ‘leads to’ Y from X ‘affects’ Y
X ‘expresses’ Y from A ‘has property’ Y
may be necessary distinctions or merely scaled refinements, in the long run, in order to separate causal ordering
from the notion of generalized influence, but parsimony suggests otherwise at this stage of understanding. What is
important and striking is how few concept types are needed to represent most conceptual scenarios, and how these
shadow spacetime relationships. More explanation of these cases, with examples, is given in [107].
5.6.2
Noun and verb concepts
Some concepts suggest an understanding of time relative to contextual ‘now’.
f ollows
f ollows
Future −−−−−+ Context/Now −−−−−+ Past
(174)
Concepts may be both things or processes (nouns or verbs, in linguistic terms). These have distinct semantics.
Although, ultimately, everything must be encodable as spacetime patterns, elemental spacetime concepts are not
enough to represent the full spectrum of meanings we would expect of knowledge representations. However, they
do take us much further than has previously been realized.
The results of the previous reduction exercise look remarkably close in form and idea to the programme of
cognitive grammar by Langacker [20]. Emotional priorities are another dimension that needs serious consideration,
even if these are indirectly encoded through spacetime structures too.
5.6.3
Irreducible positves and negatives (encoding NOT)
The well known aphorism that ‘absence of evidence is not evidence of absence’ is a a basic property reflected in
promise theory in the form: the absence of a promise is not the promise of an absence. The same is not true of
complementary sets, or the Boolean algebra used in information technology, and leads to a sometimes parodically
deterministic belief that absence of information is simply not even taken as an option.
The human interpretation of ‘not’ we seek is not a notion of set complementarity, but rather a complementary
intent. We could express this by:
expresses NOT
human NOT −−−−−−−−+ body set complement.
(175)
This is easy to see. If we divide the world into that which is carrots and that which is not carrots, then promising
to not like carrots is not the same as promising to like the complement of carrots (everything else). Thus, when
we say ‘X is NOT a generalization of Y 0 , we do not mean that X is a generalization of the complement of Y
(NOT Y ); rather, we intend to add a specific note to disregard the assertion that X is a generalization of Y from
consideration.
The question arises as to how we represent this in a knowledge network. Promise theory guides us towards the
form of (175), by defining the a promise of NOT b ≡ ¬b to be the deliberate intent to not keep the promise of b.
This guides us to represent NOT on a par with the positives, i.e. for each positive irreducible association, there is
a mirror association that expresses the negative.
84
Lemma 38 (Negative association) For each promisable association in the irredicuble types, there is a corresponding association type formed from the negation of the association promise that represents a denial of an
association.
A
C1
0
−+
C2
¬A
C20
0
∀A∃A : A = ¬A
C10
(176)
(177)
−−+
(178)
The proof of this follows from definition of negative promise [6]. Let the promise of an association be b =
A
A
¬A
C1 −+ C2 , e.g. from (124), so ¬b = ¬(C1 −+ C2 ) implies the intent not to associate, i.e. ¬b = (C1 −−+ C2 ). So
this suggests the existence of four more (effectively independent) types of irreducible association for the negation
of intent:
• Does not generalize,
• Does not lead to,
• Does not express,
• Is not close to.
More explanation of these cases is given in [107].
5.6.4
Emotional origin to logical negation?
There is a certain emphatic (even emotional) appeal to NOT statements, such as we mean them humanly. This is
not captured by the dispassionate binary nature of logical complementarity. For example:
• I am not your friend.
• X is not next to Y .
• X is not represented by Y .
It is interesting to speculate whether the origin of this reasoning attribute lies more in an emotional reaction than
in a direct programming of embedded logic. For logic to result from an emotional outrage would be ironic to
many. We may only properly understand the nuances of reasoning once emotional weights can be incorporated
into storylines.
Negations are very specific targeted promises, whose subject is a denial rather than a logical complement. It is
the binary nature of Boolean representation that muddles these ideas together in modern information technology.
If we represent such associations for each of the irreducible types, then the question arises what happens if nodes
make both promises? It is interesting to speculate whether the origin of this reasoning Clearly, it makes sense for
opposites to coexist in different contexts54 , for instance:
Alone
Alone
generalizes|company
−−−−−−−−−−−+
does NOT generalize|electricity
−−−−−−−−−−−−−−−−+
isolated
(179)
isolated
(180)
These two coexist in different usage patterns, with unambiguous semantics. However, the semantics of the following seem problematic:
Alone
Alone
generalizes|company
−−−−−−−−−−−+
does NOT generalize|company
−−−−−−−−−−−−−−−−+
isolated
(181)
isolated.
(182)
These are mutually exclusive, in a logical sense. However, in a non-deterministic system mith Byzantine causation,
it could be difficult to prevent the formation of simulatenous links, so it is important to understand the effect of such
contradictory promises. Could this mean indecision? Indecision would be more likely (and cheaply) be represented
by the absence of a relationship altogether. The response to this ‘entangled’ Schrödinger’s catlike state might be
equally spurious, sometimes exciting one possibility and sometimes the opposite. This suggests that reasoning
may itself be stochastic in nature in a sufficiently complex system, and that stability of reasoning emerges from
larger scale aggregation rather microconnectivity.
54 Contexts
play the role of different Kripke or Everett universes.
85
5.6.5
Invariant representations and its relationship to tokenization
The issue of how a knowledge representation can cope with all the angles, perspectives, and distortions inherent in
pattern recognition, in multiple dimensions, has led to the idea that brains store imagery and other sensory experiences in ‘invariant representations’ rather than as libraries of rasterized screen shots. An invariant representation
implies the decoupling of a token representation from acquired data patterns.
In mathematical physics, invariances and covariant transformations base invariance around the existence of
symmetry within a single configuration space. For a semantic spacetime, tokenization (i.e. the association of a
single point with an invariant meaning) is the simplest invariant form. Invariance is thus immediately supported by
the tokenization discussed earlier, and the invariance of semantics is closely associated with spacetime structural
stability under accretion (see figure 43). The extreme end of this idea is that a single neuron might actually encode a
person or THING; and, while this might sound extreme, it has received some empirical support [38,108], suggesting
that there is indeed some kind of matroid basis element representation going on in a brain (see also the discussion
by Hawkins [19]).
Stability
slow
ow
le
dg
e
Concepts
Kn
at
io
n
rm
In
fo
context
temporal patterns
discrimination
no
is
e
D
at
a
Timescale
spatial patterns
fast
Sensors
Instability
aggregation/semantics/invariance
Figure 43: Scale/invariance diagram template.
What remains to be understood in a learning representation is how sensory inputs accrete into generalizations
under the irreducible associations, and these superagent structures become associated with concepts. Understanding this process in a fully emergent system like gene networks or a brain is a tough problem, but not an impossible
one (see, for example, the techniques of bioinformatics [109]). For the annotation of expert input, in a humanartificial intelligent cyborg approach, one can simply insert agents amongst a mixture of partially learnt and manually inserted concepts and associations. This hybrid approach becomes possible (even practical) now that we have
a simple spacetime representation with only eight irreducible relations. The main challenge lies in linking the data
patterns to the concepts seamlessly.
Example 109 (Aggregation to invariants) Figure 44 illustrates an example of a scale/invariance diagram, in
which sensors recognize a plate by matching sensory inputs to learned data patterns and associated attributes
expressed within a knowledge representation of a plate. Initially, the focused sensory experience of a plate is too
small to single out a plate from other possibilities (e.g. flying saucer or frisbee. However, once the context of being
in a kitchen is established, the concept of kitchen feeds back into context by ‘introspection’ singling out the plate,
by context labelling, as the likely match.
The goal for stable invariant representations must be towards increasing spatial aggregation and increasing (slower)
timescale. Pattern similarities that could lead to association into clusters include:
• Digitization into similar symbols alphabets q.
• Same name or property expressed.
• Similar changes in sensory inputs dq (rising/falling).
86
kitchen
slow
INTROSPECTION
kitchenware
implement
SENSING
smooth
hard
crockery
expr by|kitchen
dish
saucer
eating
Timescale
round
plate
white
feeling
previously learnt patterns
touch
DATA STORAGE/RECALL
shape
eyes
fast
colour
movement/probing
context
aggregation/semantics/invariance
concepts
Figure 44: Scale/invariance diagram example for recognition of a plate in the context of a kitchen.
• Correlation and coactivation (timescales on dq).
• Same function or purpose of a thing verb/noun representation.
• Matching the same pattern (epitope matching) [89].
• Common activation context (origin path).
• Dividing up and recombining at random (random recombination).
• Spatial search paths (sequence).
• Stable time-sequences and spatial aggregations.
Appendix B shows a further example based on more extensive system monitoring sensors.
6
Inference, logic, and reasoning in semantic spaces
The ability to associate a sequence of concepts into a narrative path is what we call reasoning. Reasoning takes
many forms [86, 110–115], and depends intrinsically on scale, because concepts scale like any other kind of agent.
A narrative argument at one scale may make no sense at another. Since Boole’s work [110,111], Hilbert’s challenge
[116, 117], and the arrival of digital computers, Boolean algebra and first order logic, have come to dominate our
thoughts about what is meant by reasoning. It has become a mathematized form of arithmetic, for the specific
purpose of making decisions. However, these are crude forms of reasoning compared to the processes by which
humans reason. To approach human forms of reasoning, we need to think more carefully about how concepts and
associations are woven into narratives.
6.1
Reasoning from learning
Several common forms of reasoning express pathways that connect concepts. Classic steps in reasoning include
the following:
• Induction: generalization based on specific instances (concepts from exemplars). These are represented by
the coarse graining associations between concepts (figure 33):
is generalized by
Cdog −−−−−−−−+ Canimal .
87
(183)
Underlying this are spacetime promises of the form of superagency (see paper II):
inside
Adog −−−→ Sanimals .
(184)
The process of induction is an N : 1 mapping from exemplars to generalizations.
• Deduction: inferring exemplars from identification of types. These are represented by compositional associations:
has property
Cdog −−−−−−+ Cwaggly tail .
(185)
In a promise model, the superagent representing ‘dog’ makes a number of exterior promises, such as ‘waggly
tail’. The process of induction is an 1 : N mapping from generalizations to exemplars.
• Abduction: Inferring a theory from an observation (learning or model building55 ). These are represented by
compositional associations:
characterizes
Cwaggly tail −−−−−−+ Cdog .
(186)
These are retrofitted inferences of identity, analogous to the matching of receptors56 . The process of induction is an N : 1 mapping from exemplars to stories or models of explanation.
• Causation: Inferring the prerequisites for the occurrence of a concept57 . These are represented by causal
associations,
CLights
depends on
−−−−−−+
Celectricity .
(187)
(188)
which, in turn, arise from underlying conditional promises:
Alight bulb
Apower socket
+shine|power
−−−−−−−→ Aobserver .
+power
−−−−→
Alight bulb .
(189)
(190)
It is important to remember, when reasoning, that effects may have multiple causal predecessors.
• Free association / lateral thinking: the exploration of concepts that are closely related to another, by some
definition. These are represented by similarity or proximity associations, and spacetime adjacency promises.
is similar to
Cdog −−−−−−+ Cwolf .
(191)
This kind of reasoning is saying: if you are thinking about ‘dog’, then what about ‘wolf’ as a similar case.
In other words, do you need more examples of your concept, without actually going to a new level of
generalization58 .
6.2
Associative model of reasoning
We return to the model of a knowledge system as the interaction between a single test user and its exterior world,
via sensors that channel their observations into a collection of internal agencies to represent knowledge in a concept
space (see figure 45). Recall that the model is about separating timescales into short and long term recall. Sensors
collect data and transform it into short term context. Context already possesses the semantics of its sensors, and
collectively stimulates concepts by some unspecified form of approximate addressing. Meanwhile, the activation
of concepts can trigger semantically close concepts which feed back into context as memories and mix with the
input stream to trigger even more related concepts. Thus, there is a cumulative exploration of the space of concepts,
55 This
is the approach used in Bayesian learning, or machine learning.
learning is all about abduction. Abduction can be applied to any structure. Induction goes further by discretizing into qualitative
attributes and concepts.
57 This is simply a definition of causation, as used pragmatically in the physical sciences. It assumes ‘intention from within’, but does not
imply any notions of free will or the many other complications that arise in a more general philosophy.
58 It is like Amazon’s referral engine: users who bought X also bought Y .
56 Machine
88
DETECTION
DISCRIMINATION
LEARNING
Context
Assessed situation
SENSORS
SHORT
(interpretation)
Concepts
Associations
LONG
Intent/belief
(measured)
Non−determined
Now
Cyclic−determined
Figure 45: Short term memory caches the state of ‘now’ from sensor results, but the semantics are pre-supposed. Long term
memory builds up the semantics slowly over time, so that it can be used to interpret sensors as context. Modulation of slowly
varying long term memory by quickly varying short-term memory allows us to switch large amounts of information into play
(that otherwise takes a long time to assimilate) quickly, depending on context.
based on a ‘state of awareness’59 . In order to be able to use these techniques, sensory inputs first have to trigger a
concept or observe a THING, which in turn triggers an association to a property, a question, or even a lateral guess.
A linearized sequence of inferences may be accumulated into the idea of a story or narrative [14, 118]). A
narrative is closer to what we think of as everyday human reasoning than is generally credited in the literature.
This approach is based on the idea that much of what humans consider to be ‘smart’ is based on storytelling, a
tradition that goes back to the earliest forms of communication. A story is a converging network of associations
that passes through a number of specific concepts of interest, including the subject or topic of the story. By tracing
a selected path of associations linking the subjects forwards or backwards, we can can divine both predictions of
consequences and possible related causes about the subject. Associations are based on the causal and structural
spacetime associations represented by the irreducible types60 , so we already know that they encode spacetime relationships and causation. The concept of stories is thus of great importance to understanding the role of knowledge
and inference in smart behaviours (see figure 46).
Definition 69 (Story or narrative about C) A directed acyclic graph that passes through C, and whose preceding association is of a single irreducible type. The story begins with a set of axiom concepts and ends with a set of
conclusion concepts.
There may be several associations passing through C, but those of different types represent different stories. All
C1
C
Axiom set
Conclusion
C2
C4
C3
Figure 46: A story is a directed graph that converges towards a single thread at the subject of interest (i.e. it may fan out in the
past).
forms of reasoning start from an initial concept or set of concepts, converging towards a single storyline.
Definition 70 (Equivalent narratives/stories) If two sequences of associations have the same axiom set and final
conclusions, then they may be considered equivalent alternative tellings of the same story.
Since stories belong in the space of concepts, they exist independently of the learning process. A story may have
been learned wholesale, or inferred from independent sensory input events, or even imagined by introspection. The
concepts spanned by a story may have been learned at any time prior to the formation of the story.
59 This is how CFEngine policy engine, and knowledge map was constructed, with sensors evaluating to short term classes that trigger the
long term encoded ‘policy’.
60 Since the first work about promises and knowledge [69, 119], I have still to find any related work on the importance of stories to human
thought.
89
Example 110 (Can we short-circuit learning?) Is a story, learned wholesale, e.g. from a book, or around a
campfire, a proper substitute for a sensory learning experience? The amount of information within a story, based
on a context-concept-association representation, is significantly less than the information that was involved in the
original learning process, through sensory recognition and repetition. Suppose, instead, that the story was not
experienced directly, but merely passed on as a linguistic rendition, e.g. by reading a book: now, the information
in the tokens of the story may be transmitted exactly, from the book to the reader, and provide the same information
to different recipients. It might seem that identical knowledge could be passed on. What may not be the same,
however, is how the concepts used in the story link up in the network of the receiver prior to the telling, thus
extending beyond the telling of the story itself. This may differ between the storyteller and any given story receiver.
The degree of similarity in the final state depends on the extent to which the observer already has all the dependent
concepts in place to being with. Since the story modifies the network of concepts in the listener, the initial state
before the story and the final state after may not correlate at all well between storyteller and story receiver. A
full sensory learning experience will undoubtedly lead to more associations being formed around the storyline in
memory, but this depends on both the initial state and the unique sensory experience of the receiver. In short, it is
impossible to say what effect either the storytelling or sensory learning might have on different observers.
Example 111 (Stability of chained neural networks) Neural networks (ANN) are pattern discriminators, which
approximate functional mappings, and learn by supervised training. They have no other rational reasoning capabilities, so any attempt to solve an extended reasoning problem with ANN has to be solved in multiple stages,
by turning a series of tailored discriminators into a series of transformation functions. One approach is to chain
together multiple pattern discriminators in a directional hierarchy [19, 52], where early layers transmute sensory
input into a reduced dimensional (tokenized) form, and later layers identify combinatorics of these reduced forms.
This is a machine learning representation of the concept learning schema described in this section.
If independent neural nets are strongly coupled, in such an array, then they are effectively a single tightly
coupled network, with each network Na ’s hidden layers trained independently by data sets Da , but coupled as a
single structure. If there is no separate training, they are a single neural network, capable of discriminating only
a single input to output association. Each layer must be trained independently, else the combined network is just
a deeper single pattern discriminator. In such a system, the stability of a reasoned outcome builds on the stability
of all the layers. A later layer that depends on a previous layer, inputs the modulations of the earlier layer and
outputs a contextualized result to a later layer. This output is used in two ways: during operation and during
training. Referring to section 4.1.2, we can describe the timescales for deterministic operation.
Na
Layer
Nb
modulate
Nc
Layer
train
Training
data
Figure 47: Each layer in a serial network may be altered by operational modulations or by retraining. Earlier layers promise
to modulate later layers, and each layer promises to retrain from datasets, on different timescales.
+mb (da )|ma ,db
• T (mb ) ≡ T (Nb −−−−−−−−−−→ Nc ) the time it takes a layer to operate, transmuting input from layer a to
layer b, where b > a. The promise is a conditional one, depending on both the modulation of the previous
layer ma and the training data of the present layer db .
+∆d
b
• T (Db −−−−→
Nb ) the time in between changes of training data ∆da for the layer from world sources.
−∆d
b
• T (db ) ≡ T (Nb −−−−→
Db ) the time to accept and process changes of training data ∆da by the layer Na .
The switching will be semantically and dynamically stable (quasi-deterministic) iff
T (db ) T (ma ),
b>a
(192)
as described in section 4.1.2. The world must not change substantially during the training of the entire reasoning
process, else the result will be invalid before a conclusion can reached, and the work done in training will not be
reusable, i.e. it cannot keep up its evolutionary fitness for purpose. This is an evolutionary constraint: the more
primitive senses should be basically fixed compared to the plastic later stages. This is certainly true in biology,
where basic senses are encoded genetically, and brain learning of what there is to see is nurtured.
90
The training of later layers must depend on the training of the earlier layers, thus the earlier layers must
behave consistently over the time it takes to train the later ones. This means that the networks should not change
significantly, or
T (Da ) T (Db )
a<b
(193)
i.e. later stages may be more plastic than earlier ones.
However, in order to modulate later layers successfully, during training, (later layers learn from repeated
invocations of earlier layers), the operation time of an earlier layer must be significantly shorter than the training
time of the later layers, else learning cannot be repeated a sufficient number of times:
T (ma ) T (da ),
(194)
which is consistent with (192). If these inequalities are not met, the training of one layer will be rendered unstable
by a change in an earlier layer.
6.3
Couch’s rules for inference
Alva Couch made the observation that the irreducible positive associations lead to connected stories of decreasing
certainty with increasing length [118]. The essence of these rules was based on the irreducible spacetime structures
described earlier, and the uncertainty from semantics is introduced semantically rather than quantitatively (as, for
example, in statistics). From the irreducible associations, the following examples illustrate the idea:
1. If there is a route from A to B, and from B to C, then there might be a route from A to C.
2. If A is inside B, and B is inside C, then A might be considered inside C.
3. If A depends on B and B depends on C, then A might depend on C.
The appearance of ‘might’ signals an increase in uncertainty61 . There is no guarantee of the conclusion, because
the semantics of the agents need not be uniform; nevertheless, the outcome is possible or even plausible. For
instance, we might throw our hands up at 2 and say that the second of these must allow us to say that A is inside
C. However, what if we open up C then we find only B because B is opaque? How could we know that we
are supposed to open B too? And when are we allowed to stop opening boxes? It is a perfectly valid semantic
point of view that C is not really inside A, because being inside suggests being able to see or touch the boundary.
Whenever autonomous agents becomes intermediaries for a process, one cannot be certain that they will behave
uniformly62 . With these rules in mind, the process of inference over arbitrary associations lies in the ability:
• To interpret associations as aliases for the irreducible association types, so that we can always reduce chains
of arbitrary associations to chains of the irreducible types.
• To form triangle reductions that condense multiple concepts or associations into fewer based on the irreducible associations [118].
Inference, based on the irreducible associations, is a powerful tool for working in tokenized form. Simple search
algorithms may be used to generate stories. While not all these stories emerge immediately in well-formed human
readable language, the essence of a skeletal story does emerge from simple promises or rules about parsing the
association graph. This is an indication of the potential simplicity of reasoning, but at the same time the complexity
of linguistic expression.
6.4
Stories and narrative reasoning
The importance of stories, as defined, is much deeper than a cultural memory of tales told around a campfire.
Stories form patterns of functional relationships in spacetime: in other words, they are the map between algorithm,
reasoning, and machinery. We use them for reasoning by following pathways until we find concepts that match the
criteria of a question. Stories lie at the heart of processes, algorithms, and also what we humans mean when we
speak of explanation and understanding.
Definition 71 (Understanding or explanation of C) The ability for an agent tell a story that converges on C.
61 The actual rulesets were much more specific and detailed than these, based on a particular model implementation, but this captures the
principle.
62 One way to be assured of homogeneity is for them to promise homogeneity.
91
Definition 72 (Depth of understanding of concept C) The network distance, or number of associations between
concept nodes that precede the concept C in an explanatory story.
People differ in what standard of depth they consider to be sufficient to understand something. Yet these simple
definitions capture the basic essence of what is commonly meant by humans concerning imagination and understanding.
Example 112 (When do we really understand something?) Alice may say that she understands atoms, by telling
the following story: atoms are made up of protons, neutrons, and electrons. Later, she may change her mind and
add another level of predecessor: atoms are made up of protons, neutrons, and electrons, and protons and neutrons
are made up of quarks. The number of levels of explanation in terms of associations to other things may continue
indefinitely (turtles all the way down).
During the parsing of a story, the context of a knowledge agent’s mind is modified by the successive details of the
story: new context is refreshed as the stages of the story are revealed. This context may condition the associations
to particular concepts, and feedback into the stream. Thus the accumulation of context helps to build a case for an
‘answer’ or ‘conclusion’ to the story.
Lemma 39 (Story limits) If a concept C has at least one association, there is a story. There is no limit to the
possible length of a story.
6.5
Story types, classified by their promises and spacetime associations
We may classify stories into different types, according to the type of irreducible linkage they use.
Example 113 (Episodic history) A sequential timeline story (history) such as:
1. John got up
2. John went to work
3. John had lunch
4. John came home,
5. etc
may be represented at a story formed from time markers joined by precedence associations, each of which is
attached to a concept associated with the event. This is a simple, regular molecular chain: First we notice that the
precedes
concept
"up"
work
event
eats
causes lunch to be eaten
is characterized by
changes location
is characterized by
changes state
association
precedes
event
event
event
lunch
is characterized by
changes location
precedes
time
home
Figure 48: A timeline history: ordered molecules of timestamps and associated happenings.
story is made up of promises, not associations, and that each promise precedes the next.
π1 : John
got up
−−−→
went to work
π2 : John −−−−−−→
ate lunch
π3 : John
−−−−→
π4 : John
went home
−−−−−→
92
observer
(195)
observer
(196)
observer
(197)
observer
(198)
The first is a change of state, i.e. it is a scalar promise. The latter three contain both verbs (links) and nouns
(agents), i.e. they are vector promises. To turn these into associations, we separate out the concepts. In the first
case, this involves introducing a ‘state’ agent, as in the example for ‘blue’ in figure 5. The timestamps act like
road signs guiding the story forward and preserving the order. They need not be numbered, provided they preserve
their solid state order. The result is a spacetime crystalline phase in the form of a chain. Without forming these
molecular pairs, the concepts have no ability to become associated only for this one example. It would be wrong
to permanently associate up, work, lunch, and home with a precedence relationship: what we want to express is
what happened in the context of this episode.
Example 114 (Conceptual story) Look at the knowledge space or concept network in figure 49. This shows the
concepts in the mind of an observer “me”. By exploring at random, with a god’s eye view of the mind of “me”,
we can infer that “me” works for the police, and that there has been a theft at the village bakery, somewhere in
Oxford. The extent to which this picture is available to the individual agents is not important, because we know
john
police
village
works at
located in
located
happened at
works for
theft
bakery
visits
Oxford
jane
makes
works in
friend of
Me
sarah
bread
likes/wears
crumbs
wool
e.g.
leads to
e.g.
stick to
sourdough
wholegrain
Figure 49: Knowledge space belonging to “me”.
that knowledge and concepts lie in the distributed network of concepts, not in any viewpoint located inside it (there
is no neuron that knows what we are thinking: ideas are something that emerge at a larger scale).
By sensing the exterior world, police person “me” observes that Jane has crumbs on her sweater. The concepts
of Jane and crumbs and sweater are thus activated simultaneously as the current short term situation context. By
following causal links, we find a specific story about Jane, from “my” perspective is shown overlaid on figure 50.
is liked by
stick to
friend of
occurrence of
visits
crumbs −−−+ wool −−−−−+ Sarah −−−−+ Jane −−+ bakery −−−−−−−+ theft.
john
police
village
located in
located
works at
happened at
visits
Oxford
works for
jane
theft
bakery
makes
works in
friend of
Me
sarah
bread
likes/wears
wool
crumbs
e.g.
leads to
e.g.
stick to
wholegrain
Figure 50: A story about Jane.
93
sourdough
Oxford
Oxford
located in
might work in
work
property of
me
me
Figure 51: Quasi-transitive inference altering associations.
Oxford
Oxford
located in
work
located in
property of
my work
me
Figure 52: Another way for the scaling of agency to represent composite concepts.
In this example, we see how the scaling of agency and inference of quasi-transitive association, mimicking basic
spacetime concepts, allow us to anthropomorphize inanimate things, and attribute intentional language to simple
inanimate THINGS. This is a natural extension of the overlap between human and non-human concepts in an
associating network.
If stories are as important as they seem to be, our understanding of reasoning in computer science has to change
(see figure 53).
Axioms, starting point
Conclusion
Theorem proving, logic, reasoning ...
Figure 53: Logic is not a succession of inevitable mandatory steps, but rather a sequence of axioms promised and conclusions
accepted on the trust of the axioms.
6.6
Inferences from algebra of association: type symmetry transformations
Another class of stories are those that allow combinatoric inference. If we return to the example of family roles,
we see that some associations do not require an inverse to be promised: they simply follow as facts and require no
additional documentation, as both directions are properties of the ‘mind’ of an observer. However, promises have a
different logic, and do need reciprocal confirmation, because they originate from independent agents that may not
agree.
• Promise logic: Suppose exterior agents make promises:
JOHN
JANE
JOHN
+f ather of
−−−−−−−→
+daughter of
−−−−−−−−−→
−daughter of
−−−−−−−−−→
94
JANE
(199)
JOHN
(200)
JANE
(201)
John promises that he is Jane’s father, but Jane does not confirm. This could mean that he is lying, and an
observer would not be able to trust the promise without some corroboration. However, Jane promises that she
is John’s daughter and John confirms, so an implicit confirmation is in fact given. Because the relationship
between father and daughter is an independent linguistic fact, this is sufficient to conclude that the missing
promise is given63 . Thus, we may infer:
−f ather of
+daughter of
JANE −−−−−−−−−→ JOHN implies JANE −−−−−−−→ JOHN.
(202)
These two promises are not independent. Similarly, daughter, son, and child are not independent vectors:
{+male, −daughter} → f ather
(203)
Perhaps we need to use conditionals to vectorize
+daughter
=
+child|f emale
(204)
+son
=
+child|male
(205)
(206)
Note that these are conditional promise relationships, not associations.
• Association logic: Unlike the case for promises, knowledge, i.e. that which has been learnt, has already been
trusted and inferences may be predicted, even if they haven’t been observed. This is because associations are
all inside the mind of a single observer. The representation might be formed of independent agents, but we
can assume that the knowledge is likely consistent. This is how ontology arises. So, assuming that the family
concepts are known to the observer, we can make some inferences about named reducible associations that
would not necessarily follow for promises. This is because we are assuming that all the promises may be
taken as given.
male
Simon
gender
gender
father
parent
son
father
Fabian
parent
father
daughter
Dawn
gender
female
(b)
(a)
Figure 54: Different models of the same associative information, one using names, the other using roles. The named version
looks simpler, but the associations conceal a new concept: father, so they must be reducible.
As noted in section 5.4.10, we can break complex associations down into irreducible parts64 . Let’s return to
the father-child example (see figure 54). The basic associations act as transformations (functional vectors)
that enable immediate inferences to be made:
precedes
Father −−−−−+ {son, daughter, child, . . .}
(207)
There is another association to gender, so it is possible to relate the inverse to the context of the gender.
– For gender neutral, the inverse of father is child.
– For gender male, the inverse of father is son.
– For gender female, the inverse of father is daughter.
63 It
is still possible that they are both lying, but there are stronger grounds to trust the promises now. Also, an adopted daughter may lead to
further semantic quarrels about the reciprocity of the relationship, but this serves to illustrate the principle, and I’ll leave the remainder as an
exercise for the reader.
64 This is a hypothesis, demonstrated in a few cases. It remains to be proven as a theorem.
95
We can proceed in this way to deconstruct the complex association ‘is father to’ as ‘male’ and ‘precedent’,
except male is a role association like ‘father’, so it makes little difference which we choose. What is useful in
terms of spacetime inference is that the causal precedence is represented, as this can be used for potentially
transitive reasoning.
• Combining vertical and horizontal association
Deduction:
likes
−−+
general concept
has exemplar
−−−−−−+
general concept
infer:
likes
−−+
particular example
fish
(208)
particular example
(209)
fish.
(210)
fish
(211)
general concept
(212)
????
(213)
A?
(214)
motion
(215)
force
(216)
force
(217)
The reverse inference cannot be made:
likes
−−+
particular example
is generalized by
−−−−−−−−+
particular example
infer:
likes
−−+
general concept
Induction:
moves
−−−→
THING
thing
is characterized by
−−−−−−−−−+
is affected by
−−−−−−−+
motion
infer:
6.6.1
is affected by
−−−−−−−+
thing
Questions and answers: contextual stories
A facility that we associate with smart behaviour is the ability to answer questions. Answering questions is similar
to the process of deduction. Learning answers to questions is like contextualizing input. The reverse is like
contextual recall (see figure 55).
answers
question (Q)
context
Figure 55: Questions matched to answers for each context of the question is a 1:N process.
Definition 73 (Question) A proposition about a subgraph of concepts and associations (called the query set Q),
or a prompt to tell a story involving the mentioned concepts and associations.
Definition 74 (Answer) A response to a question, proposition, or situation, defining a query subgraph Q of concepts and associations, i.e. the telling of a story about the query set, in which the query set plays the role of the
subject of the story.
Answers may be given in terms of generalizations and stories about concepts stated in the question. It is not
necessarily true that the query set maps to the axiom set of story. These would be learned by induction or abduction
over time. When we answer a question, we are asked to reply with a story matching the context of the question.
Question answering is thus a form of (or at least related to) deduction.
96
6.6.2
The storytelling hypothesis
The central importance of storytelling to reasoning and problem solving suggests that no knowledge system can
be considered self-adapating without this capability. So, it seems natural to highlight this by mirroring the context
adaptation hypothesis, we can now add a promise that must be given in order for a knowledge representation to be
capable of knowledge-based reasoning.
Hypothesis 6 (Separate agent for story inference and navigation) In order to extract knowledge introspectively
from a knowledge representation, there must exist an independent mechanism or agency capable of searching and
following paths surrounding a concept and feeding back the narrative.
It is fascinating to speculate on the relationship between this narrative engine and the phenomenon of consciousness
in sufficiently complex knowledge systems.
6.7
Semantic commentary of structures
In spite of the successes of machine learning, and pattern recognition, humans excel at learning, and each individual carries vast reservoirs of knowledge and experience. Some level of understanding can be passed on without
extensive experiential training, through books, stories, or by encoded as expert systems, in order to be built on
by learning or compounded by other multiple sources. In the modern approach to so-called artificial intelligence,
much effort is expended to accomplish pattern recognition by expensive supervised training. This need not be
pursued to the exclusion of the fruits of human knowledge. Given a data set, from some kind of sensor, it may still
be helpful for experts to annotate the data, based on their broader contextual understanding of what is intended and
what is accomplished. This can be done by annotating keys and values to explain the semantic significance, at a
level that can be used for reasoning. This is a form of documentation, and may be applied to any data process to
enhance the interpretation of sensory data.
Example 115 (Semantic annotation of learning networks) Unsupervised learning, with no initial boundary semantics, have no way of developing semantic concepts over time. Without external guidance, a network that
gathers data may see patterns cluster into agencies that resemble concepts, but the names of those concepts are
unknown. Learning systems need grounding in the human realm to give meanings to their discoveries that we can
understand. The quick and simple way to assign meaning to an artificially learned network of concepts is to have
human curators name and comment on the structures. This is exactly what happens in the training of human brains
from birth.
There are two forms of semantic grounding that allow animals to learn the same concepts: i) is the evolutionary
hardcoded information that has been groomed by selection, and ii) is the reservoir of social norms and writings
that stabilizes behaviours and passes on names in a long term memory from generation to generation. Mass culture
plays a key role in the development of knowledge. If we continue to give society tools to communicate in private
channels, closed rooms, cliques and tribes, it seems likely that society would lose its ability to pass on certain kinds
of knowledge by equilibration with a mass norm.
Example 116 (Semantic commentary) Consider a fragment of programming pseudo-code:
string animal_of_the_day = "vegan lamb replacement";
string text = "Mary had a little " . animal_of_the_day;
integer result = count_words(text);
Using the core association types, we could annotate:
• animal_of_the_day is a member of string.
• text is a member of string.
• result is a member of integer.
• count_words is a member of functions.
• count_words is a member of integer.
• result depends on text.
• result depends on count_words.
• text depends on animal_of_the_day.
97
Hence we may infer that result depends on animal_of_the_day through the following story:
1. result depends on text and count_words().
2. text depends on animal_of_the_day.
3. By inference rules: text may depend on animal_of_the_day.
We don’t know this for certain, because we don’t know whether count_words() actually promises to use text.
However, we can say that it is possible.
Using this approach, one can generate semantic networks, as concept graphs, which are more efficient to construct
and update than techniques that only employ probabilistic networks for machine learning. The coordinatization
of documents allows a practical annotation mechanism to be implemented as a technology (see the example in
appendix A).
6.8
Summary
Setting aside the many deep technical issues in realizing concept networks, the basic processes of knowledge
representation are represented schematically in figure 25.
• Knowledge is acquired by learning from sensors and by thinking (introspection): an iterative process of
associating ideas to understand and react to the world.
• Concepts and associations form a tokenized model, analogous to language, of what is sensed and understood.
• Concepts and associations form a semantic space, with four irreducible association types.
• Concepts grow in significance the more often they are revisited.
• The iterative generation of context from combining sensing and introspection drives an observer’s clock
sense of time.
These ideas can now be applied to a multitude of representations, to help understand the process of adaptation, and
harness technologies for improving the value of learning to organisms or materials at all scales.
7
Smart spaces
The concept of a ‘smart space’ may refer to any kind of structure in space and time that mimics or otherwise
represents knowledge-oriented behaviours. This may be at any scale, e.g. it could include new kinds of material,
buildings, homes, vehicles, towns, etc. Thus it readily encompasses systems we do not normally associate with
intelligence. In the human realm, the aim of ‘smart spaces’ is to enhance the experience of people living in or
around them, or bring a positive benefit to society. In an industrial capacity, the aim might be to allow specialized
functional spaces (like factories, farms, etc) to become more self-sufficient in operation and adapting to contextual
challenges. In this latter view, smart spaces are a form of cybernetic organism, or what is currently referred to as
artificial intelligence. Smart spaces may be addressed from three perspectives:
• Smart environment: where space records knowledge about its exterior interactions with occupants and
environment, and may be equipped to react and interact more intelligently. If the occupants are humans,
then this is a form of ergonomics, i.e. a fitness for purpose strategy.
• Smart occupant: where occupants of a space are equipped to interact more intelligently with the space they
occupy. If the occupants are humans, then this is a cyborg strategy.
• Smart organism: as a single system in which some agents are mobile and others are fixed, but all work
cooperatively, over some bounded region.
Smart need not imply the direct insertion of information technology, as is sometimes assumed in technology circles.
It might simply refer to the ability to build valuable stories between initially unrelated concepts within the space,
including connecting people, functions, and businesses together to adapt to new challenges. From the foregoing
discussion, it should be clear that smart behaviours can emerge in all kinds of representations, without necessarily
trying to bend systems to human ways of thinking. The immune system remains a powerful example of how an
assembly of ‘dumb’ agents exhibit smart adaptive behaviour.
98
Several interpretations are possible. Should we consider occupants (human or otherwise) to be part of a process
(as in a factory), or as guests (as in a hotel), or even as intruders (as in an immune system), or as symbiotic
partners in the space (as in a shopping mall). For example, we could think of a smart home as providing a service
to its residents (like a hotel), or as an autonomous smart building embedded within an external region acting
cooperatively to share resources and minimize waste. Systems are often strongly interactive, thus boundaries may
have fluid meanings, and clean cut roles could be difficult to decide a priori. A smart space might balance all these
viewpoints simultaneously.
7.1
Interactive spaces
In studying semantic spaces, we have a choice:
• Consider how to make tools that enhance the abilities of humans.
• Consider a structure (e.g. a city, a house, a material etc) not as a external thing to be learned about in the mind
of an observer, but as representation of a mind itself, looking onto something else or perhaps introspecting
about itself.
From the perspective of the latter, we see adjacencies not as distances but as associations, places as concepts, and
the environment plus its visitors as sensory stimuli. What might this space now be thinking about? If an imaginary
deity were to look down on the mind of a city, would it appear to be smart? If so, at what scale? Would it appear
as a single organism, or as a swarm of interacting organisms [65, 120, 121]? Who is it, in reality, who judges these
behaviours, and on what basis?
We may thus pose the question: what is the relationship between smart spaces and semantic spaces? Why
should we care about such questions? One answer is that our brains are such a smart spaces that we want to understand. Another is to ask whether we could improve our interactions, to better utilize knowledge, by the formation
of new structures that go beyond the scale of an individual human: teams, costumes, rooms, communities, cities,
etc.
Example 117 (Where is the memory of a community?) In the past we have assumed that the memory of a community was the memories of the people within it, and that their personal relationships formed the adjacencies to
pass on stories representing concepts to others. Later, we built non-human structures to remember: imagery, art,
monuments, tombs, statues, writing, books, libraries, architecture, institutions, cultural habits, and so on.
Example 118 (Social networks as smart associations) A social network represents an association between parts
of a space of possibilities, mediated by humans. It forms a part of a knowledge representation. Links between
members of a social network lead to clustering of interests, i.e. concepts. Recommendations by one member of a
group get passed on to others until the group represents an interest in that topic, while at the same time expanding
the concept to overlap with other issues. Social networks lead to mixing of concepts, and humans act as message
carriers in the organism of a space/city/etc. Parts of the concept cluster may come at the topic from different
angles, triggered by different atoms of context during a concurrent activation.
7.2
What does smart mean?
Smart is an assessment made by an exterior observer. It is based prejudicially on our understanding of human
capabilities and ingenuity: we define our own idea of smart by our ability to use resources as tools, to turn learning
into practical and useful behaviours, to innovate by combining resources into new functions, to save time and
effort or optimize processes. We believe that people who ‘see more’ are smarter. From the perspectives of scale
and relevance, there are two pragmatic criteria that suggest ‘smart’:
• When we perceive an action as helpful in a given context (fit for our present purpose).
• When the outcome arrives ‘just in time’ to be useful in that same context (anyone can come up with something too late).
A smart space is about facilitating interaction that leads to economics benefits, including advancement, innovation
and happiness. Emotional aspects of systems may not be discounted as long humans make the assessments.
In the preceding sections, we’ve looked at how spacetime itself can encode the basic elements of knowledge
and reasoning. This may or may not be sufficient to justify calling a space smart. An independent assessment of
‘smart’ is a matter of perspective: each individual observer decides whether they experience something as smart or
not, relative to their own expectations. Finally, given our ability to enhance systems using Information Technology
(IT), could we make a space even smarter? We might posit that a space could be considered smart if it could
answer yes to a number of questions that include the following:
99
1. Does the space have some interpretation of sensory input, i.e. does it interact some a world beyond it?
Any space that is bounded, and which interacts with an external agency has the capability to exchange
information. Where do we draw the line? Identifying the agents on the perimeter, what promises do they
make, and how do these reveal the world as a sensory experience? Are there new agents and sensors that we
can equip the space with the extend these capabilities?
2. Does the space have a capacity to learn and encode concept and associations?
Any space, which can represent different internal states, possesses this capability, but does it have associative
and temporal memory, it is plastic in its contextual adaptation? Can it mix and combine ideas into new ones
(innovation), and is it fast enough to apply its learning in interactions with real time changes in its exterior
environment? Can we extend its capabilities with artificial means?
3. Do the normal dynamical processes of the space tell stories about itself?
Is there causal linkage between remembered events, forming a meaningful narrative? Can a space reason
independently, through its own processes? How could this reasoning benefit occupants or users? How we
answer that question probably depends on our expectations. There are very simple examples of reasoning
that we call algorithms:
Example 119 (Mechanical reasoning) The stopcock in a water closet (the switch that fills up the water
closet tank and stops filling when it is full) is a reasoning system. It has a sensor, which is the water level.
The sensor forms a sense of context: full or not full, and transduces this into a switched on a water valve.
Thus it understands two concepts: full (represented by the switch being open and the valve closed, which
usually implies that the tank is filled with liquid) and not full (represented by the switch being close and the
valve being open, for any other condition).
4. Does the structure of spacetime alter the way new sensory inputs are perceived?
Do pre-existing concepts feed back into the sensory stream, acting as anchors for new ideas or external
interactions to strengthen and expand upon?
Smart appears to be something to do with a plasticity to challenges, and innovating quickly around them, with
approximate goals. In promise theory language, we could define smart like this:
Definition 75 (Smart promise or function) We perceive an action as helpful in a given context (fit for our present
purpose). The action arrives ‘just in time’ to be useful in that same context (anyone can come up with something
too late).
7.3
Promise view
Given that we can relate promises to knowledge representations, from a promise theory viewpoint, we may begin
by asking the usual questions about the cases of interest, defining inside and outside of what we take to be the
semantic space, marked by its horizon.
1. What or who are the operational agents inside and outside the semantic space?
In an arbitrary space, we can find agents, with diverse capabilities, at all scales. Atoms, molecules, bacteria,
viruses, people, tools, vehicles, machines, building, regions, malls, warehouses, shops, streets, shop signs,
road signs, public structures, monuments, bridges, districts, parks, cities, countries, planets, etc.
How are these agents ordered and addressed?
• Should we distinguish between the semantic space and its occupants at all, or simply treat a system
as self-contained, with some agents mobile and others fixed? If so, the occupants would be formally
outside the space in our model (even though the occupants of a smart house are within it from a normal
spacetime perspective).
• What processes happen inside the space?
• How does the space sense and interact with across its boundary or with its occupants, however they are
defined?
• Who or what is doing the thinking in the system, to judge ‘smart’?
100
2. What promises do the agents make to one another, in what context?
Proximity, bonding, services, or how may we interpret these as knowledge of associations, by one of the
four irreducible types or any possible alias of those.
3. What phase(s) are the agents in?
Gases, bacteria, animals, people, and vehicles are usually in a gaseous phase, as they move around and
environment that is fixed with solid bindings. There are anchored places, forming concepts, with freely
flowing messengers to associate them.
4. What are the important scales for their interactions?
For a knowledge representation, scales are pinned by the scale of sensory apparatus. A smart space interacts
with its exterior through such sensors.
Example 120 A smart material may feel pressure on a surface; a smart city may experience trade and
migration on a timescale limited by surface area of transport hubs; a smart house is built for 1-10 people
who have a fairly predictable size and speed. Dependencies (like transportation and interface surface)
condition what happens inside a space on what happens outside it.
Based on the principles of promise theory, using only agents and promises, one may now begin to map behaviours
encoded in different spacetime media to challenges faced by a space formed from a collection of such agents.
Inside the observer horizon, we can build the notions of concepts and associations on top of these. From this, if
there is a plausible similarity between such behaviours and the behaviours we understand as smart in organisms,
then it is fair to say that we have understood the meaning of an analogous smart space.
7.4
Benefits of a knowledge representation in real space
7.4.1
Memory
The ability for a functional space to remember something depends on the shaping of structures within its observer
horizon. Any persistent spatial configurations and spacetime processes (orbits, loops, interactions) may be used
to represent an addressable memory inside the horizon, and this, in turn, can be used for learning. The speed
and spatial capacity of memory are determined by all the spacetime scales of the agents. Memory could be short
term and long term, coarse or refined. Realizing a knowledge representation in real space, is a process of iterative
learning, by direct analogy with learning by an individual, resulting in:
• Concepts: formed from any agents or patterned clusters of agents.
• Associations: links, by adjacency, or by realtime processes, e.g. transportation or messenger.
Example 121 (City memory 1) In a city, short term memory is found in the locations of mobile agents like people,
cars, taxis, work crews, or in the mutable states of agents, e.g. the state of traffic lights, the number of cars in a
parking lot, or the stock levels in a shop. The wear and tear on roads and public infrastructure recalls its usage.
The memory of sunshine could be the worn-away grass from walking or playing in the park. This imprint might be
associative, as it joins locations the start to the end of the track. On longer timescales, buildings records cultural
history. The locations of parks and monuments. In buildings, allusions to Greek styles or familiar forms (doorways,
windows, etc) root function in historical memory associations.
Example 122 (City memory 2) A cloud of pollution is a spacetime trace that recalls a process or event that
caused it. If the process continues, it will be replenished and the trace will be learnt by becoming persistent,
e.g. smog. If it is ephemeral, it will be quickly forgotten. Long term memories may survive as chemical traces
in concrete or trees. Other external sensory events may leave memories over time, e.g. an earthquake may result
in a change in building codes, and stronger buildings. This would be the memory imprint of the concept of an
earthquake.
Stored memories must also be recalled in order to be of value. In other words, recognition has to be mutual.
Example 123 (Vestigial structures) Greek columns on the face of a building recall Greek culture or symbolize
grandeur, but the allusion is meaningless unless it has resonance with observers’ and their shared ability to recognize the allusion. Similarly, the presence of a phone box, or disused railway tracks, or wartime bunkers.
We may also grow accustomed to the Greek facade and relearn its significance without any direct transmission
of intent. These are memories encoded in structures, without a doubt, but they do not make a city smarter unless
they assist other processes in the space in overcoming challenges.
101
It may be agreements, rather than promises alone that lead to memory. In every learning relationship, there is a
mutual interaction leading to a local Nash equilibrium [3, 122]. This is why bidirectional iterative relationships are
the key to knowledge.
7.4.2
Sensor types and dimensionality of context
On the outside of an observer horizon, space may extend into dimensions other than the obvious spatial dispersion
of locations joined up by vector promises; at each scale, there may be interior scalar promises, at embedded
locations that encapsulate their own semantics; these can always be transformed into a representation which uses
vector links to link them to an other agent. Human sensory perception probes not only spacetime but additional
dimensions of description, like heat and light, etc. Tables 2 and 3 shows a few sensors that could be interpreted as
inputs to a smart civic region.
Coarse grained sensory context atom
day, time, location
light, dark, red, green, blue
hot, cold, free, constrained
salt,sweet, sour, savory, etc
open, closed
open, closed
open, closed, partial
open, closed
blocked, open, snowed under
available, unavailable
parking, parked, shunt, cruising
inside, outside
incoming, outgoing
supply, demand
energy/money usage
Person by face
Vehicle by number or appearance
Pollution cloud
Flooding
Weather
sensor type
spacetime
sight
touch
taste
building door
building window
building window blinds
city bridge
city road
city shop supplies
city transport
sense of self interior/exterior
external transport
external trade
external trade/accounting balance
spacetime pattern
spacetime pattern
spacetime pattern
spacetime pattern
spacetime pattern
Table 2: A few sensors that could feed into a large scale civic smart spaces.
Sensor
Smart phone information
Lidar
Radar
Microwave
Gyroscope
Microphone
Camera
Edge detection
GPS
Seismic
Sonar
(f)MRI
Neural net pattern recognition
CANbus
Spacetime scale / usage
Laser ranging, 100m
1000m
Astronomical and atmospheric
Orientation
Audio analysis
10m
Computer vision, object classification
positional tracking
used in oil and gas, geological sink hole discovery
deep sea
brain and body
Trainable
Table 3: A few sensors that could feed into generalized smart spaces.
102
Hierarchies of encapsulation have always been an important way of separating scales, and decoupling concerns.
It is this process of aggregation (rather than fragmentation) that enables economical scaling. Current information
technology has access to a growing array of sensors, at all scales, to measure increasingly specific semantically
distinguishable patterns. In addition to simple physical sensors for scalar promises like heat, humidity, light, etc,
agents who interface to the exterior world can also act as sensors for the interior (see figure 56), and, of course,
there are artificial information channels, including sophisticated neural networks that can be trained to recognize
almost any complicated pattern, representable as ordered collections of agents, with tensor structure, that promise
a wide range of characteristics. What we need to ask, in order to design appropriate future technologies, to exploit
information, is how should we organize the spatial interactions in such a way as to make desired outcomes and
processes efficient.
City
Exterior
building X
India
customer
law firm
business
building Y
courier
messenger
power company
Figure 56: A city as a knowledge space. A business interacting with a remote customer acts as a sensor viewing the exterior
world.
7.4.3
Spatial reasoning with concept and association structures
If a space is smart, in the sense of representing concepts and associations about its external interactions, then what
do the internal structures of a space say about those external structures, and at what scale? The sensory scales in
table 2 show examples. While each space has a maximum spatial scale, limited by its size, we could extend probes
almost indefinitely at smaller scales by equipping a space with measuring devices, or at larger scales, through
satellite observation, to create something analogous to highly extensive nervous systems65 .
Do the conceptual features of the space enable it to think for itself, or merely serve up data for human consumption? In the words of the analogy, do they have an active functional significance or merely a passive archival
significance? Do they trigger new ideas and associations (processes) by virtue of being there (see the Tsunami
example 125, where the concepts became processes)?. This is what is meant by introspection.
Example 124 (Can a smart city think?) Could a city really possess a memory analogous to a human, or a computer? Memory is everywhere: from the patterns of worn grass, recalling a hot summer in the park, to the
architecture of the buildings that recall historical and cultural periods and artefacts. Could it behave like a simple
organism? I believe this is also plausible. First, a computer memory is not like human memory: it is a linear
transcription of data, with an encoding into language, or video, or some other kind of media formatting. Computer
memory is based on a model of the sensors and retrievers. When we record or play video, there is a model of a
wide-screen television involved in the coding of data; there is no interpretation of the data. That kind of memory may exist in humans too (at least for short term memory) but beyond that, there is a long term summary of
experience that transmutes raw feeds from sensors into an approximate spacetime invariant model, using a lot of
summarization, interpretation, and conceptualization. Crucially, it is possible to mimic some of the behaviours of
65 The huge extent of human sensory nerve bundles is no doubt responsible for the powerful and complex sensations we experience, both by
direct stimulation, and via the coarser emotional states that are based on them. In simpler machines, the analogy of emotional sensations would
be limited in expression by the palette of sensory inputs it has to draw on to represent sensation.
103
human cognition to imagine ‘smart’ behaviour by imagining a series of transformations from raw sensor data to
generalized concepts and associations between them.
Tsunami
Coastal
buildings
Damage, water, earthquake
Self−defense, barriers
Flood protection
continuous learning/improvement
Figure 57: Learning about Tsunami in a smart city: flood defenses are the first line of memory that accumulates.
Example 125 (A cognitively responsive smart city?) Imagine a city by the sea, which is vulnerable to storms
and Tsunami. To image a cognitively smart city, we may ignore the humans living in it, and just think in terms of
abstract processes (see figure 57). How would the city remember a Tsunami? The sensors would be the coastal
structures, which would be flooded or destroyed. Over time, the response to learning a bout Tsunami would be the
building of flood defenses. This is cumulative learning loop, and results in a memory imprint: the flood defenses.
As these are being built, the patterns are learnt and summarized, recorded in books and documents. The events are
symbolized in movies, etc. Associations form between Tsunami and related concepts: an alarm system to trigger
flood defenses and alarms on earthquakes is built. This is the physical encoding of an association between the
concept of earthquake and Tsunami. So now there are specialized sensors capable of transmuting raw data into
learning, over time, through the construction of new hardware in the city. Memory, concepts and associations have
been encoded in the physical structures of the city.
The test of a memory lies in the recall. To be somewhat like a human memory, recall should be triggered by
the occurrence of an earthquake, or the imagination of an earthquake. So now consider what happens: when an
earthquake happens, it triggers an autonomic Tsunami warning, and flood defenses activate. This signal changes
the behaviour of the city and the processes of recall are activated. If something like an earthquake occurs, then it
might trigger the alarm too (a false alarm, to be sure, but nevertheless a semantically related event that triggers
the memory of a Tsunami). Finally, the city could simply be dreaming, with no inputs to the earthquake or tsunami
detectors: if something accidentally triggers the alarm, or if all the alarms are deliberately tested by sending a
planned theta wave of activations across the city, then the defenses are once again activated as if in rehearsal, and
the memory is recalled. The autonomic response of the city configuration is the key to triggering a memory based
on a change of configuration.
Example 126 (Memory at different scales (encapsulated specialists)) If concepts are not represented directly
by the civil engineering of a space, e.g. buildings, parks, transportation, and bridges that encodes information
about the outside, then is there documentation in processes, books and institutional habits? Would it be a smart
city that paid attention to the wear and tear of usage in the park lawns, and altered its behaviour to build paths
along these trails?
The irreducible association types map to the basic processes in spacetime:
Irreducible Association
Causal association
Containment
Molecular composition
Proximity
Semantic function in smart space
Sequential spacetime processes, data storage
Conceptual generalization and accumulation
Compound roles, identities and names
Lateral association and approximation
104
Example 127 The factories supply an external region around them, and response to their needs so the internal
state of the industrial region mirrors the sensed supply or demand.
Associations can be dynamical Stories that play out as processes
• Agents and processes associated by compatibility of promises
• Ability to align for a common purpose
• Mixing concepts to bring about innovation
• Identification of similarities alternatives (functional thesaurus)
• Speed of recognition: awareness of situation.
7.4.4
Temporal reasoning with pathways and stories
Processes and connected pathways that join concepts via associative connections, within a smart space, lead to
stories, as described in section 6.4. Stories enable semantic predictions, by relating compressed tokenized concepts.
Example 128 The arrival of a load of fish or fruit to a town market stimulates processes that activate the town
and mirror the external process of harvest.
Here are some examples:
Example 129 (Episodic story) An example story about a single agent, represented in city processes:
1. A person eats breakfast,
2. then drives to the mall,
3. then goes shopping,
4. then eats lunch,
5. then sees a movie.
An episodic story describing the world-line of an agent within the space. This represents a kind of process, but
one of limited significance to the whole city or its outside environment. The story describes the process of a part
in a machine, because it is about the orbits of only a single agent. What we could say is that the single agent
introspectively binds together the concepts of breakfast, mall, shopping, lunch and movie as a single concept,
whose significance remains to be decided.
Example 130 (An associative story) Associative stories can reveal dependencies and be useful for diagnostics,
see figure 56. For example, a law firm in the city fails to perform an important transaction because:
1. The law firm is located in building X
2. The building is connected by courier to a business client
3. The business client is located in nearby building Y ,
4. Building X depends on power from an exterior company.
5. A power failure in building Y prevents the business client from receiving word from a customer in India.
These associations are causal in nature, and lead to direct predictable consequences [118] within the city, e.g.
perhaps loss of business for the city. Still, the processes of the city do not represent the world beyond.
Example 131 (Exterior agent stories) Consider a view of a city in which the internal structures are direct representations of the outside world (see figure 58).
1. Many suppliers deliver to a supply chain hub in the city. The supply chain hub encodes the tokenized concept
of supplier/supply.
105
City
Exterior
spain
food
clothing
marketplace
supplier 1
spain goods
supply
france goods
france
food
supplier 2
transport
delivery
transport
company
Figure 58: City structures are a facsimile of the world outside, thus encoding concepts of the countries beyond in a knowledge
representation using a marketplace.
2. The suppliers are variously from France of Spain, depending on context. Simultaneous activation means
that suppliers from both France and Spain become variations of the same concept, but also that France and
Spain get represented as separate entities in the city: a French food market and a Spanish food market.
3. The generalization of these is the concept of a market.
4. The city observes that deliveries depend on the transportation of goods by lorry. The shipping company
has its own support depot and business office in the city, mirroring the external activities. Thus the memory
concept of transportation is encoded by the local office. This is associated with the supply hub, by mutual
activation, when the lorry arrives at the hub, with the logo of the transport company.
Example 132 (An interior agent story) An example story represented in city processes:
1. A person eats breakfast,
2. breakfast contained coffee
3. coffee was purchased from coffeehouse
4. coffeehouse belongs to a franchise
5. the franchise is owned by the coffee group.
This is a pure pathway in a semantic graph. The story significance of the story is in the eye of the city, e.g. it might
be of benefit in diagnosing a connection involving the person concerned to a particular batch of coffee, as a way
into an abduction process. For instance, the coffee might have made some other people ill; it would therefore be
further associated with a story about visits to a hospital for medical attention, etc.
Example 133 (A process story) The arrival of snow from outside a city is a sensory experience that could trigger
drifting and road blockage (see figure 59). That is associated with snow clearing (or perhaps later flooding). Snow
forms the context for these associated concepts and the city may recall and even enact several stories. The story
roads
drifts
clear
snow
plough
flooding
Figure 59: Sensory activations and concepts for a city associated, leading to several possible stories.
may be extrapolated further by association and generalization to make predictions like occurrences of flooding or
106
hypothermia, etc. These predictions can have actionable consequences: perhaps roads, bridges, or airports are
closed.
These examples show how the processes of agents, in a sense, tell their own stories, within the contexts of a city that
faces an outside world. People and vehicles, as mobile agents, act as messengers, binding city concepts together,
or acting as sensory stimuli. As they move, leakage and random motion around the conceptual focal points leads
to an effective blurring and association of ideas.
Irrespective of whether one buys into to notion of a smart space without information technology, it is clear
to see, from these examples, how information technologies could be used to bring plausible depth to this story,
whether for the benefit of an abstract space (region, city, building, etc), or for the people who inhabit it. By
designing appropriate sensory probes (detecting specifically interesting semantics, rather than just raw physical
observables), one gives immediate relational meaning to context and long term memory.
7.4.5
Locality, centralization, and economies of scale
Not all smart spaces are equal. Some may have more knowledge than others, but we can recognize, measure,
and perhaps improve smartness by understanding how it occurs. Concepts form localized clusters, represented as a
promise graph, via all of the agents and their interactions. In semantic knowledge space, concepts may be localized
as single concept agents. This does not imply that the physical agents, which represent such concepts, have to be
localized in a single location. Culturally, there is a tendency to centralize and focus on clear locations: cognitively,
we prefer low entropy structures to represent isolated concepts, because they are easily identifiable; this may lead
to practical bottlenecks in accessing the knowledge or location.
Contextualization has the effect of dissecting a single concept into several smaller parts. Collective centralization of any shared service or function is an expression of identity: it allows us to visit, recognize, and learn
to know things, as if they were themselves agents, with their own personality66 . Thus, humans can use the same
mechanisms that we evolved, to learn and know other humans, for learning and knowing other abstract things.
Example 134 (Where is a city’s knowledge?) Does knowledge lie in a city, or in the people who live in the city?
The answer we discern from this discussion is that it lies in both. The interactions between people and things are
part of processes that lead to new concepts and associations. As automation takes on more of the jobs that drive a
city, more of the knowledge will be encoded in processes that belong either to the city, or to a placeless repository
of information.
Easy access to functional services that we depend on suggests localization. Close proximity promotes speedy
access, but context switching can also provide fast routing to a conditional part. Connection, or adjacency, may be
arranged in two ways: either by topological and geographical proximity, or the availability of messengers to act
as go betweens. Specific concepts or functions may be available to us through either one of these mechanisms.
Communication and transport play a central role in enabling processes to associate agents.
Example 135 (Cities or communities?) Traditionally we think of cities and geographical communities, but the
development of modern communications has rendered this a less useful definition. When a city resident moves
beyond its geographical domain, does he/she leave it behind, or carry it with them? In other words should we say:
X
X
space/city =
agents ? =
locations ?
(218)
Many processes, like taxation and personal health issues, are connected to our place of residence. What happens
when we are no longer within our home geo-political domain? How distributed can a representation of a process
be? Could we transform a static adjacency into a mobile utility? It is interesting to speculate on how smart
adaptations could evolve the city’s sense of awareness to cope with this.
Example 136 (Partitioning: rooms and districts) Limiting concept scope helps us to grasp concepts with less
processing. We make buildings for specific purposes: shops, houses, museums, offices, etc; we also make districts
around specific economies of scale: shopping malls, financial districts, red light districts, etc. As we embed a
concept in a larger framework, i.e. as a region of space gets bigger, central points become harder to visit due to
congestion of pathways, unless there are direct routes. Scale therefore plays a deciding role in identifying concepts
and making use of knowledge. The Dunbar limits place bounds on how large human work groups can become
while remaining effective.
66 This is how humans ‘know’ each other and know things: we form a relationships with them as if they were characters in a story, hence our
tendency to anthropomorphize concepts.
107
Example 137 (Implicit learning in a space) On a street, in jazz city, during the context of daytime, the sun shines
on the north side of the street, and south side is in shadow all day long. Over time, sun worshipers gravitate to the
sunny side of the street, while night birds flock to the shady side. Thus the street remembers the concept of sun, and
associates shops, clubs, and services with the happy sun people on the north side, by concurrent activation.
Example 138 (Economies of scale in smart cities) Recent work on what makes cities work and grow identified
the economics of infrastructure at a key part of the scaling [123, 124]. What smart knowledge base could assist in
facilitating economies of scale in a city?
Economies of scale relate to specific goods or services. To scale these by introducing more agents involved
in the keeping of promises, some measure of coordination is required. A calibrated source concept provides the
semantic basis for making a quantitative promise. The association is semantic but the physics lies in the real world
representation.
7.4.6
Central controller versus brains
For many, the intuition of what ‘smart’ means suggests a reasoning brain’, or some central guiding intelligence.
We imitate some of the the symptoms of human intelligence by using algorithms and processes, with the help of
machinery that is mechanical or computational, as if trying to place a homunculus within as the deus ex machina.
This form of mimicry uses spatiotemporal memory to emulate behaviour through processes rather than through
static structures. Our attachment to brains is clearly a liability in automating processes. Human brains are quite
over-qualified for many of the tasks to which we apply ourselves: far simpler mechanisms can easily cope with the
state spaces and constraints of isolated processes in, say, manufacturing or quasi-stable repetition (e.g. autopilots).
If smart implies quick and timesaving, then devices much simpler than ourselves may be called smart.
Turing’s halting theorem tells us that we can simulate any computable result with a brain with a central Turing
machine eventually, but this is not to say that a brain is a Turing machine, works anything like one, or that such
a simulation would work at even close to the same timescale67 . Of intelligence, we expect agents that adapt
continuously and cumulatively to their context, or have awareness of sufficient local state. The emergent behaviour
of an immune system exhibits smart behaviour, repairing us when we are sick and mostly sleeping when we are
well, yet it is the most decentralized of systems in our bodies [90].
Example 139 (IBM’s True North ‘cognitive computing’ architecture) In 2014, IBM introduced an extensible
hardware programmable chip architecture called True North [125], with a parallel cloud simulator ‘compass’.
The modular design and inbuilt routing capability allows users to build massive deep layered neural networks
with a quasi-deterministic event driven input scheme based on data oriented signal agents called ‘spikes’. As
a programmable fabric, this architecture enables massive learning capability, with connectivity approaching the
scale of a human brain, and at low power. So far demonstrations have been limited to quite simple, traditionally
recognizable tasks. Crucially, this architecture shows how non-von Neumann architecture computational fabrics
may be built as relatively dumb (but programmable) semantics spacetimes at a low level.
Example 140 (Smart organization) Many firms adopt hierarchical forms of organization. Large firms may have
up to 10 or more layers of management, arranged in a hierarchical tree. Can a management hierarchy act as a
trained neural network, i.e. a learning space that has its own knowledge representation?
The structure shown in figure 60 is drawn deliberately to suggest a layered neural network structure, with the
‘top’ of the company to the left. Customer demand is the sensory apparatus of the organization, which enters
through top level management and sales (the customer facing part of the organization). The inputs get handed
down through product managers and developers to final outputs, which represent the products or services of the
company. These, in turn, feedback back around to the inputs allowing the company to ‘think’ as a continuous
iteration of products and services68 This has the superficial structure of a semantic learning network, in which
concepts are represented both by agents and clusters of agents. Unlike what one would normally consider to be
a neural network, the agents are humans (or perhaps processes), which are themselves already quite complex
networks, so the learning capabilities are quite different to an artificial neural network structure, made from dumb
agents.
The assumption here is that there are tight horizontal connections, and no vertical connections in the layers.
In a real organization, the layers also collaborate as teams, not as individual nodes. This may help or hinder the
learning. Since, in a human organization, one assumes that the nodes are not dumb threshold devices, but contain
67 Ironically, once we’ve finished constructing automata to mimic smart behaviour, the proxies which enact programmed patterns often seem
very dumb to us, because they are frozen into behaviours that are little more sophisticated than the opening and closing of a flower in response
to sunlight.
68 In modern parlance, this process is called Continuous Delivery [126].
108
significant memory each, there is useful equilibration and redundancy by having intra-layer connections. What we
must have is stronger interlayer connections to forward learning and connect inputs to outputs.
The hierarchical structure represents a directional routing structure, or an aggregator of information, if the
agents promise to faithfully forward information up or down the hierarchy. In some companies the elements of
the layers do not map to all the elements in the next layer. There are branches for ‘separation of concerns’,
sometimes called a silos, in which each branch attempts to represent a different concept. From a semantic space
viewpoint, this is a mistake, as it tries to hardwire concepts rather than learning them. Concepts like ‘customer
type’, ‘trends’, ‘competitors’, etc, may be distributed throughout the layers in a neural network (without necessarily
being named explicitly). In a branching hierarchy, the associations between layers are encoded only through
their ‘direct reports’ or managers, leading to association pattern (b) in figure 30, which is the observation of a
coincidental correlation if a branch manager imposes coordination on the sub-managers, and may be reported
calibration (d) if the sub-managers promise an agreement with their mutual manager What is missing from a silo
developers
CEO
operators
product
manager
developers
customer demand
SALES
CTO
product
operators
outputs
product
manager
operators
developers
developers
Figure 60: Is a hierarchical organization like a learning semantic space? The structure is potentially similar if there are strong
links between layers, without siloing.
topology, or purely branching network (which is a classic computer science class hierarchy or spanning tree), is
the ability to learn and freely associate ideas. All ideas are related to a single concept representation, which is
reified as the top manager (CEO).
Example 141 (CRISPR gene editing) Clustered regularly inter-spaced short palindromic repeats (CRISPR) is a
gene editing technique for scanning and locating specific codon sequences in the genome. With centralized intent
at a large scale, and distributed cells containing centralized DNA at a small scale, decentralized molecular agents
attached to immune cells can edit the genomes of diseased cells, in a semantically specific way [127].
7.4.7
Innovation in smart spaces
The processes of innovation haven been studied from a promise theory perspective elsewhere [128]. The key
elements of innovation are the random mixing of concepts, followed by outcome gestation and selection. Mixing
of ideas can be a quick process, catalyzed by accessible localization (e.g. centralization); the latter is a slower
process, in which something is built from the new plan, functional parts align with a recombined mixture of
promised concepts, and it is then subjected to a selection process (testing), in order to judge its fitness for purpose
(see figure 61)69 .
In a smart space, the agents that define the space innovate by interacting with their exterior to learn new
ideas. This might be the occupants of the space interacting with it, or the environment beyond it. To facilitate
this interaction, catalysts can be provided to match-make the association of compatible concepts. The efficiency
of this process is accelerated by a large contact surface, so larger spaces could be smarter than smaller ones.
Matchmakers and go-betweens thus have roles to play in making a city or space smart. This kind of ‘thinking’ is
what a knowledge space does when it focuses more on acquired memories and attempts to combine them into new
ones.
• Spacetime mixing of agents is an important part of recombination. Mixing is most effective in a gas phase,
provided agents are brought together to form combinations, but could also be mediated ‘pollenated’ in a
solid phase by messengers.
69 This
is functionally equivalent to the mixing of DNA to form a zygote.
109
compound
concept
concept
catalyst
idea 1
idea 2
Figure 61: Innovation happens by mixing ideas.
• Outcome selection requires an oriented notion of a goal, or promised outcome and a judging agent, or simply
failure.
Could an inanimate space (at any scale) help humans or organizations work together, just by virtue of how it
bounds, facilitates, and shapes their interactions?
Independent researcher Meredith Belbin has studied human collaboration through teamwork, and has identified
nine abilities or roles (kinds of promise) to be played in a team collaboration (regardless of how many people
there are in the team) [129]. In Belbin’s view, human dynamics require roles for: plant (a creative ideas person
who solves problems), shaper (with the drive and courage to overcome obstacles), specialist, implementer (a
practical thinker), a resource investigator (who knows where to find help), a coordinator (an arbitrator), a monitor
(who judges progress, team worker (someone concerned with the team’s inter-personal relationships), and finisher
(someone analytical who does quality assurance). Belbin does not claim that all these roles must be taken on by
different agents (people in this case), but that these functional roles define the promises that need to be kept for an
effective teamwork. So, should we expect to find these promised roles in smart spaces too? In terms of spacetime
processes, these incorporate learning or memory (specialist), mixing (plant, coordinator, finisher). the limitation of
Belbin’s model lies in assuming human involvement, and human psychology, the following of a plan, and a notion
of progress towards a goal.
Example 142 (Revised simple view of a smart company) In example 140, we represented an organization in
the form of a neural network, deliberately, in order to contrast a simple flow discrimination network with an
organizational hierarchy. Now let’s make a more realistic sketch. A company has an outside world, with users,
customers, competitors, which it interacts with through sensor-actuators, represented roughly by sales and support.
These sensors forward their observations to the internal representations of the company: the people who remember
the relationships, including developers and operations people, i.e. the memory agents (see figure 62).
Unlike the previous example of the flow network, where agents were imagined as driving a company from a
helm, this model is more fatalistic: inputs are passed to both developer and operations agents to be interpreted.
Neither is above the other in a preferential hierarchy. Data from the outside world about successes and failures
are applicable to both equally, and they will also share associated impulses with one another.
The agents in a company are not dumb switches, they are already significantly established networks of knowledge themselves. The training of the nodes over time corresponds to their schooling and experience in doing their
jobs, as well as adapting and modernizing individually. A hybrid structure like this cannot be understood in precisely the same terms as an atomic knowledge representation: it is a hierarchy of superagents, each of which is
a learning knowledge representation. The question is whether the organization as a cooperative whole may be
considered a learning knowledge space. The stability of the organization depends on it being able to handle the
perturbations from the sensors (fix bugs and operational issues, i.e. keep its business promises) much faster than
the time it takes to train the internal nodes in their basic functions of interpreting the sensors. However, there is a
second rate of learning, which is the operational knowledge, learnt ‘on the job’, which is formed in real time from
context activations. This fits with the model of knowledge representation presented in section 5, but it does not fit
well with a neural network discriminator structure, since the latter is mainly about recognition of known patterns.
Even these simple considerations seem to indicate that a company, which interacts with the world, cannot be
steered by a single helm, in a quasi-deterministic hierarchy, else it will be unable to represent the operational
knowledge it acquires through interactions. A flatter organization, where all knowledge workers are closer to their
customers leads to a more reliable knowledge retention.
110
competitors
buyers
users
sales
requests
SENSORS
bug
tickets
KNOWLEDGE REPRESENTATION
ops
dev
concepts
Figure 62: A company as a knowledge represent ion.
7.4.8
Auto-reasoning and searching for answers
The ability to construct a coherent narrative from a collection of parts is what we call reasoning. This could be
addressed in a tokenized (story) representation or in an implicit learned network representation such as example
137. The aim of a smart space would be to improve the likelihood that a space, when faced with a challenge or
question, would automatically evolve towards a state representing a reasoned outcome. This need not guarantee
the correctness of an answer, only the convergence towards the best answer it could find. For example, in the case
of the Tsunami challenge (example 57), the responses of the city unleashed a chain of processes that automatically
protected the city. This exhibited a primitive form of autonomous reasoning. Reasoning that happens on a larger
scale would not be clearly visible to the occupants of a city, and may take place on a much slower timescale than
they normally observe. Thus, a smart city might not be apparently smart to human concerns.
The use of information technology to enhance spaces, using special sensors and training, is clearly a popular
idea today. In recent years, it has also been popular to answer questions by trawling ‘big data’, i.e. unspecific
sources of sensory information collected from a wide range of sources, in hope of discovering interest concepts
and relationships by introspection, whence to pursue reasoning based on these. Finding answers is a form of
pattern recognition, in semantic patterns (stories and narratives) built from data patterns that reveal interpretations
(concepts).
Are we making the best use of knowledge? People, machines, spaces, and communities all learn, but not all
do much to propagate or facilitate the application of their expertise. The smartest part of searching lies in setting
up the boundary conditions, i.e. in posing the right question: what pattern are we looking for, and at what scale?
Scale plays a special role in reasoning. If we search for a structure that operates at a scale much larger than our
sensory experience (like a city), it is entirely possible that behaviours might emerge that we would be entirely
oblivious to, because the benefits are not directly apparent at the scale of our own sensory apparatus. Generalized
clustering techniques, like Principal Component Analysis [17,130] can find conspicuous patterns, but cannot attach
significance to them. It is the latter process which is the most expensive long term approach. Intelligent assistants
might identify relationships that we missed with our limited human faculties, but human expertise can offer helpful
input without having to go through expensive training, using a semantic network or topic map approach.
8
Summary
Parsifal: Ich schreite kaum, doch wähn’ ich mich schon weit.
(I scarcely move, yet already it seems I have travelled far.)
Gurnemanz: Du sieh’st, mein Sohn, zum Raum wird hier die Zeit.
(You see, my son, here time becomes space.)
– R. Wagner, Parsefal
In these notes, I have tried to sketch out how to apply spacetime concepts to practical questions about the
semantics of data, and how data become knowledge. The aim is not to examine a particular approach to knowledge
representation or learning, or to describe and test a particular technology, but to rather expose the general structural
aspects, using a minimum of assumptions based on Promise Theory. This includes the emergence of what we know
as concepts and how they network into larger knowledge representations.
This effort began, motivated by observations concerning IT infrastructure [30, 87, 118, 119], and this later
carried over into the possibilities for future large-scale infrastructure. It has since assumed much wider scope.
111
Throughout the notes, we have been homing in on the conclusion that knowledge is represented principally by
concepts encoded through patterns that manipulate the various aspects of spacetime symmetries. This suggests that
information and knowledge representation are fundamentally connected to the structure of semantic spacetime,
and few other concepts are really needed (see section 5.4.10). This simple insight takes one remarkably far,
and qualifies as a basic law: ‘All representations of knowledge have semantic content proportional to the loss of
spacetime symmetry.
The detailed mechanisms for encoding of concepts, associations, and semantics can be identified with spacetime scaling, directed adjacency, and boundary interactions. A key requisite is that there be memory, with predictable and repeatable addressability over distinguishable spacetime locations. Addressability is needed for semantic continuity: recalling and updating knowledge. Although a database is a semantic space, knowledge is not
a database, but a continuous process, and a clean separation of timescales is essential for the dynamical stability of
that process (see figure 63).
T
slow
adaptation
T
Timescale
stories
T
learn
T
sample
~ Tcontext
T
token
~ Tknow
Tattractor
T
pattern
T
concurrent
fast
knowledge/learning
real time
stories
aggregation/semantics/invariance
Figure 63: Timescales overlaid onto a scale/invariance diagram.
A number of hypotheses emerged from basic promise theoretic considerations during this study:
1. The weights or significances of patterns are in inverse relation to intrinsic information of their representation.
2. Usable knowledge requires a minimum level of dynamical stabilty.
3. Context must be evaluated faster and promises less information than the derived knowledge it modulates.
4. Context evaluation suggests coordination by an independent agent.
5. Knowledge maintenance, by partially annealing of learning weights, is needed for long-term stability.
6. Story inference and navigation suggest coordination by an independent agent.
The representation of concepts and associative pathways, as spacetime structures, builds on basic irreducible associations of spacetime:
• Nested containment (scaling).
• Causal history (pre-requisite dependency).
• Composition, i.e. cooperative refinement.
• Similarity (proximity).
With the benefit of a semantic spacetime viewpoint, many apparently complex semantic issues reduce to combinations of these spacetime associations, and effectively a memory of what, where, and when.
112
Semantic tokenization is key to compressing cognition. We are most familiar with this in the context of language representation. In stable situations, we can achieve this tokenization with high levels of probability, but that
is a best case scenario. The ability to learn and crystallize concepts from ‘big’ noisy data, is strongly constrained
by the capabilities of the memory used, and hence the properties of the semantic space. Tokenization solves three
issues:
• Reduction of the dimensionality of sensory data (token formation).
• Compression of sensory information to a symbolic form of low information (context).
• Mapping variable inputs to an invariant representation (conceptualization).
This tokenization, which emerges as central to concepts, is not intentionally a ‘linguistic’ construction. It emerges
from the sequential nature of coarse grained input (both of which are just spacetime concepts). However, I believe
that the resemblence to language is not spurious. Tokens or symbols emerge from concurrent semantics of a sensory
stream. The discreteness of symbols emerges from the approximation of concurrency across multiple sensory
threads. Simultaneity of semantics is an implicit seed for binding sensory properties into prototype concepts.
The result is that there is a mutual affinity between this view of knowledge representation and familiar models of
linguistics (cognitive linguistics on the interior representation and generative linguistics on the exterior).
Present day artificial intelligence research has most recently focused on the advances in neural network for
‘deep learning’ [52, 131, 132], where one uses directional networks to learn patterns, such that weighted connections encode pre-trained semantics. While fascinating, to be sure, this approach may not be the cheapest route
to technological solutions: machine learning ignores several issues, like the problem of explicit tokenization, and
the opportunities to merge human and machine knowledge, in order to focus on trying to engineer its emergence.
It makes weak contact with existing human concepts, like knowledge bases and libraries, and then only through
arduous and expensive training. There remains the possibility of improving on this, by direct human input at the
symbolic level (like words of encouragement from a teacher), in order to imagine ‘cyborg’ systems, at all scales,
which might make the best use of both humans and machinery. The spacetime approach described here may help
to put exploit these pieces and bring them together.
Although it would be impossible to bring rigour to a subject of this scope, in a single series of notes, no matter
how detailed, I hope that there is sufficient material here to bring some clarity to a few of the larger questions
around artificial reasoning, and help formulate more of the key questions theoretically. I find it compelling that
the most fundamental of spacetime notions leads to a simple unified way of framing the subject across multiple
disciplines.
This completes the initial overview of semantic spaces as representations for organization and cognitive reasoning, and perhaps opens a door to a more complete notion of artificial intelligence. When all is done, the simple
outcome of this work seems to be that our most basic learning and abstraction capabilities have their origins,
plausibly, in the fundamental structures and processes of space and time.
Acknowledgement
I am grateful to Steve Pepper for getting me interested in many of the topics covered within this paper, and also to
Alva Couch with whom I co-developed the notion of stories and inference between 2007-2008, for a draft reading
of the manuscript. The writing of this work has been partly sponsored by Cisco Systems in 2016. None of these
necessarily endorse or agree with the presentation here.
A
Coordinatization of data streams
This shows a coordinatization of a multi-dimensional document, which could be used to describe and document
its interpretation, without alteration, post hoc.
{
"kind": "Service",
"apiVersion": "v1",
"metadata": {
"name": "myapp"
},
"spec": {
"ports": [{
113
"port": 8765,
"targetPort": 9376
}],
"selector": {
"app": "example"
},
"type": "LoadBalancer"
}
}
Dimension/Region ( 1,
0,
0,
0,
0, )
->
Dimension/Region ( 2, 1{, 0, 0, 0, ) ->
- path/proper time location[0]("kind")
- path/proper time location[1](:)
- path/proper time location[2]("Service")
- path/proper time location[3]("apiVersion")
- path/proper time location[4](:)
- path/proper time location[5]("v1")
- path/proper time location[6]("metadata")
- path/proper time location[7](:)
Dimension/Region ( 2, 2{, 1{, 0, 0, )
- path/proper time location[0]("name")
- path/proper time location[1](:)
- path/proper time location[2]("myapp")
Dimension/Region ( 2, 3{, 0, 0, 0, )
- path/proper time location[0]("spec")
- path/proper time location[1](:)
->
->
Dimension/Region ( 2, 4{, 1{, 0, 0, )
- path/proper time location[0]("ports")
- path/proper time location[1](:)
->
Dimension/Region ( 2, 4{, 2{, 1[, 1{, ) ->
- path/proper time location[0]("port")
- path/proper time location[1](:)
- path/proper time location[2](8765)
- path/proper time location[3]("targetPort")
- path/proper time location[4](:)
- path/proper time location[5](9376)
Dimension/Region ( 2, 4{, 3{, 0, 0, ) ->
- path/proper time location[0]("selector")
- path/proper time location[1](:)
Dimension/Region ( 2, 4{, 4{, 1{, 0, )
- path/proper time location[0]("app")
- path/proper time location[1](:)
- path/proper time location[2]("example")
->
Dimension/Region ( 2, 4{, 5{, 0, 0, ) ->
- path/proper time location[0]("type")
- path/proper time location[1](:)
- path/proper time location[2]("LoadBalancer")
Dimension/Region ( 2,
5{,
0,
0,
0, )
->
Dimension/Region ( 3,
0,
0,
0,
0, )
->
Dimension/Region ( 4,
0,
0,
0,
0, )
->
114
- path/proper time location[0](@)
A mixed format:
{{{one { two, three } four <alpha> text ... </alpha> }
\section{section1}xx
Nothing happenings
\section{section2}
[windows1]
<greek>
<alpha>
AAA
</alpha>
<beta>
BBB
</beta>
<gamma>
CCC
</gamma>
</greek>
[windows2]
xiug
Dimension/Region ( 1,
0,
0,
0,
0, )
Dimension/Region ( 2, 1{, 1{, 1{,
- path/proper time location[0](one)
->
0, )
->
Dimension/Region ( 2, 1{, 1{, 2{, 1{, )
- path/proper time location[0](two)
- path/proper time location[1](three)
->
Dimension/Region ( 2, 1{, 1{, 3{, 0, )
- path/proper time location[0](four)
->
Dimension/Region ( 2, 1{, 1{, 4{, 1<alpha>, )
- path/proper time location[0](text)
- path/proper time location[1](...)
Dimension/Region ( 2,
1{,
1{,
5{,
0, )
Dimension/Region ( 2, 1{, 2{, 0, 0, )
- path/proper time location[0](\section)
->
->
Dimension/Region ( 2, 1{, 3{, 1{, 0, )
- path/proper time location[0](section1)
->
Dimension/Region ( 2, 1{, 4{, 0, 0, ) ->
- path/proper time location[0](xx)
- path/proper time location[1](Nothing)
- path/proper time location[2](happenings)
- path/proper time location[3](\section)
Dimension/Region ( 2,
1{,
5{,
1{,
0, )
115
->
->
- path/proper time location[0](section2)
Dimension/Region ( 2,
1{,
6{,
0,
0, )
->
Dimension/Region ( 2, 1{, 7{, 1[, 0, )
- path/proper time location[0](windows1)
Dimension/Region ( 2,
1{,
8{,
0,
0, )
Dimension/Region ( 2,
1{,
9{,
1<greek>,
->
->
0, )
->
Dimension/Region ( 2, 1{, 9{, 2<greek>,
- path/proper time location[0](AAA)
1<alpha>, )
Dimension/Region ( 2,
0, )
1{,
9{,
3<greek>,
->
Dimension/Region ( 2, 1{, 9{, 4<greek>,
- path/proper time location[0](BBB)
1<beta>, )
Dimension/Region ( 2,
0, )
1{,
9{,
5<greek>,
1<gamma>, )
Dimension/Region ( 2,
1{,
9{,
0, )
Dimension/Region ( 2,
1{,
10{,
0,
0, )
Dimension/Region ( 2, 1{, 11{, 1[, 0, )
- path/proper time location[0](windows2)
Dimension/Region ( 2, 1{, 12{, 0, 0, )
- path/proper time location[0](xiug)
->
->
Dimension/Region ( 2, 1{, 9{, 6<greek>,
- path/proper time location[0](CCC)
7<greek>,
->
->
->
->
->
->
Each of the addressable sections of this sensory data stream can now be annotated with a specific meaning, either
by representing as a pattern or as a specific address in the data ‘document’.
B
Invariant representation by spacetime aggregation example
Figure 64 shows a scale/invariance diagram for a hypothetical cloud system monitoring system. Data are collected
from sensors, described at the bottom left of the diagram, on individual computing nodes (where only one is
shown). The data are aggregated and memorized locally, as timeseries or state caches in order to characterize a
stable (and thus approximately invariant) normal state. Changes to what has been learned to be stable normal state
are classified as anomalous high or low deviations, which are themselves invariant characterizations of a changing
state.
At any given time, invariant characteristics may or may not be active, depending on context. The tokenization
of context must therefore be complicit in the invariant representation. Invariance does not mean something which
is unable to change dynamically, but rather stability of interpretation. As we move to the right of the diagram, the
concepts become ‘larger’ in scale (space and time) and thus more stable semantically and dynamically. They accrete more support from sensory data, derivative characteristics and even external annotation (such as that provided
by the technique shown in the previous appendix). Overlapping concepts, taken in clusters may represent specific
phenomena, e.g. thrashing be exhibited by a change in load average, free memory, and several other characteristics dynamical deviations. The roles played by change, in delimiting regions of stabilty, are likely key to finding
semantics boundaries. Each independent grouping or clustering of agents can be given a new name and represent
a new concept, following the indentification of autonomous agents in promise theory.
116
sensors
memory
aggregate state
containers
derivatives
slow
weather
users
usage memory
user groups
users_high
Timescale
users_low
processes
process groups
ecosystems
efficiency
function/service
host groups
processes_high
hostX
processes_low
temperature
free_mem_high
free_mem_low
free_memory
freemem_timeseries
load_average
io_writes
load_av_timeseries
thrashing
load_av_high
load_av_low
cpu
fast
sensor data
aggregation/semantics/invariance
concepts
Figure 64: Scale/invariance diagram example.
References
[1] M. Burgess. Spacetimes with semantics (i).
http://arxiv.org/abs/1411.5563, 2014.
[2] M. Burgess. Spacetimes with semantics (ii).
http://arxiv.org/abs/1505.01716, 2015.
[3] M. Burgess. In Search of Certainty: the science of our information infrastructure. Xtaxis Press, 2013.
[4] Office of Government Commerce, editor. Best Practice for Service Support. ITIL: The Key to Managing IT
Services. The Stationary Office, London, 2000.
[5] J.A. Bergstra and M. Burgess. Local and global trust based on the concept of promises. Technical report,
arXiv.org/abs/0912.4637 [cs.MA], 2006.
[6] J.A. Bergstra and M. Burgess. Promise Theory: Principles and Applications. χtAxis Press, 2014.
[7] M. Burgess. On the theory of system administration. Science of Computer Programming, 49:1, 2003.
[8] T.M. Cover and J.A. Thomas. Elements of Information Theory. (J.Wiley & Sons., New York), 1991.
[9] C.E. Shannon and W. Weaver. The mathematical theory of communication. University of Illinois Press,
Urbana, 1949.
[10] D.C. Dennet. Consciousness Explained. Allen Lane (London), 1992.
[11] A. Einstein. Relativity, The Special and General Theory. Methuen & Co., 1920.
[12] Ruslan Salakhutdinov and Geoffrey Hinton. Semantic hashing. Int. J. Approx. Reasoning, 50(7):969–978,
July 2009.
[13] Qifan Wang, Dan Zhang, and Luo Si. Semantic hashing using tags and topic modeling. In Proceedings
of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval,
SIGIR ’13, pages 213–222, New York, NY, USA, 2013. ACM.
[14] M. Burgess. New Research on Knowledge Management Models and Methods., chapter What’s wrong with
knowledge management? The emergence of ontology. InTech, 2012.
117
[15] L. Zhang, A. Afanasyev, J. Burke, V. Jacobson, K.C. Claffy, P. Crowley, C. Papadopoulos, L. Wang, and
B. Zhang. Named data networking. SIGCOMM Comput. Commun. Rev., 44(3):66–73, July 2014.
[16] G.P. Mishra and M. Dave. A review on content centric networking and caching strategies. In Communication
Systems and Network Technologies (CSNT), 2015 Fifth International Conference on, pages 925–929, April
2015.
[17] R.O. Duda, P.E. Hart, and D.G. Stork. Pattern classification. Wiley Interscience, New York, 2001.
[18] A.G. Huth, W.A. de Heer, T.L. Griffiths, F.E. Theunissen, and J.L. Gallant. Natural speech reveals the
semantic maps that tile human cerebral cortex. Nature, 532:453458, 2016.
[19] J. Hawkins and S. Blakeslee. On Intelligence. St. Martin’s Griffin, 2004.
[20] R.W. Langacker. Cognitive Grammar, A Basic Introduction. Oxford, Oxford, 2008.
[21] D.J. Watts. Small Worlds. (Princeton University Press, Princeton), 1999.
[22] C. Berge. The Theory of Graphs. (Dover, New York), 2001.
[23] M. Burgess and A. Couch. On system rollback and totalized fields: An algebraic approach to system change.
J. Log. Algebr. Program., 80(8):427–443, 2011.
[24] S. Weinberg. Gravitation and cosmology: principles and applications of the general theory of relativity.
Wiley, 1972.
[25] S. PalChaudhuri, S. Du, A.K. Saha, and D.B. Johnson. Treecast: a stateless addressing and routing architecture for sensor networks. In Proceedings of the 4th international workshop on algorithms for wireless,
mobile, and ad hoc sensor networks (WMAN), 2004.
[26] R.C. Doss, D. Chandra, L. Pan, W. Zhou, and Mu:u Chowdhury. Dynamic addressing in wireless sensor
networks without location awareness. Journal of Information Science and Engineering, 26:443–460, 2010.
[27] Unpublished reports on earth computing.
[28] G. Box, G. Jenkins, and G. Reinsel. Time Series Analysis. Prentice Hall, New Jersey, 1994.
[29] P. Hoogenboom and J. Lepreau. Computer system performance problem detection using time series models.
Proceedings of the USENIX Technical Conference, (USENIX Association: Berkeley, CA), page 15, 1993.
[30] M. Burgess. Two dimensional time-series for anomaly detection and regulation in adaptive systems. Lecture
Notes in Computer Science, IFIP/IEEE 13th International Workshop on Distributed Systems: Operations
and Management (DSOM 2002), 2506:169, 2002.
[31] N.K. Jerne. The generative grammar of the immune system. Nobel lecture, 1964.
[32] H. Lewis and C. Papadimitriou. Elements of the Theory of Computation, Second edition. Prentice Hall, New
York, 1997.
[33] D. Wood. Theory of Computation. J. Wiley and Sons, 1987.
[34] N. Chomsky. On certain formal properties of grammars. Information and Control, 2(2):137–167, 1959.
[35] C.J. Date. Introduction to Database Systems (7th edition). Addison Wesley, Reading, MA, 1999.
[36] G.R. Grimmett and D.R. Stirzaker. Probability and random processes (3rd edition). Oxford scientific
publications, Oxford, 2001.
[37] I. Heim and A. Kratzer. Semantics in Generative Grammar. Blackwell, 1998.
[38] R. Quian Quiroga, L. Reddy, G Kreiman, C. Koch, and I. Fried. Invariant visual representation by single
neurons in the human brain. Nature, 435:1102–1107, 2005.
[39] T.H. Stang, F. Pourbayat, M. Burgess, G. Canright, K. Engø, and Å. Weltzien. Archipelago: A network security analysis tool. In Proceedings of The 17th Annual Large Installation Systems Administration Conference
(LISA 2003), San Diego, California, USA, October 2003.
118
[40] J. Strassner. Handbook of Network and System Administration, chapter Knowledge Engineering Using
Ontologies. Elsevier Handbook, 2007.
[41] J. Bjelland, M. Burgess, G. Canright, and K. Eng-Monsen. Eigenvectors of directed graphs and importance
scores: dominance, t-rank, and sink remedies. Data Mining and Knowledge Discovery, 20(1):98–151, 2010.
[42] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web.
Technical report, Stanford University, Stanford, CA, 1998.
[43] P. Arrighi and G. Dowek. Causal graph dynamics. arXiv:1202.1098 [cs.DM], 2012.
[44] K.B. Korb and A.E. Nicholson. Bayesian Artificial Intelligence. Chapman and Hall, 2004.
[45] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgen Kaufmann, San Francisco, 1988.
[46] W.R. Ashby. An introduction to cybernetics. J. Wiley & Sons, 1956.
[47] R. Desimone. Face-selective cells in the temporal cortex of monkeys. Journal of Cognitive Neuroscience,
3(1):1–8, 1991.
[48] T.L. Bjerknes, E.I. Moser, and M.B. Moser. Representation of geometric borders in the developing rat.
Neuron, 2014.
[49] M.B. Moser, D.C. Rowland, and E.I. Moser. Place cells, grid cells, and memory. Perspectives in Biology,
2016.
[50] C.F. Doeller, C. Barry, and N. Burgess. Evidence for grid cells in a human memory network. Nature,
463:657–661, 2010.
[51] M.D. Fox, A.Z. Snyder, J.L. Vincent, M. Corbetta, D.C. Van Essen, and M.E. Raichle. The human brain
is intrinsically organized into dynamic, anticorrelated functional networks. Proceedings of the National
Academy of Sciences of the United States of America, 102(27):9673–9678, 2005.
[52] B. Peng, H. Li Z. Lu, and K.F. Wong. Towards neural network-based reasoning. arXiv:1508.05508 [cs.AI],
2015.
[53] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio. Empirical evaluation of gated recurrent neural networks on
sequence modeling. arXiv:1412.3555 [cs.NE], 2014.
[54] J. Weston, A. Bordes, S. Chopra, A.M. Rush, A. Joulin B. van Merrienboer, and T. Mikolov. Towards
ai-complete question answering: A set of prerequisite toy tasks. arXiv:1502.05698v10 [cs.AI], 2015.
[55] M. Burgess. Classical Covariant Fields. Cambridge University Press, Cambridge, 2002.
[56] R.P. Feynamn. Space-time approach to quantum electrodynamics. Physical Review, 76:769, 1949.
[57] J.A. Wheeler and R.P. Feynman. Interaction with the absorber as the mechanism of radiation. Reviews of
Modern Physics, 17:157–161, 1945.
[58] J. Schwinger. Particles, Sources and Fields (Volume 1). (Addison Wesley, California), 1970.
[59] R.E. Kastner. The Transactional Interpretation of Quantum Mechanics. Cambridge, 2013.
[60] X. Zhao et. al. Toward a formal theory of belief, capability, and promise incorporating temporal aspect.
LNAI (CEEMAS 2005), 3690:296–305, 2005.
[61] J. Hawkins and D. Georgge. Hierarchical temporal memory: Concepts, theory, and terminology. Technical
report, Numenta Inc, 2006.
[62] J.C. Dainty, editor. Current Trends in Optics. Academic Press, 1994.
[63] D. Purves, G.J. Augustine, and D. Fitzpatrick et al., editors. Temporal Categories of Memory., volume
Neuroscience. 2nd edition. Sinauer Associates, 2001.
[64] M. Burgess. Analytical Network and System Administration — Managing Human-Computer Systems. J.
Wiley & Sons, Chichester, 2004.
119
[65] E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence: From Natural to Artificial Systems. Oxford
University Press, Oxford, 1999.
[66] J. Choi, J. Han, E. Cho, T. Kwon, and Y. Choi. A survey on content-oriented networking for efficient content
delivery. IEEE Communications Magazine, 49(3):121–127, March 2011.
[67] M. Burgess and K. Begnum. Voluntary cooperation in a pervasive computing environment. Proceedings of
the Nineteenth Systems Administration Conference (LISA XIX) (USENIX Association: Berkeley, CA), page
143, 2005.
[68] D. Aredo, M. Burgess, and S. Hagen. A promise theory view on the policies of object orientation and the
service oriented architecture. In submitted to Science of Computer Programming, 2006.
[69] D. Aredo and M. Burgess. On the consistency of distributed knowledge. In Proceedings of MACE 2007,
volume 6 of Multicon Lecture Notes. Multicon Verlag, 2007.
[70] P. Borrill, M. Burgess, T. Craw, and M. Dvorkin. A promise theory perspective on data networks. CoRR,
abs/1405.2627, 2014.
[71] P. Kanerva. Sparse Distributed Memory. MIT Press, 1988.
[72] P.J. Denning. Sparse distributed memory. American Scientist, pages 333–335, 1989.
[73] T. Konopka, F. Markopoulou, and L. Smolin. Quantum graphity. arXiv:hep-th/0611197v1, 2006.
[74] P. Arrighi, S. Martiel, and S. Perdrix. Reversible graph dynamics. arXiv:1502.04368 [cs.DM], 2016.
[75] Universality of causal graph dynamics. Universality of graph dynamics. arXiv:1602.01352 [cs.DM], 2016.
[76] N.A. Barricelli. Numerical testing of evolution theories : Part i theoretical introduction and basic tests. Acta
Biotheoretica, 16:6998, 1962.
[77] S. Wolfram. Statistical mechanics of cellular automata. Rev. Mod. Physics, 55:601, 1983.
[78] S. Wolfram. A New Kind of Science. Wolfram Media, 2002.
[79] M. Burgess. A promise theory approach to understanding resilience: faults, errors, and tolerance within
systems. Technical report, Available at markburgess.org, 2015-2016.
[80] X. Li and A.G. Yeh. Neural-network-based cellular automata for simulating multiple land use changes using
gis. Int. J. Geographical Information Science, 16(4):323–343, 2002.
[81] R. White and G. Engelen. Cellular automata and fractal urban form: a cellular modelling approach to the
evolution of urban land use patterns. Environment and Planning A: Planning and Design, 25:11751199,
1993.
[82] R. White, G. Engelen, and I. Uijee. The use of constrained cellular automata for high-resolution modelling
of urban land use dynamics. Environment and Planning B: Planning and Design, 24:323343, 1997.
[83] F. Wu. An experiment on the generic polycentricity of urban growth in a cellular automatic city. Environment
and Planning B: Planning and Design, 25:103–126, 1998.
[84] M. Batty and Y. Xie. From cells to cities. Environment and Planning B: Planning and Design, 21:531548,
1994.
[85] J. von Neumann. The general and logical theory of automata. Reprinted in vol 5 of his Collected Works
(Oxford, Pergamon), 1948.
[86] J. von Neumann. Probabiltistic logics and the synthesis of reliable organisms from unreliable components.
Reprinted in vol 5 of his Collected Works, 1952.
[87] M. Burgess, H. Haugerud, and A. Strandlie. Object orientation and visualization of physics in two dimensions. Computers in Physics, 12:274, 1998.
[88] S. Weinberg. Lectures on Quantum Mechanics. Cambridge University Press, 2012.
[89] A.S. Perelson and G. Weisbuch. Immunology for physicists. Reviews of Modern Physics, 69:1219, 1997.
120
[90] M. Burgess. Computer immunology. Proceedings of the Twelth Systems Administration Conference (LISA
XII) (USENIX Association: Berkeley, CA), page 283, 1998.
[91] Arun V. Holden, editor. Chaos. Manchester University Press, 1986.
[92] S. Pepper. Encyclopedia of Library and Information Sciences 4th Ed, chapter Topic Maps. CRC Press,
ISBN 9780849397127, 2015.
[93] M. Burgess. A site configuration engine. Computing systems (MIT Press: Cambridge MA), 8:309, 1995.
[94] G. Canright and K. Engø-Monsen. A natural definition of clusters and roles in undirected graphs. Science
of Computer Programming, 53:195, 2004.
[95] R. Pastor-Satorras and A. Vespignani. Epidemic spreading in scale-free networks. Phys. Rev. Lett., 86:3200–
3203, 2001.
[96] F. Karlsson. Constraints on multiple initial embedding of clauses. International Journal of Corpus Linguistics, 12(1):107–118, 2007.
[97] W.X. Zhou, S. Sornette, R.A. Hill, and R.I.M. Dunbar. Discrete hierarchical organization of social group
sizes. Proc. Royal Soc., 272:439–444, 2004.
[98] M.J. Narasimha. The batcher-banyan self-routing network: universality and simplification. Communications, IEEE Transactions on, 36(10):1175–1178, Oct 1988.
[99] G. Tononi. An information integration theory of consciousness. BMC Neuroscience, 5:42–47, 2004.
[100] M. Tegmark. Consciousness as a state of matter. arXiv:1401.1219 [quant-ph], 2014.
[101] R. Feldman and J. Sanger. The Text Mining Handbook. Cambridge, 2007.
[102] M. Burgess. Probabilistic anomaly detection in distributed computer networks. Science of Computer Programming, 60(1):1–26, 2006.
[103] W.R. Ashby. Design for a brain. J. Wiley & Sons, 1952.
[104] W3C. Defining n-ary relations on the semantic web: use with individuals. http://www.w3.org/TR/2004/WDswbp-n-aryRelations-20040721/.
[105] E.L. Robertson. Triadic relations: An algebra for the semantic web. Lecture Notes in Computer Science,
3372:91–108, 2005.
[106] R. RIZZI. Complexity of context-free grammars with exceptions and the inadequacy of grammars as models
for xml and sgml, 2002.
[107] M. Burgess. A spacetime approach to generalized cognitive reasoning in multi-scale learning.
https://arxiv.org/abs/1702.04638, 2017.
[108] D. George and J. Hawkins. A hierarchical bayesian model of invariant pattern recognition in the visual
cortex. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 3,
pages 1812–1817, July 2005.
[109] R. Durbin, S. Eddy, A. Krigh, and G. Mitcheson. Biological Sequence Analysis. Cambridge, Cambridge,
1998.
[110] G. Boole. The Mathematical Analysis of Logic: Being an Essay Towards a Calculus of Deductive Reasoning.
Forgotten Books, New York, 2010.
[111] G. Boole. An Investigation of the Laws of Thought. Forgotten Books, New York, 2010.
[112] Gregory H Moore. The emergence of first-order logic. History and Philosophy of Modern Mathematics,
page 95, 1988.
[113] B.F. Chellas. Modal Logic: An Introduction. Cambridge University Press, 1980.
[114] M. Suneel. Dynamical systems as logic gates. arXiv:nlin/0511070 [nlin.CD], 2005.
121
[115] E. Ambrosi, M. Bianchi, C Gaibisso, and G. Gambosi. A description logic based grid inferential monitoring
and discovery framework. In In The 2005 International Conference on Grid Computing and Applications.
CSREA Press, 2005.
[116] S. Pon. Hilbert’s second problem: Foundations of arithmetic. UCSD undergraduate resources, 2003.
[117] D. Hilbert and W. Ackermann. Grundzüge der theoretischen Logik. Springer, Berlin, 1928.
[118] A. Couch and M. Burgess. Compass and direction in topic maps. (Oslo University College preprint), 2009.
[119] Mark Burgess. Knowledge management and promises. Lecture Notes on Computer Science, 5637:95–107,
2009.
[120] S. Kazadi, M. Chung, B. Lee, and R. Cho. On the dynamics on clustering systems. Robotics and Autonomous
Systems (46), 2004.
[121] Kazadi S. Swarm Engineering. PhD thesis, California Institute of Technology, 2000.
[122] J.F. Nash. Essays on Game Theory. Edward Elgar, Cheltenham, 1996.
[123] M. Burgess. On the scaling of functional spaces, from smart cities to cloud computing. arXiv:1602.06091
[cs.CY], 2016.
[124] L.M.A. Bettencourt, J. Lobo, D. Helbing, C. Hühnert, and G.B. West. Growth, innovation, scaling and the
pace of life in cities. Proceedings of the National Academy of Sciences, 104(107):7301–7306, 2007.
[125] P.A. Merolla et al. A million spiking-neuron integrated circuit with a scalable communication network and
interface. Science, 345:668–673, August 2014.
[126] J. Humble and D. Farley. Continuous Delivery: Reliable Software Releases Through Build, Test, and Deployment Automation. Addison Wesley, New Jersey, 2010.
[127] E.S. Lander. The heroes of crispr. Cell, 162:18–28, 2016.
[128] M. Burgess. How innovation is promised. Technical report, 2016.
[129] M. Belbin. Team roles.
http://www.belbin.com/about/belbin, 2009.
[130] K. Begnum and M. Burgess. Principle components and importance ranking of distributed anomalies. Machine Learning Journal, 58:217–230, 2005.
[131] S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer. Scheduled sampling for sequence prediction with recurrent
neural networks. arXiv:1506.03099 [cs.LG], 2015.
[132] K. Greff, R. Kumar Srivastava, J. Koutnik, B.R. Steunebrink, and J. Schmidhuber. Lstm: A search space
odyssey. arXiv:1503.04069 [cs.NE], 2015.
122
| 3cs.SY
|
1
Recognizing Material Properties from Images
arXiv:1801.03127v1 [cs.CV] 9 Jan 2018
Gabriel Schwartz and Ko Nishino, Senior Member, IEEE
Abstract—Humans implicitly rely on properties of the materials that make up ordinary objects to guide our interactions. Grasping smooth
materials, for example, requires more care than rough ones, and softness is an ideal property for fabric used in bedding. Even when
these properties are not purely visual (softness is a physical property of the material), we may still infer the softness of a fabric by looking
at it. We refer to these visually-recognizable material properties as visual material attributes. Recognizing visual material attributes in
images can contribute valuable information for general scene understanding and for recognition of materials themselves. Unlike wellknown object and scene attributes, visual material attributes are local properties. “Fuzziness”, for example, does not have a particular
shape. We show that given a set of images annotated with known material attributes, we may accurately recognize the attributes from
purely local information (small image patches). Obtaining such annotations in a consistent fashion at scale, however, is challenging. We
introduce a method that allows us to solve this problem by probing the human visual perception of materials to automatically discover
unnamed attributes that serve the same purpose. By asking simple yes/no questions comparing pairs of image patches, we obtain
sufficient weak supervision to build a set of attributes (and associated classifiers) that, while being unnamed, serve the same function
as the named attributes, such as “fuzzy” or “rough”, with which we describe materials. Doing so allows us to recognize visual material
attributes without resorting to exhaustive manual annotation of a fixed set of named attributes. Furthermore, we show that our automatic
attribute discovery method may be integrated in the end-to-end learning of a material classification CNN framework to simultaneously
recognize materials and discover their visual material attributes. Our experimental results show that visual material attributes, whether
named or automatically discovered, provide a useful intermediate representation for known material categories themselves as well as a
basis for transfer learning when recognizing previously-unseen categories.
Index Terms—visual material attributes, human material perception, material recognition
F
1
I NTRODUCTION
P
ROPERTIES of the materials that appear in everyday scenes
inform many of the decisions we make when interacting with
things made from these materials. When cleaning a glass cup,
for example, we know not to drop it or it will break. Glass is
also often smooth, and we grasp it accordingly so it will not
slip. Examples of material properties include visual properties,
such as glossiness or translucency, as well as physical or tactile
properties, such as hardness or roughness. We can see the presence
of these material properties simply by looking at ordinary images,
suggesting that even non-visual properties can be inferred from
the visual appearance of the material. We refer to such visuallyrecognizable material properties as visual material attributes. Recognizing these visual material attributes in images would allow us
to better understand and interact with the scenes in which materials
appear.
Recognizing visual material attributes in images, however,
is particularly challenging. Unlike much prior work in object,
face, and scene attribute recognition [1], [2], [3], [4], material
attributes are local properties, as can be seen in Figure 1. While
scene, object, and face attributes are typically associated with a
characteristic shape and fixed spatial extent, material attributes
are not. A car that is “sporty” will have a typically sleek and
aerodynamic shape, and if a face has “large eyes”, such an attribute
is defined by the spatial extent of the eyes relative to the face. If a
carpet is “fuzzy”, however, the fuzziness is not associated with any
characteristic shape or scale, nor is it necessarily a consequence
of the object involved (some carpets are not fuzzy).
The authors are with the Department of Computer Science, Drexel University,
Philadelphia, PA, 19104.
E-mail: {gbs25,kon}@drexel.edu
Fig. 1. Materials are unique in that they have many characteristic locallyrecognizable visual attributes. Wool, for example, often appears “fuzzy”,
and metals are typically “shiny”. In this work we show that material
attributes can be recognized from purely local information, that we can
discover these visual attributes automatically (again from purely local
information), and that such a process can be applied to large-scale
material image datasets via an end-to-end formulation in a Convolutional
Neural Network.
In a preliminary experiment, we show that given a fullysupervised dataset containing material annotations, a set of named
semantic material attributes (such as “fuzzy”, “organic”, or “transparent”), and sparse per-pixel labels for these attributes, we can
indeed recognize visual material properties from small local image
patches. Such a method can be applied in a sliding window fashion
to produce per-pixel predictions for material attributes in arbitrary
2
images. The accuracy of these predictions supports our claim
that material attributes are indeed locally-recognizable and can
be recognized at the per-pixel level. Furthermore, we show that
these attributes alone can be used to recognize materials in a way
that separates the material itself from the surrounding context such
as objects and scenes.
The primary drawback of a straightforward fully-supervised
approach, such as the one described above, is that it requires a set
of consistent annotations for the material attributes. Some material
attributes are intuitive and challenging to precisely define. If we
wish to scale the annotation process to multiple annotators, we can
no longer assume that a single person is providing attribute labels
in a consistent and complete fashion. Additionally, such a method
implicitly depends on the choice of a fixed set of named material
attributes. This restriction gives us no way of evaluating if the
chosen set of material attributes is complete or if there are possibly
more attributes that we may implicitly associate with materials.
Rather than assuming that we can exhaustively describe the
set of attributes humans associate with materials, we show that
we can instead directly probe the human visual perception of
materials using simple yes/no questions. Using the answers to
these questions as a form of weak supervision, we derive a
method for discovering a set of unnamed locally-recognizable
visual material attributes that faithfully encodes our own human
perceptual representation of materials. Our method requires only
material annotations, and discovers unnamed attributes with the
same desirable properties as the fixed named material attributes we
described previously, while using only a small amount of easilycollected weak supervision.
Our attribute discovery method requires only simple supervision and eliminates the need to manually define a set of named
material attributes for full supervision. The training process is,
however, still not ideal for application to modern large-scale
image databases. Working well with small amounts of training
data is a benefit, but we would ideally like to leverage recent
advances in large-scale end-to-end learning as well. To this end,
we show that the same material attributes can in fact be discovered
within a Convolutional Neural Network (CNN) framework focused on material recognition in local image patches (the Material
Attribute/Category CNN, MAC-CNN). This enables us to take
advantage of potentially larger material datasets. We also find
interesting parallels with the material representation in the human
material recognition process as observed in neuroscience [5], [6].
In contrast to the intermediate representations formed by previous
attribute methods, the human material recognition process (and
our MAC-CNN) produces a perceptual representation, i.e. visual
material attributes, as a side-product of material category recognition. Our results show that we are able to discover similar
perceptual attributes using the MAC-CNN, and we additionally
demonstrate the usefulness of perceptual material attributes for
transfer learning.
2
R ELATED W ORK
In this paper we discuss the recognition of material properties
from images. There is much recent work in the areas of attribute
recognition, material recognition, and human visual perception
that is relevant to our work; we discuss these findings below.
2.1
2.1.1
Attributes
Fully-Supervised Attributes
Fully-supervised visual attributes have been widely used in object
and scene recognition, but largely at the image or scene level.
Ferrari and Zisserman [2] introduced a generative model for
certain pattern and color attributes, such as “dots”, or “stripes”.
The attributes described in their model focus on texture and color,
but are not material attributes. A paper cup, for example, may have
stripes painted on it, but “striped” is not a property of the paper
itself; it is in fact a property of the cup. Kumar et al. [3] proposed
a face search engine with their attribute-based FaceTracer framework. FaceTracer uses SVM and AdaBoost to recognize attributes
within fixed facial regions. Such fixed regions are not present
in materials, which may take on an arbitrary shape unlike the
objects which they make up. Farhadi et al. [1] applied attributes
to the problem of object recognition. Their results showed an
improvement in accuracy over a basic approach using texture
features. Lampert et al. [7] also showed that attributes transfer
information between disjoint sets of classes. These results suggest
that attributes can serve as an intermediate representation for
recognition of the categories which exhibit them. Patterson and
Hays [4] showed that they could recognize a variety of visual
attributes, some of which were in fact general material categories.
Their work, however, was not an explicit attempt at recognizing
materials.
With a few exceptions, the majority of past attribute recognition methods produced single image-wide predictions. Given
an image of a zebra, for example, the attribute prediction would
be “striped”. As our goal is to recognize materials within local
regions, we cannot rely on such global attribute predictions.
2.1.2
Weakly-Supervised Attributes
The attributes described above were all fully-supervised or “semantic” attributes. A semantic attribute is one to which we can
assign a name like “round” or “transparent”. While these attributes
were shown to be useful, it is difficult to quantify the completeness
and consistency of any given attribute set: does the set of attributes
contain everything that could help recognize the target categories,
and can the appearance (for visual attributes) be agreed upon by
a variety of annotators? Semantic attributes are also task-specific
and must be manually defined for each new recognition task.
To address the issues inherent to semantic attributes, a number
of unsupervised or weakly-supervised attribute discovery methods
have been proposed. Berg et al. [8] described a framework for
automatically learning object attributes from web data (images and
associated text). This approach learns some localized attributes.
The required text annotations are, however, image-wide and do not
guarantee locality. Patterson and Hays [4] also proposed a process
to discover and recognize scene-wide attributes in natural images.
While they are able to discover a large amount of attributes, their
learned attributes are not local. Rastegari et al. [9] learn a binary
attribute representation (binary codes) for images. As with most
existing methods, however, these attributes are image-wide and
not local. Cimpoi et al. [10] demonstrated a method for learning
an arbitrary set of describable texture attributes based on terms
derived from psychological studies. As noted by Adelson [11],
texture is only one component of material appearance, and cannot
alone describe our perception of materials. Though their results
demonstrate impressive performance on the FMD, their learned
attributes apply only globally. Most relevant to our work are the
3
attribute discovery methods of Akata et al. [12] and Yu et al. [13].
Akata et al. [12] formulated attribute discovery as a label embedding problem. Yu et al. [13] proposed a two-step procedure for
discovering and classifying attributes based on a similarity matrix.
They computed a distance matrix using Euclidean distances in the
raw feature space of labeled image patches. In contrast, we embed
the material categories in an attribute space derived from our own
human visual perception of material similarity.
2.2
Material Recognition
Adelson [11] first suggested materials as a distinct concept from
objects or simple textures when discussing “things vs. stuff”.
“Things” refers to objects, which have been the focus of much
prior work under the field of object recognition. Adelson points
out that the world does not just consist of discrete objects, but also
includes “stuff”, substances without a natural shape or fixed spatial
extent. Ice cream is one example of “stuff” that is not an object
but is still a recognizable concept in images. While materials are
not equivalent to the “stuff” discussed in his work, the work does
lay the foundation for material recognition as a vision problem.
The first collection of material category images for classification originated in Sharan et al. [14] where they introduced
a new image database (the Flickr Materials Database or FMD)
containing images from the photo sharing website Flickr. The
FMD contains a set of images each with a single material
annotation and corresponding mask identifying the presence of
that material. Building on the FMD, Liu et al. [15] created a
framework to recognize these material categories using a modified
LDA probabilistic topic model. Hu et al. [16] improved upon the
state-of-the-art FMD accuracy using kernel descriptors and largemargin nearest neighbor distance metric learning. Their experiments showed that providing explicit object detection information
to material category recognition results in a large improvement in
accuracy. Sharan et al. [17] later showed that without information
associated with objects (such as the object shape), performance
degrades significantly (from 57.1% to 42.6%). Specifically, they
note that their material category recognition method depends
heavily on non-local features such as edge contours. Given that
materials exhibit distinct locally-recognizable visual attributes (as
we show), it follows that we should be able to recognize them in
a way that does not suffer from reduced accuracy in the absence
of context.
All prior work discussed above produces a single category
prediction for each input image. This inherently assumes that there
is only one material of interest in the image, a very restrictive
assumption. To relax this assumption, recent work (including some
of our own preliminary work [18]) focuses on dense prediction:
providing a material category for each pixel in the input image.
Bell et al. introduced the OpenSurfaces [19] and MINC [20]
datasets to aid in the training of dense material recognition models.
With MINC they also describe a simple modification of the VGG
CNN architecture of Simonyan and Zisserman [21] to predict their
material categories at each pixel.
2.3 Material Perception and Convolutional Neural Networks
As the final step in the scaling of material attribute learning,
we discover perceptual material attributes within Convolutional
Neural Networks (CNNs) [22]. Convolutional neural networks are
general non-linear models that apply a set of convolution kernels
to an image in an hierarchical fashion to generate a category
probability vector. The kernel weights are model parameters that
are set via non-linear optimization (generally Stochastic Gradient
Descent) to attempt to maximize the likelihood of a set of training
data.
Recently, Shankar et al. [23] proposed a modified CNN
training procedure to improve attribute recognition. Their “deep
carving” algorithm provides the CNN with attribute pseudolabel targets, updated periodically during training. This causes
the resulting network to be better-suited for attribute prediction.
Escorcia et al. [24] show that known semantic attributes can also
be extracted from a CNN. They show that attributes depend on
features in all layers of the CNN, which will be particularly
relevant to our investigation of perceptual material attributes in
CNNs. ConceptLearner, proposed by Zhou et al. [25] uses weak
supervision, in the form of images with associated text content, to
discover semantic attributes. These attributes correspond to terms
within the text that appear in the images. All of these frameworks
predict a single set of attributes for an entire image, as opposed to
the per-pixel attributes which we introduce.
At the intersection of neuroscience and computer vision,
Yamins et al. [26] find that feature responses from highperforming CNNs can accurately model the neural response of
the human visual system in the inferior temporal (IT) cortex (an
area of the human brain that responds to complex visual stimuli).
They perform a linear regression from CNN feature outputs to
IT neural response measurements and find that the CNN features
are good predictors of neural responses despite the fact that the
CNN was not explicitly trained to match the neural responses.
Their work focuses on object recognition CNNs, not materials.
Hiramatsu et al. [5] take functional magnetic resonance imaging (fMRI) measurements and investigate their correlation with
both direct visual information and perceptual material properties
(similar to the material traits of [27]) at various areas of the
human visual system. They find that pairwise material dissimilarities derived from fMRI data correlate best with direct visual
information (analogous to pixels) at the lower-order areas and with
perceptual attributes at higher-order areas. Goda et al. [6] obtain
similar findings in non-human primates. These studies suggest the
existence of perceptual attributes in human material recognition,
but do not actually derive a process to extract them from novel
images.
3
V ISUAL M ATERIAL T RAITS
As preliminary experiment to show that materials do in fact exhibit
locally-recognizable visual properties, we use a set of named
visual material attributes (visual material traits [27]). By manually
annotating images from the FMD with per-pixel masks for 13
material traits, we can then train a set of Randomized Decision
Forest [28] classifiers to predict material traits in small local
regions. Experimental results show that visual material traits can
be recognized very accurately from small (32×32) image patches,
some traits as high as 93.1%, with an average accuracy of 78.4%.
Fig. 2 shows sample per-pixel recognition results for a selection
of material traits.
Materials, for example fabric, plastic, or metal, can be challenging to recognize due to the large variation in appearance
between instances of the same material. Despite this, looking at
the images in Fig. 3, one can see that plastic tends to have visual
material attributes that are associated with a distinct appearance,
4
Fuzzy
Organic
Smooth
Shiny
Metallic
Smooth
Fig. 2. Per-pixel material trait predictions on unseen test images. Each RGB color channel corresponds to the predicted probability of a single
selected material trait. The recognized traits clearly divide the images into regions of similar material appearance. Of particular importance is the
fact that we can successfully recognize material properties like “fuzzy” on the left, despite the fact that our training data did not include any animals.
The shape of each statue on the right is that of a person, but we see from the recognized material traits that the material is in fact metal.
Fig. 3. Materials such as the plastic in these images exhibit a wide range
of appearances depending on the object and scene. Despite this, we can
intuitively recognize visual attributes (smooth and translucent, for example) shared across different instances of the material. In our preliminary
work we show that these attributes, which we refer to as visual material
traits, are locally-recognizable and can be used to recognize material
categories like plastic or metal.
such as “smooth” and “translucent”. Our key observation is that
these visual material attributes are recognizable even when the
surrounding objects and scenes are not visible. We expect that as a
result, we should be able to recognize materials themselves from
visual material traits given only small local image patches.
Viewing the full set of material traits as an intermediate
representation, we may aggregate them within regions to describe
materials. To exploit the material properties found in locallyrecognized material traits, we treat the distribution of material
traits in a region as an image descriptor and generate a per-image
material category prediction. We do so by extracting a number of
patches within each material region and using the distributions of
traits across these patches as features for a histogram intersection
SVM [29]. The average accuracy of this method on 50-50 splits
of FMD images is 49.2%. While this is not higher than the
accuracy of methods such as [17] that implicitly use context, it is
significantly higher than both human and algorithmic performance
in the absence of context (38.7-46.9% and 33.8-42.6% respectively
respectively [17]). Furthermore, material traits learned from one
dataset can be recognized and used to extract material information
from an entirely different set, showing that the representation
generalizes well.
4
M ATERIAL ATTRIBUTE D ISCOVERY
Material trait recognition relies on a set of fully labeled material
trait examples. This assumption hinders scaling the method to
Fig. 4. Sample material image patches. Asking annotators to merely
“describe” the patches is an ambiguous question. Patches may look
similar even though the annotator does not have a concrete word to
define the similarity. We instead ask only for binary visual similarity
decisions.
larger training datasets. We also do not have a complete, mutuallyagreeable vocabulary for describing materials and their visual
characteristics; named material traits are merely one attempt at
describing material appearance. This makes scaling with multiple
annotators difficult. Considering the images in the first column
of Fig. 4, for instance, one annotator may call them fuzzy and
others may call them fluffy. People may also be inconsistent in
annotating material traits. Some may only annotate the patches in
the second column as smooth and others may only see them as
translucent. Cimpoi et al. [10] alleviate these problems for texture
recognition by preparing a pre-defined vocabulary. They may do
so by focusing on apparent texture patterns like stripes and dots.
Materials underlie these texture patterns (i.e., the stripes or dots on
a plastic cup are still plastic) and do not follow such a vocabulary.
Our goal is to discover a set of attributes that exhibit the
desirable properties of material traits. We want to achieve this
without relying on fully-supervised learning. Known material
traits, such as “smooth” or “rough,” represent visual properties
shared between similar materials. We expect that attributes that
preserve this similarity will satisfy our goal. We propose to define
a set of attributes based on the perceived distances between material categories. By working with distances rather than similarities,
we avoid any need to assume a particular similarity function. For
this, we obtain a measurement of these distances from human
annotations.
From a high-level perspective, our attribute discovery consists
of three steps:
1)
2)
3)
Measure perceptual distances between materials
Define an attribute space based on perceptual distances
Train classifiers to reproduce this space from image
patches
5
4.1
Water-Leather Subspace
Fabric
Foliage
Glass
Leather
Metal
Paper
Plastic
Stone
Water
Wood
Leather
Defining perceptual distance between material categories poses a
challenge. If each material had a single typical appearance (e.g., if
metal was always shiny and gray), we could simply compute the
difference between these typical appearances. This is not the case.
Materials may exhibit a wide variety of appearances, even sharing
appearances between categories (what we refer to as material
appearance variability). An image patch from a leaf, for example,
may appear similar to certain fabrics or plastics.
Measuring Perceptual Distances
Directly measuring distances via human annotation would be
ideal, as we have an intuitive understanding of the differences
between materials. As Sharan et al. [17] showed, this understanding persists even in the absence of object cues. It is, however, also
a difficult task to obtain these distance. Given two query image
patches, annotators would have to decide how different the patches
look on a consistent quantitative scale. We would instead like to
ask simple questions that can be reliably answered.
We propose that instead of asking how different patches look,
we reduce the question to a binary one: “Do these patches look
different or not?” We assume that this will give us sufficient information to obtain consistent and sensible perceptual information.
Our underlying assumption for this claim is that if a pair of image
patches look similar, they do so as a result of at least one shared
visual material trait.
To transform a set of binary similarity annotations into pairwise distances, we represent each material as a point defined by
the average probabilities of similarity to each material category.
The pairwise distances between these points define the material
perceptual distance matrix. This process treats each material
category as a point in a space of typical (but not necessarily
realizable) material appearances. The resulting distance between
a pair of materials depends on joint similarity with all material
categories, including the pair in question, and is thus robust to
material appearance variability.
Formally, given a set of N reference images with material
category cn ∈ {1 . . . K}, we obtain binary similarity decisions
K
sn ∈ {0, 1} for each reference image against a set of sample
images from each category. We represent each material category
in the space of typical material category appearances as K dimensional vectors pk :
1 X
pk =
sn ,
(1)
Nk
n|cn =k
where Nk = |{cn |cn = k}|. Entries dkk0 in the K × K pairwise
distance matrix D are then defined as:
dkk0 = kpk − pk0 k2 .
(2)
We obtain the required set of binary similarity annotations
through Amazon Mechanical Turk (AMT). Each task presents annotators with a reference image patch of a given material category
(unknown to the annotator) and a row of random image patches,
one from each material category. We use patches from images of
the 10 material categories from the Flickr Materials Database of
Sharan et al. [14]. Annotators are directed to select image patches
that look similar to the reference. Examples of suggested similar
image patches are given based on known material traits. Each set
of patches is shown to 10 annotators, and final results are obtained
from a vote where at least 5 annotators must agree that the patches
look similar. We collect similarity decisions for 10,000 reference
image patches.
0
0
Water
Fig. 5. Projection of materials into a 2D similarity subspace allow us
evaluate our collected annotations. Coordinates indicate perceived similarity to water and leather respectively. The locations of the two material
categories corresponding to the axes are marked. We would expect that,
in this case, water would lie furthest along the “water” axis and likewise
with leather. We can also see that water occasionally looks like leather.
Materials with common visual properties, such as the smoothness of
plastic and glass, tend to lie close to each other. Finally, no material
appears strongly similar to both leather and water. This is expected as
the two materials do generally exhibit different appearances.
The 2D projection in Fig. 5 shows that the similarity values
obtained from the AMT annotations agree with our own intuitive understanding of material appearance. The plot shows the
locations of material categories projected into one 2D subspace
of the 10-dimensional space of material appearance similarity
(each dimension corresponding to similarity with that material).
We would expect that the two materials corresponding to the axes
of their subspace will lie close to their respective axes. In this
case, water is most similar with itself, but is also similar to glass.
Leather is likewise most similar with itself, but also similar to
fabric.
To show that we do in fact obtain a consistent distance matrix,
we compute the difference between the distance matrix computed
with all annotations versus that from only n of the N total
annotations. The difference drops quickly (within the first few
hundred samples of 10,000), showing that annotators agree on a
single common set of perceptual distances.
4.2
Defining the Material Attribute Space
Discovering attributes given only a desired distance matrix poses a
challenge. First, we have no definitions for the attributes since we
are trying to discover them from examples of materials. Even if we
had such definitions, we have no knowledge of any association between the attributes and the materials. A straightforward approach
might be to directly train classifiers to output attribute values that
encode the distance matrix. This would be a particularly underconstrained problem given the aforementioned lack of labels or
associations.
We instead propose to separate attribute association and classifier learning into two steps. First, we discover attributes in an
abstract form by discovering a mapping between categories and
attribute probabilities. This mapping places each material category
into an attribute space. We ensure that the mapping preserves the
pairwise perceptual material distances, and then train classifiers to
predict the presence of these attributes on image patches.
As described in Section 4.1, we obtain a distance matrix D
from crowdsourced similarity answers for K material categories
6
C = {1 . . . K}. Using D, we find a mapping that indicates which
attributes are associated with which categories. The number of
attributes we discover is arbitrary, and we refer to it as M . The
mapping is encoded in the K × M category-attribute matrix A.
We restrict values in A to lie in the interval [0, 1] so that we may
treat them as conditional probabilities.
We impose two constraints on the category attribute mapping.
A should map categories to attributes in a way that preserves
the measured distances in D, and the mapping should contain
realizable values. If the values in A are not plausible, we will not
be able to recognize the attributes on image patches. For example,
one potential attribute mapping would be to assign each attribute
to a single category. Attribute recognition would then become
material category recognition on single image patches, which is
not feasible.
We formulate the attribute discovery process as a minimization
problem over category-attribute matrices A:
A∗ = arg min d (D; A) + wA κA (A)
(3)
A
with hyperparameter wA . d describes how well the current estimate of A encodes the pairwise perceptual differences between
material categories, and κA is a constraint that makes the discovered attribute associations exhibit a realizable distribution.
The category-attribute matrix that best encodes the desired
pairwise distances will minimize the following term defined over
rows ak of the matrix A:
d (D; A) =
X
2
(kak − ak0 k2 − Dkk0 ) .
(4)
k,k0 ∈C
4.3
Training a Material Attribute Classifier
We now must derive classifiers that recognize the attributes defined
by the category-attribute mapping. As attributes are not defined
semantically, we cannot ask for further annotation to label training
patches with attributes. Instead, we propose a model and a set of
constraints that will enable us to predict our discovered attributes
on material image patches.
We do not know a priori any particular semantics or structure
associated with the attributes, thus we model our attributes using
a general two-layer non-linear model [30]. We constrain the
predictions such that they reproduce the desired values in the
attribute matrix (in expectation) while also separating material
categories when possible.
Formally, given a training set of N image patches represented
by D-dimensional raw feature vectors xn with corresponding
material categories cn ∈ C , we train a model f with parameters Θ that maps an image patch to M attribute probabilities:
M
f (xn ; Θ) : RD → [0, 1] . Given an intermediate layer with
dimensionality H and parameters W1 ∈ RH×D , W2 ∈ RM ×H ,
b1 ∈ RH , b2 ∈ RM the prediction for an instance xn is defined
as:
f (xn ; Θ) = h (W2 h (W1 xn + b1 ) + b2 )
h (x) = min (max (x, 0) , 1) .
(7)
As additional regularization, used only during training, we mask
out a random fraction of the weights used in the model to
discourage overfitting (akin to dropout [31]).
We formulate the full classifier training process as a minimization problem:
Θ? = arg min r (X; A, Θ) + w1 κ (X; Θ) −
Θ
To discover realizable attributes, we encode our own prior
knowledge that recognizable attributes exhibit a particular distribution and sparsity pattern. We observe that semantic attributes,
specifically visual material traits, have a Beta-distributed association with material categories. Generally, a material category
will either strongly exhibit a trait or it will not exhibit it at
all. Intermediate cases occur when a material category exhibits
a particularly wide variation in appearance. Fabric, for example,
sometimes has a clear “woven” pattern but, in the case of silk or
other smooth fabrics, does not. We would like the values in A
to be Beta-distributed to match the distribution of known material
trait associations.
The canonical method for matching two distributions is to
minimize a divergence measure between them. To incorporate
this into a minimization formulation, we need a differentiable
measurement for the unknown empirical distribution of values in
A. We choose the KL-divergence and Gaussian kernel density
estimator. The Gaussian kernel density estimate at point p is:
(
)
2
1
1 X
(akm − p)
2 −2
q (p; A) =
2πh
exp −
KM k,m
2h2
(5)
The KL-divergence between the distribution of the values in
the category-attribute matrix A and the target Beta distribution
β (p; a, b) with a = b = 0.5 can then be written as:
β (p; a, b)
κA (A) =
β (p; a, b) ln
.
q (p; A)
p∈P
X
(6)
w2 π (X; A, Θ) ,
(8)
with hyperparameters w1 and w2 . r (Equation 9) is a data term
indicating the difference between predicted and expected attribute
probabilities. κ and π (Equations 10 and 11) are, respectively,
constraints on the the distribution of attribute predictions and on
the pairwise separation of material categories.
The category-attribute matrix encodes the probabilities that
each category will exhibit each attribute. We represent this in our
classifier training by matching the mean predicted probability for
each attribute to the given entry in the category-attribute matrix:
2
r (X; A, Θ) =
X
k∈C
1 X
ak −
f (xi ; Θ)
Nk
i|ci =k
.
(9)
2
Equation 9 directly encodes the desired behavior of the classifier,
but it alone is under-constrained. Each prediction for each instance
may take on any value so long as their mean matches the target
value.
We have observed that, similar to category-attribute associations, predicted probabilities for known material traits are also
Beta-distributed. Local image regions exhibiting a trait will have
uniformly high probability for that trait, only decreasing around
the trait region edges. We constrain the predicted probabilities
such that they are Beta-distributed. Using the formulation discussed in Section 4.2, we again minimize a KL-divergence of a
kernel density estimate:
X
β (p; a, b)
κ (X; Θ) =
β (p; a, b) ln
,
(10)
q (p; f (X; Θ))
p∈P
7
Fabric
Foliage
Glass
Leather
Metal
Paper
Plastic
Stone
Water
Wood
(a) Raw Features
(b) Discovered Attribute Predictions
Fig. 6. t-SNE [32] embedding of materials from the raw feature space (a) and from our discovered attributes (b). We embed a set of material image
patches into 2D space via t-SNE using raw features and predicted attribute probabilities as the input space for the embeddings. Though t-SNE has
been shown to perform well in high-dimensional input spaces, it fails to separate material categories from the raw feature space. Material categories
are, however, clearly more separable with our attribute space.
where f (X; Θ) represents the N × M matrix of attribute probability predictions for the training dataset, and q, a, b are defined
as in Equation 6.
One of the goals for our attribute representation is to discover
attributes that allow for material classification. If this were our
only goal, we could simply maximize the distance between the
predicted attributes for all pairs of different material categories.
This would conflict with our goal of preserving human perception,
as material categories do not always exhibit different appearances.
We instead modify this separation by weighting each component
of the distance based on the values in the category-attribute matrix:
X
π (X; A, Θ) =
pT
(11)
ij pij
i,j∈N |ci 6=cj
pij = 2 aci − acj − 1 (f (xi ; Θ) − f (xj ; Θ)) .
This separates the material categories in attribute space only when
the attributes dictate that there is a perceptual difference.
4.4
Analysis of Discovered Attributes
To analyze the properties of attributes discovered by our framework, we follow the procedures outlined above to collect annotations and discover a set of attributes. Since both learning steps
involve minimization of a non-linear, non-convex function, we rely
on existing optimization tools1 to find suitable estimates. As a raw
feature set, we use the local features we developed for material
trait recognition (Section 3).
If our attributes described a space that successfully separates
material categories, we would expect categories to form clusters
in the attribute space. To verify this, we compute a 2D embedding
of a set of labeled image patches. For the embedding, we use
the t-SNE method of van der Maaten and Hinton [32]. t-SNE
attempts to generate an embedding that matches the distributions
of neighboring points in the high- and low-dimensional spaces. In
Fig. 6, we represent image patches by their raw feature vectors (a)
1. Specifically, L-BFGS with box constraints for A and stochastic gradient
descent for Θ.
Input
Attribute 1
Attribute 2
Attribute 3
Random
Fig. 7. Per-pixel discovered attribute probabilities for four attributes (one
per column). These images show that the discovered attributes exhibit
patterns similar to those of known material traits. The first attribute, for
example, appears consistently within the woven hat and the koala; the
second attribute tends to indicate smooth regions. The third attribute
shows we are discovering attributes that can appear both sparsely and
densely in an image, depending on the context. These are all properties
shared with visual material traits. Attributes from a random A do not
exhibit any of these properties.
and predicted attribute probability vectors (b), and compare the 2D
embeddings resulting from each. Material categories are separated
much more clearly in our attribute space than in the raw feature
space.
Part of the usefulness of visual material traits, as we have
shown above, is derived from the fact that they each represent
a particular intuitive visual material property. This is evident in
the spatial sparsity pattern of the traits, specifically the fact that
they appear in regions and not randomly within an image. Traits
such as “shiny” are highly localized, while others such as “woven”
or “smooth” exist as coherent regions within a particular material
instance. Fig. 7 shows examples of per-pixel attribute probabilities
predicted from our discovered attribute classifiers. The attributes
exhibit both sparse and dense spatial patterns that are consistent
within local regions. Dense attributes generally correspond with
8
Accuracy vs. Train/Test Fraction
Fuzzy
Shiny
0.45
Smooth
0.30
Striped
0.15
Metallic
Organic
Accuracy (%)
Known Traits
Soft
0
Translucent
-0.15
Transparent
Rough
-0.30
Liquid
-0.45
Woven
Manmade
0
2
6
8
4
Learned Attributes
10
12
Fig. 8. Correlation between discovered attribute predictions and material
traits. Groups of attributes can collectively indicate the presence of a
material trait. Metallic, for example, correlates positively with attribute 0
and negatively with attribute 8.
smooth image regions. Sparse attributes often indicate localized
surface features such as specific texture patterns.
Randomized features have been shown to provide a viable
representation for classification tasks [33]. Despite this, we expect that such features would not be likely to encode the same
perceptual properties as our attributes. We can demonstrate this
by replacing our perceptually-derived attribute matrix A with
randomized matrix of the same shape. The last column in Fig. 7
shows typical results for such a matrix. Unlike attributes based on
human perception, these random attributes do not exhibit any of
the desired perceptual properties like spatial consistency.
We aimed to discover attributes similar to the visual material
traits that underlie human perception. We thus expect that the
discovered attributes exhibit a correlation with known traits. Fig. 8
shows the correlation between 13 discovered attributes and 13
known material traits using attributes predicted on labeled material
trait image patches. Collectively, we can indeed describe material
traits using the discovered attributes. Visually similar traits, such
as rough and woven, show similar correlations with the attributes.
Discovered attributes are also consistent with the semantic properties of material traits. Rough and smooth are mutually exclusive
traits, and we see that discovered attributes that positively correlate
with smooth do not generally correlate with rough.
We quantitatively evaluate the discovered attributes using logic
regression [34]. Given a set of image patches with known traits, we
predict our discovered attributes as binary values for use as input
variables in a logic regression model for material traits. Logic
regression from 30 attributes alone (no other features) achieves
comparable accuracy to our trait-based method and its complex
feature set. These results show that the discovered attributes do
collectively encode intuitive visual material properties. Further
details and examples of logic regression to predict material traits
can be found in Sec. 7.1.
Train/Test Fraction
Fig. 9. Accuracy vs. training set size. The line is the average of 3
random splits. Error bars indicate the minimum and maximum of the
splits. Accuracy does not continue to increase as we use larger training
datasets. This shows that we have successfully extracted as much local
information as possible from human perception.
our discovered attributes in place of labeled material traits. We
compute the histograms of these predicted probabilities across the
material region and use them as input for a histogram kernel SVM.
As we focus on local attributes, these previous local results (and
those of Sharan et al. [17] on scrambled images) serve as the
correct baseline.
To compare with our previous results using material traits, we
compute average material recognition accuracy on the Flickr Materials Database (FMD). All results are computed using M = 30
discovered attributes and 5-fold cross-validation unless otherwise
specified.
Our attributes achieve an average accuracy of 48.9% (σ =
1.2%) on FMD images using only local information. This is
comparable to our results and those of Sharan et al. [17] (using
only local information) even though we are discovering attributes
using only weak supervision.
Looking at individual class recognition rates, metal is the
most challenging category to identify while foliage is the most
accurately recognized. This follows from the results of our measurements of human perception, as annotators consistently found
that foliage image patches looked different from all other material
categories. Metal is particularly difficult to recognize locally as its
appearance depends strongly on the appearance of the surround
environment (due to specular reflection).
Fig. 9 shows that accuracy reaches a plateau as the training
dataset size increases. We also compute accuracy for varying values of M and find that past M = 30, there is little (<0.1%) gain
in accuracy from additional attributes. These results indicate that
we are in fact extracting as much perceptual material information
as we can from the available data.
5
4.5
From Discovered Attributes to Materials
Seeing that discovered attributes encode visual material properties,
we would expect them to also serve as an intermediate representation for material category recognition. To test this, we follow
our local material recognition procedure (Section 3), substituting
P ERCEPTUAL M ATERIAL ATTRIBUTES IN C ON N EURAL N ETWORKS
VOLUTIONAL
We have shown that we may use visual material traits to enable
local material recognition, and we may further scale this attributebased recognition process by automatically discovering perceptual material attributes. Our previous methods consider attributes
9
separately from category recognition. The attributes are used
solely as an intermediate representation for material categories.
Similarly for conventional object and scene recognition, attributes
like “sunset” or “natural,” have also been extracted for use as
independent features. Shankar et al. [23] generate pseudo-labels
to improve the attribute prediction accuracy of a Convolutional
Neural Network, and Zhou et al. [25] discover concepts from
weakly-supervised image data. In both cases, the attributes are
considered on their own, not within the context of higher-level
categories. In object and scene recognition, however, recent work
shows that semantic attributes seem to arise in networks that are
trained end-to-end for category recognition [35].
We would like to take advantage of the benefits of end-to-end
learning to incorporate automatically-discovered attributes with
material recognition in one seamless process. Material attribute
recognition, however, is not easily scalable. In the past we relied
on semantic attributes, such as “shiny” or “fuzzy”, that needed
careful annotation by a consistent annotator as their appearance
may not be readily agreed upon. We addressed the difficulty
in annotation scaling by automatically discovering perceptual
material attributes from weak supervision. The training process
for this method does not, however, scale well to large datasets. To
address this, we propose a novel CNN architecture that recognizes
materials from small local image patches while producing perceptual material attributes as an auxiliary output. We also introduce
a novel material database with material categories drawn from a
materials-science-based category hierarchy.
5.1 Finding Material Attributes in a Material Recognition CNN
Hiramatsu et al. [5] have shown that perceptual attributes form
an integral component of the human material recognition process.
They found that during the process of material recognition, we
form a perceptual representation of materials analogous to the
intermediate representation provided by named material traits or
automatically-discovered material attributes. The key difference is
that the perceptual representation in human material recognition
forms as an integral part of the recognition process.
Our goal is to discover unnamed material attributes while performing material recognition in one end-to-end scalable process.
Based on correlations between Convolutional Neural Network
(CNN) feature maps and human visual system neural output
discovered by Yamins et al. [26], a CNN architecture appears
to be a very suitable framework in which to discover attributes
analogous to those in human material perception. Their work
focuses on object recognition, however, and does not extract any
attributes. In this case, our automatically-discovered unnamed
material attributes are particularly relevant. We derive a novel
framework to discover perceptual attributes similar to the ones
we describe in Section 4 inside a material recognition CNN
framework.
A simple experiment to verify the feasibility of perceptual
attribute discovery in a CNN trained to recognize materials would
be to add a layer at the top of the network, immediately before the
final material category probability softmax layer, and constrain
this layer to output the same attributes we previously discovered
in a separate process. If we could predict attributes from this layer
without affecting the material recognition accuracy, this would
suggest that we could indeed combine attribute discovery and
material recognition. We implemented this approach and found
30
30
64 64
30
128 128
30
256
256
256
30
512
512
512
Fully Connected
Input Image
3x3 conv
Pooling
Material
2048
2048
Perceptual Distances
Attribute-Constrained FC
Fig. 10. Material Attribute-Category CNN (MAC-CNN) Architecture: We
introduce auxiliary fully-connected attribute layers to each spatial pooling layer, and combine the per-layer predictions into a final attribute
output via an additional set of weights. The loss functions attached to
the attribute layers encourage the extraction of attributes that match the
human material representation encoded in perceptual distances. The
first set of attribute layers acts as a set of weak learners to extract
attributes wherever they are present. The final layer combines them to
form a single prediction.
that while the material accuracy was unaffected, the attribute
predictions were less accurate than if the CNN was trained solely
to predict the attributes and not materials as well (mean average
error of 0.2 vs 0.08).
The key issue with the straightforward approach is that it is
not an entirely faithful model to the process described in [5]. They
note that the human neural representation of material categories
transitions from visual (raw image features) to perceptual (visual
properties like “shiny”) in an hierarchical fashion. This implies,
in agreement with findings of Escorcia et al. [24], that attributes
require information from multiple levels of the material recognition network. We show that this is indeed the case by successfully
discovering the attributes using input from multiple layers of the
material recognition network.
5.2
Material Attribute-Category CNN
We need a means of extracting attribute information at multiple
levels of the network. Simply combining all feature maps from
all network layers and using them to predict attributes would
be computationally-impractical. Rather than directly using all
features at once, we augment an initial CNN designed for material
classification with a set of auxiliary fully-connected layers attached to the spatial pooling layers. This allows the attribute layers
to use information from multiple levels of the network without
needing direct access to every feature map. We treat the additional
layers as a set of weak learners, each auxiliary layer discovering
the attributes available at the corresponding level of the network.
This is similar to the deep supervision of Lee et al. [36]. Their
auxiliary loss functions, however, simply propagate the same
classification targets (via SVM-like loss functions) to the lower
layers. Rather than propagating gradients, our attribute layers
discover perceptual material attributes.
For the auxiliary layer loss functions, we introduce a modified
form of the perceptual attribute loss function (Equation 8) to
the outputs of each auxiliary fully-connected layer. Specifically,
assuming the output of a given pooling layer i in the network
for image j is hij , and given categories C, |C| = K and a set
of sample points P ∈ (0, 1) for density estimation, we add the
10
following auxiliary loss functions:
ui =
1 X
1 X T
f Wi hij + bi
ak −
K k∈C
Nk
j|cj =k
di =
X
p∈P
β (p; a, b) ln
β (p; a, b)
q p; f WiT hij + bi
(12)
1
!
,
(13)
where f (x) = min (max (x, 0) , 1) clamps the outputs within
(0, 1) to conform to attribute probabilities, and weights Wi , bi
represent the auxiliary fully-connected layers we add to the network. ak represents a row in the category-attribute mapping matrix
derived as in Section 4.2. Equation 12 causes the attribute layer to
discover attributes which match the perceptual distances measured
from human annotations. As certain attributes are expected to
appear at different levels of the network, some layers will be
unable to extract them. This implies that their error should be
sparse, either predicting an attribute well or not at all. For this
reason we use an L1 error norm. Equation 13, applied only to the
final attribute layer, encourages the distribution of the attributes
to match those of known semantic material traits. It takes the
form of a KL-divergence between a Beta distribution (empirically
observed to match the distribution of semantic attribute probabilities), and a Kernel Density Estimate q (·) of the extracted attribute
probability density sampled at points p ∈ P .
The reference network we modify is based on the highperforming VGG-16 network of Simonyan and Zisserman [21].
We use their trained convolutional weights as initialization where
applicable, and add new fully-connected layers for material classification. Fig. 10 shows our architecture for material attribute
discovery and category recognition. We refer to this network as
the Material Attribute-Category CNN (MAC-CNN).
6
L OCAL M ATERIAL DATABASE
In order to train the category recognition portion of the MACCNN, we need a suitable dataset, and we find existing material
databases lacking in a few key areas. Previous material recognition
datasets [17], [19], [20] have relied on ad-hoc choices regarding
the selection and granularity of material categories. When patches
are involved, as in [20], the patches can be as large as 24% of the
image size surrounding a single pixel identified as corresponding
to a material. These patches are large enough to include entire
objects. These issues make it difficult to separate challenges
inherent to material recognition from those related to general
recognition tasks. Materials, as with their visual attributes, are
inherently local properties. While knowledge of what object the
material composes may help recognize the material, the material
is not the object. We also find that image diversity is still lacking
in these modern datasets. For these reasons, we introduce a new
local material recognition dataset to support the experiments in
this paper.
6.1
Material Category Hierarchy
Material categories in existing datasets have not been carefully selected. Examples of this appear in the recent work of
Bell et al. [20], where proposed material categories include
“mirror” and “carpet” (among others). These are in fact objects,
and their annotations reflect this. Their categories also confuse
materials and their visual properties, for example, separating
Fig. 11. One tree in our material category hierarchy. Categories at
the top level separate materials with notable differences in physical
properties. Mid-level categories are visually distinct. The lowest level
of categories are fine-grained and may require both physical and visual
properties and expert knowledge to distinguish them.
“stone” from “polished stone”. To address the issue of material
category definition, we propose a set of material categories derived
from a materials science taxonomy. Additionally, we create a
hierarchy based on the generality of each material family. Fig. 11
shows an example of one tree of the hierarchy.
Our hierarchy consists of a set of three-level material trees. The
highest level corresponds to major structural differences between
materials in the category. Metals are conductive, polymers are
composed of long chain molecules, ceramics have a crystalline
structure, and composites are fusions of materials either bonded
together or in a matrix. We define the mid-level (also referred to as
entry-level [37]) categories as groups that separate materials based
primarily on their visual properties. Rubber and paper are flexible,
for example, but paper is generally matte and rubber exhibits little
color variation. The lowest level, fine-grained categories, can often
only be distinguished via a combination of physical and visual
properties. Silver and steel, for example, may be challenging to
distinguish based solely on visual information.
Such a hierarchy is sufficient to cover most natural and manmade materials. In creating our hierarchy, however, we found that
certain categories that are in fact materials did not fit within the
strict definitions described above. For the sake of completeness,
we make the conscious decision to add these mid-level categories
to our data collection process. These categories are: food, water,
and non-water liquids. While food is both a material and an object,
we rely on our annotation process (Sec. 6.2) to ensure we obtain
examples of the former and not the latter.
6.2
Data Collection and Annotation
The mid-level set of categories forms the basis for a crowdsourced
annotation pipeline to obtain material regions from which we may
extract local material patches. We employ a multi-stage process
to efficiently extract both material presence and segmentation
information for a set of images.
The first stage asks annotators to identify materials present
in the image. Given a set of images with materials identified
in each image, the second stage presents annotators with a user
interface that allows them to draw multiple regions in an image.
Each annotator is given a single image-material pair and asked to
mark regions where that material is present. While not required,
our interface allows users to create and modify multiple disjoint
regions in a single image. Images undergo a final validation step
to ensure no poorly drawn or incorrect regions are included.
Each image in the first stage is shown to multiple annotators and a consensus is taken to filter out unclear or incorrect
identifications. While sentinels and validation were not used to
collect segmentations in other datasets, ours is intended for local
11
Fabric
Concrete
Foliage
Asphalt
Ceramic
Concrete
Fabric
Foliage
Food
Glass
Water
Metal
Plastic
Soil
Stone
Wood
Metal
Ceramic
Fig. 12. Example annotation results. Annotators did not hesitate to take
advantage of the ability to draw multiple regions, and most understood
the guidelines concerning regions crossing object boundaries. As a
result, we have a rich database of segmented local material regions.
material recognition. This implies that identified regions should
contain only the material of interest. During collection, annotators
are given instructions to keep regions within object boundaries,
and we validate the final image regions to insure this.
Image diversity is an issue present to varying degrees in
current material image datasets. The Flickr Materials Database
(FMD) [14] contains images from Flickr which, due to the
nature of the website, are generally more artistic in nature.
The OpenSurfaces and Materials in Context datasets [19], [20]
attempt to address this, but still draw from a limited variety
of sources. We source our images from multiple existing image
datasets spanning the space of indoor, outdoor, professional, and
amateur photographs. We use images from the PASCAL VOC
database [38], the Microsoft COCO database [39], the FMD [14],
and the imagenet database [40].
Examples in Fig. 12 show that our annotation pipeline successfully provides properly-segmented material regions within many
images. Many images also contain multiple regions. While the
level of detail for provided regions varies from simple polygons
to detailed material boundaries, the regions all contain single
materials.
Our database currently contains 5845 images with carefullysegmented material regions. For comparison, the MINC [20]
database does not provide any segmented material regions for
training, and only 1798 images for testing.
7
V ISUAL M ATERIAL ATTRIBUTES D ISCOVERED IN
THE MAC-CNN
To verify that the visual material attributes we seek are indeed
present in and can be extracted with our MAC-CNN, we augment
our dataset with annotations to compute the necessary perceptual
distances described in Section 4.1. Using our dataset and these
distances, we derive a category-attribute matrix A and train an
implementation of the MAC-CNN described in Section 5.2.
Fig. 13. Attribute Space Embedding via t-SNE [32]: Many categories,
such as water, food, foliage, soil, and wood, are extremely wellseparated in the attribute space. We find that this separation corresponds with per-category accuracy: well-separated categories are recognized more accurately. While other categories do overlap to some
extent, they still form separate regions in the space.
We train the network on ˜200,000 48 × 48 image patches
extracted from segmented material regions. Optimization is performed using mini-batch stochastic gradient descent with momentum. The learning rate is decreased by a factor of 10 whenever
the validation error increases, until the learning rate falls below
1 × 10−8 .
7.1
Properties of MAC-CNN Visual Material Attributes
We examine the properties of our visual material attributes by
visualizing how they separate materials, computing per-pixel attribute maps to verify that the attributes are being recognized
consistently, and linking the non-semantic attributes with known
semantic material traits (“fuzzy”, “smooth”, etc...) to visualize
semantic content. Figs. 13, 14, and 15 are generated using a test
set of held-out images.
A 2D embedding of material image patches shows that the
attributes (Fig. 13) separate material categories. A number of materials are almost completely distinct in the attribute space, while
a few form overlapping but still distinguishable regions. Foliage,
food, and water form particularly clear clusters. The quality of
the clusters correlates with category recognition accuracy, with
accurately-recognized categories forming well-separated clusters.
Visualizations of per-pixel attribute probabilities in Fig. 14
show that the attributes are spatially consistent. While overfitting
is difficult to measure for weakly-supervised attributes, we use
spatial consistency as a proxy. Spatial consistency is an indicator
that the attributes are not overly-sensitive to minute changes in
local appearance, something that would appear if overfitting were
present. The attributes exhibit correlation with the materials that
induced them: attributes with a strong presence in a material region
in one image often appear similarly in others. The visualizations
also clearly show that the attributes are representing more than
trivial properties such as “flat color” or “bumpy texture”.
Logic regression [34] is a method for building trees that
convert a set of boolean variables into a probability value via
logical operations (AND, OR, NOT). It is well-suited for collections of binary attributes such as ours. Results of performing
logic regression (Fig. 15) from extracted attribute predictions to
known semantic material traits (such as fuzzy, shiny, smooth etc...)
12
Soil
Wood
Fig. 14. The attributes from the MAC-CNN (visualized to the right of
an input image above) exhibit the same properties as our the attributes
we previously discovered independent of material recognition. The attributes form clearly delineated regions, similar to semantic attributes,
and their distributions match as well.
Metallic
Manmade
Organic
Rough
Smooth
Striped
Soft
Fuzzy
Organic
Organic
Smooth
Smooth
Fig. 15. By performing logic regression from our MAC-CNN extracted
attributes to semantic material traits, we are able to extract semantic
information from our non-semantic attributes. We can apply logic regression to material attribute predictions on patches in a sliding window
to obtain per-pixel semantic material trait information. The per-pixel trait
predictions show crisp regions that correspond well with their associated
semantic traits. Traits are independent, and thus the maps contain mixed
colors. Fuzzy and organic in the lower right image, for example, creates
a yellow tint.
show that our MAC-CNN attributes encode material traits with
the roughly same average accuracy (77%) as the our previous
attributes. We may also predict per-pixel trait probabilities in a
sliding window fashion, showing that the attributes are encoding
both perceptual and semantic material properties.
7.2
Local Material Recognition
Our results in Section 7 show that we can successfully discover unnamed visual material attributes in the combined material-attribute
network. If these discovered attributes were not complete, or if
they were not faithfully representing material properties, we would
expect material recognition accuracy to improve by removing
the attribute discovery constraints. While the attribute layers are
auxiliary, they are connected to spatial pooling layers at every
Foliage
Foliage
Fabric
Fabric
Fig. 16. Applying the MAC-CNN in a sliding-window fashion leads to a
set of material category probability maps. These material maps show
that we may obtain coherent regions using only small local patches as
input. The foliage predictions in the bottom right image are reasonable,
as the local appearance is indeed a flower. In the upper right image, the
local appearance of the fence resembles lace (a fabric).
level and thus the attribute constraints affect the entire network.
We in fact find that the average material category accuracy does
not change when the attribute layers are removed.
The average accuracy is 60.2% across all categories. Foliage
is the most accurately recognized, consistent with past material
recognition results in which foliage is the most visually-distinct
category. Paper is the least well-recognized category. Unlike the
artistic closeup images of the FMD, many of the images in our
database come from ordinary images of scenes. Paper, in these
situations, shares its appearance with a number of other materials
such as fabric.
It is important to note that we are recognizing materials
directly from single small image patches, with none of the regionbased aggregation or large patches used before and by other
methods [20]. This is a much more challenging task as the
available information is restricted. These restrictions are necessary,
however, as using large image patches (such as in the MINC
database) would cause the attributes we discover to implicitly
depend on the objects present in the patches and not simply the
materials.
In this paper, our goal is not to introduce a novel material
recognition method but rather to investigate the recognition of
material properties from images. As such, the correct baseline
for comparison is between our previously discovered attributes
and those we discover in the MAC-CNN, not material recognition
accuracy. Material recognition accuracy is used simply to demonstrate that the discovery of visual material attributes is compatible
with material recognition in a single end-to-end trainable network.
As we have shown in our parallel work [18], accurate material
recognition requires the proper integration of local and global
context. Integrating our attribute discovery method with a state-ofthe-art material recognition framework that fuses local and global
context is one promising avenue for future work.
13
While a full dense per-pixel material segmentation framework
is outside the scope of this work, we are able to use the MACCNN to produce per-pixel material probability predictions in a
sliding window fashion. Results in Fig. 16 show that we may still
generate reasonable material probability maps even from purely
local information.
7.3
Novel Material Category Recognition
One prominent application of attributes is in novel category
recognition tasks. Examples of these tasks include one-shot [41]
or zero-shot learning [7]. Zero-shot learning allows recognition
of a novel category from a human-supplied list of applicable
semantic attributes. Since our attributes are non-semantic, zeroshot learning is not applicable here. We may, however, investigate
the generalization of our attributes through a form of one-shot
learning in which we use image patches extracted from a small
number of images to learn a novel category.
To evaluate the application of visual material attributes for
novel category recognition, we train a set of attribute/material
networks on modified datasets each containing a single held-out
category. No examples of the held-out category are present during
training. The corresponding row of the category-attribute matrix
is also removed. The same number of attributes are defined based
on the remaining categories.
For the novel category training, we use a balanced dataset consisting of unseen examples of training categories and a matching
number of images from the held-out category. We also separate a
number of images of the held-out category as final testing samples.
We train a simple binary classifier (a linear SVM) to distinguish
between the training categories and the held-out category based
on either their attribute probabilities, material probabilities, or
both, computed on patches extracted from each input image. We
measure the effectiveness of novel category recognition by the
fraction of final held-out category samples properly identified as
belonging to that category.
Fig. 17 shows plots of novel category recognition effectiveness
as the number of training examples for the held-out category
varies. We can see that the accuracy plateaus quickly, indicating
that the attributes provide a compact and accurate representation
for novel material categories. The number of images we are
required to extract patches from to obtain reasonable accuracy
is generally quite small (on the order of 10) compared to full
material category recognition frameworks which require hundreds
of examples. Furthermore, we include accuracy for the same
predictions based on only material probabilities instead of attribute
probabilities, as well as using a concatenation of both. This clearly
shows that the extracted attributes can expose novel information
in the MAC-CNN that would not ordinarily be available.
8
C ONCLUSION
Material properties provide valuable cues to guide our everyday
interactions with the materials that exhibit them. We aimed to recognize such properties from images, in the form of visual material
attributes, so that we might make this information available for
general scene understanding. Our goal was not only to recognize
these properties, but to do so in a scalable fashion that allows our
method to handle modern large-scale material datasets.
By defining and recognizing a novel set of visual material
attributes – material traits – we showed that material properties are
locally-recognizable and can be aggregated to classify materials in
image regions. To scale the annotation process, we introduced a
novel method that probes our own perception of material appearance, using only weak supervision in the form of yes/no similarity
annotations, to discover unnamed visual material attributes that
serve the same function as fully-supervised material traits. Our
proposed MAC-CNN allows us to apply our attribute discovery
framework to modern large-scale image databases in a seamless
end-to-end fashion. Furthermore, the design of the MAC-CNN
exposes interesting parallels between human and computer vision.
ACKNOWLEDGMENTS
This work was supported by the Office of Naval Research grants
N00014-16-1-2158 (N00014-14-1-0316) and N00014-17-1-2406,
and the National Science Foundation awards IIS-1421094 and IIS1715251. The Titan X used for part of this research was donated
by the NVIDIA Corporation.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth, “Describing Objects by
their Attributes,” in CVPR, 2009, pp. 1778–1785.
V. Ferrari and A. Zisserman, “Learning Visual Attributes,” in NIPS, 2007,
pp. 433–440.
N. Kumar, P. Belhumeur, and S. Nayar, “FaceTracer : A Search Engine
for Large Collections of Images with Faces,” in ECCV, 2008, pp. 340–
353.
G. Patterson and J. Hays, “SUN Attribute Database: Discovering, Annotating, and Recognizing Scene Attributes,” in CVPR, 2012.
C. Hiramatsu, N. Goda, and H. Komatsu, “Transformation from ImageBased to Perceptual Representation of Materials along the Human Ventral
Visual Pathway,” NeuroImage, no. 57, pp. 482–494, 2011.
N. Goda, A. Tachibana, G. Okazawa, and H. Komatsu, “Representation
of the Material Properties of Objects in the Visual Cortex of Nonhuman
Primates,” The Journal of Neuroscience, vol. 34, no. 7, pp. 2660–2673,
2014.
C. Lampert, H. Nickisch, and S. Harmeling, “Learning to Detect Unseen
Object Classes by Between-Class Attribute Transfer,” in CVPR, 2009,
pp. 951–958.
T. L. Berg, A. C. Berg, and J. Shih, “Automatic Attribute Discovery and
Characterization from Noisy Web Data,” in ECCV, 2010, pp. 1–14.
M. Rastegari, A. Farhadi, and D. Forsyth, “Attribute Discovery via
Predictable Discriminative Binary Codes,” in ECCV, 2012, pp. 876–889.
M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” arXiv, vol. abs/1311.3618, 2013.
E. H. Adelson, “On Seeing Stuff: The Perception of Materials by Humans
and Machines,” in SPIE, 2001, pp. 1–12.
Z. Akata, F. Perronnin, Z. Harchaoui, and C. Schmid, “Label-Embedding
for Attribute-Based Classification,” in CVPR, 2013, pp. 819–826.
F. X. Yu, L. Cao, R. S. Feris, J. R. Smith, and S.-F. Chang, “Designing
Category-Level Attributes for Discriminative Visual Recognition,” in
CVPR, 2013, pp. 771–778.
L. Sharan, R. Rosenholtz, and E. Adelson, “Material Perception: What
Can You See in a Brief Glance?” Journal of Vision, vol. 9, no. 8, p. 784,
2009.
C. Liu, L. Sharan, E. H. Adelson, and R. Rosenholtz, “Exploring Features
in a Bayesian Framework for Material Recognition,” in CVPR, 2010, pp.
239–246.
D. Hu, L. Bo, and X. Ren, “Toward Robust Material Recognition for
Everyday Objects,” in BMVC, 2011, pp. 48.1–48.11.
L. Sharan, C. Liu, R. Rosenholtz, and E. H. Adelson, “Recognizing
Materials Using Perceptually Inspired Features,” International Journal
of Computer Vision, 2013.
G. Schwartz and K. Nishino, “Material Recognition from Local Appearance in Global Context,” arXiv, vol. abs/1611.09394, 2017.
S. Bell, P. Upchurch, N. Snavely, and K. Bala, “OpenSurfaces: A Richly
Annotated Catalog of Surface Appearance,” in ACM Transactions on
Graphics (SIGGRAPH 2013), 2013.
——, “Material Recognition in the Wild with the Materials in Context
Database,” in CVPR, 2015.
K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for
Large-Scale Image Recognition,” in ICLR, 2015, pp. 1–14.
14
Foliage
100
40
CNN Out put s Used
At t ribut es
Mat erials
Bot h
20
0
5
10
15
20
25
80
60
40
CNN Out put s Used
At t ribut es
Mat erials
Bot h
20
0
30
0
5
10
N
25
0
30
40
CNNBOut put sBUsed
At t ribut es
Mat erials
Bot h
0
5
10
15
0
5
10
20
25
60
40
0
20
25
30
80
CNNBOut put sBUsed
At t ribut es
Mat erials
Bot h
20
30
15
Glass
100
AccuracyB(%)
60
20
CNN Out put s Used
At t ribut es
Mat erials
Bot h
N
80
AccuracyB(%)
AccuracyB(%)
20
40
20
Soil
100
80
0
15
60
N
Fabric
100
Accuracy (%)
60
Water
100
80
Accuracy (%)
Accuracy (%)
80
0
Wood
100
0
5
10
N
15
N
20
25
60
40
CNNBOut put sBUsed
At t ribut es
Mat erials
Bot h
20
30
0
0
5
10
15
20
25
30
N
Fig. 17. Graphs of novel category recognition accuracy vs. training set size for various held-out categories. The rapid plateau shows that we need
only a small number of examples to define a previously-unseen category. The accuracy difference between feature sets shows that the attributes
are contributing novel information.
[22] Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E.
Hubbard, and L. D. Jackel, “Handwritten Digit Recognition with a BackPropagation Network,” in NIPS, 1990.
[23] S. Shankar, V. K. Garg, and R. Cipolla, “DEEP-CARVING : Discovering
Visual Attributes by Carving Deep Neural Nets,” in CVPR, 2015, pp.
3403–3412.
[24] V. Escorcia, J. C. Niebles, and B. Ghanem, “On the Relationship between
Visual Attributes and Convolutional Networks,” in CVPR, 2015, pp.
1256–1264.
[25] B. Zhou, V. Jagadeesh, and R. Piramuthu, “ConceptLearner: Discovering
Visual Concepts from Weakly Labeled Image Collections,” in CVPR,
2015, pp. 1492–1500.
[26] D. L. K. Yamins, H. Hong, C. F. Cadieu, E. A. Solomon, D. Seibert, and
J. J. DiCarlo, “Performance-optimized hierarchical models predict neural
responses in higher visual cortex,” PNAS, pp. 8619–8624, 2014.
[27] G. Schwartz and K. Nishino, “Visual Material Traits: Recognizing PerPixel Material Context,” in Color and Photometry in Computer Vision
(Workshop held in conjunction with ICCV’13), 2013.
[28] L. Breiman, “Random Forests,” Machine Learning, vol. 45, pp. 5–32,
2001.
[29] A. Barla, F. Odone, and A. Verri, “Histogram intersection kernel for
image classification,” in Proceedings of the International Conference on
Image Processing, 2003.
[30] G. Cybenko, “Approximation by superpositions of a sigmoidal function,”
Mathematics of Control Signals and Systems MCSS, vol. 2, no. 4, pp.
303–314, 1989.
[31] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and
R. R. Salakhutdinov, “Improving Neural Networks by Preventing Coadaptation of Feature Detectors,” 2012.
[32] L. van der Maaten and G. Hinton, “Visualizing Data using t-SNE,” JMLR,
vol. 9, pp. 2579–2605, 2008.
[33] A. Rahimi and B. Recht, “Random Features for Large-Scale Kernel
Machines,” in NIPS, 2008, pp. 1177–1184.
[34] I. Ruczinski, C. Kooperberg, and M. LeBlanc, “Logic Regression,”
Journal of Computational and Graphical Statistics, vol. 12, no. 3, pp.
475–511, 2003.
[35] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Object
Detectors Emerge in Deep Scene CNNs,” in ICLR, 2015.
[36] C.-Y. Lee, S. Xie, P. W. Gallagher, Z. Zhang, and Z. Tu, “DeeplySupervised Nets,” in AISTATS, 2015, pp. 562–570.
[37] V. Ordonez, J. Deng, Y. Choi, A. C. Berg, and T. L. Berg, “From Large
Scale Image Categorization to Entry-Level Categories,” in ICCV, 2013.
[38] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman, “The PASCAL Visual Object Classes
Challenge
2008
(VOC2008)
Results,”
http://www.pascalnetwork.org/challenges/VOC/voc2008/workshop/index.html.
[39] T.-Y. Lin, M. Marie, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollár, and C. L. Zitnick, “Microsoft COCO: Common Objects in
Context,” in ECCV, 2014.
[40] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang,
A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei,
“ImageNet Large Scale Visual Recognition Challenge,” International
Journal of Computer Vision (IJCV), vol. 115, no. 3, pp. 211–252, 2015.
[41] L. Fei-Fei, R. Fergus, and P. Perona, “One-Shot Learning of Object
Categories,” TPAMI, vol. 28, no. 4, pp. 594–611, 2006.
Gabriel Schwartz is a Ph.D. student at Drexel
University in the Department of Computer Science. He received his B.S. and M.S in Computer
Science from Drexel University in 2011. His research interests focus on computer vision and
machine learning.
Ko Nishino is a professor in the Department
of Computing in the College of Computing and
Informatics at Drexel University. He is also an
adjunct associate professor in the Computer and
Information Science Department of the University of Pennsylvania and a visiting associate professor of Osaka University. He received a B.E.
and an M.E. in Information and Communication
Engineering in 1997 and 1999, respectively, and
a PhD in Computer Science in 2002, all from The
University of Tokyo. Before joining Drexel University in 2005, he was a Postdoctoral Research Scientist in the Computer
Science Department at Columbia University. His primary research interests lie in computer vision and include appearance modeling and synthesis, geometry processing, and video analysis. His work on modeling
eye reflections received considerable media attention including articles
in New York Times, Newsweek, and NewScientist. He received the NSF
CAREER award in 2008.
| 1cs.CV
|
1
Undefined 1 (2009) 1–5
IOS Press
arXiv:1708.04745v2 [cs.NE] 28 Mar 2018
Weight-based Fish School Search algorithm
for Many-Objective Optimization
Fernando Buarque de Lima Neto a,∗ , Isabela Maria Carneiro de Albuquerque a and
João Batista Monteiro Filho a
a
Department of Computer Engineering, Polytechnical School of Pernambuco, Rua Benfica, 455, Recife, Brazil
E-mail: {fbln,imca,jbmf}@ecomp.poli.br
Abstract. Optimization problems with more than one objective consist in a very attractive topic for researchers due to its applicability in real-world situations. Over the years, the research effort in the Computational Intelligence field resulted in algorithms
able to achieve good results by solving problems with more than one conflicting objective. However, these techniques do not
exhibit the same performance as the number of objectives increases and become greater than 3. This paper proposes an adaptation
of the metaheuristic Fish School Search to solve optimization problems with many objectives. This adaptation is based on the
division of the school in clusters that are specialized in solving a single-objective problem generated by the decomposition of
the original problem. For this, we used concepts and ideas often employed by state-of-the-art algorithms, namely: (i) reference
points and lines in the objectives space; (ii) clustering process; and (iii) the decomposition technique Penalty-based Boundary
Intersection. The proposed algorithm was compared with two state-of-the-art bio-inspired algorithms. Moreover, a version of the
proposed technique tailored to solve multi-modal problems was also presented. The experiments executed have shown that the
performance obtained by both versions is competitive with state-of-the-art results.
Keywords: Many-objective optimization, swarm-intelligence, evolutionary algorithms, metaheuristics, optimization
1. Introduction
Real-world problems often involve optimizing more
than one objective. In many cases, these objectives are
in conflict with each other and there might not be a single solution capable of optimizing all goals at the same
time. These problems are known as multi-objective
optimization problems. In order to solve them, candidate algorithms should provide as answers a set of
good “trade-off" solutions so that a decision maker
can choose one or more [1]. Mathematically, a multiobjective optimization can be defined by means as
Pareto-dominance [2].
Not rarely, in real-world problems many (i.e. more
than 3) objectives need to be considered [3]. In those
cases, solving the optimization problem is even more
difficult because the increase of objectives number
makes it harder to distinguish a good solution from a
* Corresponding
author. E-mail: [email protected].
not so good one [4]. These many-objective problems
appear in several areas such as software engineering
[5] and automotive engine calibration [6].
The research on evolutionary and swarm intelligence algorithms for solving many-objective optimization problems features straightforward applications
of existing multi-objective algorithms for problems
with 2 or 3 objectives. Despite their success in solving problems with 2 objectives, these algorithms face
many difficulties in solving many-objective optimization problems. The most important challenge faced by
multi-objective algorithms while trying to solve manyobjective problems, mostly for the Pareto-based ones
[2], is the search ability deterioration with the increase
of the number of objectives. When the number of objectives increases, almost all solutions in the population become non-dominated, i.e. equally good [7]. In
those cases, evolutionary algorithms lose their selection pressure and swarm intelligence algorithms lose
their exploitation ability. That is, the algorithms ca-
0000-0000/09/$00.00 c 2009 – IOS Press and the authors. All rights reserved
2
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
pability of converge to good solutions is deteriorated.
Thus, multi-objective algorithms scale up poorly in
many-objective optimization problems [8]. In order to
tackle this issue, researchers in the field have proposed
many different solutions, such as adopting new preference relations as fuzzy Pareto-dominance [9] and dominance [10].
Other approach used by researchers to deal with
the loss of convergence capability is the decomposition of the many-objective problem in simpler singleobjective problems, as it is done in MOEA/D [11]. Algorithms based on this idea have shown good performance in solving many-objective optimization problems [12]. Another advantageous practice in the research of bio-inspired algorithms for solving manyobjective optimization problems is the use of reference points in the objectives space. This to guide the
search process in order to keep the diversity among
the solutions returned by and algorithm. This is an
important factor to be considered because many algorithms tend to favor dominance resistant solutions
[13], which are solutions with high performance in
at least one of the objectives, but with especially
poor performance in the other. An example of successful use of reference points is the NSGA-III algorithm [3]. In recent years some researchers are dedicating their attention to combine both decompositionbased approach and reference points for enhancing
the performance of many-objective optimization algorithms. An example of state-of-the-art algorithm that
uses both techniques is the Many-objective Evolutionary Algorithm Based on Dominance and Decomposition (MOEA/DD) [14] and the θ-Dominance Evolutionary Algorithm (θ-DEA) [4].
This work incorporates decomposition-based approach and reference points to the swarm intelligence
algorithm Fish School Search (FSS) [15] to adapt it
to solve many-objective optimization problems. FSS
versions produced so far already presented good results in optimizing multi-modal functions [16] (e.g.
multi-plateau functions [17]) and multi-objective problems [18]. Also, when compared to other swarm intelligence algorithms, one of FSS’s control mechanisms, the weights of the fishes, natively provides a
way to store how good the solutions found by the algorithm are. This is a relevant aspect because most of
the other swarm intelligence techniques proposed to
solve many-objective problems use an external archive
to store good solutions to guide the search process. The
schemes applied to manage such archives is one of the
issues that current focus of researchers in the field [1],
due to the fact the methods chosen to prune, update
and select solutions within the archive are crucial to
algorithms achieve good performances [19].
The central idea of this work is to extend a multimodal version of FSS, the weight-based Fish School
Search (wFSS) [16], to solve problems with many objectives without storing best found solutions in an external archive. The proposed technique, named weightbased many-objective Fish School Search (wmoFSS),
has the same operators as the original version of FSS
and a new clustering operator is included to promote
diversity during the searching process. This clustering
process is responsible to split the swarm in sub-groups.
Moreover, in wmoFSS the MaOP is decomposed in
sub-problems and each one is assigned to one subgroup that tries to solve it. Hence, the whole swarm
solves all sub-problems at the same time. In comparison with the original version of FSS, the most important difference in wmoFSS is the fact that each fish
has now a vector w of weights in which a single component represents the correspondent weight of a fish
considering the objectives individually. Furthermore, a
modified version of the proposed technique was also
proposed with the goal of improving its performance
on multi-modal many-objective problems.
This paper is organized as follows: Section 2 describes the main concepts within wmoFSS; Section 3
presents the novel algorithm; In Section 4 the experimental design is detailed, results are presented and analyzed; In Section 5, a modified version of wmoFSS
for solving multi-modal problems and further experiments are presented; Section 6 presents conclusions
related with this work as well as future works.
2. Background
2.1. Fish School Search algorithm
The Fish School Search (FSS) is a population-based
metaheuristic algorithm inspired in the behavior of
swimming fishes that expand and contract while looking for food. Each fish n-dimensional location represents a possible solution for the optimization problem.
The algorithm makes use of a feature named weights
for all the fishes, which represents a cumulative account on how successful has been the search for each
fish in the school.
FSS is composed by the Feeding and movement operators, the latter being split into three components: Individual, Collective-instinctive and Collective-volitive.
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
The Individual component of the movement allows every fish in the school to perform a random local search
looking for promising regions in the search space. This
component is computed using the following equation:
xi (t + 1) = xi (t) + rstepind ,
(1)
where xi (t) and xi (t+1) represent the position of a fish
i before and after the movement caused by the Individual component, respectively. r is a n-dimensional vector whose components are uniformly distributed random numbers in the interval [−1, 1]. stepind is a parameter responsible to set the maximum displacement
for this movement. A new position xi (t + 1) is only
accepted if the fitness of the fish improves within position changes. If it is not the case, the fish remains in
the same position and xi (t + 1) = xi (t).
The Collective-instinctive component of the movement is the average of Individual movements for
all fishes in the school. A vector I representing the
weighted average of displacements for each fish is calculated according to:
PS
I=
∆xi ∆fi
i=1
PN
i=1
∆fi
,
(2)
where S is the size of the school. The displacement
represented by vector I is defined in a way that fishes
which experienced a higher improvement will attract
other fishes to its position. After vector I is computed,
every fish will be encouraged to move according to:
xi (t + 1) = xi (t) + I.
(3)
The last movement component, the Collectivevolitive, is used in order to regulate school’s exploration/exploitation ability during the search process.
First of all, the barycenter B of the school is calculated based on each fish position xi and weight wi , as
described in equation 4:
PS
B(t) =
i=1 xi (t)wi (t)
.
PN
i=1 wi (t)
(4)
PS
Then, if the total school weight i=1 wi has increased
from the last to the current iteration, If school total
weight has not improved, fishes are spread away from
the barycenter.
Besides the movement operators, the Feeding operator is used in order to update weights for all fishes
3
according to:
wi (t + 1) = wi (t) +
∆fi
,
max(|∆fi |)
(5)
where wi (t) is the weight parameter for fish i, ∆fi is
the fitness variation between last and new position and
max(|∆fi |) represents the maximum absolute value
of fitness variation among all fishes in the school. wi (t)
is only allowed to vary from 1 up to wscale , which is
an user-defined attribute. The weights of all fishes are
initialized with the value wscale /2.
2.2. Weight-based Fish School Search Algorithm
Introduced in the work of Lima-Neto and Lacerda
[16], wFSS is a weight-based niching version of FSS
intended to provide multiple solutions in a single run
for multi-modal optimization problems. The niching
strategy is based on a new operator called Link Formation. This operator is responsible for defining leaders
for the fishes in order to form sub-swarms and works
according to the following: each fish a chooses randomly another fish b in the school. If b is heavier than
a, then a now has a link with b and follows b (i.e. b
leads a). Otherwise, nothing happens. However, if a
already has a leader c and the sum of weights of a followers is higher than b weight, then a stops following
c and starts to follow b. In each iteration, if a becomes
heavier than its leader, the link will be broken.
In addition to Link Formation operator inclusion,
some modifications were performed in the computation of the collective components of the movement
operators in order to emphasize leaders influence on
each sub-swarm. Also, the Collective-volitive component of the movement is also modified in a sense that
the barycenter is now calculated for each fish with relation to its leader. If the fish does not have a leader, its
barycenter will be its current position.
3. Weight-based Many-Objective Fish School
Search Algorithm
The wmoFSS algorithm incorporates a series of
modifications in FSS and some of its previously proposed versions (FSS-SAR [17] and FSS-NF [20]) in
order to adapt it to solve many-objective optimization
problems. In this section we describe in details the
framework of wmoFSS and highlight the differences
between this algorithm and the original version of FSS.
4
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
3.1. Core Idea
The main idea of wmoFSS is to split up the whole
swarm into sub-swarms, as in wFSS, and also to decompose the many-objective optimization problem
into scalar sub-problems, as in MOEA/D. Unlike
wFSS, wmoFSS substitutes the Link Formation operator by a Clustering operator similar to the proposed
by Deb and Jain [3]. Each sub-swarm is responsible
to solve one sub-problem and all sub-swarms work simultaneously. Thus, the algorithm is able to provide
multiple solutions in a single run. From the other versions of FSS, wmoFSS uses the Stagnation Avoidance
Routine [17] in the Individual component of the movement and a Feeding operator similar to the one used
in NF versions [20]. The wmoFSS algorithm also incorporates the reference points generation scheme and
normalization procedure used in NSGA-III and a comparison criterion based on the θ-dominance utilized in
θ-DEA algorithm.
In comparison with original FSS and all of its versions taken into account in this work, a difference
wmoFSS comprises is the fact that the weight of a fish
is represented by a vector w, instead of a single scalar
value. Each component of w represents the correspondent weight of a fish considering one objective individually. For a problem with m objectives, the weight
vector of a fish i is m-dimensional and represented as:
wi = [wi,1 , wi,2 , ..., wi,m ]T .
(6)
Other important modification incorporated by wmoFSS
is that the Feeding operator is also responsible for to
aggregate the weight vector into a single scalar value
named aggregated weight and denoted as w. This values represents a measure of the quality of a fish and the
θ-dominance criterion takes it into account in order to
compare fishes. The complete description of the aggregation method used to compute w is further explained
in this section.
The following pseudocode describes the main framework of wmoFSS. First, a set of reference points in the
objectives space is generated, the swarm is randomly
initialized following an uniform distribution on the decision space and split in clusters. We used the twolayer reference points generation method proposed by
Deb and Jain [3]. Then, until a stopping criterion is
not met, operators are applied in a sense that firstly
fishes perform a greedy local search, then the Feeding operator is applied to each fish and the collective
components of the movement are calculated considering each cluster individually. At last, the algorithm
sort each cluster based on the Pareto-dominance criterion [2] and return the non-dominated solutions of
each cluster.
1: Create reference points;
2: Initialize fishes randomly;
3: Run Clustering operator;
4: while Stopping condition is not met do
5:
for Each fish on the school do
6:
Move each fish according to its Individual
component;
7:
Run Feeding operator;
8:
Leaders definition;
9:
Move each fish according to its Collectiveinstinctive component;
10:
Move each fish according to its Collectivevolitive component;
11:
end for
12: end while
13: Sort each cluster based on Pareto-dominance;
14: Return non-dominated solutions from each cluster.
3.2. Clustering Operator
The Clustering operator of wmoFSS is responsible
for splitting the population into sub-swarms or clusters, as well as the Link Formation of wFSS. This operator divides the school into clusters according to the
perpendicular Euclidean distance between a fish and
a set of reference lines in the objective space. A reference line λk is defined as a line that goes from the
ideal point to a reference point generated by the approach previously explained. The assignment process
executed by the Clustering operator is similar to the
one used by NSGA-III algorithm. Considering a set of
N reference points and a set of S solutions (fishes), the
b NS0 c closest fishes of a reference line are assigned to
it. If S is not a multiple of N 0 , the S(mod N 0 ) remaining fishes are assigned to its closest reference lines.
Within this clustering procedure, we wanted to avoid
the possibility that a reference line has no fish associated with and some regions of the PF might not be
reached.
3.3. Individual Component of the Movement
As FSS and wFSS, the Individual component of
the movement in wmoFSS is responsible for each fish
greedy local search, as described by Equation 1. On
the other hand, in wmoFSS the aggregated weight (fur-
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
ther defined is this section) is the attribute considered
in the comparison between the current and candidate
positions.
The wmoFSS algorithm incorporates the Stagnation
Avoidance Routine. Hence, if the aggregated weight
of the candidate position is smaller than the current,
the fish moves towards the new position. Otherwise, a
random number in the interval [0, 1] is chosen from an
uniform distribution and compared with a number α
which decays along with the iterations. If the random
number is smaller than α, the fish moves to the candidate position, otherwise, there is no modification in its
position.
5
fj,max (t) is the maximum value found for the j-th objective until iteration t and zj∗ is the j-th coordinate of
the known ideal point or the current value found by the
algorithm so far.
After the calculation of the weight vector for all
fishes in the school, the Feeding operator computes the
aggregated weight w of each fish. In order to do this,
we define that the aggregated weight of a fish i, with
weight vector wi , belonging to the cluster associated
with the reference line λk , is the value of the achievement scalaring function (ASF) Penalty-based Boundary Intersection (PBI) obtained for this fish, which is
calculated by the following equation [11]:
3.4. Feeding Operator
In wmoFSS the weight of a fish is a vector with components corresponding to the success with respect to
each objective. The Feeding operator is then responsible to calculate this components and also to aggregate
them to provide a unique score to each solution in the
swarm.
Its first task is to normalize the solutions in the objective space. This step is necessary due to the fact
the objectives might not vary in the same range and
we want to avoid giving preference to one objective in
spite of the others during the aggregation procedure.
We decided to consider an adaptive normalization process because the extreme objectives values found by
fishes within the search varies during the execution.
Our normalization process is straightforward. All
components of the objective vector of each fish are normalized into an interval which extremes are the minimum and maximum values found so far for this objective. It is noteworthy to remember that the vector
whose coordinates are the best values for the m objectives is the ideal point and represented as z∗ . If the
ideal point of the problem is known, we use it. The vector which coordinates are the worst values found for
each objective is an approximation of the nadir point
(nadir objective vector), denoted as znad . This vector
is constructed from the worst values of each objective
considering just solutions of the Pareto front (PF) [2].
However, we use an approximation of znad due to its
difficult estimation [21].
For a fish i with objective vector fi (t), the j-th component of its weight vector, corresponding to the j-th
objective function, is computed according to the following equation:
fi,j (t) − zj∗ (t)
wi,j (t + 1) =
.
fj,max (t) − zj∗ (t)
(7)
wi = g pbi (wi |λk ) = d1 (wi ) + θd2 (wi ),
(8)
where d1 (wi ) and d2 (wi ) are given by:
d1 (wi ) =
||wTi λk ||
,
||λk ||
d2 (wi ) = ||wi − d1 (wi )
(9)
λk
||.
||λk ||
(10)
θ is a user-defined parameter and here we chose θ = 5
[14]. As we mentioned before, dk,2 consists in the Euclidean distance between wi and its orthogonal projection onto λk and represents a measure of how far wi is
from the reference line associated with its cluster. d1 is
the Euclidean distance between the orthogonal projection of wi onto λk and the ideal point.
It is important to highlight here that an ASF can be
seen as a way of represent the “success" obtained by
a solution in a determined problem and there many
ways to define those functions. Among all possible
choices for achievement scalarization function, as the
often used Weighted-sum and the Tchebycheff, we selected the PBI due to the fact this ASF have shown
to provide good results when used to solve problem
from the DTLZ test suite [3], the set of problems used
to evaluate wmoFSS in this work. Furthermore, this
function allows the user to control the balance between
convergence and diversity by changing the value of θ.
3.5. Leaders Definition
After splitting the population into clusters, to feed
the fishes and obtain the value of their aggregated
weights, the leader of each cluster is defined. We used
6
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
a criterion inspired on the θ-dominance [12] to assign
leadership in each sub-swarm. This dominance relation just considers individuals from the same cluster
and it is based on the PBI aggregation method and consider as quality measure the value of the PBI achievement scalarizing function of a normalized objective
vector in all reference directions considered by the algorithm. However, for wmoFSS, we want that each
sub-swarm is specialized in solving its respective subproblem, thus we choose to consider just the reference
direction associated with the cluster of a fish. Hence,
according to this modified θ-dominance, which we denoted as θ∗ -dominance, a fish i dominates (≺θ∗ ) a fish
j if and only if: (i) i and j belong to the same cluster;
(ii) wi < wj .
Based on the aforementioned definitions, the leaders assignment process consists in to sort each cluster according to θ∗ -dominance and assign leadership to
all θ∗ -non-dominated solutions. Information regarding
clusters, leaders and values of aggregated weight vectors is used to compute the collective components of
the movement.
to determine the value of the barycenter, that is, instead of calculate school’s barycenter, in wmoFSS a
barycenter for each cluster is computed. Given a fish i,
the barycenter of its cluster is computed as follows:
P
k∈Φi
Bi (t) = P
xk w1k
1
k∈Φi wk
.
(12)
The movement towards barycenter is computed exactly as in FSS. Clusters leaders are not affected by
this operator as well as for the Collective-instinctive
component.
4. Experimental Design
where Φi is the cluster of fish i. The parameter L is
equal to 1 if i is a cluster leader and 0 otherwise. As for
the FSS-SAR versions, only the fishes which improved
their aggregated weight are allowed to contribute to
this component computation.
To evaluate wmoFSS performance, we selected the
first four problems of the DTLZ [22] due to the fact
they appear very often in literature [23]. We compared
the Inverted Generational Distance (IGD) metric [24]
obtained by wmoFSS with the results of 2 state-of-theart algorithms: Many-objective Particle Swarm Optimization (MaOPSO) [19] and NSGA-III. We choose
both algorithms due to: (i) wmoFSS incorporates many
aspects of NSGA-III, such as the use of reference
points and clustering the population, (ii) MaOPSO
is a swarm-intelligence technique very different from
wmoFSS which uses an external archive to guide the
search. Results used for comparison for both algorithms were provided by MaOPSO authors and are the
same used in [19].
Before comparing wmoFSS with the state-of-the-art
algorithms, we performed a parameter selection step to
choose the values of stepind , stepvol and number of
fishes and also the clustering and normalization methods. In this preliminary step, we adopt a methodology
inspired on the Factorial Analysis method [25]. After
this step, the results of median, maximum and minimum values of IGD for the m-objective DTLZ1, 2,
3 and 4 problems with m = {3, 5, 10} obtained for
wmoFSS are compared with the NSGA-III algorithm
[3] and the MaOPSO [19].
3.7. Collective-Volitive Component of the Movement
Operator
4.1. Results
As in Collective-instinctive component computation, in the Collective-volitive component of the movement only fishes in the same cluster are considered for
We show in Table 1 the results obtained within
the experiments involving wmoFSS, NSGA-III and
MaOPSO.
3.6. Collective-Instinctive Component of the
Movement Operator
The Collective-instinctive component of the movement in wmoFSS is computed considering just fishes
in the same cluster. The values of aggregated weight
are used to measure the increase of success achieved
for each fish, as in FSS-NF versions. This measure is
defined as ∆w = w(t) − w(t + 1) for minimization
problems and ∆w = w(t+1)−w(t) for maximization
problems. For a fish i, its position after the collectiveinstinctive movement is computed according to:
xi (t + 1) = xi (t) + L
!
P
k∈Φi ∆xk ∆w k
P
, (11)
k∈Φi ∆w k
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
Table 1
Results of median, maximum and minimum IGD obtained on the
DTLZ1, 2, 3 and 4 test problems for m = {3, 5, 10}
3
DTLZ1
5
10
3
DTLZ2
5
10
3
DTLZ3
5
10
3
DTLZ4
5
10
wmoFSS
NSGA-III
MaOPSO
3.63E-02
7.27E-02
1.66E-02
1.85E-02
1.51E-03
1.74E-03
1.37E-03
1.59E-03
6.98E-04
6.99E-04
6.98E-04
8.25E-04
2.27E-02
9.78E-03
1.22E-02
2.32E-02
8.21E-03
1.88E-03
1.50E-03
1.42E-03
2.62E-03
1.38E-03
8.25E-04
8.24E-04
1.43E-03
1.44E-03
1.42E-03
4.44E-03
4.67E-03
4.24E-03
4.71E-03
4.80E-03
4.62E-03
6.07E-03
3.77E-03
4.11E-03
3.55E-03
1.45E-02
1.53E-02
1.10E-02
1.23E-02
2.27E-03
2.27E-03
2.27E-03
2.94E-03
2.94E-03
2.94E-03
4.95E-03
6.13E-03
5.97E-03
1.25E-02
9.93E-03
4.97E-03
4.94E-03
1.04E+00
1.41E+00
3.89E-03
4.36E-03
2.27E-03
2.67E-01
5.29E-01
4.67E-01
5.72E-01
3.12E-01
1.69E-01
2.66E-01
8.83E-02
3.52E-03
1.47E-02
1.91E-02
4.88E-03
1.20E-02
1.36E-02
6.45E-03
2.27E-03
2.94E-03
2.95E-03
2.94E-03
4.95E-03
7.54E-03
4.93E-03
8.21E-03
9.29E-03
7.54E-03
6.15E-03
6.58E-03
3.80E-03
4.14E-03
3.60E-03
5.00E-03
5.24E-03
2.45E-03
2.59E-03
2.36E-03
3.88E-03
4.09E-03
5.86E-03
6.33E-03
6.50E-03
6.22E-03
4.83E-03
5.10E-03
5.20E-03
5.02E-03
3.51E-03
4.92E-03
4.95E-03
4.90E-03
From Table 1, one can notice that, besides the multimodal problems, wmoFSS shows competitive performance in comparison with NSGA-III and MaOPSO.
Regarding NSGA-III, wmoFSS achieved results of
the same order for the 3-objective DTLZ2 and outperformed this algorithm on the 5 and 10-objective
DTLZ2. For the DTLZ4 problem, wmoFSS did not
outperformed NSGA-III in any of the cases considered, but achieved median, maximum and minimum
7
IGD of the same order as NSGA-III. In comparison
with MaOPSO, wmoFSS did not achieved better results for either DTLZ2 or DTLZ4, but the median,
maximum and minimum IGD obtained have the same
order of magnitude. From these results, it is also possible to see that wmoFSS did not achieve good results
for the DTLZ1 and DTLZ3 problems. This problems
are multi-modal instances of the DTLZ test suite and
have multiple local PFs. We attribute this difficulty of
the proposed to technique to the smaller capability of
a greedy search process does not get trapped into local
optima.
To provide a graphical analysis of results, we show
in Figures 1, 2 and 3 scatter plots of the set of solutions returned by the compared techniques for the 3objective DTLZ2 problem. In both cases the best IGD
results were selected. A random sample of the true PF
is represented by the smaller dots and the set returned
by each algorithm is represented by the bigger dots.
Fig. 1. Scatter plot of the solutions returned by wmoFSS on the 3-objective DTLZ2 problem
Fig. 2. Scatter plot of the solutions returned by MaOPSO on the
3-objective DTLZ2 problem
One can notice that, besides the fact that wmoFSS
has worse IGD in this instance, the set of solutions re-
8
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
crossover a fish and its leader to create a new position
and use it to generate movement directions during the
computation of the Individual component.
Considering a fish x(t) with leader xl (t), to calculate its candidate position of the Individual component
of the movement x(t + 1), firstly, it is necessary to create a new position in the search space with the SBX
taking as parents x and xl . The i-th component of the
generated position y(t) is given by:
Fig. 3. Scatter plot of the solutions returned by NSGA-III on the
3-objective DTLZ2 problem
0.5 (xi (t) + xl,i (t)) − β|xi (t) − xl,i (t)| , if v ≤ 0.5
(13)
0.5 (xi (t) + xl,i (t)) + β|xi (t) − xl,i (t)| , if v > 0.5
with
turned by wmoFSS is able to cover some regions of
the PF that the MaOPSO was not able to. Regarding
the NSGA-III algorithm, Figure 3 shows that this algorithm is able to multiple solutions on the true PF with
good diversity, but the returned set is not uniformly
spread as the solution returned by wmoFSS.
5. wmoFSS Solving Multi-modal DTLZ problems
As we showed in the last sections, wmoFSS algorithm is not able to achieve good results for the considered multi-modal instances of DTLZ test suite. Thus,
in this section, we propose some modifications in the
algorithm in order to increase its search effectiveness
and then check whether the proposed changes is able to
help wmoFSS to achieve better results on the DTLZ1
and DTLZ3 problems.
Within the development of this work, we notice that
there was room for improvement on FSS’s Individual component of the movement. This affirmative can
be justified by the fact this component is calculated
through a random search with depth limited by the parameter stepind . Therefore, when a fish is moved by
this component, it goes towards a direction found randomly and the algorithm does not take advantage to
the information of the search process it already has at
hand. Taking this fact into account, we decided to use
the Simulated Binary Crossover (SBX) [26] based approach in order to determine the direction a fish will
move within the Individual component of the movement, giving rise to the wmoFSS-SBX algorithm. The
SBX operator uses as parents two individuals from the
current population to create two new solutions, as the
regular crossover operators. However, the main difference of SBX is that it allows the creation of solutions
near their parents. Hence, the goal is to use SBX to
(
1
(αu) ηc +1 , if u ≤ α1
β=
1
1
( 2−αu
) ηc +1 , otherwise
(14)
where u and v are uniformly distributed real random
numbers, ηc is a parameter named Distribution Index
and the definition of α can be found at [26]. After computing y(t), the candidate position is calculated according to:
x(t + 1) = x(t) + stepind (t)
(y(t) − x(t))
. (15)
||y(t) − x(t)||
From Equation 15 one can see that in wmoFSS-SBX
fishes move to positions determined by a SBX instead
of random directions. After the calculation of x(t +
1), the algorithm follows the same steps as wmoFSS.
Thus, the aggregated weight of x(t + 1) is computed
and compared with the aggregated weight of x(t). In
wmoFSS-SBX we decided to use the SAR mode 0, that
is, there is no worsening allowance and the fish moves
to x(t + 1) only if its aggregated weight is better. The
Feeding, Clustering, leaders definition and collective
operators remains the same as in wmoFSS.
In order to evaluate wmoFSS-SBX we utilized the
same stepind , stepvol , normalization procedure and
clustering method selected for wmoFSS. As in the
work of Deb and Agrawal [26] we decided to use
ηc = 1, aiming to obtain solutions close to the fathers,
but not as close as the common use in literature, where
authors choose higher values for ηc , such as 20 [11] or
30 [3], [14], [4], [27]. This decision was taken in order
to provide to a fish a more diverse set of possibilities
for its movement and improve its chance of going to a
better position.
Another important aspect of the wmoFSS-SBX is its
greater dependence on the leader of the cluster. In the
9
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
wmoFSS, the role of the leader of a cluster is just not
to be affected by the collective operators, which represents a type of elitism. On the other hand, in wmoFSSSBX the leader plays a key role during the search process because it has a great impact on the generation of
new solutions. Hence, it is fundamental for wmoFSSSBX a cluster be able to improve its leader during the
search. In order to do so, we increase the number of
fishes per cluster by decreasing the number of reference points and also increasing the school size to 1000
fishes.
5.1. Study of the influence of collective operators on
wmoFSS-SBX
Within preliminary tests on wmoFSS-SBX, we notice that, in some cases, the Collective-instinctive component of the movement and the Collective-volitive
component of the movement seemed to decrease
the algorithm performance, mostly the Collectiveinstinctive component. Hence, this hypothesis was verified according to the following experiments. First,
three versions of wmoFSS-SBX were defined: (A):
only Individual and Collective-volitive components;
(B): only Individual component; (C): all components.
Then, two values of the θ parameter were used. This
aspect was studied in these experiments due to the
fact in wmoFSS-SBX there are more fishes per cluster,
which helps the algorithm to ensure that each subproblem will be optimally solved, increasing the diversity
of the results. Hence, we want to observe the behavior of wmoFSS-SBX when we decrease θ from 5 to 1,
emphasizing convergence in spite of diversity.
The experimental setup employed was: 20 runs and
10000 iterations, consider as preliminary performance
measures mean and standard deviation of IGD, and
analyze statistical significance of results using the
Kruskal-Wallis test with 95% of confidence. Results of
mean and standard deviation of the IGD obtained by
the three versions are presented in Table 2. Results of
Kruskal-Wallis test are also shown in Table 2, where +
indicates that a certain value of θ was significant better,
− indicates the opposite, and = indicates that IGD values achieved with both θ are indifferent. From Table
2, one can notice that, for the DTLZ1 problem, θ = 1
provided statistically significant better IGD values. On
the other hand, the same is not observed for DTLZ3.
Version A achieved statistically significant better results on DTLZ3 with θ = 1, version C with θ = 5
and, for version B, the use of both values of θ provided
statistical indifferent results.
Table 2
Results of mean and standard deviation for IGD obtained by
wmoFSS-SBX on the 5-objective DTLZ1 and DTLZ3 problems
θ
Mean
SD
1
5
1
5
3.25E-03+
3.55E-03
3.27E-03+
3.73E-03
2.57E-05+
2.42E-04
3.87E-05+
3.00E-04
C
1
5
7.11E-03+
1.36E-02
2.36E-03+
6.28E-03
A
1
5
1.81E-01+
2.50E-01
2.09E-01+
2.76E-01
1
5
1
5
2.33E-02=
2.53E-02
2.41E+00−
1.43E+00
5.19E-03=
7.12E-03
2.77E-01−
5.69E-01
A
DTLZ1
DTLZ3
B
B
C
In order to define which is the version of wmoFSSSBX more appropriated to solve DTLZ1 and DTLZ3,
versions A, B and C are compared using the best value
of θ found to solve each problem. For version A, θ = 1
for both DTLZ1 and DTLZ3. For version 3, θ = 1 for
DTLZ1 and we also selected θ = 1 for DTLZ3, as both
values are indifferent. For version C, θ = 1 for DTLZ1
and θ = 5 for DTLZ3. Results of the Kruskal-Wallis
test on IGD values with 95% of confidence are shown
in Table 3.
Table 3
Results of Kruskal-Wallis tests on the IGD values for versions A, B
and C of wmoFSS-SBX with the best value of θ
A
DTLZ1
A
B
C
=
−
DTLZ3
A
B
C
−
−
B
C
=
+
+
−
+
+
+
−
From Table 3, one can notice that version B with
θ = 1 is the best choice among the versions tested to
solve DTLZ3. For DTLZ1, the performances of versions A and B are statistically indifferent, which leads
us to conclude that the presence of the Collectivevolitive component has no effect in the search for the
considered case. We decided to tackle DTLZ1 with
version B due to the fact it demands less computational effort, as it does not necessitate to calculate the
Collective-volitive component.
10
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
Table 4
5.2. Comparison between wmoFSS and wmoFSS-SBX
Now that a version of wmoFSS-SBX that better
solve DTLZ1 and DTLZ3 was selected, we are able to
compare it with wmoFSS. As we did in the comparison of wmoFSS with NSGA-III and MaOPSO, both
algorithms solved the {3, 5, 10}- objective DTLZ1, 2,
3 and 4 problems and the obtained median, maximum
and minimum IGD values among the 20 runs are taken
into account in the analysis. Table 4 shows the results.
From Table 4, one can observe that the use of SBX
in the Individual component and all modifications incorporated by wmoFSS-SBX, in fact, improved the
results on DTLZ1 and DTLZ3. The performance on
DTLZ2 and DTLZ4 was worse. To have a more detailed analysis of the results we show in Figures 4, 5,
6, and 7 the scatter plot of the run for which wmoFSSSBX obtained the best IGD on the 3-objective DTLZ1,
2, 3 and 4. The smaller dots represent the true PF and
the bigger dots represent the set of solutions found by
wmoFSS.
Results of median, max. and min. IGD obtained by wmoFSS and
wmoFSS-SBX on DTLZ1, 2, 3 and 4 for m = {3, 5, 10}.
3
DTLZ1
5
10
3
DTLZ2
5
10
3
DTLZ3
5
10
3
Fig. 4. Scatter plot of the solutions returned by wmoFSS-SBX for
the 3-objective DTLZ1.
DTLZ4
5
10
Fig. 5. Scatter plot of the solutions returned by wmoFSS-SBX for
the 3-objective DTLZ2.
wmoFSS
wmoFSS-SBX
3.63E-02
7.27E-02
1.66E-02
1.85E-02
4.00E-03
5.16E-03
3.48E-03
3.28E-03
2.27E-02
9.78E-03
1.22E-02
2.32E-02
8.21E-03
3.37E-03
3.21E-03
2.41E-03
2.49E-03
2.35E-03
4.44E-03
4.67E-03
4.24E-03
4.71E-03
4.80E-03
4.62E-03
6.07E-03
8.44E-03
9.34E-03
6.79E-03
1.06E-02
1.23E-02
8.70E-03
9.95E-03
6.13E-03
5.97E-03
1.05E-02
8.94E-03
1.04E+00
1.41E+00
6.79E-02
1.42E-01
5.29E-01
2.18E-02
5.72E-01
3.12E-01
1.69E-01
2.66E-01
8.83E-02
3.19E-02
2.18E-02
3.93E-02
1.63E-02
2.42E-02
3.57E-02
1.53E-02
8.21E-03
9.29E-03
7.54E-03
6.15E-03
6.58E-03
2.33E-02
2.76E-02
1.37E-02
1.05E-02
1.42E-02
5.86E-03
6.33E-03
6.50E-03
6.22E-03
8.66E-03
9.31E-03
1.03E-02
8.26E-03
From Figure 4 it is possible to notice that, unlike
wmoFSS, wmoFSS-SBX is able to find solutions of
the true PF of the 3-objective DTLZ1 problem. However, we can also see that the returned set of solutions
did not cover all the PF and the extreme regions of the
PF are not represented in the PF found by wmoFSSSBX. Observing Figures 5 and 7 one can notice that
wmoFSS-SBX is also able to find solutions on the true
PF of the 3-objective DTLZ2 and DTLZ4 but with
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
11
6. Conclusion
Fig. 6. Scatter plot of the solutions returned by wmoFSS-SBX for
the 3-objective DTLZ3.
Fig. 7. Scatter plot of the solutions returned by wmoFSS-SBX for
the 3-objective DTLZ4.
poor spread within the PF. This fact is also evidenced
by the worse IGD results obtained by wmoFSS-SBX in
Table 4. We attribute this to the use of fewer reference
points by wmoFSS-SBX, which decreases the number
of sub-problems to be solved and, as consequence, the
diversity of the solutions. At last, from Figure 6, we
see that wmoFSS-SBX is able to find only one solution on the true PF of the 3-objective DTLZ3 problem,
which is not a good result, but consists in an improvement in comparison with wmoFSS, which is not able
to find any solution in the true PF of this test instance.
Summarizing, wmoFSS-SBX indeed presented a
better performance on DTLZ1 and DTLZ3 and was
also able to find optimal solutions on DTLZ2 and
DTLZ4 problems, but with a much poor diversity in
comparison with wmoFSS, mostly in the DTLZ4 problem, for which wmoFSS-SBX was able to find a small
set of solutions.
In this work we conceived, implemented and tested
the Weight-based Many-Objective Fish School Search
algorithm (wmoFSS), an adaptation of the metaheuristic Fish School Search to solve many-objective optimization problems. Our goal was to propose a competitive approach that encompasses the necessary aspects and concepts desirably the ones readily applied
by state-of-the-art algorithms.
When comparing wmoFSS with state-of-the-art algorithms, we concluded that the proposed technique
achieved competitive results in the uni-modal DTLZ
instances, as it achieved IGD values of the same magnitude order as the other algorithms, and outperformed
NSGA-III on the {5, 10}-objective DTLZ2.
To provide to the initial version of wmoFSS ways
to solve multi-modal instances of DTLZ problems,
we substituted the random search on FSS’s Individual
component of the movement by a local search guided
by directions obtained with SBX crossover taking a solution and its leader as parents. The goal of this modification is to use readily available information to improve wmoFSS search effectiveness. This version was
called wmoFSS-SBX.
wmoFSS-SBX has shown to be more effective than
wmoFSS to solve DTLZ1 and DTLZ3. On the other
hand, it decreased the diversity of the solutions found
for the DTLZ2 and DTLZ4. For DTLZ1, wmoFSSSBX was able to find a set of solutions well spread
within the true Pareto-optimal front, but the extreme
regions were not covered. For DTLZ3, wmoFSS-SBX
was able to find only one solutions in the true Paretooptimal front, which is a result with very poor diversity, even so its performance is better than wmoFSS’s,
that was not able to find a single solution on the true
Pareto-optimal front.
In summary, wmoFSS shows a promising performance in the solution of multiple instances of DTLZ
problems. Furthermore, wmoFSS-SBX seems to be a
good direction in order to improve wmoFSS to solve
multi-modal DTLZ instances. We point out in the next
section future works to improve wmoFSS performance
and to analyze its behavior in more details. As future
works, we suggest the use of a non-normalized test
suite, such as the WFG [28]. Furthermore, Estimation
of Distribution Algorithm (EDA) [29] can be used to
generate movement directions in the Individual component of the movement [8].
12
F. B. de Lima Neto et al. / Weight-based Fish School Search algorithm for Many-Objective Optimization
References
[1] M. Reyes-Sierra and C. C. Coello, “Multi-objective particle
swarm optimizers: A survey of the state-of-the-art,” International journal of computational intelligence research, vol. 2,
no. 3, pp. 287–308, 2006.
[2] K. Deb, Multi-objective optimization using evolutionary algorithms. John Wiley & Sons, 2001, vol. 16.
[3] K. Deb and H. Jain, “An evolutionary many-objective optimization algorithm using reference-point-based nondominated
sorting approach, part i: Solving problems with box constraints,” IEEE Transactions on Evolutionary Computation,
vol. 18, no. 4, pp. 577–601, 2014.
[4] Y. Yuan, H. Xu, B. Wang, and X. Yao, “A new dominance
relation-based evolutionary algorithm for many-objective optimization,” IEEE Transactions on Evolutionary Computation,
vol. 20, no. 1, pp. 16–37, 2016.
[5] K. Praditwong, M. Harman, and X. Yao, “Software module
clustering as a multi-objective search problem,” IEEE Transactions on Software Engineering, vol. 37, no. 2, pp. 264–282,
2011.
[6] R. J. Lygoe, M. Cary, and P. J. Fleming, “A real-world application of a many-objective optimisation complexity reduction
process,” in International Conference on Evolutionary MultiCriterion Optimization. Springer, 2013, pp. 641–655.
[7] H. Ishibuchi, N. Tsukamoto, and Y. Nojima, “Evolutionary many-objective optimization: A short review.” in IEEE
congress on evolutionary computation, 2008, pp. 2419–2426.
[8] O. R. Castro, R. Santana, and A. Pozo, “C-multi: A competent
multi-swarm approach for many-objective problems,” Neurocomputing, vol. 180, pp. 68–78, 2016.
[9] M. Farina and P. Amato, “A fuzzy definition of" optimality" for
many-criteria optimization problems,” IEEE Transactions on
Systems, Man, and Cybernetics-Part A: Systems and Humans,
vol. 34, no. 3, pp. 315–326, 2004.
[10] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining
convergence and diversity in evolutionary multiobjective optimization,” Evolutionary computation, vol. 10, no. 3, pp. 263–
282, 2002.
[11] Q. Zhang and H. Li, “Moea/d: A multiobjective evolutionary algorithm based on decomposition,” IEEE Transactions on
evolutionary computation, vol. 11, no. 6, pp. 712–731, 2007.
[12] Y. Yuan and H. Xu, “Multiobjective flexible job shop scheduling using memetic algorithms,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 1, pp. 336–353,
2015.
[13] K. Ikeda, H. Kita, and S. Kobayashi, “Failure of pareto-based
moeas: does non-dominated really mean near to optimal?”
in Evolutionary Computation, 2001. Proceedings of the 2001
Congress on, vol. 2. IEEE, 2001, pp. 957–962.
[14] K. Li, K. Deb, Q. Zhang, and S. Kwong, “An evolutionary
many-objective optimization algorithm based on dominance
and decomposition,” IEEE Transactions on Evolutionary Computation, vol. 19, no. 5, pp. 694–716, 2015.
[15] C. J. Bastos Filho, F. B. de Lima Neto, A. J. Lins, A. I. Nascimento, and M. P. Lima, “A novel search algorithm based on
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
fish school behavior,” in Systems, Man and Cybernetics, 2008.
SMC 2008. IEEE International Conference on. IEEE, 2008,
pp. 2646–2651.
F. Buarque De Lima Neto and M. Gomes Pereira de Lacerda,
“Weight based fish school search,” in Systems, Man and Cybernetics (SMC), 2014 IEEE International Conference on. IEEE,
2014, pp. 270–277.
J. B. Monteiro Filho, I. M. C. Albuquerque, F. B. d. L. Neto,
and F. V. Ferreira, “Optimizing multi-plateau functions with
fss-sar (stagnation avoidance routine),” in IEEE-Symposium
Series on Computational Intelligence, 2016, pp. 1–23.
C. J. Bastos-Filho and A. C. Guimarães, “Multi-objective fish
school search,” International Journal of Swarm Intelligence
Research (IJSIR), vol. 6, no. 1, pp. 23–40, 2015.
E. Figueiredo, T. Ludermir, and C. Bastos-Filho, “Many objective particle swarm optimization,” Information Sciences, vol.
374, pp. 115–134, 2016.
J. B. Monteiro Filho, I. M. C. Albuquerque, F. B. L. Neto, and
F. V. S. Ferreira, “Improved search mechanisms for the fish
school search algorithm,” in International Conference on Intelligent Systems Design and Applications. Springer, 2016, pp.
362–371.
K. Deb, K. Miettinen, and S. Chaudhuri, “Toward an estimation of nadir objective vector using a hybrid of evolutionary and
local search approaches,” IEEE Transactions on Evolutionary
Computation, vol. 14, no. 6, pp. 821–841, 2010.
K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, “Scalable multi-objective optimization test problems,” in Evolutionary Computation, 2002. CEC’02. Proceedings of the 2002
Congress on, vol. 1. IEEE, 2002, pp. 825–830.
C. von Lücken, B. Barán, and C. Brizuela, “A survey on multiobjective evolutionary algorithms for many-objective problems,” Computational Optimization and Applications, vol. 58,
no. 3, pp. 707–756, 2014.
P. A. Bosman and D. Thierens, “The balance between proximity and diversity in multiobjective evolutionary algorithms,”
IEEE transactions on evolutionary computation, vol. 7, no. 2,
pp. 174–188, 2003.
D. C. Montgomery, Design and analysis of experiments. John
Wiley & Sons, 2008.
K. Deb and R. B. Agrawal, “Simulated binary crossover for
continuous search space,” Complex systems, vol. 9, no. 2, pp.
115–148, 1995.
Y. Xiang, Z. Yuren, M. Li, and Z. Chen, “A vector angle based
evolutionary algorithm for unconstrained many-objective optimization,” IEEE Transactions on Evolutionary Computation,
2016.
S. Huband, P. Hingston, L. Barone, and L. While, “A review of multiobjective test problems and a scalable test problem toolkit,” IEEE Transactions on Evolutionary Computation,
vol. 10, no. 5, pp. 477–506, 2006.
P. Larrañaga, H. Karshenas, C. Bielza, and R. Santana, “A review on probabilistic graphical models in evolutionary computation,” Journal of Heuristics, vol. 18, no. 5, pp. 795–819,
2012.
| 9cs.NE
|
Testing-Based Forward Model Selection
Damian Kozbur
University of Zürich
Department of Economics
arXiv:1512.02666v3 [math.ST] 11 Apr 2018
Schönberggasse 1, 8001 Zürich
e-mail: [email protected].
Abstract: This paper introduces and analyzes a procedure called Testing-based forward model selection (TBFMS) in linear regression problems. This procedure inductively selects covariates that add predictive power into a working statistical model
before estimating a final regression. The criterion for deciding which covariate to include next and when to stop including covariates is derived from a profile of traditional
statistical hypothesis tests. This paper proves probabilistic bounds for prediction error
and the number of selected covariates, which depend on the quality of the tests. The
bounds are then specialized to a case with heteroskedastic data with tests derived from
Huber-Eicker-White standard errors. TBFMS performance is compared to Lasso and
Post-Lasso in simulation studies. TBFMS is then analyzed as a component into larger
post-model selection estimation problems for structural economic parameters. Finally,
TBFMS is used to illustrate an empirical application to estimating determinants of
economic growth.
MSC 2010 subject classifications: 62J05, 62J07, 62L12.
JEL subject classifications: C55.
Keywords and phrases: model selection, forward regression, sparsity, hypothesis
testing.
*The content of this paper draws from two separate papers posted on ArXiv
(https://arxiv.org/abs/1702.01000 ([42]) and https://arxiv.org/abs/1512.02666 (the
current paper)), which were originally created as two different projects. The two
projects have now been merged in preparation for the publication process. No original
material is submitted simultaneously to multiple peer-reviewed journals. A mechanical
statement of a special case of TBFMS appears in [43], which is published in the
Papers and Proceedings issue of the American Economic Review, which is not peer
reviewed. That version neither claims nor derives any theoretical results.
∗
First version: Jan 2015. This version April 12, 2018. I gratefully acknowledge helpful discussion with
Christian Hansen, Tim Conley, Dan Zhou, attendants at the ETH Zürich Seminar für Statistik Research
Seminar, attendants at the Center for Law and Economics Internal Seminar, as well as financial support
from UZH and from the ETH Fellowship program.
1
Damian Kozbur/Testing-Based Forward Model Selection
2
1. Introduction
This paper considers a procedure called Testing-Based Forward Model Selection (TBFMS)
for high-dimensional econometric problems, which are estimation problems characterized
by settings in which the number of observed characteristics per observation in the data
is large.1 High-dimensional econometrics is a leading area of current research because of
recent rapid growth in data availability and computing capacity coupled with the important
need to extract as much useful information from data in a way that allows precise and
rigorous testing of scientific hypotheses. Working within the flexibility of a high-dimensional
framework allows researchers to fully exploit richer data sets both in prediction problems
and in structural inference problems.
The primary settings of this paper are high-dimensional sparse linear regression models,
in which the number of covariates is allowed to exceed the sample size. A key challenge
with a high-dimensional data set is that estimation requires dimension reduction or regularization to avoid statistical overfitting. One regularizing structure used often in the recent
econometrics and statistics literature is sparsity. A sparsity assumption imposes that the
regression function relating the outcome and the covariates can be approximated by a regression of the outcome on a small, ex ante unknown subset of covariates. Under sparsity,
there are several consistent estimation procedures (further reviewed below) that work by
enforcing that the estimated regression function be sparse or small under an appropriate
norm.
An appealing class of techniques for high-dimensional regression problems are Greedy
algorithms. These are procedures that inductively select individual covariates into a working
model (i.e., a collection of covariates) until a stopping criterion is met. A linear regression
restricted to the final selected model is then estimated. A leading example is Simple Forward
Selection,2 which chooses the covariate that gives the highest increase of in-sample Rsquared above the previous working model. This class of techniques is widely used because
they are intuitive and simple to implement.
1
High-dimensional data arise through a combination of two ways. The data may be intrinsically high-
dimensional in that many different characteristics per observation are available. Alternatively, even when
the number of available variables is relatively small, researchers rarely know the exact functional form with
which the variables enter the model of interest and are thus faced with a large set of potential variables
formed by different ways of interacting and transforming the underlying variables.
2
Simple Forward Selection is not standard nomenclature, but is used here in order to have a parallel
language with Testing-Based Forward Model Selection. The literature is varied and uses several names
including Forward Regression and Forward Stepwise Regression.
Damian Kozbur/Testing-Based Forward Model Selection
3
In practice, deciding which covariate gives the best additional predictive power is complicated by the fact that outcomes are observed with noise or are partly idiosyncratic. For
example, in linear regression, a variable associated to a positive increment of in-sample
R-squared upon inclusion may not add any predictive power out-of-sample. Statistical hypothesis tests offer one way to determine whether a variable of interest is likely to improve
out-of-sample predictions. Furthermore, in many econometric and statistical applications,
the classical assumption of independent and identically distributed data is not always appropriate. One example of this is the presence of heteroskedastic disturbances. In such settings,
higher R-squared resulting from inclusion of one variable relative to another need not be a
signal that the first variable is a better choice. More generally, model selection procedures
tailored to the classical assumptions may have inferior performance when applied to more
realistic data-generating processes.
The availability of hypothesis tests for diverse classes of problems and settings motivates
the introduction of a testing-based strategy. Mechanically, TBFMS begins with an empty
model. The procedure then tests whether any covariates provide additional predictive capability in the population. The selection stops when no tests return a significant covariate.
Selection into the model is then based on the largest value of an associated test statistic.
Though the prior literature has not analyzed greedy algorithms that explicitly incorporate hypothesis testing, there are several earlier analyses of Simple Forward Selection.3
[68] gives bounds on the performance and number of selected covariates under a β-min
condition which requires the minimum magnitude of non-zero coefficients to be suitably
bounded away from zero. [72] and [64] prove performance bounds for greedy algorithms
under a strong irrepresentability condition, which restricts the empirical covariance matrix
of the predictors. [27] prove bounds on the relative performance in population R-squared
of the a forward selection based model (relative to infeasible R-squared) when the number of variables allowed for selection is fixed. In addition to Simple Forward Selection,
there are several related procedures in which estimation is done in stages. These include
a method that is not strictly greedy called Forward-Backward Selection, which proceeds
similarly to Simple Forward Selection but allows previously selected covariates to be discarded from the working model at certain steps (see [72]). Another related class of methods
are called boosting methods. Boosting methods inductively select covariates predictive of a
linear combination of estimated residuals and the outcome at each step (among many other
3
A mechanical statement of TBFMS in a specific case with heteroskedastic data is given in [43], which
is a Papers and Proceedings publication by the current author and is not peer-reviewed. No derivations or
theoretical properties are stated or claimed there.
Damian Kozbur/Testing-Based Forward Model Selection
4
references, see [17], see also [48] for additional results and applications in econometrics.)
As a preliminary, (before proceeding to the analysis of TBFMS), this paper proves new
bounds on the predictive performance and number of selected covariates for Simple Forward Selection. The conditions required here are weaker that those used in [72] and [64] and
impose no β-min restrictions or irrepresentability. The convergence rates here are most similar to the analysis of a Forward-Backward Selection in [73], but require markedly different
analysis since there is no chance to correct “over-selection mistakes.”
The analysis of Simple Forward Selection lays the groundwork for deriving statistical performance bounds for TBFMS. This paper derives performance bounds for TBFMS which
depend directly on the quality of the profile of tests considered, as measured by 5 constants
which characterize size and power. The abstract results for TBFMS are used to derive
asymptotic bounds for various sequences of data-generating processes. As an example, concrete tests for heteroskedastic data constructed from Huber-Eicker-White standard errors
are used to construct t-tests and explicit rates of convergence are calculated.
This paper complements an emerging branch of literature on sequential testing (see [31],
[45], [63], [28]). This literature considers hypothesis testing in stages, in which tests in later
stages can depend on testing outcomes in earlier stages. In various settings, properties like
family-wise error rates of proposed testing procedures can be controlled sequences of hypothesis tests. In all cases, the authors note that the testing procedures are complementary
to forward model selection problems as they guide which variables should be selected and
offer principled stopping rules. The interest in the current paper lies primarily in the statistical properties and performance bounds of estimates and fits based on a selected model
from a greedy algorithm that leverages testing. Future work may potentially combine the
two types of problems.
There are many other sensible approaches to high-dimensional estimation and regularization. An important and common approach to generic high-dimensional estimation problems
are the Lasso and Post-Lasso estimations. The Lasso minimizes a least squares criteria augmented with a penalty proportional to the `1 norm of the coefficient vector. This approach
favors a model with good in-sample prediction while still placing high value on parsimony
(the structure of the objective sets many coefficients are set identically to zero). The PostLasso refits based on a least squares objective function on the selected model. For theoretical
and simulation results about the performance of these two methods, see [29] [62], [35] [21]
[4], [5], [15], [19], [18] [20], [21], [36], [39], [40], [46], [47], [50], [56], [62], [65], [67], [70], [7],
[16], [7], among many more. The asymptotic estimation rates calculated for TBFMS here
match those standard for Lasso and Post-Lasso.
Damian Kozbur/Testing-Based Forward Model Selection
5
After developing several theoretical bounds, a simulation study illustrates the relative
performance of TBFMS to Lasso and Post-Lasso regression. The simulation study shows
that there are data-generating processes under which forward selection outperforms Lasso
regression in terms of prediction and estimation error.
In economic applications, models learned using formal model selection are often used
in subsequent estimation steps with the final goal of learning a structural parameter of
interest. One example is the selection of instrumental variables for later use in a first-stage
regression (see [6], [34]). Another example is the selection of a conditioning set, to properly
control for omitted variables bias when there are many control variables (see [14], [66], [11],
[41]). In both cases, bounds about the quality of the selected model are used to derive results
about the quality of post-model selection estimation and guide subsequent inference. Such
applications require a model selection procedure with a hybrid objective: (1) produce a
good fit, and (2) return a sparse set of variables. This paper addresses both objectives, and
therefore provides adequately tight bounds using strictly forward selection for application
in causal post-estimation analysis.
Finally, TBFMS is illustrated in an economic application. The application revisits the
question studied by Acemoglu, Johnson and Robinson (see [1]) of learning the effect of
institution quality on aggregate economic output in a cross section of 64 countries. [1]
propose an instrumental variables strategy, using early European settler mortality rates as
an instrument for current quality of institutions as measured the extent of protection from
expropriation. They provide an argument concluding that the effect of institutions on output
can be identified using early settler mortality as an instrument, provided that geography
is properly controlled for. In their baseline specification, [1] address this by including a
variable equal to latitude. However, geography is a broad notion and can potentially mean
many different things; for example, temperature, yearly rainfall, terrain. As a compliment
to their analysis, the set of possible controls for geography is expanded to 16. Since 16 is
not vanishingly small relative to the sample size of 64, this example lies within the space
of high-dimensional estimation problems. The most relevant geographic controls are then
chosen using TBFMS. To be robust to model selection mistakes and not suffer classical
problems known to be associated with pretesting, three model selection steps are required
(see [12], [14]), each separately applying TBFMS. These are: (1) select those geographic
variables predictive of output; (2) select those geographic controls predictive of quality of
institution; (3) select those geographic controls predictive of European settler mortality.
Final estimates of the effects of institution quality on growth are generated with standard
IV estimation applied using the union of selected controls in the three selection steps. The
Damian Kozbur/Testing-Based Forward Model Selection
6
findings about the effects of institutions on output are largely consistent with theirs when
model selection is used to determine the way to control for geography. Interestingly, this
provides further evidence supporting the robustness of the conclusions made in [1].
2. Precursor: Sharp Convergence Rates for Simple Forward Selection without
β-min or Irrepresentability Conditions
This section proves a precursory result about Simple Forward Selection which is new in the
high-dimensional econometrics and statistics literature despite the literature’s maturity. The
procedure is defined formally below and is similar to TBFMS, but uses a single threshold
rather than a profile of hypothesis tests in determining the selection of covariates. The
framework set out in this section is also helpful in terms of outlining minimal structure
needed to facilitate the method of analysis in the formal arguments that follow.4
2.1. Framework
A realization of data of sample size n is given by Dn = {(xi , yi )}ni=1 and is generated by
a joint distribution P. The data consists of a set of covariates xi ∈ Rp , as well as outcome
variables yi ∈ R for each observation i = 1, ..., n. The data satisfy
yi = x0i θ0 + εi
for some unknown parameter of interest θ0 ∈ Rp and unobserved disturbance terms εi ∈ R.
The parameter θ0 is sparse in the sense that the set of non-zero components of θ0 , denoted
S0 = supp(θ0 ), has cardinality s0 < n.
Define a loss function `(θ)
`(θ) = En [(yi − x0i θ)2 ]
where En [ · ] =
1
n
Pn
i=1 (·)
denotes empirical expectation. Note that `(θ) depends on Dn ,
but this dependence is suppressed from the notation. Define also
`(S) =
min
`(θ).
θ:supp(θ)⊆S
The estimation strategy proceeds by first searching for a sparse subset Sb ⊆ {1, ..., p}, with
cardinality sb, that assumes a small value of `(S), followed by estimating θ0 with least squares
via
4
This section draws material from the draft [42], also written by the current author, which was originally
a separate project and posted on ArXiv, but is now merged into the current paper in preparation for the
publication process.
Damian Kozbur/Testing-Based Forward Model Selection
θb ∈ arg
min
7
`(θ).
b
θ:supp(θ)⊆S
This gives the construction of the estimates x0i θb for i = 1, ..., n.
The set Sb is selected as follows. For any S define the incremental loss from the jth
covariate by
∆j `(S) = `(S ∪ {j}) − `(S).
Consider the greedy algorithm, which inductively selects the jth covariate to enter a working
model if −∆j `(S) exceeds a threshold t:
−∆j `(S) > t
and ∆j `(S) > ∆k `(S) for each k 6= j. The threshold t is chosen by the user; it is the only
tuning parameter required. This procedure is summarized formally here:
Algorithm 1. Simple Forward Regression
Initialize. Set Sb = ∅
For 1 6 k 6 p
If −∆j `(S) > t for some j ∈ {1, ..., p} \ Sb
Set b
j ∈ arg max {−∆j `(S) : −∆j `(S) > t}
Update Sb = Sb ∪ {b
j}
Else
Break
Set θb ∈ arg minθ:supp(θ)⊂Sb `(θ).
2.2. Formal Analysis
In order to analyze Algorithm 1 and state the first theorem, a few more definitions are
convenient. Define the empirical Gram matrix G by G = En [xi x0i ]. Let ϕmin (s)(G) denote
the minimal s-sparse eigenvalues given by
ϕmin (s)(G) =
min
S⊆{1,...,p}:|S|6s
λmin (GS )
where GS is the principal submatrix of G corresponding to the component set S. The
maximal sparse eigenvalues ϕmax (s)(G) are defined analogously. Let
h
i
cF (b
s) = (b
s + s0 )1/2 ϕmin (b
s + s0 )(G)−1 2kEn [εi x0i ]k∞ + t1/2 .
Damian Kozbur/Testing-Based Forward Model Selection
8
Finally, for each positive integer m, let
c0F (m) = 1 + 72 × 1.7832 × ϕmin (m + s0 )(G)−5 .
Theorem 1. Consider a data set Dn of fixed sample size n with parameter θ0 . Suppose the
normalizations En [x2ij ] = 1 hold for each j 6 p. Then under Algorithm 1 with threshold t,
b 2 ]1/2 6 cF (b
En [(x0i θ0 − x0i θ)
s).
In addition, for every integer m > 0 such that t1/2 > 2ϕmin (m + s0 )(G)−1 kEn [xi εi ]k∞ and
m 6 |Sb \ S0 |, it holds that
m 6 c0F (m)s0 .
The above theorem calculates explicit finite sample constants bounding the prediction
error norm. The second statement is a tool for bounding the number of selected covariates.
In particular, setting m∗ = min{m : m > c0F (m)s0 } implies that
sb 6 m∗ + s0
provided t1/2 > 2ϕmin (m∗ + s0 )(G)−1 kEn [xi εi ]k∞ .
The theorem is proven in detail in the Appendix. The first statement of Theorem 1 is
proven through a combination of two techniques. The first technique creates an analogue
of the Basic Inequality from standard Lasso analysis5 that states that `(Sb ∪ S0 ) 6 `(θ0 ).
b − `(Sb ∪ S0 ) in terms of
The second technique uses a Lemma from [27] which bounds `(S)
b − `(θ0 ) that
s0 , t, and the sparse eigenvalues of G. Together, these produce a bound for `(S)
translates to the statement in Theorem 1. The proof of the second statement of the theorem
involves bounding sb. In the analysis of Lasso-based estimation, bounds for sb are typically
derived with the aid of KKT conditions for the solution to the Lasso problem. Because
Simple Forward Regression is iterative, no such optimality conditions exist. Instead, the
proof leverages a pigeonhole principle which is motivated by the following reasoning. sb
being sufficiently large relative to s0 forces the existence of some subset S ⊆ Sb such that
λmin (GS )−1 is correspondingly large.6 This is possible because all covariates indexed in Sb
must exhibit some correlation to covariates indexed in S0 . At the same time λmin (GS )−1
cannot exceed ϕmin (b
s + s0 )(G)−1 . The desired bounds controlling sb are then deduced.
5
b
b 1 6 `(θ0 )+λkθ0 k1 .
θk
For Lasso estimation with penalty level λ, the Basic Inequality asserts that `(θ)+λk
Formally, this pigeonhole principle is carried out by constructing an operator L∞ (Rs0 ) → L1 (Rs0 ) which
depends on projections of covariates indexed in Sb onto covariates indexed in S0 , and then lower bounding its
6
norm with the aid of Grothendieck’s inequality ([32], see for a review, [53]; see also the exact form described
in [33]; see the appendix for more details.)
Damian Kozbur/Testing-Based Forward Model Selection
9
The statement in Theorem 1 gives finite sample bounds which are completely deterministic in the sense that they hold for every possible realization of the data. Furthermore,
the proof does not use the random nature of Dn at any step. As a result, the bounds are
very general, but it is helpful for interpretation to consider the convergence rates implied
by Theorem 1 under asymptotic conditions on Dn . Consider a sequence of random data
sets (Dn )n∈N generated by joint distributions {P = Pn }n∈N . For each n, the data again
satisfy yi = x0i θ0 + εi . In what follows, the parameters θ0 , the thresholds t, distribution P,
the dimension p of xi , etc. can all change with n.
Condition 1 (Asymptotic Regularity). The sparsity satisfies s0 = o(n). There is a sequence
Kn for which s0 = o(Kn ) and there is a bound ϕmin (Kn )(G)−1 = O(1) which holds with
probability 1 − o(1). The normalizations En [x2ij ] = 1 hold a.s. for every j 6 p. The threshold
satisfies a bound t = O(log p/n). In addition, t1/2 > 2ϕmin (Kn )(G)−1 kEn [xi εi ]k∞ with
probability 1 − o(1).
The rates assumed in Condition 3 reflect typical rates achieved under various possible
sets of low level conditions standard in the literature (ie. [6]). Condition 1 asserts three important statements. The first statement bounds the size of S0 and requires that the sparsity
level is small relative to the sample size. The second statement is a sparse eigenvalue condition useful for proving results about high-dimensional techniques like Lasso. In standard
regression analysis where the number of covariates is small relative to the sample size, a
conventional assumption used in establishing desirable properties of conventional estimators of θ is that G has full rank. In the high-dimensional setting, G will be singular if p > n
and may have an ill-behaved inverse even when p 6 n. However, good performance of many
high-dimensional estimators only requires good behavior of certain moduli of continuity of
G. There are multiple formalizations and moduli of continuity that can be considered here;
see [15]. This analysis focuses on a simple eigenvalue condition which was used in [6]. Condition 2 could be shown to hold under more primitive conditions by adapting arguments
found in [7] which build upon results in [70] and [58]; see also [57]. Condition 2 is notably
weaker than previously used irrepresentability conditions. Irrepresentability conditions require that for certain sets S and k ∈
/ S, letting xiS be the subvector of xi with components
j ∈ S, that kEn [xiS x0iS ]−1 En [xiS x0ik ]k1 is strictly less than 1.
The final statement in Condition 1 is a regularization condition similar to regularization conditions common in the analysis of Lasso. The condition, requires t1/2 to dominate
a multiple of the kEn [xi εi ]k∞ . This condition is stronger than that typically encountered
Damian Kozbur/Testing-Based Forward Model Selection
10
with Lasso, because the multiple depends on on the sparse eigenvalues of G. To illustrate
why such a condition is useful, let x̌ij denote xij residualized away from previously selected
regressors and renormalized. Then even if En [xij εi ] < t1/2 , En [x̌ij εi ] can exceed t1/2 resulting in more selections into the model. Nevertheless, using the multiple 2ϕmin (Kn )(G)−1
which stays bounded with n, is sufficient to ensure that sb does not grow faster than s0 .
Furthermore, this requirement does not implicitly impose a β-min condition and does not
implicitly impose irrepresentability.
From a practical standpoint, Condition 1 does however implicitly require the user to
know more about the design of the data in choosing an appropriate t. Choosing feasible
thresholds which satisfy a similar condition to Condition 3 is considered in the next section
where TBFMS is developed.
Theorem 2. Consider a sequence of data sets Dn indexed by n with parameters θ0 and
threshold t which satisfy Condition 1. Then for θb is obtained by Algorithm 1 there are bounds
r
b 2 ]1/2 = O
En [(x0i θ0 − x0i θ)
s0 log p
n
!
,
sb 6 O(s0 ).
which hold with probability 1 − o(1) as n → ∞.
More explicitly, the implied O constants and o sequence in bounds for Theorem 2 are
understood to depend only on the implied O constants and o sequences in Condition 1.
The theorem shows that Simple Forward Selection can obtain asymptotically the same
convergence rates in prediction error norm as other high-dimensional estimators like Lasso,
provided an appropriate threshold t is used. In addition, it selects a set with cardinality
commensurate with s0 .
Finally, two direct consequence of Theorem 2 are bounds on the deviations kθb− θ0 k1 and
kθb − θ0 k2 of θb from underlying unknown parameter θ0 . Theorem 3 above shows that deviations of θb from θ0 also achieve rates typically encountered in high-dimensional estimators
like Lasso.
Theorem 3. Consider a sequence of data sets Dn with parameters θ0 and thresholds t
which satisfy Condition 1. Suppose θb is obtained by Algorithm 1. Then there are bounds
!
!
r
r
2 log p
s
log
p
s
0
0
b 1=O
b 2=O
kθ0 − θk
and kθ0 − θk
n
n
which hold with probability 1 − o(1) as n → ∞.
Damian Kozbur/Testing-Based Forward Model Selection
11
3. Testing-Based Forward Model Selection
The previous section presented results on convergence rates Simple Forward Selection in
a very simple context. The results of Theorems 1 are useful in terms of developing intuition and proof techniques for inductive variable selection algorithms. However, in terms
of practical implementation, Section 2 leaves the question of how to choose a threshold
unanswered. This section develops TBFMS in order to analyze feasible, data-driven ways
to decide which covariates to select, and when to stop selecting.
3.1. Framework
The basic framework for this section is similar. Again, the observed data is given by Dn =
(xi , yi )ni=1 is generated by P and satisfies yi = x0i θ0 + εi for a parameter θ0 which is sparse
with s0 non-zero components supported on S0 . Define `(θ) and `(S) as before.
Define the expected loss function E : Rp → R by
E(θ) = E En (yi − x0i θ)2
and note that E(θ) = E`(θ). Extend the definition of E to apply also as a map E : 2{1,...,p} →
R by E(S) = minθ:supp(θ)⊆S E(θ). Similarly to before, for any S define the incremental loss
from the jth covariate by
∆j E(S) = E(S ∪ {j}) − E(S).
Within the class of greedy algorithms, it would be preferable to consider a greedy algorithm which inductively selects the jth covariate to enter a working model if ∆j E(S) is
large and ∆j E(S) > ∆k E(S) for each k 6= j. However, because ∆j E(S) cannot generally be
directly observed from the data, the idea that follows is to make use of statistical tests to
gauge the magnitude of ∆j E(S). Consider a set of tests given by
TjSα ∈ {0, 1} associated to H0 : ∆j E(S) = 0 and level α > 0.
Assume that the tests reject (TjSα = 1) for large values of a test statistic WjS .
The model selection procedure is as follows. Start with an empty model (consisting
b select one covariate such that
of no covariates). At each step, if the current model is S,
b
Tj Sα
b = 1, append it to S, and continue to the next step; if no covariates have Tj Sα
b = 1, then
terminate the model selection procedure and return the current model. If at any juncture,
there are two indices j, k (or more) such that TjSα = TkSα = 1, the selection is made
according to the larger value of WjS , WkS . Alternatively, additional tests TjkSα associated
Damian Kozbur/Testing-Based Forward Model Selection
12
to H0 : ∆j E(S) > ∆k E(S) could be devised to break ties. The test statistic approach is
natural for breaking potential multi-way ties. To summarize, the algorithm for forward
selection given the hypothesis tests (TjSα , WjS ) is now given formally.
Algorithm 2. Testing-Based Forward Model Selection
Initialize. Set Sb = ∅.
For 1 6 k 6 p:
b
If: Tj Sα
/ S,
b = 1 for some j ∈
Set: b
j ∈ arg max{Wj Sb : Tj Sα
b = 1},
b
b
Update: S = S ∪ {b
j}.
Else: Break.
Set: θb ∈ arg min
b
θ:supp(θ)⊂S
En (yi − x0i θ)2 .
3.2. Formal Analysis
This section formally states conditions on the hypothesis tests conditions on the data before
analyzing properties of Algorithm 1. These conditions are measures of the quality of the
given testing procedure and the regularity of the data.
Condition 2 (Hypothesis Tests). There is an integer Ktest > s0 and constants α, δtest , ctest ,
c0test , c00test > 0 such that each of the following conditions hold.
1. The tests have power in the sense that
1
P ({TjSα = 1 for every j, |S| 6 Ktest such that − ∆j E(S) > ctest }) > 1 − δtest .
3
2. The tests control size in the sense that
1
P {TjSα = 1 for some j, |S| 6 Ktest such that − ∆j E(S) 6 c0test } 6 α + δtest .
3
3. The tests are continuous in the sense that
P({WjS > WkS for each j, k, |S| 6 Ktest such that
1
TjSα = TkSα = 1 and − ∆j E(S) > −c00test ∆k E(S)}) > 1 − δtest .
3
The constants ctest and c0test measure quantities related to the size and power of the tests
and provide a convenient language for subsequent discussion. The constant c00test measures
Damian Kozbur/Testing-Based Forward Model Selection
13
the extent to which the test statistics WjS reflect the actual magnitude of ∆j E(S). Note
again that the hypothesis tests considered should not necessarily be thought of as providing
a measure of statistical significance, but more precisely, they are simply a tool for model
selection which coincidentally have many properties in common with traditional hypothesis
tests.
Condition 3 (Regularity). Normalizations E[En [x2ij ]] = 1 holds for all j. The residuals decompose into εi = εoi + εai where E[En [εoi 2 ]] < ∞, E[En [εoi xij ]] = 0 for all j, and E[En [εai 2 ]] 6
1
−1 0
2 ϕmin (Ktest )(E[G]) ctest .
Finally, (2 + 1.783 × 72ϕmin (Ktest )(E[G])−5 c00test −4 )s0 < Ktest .
Condition 3 imposes regularity conditions for the class of models considered in the following theorem. First, εi is decomposed into an orthogonal component εoi and an approximation
component εai , each of which exhibits a different kind of regularity. The orthogonal component is orthogonal to the covariates in the population. The approximation component need
not be orthogonal to the covariates, but its magnitude must be suitably controlled by the
sparse eigenvalues of E[G] and by the parameter c0test , which is a detection threshold for the
profile of hypothesis tests TjSα . This decomposition allows for approximately sparse models
similar to the framework of [6]. The fact that εai need not be orthogonal to the covariates
also allows this framework to overlay onto many problems in traditional nonparametric
econometrics.
Condition 3 also imposes conditions relating the sparse eigenvalues of E[G] with c00test , s0 ,
and Ktest . Note that Ktest measures the size of the set S ⊂ {1, ..., p} over which the hypothesis test perform well, as defined by Condition 2. Consequently, this condition requires
that the hypothesis tests TjSα perform sufficiently well over sets S, which must be larger
when E[G] has small eigenvalues or when s0 is large.
There are a few cases where Condition 3 can be simplified. Note that if p > n, even though
the empirical Gram matrix is necessarily rank deficient, the population Gram matrix may
be full rank. When E[G] is full rank, then the last statement of Condition 1 can be simplified
to (2 + 1.783 × 72λmin (E[G])−5 c00test −4 )s0 < Ktest . In addition, the condition on εai implicitly
imposes constraints on c0test and ϕmin (Ktest )(E[G])−1 . When there is no approximation error,
this requirement is no longer needed.
Damian Kozbur/Testing-Based Forward Model Selection
14
Let
cT = s0 ϕmin (Ktest )(E[G])−1 ctest
c0T = 2 + 1.783 × 72ϕmin (Ktest )(E[G])−5 c00test
−4
c00T (b
s) = ϕmax (s0 + sb)(G)1/2 ϕmin (s0 + sb)(G)−1/2 sb1/2 kEn [xi εi ]k∞
1/2
+ 3ϕmax (s0 + sb )(G)ϕmin (b
s + s0 )(G)−1/2 )s0 ctest 1/2 ϕmin (Ktest )(E[G])−1 .
Theorem 4. Consider Dn ∼ P for some fixed n and {TjSα , WjS } such that Conditions 2
and 3 hold. Then the bounds
b − E(S0 ) 6 cT
E(S)
sb 6 c0T s0
En [( x0i θ0 − x0i θb )2 ] 6 c00T (b
s)
hold with probability at least 1 − α − δtest .
Theorem 4 is proven in the Appendix. It provides finite sample bounds on the performance of TBFMS. The outline for proving Theorem 4 is similar to that for proving Theorem
1. Theorem 4 works with the fact that −∆j E(S) > ctest upon selection on the event implied
by Condition 2 instead of the simpler −∆j `(S) > t. In contrast to the proof of Theorem 1,
the proof of Theorem 4 also addresses the possibility that if covariate j is selected ahead
of covariate k, it is not necessarily the case that −∆j E(S) > −∆k E(S). This is done by
making use of the continuity constant c00test in Condition 2.
Theorem 4 can be used to derive asymptotic estimation rates by allowing the constants
to change with n. The next subsection provides an example to a linear model with heteroskedastic disturbances, where, under the stated regularity conditions, the prediction
error and estimation error attain the rate OP (s0 log p/n). These convergence rates therefore
match typical Lasso and Post-Lasso rates.
Note that the results aim to control the hybrid objective, described in the introduction, of
producing a good fit and returning a sparse set of variables. One useful implication of bounds
b 2 ] is that the results can be applied to constructing
controlling both sb and En [(x0 θ0 − xi θ)
i
uniformly valid post-model selection inference procedures (see [14]), in which for some
applications, the prediction error bound alone is insufficient. Applications to inference are
discussed in Section 4.
Damian Kozbur/Testing-Based Forward Model Selection
15
3.3. Example: Heteroskedastic Disturbances
This section gives an example of the use of Theorem 4 by illustrating an application of model
selection in the presence of heteroskedasticity in the disturbance terms εi . The conditions
required for the application of Theorem 4 are verified for a set of tests that are constructed
based on the Heteroskedasticity-Consistent standard described in [69].
For shorthand, write xijS to be the vector with components xik with k = j or k ∈ S. To
construct the tests, begin with the least squares estimate of the regression yi on xijS .
θbjS = En [xijS x0ijS ]−1 En [x0ijS yi ]
Define εbijS = yi − x0ijS θbjS . One heteroskedasticity robust estimate of the sampling variance
of θbjS , proposed in [69], is given by the expression
1
b
VbjS = En [xijS x0ijS ]−1 ΨεjS
En [xijS x0ijS ]−1
n
where
b
2
ΨεjS
= En [b
εijS
xijS x0ijS ].
Define the test statistics
−1/2
WjS = [VbjS ]jj
[θbjS ]j .
Reject H0 for large values of WjS defined relative to an appropriately chosen threshold. To
0 )0 where β
define the threshold, first let ηjS := (1 , −βjS
jS is the coefficient vector from the
least squares regression of {xij }ni=1 on {xik }ni=1,k∈S . Then define
τbjS
0 Diag(Ψεb )1/2 k
kηjS
1
jS
q
.
=
0 Ψεb η
ηjS
jS
jS
The term τbjS will be helpful in addressing the fact that many different model selection
paths are possible under different realizations of the data under P.7 Not taking this fact
7
There is an unfortunate misprint in a Papers and Proceedings version of this paper, [43], in which the
b
exponent 1/2 is missing from the term Diag(ΨεjS
). Note that [43] do not derive nor claim any theoretical
properties, and instead just state the term τbjS .
Damian Kozbur/Testing-Based Forward Model Selection
16
into account can potentially lead to false discoveries. The next condition states precisely
the hypothesis tests TjSα .
Definition 1 (Hypothesis Tests for Heteroskedastic Disturbances). Let cτ > 1 and α > 0
be parameters. Assign
TjSα = 1 ⇐⇒ Wjs > cτ τbjS Φ−1 (1 − α/p).
The term Φ−1 (1 − α/p) can be informally thought of as a Bonferroni correction term
that takes into account the fact that there are p potential covariates. The term cτ τbjS can
be informally thought of as a correction term that can account for the fact that the set S is
random and can have many potential realizations. The simulation study uses the settings
cτ = 1.01 and α = .05.
Condition 5 (Regularity for Data with Heteroskedasticity). Consider a sequence of data
sets Dn = {(xi , yi )}ni=1 ∼ P = Pn . The observations (xi , yi ) are i.n.i.d. across i and
yi = x0i θ0 + εi for some θ0 with s0 = o(n). The residuals decompose into εi = εoi + εai
such that a.s., E[εoi |xi ] = 0 and maxi |εai | = O(n−1/2 ). In addition, a.s., uniformly in
i and n, E[ε4i |xi ] are bounded above and E[ε2i |xi ] is bounded away from zero. The covariates satisfy maxj6p En [x12
ij ] = O(1) with probability 1 − o(1). There is a sequence
Kn , where s0 = o(Kn ), and bounds ϕmin (Kn )(G)−1 = O(1), ϕmax (Kn )(G)−1 = O(1),
ϕmin (Kn )(En [(εi xi )(εi xi )0 ])−1 = O(1), and max|S|6Kn ,j ∈S
/ kηjS k1 = O(1), which hold with
probability 1 − o(1). The rate condition Kn4 log3 p/n = o(1) holds.
Condition 5, as before, gives conditions on the sparse eigenvalues, this time applying to
both G and to En [(εi xi )(εi x0i )]. The requirement that ϕmin (Kn )(G)−1 stay bounded can be
relaxed, thought the cost is that the proof of Theorem 5 could not directly call upon Theorem 4 (since c00T involves a maximal sparse eigenvalue). In addition, Condition 5 assumes
a bound on ηjS that may be strong in some cases. Previous results in [64], [72] assume
the strict condition that maxj ∈S
/ 0 kηjS0 k1 < 1, which is the genuine irrepresentability condition, in analysis of inductive variable selection algorithms. Here, the requirement < 1 is
replaced by the weaker requirement = O(1). Other authors, for instance [49], use conditions analogous to max|S|6Kn ,j ∈S
/ kηjS k1 = O(1) in the context of learning high-dimensional
Damian Kozbur/Testing-Based Forward Model Selection
17
graphs, and note that the relaxed requirement is satisfied by a much broader class of datagenerating processes. Analogous bounds on kηjS k1 were not required in Theorem 1, since
c
the proof of
q Theorem 1 does not leverage bounds relating WjS to the self-normalizeds sums
En [x0ij εi ]/ En [x2ij ε2i ], j 6 p. Failure of the O(1) bound would lead only to slightly slower
convergence rates. Condition 5 also states regularity conditions on moments of εi and xi ,
which are useful for proving laws of large numbers, central limit theorems, and moderate
deviation bounds (see [38]). Finally, the rate condition assumes bounds on the relative sizes
of s0 , p, n since s0 < Kn .
Theorem 5. Consider a sequence of data sets Dn ∼ P = Pn which satisfies Condition
5. Suppose that cτ > 1 is fixed independent of n that α = o(1) with nα → ∞. Let θb be
the estimate obtained from Algorithm 2 with tests defined by Definition 1. Then there are
bounds
r
En [(x0i θ0
−
b 2 ]1/2
x0i θ)
=O
s0 log(p/α)
n
!
and
sb = O(s0 )
which hold with probability at least 1 − α − o(1) as n → ∞.
The theorem is proven in the appendix by appealing to Theorem 4.
Analogous results potentially hold for dependent data using HAC-type standard errors
(see [52], [3].) The required central limit results for such an application are beyond the
scope of this work, though using the moderate deviation results for dependent data of [24]
is a potential starting point. In addition, cluster-type standard errors for large-T -large-n
and fixed-T -large-n panels can be used by adapting arguments from [11].
Allowing α to be fixed is possible under more restrictive conditions on the approximation
error terms εai . Note that if p > n, then the rate log(p/α) becomes equivalent to simply
log(p).
As a final remark, note that there is a much simpler formulation for the hypothesis tests
that ignores the cτ τbjS terms. This results in the following definition.
Definition 2 (Simplified Hypothesis Tests for Heteroskedastic Disturbances). Let cτ > 1
and α > 0 be parameters. Assign
TjSα = 1 ⇐⇒ Wjs > Φ−1 (1 − α/p).
Damian Kozbur/Testing-Based Forward Model Selection
18
These tests are based on a simple Bonferroni-type correction. Furthermore, TBFMS using
the simpler tests is natively programmed in some statistical software, including SPSS. It is
unknown to the author at the time of this writing whether the same convergence rates can
be attained using the simpler tests. This option is explored in some finite sample settings
in the simulation study that follows. Evidence from the simulation study suggests that this
option performs better than the more complex tests defined in Definition 1.
Another way to potentially lower the significance thresholds is to consider generalized
error rates. The conditions set forth in Condition 2 require control of a notion resembling
family-wise error rate uniformly over hypothesis tests H0 : ∆j E(S) for j 6 p and |S| < Ktest
for some integer Ktest . Other types of error rates like k-family wise error rate, false discovery rate, or false discovery proportion are potentially possible as well. In particular, the
arguments in the proof of Theorem 4 would continue to be compatible with procedures that
controlled an appropriate notion of false discovery proportion. In order to keep exposition
concise, these extensions are not considered here.
3.4. TBFMS Simulation
The results in the previous sections suggest that estimation with TBFMS should produce
good results in large sample sizes. This section simulates several different data-generating
processes to evaluate the finite sample performance of TBFMS relative to select other
procedures commonly used in high-dimensional regression problems.
This simulation study draws samples from the data-generating processes in Table 1.
Table 1
TBFMS Simulation Design
Data:
DGP:
Settings:
Dn = (yi , xi , zi )n
i=1 iid
yi = xi θ0 + εi
p = dim(xi ) = 2n, θ0j = bj−1
1j6s0
0
xij ∼ N (0, 1), with corr(xij , xik ) = 0.5|j−k|
P
εi ∼ σi N (0, 1), σi = exp(ρ0 pj=1 0.75(p−j) xij )
s0 = 6
b0 ∈ {−0.5, 0.5}
ρ0 ∈ {0, 0.5}
n ∈ {100, 200, 300, 400, 500}
Damian Kozbur/Testing-Based Forward Model Selection
19
Simulations are conducted with parameter settings s0 = 6, b0 ∈ {−0.5, 0.5} , ρ0 ∈
{0, 0.5}, and n ∈ {100, 200, 300, 400, 500} according to the data-generating process displayed
in Table 1. The parameter s0 dictates the sparsity. The parameter b0 controls the nature
of the coefficient vector. When b0 = −0.5, the coefficients θ0j alternate sign in j, and
when b0 = 0.5, the coefficients are all positive. The two different settings for b0 create
different interplay between the Toeplitz correlation structure in the covariates and their
corresponding coefficients. The parameter ρ0 controls the presence of heteroskedasticity in
the disturbance terms εi . The terms εi are homoskedastic when ρ0 = 0 and heteroskedastic
otherwise. Finally, each setting of s0 , b0 , and ρ0 is simulated in sample sizes ranging between
100 and 500. The dimensionality is always taken to be double the sample size so that p = 2n.
This allows a visualization of the consistency properties of the displayed estimators in a
high-dimensional asymptotic frame. Alternative parameterizations, for example pinning
down the signal-to-noise ratio, are also interesting and possible, but are not displayed here
in favor of brevity. Each simulation design is replicated 1000 times.
The simulation study compares five different estimators. In addition to two TBFMS
estimators, two Lasso-based estimators and an infeasible estimator that knows the true
support S0 are presented.
1. TBFMS I. Implements Algorithm 2 with tests defined in Definition 1 with parameters cτ = 1.01, α = .05.
2. TBFMS II. Implements Algorithm 2 with the simplified tests defined in Definition
2 with parameters cτ = 1.01, α = .05.
3. Lasso-CV. Uses the original formulation of Lasso, implemented using glmnet, with
penalty parameter chosen using 10-fold cross validation. θb is the minimizer of the
Lasso objective function and is not refit on the selected model.
4. Post-Het-Lasso. Uses the implementation found in [6], which is designed specifically
to handle heteroskedastic disturbances. [6] requires two tuning parameters that are
directly analogous to cτ and α. These are set to cτ = 1.01 and α = .05. θb is refit on
the selected model.
5. Oracle selects the model consisting of {j : [θ0 ]j 6= 0} and estimates a subsequent
least squares regression.
The simulation results are presented in Figures 1−4. Each figure contains four plots that
track various measures of estimation quality for the five estimators for fixed s0 , b0 , and ρ0 ,
Damian Kozbur/Testing-Based Forward Model Selection
20
and for n varying along the horizontal axis. The upper left plots display the number of correctly identified covariates from S0 , averaged over simulation replications. The upper right
plots display the total number of selected covariates, averaged over simulation replications.
b 2 ]1/2 , averaged
The bottom left plots display the prediction error defined as En [(x0 θ0 − x0 θ)
i
i
over simulation replications. The bottom right plots display estimation error defined as
kθb2 − θ0 k2 , averaged over simulation replications.
In the simulations results, there is no single feasible estimator that dominates in every
setting in terms of estimation error or prediction error. However, in all settings, Lasso-CV
selects the most covariates (both in absolute terms and in terms of the number of correctly
identified covariates), followed by TBFMS I, TBFMS II, and Post-Het-Lasso. There are instead important instances when each estimator performs better. In both settings with positive coefficients (b0 = 0.5), Lasso-CV achieves the smallest estimation error with TBFMS I
and TBFMS II having slightly higher estimation error. In these settings, however, the prediction error is smallest with TBFMS II. With alternative coefficients (b0 = −0.5), however,
TBFMS I and TBFMS II dominate Lasso-CV and Post-Het-Lasso on prediction error and
estimation error. This suggests that the performance of these estimators depends on the
configuration of the signal relative to the correlation structure of the covariates. Finally, the
relative difference in performance across estimators is larger in the presence of heteroskedasticity. In the presence of heteroskedasticity, the Post-Het-Lasso exhibits faster improvement
in estimation error and prediction error with increasing n, though it is still dominated by
the other estimators. Note that each of the techniques, TBFMS I and Post-Het-Lasso, are
theoretically valid for sequences of data-generating processes with heteroskedasticity. In
addition, the properties of cross-validation with Lasso are only beginning to be understood
(see [26] for analysis of lasso with cross-validation). But it is clear from this simulation
study that Lasso-CV leads to selection of substantially more covariates to the extent that
the effects of heteroskedasticity on the performance of the estimator are still not fully clear.
Damian Kozbur/Testing-Based Forward Model Selection
21
Fig 1. Simulation Results
This figure presents simulation results for the estimators TBFMS I, TBFMS II, Post-Het-Lasso, Lasso-CV,
and Oracle. The simulation design is described fully in Table 1. This figure presents estimates for the
settings
s0 = 6: High Sparsity
b0 = 0.5: Positive Coefficients
ρ0 = 0: Homoskedastic disturbances.
Plots are based on 1000 simulation replications for every n = 100, 200, 300, 400, 500 indexed on the
horizontal axis. The upper-left plot displays the number of correctly identified covariates from S0 , averaged
over simulation replications. The upper-right plot displays the total number of selected covariates,
averaged over simulation replications. The bottom-left plot displays the prediction error defined as
b 2 ]1/2 , averaged over simulation replications. The bottom-right plot displays estimation error
En [(x0i θ0 − x0i θ)
b
defined as kθ2 − θ0 k2 , averaged over simulation replications.
Damian Kozbur/Testing-Based Forward Model Selection
22
Fig 2. Simulation Results
This figure presents simulation results for the estimators TBFMS I, TBFMS II, Post-Het-Lasso, Lasso-CV,
and Oracle. The simulation design is described fully in Table 1. This figure presents estimates for the
settings
s0 = 6: High Sparsity
b0 = 0.5: Positive Coefficients
ρ0 = 0.5: Heteroskedastic disturbances.
Plots are based on 1000 simulation replications for every n = 100, 200, 300, 400, 500 indexed on the
horizontal axis. The upper-left plot displays the number of correctly identified covariates from S0 , averaged
over simulation replications. The upper-right plot displays the total number of selected covariates,
averaged over simulation replications. The bottom-left plot displays the prediction error defined as
b 2 ]1/2 , averaged over simulation replications. The bottom-right plot displays estimation error
En [(x0i θ0 − x0i θ)
b
defined as kθ2 − θ0 k2 , averaged over simulation replications.
Damian Kozbur/Testing-Based Forward Model Selection
23
Fig 3. Simulation Results
This figure presents simulation results for the estimators TBFMS I, TBFMS II, Post-Het-Lasso, Lasso-CV,
and Oracle. The simulation design is described fully in Table 1. This figure presents estimates for the
settings
s0 = 6: High Sparsity
b0 = −0.5: Alternating Coefficients
ρ0 = 0: Homoskedastic disturbances.
Plots are based on 1000 simulation replications for every n = 100, 200, 300, 400, 500 indexed on the
horizontal axis. The upper-left plot displays the number of correctly identified covariates from S0 , averaged
over simulation replications. The upper-right plot displays the total number of selected covariates,
averaged over simulation replications. The bottom-left plot displays the prediction error defined as
b 2 ]1/2 , averaged over simulation replications. The bottom-right plot displays estimation error
En [(x0i θ0 − x0i θ)
b
defined as kθ2 − θ0 k2 , averaged over simulation replications.
Damian Kozbur/Testing-Based Forward Model Selection
24
Fig 4. Simulation Results
This figure presents simulation results for the estimators TBFMS I, TBFMS II, Post-Het-Lasso, Lasso-CV,
and Oracle. The simulation design is described fully in Table 1. This figure presents estimates for the
settings
s0 = 6: High Sparsity
b0 = −0.5: Alternating Coefficients
ρ0 = 0.5: Heteroskedastic disturbances.
Plots are based on 1000 simulation replications for every n = 100, 200, 300, 400, 500 indexed on the
horizontal axis. The upper-left plot displays the number of correctly identified covariates from S0 , averaged
over simulation replications. The upper-right plot displays the total number of selected covariates,
averaged over simulation replications. The bottom-left plot displays the prediction error defined as
b 2 ]1/2 , averaged over simulation replications. The bottom-right plot displays estimation error
En [(x0i θ0 − x0i θ)
defined as kθb2 − θ0 k2 , averaged over simulation replications.
Damian Kozbur/Testing-Based Forward Model Selection
25
4. Applications to structural inference in high-dimensional models
This section describes the application of TBFMS to inference for structural parameters in
three commonly estimated economic problems.
A. The selection of instrumental variables for the estimation of the effect of an endogenous variable on an outcome of interest from a large set of potential instruments.
B. The selection of conditioning covariates for the estimation of the effect of a covariate
of interest on an outcome from a large set of potential observable controls.
C. The selection of a conditioning set in the first stage of an instrumental variables
regression from a large set of potential observable controls.
The general pattern in each of the above three models is that they contain a low dimensional parameter of interest and a high-dimensional nuisance parameter. In this sense they
share many qualities with semiparametric estimation problems. Early works on the inference for a single coefficient in a high-dimensional regression model include [14], [71], [10], [9],
[66], and [37], while early works for high-dimensional intrumental variables models include
[6] and [30]. More detailed review of these two particular strands of literature are provided
below. More general and more recent work in [25] develops theory for inference about a
relatively low-dimensional set of prespecified target parameters when machine learning is
used to estimate some features of the models specified by general moment conditions under
weak conditions.
In all of the cases mentioned above, the quality of subsequent inference depends crucially
on the quality of estimation of the various high-dimensional components that appear in the
course of decomposing the problem into reduced form components. As a result, TBFMS is a
good candidate for input into a larger structural estimation problem whenever it performs
favorably in the relevant high-dimensional estimation steps in that problem.
The theorems that follow prove asymptotic normality for estimates of structural parameters of interest when instruments or conditioning variables are selected using TBFMS
in an appropriate way. The formal results follow directly from the performance bounds
for TBFMS given in Theorem 5 in conjuction with high-level conditions for inference in
structural inference problms (see [14], [6], [25], [12] among others). Therefore, the main
contribution of this section is in showing with simulation that TBFMS performs favorably
relative to Lasso-based methods for some very simple data-generating processes. Subsections 4.1−4.2 briefly review estimation of Models A and B and provide theoretical justification for using TBFMS to estimate components of structural problems. Subsection 4.3
then conducts a simulation study to compare the finite sample performance of the use of
Damian Kozbur/Testing-Based Forward Model Selection
26
TBFMS and alternative estimation strategies in models A and B. Alternative methods
include the two different implementations of Lasso considered in the previous section and
an infeasible benchmark that knows the most important covariates in advance.
Finally, Subsection 4.5 presents an empirical application that fits model C to a classic
data set of Acemoglu, Johnson, and Robinson.
4.1. Model A: Linear Instrumental Variables Model with High-Dimensional
Instruments
Instrumental variables techniques are widely used in applied economic research. While these
methods give an important tool for calculating structural effects, they are often imprecise.
One way to improve the precision of instrumental variables estimators is to use many
instruments or try to approximate the optimal instruments as in [2], [22], and [51].
This section follows the development [6], who consider using Post-Lasso to estimate
optimal instruments. Using post-model selection methods to form first-stage predictions in
IV estimation provides a practical approach to obtaining the efficiency gains from using
optimal instruments while simultaneously lessening the problems associated with many
instruments. An implication of Theorem 5 is that TBFMS provides good approximations
to the optimal instruments when the total number of potential instruments is large.
Consider data given by Dn = {(yi , xi , zi )}ni=1 ∼ Pn where yi ∈ R are outcome variables,
xi ∈ R are endogenous variables of interest, and zi ∈ Rp are instruments. Formally, the
class A of instrumental variables models is defined to contain joint distribution P (for all
n) such that
y i = β 0 x i + εi
xi = zi0 θ0A + ui
for some parameters β0 ∈ R, θ0A ∈ Rp such that E[εi |zi ] = E[ui |zi ] = 0 but possibly
E[εi ui ] 6= 0.8
Consider estimation of the parameter of interest β0 , the coefficient on the endogenous
regressor, using TBMS to select instruments. Assume that the first-stage follows a sparse
model with |support(θ0A )| 6 s0 < n. More precisely, consider the following condition.
8
Extending this class of models to allow approximate sparsity is trivial by decomposing ui = uai + uoi , as
was done in Conditions 4 and 5. Then E[uoi |zi ] = 0 can replace the requirement E[ui |zi ] = 0.
Damian Kozbur/Testing-Based Forward Model Selection
27
The estimate βb of β0 is defined as follows. Obtain θbA from applying Algorithm 2 to
{(xi , zi )}n using the hypothesis tests in Definition 1. Set x
bi = z 0 θbA . Set
i=1
i
b −1 En [b
b = En [xi x
βb = Q
xi yi ] where Q
bi ]
b −1 Ω
bQ
b −1 where Ω
b = En [b
Vb = Q
x2i εb2i ].
b Under regularity conThe quantity Vb is used as an estimate of the variance of β.
ditions, it will be close to the quantity V = Q−1 ΩQ−1 with high probability where
Q = E[(zi π0 )2 ], Ω = E[(zi π0 )2 ε2 ]. The next theorem shows that the estimate βb is asympi
totically Gaussian with variance V and Vb is consistent for V.
Condition 6 (Regularity for Instrument Selection.) Pn ∈ A satisfy Condition 6 for
{(xi , zi )}ni=1 and Condition SM from [6].9
Theorem 6. Consider data sets Dn = {(yi , xi , zi )}ni=1 ∼ Pn . Uniformly over all sequences
Pn ∈ A for which Condition 6 holds, as n → ∞,
n1/2 V −1/2 (βb − β0 ) →d N(0, 1) and
V − Vb →p 0.
This theorem verifies that the IV estimator formed with instruments selected by TBFMS
in a linear IV model is consistent and asymptotically normal. In addition, one can use the
result with Vb defined above. Note that this inference will be valid uniformly over a large
class of data-generating processes that include cases in which perfect instrument selection
is impossible.
The conditions assumed in Condition 6 are fairly standard. The added conditions beyond
the assertion that Condition 6 holds for {(xi , zi )}ni=1 are a simplified version of Condition
SM from [6]. Outside of moment conditions, the main restriction is the assumption that
the parameter β0 would be strongly identified if zi π0 could be observed. Coupled with the
sparse model, this condition implies that using a small number of the variables in zi is
9
For ease of reference, Condition SM from [6] is reproduced here, adapted to present notation. It states the
following. E[En [(zi θ0A )2 ]] is bounded and away from zero, uniformly in n. The conditional variance E[ε2i |zi ]
is bounded a.s. uniformly from above and away from zero, uniformly in i and n. For some q > 2 and qε > 0,
uniformly in n, maxj6p E[En [|zij εi |3 ]] + E[En [|zi0 θ0A |q |εi |2q ]] + E[En [|zi0 θ0A |1 ]] + E[En |εi |qε ]] + E[|xi |q ] = O(1).
2 2
log(p) = o(n1/3 ). s0 log(max(p, n))n2/q−1 = o(1). s20 log2 (max(p, n))/n = o(1). maxj6p En [zij
εi ] = OP (1).
Note that some of these conditions are redundant with Condition 5, but are nonetheless still stated for
completeness.
Damian Kozbur/Testing-Based Forward Model Selection
28
sufficient to strongly identify β0 , which rules out the case of weak-instruments as in [59]
and many-weak-instruments as in [23].10
4.2. Model B: Linear Model with High-Dimensional Control Variables
A different strategy for identifying structural effects in economic research is based on assuming that variables of interest are as good as randomly assigned conditional on observables.
This leads to the practical problem researchers face of choosing which control variables
to include in the model. The high-dimensional framework provides a useful setting for exploring data-dependent selection of control variables. This section considers the problem of
selecting a set of variables to include in a linear model from a large set of possible control
variables.
The structure of the TBFMS procedure ensures that any estimated coefficient that is not
set to zero can be reliably differentiated from zero relative to estimation noise. Conversely,
any coefficient that cannot be distinguished reliably from zero will be estimated to be
exactly zero. This property also complicates inference after model selection in sparse models
that may have a set of variables with small but non-zero coefficients in addition to strong
predictors. In this case, use of TBFMS may result in excluding variables with small but
non-zero coefficients, which may lead to non-negligible omitted variables bias and irregular
sampling behavior of estimates of parameters of interest. This intuition is formally developed
in [54] and [44]. Offering solutions to this problem with fully independent data is the focus
of a number of recent papers; see, for example, [8]; [6]; [71], [10], [9], [66]; [37], and [13].11
In the previous references, Lasso and related shrinkage estimators were used in place of
TBFMS.
Consider data given by Dn = {(yi , xi , wi )}ni=1 ∼ Pn where yi ∈ R are outcome variables,
xi ∈ R are variables of interest, and wi ∈ Rp are controls. Formally, the class B of linear
models is defined to contain joint distribution P (for any n) such that
yi = xi β0 + wi0 θ0B1 + εi
xi = wi0 θ0B2 + ui
for some parameters β0 ∈ R, θ0B1 , θ0B2 ∈ Rp with E[εi |wi , xi ] = 0 and E[ui |wi ] = 0.12
Here, the impact of the policy/treatment variable xi on the outcome yi is measured by the
parameter β0 which is the target of inference. The variables wi are potentially important
10
See also [34], who consider many-weak-instruments in a p > n setting.
These citations are ordered by date of first appearance on arXiv.
12
As was discussed in the case with model A, extending B to allow approximate sparsity is trivial.
11
Damian Kozbur/Testing-Based Forward Model Selection
29
conditioning variables. Data are assumed independent and identically distributed across
i. The confounding factors wi affect the policy variable via the function wi0 θ0B2 and the
outcome variable via the function wi0 θ0B1 . Both of the parameters θ0B1 and θ0B2 are unknown.
Informative inference about β0 is not possible in this model without imposing further
structure since p > n elements in wi are allowed. The additional structure is added by
assuming that a sparsity condition applies to both θ0B1 and θ0B2 . This implies that exogeneity
of di may be taken as given once one controls linearly for a relatively small number, s0 < n, of
the variables in wi whose identities are a priori unknown. Specifically, impose the following
restrictions.
Condition 7 (Regularity for Control Selection.) Pn ∈ B satisfy Condition 5 for {(xi , wi )}ni=1
and {(yi , wi )}ni=1 and Condition SM from [14].13
To estimate β0 in this environment, adopt the post-double-selection method of [9]. This
method proceeds by first substituting to obtain predictive relationships for the outcome yi
and the treatment xi in terms of only control variables:
yi = wi0 θ0RF + vi
xi = wi0 θ0FS + ui
with θ0FS = θ0B2 and θ0RF = θ0B1 + β0 θ0B2 . Next use two variable selection steps. TBFMS is
applied to each of the above two equations to select one set of variables that are useful for
predicting yi and another set of variables useful for predicting xi . Once this is done, the
union of the selected sets will index the final set of control variables.
[9] develop and discuss the post-double-selection method in detail. They note that including the union of the variables selected in each variable selection step helps address the
issue that model selection is inherently prone to errors unless stringent assumptions are
made. As noted by [44], the possibility of model selection mistakes precludes the possibility
of valid post-model-selection inference based on a single Lasso regression within a large
class of interesting models. Using both model selection steps guards against such model
13
For ease of reference, Condition SM from [14] is reproduced here, adapted only to fit in with present no-
tation. It states the following. For (y̆i , ε̆i ) = (yi , vi ) and for (y̆i , ε̆i ) = (xi , ui ) the following hold. E[En [|xi |q ]] =
O(1), E[vi2 |wi , ui ] and c E[u2i |wi ] are a.s. bounded by a constant above and away from zero, uniformly in i and
2 2
3 3
2
n. E[En [|ε̆i |q ]] + E[En [y̆i2 ]] + maxj6 E[En [wij
y̆i ]] + E[En [|wij
ε̆i |]] + 1/E[En [wij
]] = O(1). log(p) = o(n1/3 ).
2 2
maxj6p |(En − EEn )[wij 2ε̆2i ]| + |(En − EEn )[wij
y̆i ]| + maxi6n kwi k2∞ s0 log(max(p, n))/n = o(1) with proba-
bility 1 − o(1). Note that some of these conditions are redundant with Condition 5, but are nonetheless still
stated for completeness
Damian Kozbur/Testing-Based Forward Model Selection
30
selection mistakes and guarantees that the variables excluded in both model selection steps
have a negligible contribution to omitted variables bias.
Formally, the estimate for βb is given as follows. Obtain θbFS from applying Algorithm
2 to {(xi , wi )}n using the hypothesis tests in Definition 1. Set SbFS . Obtain θbRF from
i=1
applying Algorithm 2 to {(yi , wi )}ni=1 using the hypothesis tests in Definition 1. Set SbRF .
Set Sb = SbFS ∪ SbRF . Set
b θb B1 ) = Q
b −1 En [(xi , w b)yi ] where Q
b = En [(xi , w b)(xi , w b)0 ].
(β,
b
iS
iS
iS
S
bQ
b −1 ]11 where Ω
b = En [b
b −1 Ω
ε2i (xi , wiSb)(xi , wiSb)0 ]
Vb = [Q
b i − w0 θb B1 . Therefore, βb is the OLS estimate obtained by regressing yi on
and εbi = yi − βx
b S
b
iS
xi and wiSb, and Vb is the conventional heteroskedasticity-robust variance estimate. As in
b Let V = E[u2 ]−1 E[u2 v 2 ]E[u2 ]−1 .
the case of Model A, Vb is an estimate of the variance of β.
i
i i
i
The next theorem shows that βb is asymptotically Gaussian and that Vb −V vanishes asymptotically in probability.
Theorem 7. Consider sequences of data sets Dn = {(yi , xi , wi )}ni=1 ∼ Pn . Uniformly over
all sequences Pn ∈ B for which Condition 7 holds, as n → ∞,
n1/2 V −1/2 (βb − β0 ) →d N(0, 1) and
V − Vb →p 0.
This theorem verifies that the OLS estimator that regresses yi on xi and the union
of variables selected by TBFMS in the two stages described in Algorithm 4 is consistent and asymptotically normal with asymptotic variance that can be estimated with a
heteroskedasticity-robust standard error estimator. Inference based on this result will be
valid uniformly over a large class of data-generating processes, which includes cases in which
perfect variable selection is impossible.
4.3. Post Model Selection Inference: Simulation Study
This section presents simulation evidence about the finite sample performance of inference
about structural parameters. The simulation study provides a comparison with Lasso-based
estimators as was done in the previous simulation study. The simulation design is divided
into two parts, corresponding to Models A and B discussed above. The data-generating
processes considered are given by Table 2.
Damian Kozbur/Testing-Based Forward Model Selection
Table 2
Structural Estimation Simulation Design
Structural Estimation Model A: High-Dimensional Instruments.
Data:
DGP:
Settings:
Target:
Dn = (yi , xi , zi )n
i=1 iid
yi = xi β0 + εi
xi = zi0 θ0A + ui
p = dim(zi ) = 2n
A
= bj−1
1j6s0
θ0j
0
zij ∼ N(0, 1),
withcorr(zij , zik
) = 0.5|j−k|
1
0.5
(εi , ui ) ∼ σi N 0,
0.5
1
P
σi = exp(ρ0 pj=1 0.75(p−j) zij )
n ∈ {100, 500}
s0 = 6
b0 ∈ {−0.5, 0.5}
ρ0 ∈ {0, 1}
β0 = 1
Structural Estimation Model B: High-Dimensional Controls.
Data:
DGP:
Settings:
Target:
Dn = (yi , xi , wi )n
i=1 iid
yi = xi β0 + wi0 θ0B1 + εi
xi = wi0 θ0B2 + ui
p = dim(zi ) = 2n
B1
= bj−1
1j6s0
θ0j
0
B2
θ0j = sin(j)1j6s0
wij ∼ N(0, 1), with
, wik ) = 0.5|j−k|
corr(wij
1 0
(εi , ui ) ∼ σi N 0,
0 1
P
σi = exp(ρ0 pj=1 0.75(p−j) zij )
n ∈ {100, 500}
s0 = 6
b0 ∈ {−0.5, 0.5}
ρ0 ∈ {0, 1}
β0 = 1
31
Damian Kozbur/Testing-Based Forward Model Selection
32
Five estimators are considered for estimating both Models A and B, and are named
TBFMS I, TBFMS II, Lasso-CV, Post-Het-Lasso, and Oracle, analogous to the simulation
study of Section 3. The estimators differ only in that they replace TBFMS I with a different
model selection technique in selecting covariates into the final estimated model as described
in the previous sections. The respective model selection techniques used to replace TBFMS
I in A and B are identical to those described in the simulation study of Section 3.
For each estimator and each simulation setting, the bias, the standard deviation of the
point estimates, coverage probability and the average interval length are computed over the
simulation replications. In the case of Model A, the number of times that no instruments
were selected is also tracked. In these cases, no estimator of βb is defined. The bias, standard
deviation, coverage, and interval length calculations are made on only the replications for
which at least one instrument was selected. Finally, final variance estimates Vb for βb are
based on jackknife standard errors (more precisely, an approximation to jackknife standard
errors given in [60] are used for model A and HC3 standard errors [69] are used for model
B). Simulation results for Model A are shown in Table 3. Simulation results for Model B
are shown in Table 4.
The simulation results indicate that across the data-generating processes considered,
TBFMS I and TBFMS II generally achieve the best performance in terms of coverage
probabilities. In addition, their bias, standard deviation, and interval lengths closely resemble the Oracle estimator. Note that in some simulations (most notably in Model B, Panel
A) there is a large difference in coverage probabilities between the TBFMS I and II estimates and the Post-Het-Lasso estimate. Despite the fact that all the Post-Het-Lasso-based
confidence sets are asymptotically uniformly valid, this example highlights the fact that
finite sample model selection properties remains important considerations. In this example,
the signal-to-noise ratio in the first stage is low, which is a setting in which detection of
important variables is more difficult for Post-Het-Lasso relative to TBFMS I and II. Interestingly, in this case, using the relaxed penalty level with Lasso-CV does not help in terms
of coverage probability.
Damian Kozbur/Testing-Based Forward Model Selection
33
Table 3
Model A Simulation Results: Instrument Selection
n = 100
Bias
StDev
Length
n = 500
Cover
Bias
StDev
Length
Cover
A. ρ0 = 0: Homoskedastic, s0 = 6: High Sparsity, b0 = 0.5: Positive Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
0.00
-0.00
0.00
-0.03
-0.00
0.07
0.07
0.07
0.07
0.07
0.26
0.25
0.28
0.24
0.25
0.93
0.93
0.94
0.87
0.94
-0.00
-0.00
-0.00
-0.02
-0.00
0.03
0.03
0.03
0.03
0.03
0.11
0.11
0.12
0.11
0.11
0.93
0.93
0.93
0.88
0.93
B. ρ0 = 0.5: Heteroskedastic, s0 = 6: High Sparsity, b0 = 0.5: Positive Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
-0.01
-0.02
-0.06
-0.15
-0.07
0.11
0.10
0.07
0.15
0.11
0.41
0.39
0.26
0.39
0.40
0.93
0.91
0.75
0.59
0.87
0.00
-0.01
-0.00
-0.07
-0.02
0.05
0.05
0.05
0.07
0.05
0.21
0.20
0.20
0.19
0.19
0.95
0.94
0.96
0.61
0.93
C. ρ0 = 0: Homoskedastic, s0 = 6: High Sparsity, b0 = −0.5: Alternating Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
-0.00
-0.00
-0.00
-0.01
-0.00
0.04
0.04
0.04
0.04
0.04
0.15
0.15
0.16
0.15
0.15
0.95
0.94
0.94
0.93
0.95
0.00
0.00
0.00
-0.00
0.00
0.02
0.02
0.02
0.02
0.02
0.07
0.07
0.07
0.07
0.07
0.96
0.95
0.95
0.95
0.95
D. ρ0 = 0.5: Heteroskedastic, s0 = 6: High Sparsity, b0 = −0.5: Alternating Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
-0.01
-0.01
-0.03
-0.07
-0.03
0.07
0.07
0.05
0.09
0.07
0.25
0.24
0.20
0.29
0.24
0.93
0.92
0.88
0.74
0.91
-0.00
-0.00
-0.00
-0.02
-0.00
0.03
0.03
0.03
0.03
0.03
0.12
0.11
0.12
0.11
0.11
0.94
0.94
0.95
0.84
0.94
Simulation results for estimation in Model A. Simulations are conducted according to the design described
in Table 2 with s0 = 6, b0 ∈ {−0.5, 0.5}, ρ0 ∈ {0, 0.5}, and n ∈ {100, 500}. Estimates are presented for the
five estimators, TBFMS I, TBFMS II, Lasso-CV, Post-Het-Lasso, and Oracle described in the text. The
first column in each panel shows bias of the respective estimates for β0 . The second column in each panel
shows the standard deviation of the respective estimates for β0 . The third column in each panel shows
actual coverage probabilities of the respective 95% confidence intervals for β0 . The fourth column in each
panel shows length of confidence intervals for β0 . The fifth column in each panel shows the total number of
times across replications that no instruments were selected. Figures are based on 1000 simulation
replications.
The quantities in this table are calculated over replications where at least 1 instrument was selected.
TBFMS I and II each selected 0 instruments 5 times in Panel A, n = 100; 129 times in Panel B, n = 100; 7
times in Panel B, n = 500; 5 times in Panel D n = 100. Post-Het-Lasso selected 0 instruments 54 times in
Panel A, n = 100; 992 times in Panel B, n = 100; 167 times in Panel B, n = 500; 708 times in Panel C,
n = 100; 8 times in Panel C, n = 500; and 47 times in Panel D, n = 100. Lasso-CV selected 0 instruments
45 times in Panel B, n = 100; 1 time in Panel B, n = 500; 5 times in Panel D, n = 100.
Damian Kozbur/Testing-Based Forward Model Selection
34
Table 4
Model B Simulation Results: Control Selection in Linear Model
n = 100
Bias
StDev
Length
n = 500
Cover
Bias
StDev
Length
Cover
A. ρ0 = 0: Homoskedastic, s0 = 6: High Sparsity, b0 = 0.5: Positive Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
-0.05
0.01
-0.19
-0.19
0.01
0.10
0.11
0.05
0.05
0.10
0.38
0.43
0.21
0.21
0.42
0.84
0.94
0.06
0.06
0.95
0.05
0.01
-0.10
-0.11
0.01
0.18
0.17
0.09
0.09
0.32
0.98
1.20
0.55
0.64
1.17
0.88
0.95
0.00
0.00
0.95
B. ρ0 = 0.5: Heteroskedastic, s0 = 6: High Sparsity, b0 = 0.5: Positive Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
0.01
-0.01
-0.12
-0.13
-0.01
0.32
0.36
0.18
0.20
0.39
1.04
1.22
0.55
0.64
1.34
0.88
0.88
0.69
0.71
0.88
-0.00
-0.00
-0.10
-0.10
-0.00
0.24
0.26
0.15
0.15
0.38
0.77
0.81
0.43
0.43
0.83
0.91
0.90
0.63
0.63
0.91
C. ρ0 = 0: Homoskedastic, s0 = 6: High Sparsity, b0 = −0.5: Alternating Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
-0.02
-0.00
-0.08
-0.08
0.00
0.11
0.11
0.06
0.06
0.10
0.38
0.43
0.23
0.23
0.41
0.86
0.91
0.68
0.69
0.94
-0.03
0.00
-0.08
-0.08
0.00
0.06
0.04
0.02
0.02
0.04
0.18
0.17
0.10
0.10
0.17
0.78
0.95
0.09
0.09
0.95
D. ρ0 = 0.5: Heteroskedastic, s0 = 6: High Sparsity, b0 = −0.5: Alternating Sign
TBFMS I
TBFMS II
Post-Het-Lasso
Lasso-CV
Oracle
0.02
0.02
-0.13
-0.03
0.02
0.34
0.37
0.21
0.20
0.39
1.04
1.22
0.59
0.63
1.34
0.86
0.87
0.66
0.90
0.89
0.01
0.00
-0.04
-0.04
-0.00
0.23
0.24
0.13
0.13
0.25
0.76
0.81
0.42
0.42
0.82
0.89
0.90
0.86
0.86
0.90
Simulation results for estimation in Model B. Simulations are conducted according to the design described
in Table 2 with s0 = 6, b0 ∈ {−0.5, 0.5}, ρ0 ∈ {0, 0.5}, and n ∈ {100, 500}. Estimates are presented for the
five estimators, TBFMS I, TBFMS II, Lasso-CV, Post-Het-Lasso, and Oracle described in the text. The
first column in each panel shows bias of the respective estimates for β0 . The second column in each panel
shows standard deviation of the respective estimates for β0 . The third column in each panel shows actual
coverage probabilities of the respective 95% confidence intervals for β0 . The fourth column in each panel
shows length of confidence intervals for β0 . Figures are based on 1000 simulation replications.
Damian Kozbur/Testing-Based Forward Model Selection
35
4.4. Empirical Illustration: Estimating the Effects of Institutions on
Economic Output
In order to illustrate the use of testing-based forward model selection to help answer an
empirical economic question, this section revisits the problem of estimating the effect of
institution quality on aggregate economic output considered by Acemoglu, Johnson, and
Robinson in [1]. A similar exercise on this data using Lasso-based methods was performed
in [12].
To estimate the effect of institutions on output, it is necessary to address the fact that
both (1) better institutions can lead to higher output; and (2) higher output can also lead to
the development of better institutions. Because institutions and output levels both potentially affect each other, a simple correlation or regression analysis will not recover the causal
quantity of interest. [1] introduce an instrumental variable strategy, using early European
settler mortality as an instrument for institution quality. The validity of this instrument
requires first a relevance assumption that early settler mortality is predictive of quality of
current institutions. [1] argue that settlers set up lasting institutions in places where they
were more likely to establish long-term settlements. They cite several references documenting the fact that Europeans were acutely aware of mortality rates in their colonies. They
also note that the institutions set up by early European settlers tend to be highly persistent.
These arguments make the relevance assumption likely to hold. The exclusion restriction
assumption is justified in [1] by the argument that GDP, while persistent, is unlikely to be
strongly influenced by mortality rates centuries ago, except through institutions.
In their paper, [1] note that their IV strategy will be invalid if there are other factors that
are highly persistent and related to the development of institutions within a country and to
the country’s GDP. The primary candidate for such a factor discussed in [1] is geography.
In this exercise, take as given the fact that after controlling adequately for geography, it is
possible to use their instrument strategy to correctly identify the effect of institutions on
output. The outstanding problem then becomes the question of how, exactly, to adequately
control for geography. [1] controlled for the distance from the equator in their baseline
specification. They also considered specifications with continent dummies; see Table 4 in
[1].
In principle, there are many ways to construct control variables related to a broad notion
such as geography. These may include variables based on temperature, yearly rainfall, or
terrain. In order to deal with the ambiguity of the definition of geography, construct a large
set of different geographic variables. The strategy is to use TBFMS to choose from among
Damian Kozbur/Testing-Based Forward Model Selection
36
the many variables and perform a subsequent IV analysis. Let wi be a country level variable
with components consisting of the dummy variables for Africa, Asia, North America, and
South America plus the variables latitude, latitude2 , latitude3 , (latitude−.08)+ , (latitude−
.16)+ , (latitude − .24)+, ((latitude − .08)+ )2 , ((latitude − .16)+ )2 , ((latitude − .24)+ )2 ,
((latitude − .08)+ )3 , ((latitude − .16)+ )3 , ((latitude − .24)+ )3 where latitude denotes the
distance of a country from the equator normalized to be between 0 and 1, which is the same
set of controls as in [12]. Consider the model:
log(GDP per capitai ) = Protection from Expropriationi θ0 + wi0 β0 + εi .
Here, “Protection from Expropriation” is the same as was used in [1]: a measure of the
strength of individual property rights that is used as a proxy for the strength of institutions. The data here uses the same set of 64 country-level observations as in [1]. When the
set of control variables for geography, wi , is flexible enough, it is guaranteed that nothing
can be learned about the effect of interest, θ, because of lack of statistical precision. [1] do
not encounter such a problem because they assume the effect of geography is adequately
captured by one variable. Using TBFMS gives a complementary analysis that chooses controls from among the constructed set of geographic variables.
Formally, cast the estimation over a model as belonging to the class C which is characterized as follows. Consider data given by Dn = {(yi , xi , zi , wi )}ni=1 ∼ Pn where yi ∈ R is
outcome variable, xi ∈ R is an endogenous variable of interest, zi ∈ R is an instrument,
and wi ∈ Rp are controls. Define the class C of linear models to contain joint distribution
P (for all n) such that
yi = xi β0 + wi0 θ0C1 + εi
xi = zi0 π0 + wi0 θ0C2 + ui
zi = wi0 θ0C3 + vi
for parameters β0 , π0 ∈ R, θ0C1 , θ0C2 , θ0C3 ∈ Rp and E[εi |zi ] = 0.
Specializing to the current application, the fully expanded set of structural equations is
given by the following three relations.
log(GDP per capitai ) = Protection from Expropriationi θ0 + wi0 β0 + εi
Protection from Expropriationi = Settler Mortalityi π01 + wi0 Π02 + vi
Settler Mortalityi = wi0 γ0 + ui .
Damian Kozbur/Testing-Based Forward Model Selection
37
These yields three reduced form equations relating the structural variables to the controls.
log(GDP per capitai ) = wi0 β0 + ε̃i
Protection from Expropriationi = wi0 Π̃02 + ṽi
Settler Mortalityi = wi0 γ0 + ui .
Select all geographic variables according to three steps. Let S1 be the selected covariates
from running TBFMS over the data log(GDP per capitai ) on wi . Let S2 be the selected
covariates from running TBFMS over Protection from Expropriationi on, wi . Let S3 be
the selected covariates from running TBFMS over Settler Mortalityi on wi . The final set of
selected controls is given by Sb = S1 ∪S2 ∪S3 . Valid estimation and inference of the structural
parameter, θ, can then proceed by conventional IV estimation. As was the case in Models
A and B, formal validity of this procedure can be justified uniformly over sequences of data
sets generated from Model C (see [12] for example). The three model selection steps ensure
that the final estimates are robust to classical concerns about pre-test biases.
Table 5 presents estimates. Each column presents first stage coefficients and final structural coefficients using a different method for estimating the model. Heteroskedasticityrobust standard errors are shown in paranthesis. The first column of the table labeled
“Latitude” gives baseline results that control linearly for latitude, which corresponds to the
findings of [1], suggesting a strong positive effect of improved institutions on output with
a reasonably strong first stage. The second column controls for all 16 of the constructed
geography variables. This yields a visibly imprecise estimate of the effect of interest. This
is expected, since the number of control variables, 16, is large enough relative to the sample
size, 64, to prohibit precise estimation. The last column of Table 5, labeled “Forward Selection,” controls for the union of the set of variables selected by running testing-based forward
selection on each of the three reduced form equations, using heteroskedasticity-consistent
standard errors and significance thresholds as described in Section 5. The last column is
simply the IV estimate of the structural equation with the Africa dummy and the selected
latitude spline term as the control variables. Interestingly, the results are qualitatively similar to the baseline results though the first stage is somewhat weaker and the estimated
structural effect is slightly smaller.
Damian Kozbur/Testing-Based Forward Model Selection
38
Table 5
Estimates: Effect of Institutions on Economic Growth
Latitude
All Controls
TBFMS
First Stage
-0.53
(0.15)
-0.21
(0.21)
-0.38
(0.17)
Structural Estimate
0.96
(0.21)
0.98
(0.82)
0.83
(0.33)
Selected variables: 1Africa , (latitude − .16)1latitude>.16
*Note. The first two columns in this table (Latitude and All Controls) correspond to the first two columns
in Table 5 of [12]. The exact quantities estimated in [12] are slighly different despite using identical data
for the following reasons. Matlab and Stata differ in terms of coefficient estimates because of different
regularization schemes and tolerances for nearly singular matrices (which also depend on the ordering of
the variables). This difference is most relevant for the All Controls column. In order to facilitate
replication, all of the quantities in this table have now been produced within a single statistical software,
Matlab. The replication file is available from the author.
5. Conclusion
This paper has considered TBFMS for high-dimensional sparse linear regression problems.
The procedure is shown to achieve estimation rates matching those of Lasso and Post-Lasso
under a broad class of data-generating processes. In simulation studies, the method performs
well in terms of prediction and as an input into larger structural inference problems.
Damian Kozbur/Testing-Based Forward Model Selection
39
References
[1] Daron Acemoglu, Simon Johnson, and James A. Robinson. The colonial origins of
comparative development: An empirical investigation. American Economic Review,
91(5):1369–1401, 2001.
[2] Takeshi Amemiya. The non-linear two-stage least squares estimator. Journal of Econometrics, 2:105–110, 1974.
[3] Donald W. K. Andrews. Heteroskedasticity and autocorrelation consistent covariance
matrix estimation. Econometrica, 59(3):817–858, 1991.
[4] J. Bai and S. Ng. Forecasting economic time series using targeted predictors. Journal
of Econometrics, 146:304–317, 2008.
[5] J. Bai and S. Ng. Boosting diffusion indices. Journal of Applied Econometrics, 24,
2009.
[6] A. Belloni, D. Chen, V. Chernozhukov, and C. Hansen. Sparse models and methods for
optimal instruments with an application to eminent domain. Econometrica, 80:2369–
2429, 2012. Arxiv, 2010.
[7] A. Belloni and V. Chernozhukov.
Least squares after model selection in high-
dimensional sparse models. Bernoulli, 19(2):521–547, 2013. ArXiv, 2009.
[8] A. Belloni, V. Chernozhukov, and C. Hansen. Lasso methods for gaussian instrumental
variables models. 2010 arXiv:[math.ST], http://arxiv.org/abs/1012.1297, 2010.
[9] A. Belloni, V. Chernozhukov, and C. Hansen. Inference on treatment effects after
selection amongst high-dimensional controls. ArXiv, 2011. forthcoming, The Review
of Economic Studies.
[10] A. Belloni, V. Chernozhukov, and C. Hansen. Inference for high-dimensional sparse
econometric models. Advances in Economics and Econometrics. 10th World Congress
of Econometric Society. August 2010, III:245–295, 2013.
[11] A. Belloni, V. Chernozhukov, C. Hansen, and D. Kozbur. Inference in high dimensional
panel models with an application to gun contorl. ArXiv:1411.6507, 2014.
[12] A. Belloni, C. Hansen, and V. Chernozhukov. High-dimensional mehods and inference
on structural and treatmentt effectcs. Journal of Economic Perspectives, 28:29–50,
2014.
[13] Alexandre Belloni, Victor Chernozhukov, Ivan Fernández-Val, and Christian Hansen.
Program evaluation with high-dimensional data. arXiv:1311.2645, 2014.
[14] Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. Inference on treatment effects after selection amongst high-dimensional controls with an application to
Damian Kozbur/Testing-Based Forward Model Selection
40
abortion on crime. Review of Economic Studies, 81(2):608–650, 2014.
[15] P. J. Bickel, Y. Ritov, and A. B. Tsybakov. Simultaneous analysis of Lasso and Dantzig
selector. Annals of Statistics, 37(4):1705–1732, 2009.
[16] P. Bühlmann and S. van de Geer. Statistics for High-Dimensional Data: Methods,
Theory and Applications. Springer, 2011.
[17] Peter Bühlmann.
Boosting for high-dimensional linear models.
Ann. Statist.,
34(2):559–583, 2006.
[18] F. Bunea, A. Tsybakov, and M. H. Wegkamp. Sparsity oracle inequalities for the lasso.
Electronic Journal of Statistics, 1:169–194, 2007.
[19] F. Bunea, A. B. Tsybakov, , and M. H. Wegkamp. Aggregation and sparsity via `1
penalized least squares. In Proceedings of 19th Annual Conference on Learning Theory
(COLT 2006) (G. Lugosi and H. U. Simon, eds.), pages 379–391, 2006.
[20] F. Bunea, A. B. Tsybakov, and M. H. Wegkamp. Aggregation for Gaussian regression.
The Annals of Statistics, 35(4):1674–1697, 2007.
[21] E. Candès and T. Tao. The Dantzig selector: statistical estimation when p is much
larger than n. Ann. Statist., 35(6):2313–2351, 2007.
[22] G. Chamberlain. Asymptotic efficiency in estimation with conditional moment restrictions. Journal of Econometrics, 34:305–334, 1987.
[23] John C. Chao, Norman R. Swanson, Jerry A. Hausman, Whitney K. Newey, and
Tiemen Woutersen. Asymptotic distribution of jive in a heteroskedastic iv regression
with many instruments. Econometric Theory, 28(1):42–86, February 2012.
[24] X. Chen, Q.-M. Shao, and W. Biao Wu. Self-normalized Cram\’er Type Moderate
Deviations under Dependence. ArXiv e-prints, September 2014.
[25] V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. K. Newey,
and J. Robins. Double Machine Learning for Treatment and Causal Parameters. ArXiv
e-prints, July 2016.
[26] D. Chetverikov, Z. Liao, and V. Chernozhukov. On cross-validated Lasso. ArXiv
e-prints, May 2016.
[27] Abhimanyu Das and David Kempe. Submodular meets spectral: Greedy algorithms
for subset selection, sparse approximation and dictionary selection. In Lise Getoor and
Tobias Scheffer, editors, Proceedings of the 28th International Conference on Machine
Learning (ICML-11), pages 1057–1064, New York, NY, USA, 2011. ACM.
[28] W. Fithian, J. Taylor, R. Tibshirani, and R. Tibshirani. Selective Sequential Model
Selection. ArXiv e-prints, December 2015.
[29] Ildiko E. Frank and Jerome H. Friedman. A statistical view of some chemometrics
Damian Kozbur/Testing-Based Forward Model Selection
41
regression tools. Technometrics, 35(2):109–135, 1993.
[30] E. Gautier and A. Tsybakov. High-dimensional instrumental variables rergession and
confidence sets. arXiv:1105.2454v2 [math.ST], 2011.
[31] M. Grazier G’Sell, S. Wager, A. Chouldechova, and R. Tibshirani. Sequential Selection
Procedures and False Discovery Rate Control. ArXiv e-prints, September 2013.
[32] Alexander Grothendieck.
Resume de la theorie metrique des produits tensoriels
topologiques. Soc. Mat. Sao-Paulo, 8:1 – 79, 1953.
[33] O. Guédon and R. Vershynin.
Community detection in sparse networks via
Grothendieck’s inequality. ArXiv e-prints, November 2014.
[34] Christian Hansen and Damian Kozbur. Instrumental variables estimation with many
weak instruments using regularized {JIVE}. Journal of Econometrics, 182(2):290 –
308, 2014.
[35] Trevor Hastie, Robert Tibshirani, and Jerome Friedman. Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York, NY, 2009.
[36] Jian Huang, Joel L. Horowitz, and Fengrong Wei. Variable selection in nonparametric
additive models. Ann. Statist., 38(4):2282–2313, 2010.
[37] Adel Javanmard and Andrea Montanari. Confidence intervals and hypothesis testing
for high-dimensional regression. Journal of Machine Learning Research, 15:2869–2909,
2014.
[38] Bing-Yi Jing, Qi-Man Shao, and Qiying Wang. Self-normalized cramr-type large deviations for independent random variables. Ann. Probab., 31(4):2167–2215, 2003.
[39] Keith Knight. Shrinkage estimation for nearly singular designs. Econometric Theory,
24:323–337, 2008.
[40] V. Koltchinskii. Sparsity in penalized empirical risk minimization. Ann. Inst. H.
Poincar Probab. Statist., 45(1):7–57, 2009.
[41] D. Kozbur. Inference in Additively Separable Models With a High Dimensional Set of
Conditioning Variables. ArXiv e-prints, March 2015.
[42] D. Kozbur. Sharp Convergence Rates for Forward Regression in High-Dimensional
Sparse Linear Models. ArXiv e-prints, February 2017.
[43] Damian Kozbur. Testing-based forward model selection. American Economic Review,
107(5):266–69, May 2017.
[44] Hannes Leeb and Benedikt M. Pötscher. Can one estimate the unconditional distribution of post-model-selection estimators? Econometric Theory, 24(2):338–376, 2008.
[45] A. Li and R. Foygel Barber. Accumulation tests for FDR control in ordered hypothesis
testing. ArXiv e-prints, May 2015.
Damian Kozbur/Testing-Based Forward Model Selection
42
[46] K. Lounici. Sup-norm convergence rate and sign concentration property of lasso and
dantzig estimators. Electron. J. Statist., 2:90–102, 2008.
[47] K. Lounici, M. Pontil, A. B. Tsybakov, and S. van de Geer. Taking advantage of
sparsity in multi-task learning. arXiv:0903.1468v1 [stat.ML], 2010.
[48] Y. Luo and M. Spindler. L 2Boosting for Economic Applications. ArXiv e-prints,
February 2017.
[49] N. Meinshausen and P. Buhlmann. High dimensional graphs and variable selection
with the lasso. Ann. Statist., 34:1436–1462, 2006.
[50] N. Meinshausen and B. Yu. Lasso-type recovery of sparse representations for highdimensional data. Annals of Statistics, 37(1):2246–2270, 2009.
[51] Whitney K. Newey. Efficient instrumental variables estimation of nonlinear models.
Econometrica, 58:809–837, 1990.
[52] Whitney K. Newey and Kenneth D. West.
A simple, positive semi-definite het-
eroskedasticity and autocorrelation consistent covariance matrix.
Econometrica,
55(3):703–708, 1987.
[53] G. Pisier. Grothendieck’s Theorem, past and present. ArXiv e-prints, January 2011.
[54] Benedikt M. Pötscher. Confidence sets based on sparse estimators are necessarily large.
Sankhyā, 71(1, Ser. A):1–18, 2009.
[55] J. Qian, T. Hastie, R. Tibshirani, and N. Simon. Glmnet for matlab (2013).
[56] M. Rosenbaum and A. B. Tsybakov.
Sparse recovery under matrix uncertainty.
arXiv:0812.2818v1 [math.ST], 2008.
[57] M. Rudelson and S. Zhou. Reconstruction from anisotropic random measurements.
ArXiv:1106.1151, 2011.
[58] Mark Rudelson and Roman Vershynin. On sparse reconstruction from fourier and gaussian measurements. Communications on Pure and Applied Mathematics, 61:10251045,
2008.
[59] D. Staiger and J. H. Stock. Instrumental variables regression with weak instruments.
Econometrica, 65:557–586, 1997.
[60] Wurgler T. Steinhauer, A. Leverage and covariance matrix estimation in finite-sample
iv regressions. Working Paper 521, IEW, University of Zurich, 2010.
[61] T. Tao. When is correlation transitive? [https://terrytao.wordpress.com/2014/06/05/whenis-correlation-transitive/], June 2014.
[62] R. Tibshirani. Regression shrinkage and selection via the lasso. J. Roy. Statist. Soc.
Ser. B, 58:267–288, 1996.
[63] R. J. Tibshirani, J. Taylor, R. Lockhart, and R. Tibshirani. Exact Post-Selection
Damian Kozbur/Testing-Based Forward Model Selection
43
Inference for Sequential Regression Procedures. ArXiv e-prints, January 2014.
[64] Joel A. Tropp. Greed is good: algorithmic results for sparse approximation. Information Theory, IEEE Transactions on, 50(10):2231–2242, Oct 2004.
[65] S. A. van de Geer. High-dimensional generalized linear models and the lasso. Annals
of Statistics, 36(2):614–645, 2008.
[66] Sara van de Geer, Peter Bhlmann, Yaacov Ritov, and Ruben Dezeure. On asymptotically optimal confidence regions and tests for high-dimensional models. Ann. Statist.,
42(3):1166–1202, 06 2014.
[67] M. Wainwright. Sharp thresholds for noisy and high-dimensional recovery of sparsity
using `1 -constrained quadratic programming (lasso). IEEE Transactions on Information Theory, 55:2183–2202, May 2009.
[68] Hansheng Wang. Forward regression for ultra-high dimensional variable screening.
Journal of the American Statistical Association, 104:488:1512–1524, 2009.
[69] Halbert White. A heteroskedasticity-consistent covariance matrix estimator and a
direct test for heteroskedasticity. Econometrica, 48(4):817–838, 1980.
[70] C.-H. Zhang and J. Huang. The sparsity and bias of the lasso selection in highdimensional linear regression. Ann. Statist., 36(4):1567–1594, 2008.
[71] Cun-Hui Zhang and Stephanie S. Zhang. Confidence intervals for low dimensional
parameters in high dimensional linear models. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 76(1):217–242, 2014.
[72] Tong Zhang. On the consistency of feature selection using greedy least squares. Journal
of Machine Learning, 10:555–568, 2009.
[73] Tong Zhang. Adaptive forward-backward greedy algorithm for learning sparse representations. IEEE Trans. Inf. Theor., 57(7):4689–4708, July 2011.
Damian Kozbur/Testing-Based Forward Model Selection
44
Appendix A: Proof of Theorem 1
Proof. The proof of Theorem 1 is divided into seven steps. Step 1 shows the first statement
of Theorem 1. Step 2 defines a useful normalization of the selected covariates. Step 3
establishes certain bounds on the average correlation between selected covariates. Steps
4-6 show that if sb is too high, then there must exist subsets of the selected covariates
over which the average correlation must exceed what is permitted by assumption on the
sparse eigenvalues of the empirical Gram matrix G. Step 7 concludes by pulling together
the previous six steps.
Step 1
b 2 ] which depends on sb
This first section of the proof provides a bound on En [(x0i θ0 − x0i θ)
b = `(Sb ∪ S0 ) + [`(S)
b −
thereby proving the first statement of Theorem 1. First note that `(S)
b and `(Sb ∪ S0 ) 6 `(θ0 ). In addition, by Lemma 3.3 of [27],
b = `(θ)
`(Sb ∪ S0 )]. Note that `(S)
b − `(Sb ∪ S0 ) 6 ϕmin (b
`(S)
s + s0 )(G)−1
X
b 6 s0 tϕmin (b
(−∆j `(S))
s + s0 )(G)−1 .
b
j∈S0 \S
This gives
b 6 `(θ0 ) + s0 tϕmin (b
`(θ)
s + s0 )(G)−1 .
Expanding the above two quadratics in `(·) gives
b 2 ] 6 |2En [εi x0 (θb − θ0 )]| + s0 tϕmin (b
En [(x0i θ0 − x0i θ)
s + s0 )(G)−1
i
b 1 + s0 tϕmin (b
6 2kEn [εi x0i ]k∞ kθ0 − θk
s + s0 )(G)−1
b 1:
To bound kθ0 − θk
p
b2
sb + s0 kθ0 − θk
p
b 2 ]1/2 .
6 sb + s0 ϕmin (b
s + s0 )(G)−1 En [(x0i θ0 − x0i θ)
b1 6
kθ0 − θk
b 2 ]1/2 gives
Combining the above bounds and dividing by En [(x0i θ0 − x0i θ)
b 2 ]1/2 6 2kEn [εi x0 ]k∞
En [(x0i θ − x0i θ)
i
+
p
sb + s0 ϕmin (b
s + s0 )(G)−1
s0 tϕmin (b
s + s0 )(G)−1
.
b 2 ]1/2
En [(x0 θ0 − x0 θ)
i
i
Damian Kozbur/Testing-Based Forward Model Selection
45
b 2 ]1/2 6
Finally, either En [(x0i θ0 − x0i θ)
p
s0 tϕmin (b
s + s0 )(G)−1 , in which case the first statep
b 2 ]1/2 > s0 tϕmin (b
ment of Theorem 1 holds, or alternatively En [(x0i θ0 − x0i θ)
s + s0 )(G)−1 ,
in which case
p
b 2 ]1/2 6 2kEn [εi x0 ]k∞ sb + s0 ϕmin (b
En [(x0i θ − x0i θ)
s + s0 )(G)−1
i
p
+ s0 tϕmin (b
s + s0 )(G)−1
and the first statement of Theorem 1 follows.
Step 2
This section of the proof defines true and false covariates, introduces a convenient orthogonalization of all selected covariates, and associates to each false selected covariate a
parameter γ̃j on which the analysis is based.
Let xj = [x1j , ..., xnj ]0 be the vector in Rn with components xij stacked vertically. Similarly, define ε = [ε1 , ..., εn ]0 and y = [y1 , ..., yn ]0 . Let vk ∈ Rn , k = 1, ..., s0 denote true
covariates which are defined as the the vectors xj for j ∈ S0 . Define false covariates simply
as those which do not belong to S0 .
Consider any point in time in the the Simple Forward Selection algorithm when there
are m false covariates selected into the model. These falsely selected covariates are denoted
w1 , ..., wm , each in Rn , ordered according to the order they were selected.
The true covariates are also ordered according to the order they are selected into the
model. Any true covariates unselected after the m false covariate selection are temporarily
ordered arbitrarily at the end of the list. Let Mk be projection in Rn onto the space
orthogonal to span({v1 , ..., vk }). Let
ṽk =
Mk−1 vk
0
(vk Mk−1 vk )1/2
In addition, set
ε̃ =
for k = 1, ..., s0 .
Ms0 ε
.
0
(ε Ms0 ε)1/2
Let Ṽtemp = [ṽ1 , ..., ṽs0 ], ordered according to the temporary order. Note that there is
θ̃ ∈ Rs0 and θ̃ε̃ ∈ R such that
Ṽtemp θ̃temp + θ̃ε̃ ε̃ = y.
At this time, reorder the true covariates. Let b
k denote the index of the final true covariate
selected into the model when the m-th false covariate is selected. The variables ṽ1 , ..., ṽbk
maintain their original order. The unselected true covariates ṽbk+1 , ..., ṽs0 are reordered in
Damian Kozbur/Testing-Based Forward Model Selection
46
such a way that under the new ordering, θ̃k,temp > θ̃l,temp whenever l > k. Also define
Ṽ = [ṽ1 , ..., ṽs0 ] consistent with the new ordering. Redefine θ̃ by Ṽ θ̃ + θ̃ε̃ ε̃ = y so that it
is also consistent with the new ordering. Note that no new orthogonalization needs to be
done.
For any set S, Let QS be projection onto the space orthogonal to span({xj , j ∈ S}).
For each selected covariate, wj , set Spre-wj to be the set of (both true and false) covariates
selected prior to wj . Define
w̃j = cj QSpre-wj wj
where the normalization constants cj are defined in the next paragraph.
Each w̃j can be decomposed into components w̃j = r̃j + ũj with r̃j ∈ span(Ṽ ) and
ũj ∈ span(Ṽ )⊥ . The normalizations cj introduced above are then chosen so that ũ0j ũj = 1.
Associates to each false covariate w̃j , a vector γ̃j ∈ Rs0 , defined as the solution in Rs0 to
the following equation
Ṽ γ̃j = r̃j .
Set γ̃j ε̃ = ε̃0 w̃j . Assume without loss of generality that each component of θ̃ is positive
(since otherwise, the true covariates can just be multiplied by −1.) Also assume without
loss of generality that γ̃j0 θ̃ > 0.
Step 3
This section provides upper bounds on quantities related to the γ̃j defined above. The idea
guiding the argument in the next sections is that if too many covariates wj are selected,
then on average they must be correlated with each other since they must be correlated to
y. For a discussion of partial transitivity of correlation, see [61]. If the covariates are highly
correlated amongst themselves, then ϕmin (m + s0 )(G)−1 must be very high. As a result,
the sparse eigenvalues of G can be used to upper bound the number of selections. Average
correlations between covariates are tracked with the aid of the quantities γ̃j .
Divide the set of false covariates into two sets A1 and A2 where
(
)
(
)
t1/2 n1/2
t1/2 n1/2
A1 = j : |γ̃j ε̃ | 6
, A2 = j : |γ̃j ε̃ | >
.
(2ε0 Ms0 ε)1/2
(2ε0 Ms0 ε)1/2
Sections 3 - 5 of the proof bound the number of elements in A1 . Section 6 of the proof
bounds the number of elements in A2 .
Suppose the set A1 contains m1 total false selections. Collect these false selections into
W̃ = [w̃j1 , ..., w̃jm1 ]. Set R̃ = [r̃j1 , ..., r̃jm1 ], Ũ = [ũj1 , ..., ũjm1 ]. Decompose W̃ = R̃ + Ũ .
Damian Kozbur/Testing-Based Forward Model Selection
47
Then W̃ 0 W̃ = R̃0 R̃ + Ũ 0 Ũ . Since diag(Ũ 0 Ũ ) = I, it follows that the average inner product
between the ũj , given by ρ̄:
ρ̄ =
X
1
ũ0j ũl ,
m1 (m1 − 1)
j6=l∈A1
must be bounded below by
ρ̄ > −
1
m1 − 1
due to the positive definiteness of Ũ 0 Ũ . This implies an upper bound on the average offdiagonal term in R̃0 R̃ since W̃ 0 W̃ is a diagonal matrix. Since ṽk are orthonormal, the sum
P
P
P
of all the elements of R̃0 R̃ is given by k j∈A1 γ̃j k22 . Since k j∈A1 γ̃j k22 = j∈A1 kγ̃j0 k22 +
P
0
0
j6=l∈A1 γ̃j γ̃l and since W̃ W̃ is a diagonal matrix, it must be the case that
X
1
γ̃j0 γ̃l = −ρ̄.
m1 (m1 − 1)
j6=l∈A1
Therefore,
ρ̄ =
1
m1 (m1 − 1)
X
j∈A1
2
γ̃j
2
−
X
kγ̃j k22 6
j∈A1
1
.
m1 − 1
This implies that
X
2
γ̃j
j∈A1
2
6 m1 +
X
kγ̃j k22 .
j∈A1
Next, bound maxj∈A1 kγ̃j k22 . Note kγ̃j k22 = kr̃j k22 since Ṽ is orthonormal. Note that
kũj k22 /kw̃j k22 = 1/kw̃j k22 is lower bounded by ϕmin (m + s0 )(G). This follows from the fact
that you can associate kũj /cj k22 to an element of a the inverse covariance matrix for wj and
previously selected covariates. Therefore, kr̃j k22 = kw̃j k22 − 1 6 ϕmin (m + s0 )(G)−1 − 1. It
follows that
max kγ̃j k22 6 ϕmin (m + s0 )(G)−1 − 1.
j∈A1
This then implies that
X
2
γ̃j
j∈A1
2
6 m1 ϕmin (m + s0 )(G)−1 .
The same argument as above also shows that for any choice ej ∈ {−1, 1} of signs, it is
always the case that
X
j∈A1
2
ej γ̃j
2
6 m1 ϕmin (m + s0 )(G)−1 .
(In more detail, take W̃e = [w̃j1 ej1 , ..., w̃jm1 ejm1 ], etc. and rerun the same argument.)
Damian Kozbur/Testing-Based Forward Model Selection
48
Step 4
Next search for a particular choice of signs {ej }j∈A1 which give a lower bound proportional
to m1 2 /s0 on the above term. Note that this will imply an upper bound on m1 . For each
k = 1, ..., s0 , let A1k be the set which contains those j ∈ A1 such that wj is selected before
vk , but not before any other true covariate. Note that the sets A1(bk+2) , ..., A1(s0 +1) are set
empty if b
k < s0 . Also, empty sums are set to zero. Define the following two matrices:
Γ=
P
γ̃j1
j∈A11
0
P
γ̃j2
P
...
j∈A11
j∈A11
P
γ̃js0
γ̃j2
P
...
j∈A12
γ̃js0
j∈A12
..
.
..
.
..
0
0
...
..
.
.
P
γ̃js0
θ̃1
θ̃1
θ̃2
θ̃1
, B =
.
..
θ̃s0
j∈A1s0
Note that the kth row of Γ is equal to
P
j∈A1k
θ̃1
θ̃2
θ̃1
...
θ̃2
θ̃2
...
..
.
..
θ̃s0
θ̃2
...
.
θ̃s0
θ̃1
θ̃s0
θ̃2
..
.
θ̃s0
θ̃s0
γ̃k since the orthogonalization process had
enforced γ̃jl = 0 for each l < k. Therefore, the diagonal elements of the product ΓB satisfy
the equality
[ΓB]k,k =
X
γ̃j0 θ̃/θ̃k .
j∈Ak
Let C1 , C2 be constants such that
γ̃j0 θ̃/θ̃k > C1
for j ∈ A1k , and
θ̃k /θ̃l > C2
for l > k. These key constants are calculated explicitly in Section 5 of the proof. They imply
that
[ΓB]k,k > C1 |A1k | and tr(ΓB) > C1 m1 .
Further observe that whenever θ̃k > C2 θ̃l for each k, l > k, assuming without loss of
generality that C2 6 1, that (B + C2−1 I) is positive semidefinite. This can checked by
constructing auxiliary random variables who have covariance matrix B + C2−1 I: inductively
build a covariance matrix where the (k + 1)th random variable has θ̃k /θ̃k−1 covariance with
the kth random variable. Then B + C2−1 I has a positive definite symmetric matrix square
Damian Kozbur/Testing-Based Forward Model Selection
−1/2
root so let D2 = B + C2−1 I. Therefore, B = (D + C2
(and columns) of D each have norm 6
B=
1 + C2−1
−1/2
I)(D − C2
49
I). Note that the rows
and therefore B decomposes into a product
−1/2
where the rows of E, F have norms bounded by 1 + C2−1 + C2
E0F
C3 = 1 +
C2−1
+
. Therefore, let
−1/2
C2 .
Consider the set
Gs0 = {Z ∈ Rs0 ×s0 : Zij = Xi0 Yj for some Xi , Yj ∈ Rs0 , kXi k2 , kYj k2 6 1}
and observe that B̄ := C3 −1 B ∈ Gs0 . Then this observation allows the use of Grothendieck’s
inequality (using the exact form described in [33]) which gives
R
max tr(ΓZ) 6 KG
kΓ0 k∞→1 .
Z∈Gs0
R is an absolute constant which is known to be less than 1.783. It does not depend
Here, KG
R kΓ0 k
on s0 . Therefore, C1 m 6 tr(ΓB) = C3 tr(ΓB̄) 6 maxZ∈Gs0 tr(ΓZ) 6 KG
∞→1 , which
implies
R
KG
−1
C3 −1 C1 m1 6 kΓ0 k∞→1 .
R
Therefore, there is ν ∈ {−1, 1}s0 such that kν 0 Γk1 > KG
−1
C3 −1 C1 m1 . For this particular
choice of ν, it follows that
−1/2
kν 0 Γk2 > s0
R
KG
−1
C3 −1 C1 m1 .
P0 P
2
Then by definition of Γ, kν 0 Γk22 = k sk=1
j∈A1k νk γ̃j k2 . In Section 3, it was noted that
Pm1
k j=1 ej γ̃j k22 6 m1 ϕmin (m + s0 )(G)−1 for any choice of signs ej ∈ {−1, 1}m1 . It follows
that
s−1
0
R
KG
−2
C3 −2 C12 m21 6 m1 ϕmin (m + s0 )(G)−1
which yields the conclusion
2
R
m1 6 ϕmin (m + s0 )(G)−1 C1−2 C3 2 KG
s0 .
Step 5
It is left to calculate C1 , C2 which lower bound γ̃j0 θ̃/θ̃k for j ∈ A1k and θ̃k /θ̃l for l > k. A
simple derivation can be made to show that the incremental decrease in empirical loss from
the jth false selection is
−∆j `(Spre-wj ) =
1 0
1 1
y w̃j (w̃j0 w̃j )−1 w̃j0 y =
(θ̃0 γ̃j + θ̃ε̃0 γ̃j ε̃ )2
n
n w̃j0 w̃j
Damian Kozbur/Testing-Based Forward Model Selection
50
Note the slight abuse of notation in −∆j (Spre-wj ) signifying change in loss under inclusion
of wj rather than xj . Next,
(θ̃0 γ̃j + θ̃ε̃0 γ̃j ε̃ )2 6 2(θ̃0 γ̃j )2 + 2(θ̃ε̃0 γ̃j ε̃ )2
Since θ̃ε̃ = (ε0 Ms0 ε)1/2 , w̃j0 w̃j > 1, and j ∈ A1 it follows that
1 1
1 1 2
(θ̃ε̃0 γ̃j ε̃ )2 6
θ̃
0
n w̃j w̃j
n w̃j0 w̃j ε̃
t1/2 n1/2
2(ε0 Ms0 ε)1/2
!2
t
6 .
4
This implies
1 1
t
1
(−∆j `(Spre-wj )) 6
(θ̃0 γ̃j )2 + .
0
2
n w̃j w̃j
4
By the condition that the false j is selected, it holds that −∆j `(Spre-wj ) > t and so
1
4 (−∆j `(Spre-wj ))
>
t
4
which implies that
t
1
1
(−∆j `(Spre-wj )) − > (−∆j `(Spre-wj )).
2
4
4
Finally, this yields that
1
1
(γ̃j0 θ̃)2 > (−∆j `(Spre-wj )).
0
nw̃j w̃j
4
By the fact that wj was selected ahead of vk it holds that
−∆j `(Spre-wj ) > −∆k `(Spre-wj ).
Therefore, further bound the righthand side. Let z̃k be the projection of ṽk onto the space
orthogonal to all previously selected (true and false) covariates. Then
−∆k `(Spre-wj ) >
1 0
z̃ z̃k θ̃2 .
n k k
Furthermore, z̃k0 z̃k > ϕmin (m + s0 )(G)2 . This is seen by noting that z̃k results in the composition of two projections onto a span of covariates of size bounded by m + s0 .
This gives
1
11
(γ̃j0 θ̃)2 >
ϕmin (m + s0 )(G)2 θ̃k2 .
0
nw̃j w̃j
4n
Using the fact that w̃j0 w̃j > 1 implies that
1
(γ̃j0 θ̃)2 /θ̃k2 > ϕmin (m + s0 )(G)2 .
4
Now suppose no true variables remain when j is selected, then w̃j0 w̃j = ũ0j ũj = 1. Therefore,
Damian Kozbur/Testing-Based Forward Model Selection
−∆j `(Spre-wj ) =
51
1 2 2
γ̃ θ̃ > t
n j ε̃ ε̃
Note that θ̃ is given by θ̃ε̃ = ε̃0 y = ε0 Ms0 y/(ε0 Ms0 ε)1/2 = (ε0 Ms0 ε)1/2 . Therefore,
γ̃j2ε̃ > t
n
ε0 Ms0 ε
.
This implies that j ∈ A2 . Therefore, set C1 = 12 ϕmin (m + s0 )(G).
Next, construct C2 . For each selected true covariate, vk , set Spre-vk to be the set of (both
true and false) covariates selected prior to vk . Note that
θ̃k2 = −∆k `({v1 , ..., vk−1 }) > −∆k `(Spre-vk )
since {v1 , ..., vk−1 } ⊆ Spre−vk . In addition, if vk is selected before vl (or vl is not selected),
then
−∆k `(Spre-vk ) > −∆l `(Spre-vk ) > z̃l0 z̃l θ̃l2 > ϕmin (b
s + s0 )(G)2 θ̃l2 .
Therefore, taking
C2 = ϕmin (m + s0 )(G)
implies that θ̃k / θ̃l > C2 for any l > k.
Step 6
In this section, the number of elements of A2 is bounded. Recall that the criteria for j ∈ A2
is that |γ̃j ε̃ | >
t1/2 n1/2
.
(2ε0 Ms0 ε)1/2
Note also that γ̃j ε̃ is found by the coefficient in the expression
γ̃j ε̃ = ε̃0 w̃j = ε0
1
(ε0 Ms0 ε)1/2
Ms0 w̃j
Next, let H be the matrix H = [v1 , ..., vs0 , w1 , ..., wm ]. Note that
1
(ε0 Ms0 ε)1/2
Ms0 w̃j ∈ span(H)
Which implies that the above expression is unchanged when premultiplied by
H(H 0 H)−1 H 0 . Therefore,
γ̃j ε̃ = ε0 H(H 0 H)−1 H 0
1
(ε0 Ms0 ε)1/2
Ms0 w̃j .
Damian Kozbur/Testing-Based Forward Model Selection
52
Let µj be the +1 for each j ∈ A2 such that γ̃j ε̃ > 0 and −1 for each j ∈ A2 such that
t1/2 n1/2
(2ε0 Ms0 )1/2
γ̃j ε̃ < 0. By the fact that j ∈ A2 , γ̃j ε̃ µj >
X
ε0 H(H 0 H)−1 H 0
j∈A2
, summing over j ∈ A2 gives
t1/2 n1/2
w̃
µ
>
m
M
j j
2
s
(ε0 Ms0 ε)1/2 0
(2ε0 Ms0 ε)1/2
1
This implies that
(H 0 H)−1 H 0
1
X
(ε0 Ms0 ε)1/2
j∈A2
Ms0 w̃j µj
1
kε0 Hk∞ > m2
t1/2 n1/2
(2ε0 Ms0 ε)1/2
Which further implies that
√
m + s0 (H 0 H)−1 H 0
1
X
(ε0 Ms0 ε)1/2
j∈A2
Ms0 w̃j µj
2
kε0 Hk∞ > m2
t1/2 n1/2
(2ε0 Ms0 ε)1/2
Next, further upper bound the k · k2 term on the left side above by
(H 0 H)−1 H 0
6
1
X
(ε0 Ms0 ε)1/2
j∈A2
Ms0 w̃j µj
2
X
n−1/2
−1/2
ϕ
(s
+
m)(G)
kM
w̃j µj k2
min
0
s
0
(ε0 Ms0 ε)1/2
j∈A
2
next, by the fact that Ms0 is a projection (hence non-expansive) and w̃j are mutually
orthogonal,
n−1/2
6 0
ϕmin (s0 + m)(G)−1/2
(ε Ms0 ε)1/2
sX
kw̃j µj k22 .
j∈A2
In Section 3, it was shown that maxj kw̃j k22 6 ϕmin (s0 + m)(G)−1 . Therefore, putting the
above inequalities together,
n−1/2 √
t1/2 n1/2
−1 √
0
m
+
s
ϕ
(m
+
s
)(G)
m
kε
Hk
>
m
.
0 min
0
2
∞
2
(ε0 Ms0 ε)1/2
(2ε0 Ms0 ε)1/2
This implies that
m2 <
12 0
kε0 Hk2∞
(ε Ms0 ε)(m + s0 ) 0
ϕmin (m + s0 )(G)−2
nt
ε Ms0 ε
kEn [xi εi ]k2∞
ϕmin (m + s0 )(G)−2 .
t
Under the assumed condition that t1/2 > 2kEn [xi εi ]k∞ ϕmin (m + s0 )(G)−1 , it follows that
6 2(m + s0 )
1
m2 6 (m + s0 ).
2
By substituting m = m1 + m2 gives m2 6 m1 + s0 .
Damian Kozbur/Testing-Based Forward Model Selection
53
Step 7
This section concludes the proof of the second statement of the theorem by bringing together
R 2s
all of the facts proven in Steps 3-6. Combining m1 6 ϕmin (m + s0 )(G)−1 C1−2 C3 2 KG
0
and m2 6 m1 + s0 gives
h
i
R2
m 6 2ϕmin (m + s0 )(G)−1 C1−2 C3 2 KG
+ 1 s0 .
In addition,
1
C1 = ϕmin (m + s0 )(G),
2
C2 = ϕmin (m + s0 )(G),
C3 = (1 + ϕmin (m + s0 )(G)−1/2 + ϕmin (m + s0 )(G)−1 ),
R < 1.783. Therefore,
and KG
h
m 6 1 + 8 × 1.7832 × ϕmin (m + s0 )(G)−3
i
× (1 + ϕmin (m + s0 )(G)−1/2 + ϕmin (m + s0 )(G)−1 )2 s0 .
Since C32 6 9ϕmin (s0 + m)(G)−2 , the expression above can be simplified at the expense
of a slightly less tight constant, so that
m 6 [1 + 72 × 1.7832 × ϕmin (m + s0 )(G)−5 ]s0 .
Since this bound holds for each positive integer m of wrong selections, this concludes the
proof of Theorem 1.
Appendix B: Proof of Theorems 2 and 3
Theorem 2 follows by applying Theorem 1 in the following way. If sb grows faster than s0 ,
then there is m < sb such that s0 < m < Kn and m/s0 exceeds c0F (Kn ) = O(1), giving a
contradiction. The first statement of the theorem follows from applying the bound on sb.
√
√
b 2 6 sb + s0 ϕmin (b
b 1 6 sb + s0 kθ0 − θk
Theorem 3 follows by kθ0 − θk
s + s0 )(G)−1 En [(x0 θ0 −
i
b 2 ]1/2 .
x0i θ)
Appendix C: Proof of Theorem 4
Proof. First, ∆j `(S) is replaced with ∆j E(S). Second, Third, the constants C1 and C2 to
depend on ctest , c0test , c00test instead of t are made. Since
Damian Kozbur/Testing-Based Forward Model Selection
54
Step 1.
Let T be the event described by Condition 2. Then by assumption, P(T) > 1−α−3δtest /3 =
1 − α − δtest . The rest of the proof works on the event T.
Suppose that Algorithm 1 terminates before Ktest − s0 steps. Note that on T, Algorithm
b 6 ctest for every j ∈
b By application of the results
1 terminates at a step with −∆j E(S)
/ S.
of [27], Lemma 3.3, which relate the increase in R2 from inclusion of a set of regressors
to the increase in R2 from inclusion of each regressor from the set separately, noting that
b 6 s0 yields
|S0 \ S|
b 6 ϕmin (Ktest )(E[G])−1
|E(S0 ) − E(S)|
X
−∆j E(S) 6 s0 ctest ϕmin (Ktest )(E[G])−1 .
b
j∈S0 \S
Step 2.
Repeat the construction in Step 2 of the proof of Theorem 1. The inner product on
Rn is replaced by the standard inner product on L2 (Ω; Rn ) where Ω is an underlying
probability space. For example, hxj , xk iL2 (Ω;Rn ) = E[En [xij xik ]]. This yields parameters
(L2 (Ω;Rn )) s0
}k=1
θ̃, γ̃j ∈ Rs0 and θ̃ε̃ , γ̃j ε̃ ∈ R as well as an implied profile of operators {Mk
(L2 (Ω;Rn ))
{QS
}S⊆{1,...,p}
L2 (Ω; Rn ).
on
Step 3.
Divide the set of false covariates into three sets A1 , A2 , and A3 where
(
)
c0test 1/2
A1 = j < Ktest − s0 : |γ̃j ε̃ | 6
(2hε, Ms0 εiL2 (Ω;Rn ) )1/2
(
A2 =
c0test 1/2
j < Ktest − s0 : |γ̃j ε̃ | >
(2hε, Ms0 εiL2 (Ω;Rn ) )1/2
)
A3 = {j : j > Ktest − s0 } .
By arguments identical to Step 3 in the proof of Theorem 1, for any ej ∈ {−1, +1},
X
j∈A1
2
ej γ̃j
2
6 |A1 |ϕmin (Ktest )(E[G])−1 .
and
Damian Kozbur/Testing-Based Forward Model Selection
55
Step 4.
Let C1 be such that θ̃0 γ̃j /θ̃k > C1 for j ∈ A1k . Let C2 be such that θ̃k /θ̃l > C2 for l > k.
−1/2
Let C3 = 1 + C2−1 + C2
. By arguments identical to Step 4 in the proof of Theorem 1,
R 2
|A1 | 6 ϕmin (Ktest )(E[G])−1 C1−2 C32 (KG
) s0 .
Step 5.
This section calculates bounds C1 and C2 needed in Step 4 immediately above. Proceeding
as in Step 5 of the proof of Theorem 1, note that
−∆j E(Spre-wj ) = hy, w̃j i(hw̃j , w̃j i)−1 hw̃j , yi.
This allows calculation of C1 and C2 in Theorem 3 to proceed in a similar manner as in
Theorem 1. The statement ∆k `(Spre-wj ) > t is replaced by
∆k E(Spre-wj ) > c0test
for j selected. Similarly, the statement ∆j `(Spre-wj ) > ∆k `(Spre-wj ) is replaced by
∆j E(Spre-wj ) > c00test ∆k E(Spre-wj ).
Using similar reasoning
1
θ̃0 γ̃j /θ̃k > ϕmin (Ktest )(E[G])c00test = C1
2
and
θ̃k /θ̃l > ϕmin (Ktest )(E[G])c00test = C2 .
Step 6.
By similar arguments as in Step 6 in the proof of Theorem 1, |A1 ∪ A2 | 6 2|A1 | + s0 . Note
that relative to Theorem 1, some care must be taken to define what is meant by H. H is a
vector of elements of L2 (Ω; Rn ) and H 0 is a vector of elements of the dual space L2 (Ω; Rn )∗ .
Damian Kozbur/Testing-Based Forward Model Selection
56
Step 7.
By identical arguments as in Step 7 in the proof of Theorem 1,
h
i
R2
|A1 ∪ A2 | 6 2ϕmin (Ktest )(E[G])−1 C1−2 C3 2 KG
+ 1 s0 .
Using the values for C1 and C2 in Step 5,
|A1 ∪ A2 | 6 [1 + 72 × 1.7832 × ϕmin (Ktest )(E[G])−5 c00test
−4
]s0 .
Finally, if Ktest − s0 is larger than [1 + 72 × 1.7832 × ϕmin (Ktest )(E[G])−5 c00test −4 ]s0 , then it
must follow that |A3 | = 0. Then sb 6 |A1 ∪ A2 | + s0 , and from this the result follows.
Step 8.
This step proves the third statement of Theorem 4. For any S define θS∗ to be the minimizer of E(S). For any S define also dS = θS∗ − θS∗ 0 . Finally, let δ0 = θ0 − θS∗ 0 . Note that
E(S)−E(S0 ) = d0S E[G]dS . By the earlier steps, d0bE[G]dSb 6 s0 ctest ϕmin (Ktest )(E[G])−1 . But
S
√
d0bE[G]dSb > ϕmin (Ktest )(E[G])kdSbk22 . So kdSbk2 6 s0 ctest ϕmin (Ktest )(E[G])−1 . In addition,
S
δ0 is bounded by
kδ0 k2 = kE[G]−1 E[En [xi x0i θ0 + εi ]] − θ0 k2 = kE[En [x0iS0 εi ]k2
1/2
6 s0 max |E[En [xij εai ]]| 6
j
1√
s0 ctest ϕmin (Ktest )(E[G])−1
2
1/2 )
where the last bound comes from Cauchy-Schwarz (passing to E[En [x2ij ]]1/2 E[En [εa2
i ]]
along with the assumed condition on εai and the fact that c0test 6 ctest . Next,
0
∗
0
0
θb = G−1
b(xiS
b + xiS0 δ0 )]
b θS
b + εi − xiS∪S
b 0 dS
b En [xiS
S
h
i
−1
0
+
δ
)
= θS∗b + G−1
E
x
ε
+
G
E
x
x
(−d
n
n
0
b iS∪S
b
b i
b 0
b
b
iS
iS
S
S
S
h
i
⇒ kθb − θS∗bk2 6 ϕmin (b
E
x
x
(−d
+
δ
)
s)(G)−1/2 En [xiSbεi ] 2 + G−1
n
0
b iS∪S
b 0
b
b
S
iS
S
6 ϕmin (b
s)(G)
−1/2 1/2
sb
kEn [xi εi ]k∞
+ ϕmin (b
s)(G)−1/2 ϕmax (b
s + s0 )(G)1/2 (kdSbk2 + kδ0 k2 ).
2
Damian Kozbur/Testing-Based Forward Model Selection
57
Finally,
(En [(x0i θb − x0i θ0 )2 ])1/2 6 ϕmax (s0 + sb )(G)1/2 kθb − θ0 k2
6 ϕmax (s0 + sb )(G)1/2 (kθb − θS∗bk2 + kδ0 k2 + kdSbk2 )
6 ϕmax (s0 + sb)(G)1/2 ϕmin (s0 + sb)(G)−1/2 sb1/2 kEn [xi εi ]k∞
3 3
+ ϕmax (s0 + sb )(G)1/2 ( + ϕmax (s0 + sb )(G)1/2 ϕmin (b
s + s0 )(G)−1/2 )
2 2
√
× s0 ctest ϕmin (Ktest )(E[G])−1
6 ϕmax (s0 + sb)(G)1/2 ϕmin (s0 + sb)(G)−1/2 sb1/2 kEn [xi εi ]k∞
√
+ 3ϕmax (s0 + sb )(G)ϕmin (b
s + s0 )(G)−1/2 ) s0 ctest ϕmin (Ktest )(E[G])−1 .
Appendix D: Proof of Theorem 5
Step 1.
This section sets up basic notation and outlines the proof. The strategy is to apply Theorem 4 using the conditional distribution Px on for Dn , conditional on x. The uncon∗|x
ditional result is then shown to follow. Let Ex (S) = E[`(S)|x]. In addition, let θjS =
(xj QS xj )−1 E[xj QS (xθ0 + εa )|x].
Next let
bjS = Vb −1/2 ([θbjS ]j − [θ∗|x ]j ).
Z
jS
jS
Let tα = Φ−1 (1 − α/p). Let A be the event given by
1 + cτ
b
A = |ZjS | 6
τbjS tα for all j, |S| < Kn .
2
Note that
∗|x
∆j Ex (S) = [θjS ]2j AjS
for AjS defined by AjS = [G−1
jS ]jj .
The next three steps calculate the extent to which the event A implies that the tests
have power, control size, and exhibit continuity. Step 2 performs a power calculation. Step
3 performs a size calculation, and Step 4 performs a continuity calculation for the test
statistics.
Damian Kozbur/Testing-Based Forward Model Selection
58
Step 2.
This section calculates power properties. Suppose that
2 2
tα .
−∆j Ex (S) > AjS VbjS (cτ + 1)2 τbjS
Then on A, and for |S| < Kn ,
∗|x
2 2
tα
[θjS ]2j AjS > AjS VbjS (cτ + 1)2 τbjS
∗|x
1/2
|[θjS ]j | > VbjS (cτ + 1)b
τjS tα
1/2
∗|x
|[θbjS ]j | > VbjS (cτ + 1)b
τjS tα − |[θjS ]j − [θbjS ]j |
1
+
c
1/2
1/2
τ
b
|[θjS ]j | > VbjS (cτ + 1)b
τjS tα − VbjS
τbjS tα
2
1/2
|[θbjS ]j | > Vb cτ τbjS tα
jS
which implies TjSα = 1.
Step 3.
−1/2
This step performs a size calculation. By construction, if TjSα = 1 then |VbjS [θbjS ]j | >
cτ τbjS tα , which is equivalent to
1/2
|[θbjS ]j | > cτ τbjS tα VbjS .
On A, for |S| < Kn , note that
∗|x
1/2
|[θbjS ]j − [θjS ]j | 6 VbjS
1 + cτ
2
τbjS tα .
Then TjSα = 1 ⇒
∗|x
|[θjS ]j |
1/2
1/2 1 + cτ
b
b
>cτ τbjS tα VjS − VjS
τbjS tα
2
cτ − 1
1/2
= VbjS τbjS tα
2
Then
2 2
−∆j Ex (S) > AjS VbjS τbjS
tα
cτ − 1
2
2
.
Damian Kozbur/Testing-Based Forward Model Selection
59
Step 4.
This section calculates continuity properties of the tests. Suppose that TjSα = TkSα = 1,
−1/2 b
−1/2 b
and that WjS 6 WkS . Note that WjS 6 WkS implies Vb
|[θjS ]j | 6 Vb
|[θkS ]k |. Then
jS
kS
on A, and for |S| < Kn ,
−1/2
∗|x
VbjS |[θjS ]j | −
1 + cτ
2
−1/2
∗|x
τbjS tα 6 VbkS |[θkS ]k | +
−1/2
∗|x
−1/2
∗|x
VbjS |[θjS ]j | 6 VbkS |[θkS ]k | + 2
1 + cτ
2
1 + cτ
2
τbkS tα
τbkS tα
−1/2 −1/2
−1/2 −1/2
VbjS AjS (−∆j Ex (S))1/2 6 VbkS AkS (−∆k Ex (S))1/2 + 2
1 + cτ
2
τbkS tα
τ 2 2
τbkS t2α derived in
Using the fact that −∆k Ex (S) > AkS VbkS 1−c
2
Step 2 and the fact that
−1/2
−1/2
1+cτ
1+cτ
1/2
b
TkSα = 1, gives that
τbkS tα 6 V
A
(−∆k Ex (S))
.
kS
2
cτ −1
kS
Then
−1/2 −1/2
VbjS AjS (−∆j Ex (S))1/2
6
−1/2 −1/2
VbkS AkS (−∆k Ex (S))1/2
+
−1/2 −1/2
2VbkS AkS (−∆k Ex (S))1/2
This gives
VbjS AjS
1 + cτ 2
−∆j Ex (S) 6
1+2
(−∆k Ex (S))
cτ − 1
VbkS AkS
Step 5
This step summarizes Steps 2-4 and outlines the next steps. By Steps 2-4, the following
implications are valid on A for all j, |S| < Kn :
2 t2 .
1. TjSα = 1 if − ∆j Ex (S) > AjS VbjS (cτ + 1)2 τbjS
α
1−cτ 2 2 2
τbjS tα
2
2. −∆j Ex (S) > AjS VbjS
3. −∆j Ex (S) 6
VbjS AjS
VbkS AkS
τ
1 + 2 1+c
cτ −1
2
if TjSα = 1.
(−∆k Ex (S)) if TjSα = TkSα = 1, WjS 6 WkS .
Next define a sequence of sets X = Xn which will be shown to have the property that
both P(x ∈ X) → 1 and
1 + cτ
.
cτ − 1
Damian Kozbur/Testing-Based Forward Model Selection
60
PX (A) = inf P(A|x) → 1.
x∈X
In addition, there will be constants c̃test , c̃0test , c00test > 0 which are independent of n and
the realization of x, such that for ctest = n1 c̃test , c0test = n1 c̃0test and for the set B defined by
2 t2 6 c
1. AjS VbjS (cτ + 1)2 τbjS
test
α
τ 2 2 2
B=
τbjS tα > c0test
|S| < Kn
2. AjS VbjS 1−c
2
2
3. AjS VbjS 1 + 2 cτ −1 > c00test
b
cτ +1
AkS VkS
it holds that PX (B) → 1.
Define sets X = Xn as follows.
X = X1 ∩ X2 ∩ X3 ∩ X4
X1 = {x : maxj6p En [x12
ij ] = O(1)}
X2 = {x : ϕmin (Kn )(G)−1 = O(1)}
X3 = {x : kηjS k1 = O(1)}
X4 = {x : P(ϕmin (Kn )(En [ε2i xi x0i ])−1 = O(1)|x) = 1 − o(1)}
Note that P(X1 ), P(X2 ), P(X3 )
P(X4 )
→
→
1 by assumption. In addition, failure of
1 would contradict the unconditional statement in Condition 5 that
P(ϕmin (Kn )(En [ε2i xi x0i ])−1 = O(1)) = 1 − o(1). Therefore, P(X) → 1.
Next, Step 6 proves that PX (A) → 1 and Step 7 proves that PX (B) → 1. Note that these
steps allow the aplication of Theorem 4 conditionally on x. Step 8 uses this fact, and then
∗|x
concludes the proof by showing that θ0 is bounded to θS0 .
Damian Kozbur/Testing-Based Forward Model Selection
61
Step 6.
Note that
bjS = Vb −1/2 ([θbjS ]j − [θ∗|x ]j )
Z
jS
jS
−1/2
= VbjS (x0j QS xj )−1 x0j QS (ε − E[ε|x])
= ((x0j QS xj )−1 En [b
ε2ijS [QS xjS ]2i ](x0j QS xj )−1 )−1 (x0j QS xj )−1 x0j QS (ε − E[ε|x])
= En [b
ε2ijS [QS xjS ]2i ]−1 x0j QS (ε − E[ε|x])
0
0
= En [b
ε2ijS (ηjS
xijS )2 ]−1 ηjS
xjS (ε − E[ε|x]).
0
0
= En [b
ε2ijS (ηjS
xijS )2 ]−1 ηjS
xjS (εo + εa − E[εa |x]).
Let ε̈ = εo + εa − E[εa |x].
Define the Regularization Event by
| Pn x ε̈ |
ik i
6 tα for every k 6 p
R = qPi=1
n
x2 ε̈2
i=1
ik i
In addition, define the Variability Domination Event by
V=
( n
X
x2ik ε̈2i 6
i=1
1 + cτ
2
2 X
n
)
x2ik εb2ijS for every k ∈ jS, for every |S| < Kn
i=1
The definition of the Regularization Event and the Variability Domination Event are
useful since
R ∩ V ⇒ A.
To see this, note that on R, the following inequality holds for any conformable vector ν:
n X
X
i=1 k∈jS
2
νk xik ε̈i 6 tα
X
k∈jS
v
2
u n
uX
|νk |t
x2ik ε̈2i
i=1
Damian Kozbur/Testing-Based Forward Model Selection
62
Furthermore, on V, the previous expression can be further bounded by
6
1 + cτ
2
2
X
tα
v
2
u n
uX
|νk |t
x2ik εb2ijS
k∈jS
i=1
2
qP
2
n
2 ε
2
2 tα
n
x
b
|ν
|
X
X
i=1 ik ijS
k∈jS k
1 + cτ
=
νk xik εb2ijS
2
P n P
2
2
i=1
εbijS
k∈jS
i=1
k∈jS νk xik
2
0
εb 1/2 2 n
1 + cτ 2 2 kν Diag(ΨjS ) k1 X X
=
tα
νk xik εb2ijS .
b ν
2
ν 0 ΨεjS
i=1
k∈jS
P
Specializing to the case that ν = ηjS , and using τbjS =
bjS | 6
|Z
b )1/2 k
kν 0 Diag(ΨεjS
1
q
b ν
0
ν ΨεjS
gives that
1 + cτ
τbjS tα on R ∩ V.
2
It is therefore sufficient to prove that R and V have probability → 1 under PX . PX (R) → 1
follows immediately from the moderate deviation bounds for self-normalized sums given in
[38]. For details on the application of this result, see [6].
∗|x
Therefore, it is only left to show that PX (V) → 1. Define εijS = yi −x0ijS θjS . Furthermore,
define ξijS through the decomposition εijS = ε̈i + ξijS . Let εjS and ξjS be the respective
stacked versions. Let c̃τ = ((1 + cτ )/2)2 .
Then
c̃τ
n
X
x2ik εb2ijS = c̃τ
" n
X
i=1
> c̃τ
x2ik (b
ε2ijS − ε2ijS ) +
n
X
i=1
i=1
i=1
" n
X
n
X
n
X
x2ik (b
ε2ijS − ε2ijS ) +
i=1
=
x2ik ε̈2i + 2
n
X
n
X
x2ik ε̈2i + 2
i=1
x2ik ε̈i ξijS +
i=1
#
x2ik ε̈i ξijS
i=1
x2ik ε̈2i
i=1
+ c̃τ
n
X
n
x2ik (b
ε2ijS − ε2ijS ) +
i=1
n
X
+ 2c̃τ
i=1
(c̃τ − 1) X 2 2
xik ε̈i
2
i=1
x2ik ε̈i ξijS +
(c̃τ − 1)
2
n
X
i=1
x2ik ε̈2i .
n
X
#
2
x2ik ξijS
Damian Kozbur/Testing-Based Forward Model Selection
63
Define the two events
(c̃τ − 1)
0
2
2
2
2 2
V = c̃τ En [xik (b
εijS − εijS )] +
En [xik ε̈i ] > 0 for all j, k 6 p, |S| < Kn
2
(c̃τ − 1)
V00 = 2c̃τ En [x2ik ε̈i ξijS ] +
En [x2ik ε̈2i ] > 0 for all j, k 6 p, |S| < Kn
2
Therefore V0 ∩ V00 ⇒ V.
Note
that
En [x2ik ε̈2i ]
>
1
2 2
2 En [xik εi ]
− En [x2ik E[εai |x]]
>
1
2 2
2 En [xik εi ]
−
1/2 E [x4 ]1/2 . This is bounded below with PX → 1 by a positive conmaxi6n E[εa2
n ik
i |x]
stant independent of n. Therefore, to show that PX (V0 ) → 1, PX (V00 ) → 1, it suffices to
ε2ijS − ε2ijS )] and En [x2ik ε̈i ξijS ] respectively are suitably smaller order.
show En [x2ik (b
ε2ijS − ε2ijS )]. It is convenient to bound the slightly more general
First consider En [x2ik (b
sum En [xik xil (b
ε2ijS − ε2ijS )], since this will show up again in Step 8.
i
h
i
h
∗|x
∗|x
En [xik xil (b
ε2ijS − ε2ijS )] = 2En xik xil εijS x0ijS (θjS − θbjS ) + En xik xil (x0ijS (θjS − θbjS ))2
∗|x
∗|x
6 2kEn [xik xil εijS x0ijS ]k2 kθjS − θbjS k2 + λmax En [xik xil xijS x0ijS ]kθjS − θbjS )k22
√
∗|x
Standard reasoning gives that kθjS − θbjS k2 6 ϕmin (Kn )(G)−1/2 Kn kEn xijS εijS k∞ . There-
fore, the bound continues
6 2kEn [xik xil εijS x0ijS ]k2 ϕmin (Kn )(G)−1/2
p
Kn kEn xijS εijS k∞
+λmax En [xik xil xijS x0ijS ]ϕmin (Kn )(G)−1 Kn kEn xijS εijS k2∞ .
Note that λmax En [xik xil xijS x0ijS ] 6 Kn maxj6p En [x4ij ].
6 2kEn [xik xil εijS x0ijS ]k2 ϕmin (Kn )(G)−1/2
p
Kn kEn xijS εijS k∞
+Kn2 max En [x4ij ]ϕmin (Kn )(G)−1 kEn xijS εijS k2∞
j6p
An application of Cauchy-Schwarz to the top line gives
62
p
p
Kn max En [x4ik ]1/2 max En [ε2ijS x2ij ]1/2 ϕmin (Kn )(G)−1/2 Kn kEn xijS εijS k∞
j
j,S
+Kn2 max En [x4ij ]ϕmin (Kn )(G)−1 kEn xijS εijS k2∞
j6p
Next, kEn xijS εijS k∞ and En [ε2ijS x2ij ]1/2 are bounded using εijS = εi − E[εi |x] + ξijS . Note
that by construction kEn [xijS ξijS ]k∞ = 0. Then
kEn [xijS εijS ]k∞ 6 kEn [xi εi ]k∞ + kEn [xi E[εai |x]k∞
Damian Kozbur/Testing-Based Forward Model Selection
64
p
6 kEn [xi εi ]k∞ + max En [x2ij ]1/2 En [E[εai |x]2 ]1/2 = O( log p/n)
j6p
with PX → 1. Next,
2
2
2
En [ε2ijS x2ij ] 6 3En [ε2i x2ij ] + 3En [E[εa2
i |x]xij ] + 3En [ξijS xij ]
4 1/2
6 3En [ε2i x2ij ] + 3En [x2ij ] max E[εa2
En [x4ij ]1/2
i |x] + 3En [ξijS ]
i6n
1/2
P0
4 ]
QjS xl θ0,l =
Next, En [ξijS
6 O(1)s20 on X1 ∩X3 . To see this, note ξjS = QjS xθ0 = sl=1
Ps0
l=1 ηl,(jS) xljS = η̃jS xS0 ∪jS for some new linear combination η̃jS . Note that kη̃jS k1 6
1/4
4 ]
s0 O(1). Then En [ξijS
6 kη̃jS k1 maxk6p En [x4ik ]1/4 from which the bound follows.
Next consider En [x2ik ε̈i ξijS ]. Consider two cases. In Case 1,
2 1/2
] 6
En [x4ik ξijS
1 (c̃τ − 1) En [x2ik ε̈2i ]
.
2c̃τ
2
En [ε̈2i ]1/2
2 ]1/2 E [ε̈2 ]1/2 6
In this case, 2c̃τ En [x2ik ε̈i ξijS ] 6 En [x4ik ξijS
n i
c̃τ −1
2 ,
and the requirement of V00
for k, j, S holds.
2 2
1 (c̃τ −1) En [xik ε̈i ]
holds. Then
2c̃τ
2
En [ε̈2i ]1/2
2
2
2
4
E[En [xik ξijS ε̈i ]|x] is bounded away from zero by conditions on E[εi |x] and maxi |εai |. In
addition, E[En [|xik |6 |ξijS |3 |ε̈i |3 ]|x] 6 maxi E[|ε¨i |3 |x]En [|xik |6 |ξijS |3 ] 6 O(1)En [|xik |6 |ξijS |3 ].
For Case 2, suppose the alternative that En [x4ik x2ijS ] >
This term is further bounded by
1/2
En [|ξijS |6 ]1/2
O(1)En [x12
ik ]
4 ] earlier, it follows that E [|ξ
6 1/2 =
Using the same reasoning as bounding En [ξijS
n ijS | ]
O(1)s30 . In addition, En [x12
ik ] = O(1). As a result, for those k, j, S which fall in Case 2,
the self-normalized sum
√
n|En [x2ik ξijS ε̈i ]|
q
j,k,S∈Case 2
2 ε̈2 ]
En [x4ik ξijS
i
p
is O(log(pKn )) with probability 1 − o(1) provided
log(pKn )
=
max
=
o(n1/6 /(s30 )1/3 ).
2
This holds under the assumed rate
q conditions. Then maxj,k,S |En [xik ξijS ε̈i ]| is
2 ε̈2 ]. Furthermore, E [x4 ξ 2 ε̈2 ]
bounded by √1n O(log(pKn ) maxj,k,S En [x4ik ξijS
6
n ik ijS i
i
4 ]1/2 E [ε̈4 ]1/2 6 (E [x12 ]2/3 E [ξ 12 ]1/3 )1/2 E [ε̈4 ]1/2 6 O(1)s2 E [ε̈4 ]1/2 . Note that
En [x8ik ξijS
n i
n ik
n ijS
n i
0 n i
En [ε̈4i ]1/2 6 O(1) with PX → 1. Together, these give that maxj,k,S En [x2ik ε̈i ξijS ] = o(1) with
PX → 1. Finally, PX (V) → 1.
Damian Kozbur/Testing-Based Forward Model Selection
65
Step 7.
This step shows that PX (B) → 1. First, AjS depend only on x and are bounded above and
below by constants which do not depend on n on X from the assumption on the sparse
eigenvalues of G. For bounding τbjS above and away from zero, since 1 6 kηjS k1 , kηjS k2 6
b = E [x
0 b2 ] remain
O(1) on X, it is sufficient to show that the eigenvalues of ΨεjS
n ijS xijS ε
ijS
b remain bounded
bounded above and away from zero and that the diagonal terms of ΨεjS
above and away from zero. Note that by arguments in last step (Step 6), it was shown
p
that En [xik xil (b
εijS − εijS )] = O( log p/n) with PX → 1. Therefore, kEn [xijS x0ijS εb2ijS ] −
p
En [xijS x0ijS ε2ijS ]kF = O(Kn log p/n ) with PX → 1. Here, F is the Frobenius norm. By the
assumed rate condition, the above quantity therefore vanishes with PX → 1.
Next,
En [xijS x0ijS ε2ijS ] = En [xijS x0ijS ε2i ] + 2En [xijS x0ijS εi (ξijS + E[εai |x])]
+En [xijS x0ijS (ξijS + E[εai |x])2 ]
The first term above, En [xijS x0ijS ε2i ], has eigenvalues bounded away from zero for all j, S
with PX → 1. The third term above, En [xijS x0ijS (ξijS + E[εai |x])2 ] is positive semidefinite
by construction. The second term above has Frobenius norm tending to zero for all j, S
with PX → 1. This, in conjunction with the fact that the eigenvalues of En [xijS x0ijS εbijS ]
are bounded above and away from zero with PX → 1 shows that the eigenvalues of
b = E [x
0 b2 ] are bounded above and away from zero with PX → 1. Finally,
ΨεjS
n ijS xijS ε
ijS
for bounding VbjS , it is sufficient to show that maxk6p En [ε2 (η 0 xijS )2 ] be bounded above.
i
This follows immediately from
and
maxk6p En [x4ik ]
E[ε4i |x]
jS
being uniformly bounded and maxj,S kηjS k1 = O(1)
= O(1). These imply that PX (B) → 1.
Step 8
The previous steps show that Theorem 4 can be applied conditionally on x. (Note that
renormalizing the covariates to satisfy En [x2ij ] = 1 does not affect Ex (S) and therefore does
not affect the conclusion of Step 5.) Therefore,
p
|x
b 2 ]1/2 6 O( s0 log p/n) → 1.
PX En [(x0i θ0 − xi θ)
To show Theorem 5, it is left to show that
p
|x
PX En [(x0i θ0 − xi θ0 )2 ]1/2 6 O( s0 log p/n) → 1.
Damian Kozbur/Testing-Based Forward Model Selection
66
This follows from from assumptions on εa Note that
|x
θ0 − θ0 = (x0S0 xS0 )−1 x0S0 E[εa |x].
As a result,
|x
kθ0 − θ0 k2 6 ϕmin (s0 )(G)−1/2 kEn [xis0 E[εai |x]]k2
√
6 ϕmin (s0 )(G)−1/2 s0 kEn [xij E[εai |x]]k∞ .
By the assumed rate conditions and by maxi E[εai ] = O(n−1/2 ), the bound follows.
Appendix E: Proof of Theorems 7 and 8
Proof. Theorem 7 follows by applying Theorem 4 in [6]. Theorem 8 follows by applying
Theorem 2 of [14].
Appendix F: Additional computation details
Computation of the simulation studies and empirical example are performed using the
software Matlab R2015a. The simulation study uses in addition the package Glmnet for
Matlab ([55]). Replication files as well as detailed instructions are available from the author.
| 10math.ST
|
1
A Generalized Framework for Kullback-Leibler
Markov Aggregation
Abstract—This paper proposes an information-theoretic cost
function for aggregating a Markov chain via a (possibly stochastic) mapping. The cost function is motivated by two objectives:
1) The process obtained by observing the Markov chain through
the mapping should be close to a Markov chain, and 2) the
aggregated Markov chain should retain as much of the temporal
dependence structure of the original Markov chain as possible.
We discuss properties of this parameterized cost function and
show that it contains the cost functions previously proposed by
Deng et al., Xu et al., and Geiger et al. as special cases. We
moreover discuss these special cases providing a better understanding and highlighting potential shortcomings: For example,
the cost function proposed by Geiger et al. is tightly connected
to approximate probabilistic bisimulation, but leads to trivial
solutions if optimized without regularization. We furthermore
propose a simple heuristic to optimize our cost function for
deterministic aggregations and illustrate its performance on a
set of synthetic examples.
Index Terms—Markov chain, lumpability, predictability, bisimulation, model reduction
I. I NTRODUCTION
Markov aggregation is the task of representing a Markov
chain with a large alphabet by a Markov chain with a smaller
alphabet, thus reducing model complexity while at the same
time retaining the computationally and analytically desirable
Markov property (see Fig. 1). Such a model reduction is
necessary if the original Markov chain is too large to admit
simulation, estimating model parameters from data, or control
(in the case of Markov decision processes). These situations
occur often in computational chemistry (where aggregation is
called coarse-graining, e.g., [1]), natural language processing,
and the simulation and control of large systems (giving rise
to the notion of bisimulation, e.g., [2]). Additionally, Markov
aggregation can be used as a tool in exploratory data analysis,
either to discover groups of “similar” states of a stochastic
process or to cluster data points, cf. [3], [4].
Information-theoretic cost functions were proposed for
Markov aggregation in [5]–[8]. Specifically, the authors of [5]
proposed a cost function linked to the predictability of the
aggregated Markov chain. Such an approach is justified if the
original model is nearly completely decomposable, i.e., if there
is a partition of the alphabet such that transitions within each
element of the partition occur quickly and randomly, while
transitions between elements of the partition occur only rarely.
Building on this work, the authors of [7] proposed a cost
function linked to lumpability, i.e., to the phenomenon where
a function of a Markov chain is Markov. Such an approach
X ∼ Mar(X , P)
Observation pY |X
arXiv:1709.05907v1 [cs.SY] 18 Sep 2017
Rana Ali Amjad, Clemens Blöchl, Bernhard C. Geiger
Institute for Communications Engineering, Technical University of Munich, Germany
[email protected], [email protected]
Ag
gr
eg
a
tio
n
D̄(Y||Ỹ)
Y
Ỹ ∼ Mar(Y, Q)
Fig. 1. Illustration of the aggregation problem: A stationary first-order Markov
chain X is given. We are interested in finding a conditional distribution pY |X
and an aggregation of X, i.e., a Markov chain Ỹ on Y. The conditional
distribution pY |X defines a stationary process Y, a noisy observation of X.
Y might not be Markov of any order, but can be approximated by a Markov
chain Ỹ.
is justified whenever there are groups of states with similar
probabilistic properties (in a well-defined sense). Both [5]
and [7] focus on deterministic aggregations, i.e., every state
of the original alphabet is mapped to exactly one state of the
reduced alphabet. Moreover, the authors of both references
arrive at their cost functions by lifting the aggregated Markov
chain to the original alphabet. The authors of [6] present an
information-theoretic cost function for stochastic aggregations,
but they do not justify their choice by an operational characterization (such as predictability or lumpability). Instead, they
arrive at their cost function via the composite of the original
and the aggregated Markov chain.
In this paper, we extend the works [5]–[8] as follows:
1) We present a two-step approach to Markov aggregation (Section III): Observing the original Markov chain
through a (stochastic or deterministic) channel, and
then approximating this (typically non-Markov) process
as a Markov chain (see Fig. 1). This approach has
already been taken by [7], albeit only for deterministic
aggregations.
2) Using this two-step approach, we propose a parameterized, information-theoretic cost function for Markov
aggregation (Section IV). We arrive at this cost function
neither via lifting nor via the composite model, but via
requiring specific operational qualities of the process
observed through the channel: It should be close to a
Markov chain and it should retain the temporal dependence structure of the original Markov chain.
2
3) We show that our cost function contains the cost functions of [5]–[8] as special cases (Section V). We also
discuss previous algorithmic approaches to the Markov
aggregation problem.
4) We propose a simple, low-complexity heuristic to minimize our generalized cost function for deterministic
aggregations (Section VI).
5) As a side result, we justify the cost function proposed
in [7] by showing a tight connection to approximate
probabilistic bisimulation (Section III-A).
We illustrate our cost function for various examples in
Section VII. Specifically, we investigate the aggregation of
quasi-lumpable and nearly completely decomposable Markov
chains, and we look at a toy example from natural language
processing. We also take up the approach of [3], [4] to perform
clustering via Markov aggregation. In future work, we shall
extend our efforts to Markov decision processes, and provide
a theory for lifting stochastic aggregations as indicated in [6,
Remark 3].
II. N OTATION AND D EFINITIONS
1
(1)
The Markov chain is time-homogeneous if, the right-hand side
of (1) does not depend on n, i.e., if
pZn |Zn−1 (zn |zn−1 ) = pZ2 |Z1 (zn |zn−1 ) := Pzn−1 →zn .
z∈Z
H(Z2 |Z1 ) :=
X
H(Z2 |Z1 = z)pZ1 (z)
I(Z1 ; Z2 ) := H(Z2 ) − H(Z2 |Z1 ).
(3c)
The entropy rate and redundancy rate of a stationary stochastic
process Z (not necessarily Markov) are
H(Z1n )
= lim H(Zn |Z1n−1 )
n→∞
n→∞
n
R̄(Z) := lim I(Zn ; Z1n−1 ) = H(Z) − H̄(Z).
H̄(Z) := lim
n→∞
(3d)
(3e)
The Kullback-Leibler divergence rate (KLDR) between two
stationary stochastic processes Z and Z′ on the same finite
alphabet Z is [10, Ch. 10]
pZ ′n1 (z1n )
1 X
pZ ′n1 (z1n ) log
n→∞ n
pZ1n (z1n )
n
n
(3f)
z1 ∈Z
provided the limit exists. If the limit exists, it is finite if, for
all n and all z1n , pZ1n (z1n ) = 0 implies pZ ′ n1 (z1n ) = 0 (short:
pZ ′ n1 ≪ pZ1n ). In particular, if Z′ ∼ Mar(Z, P′ ) and Z ∼
Mar(Z, P), then [11]
D̄(Z′ ||Z) =
X
′
µz Pz→z
′ log
z,z ′ ∈Z
′
Pz→z
′
.
Pz→z′
(3g)
provided P ≪ P′ .
These information-theoretic quantities can be used to give
an equivalent definition of Markovity:
Lemma 1 ([12, Prop. 3]). Suppose the stochastic process Z
is stationary. Then, Z is Markov iff H̄(Z) = H(Z2 |Z1 ).
If Z is a stationary process on Z (not necessarily Markov),
then one can approximate this process by a Markov chain
Z̃ ∼ Mar(Z, P):
Lemma 2 ([10, Cor. 10.4]). Let Z be a stationary process
on Z, and let Z′ ∼ Mar(Z, P′ ) be any Markov chain on Z.
Then,
P = arg min D̄(Z||Z′ )
(2)
If the transition probability matrix P = [Pzn−1 →zn ] of a
time-homogeneous Markov chain is irreducible and aperiodic
(see [9] for terminology), then there exists a unique vector µ
such that µT = µT P, which represents the invariant distribution of Z. If the initial distribution pZ1 coincides with µ,
then Z is stationary and we denote this stationary, irreducible
and aperiodic first-order Markov chain by Z ∼ Mar(Z, P).
In this work we deal exclusively with first-order stationary,
irreducible and aperiodic Markov chains.
We use information-theoretic cost functions for the aggregation problem. The entropy of Z, the conditional entropy of
(3b)
z∈Z
D̄(Z′ ||Z) := lim
We denote vectors and matrices by bold lower case and
blackboard bold upper case letters, e.g., a and A. A diagonal
matrix with vector a on the main diagonal is denoted by
diag(a). The transpose of A is AT .
We denote random variables (RVs) by upper case letters,
e.g., Z, and their alphabet by calligraphic letters, e.g., Z.
In this work we will restrict ourselves to RVs with finite
alphabets, i.e., |Z| < ∞. Realizations are denoted by lower
case letters, e.g., z, where z ∈ Z. The probability mass
function (PMF) of Z is pZ , where pZ (z) := Pr(Z = z) for all
z ∈ Z. Joint and conditional PMFs are defined accordingly.
We denote a one-sided, discrete-time, stochastic process
with Z := (Z1 , Z2 , . . . ), where each Zk takes values from the
n
:= (Zm , Zm+1 , Zn ).
(finite) alphabet Z. We abbreviate Zm
We consider only stationary processes, i.e., PMFs are invariant
w.r.t. a time shift. In particular, the marginal distribution of Zk
is equal for all k and shall be denoted as pZ .
A first-order Markov chain is a process that satisfies, for all
n > 1 and all z1n ∈ Z n ,
pZn |Z n−1 (zn |z1n−1 ) = pZn |Zn−1 (zn |zn−1 ).
Z2 given Z1 , and the mutual information between Z1 and Z2
are defined by
X
H(Z) := −
pZ (z) log pZ (z)
(3a)
(4a)
P′
where
Pz→z′ = pZ2 |Z1 (z ′ |z).
(4b)
Moreover, for Z̃ ∼ Mar(Z, P),
D̄(Z||Z̃) = H(Z2 |Z1 ) − H̄(Z).
(4c)
By Lemma 1 we know that right-hand side of (4c) is 0 iff
Z is Markov. Hence, one can view the KLDR D̄(Z||Z̃) as a
measure of how close a process Z is to a Markov chain.
3
III. M ARKOV C HAIN AGGREGATION
Given a Markov chain X, Markov aggregation deals with
the problem of finding a Markov chain Ỹ on a given smaller
alphabet Y which is the optimal representation of X in the
¯
sense of minimizing a given cost function C(X,
Ỹ). This is
depicted in Fig. 1 by the diagonal arrow and is summarized
in the following definition:
Definition 1 (Markov Aggregation Problem). Let X ∼
¯ ·) be given.
Mar(X , P), Y, and an arbitrary cost function C(·,
The Markov aggregation problem concerns finding a minimizer of
¯
Ỹ)
(5)
min C(X,
Ỹ
where the optimization is over Markov chains on Y.
In this work we address the Markov aggregation problem
using the two-step approach depicted in Fig 1. The first step is
to use a (possibly stochastic) mapping from X to Y. Applying
this mapping to X leads to a stationary process Y which may
not be Markov (in fact, Y is a hidden Markov process). In
the second step we look for the optimal approximation Ỹ of
Y in the sense of Lemma 2.
This two-step approach is a popular method of Markov
aggregation and has been employed in various works including
[5], [7], [8]. In these references, the mapping in the first
step was restricted to be deterministic whereas in this work
we allow it to be stochastic. In other words, while these
references were looking for a partition of X induced by
a function g: X → Y, in this work we permit stochastic
mappings induced by a conditional distribution pY |X . We
represent pY |X as a row stochastic matrix W = [Wx→y ],
where Wx→y = pY |X (y|x).
With this notation, the following corollary to Lemma 2
solves the second of the two steps in our approach, i.e.,
it characterizes the optimal approximation Ỹ of the hidden
Markov process Y that we obtain by observing X through
W:
Corollary 1. Let X ∼ Mar(X , P) and let W denote a
conditional distribution from X to Y. Let Y be the hidden
Markov process obtained by observing X through W, and let
Ỹ ∼ Mar(Y, Q) be its best Markov approximation in the
sense of minimizing D̄(Y||Ỹ) (cf. Lemma 2). Then,
Q = UPW
(6)
where U := diag(ν)−1 WT diag(µ) with ν T := µT W being
the marginal distribution of Yk .
Note that this corollary extends [7, Lem. 3] from deterministic to stochastic mappings.
With the second step solved, the two-step approach to the
optimization problem stated in Definition 1 boils down to
optimization over the mapping W. We can thus restate the
Markov aggregation problem as follows:
Definition 2 (Markov Aggregation Problem Restated). Let
¯ ·) be
X ∼ Mar(X , P), Y, and an arbitrary cost function C(·,
given. Let
¯
C(X, W) = C(X,
Ỹ)
(7)
where Ỹ is the Markov approximation of the hidden Markov
process Y that is obtained by observing X through the
stochastic mapping W. The Markov aggregation problem
using the two-step approach concerns finding a minimizer of
min C(X, W)
W
(8)
where the optimization is over stochastic mappings from X to
Y. If the optimization is restricted over deterministic mappings
g, we abuse notation and write C(X, g) for the cost.
Note that Definition 1 and Definition 2 are not equivalent
in general, i.e., the optimal aggregated chain Ỹ obtained
by solving (5) is not the same as the optimal aggregated
chain Ỹ obtained by solving (8). The two formulations only
become equivalent when we restrict the optimization in (5) to
aggregated Markov chains which can be obtained as a result
of the aforementioned two-step approach.
A. Markov Aggregation via Lumpability
¯
The optimal mapping W depends on the cost function C.
One possible choice in the light of Lemma 2 is
¯
C(X,
Ỹ) = D̄(Y||Ỹ).
(9)
In other words, we wish to find a mapping W such that the
hidden Markov process Y is as close to a Markov chain
as possible in an information-theoretic sense. This may be
reasonable since it states that data obtained by simulating the
aggregated model Ỹ differs not too much from data obtained
by simulating the original model in conjunction with the
stochastic mapping, i.e., data obtained from Y.
There are two shortcomings of the cost (9). First, (9) focuses
only on getting Y close to Ỹ but not on preserving any form
of information in X. This gives rise to trivial optimal solutions:
If W is such that the conditional distribution does not depend
on the conditioning event, i.e., pY |X = pY , or W = 1αT
for some probability vector α, then Y is independent and
identically distributed (i.i.d.) and hence Markov. Indeed, in this
case H̄(Y) = H(Y2 |Y1 ) = H(Y ), from which D̄(Y||Ỹ) = 0
follows. The cost function is thus inappropriate for Markov
aggregation, unless it is regularized appropriately. The second
shortcoming is linked to the fact that D̄(Y||Ỹ) requires,
by (4c), the computation of the entropy rate H̄(Y) of a hidden
Markov process. This problem is inherently difficult [13],
and analytic expressions do not even exist for simple cases
(cf. [14]). In the following, we discuss two previously proposed relaxations of the Markov aggregation problem for
¯
C(X,
Ỹ) = D̄(Y||Ỹ).
The authors of [7] addressed the second shortcoming by
relaxing the cost via
CL (X, W) := H(Y2 |Y1 ) − H(Y2 |X1 ) ≥ D̄(Y||Ỹ).
(10)
This cost does not require computing H̄(Y) and is linked to
the phenomenon of lumpability, the fact that a function of
a Markov chain has the Markov property [12, Thm. 9]: If
CL (X, W) = 0, then Y is a Markov chain.
We now show that, at least for deterministic mappings, this
cost function also has a justification in approximate probabilistic bisimulations, or ε-bisimulations. More specifically, the
4
authors of [15] discussed bisimilarity of Markov processes and
showed that two Markov chains are bisimilar if one can be
described as a function of the other (see discussion after [15,
Def. 5.2]). In other words, if X ∼ Mar(X , P) is a Markov
chain, g: X → Y a surjective function, and Y ∼ Mar(Y, Q)
satisfies Yk = g(Xk ), then X and Y are bisimilar. Since
this is equivalent to lumpability, bisimilarity is implied by
CL (X, g) = 0.
Extending this line of reasoning, we give a justification of
the cost function CL (X, g) in terms of ε-bisimulation of pairs
of Markov chains, even in case X is not lumpable w.r.t. g. To
this end, we adapt [16, Def. 4 & 5] for our purposes:
Definition 3 (ε-Bisimulation). Consider two finite Markov
chains X ∼ Mar(X , P) and Ỹ ∼ Mar(Y, Q) and assume
w.l.o.g. that X and Y are disjoint. We say that X and Ỹ are
ε-bisimilar if there exists a relation Rε ⊆ (X ∪ Y) × (X ∪ Y)
such that for all x ∈ X and y ∈ Y for which (y, x) ∈ Rε ,
and all T ⊆ X ∪ Y we have
X
X
Px→x′ ≥
(11)
Qy→y′ − ε
x′ ∈Rε (T )∩X
y ′ ∈T ∩Y
where Rε (T ) := {s2 ∈ X ∪ Y: s1 ∈ T, (s1 , s2 ) ∈ Rε }.
The definitions of ε-bisimulations are typically given for
labeled [16, Def. 4 & 5] or controlled [2, Def. 4.4] Markov
processes with general alphabets and thus contain more restrictive conditions than our Definition 3. Our definition is
equivalent if the alphabets are finite and if the set of labels is
empty. We are now ready to state
Proposition 1. Let X ∼ Mar(X , P) and the surjective
function g: X → Y be given. Let Q be as in Corollary 1,
where Wx→y = 1 iff y = g(x). Let Ỹ ∼ Mar(Y, Q). Then,
X and Ỹ are ε-bisimilar with
s
ln(2)CL (X, g)
ε=
.
(12)
2 minx∈X µx
Proof: See Section VIII-A.
Despite this justification, the cost function CL (X, W) is
mainly of theoretical interest. The reason is that the shortcoming of leading to trivial solutions is inherited by CL (X, W),
since for W = 1αT one gets CL (X, W) = 0, regardless of α
and P. Even restricting W to be a deterministic partition, as
considered in [7], does not solve this problem: The combinatorial search over all partitions may have its global optimum at a
partition that makes Y close to an i.i.d. process. Indeed, if the
cardinality of Y is not constrained (or if g is not required to be
surjective), then the constant function g yields CL (X, g) = 0.
B. Markov Aggregation by Predictability
A different approach was taken by the authors of [5] who
proposed the following cost function (again with the focus on
deterministic partitions):
CP (X, W) := I(X1 ; X2 ) − I(Y1 ; Y2 )
(13)
The computation of CP (X, W) is simple as it does not require
computing H̄(Y). Furthermore, CP (X, W) reflects the wish
to preserve the temporal dependence structure of X, i.e., it is
connected to predicting future states of Y based on knowledge
of past states of Y. Since X is not i.i.d., observing Xk reveals
some information about Xk+1 . Minimizing CP (X, W) thus
tries to find a W such that Yk reveals as much information
about Yk+1 as possible, and hence does not lead to the
same trivial solutions as CL (X, W) and D̄(Y||Ỹ): A constant
function g or a soft partition W = 1αT render Y1 and Y2
independent, hence the cost is maximized at CP (X, W) =
I(X1 ; X2 ). Unfortunately, as it was shown in [7, Thm. 1],
we have CP (X, W) ≥ CL (X, W), i.e., (13) does not capture
Markovity of Y as well as the relaxation proposed by [7].
Note that although CP (X, W) does not lead to the same
trivial solutions as CL (X, W) and D̄(Y||Ỹ), it still tries to
preserve only part of the information about temporal dependence in X, i.e., information which is helpful in predicting the
next sample. Such a goal is justified in scenarios in which X
is quasi-static, i.e., runs on different time scales: The process
X moves quickly and randomly within a group of states, but
moves only slowly from one group of states to another.
Since all other information contained in X is not necessarily
preserved by minimizing CP (X, W), this cost function can
also lead to undesired solutions: For example, if X is i.i.d.
and hence does not contain any temporal dependence structure,
then I(X1 ; X2 ) = 0 and CP (X, W) = 0 for every mapping
W.
A third objective for optimization may be worth mentioning.
The information contained in X splits into a part describing its
temporal dependence structure (measured by its redundancy
rate I(X1 ; X2 )) and a part describing its new information
generated in each time step (measured by its entropy rate
H̄(X)). Indeed, we have
H(X) = H̄(X) + I(X1 ; X2 ).
(14)
While in this work we focus on preserving Markovity
via CL (X, W) and the temporal dependence structure via
CP (X, W), the authors of [12] investigated conditions such
that the newly generated information (measured by H̄(X)) is
preserved. Developing a Markov aggregation framework that
trades between three different goals – Markovity, temporal
dependence, generated information – is the object of future
work.
IV. R EGULARIZED M ARKOV C HAIN AGGREGATION
In this section we combine the approaches in [7] and [5]
to obtain a new cost function for Markov aggregation. As
discussed in Section III, the aim of Markov aggregation in
[7] is to get a process Y which is as Markov as possible.
This can be captured well by the cost function D̄(Y||Ỹ)
in the light of Lemma 2. The authors of [5] define the
Markov aggregation problem in terms of finding a mapping
W which preserves the temporal dependence structure in X.
This temporal dependence is captured well by the redundancy
rate of the process, which for a Markov chain X equals
R̄(X) = I(X1 ; X2 ). Preserving this temporal dependence
structure is thus well captured by maximizing the redundancy
rate of Y, i.e., by the following optimization problem:
min R̄(X) − R̄(Y)
W
(15)
5
Hence, to combine both the goal of Markovity and the goal of
preserving temporal information, one can define the following
Markov aggregation problem
min (1 − β)D̄(Y||Ỹ) + β(R̄(X) − R̄(Y))
W |
{z
}
(16)
=:δβ (X,W)
where 0 ≤
Definition 1.
ensures that
negative. We
β ≤ 1. Clearly, for β = 0 we are back at
For a general β, the data processing inequality
R̄(X) ≥ R̄(Y), hence the cost (16) is nonmoreover have
Lemma 3. δβ (X, W) is non-decreasing in β.
Proof: See Section VIII-B.
Although minimizing δβ (X, W) tries to preserve both
Markovity and the temporal information in X, the computation
of (16) requires computing H̄(Y). We thus take the approach
of [7] to relax (16). By [7, Thm. 1], we have
D̄(Y||Ỹ) ≤ CL (X, W).
(17)
By the data processing inequality we also have
R̄(X) − R̄(Y) ≤ CP (X, W).
were proposed to solve the respective Markov aggregation
problem.
1
• For β = 2 , optimizing (20) is equivalent to optimizing
CP (X, W). The authors of [5] proposed this cost function for deterministic aggregations, i.e., they proposed
optimizing CP (X, g). Note that this restriction to deterministic aggregations comes at the loss of optimality:
In [17, Example 3] a reversible, three-state Markov chain
was given for which the optimal aggregation to |Y| = 2
states is stochastic. For the bi-partition problem, i.e., for
|Y| = 2, the authors of [5] propose a relaxation to
a spectral, i.e., eigenvector-based optimization problem,
the solution of which has a computational complexity of
O(|X |3 ). In general, this relaxation leads to a further loss
of optimality, even among the search over all deterministic bi-partitions. For a general Y, they suggest to solve
the problem by repeated bi-partitioning, i.e., splitting sets
of states until the desired cardinality is achieved.
• For β = 1, the problem becomes equivalent to maximizing I(X1 ; Y2 ). This is exactly the information bottleneck
problem [18] for a Lagrangian parameter γ → ∞:
(18)
I(X2 ; Y2 ) − γI(X1 ; Y2 ).
Hence, combining the two we get an upper bound on
δβ (X, W) that does not require computing H̄(Y). Indeed, for
β ∈ [0, 1],
δβ (X, W) ≤ (1 − β)CL (X, W) + βCP (X, W).
(19)
•
Rather than the right-hand side of (19), we propose the
following cost function for Markov aggregation:
Cβ (X, W) := (1 − 2β)CL (X, W) + βCP (X, W).
where again β ∈ [0, 1]. One can justify going from (19) to
(20) by noticing that for every 0 ≤ β ≤ 1 for (19), one can
find a 0 ≤ β ≤ 0.5 for (20) such that the two optimization
problems are equivalent, i.e., they have the same optimizer W.
Furthermore, for β = 1, the cost function in (20) corresponds
to information bottleneck problem, a case that is not covered
by (19). Hence, not only is Cβ a strict generalization of δβ but
also has the information bottleneck problem as an interesting
corner case. In the following we summarize some of the
properties of Cβ .
Proof: See Section VIII-C.
V. R ELATED W ORK : S PECIAL C ASES
OF
Cβ (X, W)
We now show that specific settings of β lead to cost
functions that have been proposed previously in the literature.
We list these approached together with the algorithms that
Algorithmic approaches to solving this optimization problem are introduced in [19]. Note, that in this case the
optimal aggregation will be deterministic [17, Thm. 1].
For β = 0, the authors of [7] relaxed their cost function
C0 (X, g) = CL (X, g) as
C0 (X, g) = H(Y2 |Y1 ) − H(Y2 |X1 ) = I(Y2 ; X1 |Y1 )
≤ I(X2 ; X1 |Y1 ) = I(X2 ; X1 ) − I(X2 ; Y1 )
(22)
(20)
Lemma 4. For Cβ and 0 ≤ β ≤ 1 we have:
1) Cβ (X, W) ≥ 0
2) δ0.5 (X, W) = C0.5 (X, W) = 21 CP (X, W)
3) C1 (X, W) = CIB (X, W) := I(X1 ; X2 |Y2 )
4) For β ≤ 12 , βCP (X, W) ≤ δβ (X, W) ≤ Cβ (X, W)
5) For β ≥ 12 , Cβ (X, W) ≤ δβ (X, W) ≤ βCP (X, W)
6) If X is reversible, then Cβ (X, W) is non-decreasing in
β
(21)
•
and proposed using the agglomerative information bottleneck method [20] with the roles of X1 and X2 in (21) exchanged to solve this relaxed optimization problem. The
method has a computational complexity of O(|X |4 ) [19,
Sec. 3.4]. While the mapping minimizing CL (X, W)
may be stochastic, the mapping minimizing (22) will
be deterministic; hence, with this relaxation in mind,
the restriction to deterministic aggregations made in [7]
comes without an additional loss of optimality compared
to what is lost in the relaxation.
The authors of [6] proposed minimizing
I(X1 ; X2 ) − I(X2 ; Y1 ) − γH(Y2 |X1 ).
(23)
They suggested using a deterministic annealing approach,
reducing γ successively until γ = 0. In the limiting case,
the cost function then coincides with (22) and the optimal
aggregation is again deterministic. Note that, for reversible Markov chains, we have I(X1 ; Y2 ) = I(X2 ; Y1 ),
hence both (22) and (23) (for γ = 0) are equivalent to
C1 (X, W). Analyzing [6, Sec. III.B] shows that in each
annealing step the quantity
D(pX2 |X1 =x ||pX2 |Y1 =y )
X
pX |X (x′ |x)
:=
pX2 |X1 (x′ |x) log 2 1 ′
pX2 |Y1 (x |y)
′
x ∈X
(24)
6
has to be computed for every x and y. Hence, the computational complexity of this approach is O(|Y| · |X |2 )
in each annealing step.
VI. M ARKOV C HAIN AGGREGATION A LGORITHMS
We now propose an iterative method for optimizing (20)
over deterministic aggregations for general values of β. The
method consists of a sequential optimization algorithm (Algorithm 1) and an annealing procedure for β (Algorithm 2)
that prevents getting stuck in local optima. Since we focus
only on deterministic aggregations, in the remainder of this
section we can replace Cβ (X, W) by Cβ (X, g) for some
g: X → Y. Our algorithm has a computational complexity of
O(|Y| · |X |2 ) per iteration. Note, however, that the restriction
to deterministic aggregation functions comes, at least for some
values of β, with a loss of optimality, i.e., in general we have
minW Cβ (X, W) ≤ ming Cβ (X, g).
A. Sequential Algorithm
We briefly illustrate an iteration of Algorithm 1: Suppose
x ∈ X is mapped to the aggregate state y ∈ Y, i.e., g(x) = y.
We remove x from aggregate state y. We then assign x to
every aggregate state y ′ , y ′ ∈ Y, while keeping the rest of the
mapping g the same and evaluate the cost function. Finally, we
assign x to the aggregate state that minimized the cost function
(breaking ties, if necessary). This procedure is repeated for
every x ∈ X .
Algorithm 1 Sequential Generalized Information-Theoretic
Markov Aggregation.
1: function g = S GITMA(P, β, |Y|, #itermax , optional:
initial aggregation function ginit )
2:
if ginit is empty then
⊲ Initialization
3:
g ← Random Aggregation Function
4:
else
5:
g ← ginit
6:
end if
7:
#iter ← 0
8:
while #iter < #itermax do
⊲ Main Loop
9:
for all elements x ∈ X do
⊲ Optimizing g
10:
for all aggregate
( states y ∈ Y do
g(x′ )
x′ 6= x
11:
gy (x′ ) =
⊲ Assign x
y
x′ = x
to aggregate state y
12:
Cgy = Cβ (X, gy )
13:
end for
⊲ (break ties)
14:
g = arg min Cgy
gy
15:
16:
17:
18:
end for
#iter ← #iter + 1
end while
end function
It is easy to verify that the cost function is reduced in each
step of Algorithm 1, as a state is only assigned to a different
aggregate state if the cost function is reduced. Hence, the
algorithm modifies g in each iteration in order to reduce the
cost until it either reaches the maximum number of iterations
or until the cost converges.
Note that the algorithm is random in the sense that it
is started with a random aggregation function g. Depending
on the specific application, though, a tailored initialization
procedure may lead to performance improvements.
Finally, it is worth mentioning that for β = 1 our Algorithm 1 is equivalent to the sequential information bottleneck
algorithm proposed in [19, Sec. 3.4].
B. Annealing Procedure for β
Although Algorithm 1 is guaranteed to converge (with
proper tiebreaking), convergence to a global optimum is not
ensured. The algorithm may get stuck in poor local minima.
This happens particularly often for small values of β, as our
experiments in Section VII-B show. The reason is that, for
small β, Cβ (X, W) has many poor local minima and, randomly
initialized, the algorithm is more likely to get stuck in one of
them. In contrast, our results suggest that for larger values of
β the cost function has only few poor local minima and that
the algorithm converges to a good local or a global minimum
for a significant portion of random initializations.
A solution for small β would thus be to choose an initialization that is close to a “good” local optimum. A simple idea is
thus to re-use the function g obtained for a large value of β as
initial aggregation for smaller values of β. We thus propose the
following annealing algorithm: We initialize β = 1 to obtain
g. Then, in each iteration of the annealing procedure, β is
reduced and the aggregation function is updated, starting from
the result of the previous iteration. The procedure stops when β
reaches the desired value, βtarget . The β-annealing algorithm is
sketched as Algorithm 2. As is clear from the description, the
β-annealing algorithm closely follows graduated optimization
in spirit [21]. The results for synthetic datasets with and
without β-annealing are discussed in Section VII-B, which
show that without restarts one keeps getting stuck in bad local
optima for small β, while with β-annealing one is able to avoid
them. Furthermore in our experiments we have observed that
β-annealing achieves good results for random initializations,
hence tailoring initialization procedures is not necessary at
least for the scenarios we considered.
Algorithm 2 β-Annealing Information-Theoretic Markov Aggregation
1: function g = A NN ITMA(P, βtarget , |Y|, #itermax , ∆)
2:
β←1
3:
g = sGITMA(P, β, |Y|, #itermax )
⊲ Inizialization
4:
while β > βtarget do
5:
β ← max{β − ∆, βtarget }
6:
g = sGITMA(P, β, |Y|, #itermax , g)
7:
end while
8: end function
Note that the β-annealing algorithm admits producing results for a series of values of β at once: Keeping all intermediate aggregation functions, one obtains aggregations
for all values of β in the set {1, 1 − ∆, 1 − 2∆, . . . , 1 −
7
1−β
∆⌈ ∆target ⌉, βtarget }. The aggregations one obtains are exactly
those one would obtain from restarting A NN ITMA for each
value in this set, each time with the same random initial
partition. We used this fact in our experiments: If we were
interested in results for βtarget ranging between 0 and 1 in
steps of 0.05, rather than restarting A NN ITMA for each value
in this set, we started A NN ITMA for βtarget = 0 and ∆ = 0.05
once, keeping all intermediate results.
C. Computational Complexity of the Sequential Algorithm
Note that the asymptotic computational complexity of Algorithm 2 equals that of Algorithm 1, since the former simply
calls the latter ⌈(1 − βtarget )/∆⌉ + 1 times. We thus only
evaluate the complexity of Algorithm 1. To this end, we first
observe that the cost Cβ (X, W) can be expressed with only
three mutual information terms for any β and W:
Cβ (X, g) = βI(X1 ; X2 ) + (1 − 2β)I(X1 ; g(X2 ))
− (1 − β)I(g(X1 ); g(X2 )). (25)
The first term I(X1 ; X2 ) is constant regardless of the aggregation hence the computation of Cβ (X, g) depends upon the
computation of the other two terms.
In each iteration of the main loop, we evaluate Cβ (X, gy )
in line 12 for each x ∈ X and y ∈ Y. Note that gy differs
from the current g only for one element as defined in line
11. Thus, the joint PMF pX1 ,gy (X2 ) differs from pX1 ,g(X2 ) in
only two rows and hence can be computed from pX1 ,g(X2 )
in O(|X |) computations. Moreover, I(X1 ; gy (X2 )) can be
computed from I(X1 ; g(X2 )) in O(|X |) computations, cf. [20,
Prop. 1]. This is due to the fact that we can write [22,
eq. (2.28)]
I(X1 ; gy (X2 )) = I(X1 ; g(X2 ))
X
pX1 ,gy (X2 ) (x1 , y2 )
+
pX1 ,gy (X2 ) (x1 , y2 ) log
pX1 (x1 )pgy (X2 ) (y2 )
x1 ∈X
y2 ∈{y,g(x)}
−
X
pX1 ,g(X2 ) (x1 , y2 ) log
x1 ∈X
y2 ∈{y,g(x)}
pX1 ,g(X2 ) (x1 , y2 )
pX1 (x1 )pg(X2 ) (y2 )
.
(26)
The term I(gy (X1 ); gy (X2 )) can be computed from
I(g(X1 ); g(X2 )) in O(|Y|) computations, but requires the
updated joint PMF pgy (X1 ),gy (X2 ) . This PMF can be computed
from pg(X1 ),g(X2 ) in O(|X |) computations. Combining this
with the fact that line 12 is executed once for each aggregate
state in Y and once for each state in X in every iteration,
we get that optimizing g has a computational complexity of
O(|Y| · |X |2 ) per iteration.
VII. E XPERIMENTS
AND
R ESULTS
A. A Non-Reversible Markov Chain
The last property of Lemma 4 cannot be generalized to
non-reversible Markov chains. Specifically, as the proof of
Lemma 4 shows, Cβ is non-decreasing in β iff CP ≥ 2CL .
Since one can find also non-reversible Markov chains for
which this holds, reversibility is sufficient but not necessary
for Cβ to be non-decreasing in β. We next consider a nonreversible Markov chain X ∼ Mar({1, 2, 3}, P) with
0.4
0.3
0.3
P = 0.25 0.3
0.45
(27)
0.15 0.425 0.425
and let g be such that g(1) = 1 and g(2) = g(3) = 2. Then,
CL = 0.0086 and CP = 0.0135, i.e., CP < 2CL . In this case,
Cβ is decreasing with increasing β.
B. Quasi-Lumpable and Nearly Completely Decomposable
Markov Chains
Suppose we have a partition {Xi }, i = 1, . . . , M , of X
with |Xi | = Ni . Then for any A′ and P′ij that are M × M
and Ni × Nj row stochastic matrices, respectively, define A =
[aij ] = (1 − α)A′ + αI, α ∈ [0, 1], and let
a11 P′11
a12 P′12 · · · a1M P′1M
a21 P′21
a22 P′22 · · · a2M P′2M
P′ =
. (28)
..
..
..
..
.
.
.
.
′
′
′
aM1 PM1 aM2 P22 · · · aMM PMM
Let further X ∼ Mar(X , P′ ). If g induces the partition
{Xi }, then it can be shown that Y is Markov with transition
probability matrix A, i.e., Ỹ ≡ Y, and C0 (X, Ỹ) = 0. The
Markov chain X is lumpable w.r.t. the partition g. The matrix
P′ is block stochastic and the parameter α specifies how
dominant the diagonal blocks are. Specifically, if α = 1, then
P′ is block diagonal and we call X completely decomposable.
Such a Markov chain is not irreducible. We hence look at
Markov chains X ∼ Mar(X , P) with
P = (1 − ε)P′ + εE
(29)
where ε ∈ [0, 1] and where E (which can be interpreted as
noise) is row stochastic and irreducible. For small values of
ε we call X nearly completely decomposable (NCD) if α is
close to one, otherwise we call it quasi-lumpable.
We now perform experiments with these types of Markov
chains. We set M = 3, N1 = N2 = 25, and N3 = 50,
and chose the parameters from α ∈ {0, 0.5, 0.95} and ε ∈
{0, 0.4, 0.8}. For each pair (α, ε), we generated 250 random
matrices A′ and P′ij . A selection of the corresponding matrices
P is shown in Fig. 2.
Note that in practice the states of even a completely decomposable Markov chain X are rarely ordered such that the
transition probability matrix is block diagonal. Rather, the state
labeling must be assumed to be random. In this case, P is
obtained by a random permutation of the rows and columns
of a block diagonal matrix (see Fig. 2(d)), which prevents
the optimal aggregation function being “read off” simply by
looking at P. That P has a block structure in our case does
not affect the performance of our algorithms, since they 1) are
unaware of this structure and 2) are initialized randomly.
We applied our aggregation algorithm both with and without
the annealing procedure for β ∈ {0, 0.1, . . . , 0.9, 1} with the
goal of retrieving the partition {Xi }. We measure the success,
i.e., the degree to which the function g obtained from the
8
10
20
20
20
40
40
40
60
60
60
80
80
80
100
100
100
20
30
40
50
60
70
80
90
20
40
60
80
100
20
(a) α = ε = 0
40
60
80
100
20
(b) α = 0.95, ε = 0.8
40
60
80
100
(c) α = 0.95, ε = 0.4
100
20
40
60
80
100
(d) α = 0.95, ε = 0.4; rows and
columns are permuted
0.2
ARI
Cβ
Cβ
1
0.2
0.15
0.1
0.1
0.5
5 · 10−2
0
0
0.2
0.4
0.6
0.8
0
1
0
0
0.2
0.4
β
0.8
0
1
α=0 (lumpable)
α=0.5
α=0.95 (NCD)
0.2
0.4
0.6
0.8
0.5
0
1
1
0.5
0
0
0.2
0.4
β
(h) ARI with β-annealing, ε = 0
0.8
1
ARI
ARI
0.5
0.6
(g) ARI without β-annealing, ε = 0.4
1
0
0.4
β
(f) Cost without β-annealing, ε = 0.4
1
0
0.2
β
(e) Cost with β-annealing, ε = 0.4
ARI
0.6
0.6
0.8
β
(i) ARI with β-annealing, ε = 0.4
1
0
0.2
0.4
0.6
0.8
1
β
(j) ARI with β-annealing, ε = 0.8
Fig. 2. (a)-(c): Colorplots of the transition probability matrices P for different values of α and ε. For large α the block diagonal structure becomes more
dominant. (d): A random permutation of the rows and columns hides the block structure. (e)-(j): Curves showing the cost function Cβ and the adjusted Rand
index (ARI) for different settings. Mean values (solid lines) are shown together with the standard deviation (shaded areas).
algorithm agrees with the partition {Xi }, using the adjusted
Rand index (ARI). An ARI of one indicates that the two
partitions are equivalent. Note that we always assume that the
number M of sets in the partition {Xi } is known.
The results are shown in Fig. 2. Specifically, Fig. 2(e) shows
that the cost for the aggregation found by our algorithm with
β-annealing decreases monotonically with decreasing β: We
obtain a partition for a given value of β. This partition has,
by assuming CP ≥ 2CL (cf. Section VII-A), an even lower
cost for a smaller value of β. Further optimization for this
smaller value of β further reduces the cost, leading to the
depicted phenomenon. In contrast, the sequential Algorithm 1
without the annealing procedure fails for values of β less than
0.5. This is apparent both in the cost in Fig. 2(f) (which has
a sharp jump around β = 0.5) and in the ARI in Fig. 2(g)
(which drops to zero). Apparently, the algorithm gets stuck in
a bad local optimum.
Figs. 2(h) to 2(j) show the ARI of the aggregations obtained
by our algorithm with β-annealing. First of all, it can be
seen that performance improves with increasing α, since
the dominant block structure makes discovering the correct
partition more easy. Moreover, it can be seen that for α = 0 the
optimum β lies at smaller values, typically smaller than 0.5.
The position of this optimum increases with increasing noise:
While in the noiseless case the correct partition is typically
obtained for β close to zero, in the highly noisy case of
ε = 0.8 we require β ≈ 0.4 to achieve good results. The
reason may be that the higher noise leads to more partitions
9
A GGREGATING A
β Value
Ref.
β=1
β = 0.8
β = 0.5
β=0
ARI
–
0.43
0.46
0.35
0.12
A GGREGATING A
β Value
Ref.
β=1
β = 0.8
β = 0.5
β=0
ARI
–
0.2
0.34
0.24
0.35
0.15
0.31
0.01
0.02
LETTER BI - GRAM MODEL .
TABLE I
T HE PARTITIONS ARE SHOWN TOGETHER WITH THE ARI ARI W. R . T. THE REFERENCE PARTITION ( FIRST
ROW ) FOR |Y| = 4
Partitions, shown for |Y| = 4
{ },{!"$’(),-.:;?[]},{aeiou},{0123456789},{AEIOU},{BCDFGHJKLMNPQRSTVWYZ},{bcdfghjklmnpqrstvwxyz}
{ !’),-.0:;?]},{aeioy},{"$(123456789ABCDEFGHIJKLMNOPQRSTUVWY[h},{Zbcdfgjklmnpqrstuvwxz}
{ !’),-.:;?Z]},{aeiouy},{"$(0123456789ABCDEFGHIJKLMNOPQRSTUVWY[h},{bcdfgjklmnpqrstvwxz}
{ !3?Z},{’2456789AOUaeiou},{"$(-01BCDEFGHIJKLMNPQRSTVWY[bhjqw},{),.:;]cdfgklmnprstvxyz}
{ -2CEFMPSTcfgopst},{"’456789AOUZaeiu},{!$1?BDGHJLNQRVW[bhjklmqrvwz},{(),.03:;IKY]dnxy}
LETTER BI - GRAM MODEL .
TABLE II
T HE PARTITIONS ARE SHOWN TOGETHER WITH THE ARI ARI W. R . T. THE REFERENCE PARTITION ( FIRST
ROW ) FOR |Y| ∈ {2, 7}
Partitions, shown for |Y| ∈ {2, 7}
{ },{!"$’(),-.:;?[]},{aeiou},{0123456789},{AEIOU},{BCDFGHJKLMNPQRSTVWYZ},{bcdfghjklmnpqrstvwxyz}
{ !"’),-.01235689:;?KU]aehioy},{$(47ABCDEFGHIJLMNOPQRSTVWYZ[bcdfgjklmnpqrstuvwxz}
{ !’),-.:;?]},{aeioy},{"$(0123456789ABCDEFGHIJLMNOPQRSTVWY[},{bcfjmpqstw},{dgx},{KUh},{Zklnruvz}
{ !"’),-.01235689:;?EU]aehiouy},{$(47ABCDFGHIJKLMNOPQRSTVWYZ[bcdfgjklmnpqrstvwxz}
{ !’),-.:;?]},{aeioy},{$"(0123456789ABCDEFGHIJLMNOPQRSTUVWYZ[},{Kh},{bcfjkmpqstw},{dg},{lnruvxz}
{ !’-12368?EOUZaeiou},{"$(),.04579:;ABCDFGHIJKLMNPQRSTVWY[]bcdfghjklmnpqrstvwxyz}
{ ’},{!),.:;?]dy},{aeiou},{"$(-0123589ACEIMOPRSTUWZ},{BDFGHJKLNQVYhj},{7[bcfgkmpqstw},{46lnrvxz}
{ !$(-0124578?ABCFHLMNOPRSTUVWaceglnostuwxz},{"’),.369:;DEGIJKQYZ[]bdfhijkmpqrvy}
{ 4689ao},{$’AKOiux},{!?HVZhjkmvz},{"(-25CEFLMNRUWY[egnprs},{37BPQbl},{1:;STctw},{),.0DGIJ]dfy}
being quasi-lumpable by leading to an i.i.d. Y, hence for small
values of β one may get drawn into these “false solutions”
more easily. In contrast, for NCD Markov chains (i.e., for
α = 0.95) sometimes noise helps in discovering the correct
partition. Comparing Figs. 2(h) and 2(i), one can see that a
noise of ε = 0.4 allows us to perfectly discover the partition.
We believe that a small amount of noise helps in escaping bad
local minima.
The fact that the β for which the highest ARI is achieved not
necessarily falls together with the values 0, 0.5, or 1 indicates
that our generalized aggregation framework has the potential to
strictly outperform aggregation cost functions and algorithms
that have been previously proposed (cf. Section V).
C. An Example from Natural Language Processing
We took the letter bi-gram model from [23], which was
obtained by analyzing the co-occurrence of letters in F. Scott
Fitzgerald’s book “The Great Gatsby”. The text was modified
by removing chapter headings, line breaks, underscores, and
by replacing é by e. With the remaining symbols, we obtained
a Markov chain with an alphabet size of N = 76 (upper and
lower case letters, numbers, punctuation, etc.).
We applied Algorithm 2 for |Y| ∈ {2, . . . , 7} and β ∈
{0, 0.1, . . . , 0.9, 1}. To get consistent results, we restarted the
algorithm 20 times for β = 1 and chose the aggregation g
that minimized C1 (X, g); we used this aggregation g as an
initialization for the β-annealing procedure.
Looking at the results for |Y| = 4 in Table I, one can
observe that the results for β = 0.8 appear to be most
meaningful when compared to other values of β such as
β = 1 (information bottleneck), β = 0.5 (as proposed
in [5]), and β = 0 (as proposed in [7]). Specifically, for
β = 0 it can be seen that not even the annealing procedure
was able to achieve meaningful results. This conclusion is
supported by calculating the ARI of these aggregations for
a plausible reference aggregation of the alphabet into upper
case vowels, upper case consonants, lower case vowels, lower
case consonants, numbers, punctuation, and the blank space as
shown in the first row of the Table I. The absolute ARI values
are not a good performance indicator in this case since we are
comparing to a reference partition with seven sets whereas
|Y| = 4.
In Table II the same experiment is repeated for |Y| ∈ {2, 7}.
We again observe that β = 0.8 leads to the most meaningful
results which is also supported by ARI values.
D. Clustering via Markov Aggregation
Data points are often described only by pairwise similarity
values, and these similarity values can be used to construct the
transition probability matrix of a Markov chain. Then, with
this probabilistic interpretation, our information-theoretic cost
functions for Markov aggregation can be used for clustering.
This approach has been taken by [3], [4].
We considered two different data sets: three linearly separable clusters and three concentric circles, as shown in Fig 3. The
three linearly separable clusters were obtained by placing 40,
20, and 40 points, drawn from circularly symmetric Gaussian
distributions with standard deviations 2.5, 0.5, and 1.5 at
horizontal coordinates -10, 0, and 10, respectively. The three
concentric circles were obtained by placing 40 points each at
uniformly random angles at radii {0.1, 7, 15}, and by adding
to each data point spherical Gaussian noise with a standard
deviation of 0.3. In both cases, we computed the transition
probability matrix P according to
Pi→j ∝ e
−
kxi −xj k2
2
σk
(30)
where xi and xj are the coordinates of the i-th and j-th data
point, k ·k22 is the squared Euclidean distance, and where σk is
a scale parameter. We set σk to the average squared Euclidean
distance between a data point and its k nearest neighbors (and
10
20
b
b
b
b
80
40
60
80
100
b b
b
bb
b
b
b
b
bb
bb
b
b
b
bb
b
b
20
30
30
40
40
50
50
60
60
70
70
80
80
b
b
b bb b
b
b b
b
bb
b
bb
b
bb
bb
40
60
80
100
(c) P, k = 120
b
b
b
b
b
b
b
b
b
b
b
b
b bb
b
b bb
b
b
b bb b
b
b
b
b
b bb b
b b
b
b
b
b
b
b
b
bb
b b b b bb bb b bb b b bb b b
b b
b bb
b
bb
b
bb
b
b
b
bb
b
b
bb
bb b
b
b
bbb
b
b
b
bb
b
b
b
b
b
b b
b
b
bb b
{0.2, 0.5, 0.8},
10
90
b
120
20
b
b
bb
b b b b bb bb b bb b b bb b b
b b
b
10
b bb
b
b bb
b
100
bbb
b
b bb
b
b
80
b bb b
b
b
60
b
20
(b) β ∈
k = 15
b
b
120
(a) P, k = 15
b
b
bb
b
b
b
b
40
b
b
b
b
b b
b
b
b bb b
b
20
b
b
b
bb
b b b b bb bb b b b bb bb b b
b b
b
bb b
120
b bb
b
b bb
b
100
b
b
20
b
b bb
b
b
b
b bb b
b
40
60
b
b
b
b b
bb
b
b
b
b
b
bbb
b
120
(d) β = 0.2, k = 120
b
b
b
b
b
b
b
b
b
b
bb
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
b
(e) β = 0.8, k = 120
b
b
b
b
b
b b b b b b bb
b bb b b b
b b
b
b
b
b
b b
b
b
b
b
bb
b
b
b
b
b b
b
b b b
b b bb bb
b
b
b bb
b
b
b b
b
b
90
100
100
20
40
60
80
(f) P, k = 15
100
20
40
60
80
100
(h) β ∈ {0.2, 0.5, 0.8}, k ∈ {15, 100}
(g) P, k = 100
Fig. 3. Clustering three circles (first row) and three linearly separable clusters (second row). For k = 15, the transition probability matrices (shown in (a)
and (f)) are nearly completely decomposable. The result for the three circles depends strongly on a careful setting of the parameters β and k ((b), (d), and
(e)), while the three linearly separable clusters were separated correctly for all parameter choices (h).
averaged this quantity over all data points). We set k either to
15 or to the total number of data points.
We applied our Algorithm 2 with the annealing procedure
for β. As in the previous experiment, we restarted the algorithm 50 times for β = 1 and chose the aggregation g
that minimized C1 (X, g); we used this aggregation g as an
initialization for the β-annealing procedure.
The results are shown in Fig. 3, together with a colorplot
of the respective transition probability matrices. It can be
seen that the three linearly separable clusters were detected
correctly for all chosen parameter values. This is not surprising
for k = 15, since in this case the resulting Markov chain
is nearly completely decomposable. Interestingly, though, the
same results were observed for k = 100 for which P is
structured, but not block diagonal. One may claim that these
results are due to Algorithm 2 getting stuck in a local optimum
for β = 1 which accidentally coincides with the correct
clustering, and that optimizing our cost function for values
of β larger than 0.5 but smaller than 1 may fail. Since we
reproduced these results by using Algorithm 1 (with 50 restarts
to escape bad local optima) for values of β greater than 0.5,
this claim can be refuted.
For the three concentric circles, things look different. We
correctly identified the clusters only for a nearly completely
decomposable P, i.e., for a careful setting of k (and we were
able to reproduce these results for β greater than 0.5 using
Algorithm 1). For k = 120, i.e., equal to the number of data
points, the three circles were not identified correctly.
Since we have reason to believe that the optimal k depends
strongly on the data set, we are hesitant to recommend this approach to cluster data points that are not linearly separable (in
which case a simpler method such as k-means would suffice).
Our preliminary analysis of [3] suggests that their approach
(in which X is a random walk on the k-nearest neighbor graph
of the data set and in which the authors chose β = 0.5) suffers
from similar problems. Finally, the authors of [4] suggest to
let X “relax” to some metastable point, i.e., take an r-th
power of P such that Pr is approximately a projection; their
approach is equivalent to ours for β = 1, with P replaced
by Pr . Nevertheless, also this approach requires setting r and
k for (30). Whether this relaxation to metastability can be
successfully combined with our generalized cost function for
Markov aggregation will be deferred to future investigations.
VIII. P ROOFS
A. Proof of Proposition 1
Consider the relation Rε = {(g(x), x): x ∈ X }. It can be
shown that
∀T ⊆ X ∪ Y: Rε (T ) = g −1 (T ∩ Y) ⊆ X .
(31)
We thus need to show that, for all x and all B ⊆ Y,
X
X
Px→x′ ≥
Qg(x)→y − ε.
(32)
Now let R = [Rx→y ] = PW, i.e., we have
X
Rx→y =
Px→x′ .
(33)
x′ ∈g−1 (B)
y∈B
x′ ∈g−1 (y)
One can show along the lines of [7, (65)–(68)] that
X
X
Rx→y
CL (X, W) =
µx
Rx→y log
Qg(x)→y
x∈X
y∈Y
{z
}
|
(34)
=:D(Rx→· ||Qg(x)→· )
from which we get that, for every x,
D(Rx→· ||Qg(x)→· ) ≤
CL (X, W)
.
minx µx
(35)
11
With Pinsker’s inequality [22, Lemma 12.6.1] and [22,
(12.137)] we thus get that, for every x and every B ⊆ Y,
s
X
ln(2)CL (X, W)
.
(36)
Rx→y − Qg(x)→y ≤
2 minx µx
y∈B
Combining this with (33) thus shows that (32) holds for
s
ln(2)CL (X, W)
.
(37)
ε=
2 minx µx
δβ (X, W) − βI(X1 ; X2 )
= (1 − β)H(Y2 |Y1 ) − (1 − 2β)H̄(Y) − βH(Y )
≤ (1 − β)H(Y2 |Y1 ) − (1 − 2β)H(Y2 |X1 ) − βH(Y )
= (1 − 2β)H(Y2 |Y1 ) − (1 − 2β)H(Y2 |X1 ) − βI(Y1 ; Y2 )
= (1 − 2β)CL (X, W) − βI(Y1 ; Y2 ).
The inequality is reversed for β ≥ 0.5.
For the fifth property, we repeat the last steps with
− (1 − 2β)H̄(Y) ≤ −(1 − 2β)H(Y2 |Y1 )
This completes the proof.
B. Proof of Lemma 3
We show that the derivative of δβ (X, W) w.r.t. β is positive.
Indeed,
d
δβ (X, W)
dβ
= R̄(X) − R̄(Y) − D̄(Y||Ỹ)
(38)
= I(X1 ; X2 ) − H(Y ) + H̄(Y) − H(Y2 |Y1 ) + H̄(Y). (39)
The entropy rate of the reversed process equals the entropy
rate of the original process, i.e.,
H̄(Y) = lim H(Yn |Y1n−1 ) = lim H(Y1 |Y2n ).
n→∞
The fourth property is obtained by observing that, if β ≤ 0.5
n→∞
(40)
We can now apply [22, Lem. 4.4.1] to both sides to get
H̄(Y) ≥ H(Y2 |X1 ) and H̄(Y) ≥ H(Y1 |X2 ). We use this
in the derivative to get
d
δβ (X, W)
dβ
≥ I(X1 ; X2 ) − H(Y ) + H(Y1 |X2 ) − H(Y2 |Y1 ) + H(Y2 |X1 )
(41)
= H(X|Y ) − H(X1 |Y1 , X2 ) − H(Y2 |Y1 ) + H(Y2 |X1 )
(42)
= I(X1 ; X2 |Y1 ) − I(X1 ; Y2 |Y1 ) ≥ 0
(43)
by data processing.
C. Proof of Lemma 4
The first property follows by recognizing that
Cβ (X, W)
= (1 − β)CL (X, W) + β(CP (X, W) − CL (X, W))
(44)
and that CP (X, W) ≥ CL (X, W).
The second property follows immediately from the definition of δβ (X, W) and CP (X, W).
For the third property, note that
C1 (X, W) = CP (X, W) − CL (X, W)
= I(X1 ; X2 ) − H(Y ) + H(Y2 |X1 )
= I(X1 ; X2 ) − I(X1 , Y2 ) = I(X1 ; X2 |Y2 ).
(45)
noticing that (1 − 2β) ≤ 0 if β ≥ 0.5. Again, the inequality
is reversed for β ≤ 0.5.
If X is reversible, then the PMFs do not change if the
order of the indices is reversed. As a consequence, we
have I(X1 ; X2 |Y2 ) = I(X2 ; X1 |Y1 ) = C1 (X, W). But
CL (X; W) = I(Y2 ; X1 |Y1 ) ≤ C1 (X, W) by data processing.
Thus, the sixth property follows by noting that, with (44),
Cβ (X, W) = (1 − β)CL (X, W) + βC1 (X, W).
ACKNOWLEDGMENTS
The authors thank Matthias Rungger and Majid Zamani,
both from Hybrid Control Systems Group, Technical University of Munich, for discussions suggesting the connection
between lumpability and bisimulation. The work of Rana Ali
Amjad was supported by the German Ministry of Education
and Research in the framework of an Alexander von Humboldt
Professorship. The work of Bernhard C. Geiger was funded
by the Erwin Schrödinger Fellowship J 3765 of the Austrian
Science Fund.
R EFERENCES
[1] M. A. Katsoulakis and J. Trashorras, “Information loss in coarsegraining of stochastic particle dynamics,” J. Stat. Phys., vol. 122, no. 1,
pp. 115–135, 2006.
[2] A. Abate, “Approximation metrics based on probabilistic bisimulations
for general state-space Markov processes: A survey,” Electronic Notes
in Theoretical Computer Science, vol. 297, pp. 3 – 25, 2013, Proc.
Workshop on Hybrid Autonomous Systems.
[3] A. Alush, A. Friedman, and J. Goldberger, “Pairwise clustering based on
the mutual-information criterion,” Neurocomputing, vol. 182, pp. 284–
293, 2016.
[4] N. Tishby and N. Slonim, “Data clustering by Markovian relaxation and
the information bottleneck method,” in Advances in Neural Information
Processing Systems (NIPS), Denver, CO, Nov. 2000. [Online]. Available:
http://citeseer.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.3488
[5] K. Deng, P. G. Mehta, and S. P. Meyn, “Optimal Kullback-Leibler
aggregation via spectral theory of Markov chains,” IEEE Trans. Autom.
Control, vol. 56, no. 12, pp. 2793–2808, Dec. 2011.
[6] Y. Xu, S. M. Salapaka, and C. L. Beck, “Aggregation of graph models
and Markov chains by deterministic annealing,” IEEE Trans. Autom.
Control, vol. 59, no. 10, pp. 2807–2812, Oct. 2014.
[7] B. C. Geiger, T. Petrov, G. Kubin, and H. Koeppl, “Optimal KullbackLeibler aggregation via information bottleneck,” IEEE Trans. Autom.
Control, vol. 60, no. 4, pp. 1010–1022, Apr. 2015, open-access:
arXiv:1304.6603 [cs.SY].
[8] M. Vidyasagar, “Reduced-order modeling of Markov and hidden Markov
processes via aggregation,” in Proc. IEEE Conf. on Decision and Control
(CDC), Atlanta, GA, Dec. 2010, pp. 1810–1815.
[9] J. G. Kemeny and J. L. Snell, Finite Markov Chains, 2nd ed. Springer,
1976.
[10] R. M. Gray, Entropy and Information Theory. New York, NY: Springer,
1990.
12
[11] Z. Rached, F. Alajaji, and L. L. Campbell, “The Kullback-Leibler
divergence rate between Markov sources,” IEEE Trans. Inf. Theory,
vol. 50, no. 5, pp. 917–921, May 2004.
[12] B. C. Geiger and C. Temmel, “Lumpings of Markov chains, entropy rate preservation, and higher-order lumpability,” J. Appl. Probab.,
vol. 51, no. 4, pp. 1114–1132, Dec. 2014, extended version available:
arXiv:1212.4375 [cs.IT].
[13] D. Blackwell, “The entropy of functions of finite-state {M}arkov
chains,” in Trans. first {P}rague Conf. Inf. theory, {S}tatistical Decis. Funct. random Process. held {L}iblice near {P}rague from
{N}ovember 28 to 30, 1956.
Prague: Publishing House of the
Czechoslovak Academy of Sciences, 1957, pp. 13–20.
[14] O. Ordentlich, “Novel lower bounds on the entropy rate of binary hidden
Markov processes,” in Proc. IEEE Int. Sym. on Information Theory
(ISIT), Jul. 2016, pp. 690–694.
[15] J. Desharnais, A. Edalat, and P. Panangaden, “Bisimulation for labelled
Markov processes,” Information and Computation, vol. 179, no. 2, pp.
163 – 193, 2002.
[16] G. Bian and A. Abate, “On the relationship between bisimulation and
trace equivalence in an approximate probabilistic context,” in Proc. Int.
Conf. on Foundations of Software Science and Computation Structure
(FOSSACS), J. Esparza and A. S. Murawski, Eds. Uppsala: Springer
Berlin Heidelberg, Apr. 2017, pp. 321–337.
[17] B. C. Geiger and R. A. Amjad, “Mutual information-based clustering:
Hard or soft?” in Proc. Int. ITG Conf. on Systems, Communications and Coding (SCC), Hamburg, Feb. 2017, pp. 1–6, open-access:
arXiv:1608.04872 [cs.IT].
[18] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck
method,” in Proc. Allerton Conf. on Communication, Control, and
Computing, Monticello, IL, Sep. 1999, pp. 368–377.
[19] N. Slonim, “The information bottleneck: Theory and applications,” Ph.D.
dissertation, Hebrew University of Jerusalem, 2002.
[20] N. Slonim and N. Tishby, “Agglomerative information bottleneck,” in
Advances in Neural Information Processing Systems (NIPS), Denver,
CO, Nov. 1999, pp. 617–623.
[21] A. Blake and A. Zisserman, Visual Reconstruction. Cambridge, MA,
USA: MIT Press, 1987.
[22] T. M. Cover and J. A. Thomas, Elements of Information Theory, 1st ed.
Wiley Interscience, 1991.
[23] B. C. Geiger and Y. Wu, “Higher-order optimal Kullback-Leibler
aggregation of Markov chains,” in Proc. Int. ITG Conf. on Systems,
Communications and Coding (SCC), Hamburg, Feb. 2017, pp. 1–6,
open-access: arXiv:1608.04637 [cs.IT].
| 7cs.IT
|
1
Max Consensus in Sensor Networks: Non-linear
Bounded Transmission and Additive Noise
arXiv:1602.01128v1 [cs.SY] 2 Feb 2016
Sai Zhang, Cihan Tepedelenlioğlu, Senior Member, IEEE, Mahesh K. Banavar, Member, IEEE, and Andreas
Spanias, Fellow, IEEE
Abstract—A distributed consensus algorithm for estimating
the maximum value of the initial measurements in a sensor
network with communication noise is proposed. In the absence of
communication noise, max estimation can be done by updating
the state value with the largest received measurements in every
iteration at each sensor. In the presence of communication noise,
however, the maximum estimate will incorrectly drift and the
estimate at each sensor will diverge. As a result, a soft-max
approximation together with a non-linear consensus algorithm
is introduced herein. A design parameter controls the trade-off
between the soft-max error and convergence speed. An analysis of
this trade-off gives a guideline towards how to choose the design
parameter for the max estimate. We also show that if some prior
knowledge of the initial measurements is available, the consensus
process can converge faster by using an optimal step size in
the iterative algorithm. A shifted non-linear bounded transmit
function is also introduced for faster convergence when sensor
nodes have some prior knowledge of the initial measurements.
Simulation results corroborating the theory are also provided.
tions in different areas [2], [3], [12]–[15]. A comprehensive
review of the consensus literature is provided in [16]. While
average consensus is well studied in literature (e.g., [2], [3],
[17]), estimating the average is not always the goal. In various
applications, estimating the maximum measured value in the
network is necessary [15], [7]. For example, spectrum sensing
algorithms that use the OR-rule for cognitive radio applications
can be implemented using max consensus [10]. Also, max
consensus can be used to estimate the maximum and minimum
degrees of the network graph, which are useful in optimizing
consensus algorithms [17]. In [4], it is also mentioned that
max consensus and min consensus have a broad range of
applications in distributed decision-making for multi-agent
systems. In [11], max consensus is used to compensate for
clock drift and is used to time-synchronize wireless sensor
network nodes.
Index Terms—Max Consensus, Soft-max, Bounded Transmissions, Asymptotic Covariance, Adaptive Step Size.
To deal with the problem of finding a unique leader in a
group of agents in a distributed way, a max consensus problem
in a noise-free environment is proposed in [4], where each
node in the network collects data from all of its neighbors
and find the largest received data. At each iteration, after
comparing its own state and the largest received data, each
node updates its state with the max of the two values.
I. I NTRODUCTION
In sensor networks with no fusion center, consensus is a
process where all the sensors in the network achieve global
agreement using only local transmissions. There are many
advantages of using decentralized wireless sensor networks,
including robustness to link failures and scalability [2], [3].
Max consensus algorithms in the absence of noise are
considered in literature [4]–[9] with applications detailed in
[10], [11]. To the best of our knowledge, all past research
on max consensus has assumed no communication noise.
In the presence of noise, however, existing max consensus
approaches will suffer from divergence.
In this paper, we design, for the first time in the literature,
a fully distributed iterative algorithm to let nodes reach consensus on the maximum of the initial measurements in the
presence of noise. The results provide guidelines on how to
select a design parameter that balances convergence time with
estimator accuracy.
A. Literature Review on Existing Max Consensus Algorithms
The distributed consensus problem has a long history and
has attracted many researchers recently. It has broad applicaThe authors from Arizona State University are funded in part by the NSF
award ECCS - 1307982 and the SenSIP Center, School of ECEE, Arizona
State University. Banavar is funded in part by the NSF CRII award 1464222.
(Email: {szhang98, cihan, spanias}@asu.edu, [email protected]). Parts
of the material in this paper were presented at the 2013 Asilomar Conference
on Signals, Systems and Computers [1].
Max consensus algorithms using a similar approach as in
[4] are proposed in [5]–[9]. At each time step, every sensor
in the network updates its state with the largest measurements
it has recovered so far. Reference [7] considers both pairwise
and broadcast communications, and analyzes the convergence
time. A Max-plus algebra is used in [6] to analyze the max
consensus algorithm in a directed graph. Time dependent
graphs are considered in [8], where it is shown that strong
connectivity is required for reaching max consensus. A general
class of algorithms which can be used for both average and
min consensus algorithms is also mentioned in [5] by the
selection of a design parameter.
In [18], the authors extend the work of the weighted power
mean algorithm originally proposed by [19] and show that this
class of algorithms can also be used to calculate the maximum
of the initial measurements when the design parameter is
chosen to be infinity. A similar max approximation algorithm
is also mentioned in [15] to compute the maximum of the
initial measurements in a centralized sensor network with a
fusion center. Reference [18] also describes another distributed
coordination algorithm for max consensus. None of these
approaches addresses the issue of additive noise, or limited
transmit power in their setup.
2
B. Statement of Contributions
In wireless sensor networks, information exchange between
nodes usually involves communication noise and wireless
fading channels, therefore it is practical to assume that communications between sensor nodes is noisy [2], [20]. To the
best of our knowledge, all past research on max consensus
has assumed no communication noise. In the presence of
noise, if the max consensus algorithm at each node keeps the
largest received data as in [4]–[9], the estimated maximum
value at each node will drift incorrectly at each iteration
and states of nodes will not converge. The reason is that
by using the max operator at each node, the positive noise
samples will increase the maximum leading to divergence.
Even with channel coding, codeword over the channel may
cause divergence. Some max approximation functions are
considered in [5], [15], [18], [19] for max consensus. However,
the design parameters are chosen to be extreme values to get a
small estimator error, without considering its effect on transmit
power, or convergence time. If the design parameter is set at
extreme values, the the dynamic range of the state value will
be extremely large, which results in large transmit power if
linear consensus is used or slow convergence if non-linear
consensus algorithm is used.
The contribution of this paper is in both design and
analysis of a max consensus algorithm in wireless sensor
networks in the presence of communication noise. Regarding
design, the soft maximum, together with non-linear bounded
transmissions is proposed. The soft maximum of a vector
x := [x1 . . . xN ] is denoted as:
N
X
1
smax(x) = log
eβxi ,
β
i=1
(1)
where N is the number of elements in x and β > 0 is a design
parameter. The soft-max of x approximates the maximum
value of x for large β, and is equal to the maximum value
of x as β goes to infinity. The soft-max of x is always larger
than the maximum value of x and the difference between the
two values is less than β −1 log N . Note that if β < 0, then
(1) approximates the minimum value of x.
In the proposed max consensus algorithm, every sensor in
the network evaluates a function eβxi of its initial observation
xi and a non-linear average consensus algorithm such as those
in [3] can be used with a judicious choice of β. The non-linear
bounded transmission is used in the max consensus approach
because an accurate max estimation result can be obtained only
when the design parameter β is chosen to be large. A large
β may result in a very large transmit power if the traditional
linear consensus as in [17] is used. Since sensors typically have
low transmit power, the bounded transmission assumption is a
practical necessity. Note that since average consensus is used,
the proposed max consensus algorithm will work even when
the structure of the graph changes over time [2].
Regarding analysis, three sources of errors in the proposed
max consensus algorithm are presented. We show that the
parameter of the soft-max function that makes the soft-max
approximation accurate, also makes the convergence slow. The
technical novelty in the analysis is the analytical study of
this trade-off. By bounding the error between the convergence
result and the true max, we show that an approximation
of the needed convergence time can be calculated. The last
contribution of the paper is the introduction of the shifted nonlinear bounded function, which makes the convergence faster.
The analyses provide guidelines for nonlinear transmission
design, and algorithm parameter settings to trade-off between
estimation error and faster convergence.
C. Organization
The rest of this paper is organized as follows. In Section
II, the system model is provided. We describe the problem
statement in Section III, and prove that the sensors reach
consensus with the proposed solution. The error analysis
and convergence speed are also analyzed in Section III. In
Section IV, a shifted non-linear function is proposed for faster
convergence. In Section V, simulation results for all proposed
consensus algorithms are presented. Finally, concluding remarks are given in Section VI.
II. S YSTEM M ODEL
A. Graph Representation
Consider an undirected connected graph G = (N, E) containing a set of nodes N = {1, . . . , N } and a set of edges E.
The set of neighbors of node i is denoted by Ni , i.e., Ni =
{j|{i, j} ∈ E}. Two nodes can communicate with each other
only if they are neighbors. The number of neighbors of node
i is di . We use a degree matrix, D = diag[d1 , d2 , . . . , dN ],
which is a diagonal matrix containing the degrees of each
node. The connectivity structure of the graph is characterized
by the adjacency matrix A = {aij } such that aij = 1 if
{i, j} ∈ E and aij = 0 otherwise. The graph Laplacian of
the network L is defined as L = D − A. For a connected
graph, the smallest eigenvalue of the graph Laplacian is always
zero, i.e., λ1 (L) = 0 and λi (L) > 0, i = 2, · · · , N . The zero
eigenvalue λ1 (L) = 0 corresponds to the eigenvector with
all entries one, i.e. L1 = 0. The performance of consensus
algorithms often depends on λ2 (L), which is also known as the
algebraic connectivity [21]. Algebraic connectivity of simple
and weighted graphs are discussed in [22], where several upper
and lower bounds to λ2 (L) are also given.
B. Assumptions on Wireless Sensor Network Model
We have the following assumptions on the system model: i)
nodes in the distributed sensor network have their own initial
measurements, and the nodes do not know if they have the
maximum; ii) the communications in the network are synchronized, and at each iteration, nodes are broadcasting their
state values to their neighbors; iii) communications between
nodes is analog following [16], [17], [23] and is subject to
additive noise; and iv) each node updates its state based on
the received data.
3
soft-max method requires the parameter β to be large, which
can result in a large dynamic range of the mapped initial
measurements and large transmit power. Moreover, using a
linear transmit amplifier is power-inefficient. As a result, a
non-linear consensus (NLC) algorithm can be implemented
[3]. The consensus on the soft maximum is achieved by letting
each sensor map its state value at time t through a bounded
function h(·) before transmission to ensure bounded transmit
power. To describe the communications between nodes, we use
the standard Gaussian MAC so that each node receives a noisy
version of the superposition of the transmitted signal from its
neighbors. This is because the step sizes are the same across
different network links and there is no need to recover the
transmitted data separately. Consider the following algorithm
with additive noise at the receiver:
1
tanh(x)
0.8
√ x
1+x2
0.6
0.4
π
2
π arctan( 2 x)
u(x)
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
−5
−4
−3
−2
−1
0
x
1
2
3
4
5
Fig. 1: Bounded Transmission Functions.
III. M AX C ONSENSUS U SING
THE
S OFT- MAX
A. Problem Statement
Consider a wireless sensor network with N sensor
nodes, each with a real-valued initial measurement, xi , i =
1, 2, · · · , N . It is desired that the nodes reach consensus on the
maximum value of the initial measurements, xmax := maxi xi ,
under the assumption that the sensors have a single state
that they update based on local received measurements. Max
consensus in the absence of noise is straight forward: the nodes
update their states with the largest received measurement thus
far in each iteration. Consider the following algorithm at each
node:
xi (t+1) = max xi (t), max xj (t) , x̂max,i (t+1) = xi (t+1).
j∈Ni
(2)
However, in the presence of noise, such algorithms will diverge
due to positive noise samples, which can be shown in Figure 3
in Section V. An intuitive explanation is that and any positive
noise sample will always make the maximum larger if the max
operator is used in the max consensus algorithm.
Average consensus is well studied in literature. Existing
average consensus algorithms converge to the sample mean
of the initial measurements. As a result, the soft-max can be
used to calculate the maximum. To relate the soft-max to the
sample mean of {eβxi }, we have,
ȳ =
N
N
1 X
1 X βxi
e
=
yi (0),
N i=1
N i=1
(3)
where ȳ is the sample mean of the mapped initial measurements and yi (0) := eβxi . The quantity ȳ is computed
using an iterative distributed algorithm, in which each sensor
communicates only with its neighbors. If the states of all the
sensor nodes converge to ȳ, then the network is said to have
reached consensus on the sample average of the mapped initial
measurements. The relation between ȳ and the soft-max value
is given by
N
smax(x) =
X βx
1
1
log
e i = (log N + log ȳ).
β
β
i=1
(4)
The average consensus algorithms like in [2], [3] can be
used to achieve consensus in the sensor network. Sensors may
adopt either a digital or analog method for transmitting their
information to their neighbors. One such method is the linear
amplify-and-forward (AF) scheme in which sensors transmit
scaled versions of their measurements to their neighbors where
the iterative algorithm may be chosen as the linear consensus
algorithm of [2]. However, using the AF technique is not
a viable option for consensus on the soft-max. The reason
is that accurate approximation of the max value using the
h
i
X
yi (t+1) = yi (t) − α(t) di h(yi (t)) −
h(yij (t)) + ni (t) , (5)
j∈Ni
where i = 1, 2, . . . , N , and t = 0, 1, 2, . . ., is the time index.
The value yi (t + 1) is the state update of node i at time
t + 1, yij (t) is the state value of the j th neighbor of node
i at time t, and ni (t) is the additive noise at node i, which is
assumed to be independent across time and space with zero
mean and
σn2 . α(t) is P
a positive step size which
Pvariance
∞
∞
2
satisfies
t=0 α (t) < ∞ and
t=0 α(t) = ∞. The node
j transmits its information yij (t) by mapping it through the
non-linear function h(·) to constrain the transmitted power.
We assume that
√
h(x) =
γ u(ωx),
(6)
where u(x) is a normalized non-linear bounded function as in
Figure 1 and we make the following assumption on u(x):
Assumptions
(A1): u(0) = 0. u(x) = −u(−x).
(A2): max (u(x)) = 1.
(A3): The function u(·) is differentiable and invertible,
u′ (0) = 1 and 0 < du(x)
dx ≤ 1.
Parameters γ and ω are constants used to control the shape of
h(·). Note that invertibility of h(·) is needed for convergence,
however there is no need to apply the inverse of h(·) in
equation (5).
PNode i receives a noisy version of the superposition
j∈Ni h(yij (t)). The recursion in equation (5) can be expressed in vector form as,
y(t + 1) = y(t) − α(t)[Lh(y(t)) + n(t)],
(7)
T
where y(t) = [y1 (t) y2 (t) · · · yN (t)] and h(y(t)) =
[h(y1 (t)) h(y2 (t)) · · · h(yN (t))]T . L is the Laplacian matrix
of the graph and n(t) is the vector containing the additive
reception noise at nodes. Since the noise is i.i.d. with variance
σn2 , the covariance of n(t) is σn2 I. Since (5) converges to
a value that approximates (3), the consensus estimate of the
maximum at node i can be written using (4) as
x̂maxi (t∗ ) =
1
(log N + log yi (t∗ )) ,
β
(8)
where t∗ is the iteration at which the algorithm is stopped.
B. Proof of Convergence
Since the non-linear average consensus approach is used
in the max consensus algorithm, the convergence proof will
follow the proof in [3] which uses a discrete time Markov
process approach [24] (also see Theorem 5 in [3]). Therefore,
there exists a finite real random variable θ∗ such that,
h
i
Pr lim y (t) = θ∗ 1 = 1,
t→∞
(9)
4
where 1 is a column vector with all ones. Equation (9) shows
that the convergence is reached when t → ∞. The following
theorem characterizes the random variable θ∗ .
∗
∗
Theorem 1. θ in (9) is an unbiased estimate of ȳ, E[θ ] = ȳ.
Its mean square error ξN = E[(θ∗ − y)2 ], and is finite which
can be expressed as,
ξN =
∞
σn2 X 2
α (t).
N t=0
(10)
Proof: The proof is a straightforward adaptation of Theorem
3 in [3].
The nodes in the sensor network reach consensus on the
random variable θ∗ which is an unbiased estimator of the
average of the mapped initial measurements, E[θ∗ ] = ȳ. Then
the soft-max of the initial measurements can be obtained using
equation (4).
In the following subsection, we will analyze the performance of the system by considering the three sources of error
between the max estimate at each node x̂maxi (t∗ ) in (8), and
the true maximum of the initial measurements xmax .
C. Error Analysis
Let θ0 be a realization of θ∗ . From (9) we have that the
states of nodes in the sensor network are converging to θ0 as
t → ∞. However, in practice, we need to stop the algorithm
at a finite iteration time t∗ . There are three sources of error
between the true maximum xmax and x̂maxi (t∗ ) in (8): i)
(smax(x) − xmax ) = β1 (log N + log ȳ) − xmax , due to the
fact that soft-max approximation will always be larger than
the true max; ii) (θ0 − ȳ) caused by communication noise and
iii) (yi (t∗ ) − θ0 ) cause by finite number of iterations.
The size of these errors depend on β, N, t∗ and σn2 . In the
following subsection, we are going to characterize and analyze
these errors.
1) Soft-max error: This is a deterministic error which
depends on β, N , and the value of x. We have:
xmax ≤ smax(x) ≤ xmax +
1
log N,
β
(11)
Both inequalities are clearly tight for large β.
2) MSE of the algorithm: The second term (θ0 − ȳ) is
due to the presence of communication noise: the state of the
sensors does not converge to the sample mean of the mapped
initial measurements, instead it converges to a random variable
θ∗ whose expectation is the sample mean of the mapped initial
measurements, ȳ from (3). This occurs also in linear average
consensus in the presence of noise. The mean square error of
θ∗ is defined as ξN = E[(θ∗ − y)2 ] and is characterized as
(10) in Theorem 1. From (10), we
see that the mean square
P∞
error is finite and is small when t=0 α2 (t) or σn2 small.
3) Convergence speed: The third cause of error is due to a
finite number of iterations: even though limt→∞ y(t) = θ0 ,
y(t∗ ) 6= θ0 . However, with a judicious choice of nonlinear function h(·) and step size α(t), one can reduce
the convergence time. From now on, we
will assume that
P∞
a
2
,
a
>
0,
which
satisfies
α(t)
=
t=0 α (t) ≤ ∞,
t+1
P∞
= ∞. The convergence speed is analyzed by
t=0 α(t) √
establishing t(y(t)−θ0 1) is asymptotically normal with zero
mean and some covariance matrix C. The next theorem further
quantifies the convergence speed.
Theorem 2. Let 2aλ2 (L)h′ (θ0 ) > 1 so that the matrix
[ah′ (θ0 )B + I/2] is stable (every eigenvalue of the square
matrix has strictly negative real part) and I is the identity
matrix, and B is a diagonal matrix containing all the nonzero eigenvalues of −L. Define U = [N −1/2 1 Φ] which is a
unitary matrix whose columns are the eigenvectors of L. Let
[ñ(t) ñ(t)] = N −1 UT n(t) and Cñ = E[ññT ] is a diagonal
matrix, Cñ ∈ R(N −1)×(N −1) . Then as t → ∞,
√
t(y(t) − θ0 1) ∼ N (0, C),
(12)
where the asymptotic covariance matrix C = N −1 a2 σn2 11T +
N −1 ΦSθ0 ΦT ,R
′
′
∞
and Sθ0 = a2 0 e(ah (θ0 )B+I/2)t Cñ e(ah (θ0 )B+I/2)t .
The proof is the same as given in Theorem 5 in [3].
The convergence speed is quantified by kCk, which is
defined to be the largest eigenvalue of the covariance matrix.
We show in Appendix that the l2 norm of the covariance matrix
can be expressed as
kCk = max xT Cx
kxk≤1
a2 σn2
1
.
= max a2 σn2 ,
N 2ah′ (θ0 )λ2 (L) − 1
(13)
This norm, kCk, can be optimized with respect to a, and the
′
value that minimizes kCk is a∗ = (N + 1)/[2N λ2 (L)h (θ0 )].
The optimal value for the l2 norm of the covariance matrix
denoted as kC∗ k can be represented as
2
1
h (θ0 )
2
2
2
N +1
σn
1
=
,
2N
λ22 (L)γ
ωu′ (ωθ0 )
kC∗ k =
N +1
2N
2
σn2
λ22 (L)
′
(14)
which is proved in Appendix. The interpretation is that convergence is slower when kC∗ k is larger.
It is clear form equation (14) that convergence will be fast
if λ2 (L) large, which implies faster convergence in a more
′
connected graph. Also the value of kC∗ k decreases as h (θ0 )
increases, which shows that the convergence speed depends
on the non-linear function and the convergence point.
By observing the three sources of error mentioned above,
we find there is a trade-off between the convergence speed
and the soft-max error. To see this, recall that the convergence
speed is quantified by kC∗ k. From the analysis of sources
of error, choosing a larger β would reduce the deterministic
bias caused by the soft-max mapping (11), but degrades the
variance term in (14). The reason is that h(·) is chosen to
be an odd bounded transmission function as in Figure 1 with
a zero-crossing and steepest slope at the origin, with h′ (x)
′
decreasing for x ≥ 0. Since θ0 ≥ 0, h (θ0 ) will be small when
θ0 gets larger which increases the value of kC∗ k and makes the
convergence slower. The convergence point
be large
P θ0βxwill
i
.
Therefore
e
when β is chosen large since θ0 ≈ N1
i
a trade-off between the convergence speed and the soft-max
error exists: a more accurate soft-max can be obtained by
choosing a large β, but this degrades the convergence speed.
5
p( xmax −1)
ε2
√
ε−2 N
= −p γ (log N ) xmax 2 xmax
2
p( ε
−1)
2
N
+1
√
p γ (log N ) xmax
≈−
.
−1)
p( xmax
ε2
ε22 N
D. Bound on Convergence Time
The convergence speed of the max-consensus algorithm is
quantified by the asymptotic covariance matrix. If some prior
knowledge about the distribution of the initial measurements
is known, the step size can be set based on the expression of
′
a∗ = (N +1)/[2N λ2 (L)h (θ0 )] and α(t) = a∗ /(t+1) to make
the convergence fast. In this section, we assume that the step
size is set to be a∗ as mentioned. The trade-off controlled by
β balances soft-max error and convergence speed. How much
time t∗ is needed for the nodes to reach consensus is always
an important problem. In this subsection, we will show that
by upper bounding the three sources of error in Section III-C,
an approximation on the iteration time for reaching consensus
can be calculated.
The estimate of the max at iteration time t∗ is expressed
as (8), where yi (t∗ ) is the state at node i at time t∗ . Of the
three errors in Section III-C, note that the error (θ0 − ȳ) can be
ignored when the noise variance σn2 is small, or can be reduced
by running the consensus several times and taking the average
of the results. In the following, we ignore the error (θ0 − ȳ)
and calculate the iteration time by bounding the soft-max error
denoted by ε2 and error caused by a finite stopping time t∗
denoted by ε1 . When a = a∗ , the norm of the asymptotic
covariance matrix of (y(t∗ ) − θ0 1) can be bounded by ε1 if
∗
∗
kC k
kC k
.
≤ ε 1 ⇒ t∗ ≥
t∗
ε1
(15)
The soft-max error is bounded by bounding the upper bound
in equation (11), which can be expressed as:
log N
log N
.
≤ ε2 ⇒ β ≥
β
ε2
(16)
By substituting (16) into kC∗ k, a lower bound of the iteration
time needed for reaching consensus can be calculated using
(15):
t∗ ≥
=
∗
kC k
=
ε1
N +1
2N
N +1
2N
2
σn2
2
λ2 (L)
2
σn2
λ22 (L)
′
1
h (N
1
P
βxi )
ie
ε1
2 2
N +1
σn
1
≥
2N
λ22 (L)
ε1
1
h (θ0 )
′
ε1
2
2
1
h′
1
N
P
i
e
log N
xi
ε2
2
.
(17)
The last inequality holds because of (16) and using that h′ (x)
is decreasing function when x > 0. We now study how the
final lower bound depends on the convergence error, ε1 , and
soft-max error, ε2 . It is clear that the bound is inversely related
to ε1 . How ε2 affects the bound depends on the choice of h(·).
In the following, we provide two examples of h(·) and
show that how equation (17) is affected by ε2 . We will
consider two cases. In the first case, h(x) converges to its
maximum value polynomially fast and in the second case it
converges exponentially.
First
consider the polynomial case
√
and let h(x) ≈ γ 1 − xp1+1 for large x > 0 for p > 0.
Note that p controls the value of ω in the definition of h(x)
in (6). Then,
′
h
1 X logε N xi
e 2
N i
!
≈ h′
1 logε N xmax
e 2
N
(18)
(19)
(20)
Equation (18) holds since when ε2 is small, the term xi =
xmax dominates. Equation (20) shows
how ε2 affects the
√
convergence time when h(x) ≈ γ 1 − xp1+1 for large
x > 0 for p > 0.The asymptotically
optimal p that minimizes
−
1
, and the lower bound
(17) is p∗ = 1/ (log N ) xmax
ε2
of the iteration time needed can be calculated as,
2
2 2 ε4 e2 xmax − 1
2
ε
N
+1
σ
1
2
n
t∗ ≥
.
2N
λ22 (L)
ε1
γx2max
(21)
√
On the other hand, if h(x) converges to its final value, γ,
√
−qx
exponentially fast, we have h(x) ≈ γ (1 − e
) for large
x > 0, with q > 0 which controls the value of ω in the
definition of h(x) in (6). Then,
′
h
1 X logε N xi
e 2
N i
!
≈ h′
1 logε N xmax
e 2
N
xmax
√ log N
ε2
e−qN
xmax ε−2
= −q γ
2 N
N
=−
√
q γ (log N ) xmax N
ε2 eqN
xmax
−1
ε
(22)
xmax
−1
ε2
!
(23)
2
xmax
−1
ε2
.
(24)
The
optimal q that minimizes (17) is q ∗ =
asymptotically
1− xmax
ε2
N
, and the lower bound of the iteration time needed
can be calculated as,
t∗ ≥
N +1
2N
2
σn2
2
λ2 (L)
1
ε1
ε42 e2
2
γxmax (log N )2
.
(25)
When choosing the non-linear bounded function as mentioned above, we have the following observations based on
equation (17), (20) and (24): (i) the required convergence
time will be longer when the error requirements ε1 and ε2
are smaller; (ii) the soft-max error term ε2 dominates the
convergence time in equation (17) in both examples; (iii) the
required convergence time will be longer when xmax or the
system size N is larger; (iv) by comparing equation (21)
and (25), the convergence will be faster when choosing h(·)
that converges to its maximum value exponentially fast is
appropriate if
2
xmax
− 1 > (log N )−2 ,
ε2
(26)
and h(·) that converges to its maximum value polynomially
should be chosen otherwise.
Finally, note that estimating the minimum value of the
local measurements is also sometimes necessary. The minconsensus can be achieved using the similar initial mapping
but choosing β < 0.
IV. S HIFTED N ON - LINEAR B OUNDED F UNCTION U SED
M AX C ONSENSUS
IN
An accurate max estimation using the soft-max approach
requires the design parameter β to be large. As a result, the
6
1
250
0.9
0.8
State of Sensor Nodes
200
0.7
0.6
0.5
0.4
0.3
0.2
150
100
50
0.1
0.6
0.8
0
1
Fig. 2: Graph Representation of the Sensor Network, N = 75.
exponential function used for initial measurements mapping
expands the dynamic range of the initial measurements and the
convergence speed is slow. We now give a modified non-linear
distributed average consensus method using a shifted nonlinear bounded function that can make the convergence process
faster if some prior knowledge of the initial measurements is
available.
The method is based on′ the fact that the convergence is
faster when the value of h (θ0 ) is large from equation (14).
h(·) in the iterative algorithm is replaced by a shifted nonlinear function g(·) defined as g(x) = h(x − T ), where
T is a shift constant. In this case, the
optimal step
size
+1
N +1
1
1
a∗s = N2N
=
λ2 (L)g′ (θ0 )
2N
λ2 (L)h′ (θ0 −T ) . The
convergence speed is quantified by the norm of the asymptotic
covariance matrix kC∗s k and can be expressed as,
kC∗s k =
N +1
2N
2
σn2
2
λ2 (L)
0
1
2
3
4
5
6
7
8
9
10
Iterations, t
1
h′ (θ0 − T )
2
.
(27)
′
From (27), convergence will be faster when h (θ0 − T ) is
′
larger. h (θ0 − T ) reaches its largest value when T = θ0
if h(·) is chosen as a sigmoid function with steepest slope
at origin. Note that θ0 is unknown in practice, but one can
use prior information on the initial measurements to choose
T . If the distribution of the initial measurements is known
at the sensor nodes, a reasonable choice of T is to choose
it as the expected P
value of the mapped initial measurements:
N
T = E[eβxi ] ≈ N1 i=1 eβxi .
V. S IMULATIONS
In this section, simulation results for max consensus algorithms are presented. Different β values are used to trade-off
between convergence speed and error between the proposed
approach and the true max of the initial measurements.
A. Performance of Max Consensus
In the max consensus simulations, the initial measurements
{xi } are chosen to be uniformly distributed over (0, 1) in
Figure 3 and 4. Gaussian noise with zero mean and unit
variance is added to receiver nodes. The graph of the sensor
network is fixed for all simulations with N = 75 sensors. as
shown in Figure 2. In Figure 3, the traditional max consensus
algorithm is used and each node always keeps the largest
Fig. 3: Entries of traditional max consensus result versus
iterations t (Keep the largest measurement at each iteration).
2
+ log(yi (t))), i = 1, 2, . . . , N
0.4
1
β (log N
0.2
1.8
1.6
1.4
1.2
1
xmax
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50
60
70
80
90
100
Iterations, t
(a) Entries of the consensus
soft max result versus iterations t, β =
√
5, ω = 0.015, h(x) = γ tanh(ωx), α(t) = 4.4473
, a∗ ≈ 4.4473.
t+1
1.8
+ log(yi (t))), i = 1, 2, . . . , N
0
1
β (log N
0
1.6
1.4
1.2
1
xmax
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50
60
70
80
90
100
Iterations, t
(b) Entries of the consensus soft max result versus iterations t, β
√
= 7, ω = 0.015, h(x) = γ tanh(ωx), α(t) = 61.7513
, a∗ ≈
t+1
61.7513.
Fig. 4: Simulation Results for Max Consensus.
measurement from its neighbors. We can see from Figure 3
that in the presence of noise, the states of nodes will diverge.
In Figure 4, x̂maxi (t∗ ) from equation (8) for all nodes are
plotted to illustrate the convergence of the soft-max result for
different β and a values. Note that the actual maximum value
is 0.9561 in the simulations. In each of Figures 4, h(x) =
√
γ tanh(ωx), with ω = 0.015 and γ = 7.5dB, note that γ
controls the peak transmit power and ω controls the shape of
h(x); β is 5 in Figure 4(a) and 7 in Figure 4(b). a is chosen as
(N +1)/[2N λ2 (L)h′ (ȳ)] ≈ a∗ and the following observations
can be made by comparing the two figures: (i) As β increases,
7
1.2
1
xmax
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50
60
70
80
90
+ log(yi (t))), i = 1, 2, . . . , N
1.6
1.4
1
β (log N
1
β (log N
+ log(yi (t))), i = 1, 2, . . . , N
1.4
0.4
1.2
1
xmax
0.8
0.6
0.2
0
100
0
10
20
30
B. Performance of Max Consensus with Shifted Non-linear
Bounded Function
In the simulation of max consensus using a shifted nonlinear bounded function, the initial measurements {xi } are
chosen to be uniformly distributed over (0, 1). Gaussian noise
with zero mean and unit variance is added to the receiver
nodes, γ = 7.5dB in all simulations. The graph is the same
as the max consensus simulation with N = 75 sensors. In
Figure 6, x̂maxi (t∗ ) for all the nodes are plotted. In each of
Figures, a = 12, ω = 0.015 and β is 7. In Figure 6(b), shifted
non-linear bounded functions are used in transmission, and
T is chosen to be the sample mean of the mapped initial
measurements. By comparing Figure 6(a) and 6(b), it is shown
that using the shifted non-linear bounded function can improve
the convergence speed.
VI. C ONCLUSIONS
A practical approach for reliable computation of the maximum value of local measurements over autonomous sensor
networks with no fusion center is proposed. The trade-off between estimation accuracy and convergence time is quantified.
It is proved that the sensor network will reach consensus,
that is, the state values converge to a random variable whose
expectation is the sample mean of the mapped function, and
the soft-max can be calculated using the consensus result. The
shifted non-linear function used to adjust the transmit nonlinearity is also introduced to make the convergence speed faster.
The results provide guidelines towards nonlinear transmission
design, and algorithm parameter settings to trade-off between
estimation error and faster convergence.
60
70
80
90
100
1.8
+ log(yi (t))), i = 1, 2, . . . , N
the estimates of the soft-max of x are closer to the actual
value of the maximum value of x and (ii) As β increases, the
convergence is slower, which matches the result in equation
(14). In Figure 5, an accurate max estimate is obtained by set
β = 30. It is shown that by properly choosing the non-linear
function h(·) and step size a, an accurate max consensus can
be attained within a few iterations. From Figure 5, we can
see that the error between the convergence result and the true
max is around 0.06, therefore, Figure 5 can be a recommended
solution for max estimation in sensor networks.
50
(a) Entries of the consensus soft max
√ result versus iterations t, β
= 7, ω = 0.015, N = 75, h(x) = γ tanh(ωx), α(t) = 12/(t+1).
1
β (log N
Fig. 5: Entries of the consensus soft max result versus
√
iterations t, β = 30, ω = 10−11 , h(x) = γ tanh(ωx),
10
, a∗ ≈ 5.03 × 1010 .
α(t) = 5.03×10
t+1
40
Iterations, t
Iterations, t
1.6
1.4
1.2
1
xmax
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50
60
70
80
90
100
Iterations, t
(b) Entries of the consensus soft
√ max result versus iterations t, β =
7, ω = 0.015, N = 75, h(x) = γ tanh(ω(x−T )), T = 138.1045,
α(t) = 12/(t+1).
Fig. 6: Simulation Results for Max Consensus Using Shifted
Non-linear Bounded Functions.
A PPENDIX
The Convergence will be slow when kCk is large, where
kCk is the max eigenvalue of C. The problem can be
formulated as,
kCk =
max
{x|x∈RN kxk≤1}
xT Cx.
(28)
Let U = [ √1N Φ], the columns of U are the eigenvectors
of L. Since L is an Hermitian matrix, the columns of U form
an orthonormal basis of RN . Let x = Uz with kzk ≤ 1, we
have
xT Cx = (Uz)T C(Uz)
2 2 T T
a σn U 11 U
UT ΦSθ0 ΦT U
= zT
+
z
N
N
= zT {A1 + A2 } zT = zT A3 zT ,
(29)
(30)
(31)
where
A1
=
diag a2 σn2 , 0, · · · , 0 N ×N ,
1
A2
=
diag 0, N S1,1 , · · · , N1 Sn−1,n−1 N ×N and
a2 σ 2
n
,
i = 1, 2, · · · , N − 1.
Si,i = 2ah′ (θ0 )λi+1
2 2 1 (L)−1
A3 = diag a σv , N S1,1 , · · · , N1 Sn−1,n−1 N ×N . Equality
in (31) holds since the columns of Φ are orthogonal to 1 and
Sθ0 is a diagonal matrix which can be calculated as,
Sθ0 = a2
Z
∞
0
e[ah (θ0 )B+I/2]t Ce[ah (θ0 )B+I/2]t dt
′
′
(32)
8
Z ∞
= a2 σn2
eHt dt
(33)
0
a2 σn2
a2 σn2
= diag
,··· ,
, (34)
′
′
2ah (θ0 )λ2 (L) − 1
2ah (θ0 )λN (L) − 1
where H is an (N − 1) × (N − 1) diagonal matrix and
Hi,i = 2ah′ (θ0 )λi+1 (L) − 1. Note that (34) holds under the
assumption that 2ah′ (θ0 )λi (L)−1 > 0 for all i, which is same
as the requirement in Theorem 5 in [3] that [ah′ (θ0 )B + I/2]
is stable.
Since λ2 (L) is the smallest non-zero eigenvalue, we have
a2 σn2
a2 σn2
1
1
≥
, for i > 2. (35)
N 2ah′ (θ0 )λ2 (L) − 1
N 2ah′ (θ0 )λi (L) − 1
Therefore, from equation (31), (34) and (35), we get,
max
xT Cx
a2 σn2
1
2 2
.
= max a σn ,
N 2ah′ (θ0 )λ2 (L) − 1
kCk =
{x|x∈RN kxk≤1}
(36)
In the following, the optimal kC∗ k is calculated together
with the corresponding optimal a = a∗ .
kC∗ k =
=
min
max
{a|2ah′ (θ0 )λ2 (L)>1} {x|x∈RN kxk≤1}
xT Cx
a2 σn2
2 2 1
max
a
σ
,
. (37)
n
{a|2ah′ (θ0 )λ2 (L)>1}
N 2ah′ (θ0 )λ2 (L)−1
min
We noticed that the first term a2 σn2 in equation (37) is a
monotonic increasing function of a. The monotonicity for the
second term can be check by taking the derivative respect to a,
it is easy to check that the term is decreasing if 2h′ (θ01)λ2 (L) <
1
1
, and the term is increasing if a > h′ (θ0 )λ
.
a ≤ h′ (θ0 )λ
2 (L)
2 (L)
∗
By checking the value of kC k for marginal a, we find that
the problem in equation (37) is solved by letting,
a
2
σn2
=
1
N
a2 σn2
′
2ah (θ0 )λ2 (L) − 1
.
(38)
.
(39)
By solving equation (38), we get,
a = a∗ =
N +1
2N
1
λ2 (L)h′ (θ0 )
1
. Plug
It is easy to check that 2h′ (θ01)λ2 (L) < a∗ ≤ h′ (θ0 )λ
2 (L)
the optimal a∗ into the expression of kCk, the corresponding
optimal value, kC∗ k is given by,
kC∗ k =
N +1
2N
2
σn2
λ22 (L)
1
h (θ0 )
′
2
.
(40)
R EFERENCES
[1] S. Zhang, C. Tepedelenlioglu, M. Banavar, and A. Spanias, “Maxconsensus using the soft maximum,” in 2013 Asilomar Conference on
Signals, Systems and Computers, Nov. 2013, pp. 433 – 437.
[2] S. Kar and J. Moura, “Distributed consensus algorithms in sensor
networks with imperfect communication: Link failures and channel
noise,” IEEE Transactions on Signal Processing, vol. 57, no. 1, pp.
355 –369, Jan. 2009.
[3] S. Dasarathan, C. Tepedelenlioglu, M. Banavar, and A. Spanias, “Nonlinear distributed average consensus using bounded transmissions,” IEEE
Transactions on Signal Processing, vol. 61, pp. 6000 – 6009, Dec. 2013.
[4] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Transaction on
Automatic Control, vol. 49, no. 9, pp. 1520 – 1533, Sept. 2004.
[5] A. Tahbaz-Salehi and A. Jadbabaie, “A one-parameter family of distributed consensus algorithms with boundary: From shortest paths to
mean hitting times,” in 45th IEEE Conference on Decision and Control,
Dec. 2006, pp. 4664 – 4669.
[6] B. Nejad, S. Attia, and J. Raisch, “Max-consensus in a max-plus
algebraic setting: The case of fixed communication topologies,” in International Symposium on Information, Communication and Automation
Technologies, Oct. 2009, pp. 1–7.
[7] F. Iutzeler, P. Ciblat, and J. Jakubowicz, “Analysis of max-consensus
algorithms in wireless channels,” IEEE Transactions on Signal Processing, vol. 60, pp. 6103 – 6107, Nov. 2012.
[8] G. Shi and K. H. Johansson, “Finite-time and asymptotic convergence of
distributed averaging and maximizing algorithms,” arXiv:1205.1733v2,
Sept. 2012.
[9] S. Giannini, D. D. Paola, A. Petitti, and A. Rizzo, “On the convergence
of the max-consensus protocol with asynchronous updates,” in 52nd
IEEE Conference on Decision and Control, Dec. 2013.
[10] Z. Li, R. Yu, and M. Huang, “A distributed consensus-based cooperative
spectrum-sensing scheme in cognitive radios,” IEEE Transactions on
Vehicular Technology, vol. 59, no. 1, pp. 383 – 393, Jan. 2010.
[11] D. Zhao, Z. An, and Y. Xu, “Time synchronization in wireless sensor
networks using max and average consensus protocol,” International
Journal of Distributed Sensor Networks, Feb. 2013.
[12] P. A. and A. Jadbabaie, “Synchronization in oscillator networks: Switching topologies and non-homogeneous delays,” in 44th IEEE Conference
on Decision and Control, 2005 and 2005 European Control Conference,
Dec. 2005, pp. 5692 – 5697.
[13] R. Olfati-Saber, “Flocking for multi-agent dynamic systems: algorithms
and theory,” IEEE Transaction on Automatic Control, vol. 51, no. 3, pp.
401 – 420, March 2006.
[14] W. Ren, R. W. Beard, and E. M. Atkins, “A survey of consensus
problems in multi-agent coordination,” in 2015 American Control Conference, June, 2005, pp. 1859 – 1864.
[15] M. Goldenbaum, S. Stanczak, and M. Kaliszan, “On function computation via wireless sensor multiple-access channels,” in Wireless
Communications and Networking Conference, April 2009, pp. 1 – 6.
[16] R. Olfati-Saber, J. Fax, and R. Murray, “Consensus and cooperation in
networked multi-agent systems,” IEEE Proc., vol. 95, no. 1, pp. 215–
233, Jan. 2007.
[17] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,”
in Proc. 42nd IEEE Conference on Decision and Control, vol. 5, Dec.
2003, pp. 4997 – 5002.
[18] J. Cortes, “Distributed algorithms for reaching consensus on general
functions,” Automatica, vol. 44, no. 3, pp. 401 – 420, 2008.
[19] D. Bauso, L. Giarre, and R. Pesenti, “Nonlinear protocols for optimal
distributed consensus in networks of dynamic agents,” System and
Control Letters, vol. 55, no. 11, pp. 918 – 928, June 2006.
[20] M. Huang and J. Manton, “Stochastic Lyapunov analysis for consensus
algorithms with noisy measurements,” in American Control Conference,
July 2007, pp. 1419 – 1424.
[21] R. Olfati-Saber and R. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Transactions on
Automatic Control, vol. 49, no. 9, pp. 1520 – 1533, Sept. 2004.
[22] N. M. M. de Abreu, “Old and new results on algebraic connectivity of
graphs,” Linear Algebra and its Applications, vol. 423, pp. 53–73, May.
2007.
[23] R. Olfati-Saber, “Ultrafast consensus in small-world networks,” in Proceeding of the 2005 American Control Conference, June 2005, pp. 2371
– 2378.
[24] M. Nevelson and R. Khasminskil, Stochastic Approximation and Recursive Estimation. Amer. Math Soc., 1973.
| 3cs.SY
|
A general estimator for the right endpoint
with an application to supercentenarian women’s records
∗
arXiv:1412.3972v3 [math.ST] 16 Jun 2016
Isabel Fraga Alves (CEAUL, University of Lisbon )
Cláudia Neves (University of Reading, UK )
Pedro Rosário (CEAUL, University of Lisbon )
Abstract
We extend the setting of the right endpoint estimator introduced in Fraga Alves and
Neves (Statist. Sinica 24:1811–1835, 2014) to the broader class of light-tailed distributions
with finite endpoint, belonging to some domain of attraction induced by the extreme
value theorem. This stretch enables a general estimator for the finite endpoint, which
does not require estimation of the (supposedly non-positive) extreme value index. A new
testing procedure for selecting max-domains of attraction also arises in connection with
asymptotic properties of the general endpoint estimator. The simulation study conveys
that the general endpoint estimator is a valuable complement to the most usual endpoint
estimators, particularly when the true extreme value index stays above −1/2, embracing
the most common cases in practical applications. An illustration is provided via an
extreme value analysis of supercentenarian women data.
KEY WORDS AND PHRASES: Extreme value theory, Semi-parametric estimation, Tail
estimation, Regular variation, Monte Carlo simulation, Human lifespan
1
Introduction
The extreme value theorem (with contributions from Fisher and Tippett, 1928; Gnedenko,
1943; de Haan, 1970) and its counterpart for exceedances above a threshold (Balkema and
de Haan, 1974) ascertain that inference about rare events can be drawn on the larger (or
lower) observations in the sample. While restricting attention to the large rare events,
the theoretical framework provided by the extreme value theorem reads as follows. If
a non-degenerate limit G is achieved by the distribution function (d.f.) of the partial
maxima Xn,n of a sequence {Xn }n≥1 of independent and identically distributed (i.i.d.)
random variables (r.v.) with common d.f. F , and if there exist an > 0 and bn ∈ R such
that limn→∞ F n (an x + bn ) = G(x), for every continuity point of G, then G is one of the
three distributions
Λ(x) = exp{− exp(−x)},
−α
Φα (x) = exp{−x
x ∈ R,
}, x > 0,
α
Ψα (x) = exp{−(−x) }, x < 0,
∗
α > 0,
(1)
(2)
α > 0.
Funded by FCT - Fundação para a Ciência e a Tecnologia, Portugal, through the project
UID/MAT/00006/2013
1
Introduction
These can be nested in the Generalized Extreme Value (GEV) d.f.
Gγ (x) := exp{−(1 + γx)−1/γ }, 1 + γx > 0, γ ∈ R.
(3)
We then say that F is in the (max-)domain of attraction of Gγ , for some extreme value
index (EVI) γ ∈ R [notation: F ∈ D(Gγ )]. For γ = 0, the right-hand side of (3) is read
as exp (−e−x ). The theory of regular variation (Bingham et al., 1987; de Haan, 1970;
de Haan and Ferreira, 2006), provides necessary and sufficient conditions for F ∈ D(Gγ ).
Let U be the tail quantile
function defined by the generalized inverse of 1/(1 − F ), i.e.
U (t) := F ← 1 − 1/t , for t > 1. Then, F ∈ D(Gγ ) if and only if there exists a positive
measurable function a(·) such that the condition of extended regular variation
U (tx) − U (t)
xγ − 1
=
,
t→∞
a(t)
γ
lim
(4)
holds for all x > 0 [notation: U ∈ ERVγ ]. The limit in (4) coincides with the U -function of
the Generalized Pareto distribution (GPD), with distribution function 1 + log Gγ . Hence,
for extrapolating beyond the range of the available observations, the statistics of extremes
will be exclusively focused on those observations over a sufficiently high threshold. Then
the excesses above this threshold are expected to behave as observations drawn from the
GPD.
The right endpoint of the underlying distribution function F is defined as
xF := sup{x : F (x) < 1} ≤ ∞,
which in terms of high quantiles is given by xF = limt→∞ U (t) = U (∞). For estimating
the right endpoint xF we will follow a semi-parametric approach, that is, our focus is
on the domain of attraction rather than on the limiting GEV distribution. We will also
assume that k is an intermediate sequence of positive integers k = kn such that k → ∞
and k/n → 0, as n → ∞. This is our large sample assumption for the moment. Other
mild yet reasonable conditions in the context of extreme value estimation will come forth
in section 3, which essentially convey suitable bounds on the intermediate sequence kn .
This paper deals with a unifying semi-parametric approach to the problem of estimating the finite right endpoint xF when F belongs to some domain of attraction where a
finite endpoint is admissible, more formally F ∈ D(Gγ )γ≤0 . We term this estimator x̂F the
general endpoint estimator. We will provide evidence that despite x̂F not being asymptotically normal for all values of γ < 0 (a drawback if one wishes to construct confidence
intervals) it proves nonetheless to be a valuable tool in terms of applications. One of the
most obvious estimators of the right endpoint is the sample maximum. In fact, de Haan
and Ferreira (2006) point out in their Remark 4.5.5 that using the sample maximum Xn,n
to estimate xF in case γ < −1/2 is approximately equivalent to using the moment related
estimator for the endpoint. The striking feature of the general endpoint estimator is that
it avoids the nuisance of changing “tail estimation-goggles” each time we are dealing with
yet another sample, possibly from a distribution in a different domain of attraction. We
exemplify this point by referring the study by Einmahl and Magnus (2008), which could
well benefit from using the same endpoint estimator at all instances, in all athletic events.
This freedom of constraint about γ ≤ 0 motivates the present general estimator, alongside
with its preceding application to the long jump records in (Fraga Alves et al., 2013).
The outline of the paper is as follows. Section 2 introduces the general estimator
F
x̂ and its theoretical assumptions, aligned with the usual semi-parametric framework.
Large sample results for x̂F are presented in Section 3, as well as a new test statistic
2
Semi-parametric approach to endpoint estimation
based on x̂F aimed at selecting max-domains of attraction. Section 4 is dedicated to
a comparative study via simulation, involving common parametric and semi-parametric
inference approaches in extremes. Section 5 provides an illustration of the exact behaviour
of the general endpoint estimator, using the supercentenarian women data set. Here we
consider two alternative settings: estimation of the right endpoint with a link to the EVI
estimation, estimation of the endpoint when this link to the EVI is broken and finally in
Section 6, we list several concluding remarks. Appendix 7 encloses all the proofs of the
large sample results in Section 3 and Appendix 8 encompasses the finite sample properties
for a consistent reduced bias estimator of the endpoint, for an EVI in (−1/2, 0), being
compared with POTML and Moment methodologies.
2 Semi-parametric approach to endpoint estimation
We now introduce some notation. Let F be the d.f. of the r.v. X and X1,n ≤ X2,n ≤ . . . ≤
Xn,n be the n-th ascending order statistics (o.s.) associated with the sample X1 , . . . , Xn
of n i.i.d. copies of X. We assume F ∈ D(Gγ ), for some γ ≤ 0, and that xF < ∞.
Several estimators for the right endpoint xF of a light-tailed distribution attached to
an EVI γ < 0 are available in the literature (e.g. Hall, 1982; Cai et al., 2013; de Haan
and Ferreira, 2006). These estimators often bear on the extreme value condition (4) with
x = x(t) → ∞, as t → ∞: since γ < 0 entails that limt→∞ U (t) = U (∞) exists finite,
then relation (4) rephrases as
U (∞) − U (t)
1
=− .
t→∞
a(t)
γ
lim
A valid estimator for the right endpoint xF = U (∞) thus arises by making t = n/k in the
approximate equality U (∞) ≈ U (t) − a(t)/γ , replacing U (n/k), a(n/k) and γ by suitable
consistent estimators, i.e.
n â nk
∗
x̂ = Û
−
k
γ̂
(cf. Section 4.5 of de Haan and Ferreira, 2006). Typically we consider the class of endpoint
estimators
â nk
∗
x̂ = Xn−k,n −
.
(5)
γ̂
There is however one estimator for the right endpoint xF that does not depend on the
estimation of the EVI γ. This estimator, introduced in Fraga Alves and Neves (2014),
is primarily tailored for distributions with finite right endpoint in the Gumbel domain of
attraction. The study of consistency and asymptotic distribution of this same endpoint
estimator is the main objective in this paper, while it aims at covering the whole scenario
in extremes, thus providing a unified estimation procedure for the right endpoint in the
case of γ ≤ 0.
The general right endpoint estimator from Fraga Alves and Neves (2014) is defined as
k−1
x̂F := Xn,n + Xn−k,n −
1
1 X
log 1 +
Xn−k−i,n .
log 2
k+i
(6)
i=0
With ai,k := log
k+i+1
k+i
/ log 2, the endpoint estimator x̂F in (6) can be expressed in the
3
Endpoint estimation and testing
equivalent form
F
x̂ := Xn,n +
k−1
X
ai,k (Xn−k,n − Xn−k−i,n )
with
k−1
X
i=0
ai,k = 1.
(7)
i=0
From the non-negativeness of the weighted spacings in the sum in (7), it is clear that
estimator x̂F is greater than Xn,n , which constitutes a major advantage to the usual semiparametric right endpoint estimators in the Weibull max-domain of attraction. Therefore, the estimator x̂F defined in (6) can be seen as a real asset in the context of semiparametric estimation of the finite right endpoint, embracing all distributions connected
with a non-positive EVI γ, which gains by far a broader spectrum of application to the
usual alternatives.
3
Endpoint estimation and testing
This section contains the main results of the paper, giving accounts of strong consistency
and sometimes asymptotic normality (we will see that the limiting normal distribution
is only attained if γ < −1/2) of the general endpoint estimator x̂F defined in (6). A
second order reduced bias version of x̂F is also devised. Additionally, we provide the
asymptotic framework for a statistical test aimed at discriminating between max-domains
of attraction. The new test statistic builds on the general endpoint estimator x̂F . All the
proofs are postponed to the Appendix 7.
Proposition 1. Suppose xF exists finite. Assume that the extended regular variation
property (4) holds with γ ≤ 0. If k = kn → ∞, kn /n → 0, as n → ∞, then the following
almost sure convergence holds with respect to x̂F defined in (6):
a.s.
x̂F −→ xF ,
n→∞
p
then x̂F is a consistent estimator for xF < ∞, i.e. x̂F −→ xF .
n→∞
Note that if F ∈ D(Gγ ) with γ > 0, then x̂F converges almost surely to infinity.
We now require a second order refinement of condition (4) and auxiliary second order
conditions in order to have a grasp at the speed of convergence in (4). In particular, we
assume there exists a positive or negative function A0 with limt→∞ A0 (t) = 0 such that
for each x > 0,
γ
U (tx)−U (t)
− x γ−1
a0 (t)
lim
= Ψ?γ,ρ (x),
(8)
t→∞
A0 (t)
where ρ is a non-positive parameter and with
γ+ρ
x
−1
,
γ+ρ
log x,
Ψ?γ,ρ (x) :=
1 γ
γ x log x,
1 (log x)2 ,
2
γ + ρ 6= 0, ρ < 0,
γ + ρ = 0, ρ < 0,
ρ = 0 6= γ,
γ = ρ = 0,
a(t) 1 − A0 (t) ,
a(t) 1 − A0 (t)/γ ,
a0 (t) :=
a(t),
4
ρ < 0,
ρ = 0 6= γ,
γ = ρ = 0.
Endpoint estimation and testing
Moreover, |A0 | ∈ RVρ and
lim
a0 (tx)
a0 (t)
− xγ
A0 (t)
t→∞
= xγ
xρ − 1
,
ρ
(9)
for all x > 0 (cf. Theorem 2.3.3 and Corollary 2.3.5 of de Haan and Ferreira, 2006). Denote
U (∞) := limt→∞ U (t) (= xF ); if (8) holds with γ < 0 then, provided x = x(t) → ∞,
lim
U (∞)−U (t)
a0 (t)
t→∞
+
1
γ
= Ψ?γ,ρ (∞) := −
A0 (t)
1
I
γ + ρ {ρ<0}
(10)
by similar arguments of Lemma 4.5.4 of de Haan and Ferreira (2006), with IA denoting
the indicator function which is equal to 1 if A holds true and is equal to zero otherwise.
Theorem 2. Let F be a d.f. in the Weibull domain of attraction, i.e., F ∈ D(Gγ ) with
γ < 0. Suppose U satisfies condition (8) with γ < 0 and, in this sequence, assume that
(10) holds. We define
1 2−γ − 1
h(γ) :=
+1 .
(11)
γ γ log 2
√
If the intermediate sequence k = kn is such that kn A0 (n/kn ) → λ∗ ∈ R, then
k min(−γ,1/2)
x̂F − xF
d
∗
−
h(γ)
−→
W
I
+
N
−
λ
b
I{γ≤−1/2} ,
γ,ρ
{γ≥−1/2}
n→∞
a0 nk
where W is a max-stable Weibull r.v., with d.f. exp{−(γx)−1/γ } for x < 0, N is a normal
r.v. with zero mean and variance given by
V ar(N ) = 1 +
2−(2γ+1) − 1 2−(γ+1) − 1 log 2
2
−γ
√
−
+
(2
−
1)
.
γ (log 2)2
2γ + 1
γ+1
2
(12)
and bγ,ρ is defined as
bγ,ρ
1
:=
log 2
Z
1
1/2
Ψ?γ,ρ
1 ds
=
2s s
1
1 1−2−(γ+ρ)
γ+ρ log
2 γ+ρ
1
2−γ (log 2γ
γ 3 log 2
−1 ,
ρ < 0,
+ 1) − 1 ,
ρ = 0.
Moreover, the r.v.s W and N are independent.
Remark 3. The same normalization by (a0 (n/k))−1 , with respect to γ = 0, is obtained
in Fraga Alves and Neves (2014) towards the Gumbel limit.
Corollary 4. Under the conditions of Theorem 2,
√ x̂F −xF
− h(γ)
k
n
a0
k
k (γ+1/2)+
d
−→ R,
n→∞
where a+ := max(a, 0) and R is a random variable with the following characterization:
1. Case −1/2 < γ < 0: R is max-stable Weibull, with d.f. exp{−(γx)−1/γ } for x < 0,
with mean Γ(1 − γ)/γ and variance equal to γ −2 Γ(1 − 2γ)R − Γ2 (1 − γ) . Here and
∞
throughout, Γ(.) denotes the gamma function, i.e. Γ(a) = 0 ta−1 e−t dt, a > 0.
2. Case γ < −1/2: R has normal distribution with mean −λ∗ bγ,ρ and variance given
in (12).
5
Endpoint estimation and testing
3. Case γ = −1/2: R is the sum of the two cases above, taken as independent compo√
nents, which yields a random part with mean Γ(1/2) − λ∗ b−1/2,ρ = π − λ∗ b−1/2,ρ
√
and variance 5 − π + 4 1 + (1/ 2 − 1)(2 + log 2)/ log 2 / log 2.
Remark 5. The function h(γ) is monotone decreasing for all γ < 0. Taking into account
the statement of Theorem 2, an adaptive reduced bias estimator based on the general
estimator x̂F is given by x̂FRB1 = x̂F − h(γ̂)â(n/k), with consistent estimators γ̂ and
â(n/k). The dominant component of the bias comes from the scale function a(n/k) which,
in case γ is close to 0, determines a very slow convergence. We have conducted several
simulations in this respect, indicating that this bias correction (of first order) has a very
limited effect.
In addition, we consider an adaptive second order reduced bias estimator developed
on the general estimator x̂F and supported on the asymptotic statement in Theorem 2
for γ ∈ (−1/2, 0). The limiting Weibull random variable W has a non-null mean equal
to Γ(1 − γ)/γ. We note that, for small negative values of γ, the convergence of the
normalized general estimator x̂F towards the Weibull limit can be very slow since it is
essentially governed by the function a ∈ RVγ and by the power transform k γ . Formally,
the general endpoint estimator x̂F satisfies the distributional representation
n
n γ
n γ
x̂F = xF + h(γ)a0
+ a0
k W + op a0
k ,
k
k
k
with h(γ) defined in (11). We thus develop an adaptive second order reduced bias estimator as follows:
Γ(1 − γ̂)
n
x̂FRB2 = x̂FRB1 −
â0 ( ) k γ̂
(13)
γ̂
k
(see Remark 5 for the definition of x̂FRB1 ). Furthermore, an approximated 100(1 − α)%confidence upper bound for xF is given by
n
xF < x̂F − â0 ( ) h(γ̂) + k γ̂ qα ,
k
(14)
with estimated α-quantile of the Weibull limit distribution qα := (− log α)−γ̂ /γ̂.
In practice, it is often advisable to perform statistical tests on the EVI sign so as to
prevent against an actual infinite endpoint. In Section 5, the testing procedures by Neves
et al. (2006) and Neves and Fraga Alves (2007) are applied with independent interest
from the particular EVI estimation problem inherent to the endpoint estimation. The
hypothesis-testing problem regarding the suggested max-domain of attraction selection is
stated as follows:
H0 : F ∈ D(G0 ) vs H1 : F ∈ D(Gγ )γ6=0 .
(15)
We will introduce another test statistic for tackling this testing problem. The new statistic Gn,k arises in connection with the general endpoint estimator x̂F , thus in complete
detachment of any extreme value index estimation procedure. It is given by
Gn,k :=
x̂F − Xn−k,n
.
Xn−k,n − Xn−2k,n
(16)
The next Theorem comprises the testing rule and ascertains consistency of the new testing
procedure with prescribed significance level α.
6
Comparative study via simulation
Theorem 6. Assume F ∈ D(Gγ ), for some γ ∈ R. Furthermore assume that the tail
quantile function U satisfies the second order conditions (8) up to (9). We define
G∗n,k (0) := log 2 Gn,k − log k +
log 2
,
2
where Gn,k is given in (16). If k = kn is an intermediate sequence such that
λ∗ ∈ R, as n → ∞, then
(17)
√
k A0 (n/k) →
d
• G∗n,k (0) −→ Z, if γ = 0. Here Z has Gumbel distribution function Λ = G0 ;
n→∞
P
• G∗n,k (0) −→ + ∞ , if γ > 0;
n→∞
P
• G∗n,k (0) −→ − ∞ , if γ < 0.
n→∞
Denoting by ξp := − log(− log(p)) the p-quantile of the Gumbel distribution, a critical
region for the two-sided test postulated in (15), at an approximate α-level, is deemed by
Theorem 6. The statement is that we reject H0 if either G∗n,k (0) ≤ ξα/2 or G∗n,k (0) ≥
ξ1−α/2 . Theorem 6 also allows testing for the one-side counterparts:
• H0 : F ∈ D(G0 ) vs H1 : F ∈ D(Gγ )γ>0 ,
thus rejecting H0 in favour of heavy-tailed distributions, if G∗n,k (0) ≥ ξ1−α ;
• H0 : F ∈ D(G0 ) vs H1 : F ∈ D(Gγ )γ<0 ,
by rejecting H0 in favour of short-tailed distributions, if G∗n,k (0) ≤ ξα .
Denoting the power function for the testing problem (15) (and subsequent one-sided
alternatives) by βn (γ) := Pγ [reject H0 ], it follows immediately from Theorem 6 that all the
designed tests are consistent tests since, as n → ∞, βn (γ)|γ6=0 → 1, with an approximate
level α given by βn (0) → α.
4
Comparative study via simulation
After dealing with the consistency and asymptotic distribution of the general endpoint
estimator x̂F defined in (6), we are now ready to find out how these properties carry over
to the finite sample setting. The finite sample properties of the test statistic G∗n,k (0),
defined in (17), are also investigated, assessing how it performs at either discerning the
presence of a heavy-tailed model (with d.f. F ∈ D(Gγ )γ>0 ), or at detecting a short-tailed
model (with F ∈ D(Gγ )γ<0 ). Consistency of the estimator (cf. Proposition 1) justifies
our belief that a larger sample leads to more accurate estimation about the true value
xF , whereas consistency of the test (cf. Theorem 6) connects a larger sample with a more
powerful testing procedure. Of course how large “sufficiently large” is, in terms of both
k and n, depends on the particular circumstances. The number of upper order statistics
k ∗ (yet to be determined) can be viewed as the effective sample size for extrapolation
beyond the range of the available observations. In particular, if k ∗ is too small, then the
endpoint estimator tends to have a large variance, whereas if k ∗ is too large, then the bias
tends to dominate. This typical feature will be reflected in our simulation results. We
argue comparison with other well-known endpoint estimators by means of their estimated
absolute bias and mean squared errors.
To this end, we have generated N = 300 samples from each of the four models listed
below, taken as key examples:
7
4.1 Endpoint estimation
• Model 1, with d.f. F1 (x) = 1 − [1 + (−x)−τ1 ]
γ = −1/(τ1 τ2 ) and the endpoint xF1 = 0.
• Model 2, with d.f. F2 (x) = 1 −
R log(1−1/x)
−∞
−τ2
, x < 0, τ1 , τ2 > 0. The EVI is
λ2 te−λt dt, x < 0, λ > 0. The EVI
d
is γ = −1/λ and the endpoint xF2 = 0. Moreover, X = − 1/(eZ − 1), where Z is
Gamma(shape = 2, rate = λ) distributed.
−τ
• Model 3, with d.f. F3 (x) = 1 − 1 + ( x1 − 1)−τ1 2 , x ∈ (0, 1), τ1 , τ2 > 0. The EVI
is γ = −1/(τ1 τ2 ) and the endpoint is xF3 = 1.
• Model 4, with d.f. F4 (x) = 1 − (1 − x)−1/γ , x ∈ (0, 1), γ < 0. The EVI is γ and
the endpoint is xF4 = 1. This corresponds to a Beta(1, −1/γ) model.
Each one of these models satisfies the main assumption that F ∈ D(Gγ ), for some EVI
γ < 0, which immediately entails a finite right endpoint xF .
Models 1, 2 and 3 are the same ones as in Girard et al. (2011, 2012), although these
works only tackle the EVI equal to −1. Model 4 is a Beta distribution parameterized
in γ < 0. At the present stage we are interested in studying the exact performance of
the general endpoint estimator x̂F for different ranges of the negative EVI. The extreme
value index γ is therefore a design parameter in the present simulation study, albeit
under the restriction to γ < 0. The second design parameter is of course the true right
endpoint xF , thus assumed finite. A number of combinations between model and design
factors are assigned in order to obtain distinct values of the negative EVI, particularly
γ = −1/2, −1/5, together with two possibilities for the right endpoint, xF = 0 and xF = 1.
The case γ = 0 has been extensively studied in Fraga Alves and Neves (2014), thus being
obviated in the present setting.
4.1
Endpoint estimation
The finite sample performance of the general estimator (notation: FAN) is here compared
with the naı̈ve maximum estimator Xn,n (notation: MAX) and with the estimator x̂∗
(notation: MOM.inv) introduced in equation (2.21) from Ferreira et al. (2003):
x̂∗ := Xn−k,n −
â(n/k)
.
−
γ̂n,k
(18)
We note that the above estimator evolves from (5) by using the consistent estimators for
the EVI and scale function, respectively defined as
−
γ̂n,k
:= 1 −
1
1−
2
(1) 2
−1
Nn,k
(19)
(2)
Nn,k
and
(1)
−
),
â(n/k) := Nn,k (1 − γ̂n,k
where
(r)
Nn,k
k−1
1X
:=
(Xn−i,n − Xn−k,n )r ,
k
i=0
8
(20)
r = 1, 2.
(21)
4.1 Endpoint estimation
Note that EVI estimator (19) is shift and scale invariant. The class (5) of estimators
for the finite right endpoint has already been applied to a lifespan study by Aarssen and
de Haan (1994) and Ferreira et al. (2003), under the assumptions of a finite endpoint and
that the EVI lies between −1/2 and 0. Another method for estimating the right endpoint
for negative EVI is via modeling the exceedances over a certain high threshold exactly
by a Generalized Pareto distribution (GPD). The result underpinning this parametric
approach establishes that F ∈ D(Gγ ) is equivalent to the relation
lim
sup
t↑xF 0<x<xF −t
x
F (t + x) − F (t)
− Hγ
= 0,
1 − F (t)
σt
(22)
where Hγ (x) := 1 + log Gγ (x) is the GPD (see e.g. de Haan and Ferreira, 2006). In
particular, if γ = 0, the GPD reduces to the exponential d.f. H0 (x) := 1−exp(−x), x ≥ 0.
In brief, condition (22) states that F ∈ D(Gγ ) if and only if the excesses Y := X −t above a
high threshold t are asymptotically Generalized Pareto distributed. Next to introducing
the class of GP distributions, relation (22) also enables to step away from the maxdomain of attraction, towards the actual fit of the GPD to the sample excesses, providing
a natural entry point to the POT approach. Once selected a high threshold t, the POT
method deems the shape parameter γ ∈ R (analogue to the EVI) and scale parameter
σt > 0 (which ultimately accommodates the influence of the threshold t) as the two indices
characterizing the excess distribution function over t. Then we can proceed via maximum
likelihood (ML), the methodology at the core of the POTML.GPD procedures. We note
that, if γ < 0, the parametric GPD fit corresponds to modeling the exceedances X over t
by a Beta distribution with finite right endpoint estimated by x̂FP OT = t− σ̂ M L /γ̂ M L , with
γ̂ M L < 0, σ̂ M L > 0. Of course, in the case γ = 0, a finite right endpoint is not allowed
while fitting the exponential distribution. References about the POTML.GPD approach
are the seminal works by Smith (1987) and Davison and Smith (1990).
In the semi-parametric setting, i.e. while working in the domain of attraction rather
than dealing with the limiting distribution itself, the upper intermediate o.s. Xn−k,n plays
the role of the high threshold t. For the asymptotic properties of the POTML estimator
of the shape parameter γ under a semi-parametric approach, see e.g. Drees et al. (2004),
Li and Peng (2010) and Zhou (2010).
In case F ∈ D(Gγ ) with γ < 0, we have mentioned before endpoint estimators arising
from the class (5). Now, let {Yi := Xn−i+1,n − Xn−k,n }ki=1 be the excesses above the
supposedly high random threshold Xn−k,n . Furthermore, and building on the relation
above, the sample excesses {Yi }ni=1 are assumed to follow a GPD. Then, the ML estimator
(σ̂kM L , γ̂kM L ) can be worked out as the solution of
arg max
γ<0, σ>0
k
Y
i=1
hγ (Yi /σ)/σ = arg max
γ<0, σ>0
k
Y
i=1
1+
γ −(1/γ+1) −k
Yi
σ ,
σ
∂
with hγ (x) := ∂x
Hγ (x) (see p.19 of de Haan and Ferreira, 2006, for a detailed explanation).
The POTML.GPD estimator of the right endpoint is defined as
x̂FM L := Xn−k,n − σ̂kM L /γ̂kM L , with γ̂kM L < 0 and σ̂kM L > 0,
(23)
showing a close similarity with the semi-parametric class (5) of right endpoint estimators.
We refer to section 4.5.1 of de Haan and Ferreira (2006) and Qi and Peng (2009) for the
semi-parametric handling of (23).
The POTML.GPD endpoint estimator relies on the shift and scale invariant ML-
9
4.1 Endpoint estimation
estimator of the shape parameter γ < 0, a restriction strictly to ensure that (23) is a
valid endpoint estimator. However there is no explicit formula for the ML-estimator.
An accumulating literature has pointed out this disadvantage. Maximization of the loglikelihood, reparameterized in (τ := γ/σ, γ), has been discussed in Grimshaw (1993).
Although theoretically well determined, even when γ ↑ 0, the non-convergence to a MLsolution can be an issue when γ is close to zero. There are also irregular cases which
may compromise the practical applicability of ML. Theoretical and numerical accounts
of these issues can be found in Castillo and Daoudi (2009) and Castillo and Serra (2015)
and references therein.
Inspired by the numeric examples in Girard et al. (2011, 2012), we have generated
N = 300 replicates of a random sample with size n = 1000 and computed the average
L1 -error given by
E(k ∗ ) :=
N
1 X
|ε(j, k ∗ )|,
N
where
ε(j, k ∗ ) := x̂k∗ (j) − xF , k ∗ ≤ n,
j=1
where xk∗ (j) denotes the endpoint estimator evaluated at the j-th replicate, for every k ∗ .
We borrow models 1-3 from Girard et al. (2011, 2012) and their performance measures
but we will not proceed with their proposals for endpoint estimation. Unlike their high
order moment estimators, none of the endpoint estimators adopted in the present simulation study (MAX, FAN, MOM.inv, and POTML.GPD) require the knowledge of the
original sample size n, since these rely on a certain number k ∗ of top o.s. only. For the
purpose of simplicity, the number k ∗ will be viewed as the effective sample size. In the
sequel, the naı̈ve estimator MAX is attached to k ∗ = 1 since it coincides with the first top
o.s.; the POTML.GPD and MOM.inv endpoint estimators are functions of the k ∗ = k + 1
upper o.s.; finally, the FAN estimator requires k ∗ = 2k top observations. We can also
compute the “optimal” values of k ∗ in the sense of minimizing the average L1 -error, i.e.
k0∗ := arg min{E(k ∗ ), k ∗ ≤ n}. Since the MAX entails k ∗ = 1, the associated function E
is constant and this optimality criterium has no effect on the naı̈ve estimator. Table 1
displays the simulation results, where we have considered parameter combinations with
respect to γ = −0.5, −0.2 and xF = 0, 1. The POTML.GPD endpoint estimates were
found by maximizing the log-likelihood over γ < 0.
The relative performance of the adopted endpoint estimators on the “optimal” k0∗ is
depicted in the box-plots of Figures 1 and 2, in terms of their associated errors ε(j, k0∗ ), j =
1, . . . , N , with N = 300. Apart from the obvious conclusion that the MAX tends to
underestimate the true value of the endpoint xF , we find that the POTML.GPD, MOM.inv
and FAN estimators have distinct behaviors with respect to the optimal levels k0∗ . In
particular, FAN estimates are not so spread out as the ones returned by POTML.GPD
or by the MOM.inv endpoint estimators, the latter showing larger variability.
Figures 3 and 4 display the plain average L1 -error (i.e. without the optimality assessment) against the number k ∗ of upper o.s. used in the corresponding estimation process.
The pertaining mean squared errors (MSE) are depicted in Figures 5 and 6, respectively.
The four models here addressed are set with EVI= −1/2, −1/5. From these results, it
is clear that the MOM.inv and POTML.GPD endpoint estimators are very unstable for
small values of k ∗ (k ∗ ≤ 100), contrasting with the small variance of the proposed FAN
estimator along the entire trajectory. On the other hand, both MOM.inv and FAN estimators show increasing L1 -error with increasing of k ∗ , a common feature to extreme
semi-parametric estimators. The FAN estimator seems to perform best in those regions
where other estimators exhibit high volatility, which may range from small to moderately
large values of k ∗ . This feature is more severe when γ = −0.2 (see second row of Figures
10
4.1 Endpoint estimation
Table 1: Average L1 -errors. The lowest values appear in bold.
Model
MAX
FAN
MOM.inv
POTML.GP
Model 1 (xF = 0)
(τ1 , τ2 ) = (2, 1)
(τ1 , τ2 ) = (5, 1)
0.028
0.231
0.013
0.041
0.024
0.145
0.019
0.128
Model 2 (xF = 0)
λ=2
λ=5
0.009
0.173
0.005
0.044
0.009
0.133
0.008
0.133
Model 3 (xF = 1)
(τ1 , τ2 ) = (2, 1)
(τ1 , τ2 ) = (5, 1)
0.027
0.187
0.012
0.129
0.022
0.120
0.014
0.095
Model 4 (xF = 1)
γ = −1/2
γ = −1/5
0.029
0.234
0.014
0.171
0.029
0.103
0.014
0.085
3 and 4), where the instability persists until an impressive k ∗ = 300 is reached. Once
attained a plateau of stability, the POTML.GPD tends to perform very well in general.
The best way to apply MOM.inv seems to dwell in a precise choice of k ∗ , which should be
selected at the very end of the very erratic path, just before bias sets in. The general endpoint estimator (FAN) tends to return values with a low average L1 -error and low MSE.
In fact, Figures 3 up to 6 not only provide us with a snapshot for this specific choice of
EVI values (−1/2 and −1/5), but also allow to foresee the estimates behavior with respect
to other EVIs in between, once we screen the plots from the top to the bottom in each
Figure. The boxplots in Figures 1 and 2 already suggested this possibility: the outliers
marked in these boxplots seem to move from lower to larger values of optimal bias ε(·, k0∗ )
as we progress on increasing EVI.
Altogether, the general endpoint estimator FAN seems to be an improvement to the
naı̈ve MAX estimator and tends to surpass the MOM.inv and POTML.GPD estimators,
by delivering low biased estimates quite often, while showing a low variance component.
This is particularly true for a EVI close to zero, as it would be expected from the one
estimator primarily tailored to tackle endpoint estimation in the Gumbel max-domain
of attraction (cf. Fraga Alves and Neves, 2014). Furthermore, the FAN estimator seems
to work remarkably well under a fairly negative EVI, considering that this is a general
estimator which does not accommodate any specific information about the true value of the
EVI. The overall performance of the MOM.inv and POTML.GPD endpoint estimators
is clearly damaged by their large variance in the top of the sample. It is worthy to
notice that the presented estimation procedure x̂F acts as a complement to numerical
POTML methods for endpoint estimation, the latter only available for strict negative
shape parameter γ; in contrast, x̂F presents an explicit simple expression, unifying the
estimation method to distributions with non-positive EVI, γ ≤ 0.
11
4.1 Endpoint estimation
Figure 1: Boxplots of the optimal bias ε(j, k0∗ ), j = 1, . . . , N , with N = 300. Endpoint estimates are drawn
from Model 1 (left) and Model 2 (right), with true value xF = xF1 = xF2 = 0.
Figure 2: Boxplots of the optimal bias ε(j, k0∗ ), j = 1, . . . , N , with N = 300. Endpoint estimates are drawn
from Model 3 (left) and Model 4 (right), with true value xF = xF3 = xF4 = 1.
12
4.1 Endpoint estimation
Figure 3: Average L1 -error, E(k∗ ), plotted against k∗ ≤ n/2. Endpoint estimates are drawn from Model 1
(left) and Model 2 (right), with true value xF = xF1 = xF2 = 0.
Figure 4: Average L1 -error, E(k∗ ), plotted against k∗ ≤ n/2. Endpoint estimates are drawn from Model 3
(left) and Model 4 (right), with true value xF = xF3 = xF4 = 1.
13
4.1 Endpoint estimation
Figure 5: Mean squared error (MSE) as function of the number of upper o.s. k∗ , k∗ ≤ n/2. Endpoint estimates
are drawn from Model 1 (left) and Model 2 (right), with true value xF = xF1 = xF2 = 0.
Figure 6: Mean squared error (MSE) as function of the number of upper o.s. k∗ , k∗ ≤ n/2. Endpoint estimates
are drawn from Model 3 (left) and Model 4 (right), with true value xF = xF3 = xF4 = 1.
14
4.2 A testing procedure built on the general endpoint estimator
4.2 A testing procedure built on the general endpoint estimator
This section concerns the finite sample performance of Gn,k , presented in Theorem 6, as
a convenient tool for either discarding heavy-tailed models or for detecting short-tailed
models F ∈ D(Gγ ), γ < 0. One grounding result in this respect is that all the semiparametric endpoint estimators we are adopting, are consistent under the assumption
that k = kn is an intermediate sequence of positive integers, i.e. k = kn → ∞ and
kn /n → 0, as n → ∞. The testing procedures we wish to apply also bear on this usual
assumption in statistics of extremes. There are many proposals for testing procedures
aiming at the selection of a suitable max-domain of attraction. For a wide view on this
topic, we refer the surveys on testing about extreme values conditions available in Hüsler
and Peng (2008) and Neves and Fraga Alves (2008). We recall that EVI estimation is not
a requirement for the general endpoint estimation defined (6) and emphasize that both
Weibull and Gumbel domain are allowed. Thus, for the time being, we will rely on testing
procedures which do not require external estimation of the EVI. We will compare the new
test statistic Gn,k , defined in (16), with the Ratio and Greenwood statistics introduced in
Neves et al. (2006) and Neves and Fraga Alves (2007) for the one-side alternatives
H0 : F ∈ D(G0 )
H1 : F ∈ D(Gγ )γ>0
vs
[or
H10 : F ∈ D(Gγ )γ<0 ] .
The Ratio(R) and Greenwood(Gr) statistics are defined as
R :=
Xn,n − Xn−k,n
N1
Gr :=
N2
,
(N1 )2
P
(j)
j
with Nj = Nn,k = k1 k−1
i=0 (Xn−i,n − Xn−k,n ) , j = 1, 2. Under H0 , the standardized
∗
version of Ratio statistic, R := R − log k,
pis asymptotically Gumbel, whereas the suitably
∗
normalized Greenwood statistic, Gr := k/4(Gr−2), is asymptotically standard normal.
Formally, approximated α-significant tests against the alternative H1 [resp. H10 ] render
the rejection regions R∗ ≥ ω1−α [resp. R∗ ≤ ωα ] and Gr∗ ≥ z1−α [resp. Gr∗ ≤ zα ],
where ω := Λ← () and z := Φ← (). Here, Φ denotes the d.f. of the standard normal.
Corresponding p-values of the test are p = 1 − Λ(g ∗ ) against the heavy-tailed alternative,
and p = Λ(g ∗ ) against the short-tailed alternative, for the observed values r∗ and gr∗ of the
test statistics R∗ and Gr∗ , respectively. The approximated p-values against heavy-tailed
alternatives H1 [resp. short-tailed alternatives H10 ] are given by 1 − Λ(r∗ ) and 1 − Φ(gr∗ )
∗ (0) of the test statistic G∗ (0).
[resp. Λ(r∗ ) and Φ(gr∗ )], for an observed value g ∗ := gn,k
n,k
Figures 7-8 summarize the comparison between the performance of the test based on
G∗n,k (0) (cf. Theorem 6) with the above mentioned tests R∗ and Gr∗ . The simulations
yield large p-values in connection with the heavy-tailed alternatives H1 , meaning that
heavy-tailed distributions are likely to be detected by these tests. On the opposite side,
the test are not so sharp against short-tailed alternatives in H10 . The new test statistic
G∗n,k (0) rejects on smaller values of k ∗ than the R∗ statistic, thus revealing more powerful
than the Ratio-test.
The Greenwood test compares favourably to G∗n,k (0) in terms of power, against the
heavy-tailed alternative. However, it tends to be a more conservative test than the new
proposal, often returning p-values much less than 5%.
15
4.2 A testing procedure built on the general endpoint estimator
Figure 7: Average p-values of the simulated G∗n,k (0), R∗ and Gr∗ either against a heavy-tailed alternative
(left) or a short-tailed alternative (right), with respect to k ∗ ≤ n = 1000.
Figure 8: Average p-values of the simulated G∗n,k (0), R∗ and Gr∗ either against a heavy-tailed alternative
(left) or a short-tailed alternative (right), with respect to k ∗ ≤ n = 1000.
16
Case study: supercentenarian women lifespan
5
Case study: supercentenarian women lifespan
This section is devoted to the practical illustration of our methodology for statistical
inference about the endpoint. Our data set of oldest people comprises records of lifetimes
in days of verified supercentenarians (women), with deaths in the time window 1986-2012.
The data set was extracted from Table B of Gerontology Research Group (GRG), as of
January 1, 2014, merged with Tables C and E, as of June 29, 2015, available at http:
//www.grg.org/Adams/Tables.htm. Although the referred database includes lifespan
records tracing back to 1903, these are often sparse and with a low average number of
yearly records, which is not surprising since the GRG was only founded in 1990. The
later two years of records 2013-2014 are not yet closed. Therefore we settle with the
1272 supercentenarian women lifetimes, recorded from 1986 to 2012, and corresponding
to approximately 90% of the total number of records since 1903.
The terms “life expectancy” and “lifespan” describe two entirely different concepts,
although people tend to use these terms interchangeably. Life expectancy refers to the
number of years a person is yet expected to live at any given age, based on the statistical
average. Lifespan, on the other hand, refers to the maximum number of years that a
person can potentially expect to live based on the greatest number of years anyone has
lived. We are interested in the latter.
Formally, in gerontology literature, maximum lifespan potential (MLSP) is the operative definition for the verified age of the longest lived individual for a species (Olshansky
et al., 1990b) and, in this sense, can be viewed as a theoretical upper limit to lifetime.
The oldest documented age reached by any living individual is 122 years, meaning humans
are said to have a MLSP of 122 years. In Biology, theories of ageing are mainly divided
into two groups: damage theories and program theories. According to damage theories,
we age because our systems break down over time; so, if damage theories hold true, we
can survive longer by avoiding damaging our organism. Program theories consider that
we age because there is an inbuilt mechanism that tells us to die; according to that, we
cannot survive longer than the upper limit of longevity despite of our best efforts (see
Hanayama (2013)). Kaufmann and Reiss (2007) also discussed the issue of whether the
right endpoint of the life span is infinite, for which they analyzed mortality data from
West Germany. Their estimated shape parameter within the Generalized Pareto model is
equal to −0.08 and the right endpoint of the estimated beta distribution is equal to 122
years, an estimate that fits well to the worldwide reported life span of the most famous
record-holder, the Frenchwoman Jeanne-Louise Calment (122 years and 164 days) who
was born in Arles on Feb. 21, 1875 and died in Arles on Aug. 4, 1997. However, Kaufmann and Reiss (2007) did not conclude categorically that human lifespan has a finite
upper limit, arguing that by using the concept of penultimate distributions we can show
that an infinite upper limit is well compatible with extreme value theory. They carry
on pointing out a Beta distribution as a suitable model. Aarssen and de Haan (1994)
analyzed lifespan data from the Netherlands using statistical methods under the extreme
value theory umbrella. Aarssen and de Haan (1994) showed that there is a finite age
limit, tackled with reasonable confidence bounds in the 113 − 124 year span, a conclusion
confined to the years of birth 1877 − 1881 in the Netherlands.
Stephen Coles, a specialist in tracking human supercentenarians and co-founder of
the Supercentenarian Research Foundation (SRF), refers to the supercentenarians as “the
most extreme example of human longevity that we know about, the oldest old”. In
Coles (2011), the value 122 is referred to as the “Calment Limit” for human longevity
(what is designated here as the MLSP), which is supported on the fact that nobody
has come even close that extreme age over the last 19 years. It is also mentioned that,
17
Case study: supercentenarian women lifespan
from his research experience, “supercentenarians had virtually nothing in common: they
had different occupations, lifestyles, religions and so on, regardless the common factor of
long-lived relatives.” Also, according to Vaupel (2011) “the explosion in very long life has
already begun”, although by his perspective “we cannot see much beyond 122.”
Several authors have stated that despite of the increasing “life expectancy”, the “maximum human lifespan” has not much changed. According to Troen and Cristafalo (2001)
some biodemographic estimates predict that elimination of most of the major diseases
such as cancer, cardiovascular disease, and diabetes would add no more than 10 years to
the average life expectancy, but would not affect MLSP (Olshansky et al., 1990a; Troen
and Cristafalo, 2001). Other researchers go further enough to hypothesize that mortality
will be compressed against that fixed upper limit to life time (Compression Theory by Fries,
1980). On the other hand, Wilmoth and Robine (2003) argue a possible world trend in
maximum lifespan, based on a long series of Swedish data. Above all, there is still plenty
of scope to assess significance of other covariates, like the negative impact of obesity and
epidemic diseases on the rise in life expectancy trends and the possible impact on the
MLSP.
What researchers seem to agree on is the need for better data, since at present, there
is insufficient data available on the extreme elderly population. We should keep in mind
that age is often misreported and at the time the centenarians (and supercentenarians)
were born, record keeping was less complete than it is nowadays.
With this illustrative example of estimation of the ultimate lifespan by adopting the
general endpoint estimator, it is not our aim to make conclusions for a specific cohort of
individuals in time or space, nor any other type of serial studies. Instead, the interest will
be on the question of what are sensible bounds for the MLSP, at the current state of the
art.
At this point, several assumptions are needed about the right tail of the lifetime
distribution, which is the focus of our extreme value analysis. The first assumption is
that the available data comprises a sample of i.i.d. observations. We find reasonable to
assume independence in our data, since we have one record for each individual person. The
stationarity assumption is preliminary asserted from the plot in Figure 9 displaying the
comparative boxplots for the larger women’s lifetimes by the year. A common feature to all
the boxplots is the presence of very large observations classified as extreme outcomes. The
boxplots also suggest an increasing third quartile as we progress in time. Such an increase
is not so apparent in the annual maxima as we move across the 27 time points (years). A
possible interpretation is that an increase in the mean of the supercentenarians’ lifetimes
may not be connected to an increasing lifespan over time, but rather to a possible trend
in the frequency of the highest lifetime observations. There is a recent semi-parametric
development by de Haan et al. (2015), suitable for assessing the presence of a trend in
the frequency of extreme observations, which is also reflected in the scale of extremes.
However, their inference techniques require a large number of replicates per year, and this
is not at our grasp given the limited amount of observations available within the same
year. For instance in the top part of Figure 9, that we have always less than 40 yearly
observations until 1997 (the year of Calment), with the lowest number of 12 observations
for 1991. Although the number of observations virtually doubles for the later years, the
absolute record still stands on the Calment limit of 122 years, the overall sample maximum.
The plot in Figure 10 shows the Loess fit (Local scatterplot smoothing) to the yearly data,
given by blue curve overlay. This is almost parallel to the horizontal axis. Confidence
bands are also presented with a preassigned 95% confidence level. These seem rather
narrow. Hence, we find no evidence of a particular trend through this nonparametric
18
Case study: supercentenarian women lifespan
Figure 9: Comparative boxplots of the lifetimes of supercentenarians (Women) reported from 1986 to 2012.
Table 2: Maximum-likelihood parametric GPD(σt ) and GPD(σ) fit to the lifetime exceedances of supercentenarian women for the time period 1986-2012.
σt
exp{β0 + β1 t}
exp{β0 }
exp{β0 + β1 t}
exp{β0 }
exp{β0 + β1 t}
exp{β0 }
exp{β0 + β1 t}
exp{β0 }
exp{β0 + β1 t}
exp{β0 }
`i , i = 0, 1
β̂0
-1739.324 0.327
-1740.238 0.432
-867.215 0.497
-867.374 0.436
-424.197 0.558
-425.542 0.304
-192.652 0.567
-193.942 0.198
-75.904 0.150
-76.495 -0.237
β̂1
0.006
–
-0.003
–
-0.013
–
-0.019
–
-0.021
–
γ̂ thresh/#exc x̂FP OT p-value
-0.056
110/1272
-0.061
135.36 0.1763
-0.083
111/639
-0.078
130.63 0.5729
-0.062
112/338
-0.045
142.05 0.1010
-0.044
113/164
-0.016
191.21 0.1083
0.141
114/82
0.170
–
0.2769
technique.
Despite the above, a POT parametric approach is applied to detect a possible trend in
the scale. Here, the GPD is fitted to the threshold excesses, via a ML fit to Hγ (x/σ), GPD
is fitted to the threshold excesses, considering a trend through time as σ = σt = exp{β0 +
β1 t}. The resulting non-stationary model is denoted by M1 , whereas the corresponding
stationary version GPD with σ = exp{β0 } is denoted by M0 . Let `1 and `0 be the
maximized log-likelihoods for models M1 and M0 , respectively. The Likelihood Ratio
test for H0 : M1 = M0 using the deviance statistic D = 2{`1 − `0 } to formally compare
models M1 and M0 , returns the results summarized in Table 2, for the lifetime data
over the selected thresholds 110, 111, 112, 113 and 114. The last column contains the pvalues, related to the different threshold selection. Again, we find no strong evidence of
a linear trend in the log-scale parameter σ. Moreover, for each threshold, the EVI and
endpoint ML estimates for women lifetimes are also listed for the sake of comparison with
subsequent results in Section 5.3.
The previous preliminary parametric analysis does not exhaust all the possible choices
of parametric models encompassing a trend in extremes. The interest is not in the se-
19
Case study: supercentenarian women lifespan
Annual lifetimes of supercentenarians (Women)
122
Age in years
120
118
116
114
112
110
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
Year of death
Figure 10: Loess fit to the lifetime of supercentenarians (Women) reported from 1986 to 2012.
lection of the most suitable parametric model for extremes, but in being able to ascertain, with a certain degree of confidence, that dropping out of the time covariate does
not affect our subsequent analysis under the assumption of stationary supercentenarian
women’s lifetimes. For recent applications incorporating information over time, we refer
the works of Stephenson and Tawn (2013) and de Haan et al. (2015). The latter comprises
a comparative analysis with existing methodologies in a similar context.
The available supercentenarian women data should be regarded as the greatest lifetimes collection ever attained by the human population. We have 1272 observations
available, which we have found to satisfy the i.i.d. assumption. Figure 11 contains the
boxplot and the histogram for the whole univariate data set.
Figure 11: Boxplot and histogram built on the lifetimes of supercentenarian women, for the time interval
1986-2012.
20
5.1 Testing finiteness in the right endpoint
5.1
Testing finiteness in the right endpoint
Our first aim is to assess finiteness in the right endpoint of the d.f. F underlying the
women lifespan data. The detection of a possibly finite upper bound on our data follows
a semiparametric approach, meaning that we essentially assume that F belongs to some
max-domain of attraction. We then consider the usual asymptotic setting, where k =
kn → ∞ and kn /n → 0, as n → ∞, and hence Xn−k,n → xF a.s. According to this setting,
it is only natural to expect that any statistical approach to the problem of whether there
is a finite endpoint or not, will depend on the extent of the dip into the original sample of
supercentenarian women’s records. The baseline to this issue is mostly driven by a second
and more operative question: how to select the adequate top sample fraction to use with
both our testing and estimation methods? A suitable choice of k comes from a similar
approach to the one in Wang (1995), where kopt is deemed to be selected at the value k
from which the null hypothesis is rejected.
For a more definite judgment about the existence of a finite upper bound on the
supercentenarian lifetimes, we are going to apply the testing procedure introduced in
Neves and Pereira (2010). The purpose now is to detect finiteness in the right endpoint
of the underlying distribution which may belong to either Weibull or Gumbel domains.
More formally, the testing problem
H0 : F ∈ D(G0 ), xF = ∞ vs
H1 : F ∈ D(Gγ )γ≤0 , xF < ∞
(r)
is tackled using the log-moments Nr ≡ Nn,k , r = 1, 2, defined in (21), but now replacing
the observations Xi,n by their log-transform log(Xi,n ). It is possible to do so because
we are dealing with positive observations. We point out however that this leads to a
non-location invariant method. The test statistic T1 being used is defined as
k
1 X Xn−i,n − Xn−k,n − T
N1
T1 :=
, with T := Xn−k,n
k
Xn,n − Xn−k,n
2
i=1
[N1 ]2
1−
N2
!−1
.
√
Under H0 the standardized version of the test, T1∗ := k log k T1 , is asymptotically normal. Moreover, T1∗ tends to inflect to the left for bounded tails in the Weibull domain and
to the right if the underlying distribution belongs to the Gumbel domain. The rejection
region of the test is given by |T1∗ | ≥ z1−α/2 , for an approximate α significance level. Figure
12 displays the sample path of T1∗ . The most adequate choice of the intermediate number
k (which carries over to the subsequent semi-parametric inference) is set on the lowest k
at which the critical barriers with a α = 5% significance level are crossed. This optimality
criterium yields k0N P = 487, spot on the smallest value where we find enough evidence of
a finite endpoint.
It remains to be assess whether the distribution underlying the supercentenarian data
(now assumed bounded from above) belongs to the Gumbel domain or to the Weibull maxdomain of attraction. This will be carried out by a proper hypothesis-testing problem,
termed statistical choice of extremes domains. In view of our specific interest on the finite
endpoint, we are using the one-sided version of the test. The hypotheses are:
H0 : F ∈ D(G0 )
vs
H10 : F ∈ D(Gγ )γ<0 .
Figure 13 depicts the sample paths of the new test statistic G∗n,k (0) from Theorem 6,
the Greenwood Gr∗ statistic, and the Ratio R∗ statistic, as well as their α = 5% asymptotic critical values. The Greenwood test finds enough evidence to reject the Gumbel
domain hypothesis in favour of a bounded short-tail in Weibull domain, at the suitable
21
5.1 Testing finiteness in the right endpoint
Figure 12: Detecting finiteness in right endpoint for verified supercentenarians data set: sample paths of the
normalized statistics. The horizontal dashed lines correspond to the α = 5% critical barriers.
Figure 13: Testing max-domains of attraction with short tailed alternative, for verified supercentenarians
data set: sample paths of the normalized test statistics. The horizontal dashed lines correspond to the α = 5%
critical barriers.
intermediate value of k0G = 465. The two other statistics, new G∗n,k (0) and ratio R∗ , lead
to a more conservative conclusion, with both tests leaning towards the non rejection of
the null hypothesis. This conservative aspect also crops up in the simulations section
(Section 4), where the Greenwood test is found to be more powerful against short-tailed
alternatives attached to γ < 0.
From the previous analysis, we find it reasonable to assume a finite right endpoint
for some distribution in the Weibull domain simultaneously for all the adopted testing
methods at the maximum number of upper extremes
kopt := max{k0G , k0N P } = 487.
Therefore, based not only on the testing procedures but also on the complementary
EVI estimation presented in Section 5.3, it seems reasonable to conclude that the lifespan distribution belongs to the Weibull domain of attraction with finite right endpoint xF .
For definitely discarding the presence of a heavy-tailed distribution underlying our
data set, it is also important to test:
H0 : F ∈ D(G0 )
vs
H1 : F ∈ D(Gγ )γ>0 .
22
5.2 Endpoint estimation for women’s records without EVI knowledge
Figure 14: Testing max-domains of attraction for verified supercentenarians data set: p-values for the normalized test statistics, with heavy tailed alternatives. The horizontal dashed line corresponds to the α = 5%
nominal level.
Figure 15: Endpoint estimation for verified supercentenarians data set: estimator x̂F and Calment limit.
In Figure 14 we observe that all the p-values determined by all the three test statistics
are increasing with k. Again, the conservative behavior of the two tests seems to emerge.
Despite all p-value paths begin at very small values around zero, this only lingers for a tight
range of higher thresholds which may not be in good agreement with the requirement of a
sufficiently large k. The Greenwood statistic returns p-values very close to 1, from about
k = 400 onwards. The other two statistics (G∗n,k (0) and R∗ ) yield moderate p-values,
with larger values returned by G∗n,k (0). It seems sensible then to discard a heavy-tailed
distribution for the supercentenarian women lifespan, a conclusion clearly verified by the
Greenwood test.
5.2 Endpoint estimation for women’s records without EVI
knowledge
Following the testing procedures in 5.1 and the reported optimal number kopt = 487,
we will present analogous graphical tools with respect to endpoint estimation. The first
purpose is to illustrate the smooth behaviour of x̂F defined in (6), already anticipated in
the simulations section (Section 4).
Figure 15 displays the comparative finite-sample behaviour of x̂F (notation: FAN) with
23
5.3 Linking endpoint estimation of women’s records to EVI estimation
Figure 16: EVI estimation with the verified supercentenarian data.
the naı̈ve Calment limit (notation: MAX) for the supercentenarian’s data set. Recall that
our database is regarded as a collection of the greatest lifetimes of women population and
that the general endpoint estimator x̂F always returns values above the naı̈ve endpoint
estimate, i.e. greater than “Calment limit” of 122.4 years.
After the initial rough path in the range of approximately one hundred top observations, the estimates trajectory of x̂F then becomes very flat. Once we dip into the
intermediate range of extremes, the sample path becomes smoother.
At this point we find it sensible to provide similar information provided by the other
semi-parametric and parametric endpoint estimators already intervening in the simulation
study, as they constitute a good complement to a more thorough insight about the true
endpoint. This is the subject of the next section.
5.3 Linking endpoint estimation of women’s records to EVI
estimation
In this section, the endpoint estimation is tackled using methodologies that require an
estimator for the extreme value index γ < 0. In this sequence, we adopt the moment
−
related estimator γ̂n,k
(notation: MOM.inv), defined in (19), and the POTML.GPD esM L to plug in (23) (subject to γ < 0). Figure 16 is
timator of the shape parameter γ̂n,k
the estimates plot of the chosen estimators for γ, as function of k. We note that these
two estimators enjoy the location and scale invariance property. Retaining the kopt = 487
larger observations (a value delivered by the testing procedures) as the effective sample,
−
M L = −0.059, both coherent with a
we find the point estimates γ̂n,487
= −0.087 and γ̂n,487
short-tailed distribution attached to some γ ∈ (−1/2, 0).
Figure 17 depicts the results for estimators x̂FM L , x̂∗ (defined in (23) and (18), respectively). We are also including the second order reduced bias version of the general
−
endpoint estimator x̂FRB2 defined in (13), by plugging in the estimators γ̂ = γ̂n,k
and â0 nk
provided in (19) and (20), respectively. The trajectory of x̂FRB2 lies in the middle range
between that of x̂FM L and of x̂∗ .
The optimal value kopt = 487 can viewed as benchmark value (or change point) since it
breaks the disruptive estimates pattern. It actually pinpoints where the graph stops being
too rough to make inference and starts being more stable, so that we can infer about the
endpoint. The latter applies in particular to estimators x̂∗ , x̂FRB2 and x̂FM L , which return
24
5.3 Linking endpoint estimation of women’s records to EVI estimation
Figure 17: Endpoint estimation with the verified supercentenarian data.
x̂∗kopt = 128.32, x̂FRB2,kopt = 131.82 and x̂FM L,kopt = 135.62. The simulations also outline
the plain general endpoint estimator x̂F as being relatively efficient (in terms of bias and
MSE) on those moderate values of k ∗ where other estimators fall short. In this respect,
Figure 15 shows a rather flat plateau from which we find safe to draw an estimate for the
endpoint, and this portion of the graph includes x̂Fkopt = 122.94. The asymptotic results in
Theorem 4 for γ ∈ (−1/2, 0) equip us with a tool for finding an approximate 100(1 − α)%
upper bound for the true endpoint xF . Obviously, this process calls for the estimation
of γ since we need some guidance about the proper interval where the EVI lies within.
Hence, we have just introduced the most direct link to the estimation of γ (see Figure
16) in connection with the general endpoint estimator x̂F . Selecting kopt = 487, the 95%
confidence upper bound for xF delivered by (14) is 133.23 years.
We also note that the simulation outcomes relate well to the present results arising
from the available data of supercentenarian women’s records. For instance, we observe in
Figure 16 that small values of k (k ≤ 110, approximately) find positive estimates for the
MOM.inv estimator which could on their own account for the erratic pattern of x̂∗k in the
plot of Figure 17, but the simulations have yield this rough pattern very often in connection
with a true negative EVI. Furthermore, the shape parameter γ is estimated via POTML
subject to γ < 0 (cf. Figure 16) and this returns endpoint estimates x̂FM L well aligned with
the ones pertaining to x̂FRB2 and x̂∗k (cf. Figure 17). Therefore, heeding to the simulation
results, we find reasonable to conclude that the misrepresentation of the negative EVI for
moderate values of k is not the main factor compromising the performance of the adopted
endpoint estimators in the early part of the plot (amounting to about 10% of the original
sample size n, say).
Finally, this practical application also seems to suggest that removing the bias component from x̂F causes an increase in the variance. The bias/variance trade-off effect is
grasped more thoroughly in the Appendix 8, where the finite sample properties of the
second order reduced bias estimator x̂FRB2 are studied by taking Models 1 and 4, from
Section 4, as parent distributions. The brief simulation study in the Appendix 8 is expected to reinforce the suggested competitive performance amongst the above-mentioned
endpoint estimators.
25
5.4 An upper limit to lifespan and probability of surpassing Calment limit
5.4 An upper limit to lifespan and probability of surpassing
Calment limit
From the previous data analysis, one would say that the ultimate human lifespan would
not be greater than 133.23 years (the estimated upper bound from (14), obtained in section
5.3). This gives some insight beyond Calment’s achievement: the absolute record of 122.4
years, still holding to the present date (for the last 19 years). We thus expect that the
probability of exceeding the “Calment limit”, even for a supercentenarian women, will be
extremely low.
Figure 18: Probability of exceeding the “Calment limit”, for a supercentenarian women, given today’s state
of the art.
Figure 19: Probability of exceeding the “Calment limit” using the POTML.GPD approach and corresponding
95% confidence bound: all sample (top); 350 ≤ k ≤ 700 (bottom).
26
Concluding remarks
This tail probability can be estimated using the following semi-parametric estimator:
k
Pbn (Xs > 122.4) :=
n
max 0, 1 +
− 122.4 − Xn−k,n
γ̂n,k
â(n/k)
−1/γ̂ −
n,k
,
(24)
(cf. (4.4.1) in de Haan and Ferreira, 2006) where Xs := X|X ≥ 110 denotes the lifetime
−
of a supercentenarian women, and γ̂n,k
, â(n/k) are the related estimators defined in (19)
and (20), respectively.
Figure 18 depicts the probability estimates from (24), together with their POTML.GPD
analogues, for a wide range of larger values of the 1272 verified supercentenarians data
set. In contrast with the previous statistical analysis, the sample size n now intervenes in
(24). Therefore, any inference drawn on this account will apply to the subpopulation of
supercentenarian women under study. All point estimates are very close to zero. Figure 19
displays again the POTML.GPD estimates but with respective 95% upper bounds. Since
we are dealing with very small probabilities the asymptotic bounds are not so sharp, as
opposed to the case of an underlying distribution with infinite right endpoint.
6
Concluding remarks
The scope for application of the right endpoint estimator introduced in Fraga Alves and
Neves (2014), primarily designed for the Gumbel domain of attraction, is here extended to
the case of an underlying distribution function F in the Weibull domain. The consistency
property and asymptotic distribution of this general endpoint estimator x̂F renders a unified estimation procedure for the right endpoint under the assumption that F ∈ D(Gγ )γ≤0 .
A new test statistic arises tied-up with x̂F thus incrementing the range of available testing
procedures for selecting max-domains of attraction. Our main findings are listed below.
• The general endpoint estimator does not require the estimation of the EVI, unlike
the widely-used semi-parametric alternatives.
• By construction, the estimator x̂F always returns larger values than the sample
maximum Xn,n , a property not shared by other semi-parametric methodologies we
have encountered so far, particularly those predicated on the Weibull max-domain
of attraction.
• The simulation study conveys a good finite sample performance of the general endpoint estimator, ascertaining competitiveness to benchmark endpoint estimators
specifically tailored for the Weibull domain.
• Related to the previous, the general endpoint estimator performs better for distributions with some γ > −1/2, which corresponds to the most common situation in
practical applications.
• The problem of choosing the most adequate number k of upper order statistics is here
mitigated by the usual flat pattern of the estimates trajectories, a typical feature of
the general endpoint estimator.
• The application to the supercentenarian women’s lifetimes illustrates how we can
easily establish a confidence upper bound to the right endpoint, building on the
asymptotic results for γ > −1/2.
27
APPENDIX A: Proofs
7
APPENDIX A: Proofs
This section is entirely dedicated to the proofs of the results introduced in Section 3. In
what follows we find more convenient to consider the estimator x̂F in the functional form
Z 1
ds
1
F
Xn−[2ks],n − Xn−[ks],n
x̂ = Xn−k,n −
,
(25)
log 2 0
s
where [a] denotes the integer part of a ∈ R (more details about the representation (25)
can be obtained in Fraga Alves and Neves, 2014).
We note that if s ∈ [0, 1/(2k)[, then the integral in (25) is equal to zero. Bearing this
in mind, we write
Z 1
ds
1
x̂F = Xn−k,n −
Xn−[2ks],n − Xn−[ks],n
.
log 2 1
s
2k
Moreover, if s ∈ [1/(2k), 1/k[ then [ks] = 0 (not depending on s) and thus Xn−[ks],n =
Xn,n . Therefore, we have that
Z 1
Z 1
Z 1
k ds
1 n k
ds
ds o
.
Xn−[2ks],n
− Xn,n
+
(Xn−[2ks],n − Xn−[ks],n )
1
1
1
log 2
s
s
s
2k
2k
k
(26)
With a suitable variable transform on the last integral, we can reassemble (26) in a tidy
manner:
Z 1
1
ds
x̂F = Xn,n + Xn−k,n −
Xn−[2ks],n .
(27)
log 2 1
s
x̂F = Xn−k,n −
2
This is the main algebraic expression that will be used to derive the asymptotic distribution of x̂F in the proof of Theorem 2, which is a natural consequence of the three random
contributions in (27).
Proof of Proposition 1: We see that the integral in the functional form (27) satisfies
the inequalities
Z 1
ds
(log 2) Xn−k,n ≤
Xn−[2ks],n
≤ (log 2) Xn−2k,n .
1
s
2
Therefore, we obtain the following upper and lower bounds involving x̂F − xF ,
Xn,n − xF ≤ x̂F − xF ≤ (Xn,n − xF ) + Xn−k,n − Xn−2k,n ,
and the result thus follows easily because the three o.s. Xn,n , Xn−k,n and Xn−2k,n all
converge almost surely to xF , provided the intermediate nature of k = kn .
Remark 7. Alternative proof based on the functional form (7) of the k ∗ := 2k top o.s.:
strong consistency of the general endpoint estimator comes from the lower and upper
28
APPENDIX A: Proofs
bounds of (7) given below.
x̂F − xF
!
k−1
1 X
k+i+1
Xn−k,n −
log(
) Xn−k−i,n
log 2
k+i
i=0
!
k−1
k+i+1
1 X
log(
) = Xn,n − xF
Xn−k,n − Xn−k,n
log 2
k+i
= (Xn,n − xF ) +
≥ (Xn,n − xF ) +
i=0
and on the other hand,
k−1
F
x̂ − x
F
F
≤ (Xn,n − x ) +
Xn−k,n − Xn−2k+1,n
1 X
k+i+1
)
log(
log 2
k+i
!
i=0
F
= (Xn,n − x ) + (Xn−k,n − Xn−2k+1,n ) ;
since for any intermediate k = kn the o.s. Xn,n , Xn−k,n , Xn−2k,n converge almost surely
to xF , the result follows.
Before getting under way to the proof of the main Theorem, we need to lay down some
ground results. These comprise a Proposition regarding γ < 0 and a Lemma for general
γ.
Proposition 8. Suppose Xn,n is the maximum of a random sample whose parent d.f.
F detains finite right endpoint of F , i.e. xF = U (∞) < ∞. Assume the second order
condition
(8) holds with γ < 0. If k = kn is such that, as n → ∞, k → ∞, k/n → 0 and
√
k A0 (n/k) → λ∗ ∈ R, then
1. for γ ≥ −1/2, for each ε > 0,
k −γ−ε
Xn,n − xF
a0 nk
p
−→ 0.
(28)
n→∞
Moreover,
k −γ
Xn,n − xF d Z γ
−→
,
n→∞ γ
a0 nk
where Z denotes a standard Fréchet with d.f. Φ1 as in (2).
2. for γ < −1/2,
√
k
Xn,n − xF
a0 nk
p
−→ 0.
n→∞
d
Proof:
to the well-known equality in distribution that Xi,n = U (Yi,n ), i = 1, 2, . . . , n,
Owing
n
with Yi,n i=1 the n-th o.s. from a sample of n independent r.v.s with common (standard)
Pareto d.f. given by 1 − x−1 , x ≥ 1, then the following equality in distribution holds:
Xn,n − xF d n U nk Yn,n nk − U nk
1 o n U (∞) − U nk
1o
=
+
+
−
.
γ
γ
a0 nk
a0 nk
a0 nk
29
APPENDIX A: Proofs
Now we use conditions (8) and (10) with t replaced by n/k everywhere:
γ
γ −1
k n Yn,n − 1 1
Xn,n − xF d
n ? k
Ψ
=
+ + A0
Yn,n 1 + op (1)
γ
γ
k γ,ρ n
a0 nk
n ?
− A0
Ψ (∞) 1 + o(1)
k γ,ρ
γ
o
k γ n−1 Yn,n
n n ? k
1
n
=
+ A0
Ψγ,ρ
Yn,n +
I{ρ<0} + op A0
γ
k
n
γ+ρ
k
We note at this stage that n−1 Yn,n is asymptotically a Fréchet r.v. with d.f. given by Φ1
in (2). This non-degenerate limit
n,n going to infinity with probability one,
yields (k/n)Y
?
−1
which implies in turn that Ψγ,ρ k n Yn,n → − (γ + ρ)−1 I{ρ<0} , as n → ∞. Therefore,
we obtain for γ ≥ −1/2,
γ
−1 Y
F
n
X
−
x
n,n
d
n,n
−γ−1/2
k −γ
=
+
o
k
,
(29)
p
γ
a0 nk
√
by virtue of kA0 (n/k) = O(1), and (28) thus follows directly for each ε > 0. The second
part in point 1. is ensured from (29) by the continuos mapping theorem. For γ < −1/2,
we observe from (29) that
√ Xn,n − xF d 1/2+γ n−1 Yn,n γ
=k
k
+ op (1).
γ
a0 nk
Since we are addressing the case γ+1/2 < 0, the fact that n−1 Yn,n converges in distribution
to a Fréchet r.v. suffices to conclude the proof.
Lemma 9. Suppose that U satisfies the second order
√ condition (8) with γ ∈ R and ρ ≤ 0.
If k = kn is an intermediate sequence such that k A0 (n/k) = O(1), then
n
√
√ Z 1 Xn−[2ks],n − U 2ks
ds Xn−k,n − U nk
,
(30)
k Pn , Qn := k
s
a0 nk
a0 nk
1/2
converges in distribution to the bivariate normal (P, Q) random vector with zero mean
and covariance structure given by
( −(2γ+1)
2 2
−1
2−(γ+1) −1
−
, γ 6= 0,
2
γ
2γ+1
γ+1
E(P ) =
1 − log 2,
γ = 0,
(
−γ
− √12 2 γ−1 , γ 6= 0,
E(P Q) =
log
√ 2,
γ = 0,
2
E(Q2 ) = 1.
Proof: The first component in (30) shall be tackled by Theorem 2.4.2 of de Haan and
Ferreira (2006) with k replaced by 2k therein. In particular,
n
√ Z 1 Xn−[2ks],n − U 2ks
ds
2k
n
s
a
0 k
1/2
Z
n
n √
n
n o
1 nX
a0 2k
U 2ks
− U 2k
ds
n−[2ks],n − U 2k
2k
−
=
(31)
n
n
n
s
a0 k
a0 2k
a0 2k
1/2
30
APPENDIX A: Proofs
Then, under the second order conditions (8) and (9), Theorem 2.4.2 of de Haan and
Ferreira (2006) yields for the definite integral on the right hand-side of (31):
n
√ Z 1 Xn−[2ks],n − U 2ks
ds
2k
n
s
a0 k
1/2
Z 1n
√
1
n o ds
n
−γ−1
−γ−1/2−ε
,
=
s
W
(s)
+
o
(1)s
+
o
2k
A
+
O
A
n
p
0
p
0
2γ 1/2
2k
s
k
where√{Wn (s)}n≥1 ,s > 0, denotes a sequence of Brownian motions. Under the assumption
that kA0 n/(2k) = O(1), we obtain as n → ∞,
Z 1
√
1
ds
n
k Pn = √
(2s)−γ Wn (s) 2 + Op A0
+ op (1).
s
k
2 1/2
R1
If γ = 0, the integral on the right hand side becomes 1/2 Wn (s) ds/s2 . In either case, this
integral corresponds to the sum of asymptotically multivariate normal random variables.
Now, the second component of the random vector (30) is asymptotically standard normal
(cf. Theorem 2.4.1 of de Haan and Ferreira, 2006). Finally, the covariance for the limiting
bivariate normal, E(P Q), is calculated in a straightforward way using similar calculations
to the ones in p.163 of de Haan and FerreiraR (2006).
1
Proof of Theorem 2: Let h(γ) = (log 2)−1 1/2 (2s)−γ −1)/(−γ) ds/s, which is defined
in (11). Taking the auxiliary function a0 from the second order condition (8) we write
the following normalization of x̂F (cf. (27) and (31)):
!
Z 1
n
U 2ks
− U nk
x̂F − xF
1
(2s)−γ − 1 ds
1
− h(γ) = Wn −
−
Pn + Qn −
,
log 2
log 2 1/2
γ
s
a0 nk
a0 nk
with (Pn , Qn ) defined in Lemma 9 and Wn := Xn,n −xF /a0 (n/k). Now, Lemma 9 entails
√
that k(Pn , Qn ) is asymptotically bivariate normal distributed as (P, Q). Proposition 8
expounds √
the limiting distribution of Wn provided suitable normalization, possibly different than k. Hence, the crux of the proof is in the following distributional expansion,
under the second order condition (8), for large enough n:
√
Z
F
n√
F
k
n 1 ?
1 ds o
−γ x̂ − x
−γ
−(γ+1/2)
k
−h(γ)
=
k
W
+k
kQ
−
P
+A
Ψ
.
n
n
n
0
γ,ρ
log 2
k 1/2
2s s
a0 nk
(32)
We shall consider the cases γ > −1/2, γ = −1/2 and γ < −1/2 separately.
Case γ > −1/2: Proposition 8(1) and Lemma 9 upon (32) ascertain the result, by virtue
that W = Z γ /γ with Z a standard Fréchet random variable.
Case γ = −1/2: The random component Wn is asymptotically independent of the remainder Pn and Qn . This claim is supported on Lemma 21.19 of van der Vaart
(1998). Hence, the combination of Proposition 8 with Lemma 9 ascertains the result.
Case γ < −1/2: It is now convenient to rephrase (32) with a suitable normalization in
view of Proposition 8 and the precise statement thus follows:
Z 1
√ n
√ x̂F − xF
1 ds o
Pn
n 1
?
k
−h(γ)
=
k
Q
−
−
A
Ψ
+Op (k γ+1/2 ).
n
0
log 2
k log 2 1/2 γ,ρ 2s s
a0 nk
31
APPENDIX A: Proofs
Proof of Corollary 4: The result follows immediately from Theorem 2, provided W
and N are independent random variables. Lemma 21.19 of van der Vaart (1998) ensures
the latter.
Proof of Theorem 6: The test statistic
Gn,k :=
x̂F − Xn−k,n
Xn−k,n − Xn−2k,n
expands as
Xn,n −U (n/k)
a0 (n/k)
−
U n/(2ks) −U (n/k) ds
−
s
1/2
a0 (n/k)
Xn−2k,n −U (n/k)
Qn −
a0 (n/k)
1
log 2 Pn
1
log 2
R1
,
(33)
where Pn and Qn are defined and accounted for in Lemma 9. Under the stated conditions
in the Theorem, in particular condition (8) of regular variation of second order, we have
for the remainder building blocks:
( γ
k (Yn,n /n)γ −1
Xn,n − U (n/k) d
1 + op (1) ,
γ 6= 0,
γ
=
a0 (n/k)
log(Yn,n /n) + log k 1 + op (1) , γ = 0,
( −γ
√
2 −1
Xn−2k,n − U (n/k) d
+ Op 1/ k ,
γ 6= 0,
γ
√
=
a0 (n/k)
log(1/2) + Op 1/ k , γ = 0
(cf. proof of Proposition 8), and
1
b(γ) := −
log 2
Z
1
1/2
(
U n/(2ks) − U (n/k) ds
=
a0 (n/k)
s
γ log 2−1+2−γ
+ O A0 (n/k) ,
γ 2 log 2
log 2
2 + O A0 (n/k) ,
γ=
6 0,
γ = 0.
Plugging all the blocks above back in expression (33) for Gn,k , we therefore obtain:
if γ = 0,
log 2
2
log(Yn,n /n) + b(0) − log2 2 + Op
G∗n,k (0) = log 2 Gn,k − log k +
= log 2
log 2 +
log
√k
k
Op √1k
+ Op A0 (n/k)
,
whereas, if γ 6= 0,
log 2
2
+ b(γ) −
G∗n,k (0) = log 2 Gn,k − log k +
= log 2
kγ (Yn,n /n)γ −1
γ
log 2
log
√k
2 − log k + Op
k
2−γ −1
+ Op √1k
γ
+ Op A0 (n/k)
.
Finally, since Yn,n /n is a non-degenerate random variable, eventually, as it converges to
a unit Fréchet, the statement follows for γ ∈ R.
32
APPENDIX B: Finite sample properties of x̂FRB2 , −1/2 < γ < 0
REFERENCES
8 APPENDIX B: Finite sample properties of x̂FRB2,
−1/2 < γ < 0
The illustrative example about the supercentenarian women’s records in section 5 also
shows how the second order expansion of the general endpoint estimator can be used
to remove some contribution to the asymptotic bias. This is the idea underpinning the
reduced bias version (13), provided the true negative EVI stays above −1/2.
Some finite sample results for x̂FRB2 are displayed in Figure 20. We have generate
N = 300 samples of size n = 1000 from the parent Models 1 (xF1 = 0) and 4 (xF4 = 1).
These models were introduced in the simulation study comprising section 4. Here, we have
chosen to set the EVI at the values −0.4 and −0.2. The practical application in section
5 allows to foresee (cf. Figure 17) that by reducing the bias in the general endpoint
estimator, we end up with a new estimator with larger variance. To this extent, we
have furthermore anticipated a new estimator with very similar features to the designated
MOM.inv and POTML.GPD endpoint estimators. Now, the simulation results seem to
support our “guess”. The comparative box-plots in Figure 20 show a close resemblance
of patterns within the group encompassing the three estimators FAN.RB2, MOM.inv and
POTML.GPD, although there are situations in which the reduced bias version can serve
as a good complement to the MOM.inv and POTML.GPD, particularly for the cases of
anomalous behaviour of the likelihood surface, often encountered for the GPD. Figure 20
also illustrates the distinctive behavior of the general endpoint estimator, emphasizing
lower MSE delivered by this estimator.
References
Aarssen, K. and de Haan, L. (1994). On the maximal life span of humans. Mathematical
Population Studies, 4:259–281.
Balkema, A. A. and de Haan, L. (1974). Residual life time at great age. Ann. Probab.,
2:792–804.
Bingham, N. H., Goldie, C., and Teugels, J. L. (1987). Regular Variation. Encyclopedia
of Mathematics and its Applications, vol. 27, Cambridge University Press.
Cai, J. J., de Haan, L., and Zhou, C. (2013). Bias correction in extreme value statistics
with index around zero. Extremes, 16:173–201.
Castillo, J. and Daoudi, J. (2009). Estimation of generalized Pareto distribution. Statistics
& Probabiliy Letters, 79:684–688.
Castillo, J. and Serra, I. (2015). Likelihood inference for generalized Pareto distribution.
Computational Statistics and Data Analysis, 83:116–128.
Coles, S. (2011). Is there an upper limit to human longevity? AXA Global Forum for
Longevity. Conference Proceedings from initial meetings, pages 96–103.
Davison, A. and Smith, R. (1990). Models for exceedances over high thresholds. Journal
of the Royal Statistical Society. Series B (Methodological), 52(3):393–442.
de Haan, L. (1970). On regular variation and its application to the weak convergence of
sample extremes. Mathematisch Centrum Amsterdam.
33
REFERENCES
REFERENCES
Figure 20: Boxplots of the errors ε(j, k0∗ ), j = 1, . . . , N = 300 (top) and Mean Squared Errors (MSE) plotted
against k ∗ , k ∗ ≤ n, (down) for MAX, FAN, FAN.RB2, MOM.inv and POTML.GPD endpoint estimators.
34
REFERENCES
REFERENCES
de Haan, L. and Ferreira, A. (2006). Extreme Value Theory: An Introduction. Springer.
de Haan, L., Klein Tank, A., and Neves, C. (2015). On tail trend detection: modeling
relative risk. Extremes, 18:141–178.
Drees, H., Ferreira, A., and de Haan, L. (2004). On maximum likelihood estimation of
the extreme value index. Ann. Appl. Probab., 14(3):1179–1201.
Einmahl, J. H. J. and Magnus, J. R. (2008). Records in athletics through extreme-value
theory. JASA, 103:1382–1391.
Ferreira, A., de Haan, L., and Peng, L. (2003). On optimising the estimation of high
quantiles of a probability distribution. Statistics, 37(5):40–434.
Fisher, R. A. and Tippett, L. H. C. (1928). Limiting forms of the frequency distribution
of the largest and smallest member of a sample. Math. Proc. Cambridge Philos. Soc.,
24:180–190.
Fraga Alves, M. I., de Haan, L., and Neves, C. (2013). How far can Man go? In Torelli, N.,
Pesarin, F.,and Bar-Hen, A., editors, Advances in Theoretical and Applied Statistics.
Springer, pages 185–195.
Fraga Alves, M. I. and Neves, C. (2014). Estimation of the finite right endpoint in the
Gumbel domain. Statist. Sinica, 24:1811–1835.
Fries, J. (1980). Aging, natural death, and the compression of morbidity. New England
Journal of Medicine, 303:130–135.
Girard, S., Guillou, A., and Stupfler, G. (2011). Estimating an endpoint with high order
moments. Test, 21:697–729.
Girard, S., Guillou, A., and Stupfler, G. (2012). Estimating an endpoint with high order
moments in the Weibull domain of attraction. Statist. Probab. Lett., 82:2136–2144.
Gnedenko, B. V. (1943). Sur la distribution limite du terme maximum d’une série
aléatoire. Ann. of Math., 44:423–453.
Grimshaw, S. (1993). Computing maximum likelihood estimates for the generalized pareto
distribution. Technometrics, 35(2):185–191.
Hall, P. (1982). On estimating the endpoint of a distribution. Ann. Statist., 10:556–568.
Hanayama, N. (2013). A discussion of the upper limit of human longevity based on study
of data for oldest old survivors and deaths in japan. In Proceedings 59th ISI World
Statistics Congress, pages 4518–4522.
Hüsler, J. and Peng, L. (2008). Review of testing issues in extremes: in honor of Professor
Laurens de Haan. Extremes, 11:99–111.
Kaufmann, E. and Reiss, R. (2007). About the longevity of humans. Section 19.1 of
Statistical Analysis of Extreme Values, pages 453–463.
Li, D. and Peng, L. (2010). Comparing extreme models when the sign of the extreme
value index is known. Statistics & Probability Letters, 80(7-8):739–746.
35
REFERENCES
REFERENCES
Neves, C. and Fraga Alves, M. I. (2007). Semi-parametric approach to Hasofer-Wang and
Greenwood statistics in extremes. TEST, 16:297–313.
Neves, C. and Fraga Alves, M. I. (2008). Testing extreme value conditions – an overview
and recent approaches. REVSTAT - Statistical Journal, 6:83–100.
Neves, C. and Pereira, A. (2010). Detecting finiteness in the right endpoint of light-tailed
distributions. Statistics and Probability Letters, 80:437–444.
Neves, C., Picek, J., and Fraga Alves, M. (2006). Contribution of the maximum to the
sum of excesses for testing max-domains of attraction. Journal of Statistical Planning
and Inference, 136:1281–1301.
Olshansky, J., Carnes, B., and Cassel, C. (1990a). In search of methuselah: estimating
the upper limits to human longevity. Science, 250:634–640.
Olshansky, S., Carnes, B., and Cassel, C. (1990b). In search of methuselah: estimating
the upper limits to human longevity. Science, 250:634–640.
Qi, Y. and Peng, L. (2009). Maximum likelihood estimation of extreme value index for
irregular cases. Journal of Statistical Planning and Inference, 139:3361–3376.
Smith, R. (1987). Estimating tails of probability distributions. The Annals of Statistics,
15(3):1174–1207.
Stephenson, A. and Tawn, J. (2013). Determining the best track performances of all time
using a conceptual population model for athletics records. Journal of Quantitative
Analysis in Sports, 9(1):67–76.
Troen, B. and Cristafalo, V. (2001). Principles and practice of geriatric surgery, chapter
Cell and molecular aging, pages 8–23. Springer, New York.
van der Vaart, A. W. (1998). Asymptotic Statistics. Cambridge University Press.
Vaupel, S. (2011). Longevity : What is it all about? AXA Global Forum for Longevity.
Conference Proceedings from initial meetings, pages 2–9.
Wang, J. (1995). Selection of the k largest order statistics for the domain of attraction of
the Gumbel distribution. Journal of American Statistical Association, 90:1055–106.
Wilmoth, J. and Robine, J.-M. (2003). The world trend in maximum lifespan. Population
and Development Review, 29:239–24.
Zhou, C. (2010). The extent of the maximum likelihood estimator for the extreme value
index. Journal of Multivariate Analysis, 101(4):971 – 983.
36
| 10math.ST
|
A Study of Deep Learning Robustness
Against Computation Failures
Jean-Charles Vialatte and François Leduc-Primeau
arXiv:1704.05396v1 [cs.NE] 18 Apr 2017
Abstract
For many types of integrated circuits, accepting larger failure rates in computations can be used to improve
energy efficiency. We study the performance of faulty implementations of certain deep neural networks based on
pessimistic and optimistic models of the effect of hardware faults. After identifying the impact of hyperparameters
such as the number of layers on robustness, we study the ability of the network to compensate for computational
failures through an increase of the network size. We show that some networks can achieve equivalent performance
under faulty implementations, and quantify the required increase in computational complexity.
I. I NTRODUCTION
Deep neural networks achieve excellent performance at various artificial intelligence tasks, such as speech
recognition [1] and computer vision [2]. Clearly, the usefulness of these algorithms would be increased significantly
if state-of-the-art accuracy could also be obtained when implemented on embedded systems operating with reduced
energy budgets. In this context, the energy efficiency of the inference phase is the most important, since the learning
phase can be performed offline.
To achieve the best energy efficiency, it is desirable to design specialized hardware for deep learning inference.
However, whereas in the past the energy consumption of integrated circuits was decreasing steadily with each new
integrated circuit technology, the energy improvements that can be expected from further shrinking of CMOS circuits
is small [3]. A possible approach to continue improving the energy efficiency of CMOS circuits is to operate them
in the near-threshold regime, which unfortunately drastically increases the amount of delay variations in the circuit,
which can lead to functional failures [4]. There is therefore a conflict between the desire to obtain circuits that
operate reliably and that are energy efficient. In other words, tolerating circuit faults without degrading performance
translates into energy gains. Neural networks are interesting algorithms to study in this context, because their ability
to process noisy data could also be useful to compensate for hardware failures. It is interesting to draw a parallel
with another important class of algorithms that process noisy data: decoders of error-correction codes. For the case
of low-density parity-check codes, it was indeed shown that a decoder can save energy by operating unreliably
while preserving equivalent performance [5].
Of course, a perhaps more straightforward way to decrease the energy consumption is to reduce the computational
complexity. For example, [6] proposes an approach to decrease the number of parameters in a deep convolutional
neural network (CNN) while preserving prediction accuracy, and [7, 8] propose approaches to replace floating-point
operations with much simpler binary operations. The approach of this paper is the opposite. We consider increasing
the number of parameters in the model to provide robustness that can then be traded for energy efficiency. The
two aims are most likely complementary, and the situation is in fact similar in essence to the problem of data
transmission, where it is in many practical cases asymptotically (in the size of the transmission) optimal to first
compress the data to be transmitted, and then to add back some redundancy using an error-correction code.
In this preliminary study, we consider the ability of simple neural network models to increase their robustness to
computation failures through an increase of the number of parameters. It was shown previously for an implementation
of a multilayer perceptron (MLP) based on stochastic computing that it is possible to compensate for a reduced
precision by increasing the size of the network [9]. The impact of hardware faults on CNNs is also considered in
This is an extended version of the paper with the same title published in Proc. 9th Int. Conf. on Advanced Cognitive Technologies and
Applications, Feb. 2017.
Jean-Charles Vialatte and François Leduc-Primeau are with IMT Atlantique, Brest, France. Jean-Charles Vialatte is also with Cityzen
Data, Guipavas, France.
Emails: {jc.vialatte, francois.leduc-primeau}@imt-atlantique.fr.
[10] using a different approach that consists in adding compensation mechanisms in hardware while keeping the
same network size. The deviation models that we consider in this paper attempt to represent the case in which
computation circuits are not protected by any compensation mechanism. We consider a pessimistic model and a more
optimistic deviation model, but in both cases the magnitude of the error is only limited by the bounded codomain
of the computations. We show that despite this, increasing the size of the network can sometimes compensate for
faulty computations. The results provide an idea of the reduction in energy that must be obtained by a faulty circuit
in order to reduce the overall energy consumption of the system.
The remainder of this paper is organized as follows. Section II provides a quick summary of the neural network
models used in this paper, and Section III presents the methodology used for selecting hyperparameters of the
models and for training. Section IV then describes the modeling of deviations, that is of the impact of circuit
faults on the computations. Section V discusses the robustness of the inference based on simulation results. Finally,
Section VI concludes the paper.
II. BACKGROUND AND NAMING C ONVENTIONS
A neural network is a neural representation of a function f that is a composition of layer functions fi . Each layer
function is usually composed of a linear operation followed by a non-linear one. A dense layer is such that its inputs
and outputs are vectors. Its linear part can be written as a matrix multiplication between an input x and a weight
matrix W : x 7→ W x. The number of neurons of a dense layer refers to the number of rows of W . A n-dimensional
convolutional layer is such that its inputs and outputs are tensors of rank n. Its linear part canPbe written as a
n-dimensional convolution between an input x and a weight tensor W of rank n + 2: x 7→ ( Wpq ∗n xp )∀q ,
p
where p and q index slices at the last ranks and ∗n denotes the n-dimensional convolution. The output tensor slices
indexed by p and q are called feature maps. A pooling operation is often added to convolutional layers to scale
them down.
An MLP [11] is composed only of dense layers. A CNN [12] is mainly composed of convolutional layers. For
both network types, to perform supervised classification, a dense output layer with as many neurons as classes is
usually added. Then, the weights are trained according to an optimization algorithm based on gradient descent.
III. N EURAL N ETWORK M ODELS
We consider two types of deep learning models and train them in the usual way, assuming reliable computations.
To simplify model selection, we place some mild restrictions on the hyperparameter space, since we are more
interested in the general robustness of the models, rather than in finding models with the very best accuracy. As
described below, we restrict most layers to have the same number of neurons, and the same activation function.
We also try only one optimisation algorithm, and consider only one type of weight initialization.
The first model type that we consider is an MLP network composed of L dense layers, each containing N neurons,
that we denote as MLP–L–N . The activation function used in all the layers is chosen as the rectified linear unit
(reLU) [13]. In fact, since a circuit implementation (and particularly, a fixed-point circuit implementation) can only
represent values over a bounded range, we take this into account in the training by using a clipped-reLU activation,
which adds a saturation operation on positive outputs. We note that such an activation function has been used before
in a different context, in order to avoid the problem of diverging gradients in the training of recurrent networks [14].
The use of the tanh activation was also considered, but reLU was observed to yield better accuracy.
The second model type is a CNN network composed of L convolutional C × C layers with P × P pooling of
type pool, ultimately followed by a dense layer of 200 neurons [12]. This class is denoted as CNN–L–C –P –F –
pool, where F is the number of feature maps used in each convolutional layer. Clipped-reLU activations are used
throughout. The type of pooling pool can be either “max” pooling or no pooling.
For simplicity, we opted to train the networks on the task of digit classification using the MNIST dataset [15]. In
addition to the layers defined above, each model is terminated by a dense “softmax” layer used for classification,
and containing one neuron per class. Initialization was done as suggested in [16]. To prevent overfitting during
training, a dropout [17] of 25% and 50% of the neurons have been applied on convolutional and dense layers,
respectively, except on the first layer. We used categorical crossentropy as the loss function and the “adadelta”
optimizer [18]. The batch size was 128 and we trained for 15 epochs. The saturation value of the clipped-reLU
activation is chosen as 1 as this provides a good balance between performance and implementation complexity.
Note that as the range is increased, more bits must be used to maintain the same precision, leading to larger circuits.
IV. D EVIATION M ODELS
We consider that all the computations performed in the inference phase are unreliable, either because they are
performed by unreliable circuits, or because the matrix or tensor W is stored in an unreliable memory. We do,
however, assume that the softmax operation performed at the end of the classification layer (i.e. the output layer)
is computed reliably. We assume that each layer is affected by deviations independently, and that deviations occur
independently for each scalar element of a layer output. Note that in the case of a convolutional layer, it would be
natural in practice to store each convolution kernel only once, even though each kernel is used to generate multiple
output scalar elements. The assumption that deviations occur independently for each scalar output is therefore a
simplifying assumption for the case where the tensor W of a convolutional layer is stored unreliably. However, it
remains realistic for the case where deviations are caused by the processing circuits.
We are interested in circuits with a reasonably low amount of unreliability, and therefore it is useful to partition
the outcome of a computation in terms of the occurrence or non-occurrence of a deviation event. We say that a
deviation event occurs if the output of a circuit is different from the output that would be generated by a fully
reliable circuit. The probability of a deviation event is denoted by p.
Obtaining a precise characterization of the output of a circuit when timing violations or other circuit faults are
possible requires knowledge of the specific circuit architecture and of various implementation parameters. We will
rely here on two simplified deviation models. We assume that the circuits operate on fixed-point values defined over
a certain range, which is motivated by the fact that fixed-point computations are much simpler than floating-point
ones, and reducing circuit complexity is the obvious first step in trying to improve energy efficiency. The first
deviation model is a pessimistic model that assumes that when a deviation occurs, the output is sampled uniformly
at random from the bounded output domain of that circuit. We call this deviation model conditionally uniform.
Note that we assume that the circuit output is continuous within the bounded range to avoid the need to take into
account the number of quantization levels. The second model is more optimistic, and assumes that the occurrence
of a deviation can be detected. Therefore, when a deviation occurs, we replace the output by a “neutral” value, in
this case 0. We call this deviation model the erasure model.
In a synchronous circuit, the deviations can be observed in the memory elements (registers) that separate the
logic circuits. By changing the placement of these registers, we can in effect select the point where deviations will
be taken into account. Note however that register placement cannot be arbitrary, as we also seek to have similar
logic depth separating all registers. When considering the effect of the deviations on the inference, we noticed
that robustness can be increased if deviations are sampled before the activation function of each layer. This is not
surprising, since these activation functions act as denoising operations. All simulation results presented in this paper
therefore consider that deviations are sampled before the activation function. In the case of convolutional layers,
the pooling operation is also performed after the sampling of deviations.
V. R ESULTS
The classification performance of the MLP and CNN models was evaluated using Monte-Carlo (MC) simulations
that sample deviations according to the deviation models described in Section IV. Because we wish to evaluate a large
number of neural network models, each MC simulation is performed by sampling only 10 deviation realizations.
A. Effect of some hyperparameters on robustness
Robustness refers to the ability to maintain good classification accuracy in the presence of deviations. A model
has better robustness if it achieves a lower classification error at a given p. We investigated the impact of several
hyperparameters on the robustness of the inference. We evaluated the performance using the clipped-reLU and
tanh activation functions, and found robustness to be similar in both cases. For CNN models, we also evaluated
the impact of the choice of pooling function. In this case, we found that using no pooling, rather than max pooling,
provided a slight improvement in robustness. Finally, we considered the impact of the number of layers L. The
effect of L on classification error is shown in Figure 1 for an MLP-L-N network and in Figure 2 for a CNN-L-C P -F -pool network, for the case where the inference is affected by conditionally uniform deviations. For each value
of L, we select the 5 best models based on their deviation-free performance to show the variability of the model
optimization. In both cases, the number of layers does not have a clear impact on robustness, but depending on the
value of p, increasing the number of layers can improve robustness, even when a larger number of layers decreases
·10−2
1-layer
2-layer
3-layer
4-layer
5-layer
6
error rate
5
mlp
mlp
mlp
mlp
mlp
4
3
2
1
0
1
2
3
4
p
5
·10−3
Fig. 1. Error rate of MLP under conditionally uniform deviations with prob. p. (Best viewed in color.)
·10−2
1-layer
2-layer
3-layer
4-layer
5-layer
error rate
6
4
CNN
CNN
CNN
CNN
CNN
2
0
0
1
2
3
p
4
5
·10−3
Fig. 2. Error rate of CNN under conditionally uniform deviations with prob. p. (Best viewed in color.)
performance at p = 0 (reliable computation). For example, in the case of MLP models, the best performance at
p = 0 is obtained by a 2-layer model, but using a larger number of layers improves performance for p ≥ 5 · 10−4 .
B. Fault tolerance
Since neural networks are designed to reject noise in their input, we might expect them to also be able to
reject the “noise” introduced by faulty computations. We investigate this ability for networks trained with standard
procedures, in order to provide a baseline for future targeted approaches.
We first choose an error rate target that we want the network to achieve. We then consider various deviation
probability values, and for each, look for the model with the smallest number of parameters that can achieve
or outperform the performance target under the deviation constraint. For MLP networks, the results are shown
in Figure 3 for the case of conditionally uniform deviations, and in Figure 4 for the case of erasure deviations.
Similarly, Figures 5 and 6 show the results for CNN networks, respectively for conditionally uniform and erasure
deviations.
Note that the finite set of models trained implies upper and lower bounds on the number of parameters. As a
result, the lack of robust models for some performance targets (represented by single data points at p = 0) does
not necessarily mean that deviations cannot be tolerated at that performance target, but merely that no sufficiently
4
·106
nb of params
3
2
1
0
0
0.5
1
p
1.5
2
·10−3
error=0.015
error=0.016
error=0.017
error=0.018
error=0.019
error=0.02
error=0.021
error=0.022
error=0.023
error=0.024
error=0.025
error=0.026
error=0.027
error=0.028
error=0.029
Fig. 3. Number of model parameters needed to achieve error rate target using MLP under conditionally uniform deviations.
2
·106
error=0.015
error=0.016
error=0.017
error=0.018
error=0.019
error=0.02
error=0.021
nb of params
1.5
1
0.5
0
1
2
3
p
4
5
·10−3
Fig. 4. Number of model parameters needed to achieve error rate target using MLP under erasure deviations.
large model were found within the set of models that were trained. Similarly, the saturation to a minimum number
of parameters that is observed as the performance target is decreased corresponds to the number of parameters in
the smallest trained models.
We simulate deviation probabilities on the order of 10−3 , which are already considered quite large for digital
circuit design. We can see that for both network types and for both deviation models, there are indeed many
performance targets at which performance can be preserved in the presence of computation failures by picking
a larger model from the set of pre-trained models. As might be expected, the necessary redundancy grows more
slowly for the optimistic erasure deviation model. In fact for this deviation model, CNN models are able to preserve
performance even for deviation probabilities on the order of 10−2 , as seen in Fig. 6.
We expect all the curves in Figures 3–6 to be non-decreasing, since computation noise is unlikely to improve
performance in a feedforward network. The apparent inconsistency observed in Fig. 4 for the 0.016 error target is
likely an artefact caused by the small number of deviations that are sampled in the MC simulation.
In addition to considering the absolute number of parameters required to achieve a performance target in the
presence of deviations, it is interesting to compare the number of parameters required by the faulty and reliable
implementations, to obtain the fault-tolerance efficiency of the network. Let M be a network model that achieves
our performance target when deviations occur with probability p, and Mo be the smallest model that achieves the
target under reliable computation. We define the fault-tolerance efficiency of M as n(Mo )/n(M ), where n(X)
3.5
·104
error=0.006
error=0.008
error=0.012
error=0.014
error=0.016
error=0.018
error=0.02
error=0.022
error=0.024
error=0.026
error=0.028
error=0.03
nb of params
3
2.5
2
1.5
0
0.5
1
1.5
2
p
2.5
·10−3
Fig. 5. Number of model parameters needed to achieve error rate target using CNN under conditionally uniform deviations.
·104
error=0.006
error=0.008
error=0.01
error=0.012
error=0.014
error=0.016
error=0.018
error=0.02
nb of params
3.5
3
2.5
2
1.5
0
1
2
3
p
4
5
·10−2
Fig. 6. Number of model parameters needed to achieve error rate target using CNN under erasure deviations.
denotes the number of parameters used by a model X . To evaluate efficiency empirically, we take Mo to be the
smallest available model that achieves the performance target, even though it is possible that a smaller model exists
but was not found during the training phase.
Efficiency curves for the MLP models are shown in Figures 7 and 8, and for the CNN models in Figures 9
and 10. Two characteristics of these results are worth pointing out. First, even if, in the case of CNN models,
a significant amount of deviations can be tolerated while maintaining an efficiency close to 1, there exist clear
deviation thresholds after which the efficiency starts degrading rapidly. Second, this deviation threshold becomes
smaller as the error rate target is decreased. This suggests that with this naive fault-tolerance approach, we are
forced to compromise on fault-tolerance efficiency to achieve a lower classification error rate.
VI. C ONCLUSION
In this paper, we have considered the effect of faulty computations on the performance of MLP and CNN inference,
using a pessimistic and an optimistic deviation model. We showed that using standard training procedures, it is in
many cases possible to find models that will compensate for computation failures, when the deviation probability is
on the order of 10−3 and at the cost of an increase in the number of parameters. We also studied the efficiency of
the fault-tolerance that is achieved using standard model training. Our results show that the fault-tolerance efficiency
decreases as the performance target is increased. It seems reasonable to expect that networks designed for fault-
error=0.019
error=0.021
error=0.023
error=0.025
error=0.027
error=0.029
1
efficiency
0.8
0.6
0.4
0.2
0
0
0.5
1
p
1.5
2
·10−3
Fig. 7. Fault-tolerance efficiency under an error rate target (MLP, conditionally uniform deviations).
error=0.015
error=0.016
error=0.017
error=0.018
error=0.019
error=0.02
1
efficiency
0.8
0.6
0.4
0.2
0
1
2
3
4
p
5
·10−3
Fig. 8. Fault-tolerance efficiency under an error rate target (MLP, erasure deviations).
error=0.012
error=0.014
error=0.016
error=0.018
error=0.02
error=0.022
error=0.024
error=0.026
error=0.028
error=0.03
error=0.032
error=0.034
1
efficiency
0.8
0.6
0.4
0.2
0
0.5
1
1.5
p
2
2.5
3
·10−3
Fig. 9. Fault-tolerance efficiency under an error rate target (CNN, conditionally uniform deviations).
error=0.006
error=0.008
error=0.01
error=0.012
error=0.014
error=0.016
error=0.018
error=0.02
efficiency
1
0.8
0.6
0.4
0
1
2
3
4
p
5
·10−2
Fig. 10. Fault-tolerance efficiency under an error rate target (CNN, erasure deviations).
tolerance could achieve an efficiency that only depends on the amount of deviations occurring during the inference
process, and not on the performance target. These results therefore provide a baseline for future work seeking to
identify systematic ways of designing robust deep neural networks, and highlight that one of the objectives of the
network design should be to decouple the fault tolerance from the classification performance.
R EFERENCES
[1] D. Amodei et al., “Deep speech 2: End-to-end speech recognition in english and mandarin,” CoRR, vol. abs/1512.02595, 2015.
[Online]. Available: http://arxiv.org/abs/1512.02595
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in Neural
Information Processing Systems, 2012, pp. 1097–1105.
[3] R. Dreslinski, M. Wieckowski, D. Blaauw, D. Sylvester, and T. Mudge, “Near-threshold computing: Reclaiming Moore’s law through
energy efficient integrated circuits,” Proc. of the IEEE, vol. 98, no. 2, pp. 253–266, Feb. 2010.
[4] M. Alioto, “Ultra-low power VLSI circuit design demystified and explained: A tutorial,” Circuits and Systems I: Regular Papers, IEEE
Transactions on, vol. 59, no. 1, pp. 3–29, 2012.
[5] F. Leduc-Primeau, F. R. Kschischang, and W. J. Gross, “Modeling and energy optimization of LDPC decoder circuits with timing
violations,” CoRR, vol. abs/1503.03880, 2015. [Online]. Available: http://arxiv.org/abs/1503.03880
[6] F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer, “SqueezeNet: AlexNet-level accuracy with 50x
fewer parameters and <1mb model size,” CoRR, vol. abs/1602.07360, 2016. [Online]. Available: http://arxiv.org/abs/1602.07360
[7] M. Courbariaux and Y. Bengio, “Binarynet: Training deep neural networks with weights and activations constrained to +1 or -1,”
CoRR, vol. abs/1602.02830, 2016. [Online]. Available: http://arxiv.org/abs/1602.02830
[8] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, “XNOR-Net: ImageNet classification using binary convolutional neural
networks,” CoRR, vol. abs/1603.05279, 2016. [Online]. Available: http://arxiv.org/abs/1603.05279
[9] A. Ardakani, F. Leduc-Primeau, N. Onizawa, T. Hanyu, and W. J. Gross, “VLSI implementation of deep neural network using integral
stochastic computing,” IEEE Trans. on VLSI Systems, 2017, to appear.
[10] Y. Lin, S. Zhang, and N. R. Shanbhag, “Variation-tolerant architectures for convolutional neural networks in the near threshold voltage
regime,” in 2016 IEEE International Workshop on Signal Processing Systems (SiPS), Oct 2016, pp. 17–22.
[11] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural networks, vol. 2,
no. 5, pp. 359–366, 1989.
[12] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE,
vol. 86, no. 11, pp. 2278–2324, 1998.
[13] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural networks,” in International Conference on Artificial Intelligence
and Statistics, 2011, pp. 315–323.
[14] A. Y. Hannun et al., “Deep speech: Scaling up end-to-end speech recognition,” CoRR, vol. abs/1412.5567, 2014. [Online]. Available:
http://arxiv.org/abs/1412.5567
[15] Y. LeCun, C. Cortes, and C. J. Burges, “The MNIST database of handwritten digits,” 1998.
[16] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in International conference on
artificial intelligence and statistics, 2010, pp. 249–256.
[17] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from
overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.
[18] M. D. Zeiler, “ADADELTA: an adaptive learning rate method,” CoRR, vol. abs/1212.5701, 2012. [Online]. Available:
http://arxiv.org/abs/1212.5701
| 9cs.NE
|
1
Rate-cost tradeoffs in control
arXiv:1612.02126v2 [cs.IT] 19 Oct 2017
Victoria Kostina, Babak Hassibi
Abstract—Consider a control problem with a communication
channel connecting the observer of a linear stochastic system
to the controller. The goal of the controller is to minimize
a quadratic cost function in the state variables and control
signal, known as the linear quadratic regulator (LQR). We
study the fundamental tradeoff between the communication
rate r bits/sec and the expected cost b. We obtain a lower
bound on a certain rate-cost function, which quantifies the
minimum directed mutual information between the channel
input and output that is compatible with a target LQR cost.
The rate-cost function has operational significance in multiple
scenarios of interest: among others, it allows us to lowerbound the minimum communication rate for fixed and variable
length quantization, and for control over noisy channels. We
derive an explicit lower bound to the rate-cost function, which
applies to the vector, non-Gaussian, and partially observed
systems, thereby extending and generalizing an earlier explicit
expression for the scalar Gaussian system, due to Tatikonda
el al. [2]. The bound applies as long as the system noise has
a probability density function. It can be closely approached
by a simple lattice quantization scheme that only quantizes
the innovation, that is, the difference between the controller’s
belief about the current state and the true state. Via a separation principle between control and communication, similar
results hold for causal lossy compression of additive noise
Markov sources. Apart from standard dynamic programming
arguments, our technical approach leverages the Shannon
lower bound, develops new estimates for data compression
with coding memory, and uses some recent results on high
resolution variable-length vector quantization to prove that
the new converse bounds are tight.
Index Terms—Linear stochastic control, LQR control, distributed control, rate-distortion tradeoff, high resolution,
causal rate-distortion theory, Shannon’s lower bound.
matrices of dimensions n × n and n × m, respectively. At
time i, the controller observes output Gi of the channel, and
chooses a control action Ui based on the data it has observed
up to time i. At time i, the encoder observes the output of
the sensor Yi ∈ Rk :
Yi = CXi + Wi ,
(2)
k
where C is a k × n deterministic matrix, and Wi ∈ R is
the observation noise. The encoder forms a codeword Fi ,
which is then passed on to the channel. Like the controller,
the encoder has access to the entire history of the data it
has observed. See Fig. 1.
We assume that system noises V1 , V2 . . . are zero-mean,
i.i.d. independent of everything else, and that observation
noises W1 , W2 , . . . are also zero-mean, i.i.d. independent
of everything else. We assume that X1 is zero-mean and
independent of {Vi , Wi }∞
i=1 . Furthermore, we make the
usual assumption that the pair (A, B) is controllable and
that the pair (A, C) is observable. If the encoder observes
the full system state, i.e. Yi = Xi (rather than its noisecorrupted version as in (2)), then we say that the system is
fully observed (rather than partially observed).
CONTROLLER
Ui
SYSTEM
Gi
CHANNEL
Fi
I. I NTRODUCTION
ENCODER
Yi
SENSOR
A. System model
Consider the following discrete time stochastic linear
system:
Xi+1 = AXi + BUi + Vi ,
n
Fig. 1: The distributed control system.
(1)
n
where Xi ∈ R is the state, Vi ∈ R is the process noise,
Ui ∈ Rm is the control action, and A and B are fixed
The authors are with California Institute of Technology (e-mail:
[email protected], [email protected]).
A part of this work was presented at the 54th Annual Allerton Conference
on Communication, Control and Computing [1].
The work of Victoria Kostina was supported in part by the National
Science Foundation (NSF) under Grant CCF-1566567. The work of Babak
Hassibi was supported in part by the National Science Foundation under
grants CNS-0932428, CCF-1018927, CCF-1423663 and CCF-1409204, by
a grant from Qualcomm Inc., by NASA’s Jet Propulsion Laboratory through
the President and Director’s Fund, and by King Abdullah University of
Science and Technology.
Notation: Capital letters X, Y denote (spacial) random
vectors; X t , (X1 , . . . , Xt ) denotes temporal random
vectors, or the history of vector samples up to time t; Xit ,
(Xi , . . . , Xt ) (empty if t < i); X ∞ , (X1 , X2 , . . .); for i.i.d.
random vectors, X denotes a random vector distributed the
same as each of X1 , X2 , . . .; Drepresents a delay by one, i.e.
DX t , (0, X t−1 ); ΣX , E (X − E [X])(X − E [X])T
denotes the covariance matrix of random vector X. Sansserif capitals A, B denote constant matrices; A ≥ B (> B)
signifies that A − B is positive semidefinite (definite); In is
the n × n identity matrix; 0k×n is the k × n all-zero matrix.
Lowercase letters a, b, . . . denote known scalars.
show that R(b) provides a lower bound on the minimum
capacity of the channel necessary to sustain LQR cost b.
We will also show that over noiseless channels, R(b) can be
closely approached by a simple variable-length lattice-based
quantization scheme that transmits only the innovation.
B. The rate-cost tradeoff
The efficiency of a given control law at time t is measured
by the linear quadratic regulator (LQR) cost function:1
(3)
LQR X t , U t−1
#
" t−1
X
XiT QXi + UiT RUi + XtT St Xt ,
,E
C. Prior art
i=1
The analysis of control under communication constraints
has a rich history. The first results on the minimum data rate
required for stabilizability appeared in [6], [7]. These works
analyze the evolution of a scalar system from a worst-case
perspective. In that setting, the initial state X1 is assumed
to belong to a bounded set, the process noise V1 , V2 , . . . is
assumed to be bounded, and the system is said to be stabilizable if there exists a (rate-constrained) control sequence
such that the worst-case deviation of the system state from
the target state 0 is bounded: lim supt→∞ kXt k < ∞. In
[6], [7], it was shown that a fully observed unstable scalar
system can be kept bounded by quantized control if and
only if the available data rate exceeds log A bits per sample.
Tatikonda and Mitter [8] generalized this result to vector
systems; specifically, they showed that the necessary data
rate to stabilize a vector system with bounded noise is at
least
X
log |λi (A)|,
(7)
r>
where Q ≥ 0, R ≥ 0 and St ≥ 0. The LQR cost balances
between the deviation of the system from the desired state 0,
and the required control power defined with respect to the
norms induced by the matrices Q, R and St . In the special
case Q = In , R = 0 and St+1 = In , the cost function in (3)
is the average mean-square
deviation
of the system from the
hP
i
t
2
desired state 0, E
i=1 kXi k .
Given a joint distribution of random vectors F t and Gt ,
the directed mutual information is defined as [3]
t
X
I(F i ; Gi |Gi−1 ).
(4)
I(F t → Gt ) ,
i=1
Directed mutual information, which captures the information
due to causal dependence of Gt on F t , and which is less
than or equal to the full mutual information I(F t ; Gt ), has
proven useful in communication problems where causality
and feedback play a role. Given a joint distribution PF t Gt , it
is enlightening to consider causally conditional probability
kernel [4]
t
Y
PGi |Gi−1 ,F i
(5)
PGt ||F t ,
i : |λi (A)|≥1
where the sum is over the unstable eigenvalues of A,
i.e. those eigenvalues whose magnitude exceeds 1. Compellingly, (7) shows that only the nonstable modes of A
i=1
matter; the stable modes can be kept bounded at any
Note that PF t Gt = PF t kDGt PGt kF t .
arbitrarily small quantization rate (and even at zero rate if
The following information-theoretic quantity will play
Vt ≡ 0). In a setting without system noise, Yüksel and Başar
a central role in determining the operational fundamental
[9] showed that variants of (7) remain necessary under three
limits of control under communication constraints.
different stability criteria. Using a volume-based argument,
Definition 1 (rate-cost function). The rate-cost function for Nair et al. [10] showed a lower bound to quantization rate
the system in (1), (2), is defined as
in order to attain lim supt→∞ kXt k ≤ d, thereby refining
Nair et al. [10] also presented an achievability scheme
(7).
P
:
1
t
t
U kY
I(Y t → U t ) : 1 LQR(X t , U t−1 )≤b (6) confirming that for scalar systems, that bound is tight.
R(b) , lim inf
t→∞
t
t
Nair and Evans [11] showed for systems with unbounded
process and observation disturbances, Tatikonda and Mitter’s
Note that the system dynamics (1), (2) fixes the kernel condition on the rate (7) continues to be necessary and
PY t kDU t , while the optimization in (6) is over the causal sufficient in order to keep the mean-square deviation of
channels PU t kY t (i.e. the conditional probability distribu- the plant state from 0 bounded, that is, in order to satisfy
tions PUi |U i−1 ,Y i , i = 1, . . . , t) comprising the encoder, the lim sup
2
< ∞.
t→∞ E kXt k
channel and the decoder in Fig. 1 that satisfy the LQR cost
Nair and Evans’ converse bound [11] applies to fixed-rate
constraint.
quantizers, that is, to compressors whose outputs can take
In this paper, we will show a simple lower bound to one of 2r values. Nair and Evans [11] showed that timethe rate-cost function (6). Although R(b) does not have invariant fixed-rate quantizers are unable to attain bounded
a direct operational interpretation unless the channel is cost if the noise is unbounded [11], regardless of their
probabilistically matched [5] to the system, it is linked to the rate. The reason is that since the noise is unbounded, over
minimum data rate required to keep the system at LQR cost time, a large magnitude noise realization will inevitably
b, over both noiseless and noisy channels. Namely, we will be encountered, and the dynamic range of the quantizer
will be exceeded by a large margin, not permitting re1 As common in information theory, here we abuse the notation slightly
t
t−1
covery. Brockett and Liberzon [12] proposed an adaptive
and write LQR X , U
to mean that LQR () is a function of the
quantization scheme, which “zooms out” (i.e. increases the
joint distribution of X t , U t−1 .
2
Causal rate-distortion function is challenging to evaluate
exactly, and beyond scalar Gauss-Markov source [2], [26],
no closed-form expression is known for it. For stationary
scalar Gaussian processes, Derpich and Ostergaard [34]
showed an upper bound and Silva et al. [32] a lower bound.
For vector Gauss-Markov sources, Tanaka et al. developed
a semidefinite program to compute exactly the minimum
directed mutual information in quantization [35] and control
[36].
quantization intervals) when the system is far from the
target and “zooms in” to drive the system state further to
0 when the system is close to the target. Nair and Evans
[11] proposed a stabilizing quantization scheme in which the
number of quantization levels is finite at each step but varies
with time. Yüksel [13], [14] considered fixed-rate quantizers
with adaptive bin sizes and showed that there exists a
unique invariant distribution for the system state and the
quantizer parameters in [13], and studied the existence and
the structure of optimal quantization and control policies in
[14]. Structural properties of optimal zero-delay quantizers
for the compression Markov sources were investigated in
[15]–[23]. Structural properties of quantizers embedded in
control feedback loop are studied in [24].
D. Our contribution
In this paper, we show a lower bound to R(b) of a fully
observed system, which holds as long as the system noises
Vi are continuous random variables. We do not require the
noise to be bounded or Gaussian. We also show that (7)
remains necessary to keep the LQR cost bounded, even if
the system noise is non-Gaussian, generalizing previously
known results. Although our converse lower bound holds
for a general class of codes that can take full advantage of
the memory of the data observed so far and that are not
constrained to be linear or have any other particular structure, we show that the new bound can be closely approached
within a much more narrow class of codes. Namely, a simple
variable-length quantization scheme, which uses a lattice
covering and which only transmits the difference between
the controller’s estimate about the current system state and
the true state, performs within a fraction of a bit from
the lower bound, with a vanishing gap as b approaches its
minimum attainable value, bmin . The scheme is a variant
of a classical differential pulse-code modulation (DPCM)
scheme, in which a variable-length lattice code used to
encode the innovation process. Unlike previously proposed
quantization schemes with provable performance guarantees,
our scheme does not use dither.
Our results generalize to partially observed systems,
where the encoder does not have access to Xi but only to
its noise-corrupted version, Yi . For those results to hold, we
require the system and observation noises to be Gaussian.
Our approach is based on a new lower bound to causal
rate-distortion function, termed the causal Shannon lower
bound (Theorem 9 in Section III below), which holds
for vector Markov sources with continuous additive disturbances, as in (1). For scalar Gauss-Markov source, the bound
coincides with a previously known expression [26].
In variable-rate (or length) quantization, the quantizer
can have a countably infinite number of quantization cells.
Entropy coding is applied to encode the indices of quantization cells, so that the more likely quantization cells have
a shorter description and the less likely ones a longer one.
That way, once in a while a rare source realization might
cause a longer string of bits to be transmitted. Elia and
Mitter [25] considered stabilization of a noiseless linear
system controlled with a variable-length scalar quantizer,
and showed that for a certain notion of coarseless, the
coarsest quantizer is logarithmic (i.e. its levels follow a
logarithmic law).
Beyond worst case and mean-square stabilizability,
Tatikonda et al. [2] considered a setting known as linear
quadratic Gaussian (LQG) control ((1), (2) with Gaussian
disturbances and quadratic cost function in (3)) with communication constraints and tied the minimum attainable
LQG cost to the Gaussian causal rate-distortion function,
introduced decades earlier by Gorbunov and Pinsker [26],
which is equal to the minimal (subject to a distortion constraint) directed mutual information between the stochastic
process and its representation [27]. Stabilizability of LQG
systems under a directed mutual information constraint was
studied in [28]. The problem of minimizing an arbitrary cost
function in control of a general process under a directed mutual information constraint was formulated in [29]. Control
of general Markov processes under a mutual information
constraint was studied in [30]. Silva et al. [31] elucidated
the operational meaning of directed mutual information, by
pointing out that it lower-bounds the rate of a quantizer
embedded into a feedback loop of a control system, and
by showing that the bound is approached to within 1 bit E. Technical approach
The main idea behind our approach to show a conby a dithered prefix-free quantizer, a compression setting
verse
(impossibility) result is to recursively lower-bound
in which both the compressor and the decompressor have
distortion-rate
functions arising at each step. We apply the
access to a common dither - a random signal with special statistical properties. More recently, Silva et al. [32] computed classical Shannon lower bound to distortion-rate function
a lower bound to the minimum quantization rate in scalar [37], which bounds the distortion-rate function X in terms
Gaussian systems with stationary disturbances and proposed of the entropy power of X, and we use the entropy power
a dithered quantization scheme that performs within 1.254 inequality [38], [39] to split up the distortion-rate functions
bits from it. Tanaka et al. [33] generalized the results of [32] of the sums of independent random variables. Since Shanto vector systems.
non’s lower bound applies as long as the source random
3
in Section V, and the achievability schemes are presented
in Section VI.
variable has a density, our technique circumvents a precise
characterization of the distribution of the state at each time
instant. The technique also does not restrict the system
noises to be Gaussian.
To show that our bound can be approached at high rates,
we build on the ideas from high resolution quantization
theory. A pioneering result of Gish and Piece [40] states that
in the limit of high resolution, a uniform scalar quantizer
incurs a loss of only about 12 log2 2πe
12 ≈ 0.254 bits per
sample. Ziv [41] showed that regardless of target distortion,
the normalized output entropy of a dithered scalar quantizer
exceeds that of the optimal vector quantizer by at most
4πe
1
2 log 12 ≈ 0.754 bits per sample. A lattice quantizer
presents a natural extension of a scalar uniform quantizer
to multiple dimensions. The advantage of lattice quantizers
over uniform scalar quantizers is that the shape of their
quantization cells can be made to approach a Euclidean
ball in high dimensions. Indeed, relying on a fundamental
result by Rogers [42] and crediting Poltyev, Zamir and
Feder [43, (25)] showed that the logarithm of the normalized
second moment of the best n-dimensional lattice quantizer
1
, at a speed O logn n .
converges to that of a ball, log 2πe
The works of Gersho [44], Zamir and Feder [45] and Linder
and Zeger [46] established that the entropy rate of dithered
lattice quantizers converges to Shannon’s lower bound in the
limit of vanishing distortion.
While the presence of a dither signal both at the encoder
and the decoder greatly simplifies the analysis and can
improve the quantization performance, it also complicates
the engineering implementation. In this paper, we do not
consider dithered quantization. Neither do we rely directly
on the classical heuristic reasoning by Gish and Piece [40].
Instead, we use a non-dithered lattice quantizer followed by
an entropy coder. To rigorously prove that its performance
approaches our converse bound, we employ a recent upper
bound [47] on the output entropy of lattice quantizers in
terms of the differential entropy of the source, the target
distortion and a smoothness parameter of the source density.
II. M AIN
RESULTS
A. Fully observed system
In the absence of communication constraints, the minimum quadratic cost due to the system noise attainable in
the limit of infinite time is well known to be
bmin = tr(ΣV S),
(8)
where ΣV is the covariance matrix of each of the V1 , V2 , . . .,
and S is the solution to the algebraic Riccati equation
S = Q + AT (S − M) A,
T
(9)
T
M , L (R + B SB)L
−1
T
= SB(R + B SB)
T
−1
L , (R + B SB)
T
(10)
T
B S,
B S.
(11)
(12)
Our results quantifying the overhead over (8) due to
communication constraints are expressed in terms of the
entropy power of the system and observation noises. The
entropy power of an n-dimensional random vector X is
defined as2
1
2
N (X) ,
exp
h(X) ,
(13)
2πe
n
R
where h(X) = − Rn fX (x) log fX (x)dx is the differential
entropy of X, and fX (·) is the density of X with respect
to the Lebesgue measure on Rn .
The entropy power satisfies the following classical inequalities:
1
1
(14)
N (X) ≤ (det ΣX ) n ≤ Var [X] .
n
The first equality in (14) is attained if and only if X
is Gaussian and the second if and only if X is white.
Thus, N (X) is equal to the normalized variance of a white
Gaussian random vector with differential entropy h(X).
Our first result is a lower bound on the rate-cost function.
Theorem 1. Consider the fully observed linear stochastic
system (1). Suppose that V has a density. At any LQR cost
The rest of the paper is organized as follows. In Section II, b > tr(ΣV S), the rate-cost function is bounded below as
!
1
we state and discuss our main results: Section II-A focuses
N (V )| det M| n
n
, (15)
on the scenario where the observer sees the system state R(b) ≥ log | det A| + log 1 +
2
(b − tr(ΣV S))/n
(fully observed system), Section II-B discusses a generalization to the scenario where the observer sees a noisy where M is defined in (11).
measurement of the system state (partially observed system),
The bound in Theorem 1 is nontrivial if M > 0, which
and Section II-C discusses the operational implications
happens if rank B = n and either Q > 0 or R > 0. The
of our bounds in the settings of fixed-rate quantization,
bound in Theorem 1 continues to hold whether or not at
variable-rate quantization and joint source-channel coding.
time i the encoder observes the previous control inputs
In Section III, we introduce the causal lossy compression
U1 , U2 , . . . , Ui−1 .
problem, and we state the causal Shannon lower bound,
The right-hand side of (15) is a decreasing function of
together with a matching achievability result. In Section IV,
b, which means that the controller needs to know more
we discuss separation between control, estimation and communication, a structural result that allows us to disentangle
2 All log’s and exp’s are common arbitrary base specifying the informathe three tasks. The proofs of the converse results are given tion units.
F. Paper organization
4
information about the state of the system to attain a smaller
target cost. As an important special case, consider the ratecost tradeoff where the goal is to minimize the mean-square
deviation from the desired state 0. Then, Q = In , R = 0,
S = M = In , bmin = Var [V ], and (15) particularizes as
n
N (V )
R(b) ≥ log | det A| + log 1 +
. (16)
2
(b − Var [V ])/n
In another important special case, namely Gaussian V , (15)
particularizes as
!
1
n
| det ΣV M| n
R(b) ≥ log | det A| + log 1 +
. (17)
2
(b − tr(ΣV S))/n
For the scalar Gaussian system, (17) holds with equality.
This is a consequence of known analyses [26], [2].
A typical behavior of (16) is plotted in Fig. 2 as a function
of target cost b. As b ↓ bmin , the required rate R(b) ↑ ∞.
Conversely, as b ↑ ∞, the rate monotonically decreases and
approaches log | det A|. The rate-cost tradeoff provided by
Theorem 1 can serve as a gauge for choosing an appropriate
communication rate in order to meet the control objective.
For example, in the setting of Fig. 2, decreasing the data
rate below 1 nat per sample incurs a massive penalty in cost,
because the rate-cost function is almost flat in that regime.
On the other hand, increasing the rate from 1 to 3 nats per
sample brings a lot of improvement in the attainable cost,
while further increasing it beyond 3 nats results in virtually
no improvement.
Also plotted in Fig. 2 is the output entropy of a variablerate uniform scalar quantizer that takes advantage of the
memory of the past only through the innovation, i.e. the
difference between the controller’s prediction of the state at
time t given the information the controller had at time t − 1
and the true state (see Section VI for a precise description
of the quantizer). Its performance is strikingly close to the
lower bound, being within 0.5 nat even at large b, despite
the fact that quantizers in this class cannot attain the optimal
cost exactly [48]. The gap further vanishes as b decreases.
The gap can be further decreased for multidimensional
systems by taking advantage of lattice quantization. These
effects are formally captured by the achievability result we
are about to present, Theorem 2.
The directed entropy is defined as [4]
t
X
t
t
H(Gi |Gi−1 , F i ).
(18)
H(G kF ) ,
Fig. 2: The minimum quantizer entropy compatible with cost b
in a fully observed system (1) with parameters n = 1, A = 2,
B = Q = R = 1, V has Laplace distribution with variance 1.
On the other hand, there exists a variable-rate quantizer
that keeps the system at cost b and whose average encoded
length does not exceed H(b) (see Section II-C for details).
Thus, unlike R(b), the function H(b) has a direct operational
interpretation.
Theorem 2, presented next, holds under the assumption
that the density of the noise is sufficiently smooth. Specifically, we adopt the following notion of a regular density.
Definition 3 (Regular density, [49]). Let c0 ≥ 0, c1 ≥ 0.
Differentiable probability density function fX of a random
vector X ∈ Rn is called (c0 , c1 )-regular if 3
k∇fX (x)k ≤ (c1 kxk + c0 )fX (x),
Replacing in Definition 1 the directed mutual information
by the directed entropy, we introduce
Theorem 2. Consider the fully observed linear stochastic
system (1). Suppose that M > 0 and that V has a regular
density. Then, at any LQR cost b > tr(ΣV S)), the entropy-
Definition 2 (entropy-cost function). The entropy-cost function is defined as
o
n
PU t kY t :
(19)
H(b) , lim inf H(U t kY t ) : 1 LQR(X t , U t−1 )≤b
i
Since I(Y ; Ui |U
t
i−1
) ≤ H(Ui |U i−1 , Y i ), we have
H(b) ≥ R(b).
(21)
A wide class of densities satisfying smoothness condition
(21) is identified in [49]. Gaussian, exponential, uniform,
Gamma, Cauchy, Pareto distributions are all regular. Convolution with Gaussians produces a regular density: more
precisely, the density of B + Z, with B ⊥⊥ Z and Z ∼
N (0, σ 2 I), is ( σ42 E [kBk] , σ32 )-regular. 4 Likewise, if the
density of Z is (c0 , c1 )-regular, then that of B + Z, where
kBk ≤ b a.s., B ⊥⊥ Z is (c0 + c1 b, c1 )-regular.
i=1
t→∞
∀x ∈ Rn .
3 As
usual, ∇ denotes the gradient.
B⊥
⊥ Z reads “B is independent of Z”.
4 Notation
(20)
5
blocks. Consider the orthogonal projection matrix onto that
space, Πℓ ΠTℓ . Specifically, Πℓ is a 0-1-valued n × ℓ matrix
given by:
I
(24)
Πℓ , ℓ
0
cost function is bounded by
1
n
!
n
N (V )| det M|
log 1 +
2
(b − tr(ΣV S))/n
1
(22)
+ O1 (log n) + O2 (b − tr(ΣV S)) 2 ,
H(b) ≤ log | det A| +
where O1 (log n) ≤ C1 log n and O2 (ξ) ≤ C2 min {ξ, c2 }
The improvement of Theorem 1 can now be stated as
for some nonnegative constants C1 , C2 and c2 . Furthermore, follows.
(22) is attained by a DPCM scheme with asymptotically
Theorem 3. Consider the fully observed linear stochastic
uniform temporal rate and distortion allocations.
system (1). Suppose that V has a density. Let Λ be a
The first two terms in (22) match the first two terms in diagonal matrix such that M′ , JT MJ ≥ Λ, where J is
(15). The O1 (log n) term is the penalty due to the shape defined in (23). At any LQR cost b > tr(ΣV S), the rateof lattice quantizer cells not being exactly spherical. In cost function is bounded below as
Section VI, we provide aprecise expression
for that term
1
ℓ
µ′ N (ΠTℓ J−1 V )
R(b) ≥ ℓ log a′ + log 1 +
. (25)
for n = 1, 2, . . .. The O2 (b − tr(ΣV S)) 2 is the penalty
2
(b − tr(ΣV S))/ℓ
due to the distribution of the innovation not being uniform.
It becomes negligible for small b − tr(ΣV S), and the speed where Πℓ is defined in (24),
1
of that convergence depends on the smoothness parameters
a′ , det ΠTℓ J−1 AJΠℓ ℓ ,
(26)
of the noise density.
1
(27)
µ′ , (det Λ) ℓ
Theorem 2 implies that if the channel Fi → Gi is
noiseless, then there exists a quantizer with output entropy
In particular, by letting ℓ be the number of unstable
given by the right side of (22) that attains LQR cost eigenvalues, Theorem 3 implies that
X
b > tr(ΣV S)), when coupled with an appropriate controller.
log |λi (A)|.
(28)
R(b) ≥
In fact, the bound in (22) is attainable by a simple lattice
i : |λi (A)|≥1
quantization scheme that only transmits the innovation of
If ℓ = n, then a′ = |det A|, and the bound in (25) reduces
the state (a DPCM scheme). The controller acts as if it
observed a noise-corrupted version of the state: it forms its to (15).
best estimate of the state based on the quantized data, and
If the dimensionality of control is less than that of the
computes the control action based on that estimate.
system, m < n, the bounds in Theorem 1 and Theorem 3
Theorem 1 gives a lower (converse) bound on the output reduce to log | det A| and ℓ log a′ , respectively, losing the
entropy of quantizers that achieve the target cost b, without dependence on b. The bound in Theorem 4 below is a
making any assumptions on the quantizer structure and decreasing function of b, even if m < n.
permitting the use of the entire history of observation data.
Theorem 4. Consider the fully observed linear stochastic
Theorem 2 proves that the converse can be approached
system (1). Suppose that V has a density. At any LQR cost
by a strikingly simple quantizer coupled with a standard
b > tr(ΣV S), the rate-cost function is bounded below as
controller, without common randomness (dither) at the en
n
µN (V ) m m
a2
coder and the decoder. Furthermore, although nonuniform R(b) ≥ log | det A| + m log
,
+
2
| det A|2
b − tr(ΣV S)
rate allocation across time is allowed by Definition 2, such
(29)
freedom is not needed to achieve (22); the scheme that
achieves (22) satisfies H(Ut |U t−1 , Y t ) → r in the limit where
1
of large t.
! 2im
i
iT T
det
LA
Σ
A
L
V
Although the bound in Theorem 1 is tight at low b
,
(30)
a , inf
i≥1
det ΣV det (LLT )
(as demonstrated by Theorem 2), if A has both stable
1
and unstable eigenvalues and b is large, it is possible to
µ , det R + BT SB det LLT m .
(31)
improve the bound in Theorem 1 by projecting out the stable
modes of the system. Towards this end, consider Jordan
If m = n, Theorem 4 reduces to Theorem 1.
decomposition of A:
We conclude Section II-A with a few technical remarks.
A = JA′ J−1 ,
(23) Remark 1. Since
is bounded
hP the running
i mean-square cost
t
1
2
2
≤
max
E
kX
kX
k
where A′ is the Jordan form of A, and J is invertible. Without above as t E
1≤i≤t
i k , our
i
i=1
loss of generality, assume the eigenvalues of A′ are ordered Theorem 3 implies (via (28)) the weaker result of Nair
in the decreasing magnitude order. Write A′ as a direct sum and Evans
[11],
who showed the necessity of (28) to keep
of its Jordan blocks, A′ = A1 ⊕ . . . ⊕ As̄ . For some s such supt E kXt k2 bounded. Note also that the approach of
that 1 ≤ s ≤ s̄ ≤ n, let ℓ = dim(A1 ) + . . . + dim(As ) be Nair and Evans [11] applies only to fixed-rate quantization,
the dimension of the the column space of the first s Jordan while our approach encompasses both fixed- and variable6
all Gaussian. At any target LQR cost b > bmin , the rate-cost
function is bounded below as
!
1
n
(det NM) n
R(b) ≥ log | det A| + log 1 +
. (38)
2
(b − bmin )/n
rate quantization, as well as control over noisy channels
and non-Gaussian system disturbances.
Remark 2. Tatikonda et al. [2, (17)] proposed to apply
the classical reverse waterfilling, known to achieve the
noncausal Gaussian rate-distortion function, at each step
The bound in Theorem 5 is nontrivial if both M > 0
i, to compute the causal Gaussian rate-distortion function
and
N > 0. The necessary condition for that is rank B =
for vector Xi . Unfortunately, reverse waterfilling is only
rank
C = n. In the fully observed system, C = K = I,
suboptimal, as can be verified numerically by comparing
P
=
N = ΣV , ΣW = 0, and (38) reduces to (17). As
[2, (17)] to the semidefinite program of Tanaka et al. [35].
we
will
see in Section V, the key to proving Theorem 5 is
The reason reverse waterfilling does not apply is that λi (t)
to
represent
the evolution of the current encoder’s estimate
in [2, (15)] depend on the distortion threshold, an effect
of
the
state
in terms of its previous best estimate and an
not present in the computation of classical non-causal rateindependent
Gaussian, as carried out by the Kalman filter
distortion function.
recursion. It is for this technique to apply that we require the
noises to be Gaussian.5 As in the fully observed case, the
B. Partially observed system
observer quantizes the innovation, only now the innovation
Consider now the scenario in which the encoder sees is the difference between the observer’s estimate and the
only a noisy observation of the system state and forms controller’s estimate. The scheme operates as follows. The
a codeword to transmit to the controller using its present observer sees Yi , recursively computes its state estimate,
and past noisy observations. If the system noise and the computes the innovation, quantizes the innovation. The
observation noise are jointly Gaussian, our results in Section controller receives the quantized value of the innovation,
recursively computes its state estimate, and forms the control
II-A generalize readily.
In the absence of communication constraints, the cele- action that is a function of controller’s state estimate only.
brated separation principle of estimation and control states Accordingly, the tasks of state estimation, quantization and
that the minimum cost in (32) can be attained by separately control are separated. The momentous insight afforded by
estimating the value of Xi using the noisy observations Theorem 6 below is that in the high rate regime, this simple
Y i , and by applying the optimal control to the estimate as scheme performs provably close to R(b), the best rate-cost
if it was the true state of the system. The minimum cost tradeoff theoretically attainable.
decomposes into two terms, the cost due to the noise in
Theorem 6. Consider the partially observed linear stochasthe system (see (8)), and the cost due to the noise in the
tic system (1), (2). Suppose that M > 0, N > 0, and that
observation of the state:
X1 , V and W are all Gaussian. At any LQR cost b > bmin ,
bmin = tr(ΣV S) + tr ΣAT MA ,
(32) the entropy-cost function is bounded above as
!
1
n
where Σ is the covariance
matrix
of
the
estimation
error
n
|det
NM|
H(b) ≤ log | det A| + log 1 +
Xt − E Xt |Y t , U t−1 in the limit t → ∞. If system
2
(b − bmin )/n
noise and the observation noise are Gaussian, the optimal
1
+ O1 (log n) + O2 (b − bmin ) 2 ,
(39)
estimator admits a particularly elegant implementation via
the Kalman filter. Then, Σ is given by
where O1 (log n) ≤ C1 log n and O2 (ξ) ≤ C2 min {ξ, c2 }
(33) for some nonnegative constants C1 , C2 and c2 . Furthermore,
Σ = P − K CPCT + ΣW KT ,
where ΣW is the covariance matrix of each of the (39) is attained by a DPCM scheme with asymptotically
W1 , W2 , . . ., P is the solution to the algebraic Riccati uniform temporal rate and distortion allocations.
equation
Theorems 3 and 4 generalize as follows.
P = APAT − AK CPCT + ΣW KT AT + ΣV ,
and K is the steady state Kalman filter gain:
−1
.
K , PCT CPCT + ΣW
(34)
Theorem 7. Consider the partially observed linear stochastic system (1), (2). Suppose that (A, B) is controllable and
(A, C) is observable. Suppose further that X1 , V and W
are all Gaussian. At any LQR cost b > bmin , the rate-cost
function is bounded from below as
1
η ′ µ′
R(b) ≥ log a′ + log 1 +
.
(40)
2
(b − bmin )/ℓ
(35)
As Theorem 5 below shows, the rate-cost function is
bounded in terms of the steady state covariance matrix of
the innovation in encoder’s state estimate:
N , K CPCT + ΣW KT
(36)
T
= AΣA − Σ + ΣV
where a′ , µ′ are defined in (26)-(27), and
1
η ′ , det ΠTℓ J−1 NJ−1 T Πℓ ℓ ,
(37)
Theorem 5. Consider the partially observed linear stochastic system (1), (2). Suppose further that X1 , V and W are
5 Note
7
that Theorem 1 makes no such restrictions.
(41)
where J and Πℓ are defined in (23), (24), respectively.
significant performance improvement. One such technique
that approaches (45) in a particular scenario is discussed
Theorem 8. Consider the partially observed linear stochasin [53]. In general, how closely the bound in (45) can be
tic system (1), (2). Suppose that (A, B) is controllable, (A, C)
approached over noisy channels remains an open problem.
is observable, and that m ≤ k ≤ n. Suppose further that
2) Control under fixed-rate quantization: If the channel
X1 , V and W are all Gaussian. At any LQR cost b > bmin ,
connecting the encoder to the controller is a noiseless bit
the rate-cost function is bounded from below as
pipe that accepts a fixed number of r bits per channel use
ηµ
a2
m
(so
that C = r), both R(b) and H(b) are lower bounds on
,
+
R(b) ≥ log | det A| + log
2
| det A|2
(b − bmin )/m
the minimum quantization rate r required to attain cost b:
(42)
R(b) ≤ H(b) ≤ r.
(46)
where µ is defined in (31), and
Therefore, the converse results in Theorems 1, 3, 4, 5, 7
1
! 2im
det LAi K CPCT + ΣW KT Ai T LT
and 8 give sharp lower bounds on the minimum size of
,
a , inf
T
T
T
a fixed-rate quantizer compatible with LQR cost b. The
i≥1
det (CPC + ΣW ) det (LL ) det (K K)
achievability
results in Theorems 2 and 6 are insufficient
1
η , det CPCT + ΣW det KT K m .
(43) to establish the existence of a fixed-rate quantizer of rate
If k = m = n, the result in Theorem 8 reduces to that in approaching R(b). While attempting to find a time-invariant
fixed-rate quantizer operating at any finite cost is futile
Theorem 5.
[11], determining whether there exists an adaptive fixed-rate
quantization scheme approaching the converses in Theorems
C. Operational implications
1 and 5 remains an intriguing open question.
3) Control under variable-length quantization: Assume
So far we have formally defined the rate-cost / entropycost functions as the limiting solutions of a minimal directed now that the channel connecting the encoder to the controller
mutual information / entropy subject to an LQR cost con- is a noiseless channel that accepts an average of r bits per
straint (Definition 1 / Definition 2) and presented lower and time unit. That is, the channel input alphabet is the set
upper bounds to those functions. In this section we discuss of all binary strings, {∅, 0, 1, 00, 01, . . .}, and the encoding
the operational implications of the results in Section II-A function must be such that
t
and Section II-B in several communication scenarios.
1X
E [ℓ(Fi )] ≤ r,
(47)
1) Control over a noisy channel: Consider first a general
t i=1
scenario in which the channel in Fig. 1 is a dynamical
channel defined by a sequence of conditional probabilities where ℓ(·) denotes the length of the binary string in its
argument.
{PGi |Gi−1 ,F i }∞
i=1 . For the class of directed information
The minimum encoded average length L⋆ (X) in lossless
stable channels, the capacity of the channel is given by [50],
compression of object X is bounded as [54], [55]
[51]
1
t
t
L⋆ (X) ≤ H(X)
(48)
(44)
C = lim sup I(F → G ),
t→∞ t
≤ L⋆ (X) + log2 (L⋆ (X) + 1) + log2 e.
(49)
where the sup is over all causally conditioned kernels
PF t kDGt . If past channel outputs are not available at the Note that (48) states that the optimum compressed length
encoder, then the sup is over all PF t , and I(F t → Gt ) = is below the entropy. This is a consequence of lifting the
prefix condition: without prefix constraints one can indeed
I(F t ; Gt ).
The following result, the proof of which is deferred until compress at an average rate slightly below the entropy [54],
Section IV, implies that the converse results in Theorems [56].
Consider the operational rate-cost function for control
1, 3, 5 and 7 present lower bounds on the capacity of the
under variable-rate quantization, Rvar (b), defined as the
channel necessary to attain b.
liminf of r such that LQR cost b and (47) are achievable in
Proposition 1. A necessary condition for stabilizing the the limit of large t. It follows from (48) that
system in (1), (2) at LQR cost b is,
Rvar (b) ≤ H(b),
(50)
R(b) ≤ C.
(45)
which implies the existence of a variable-rate quantizer
One remarkable special case when equality in (45) is whose average rate does not exceed the expressions in
attained is control of a scalar Gaussian system over a scalar Theorems 2 and 6. Likewise, by (49) and Jensen’s inequality,
memoryless AWGN channel [2], [52], [53]. In that case,
H(b) ≤ Rvar (b) + log2 (Rvar (b) + 1) + log2 e,
(51)
the channel is probabilistically matched to the data to be
transmitted [5], and no coding beyond simple amplification which leads via (20) to the lower bound Rvar (b) ≥
is needed. In practice, such matching rarely occurs, and intel- ψ −1 (R(b)), where ψ −1 (·) is the functional inverse of
ligent joint source-channel coding techniques can lead to a ψ(x) = x + log2 (x + 1) + log2 e. Thus the bounds
8
in Theorems 1, 3, 4, 5, 7 and 8 provide lower bounds
to Rvar (b). Consequently, our converse and achievability
results in Section II-A and Section II-B characterize the
operational rate-cost tradeoff for control with variable rate
quantizers.
by subtracting fi (Z i ) prior to encoding and adding it back
after decoding. Formally,
Remark 3. The minimum average compressed length L⋆p (X)
among all prefix-free lossless compressors satisfies H(X) ≤
L⋆p (X) ≤ H(X) + 1. Therefore, the minimum average
rate of a prefix-free quantizer compatible with cost b is
bounded as R(b) ≤ Rvar,p (b) ≤ H(b) + 1. Accordingly, the
theorems in Section II-A and Section II-B also characterize
the operational rate-cost tradeoff for control with variable
rate prefix-free quantizers.
Classical DPCM schemes, described next, fit in the
model (57).
DPCM encoder, upon observing the current source sample Si and the history of decoder’s outputs Ŝ i−1 , computes
the state innovation S̃i recursively using the formula
III. C AUSAL S HANNON
LOWER BOUND
i=1
(59)
i
i
DPCM decoder, having recovered S̃ˆi , forms its state
estimate as
Ŝi = Ŝi|i−1 + S̃ˆi .
(61)
Noting that (60) is just a special case with (57) for Z ∞ =
DŜ ∞ , we conclude:
i=1
H(Ŝi |Ŝ i−1 , S i , Z i ).
HS ∞ kZ ∞ (d) = HS̄ ∞ kZ ∞ (d).
S̃i , Si − Ŝi|i−1 ,
(60)
i
where Ŝi|i−1 , E Si |Ŝ i−1 is the a priori (predicted) state
estimate at the decoder, and sends S̃ˆi , an encoded value of
the innovation S̃ . We allow S̃ˆ to depend on past history.
where Wi ≥ 0. Understanding causal compression bears
great independent interest and is also vital to understanding
control, as clarified in Section IV below.
Causally conditioned directed information and causally
conditioned directed entropy are defined as:
t
X
t
t
t
I(S i ; Ŝi |Ŝ i−1 , Z i ),
(53)
I(S → Ŝ kZ ) ,
H(Ŝ t kS t , Z t ) ,
(58)
h
In this section, we consider causal compression of a
discrete-time random process, S ∞ = (S1 , S2 , . . .), under
the weighted mean-square error:
t
h
i
X
E (Si − Ŝi )T Wi (Si − Ŝi ) ,
(52)
WSE S t , Ŝ t ,
i=1
t
X
RS ∞ kZ ∞ (d) = RS̄ ∞ kZ ∞ (d),
Proposition 2. Both the causal rate-distortion and the dentropy function are attained in the class of DPCM.
Another implication of (58)–(59) is the following. Consider the Markov process
Si+1 = ASi + Vi ,
(62)
where V1 , V2 , . . . are i.i.d. random vectors. The uncontrolled
process (62) and the controlled system process (1) are
related through
(54)
X t = St +
Definition 4 (causal rate- and entropy-distortion functions).
The causal rate- and entropy-distortion functions under the
weighted mean-square error in (52) with side information
Z ∞ causally available at both encoder and decoder are
defined as, respectively,
1
I(S t → Ŝ t kZ t ), (55)
RS ∞ kZ ∞ (d) , lim
inf
t→∞ PŜ t kS t ,Z t : t
t
t
1
t WSE(S , Ŝ )≤d
1
H(Ŝ t kS t , Z t ). (56)
HS ∞ kZ ∞ (d) , lim
inf
t→∞ PŜ t kS t ,Z t : t
t
t
1
t WSE(S , Ŝ )≤d
t−1
X
At−1−i BUi .
(63)
i=0
Provided that both encoder and decoder have access to past
controls, we have
RS ∞ (d) = RX ∞ kDU ∞ (d).
(64)
HS ∞ (d) = HX ∞ kDU ∞ (d),
(65)
where DU ∞ signifies causal availability of past controls
at both encoder and decoder. Furthermore, using Ŝi|i−1 =
AŜi−1 , X̂i|i−1 = AX̂i−1 + BUi−1 , it is easy to show that
both processes (62) and (63) have the same innovation
process: X̃i = S̃i . Thus the same DPCM scheme can be
In the absence of side information, the conditional di- used to encode both.
rected mutual information / entropy reduce to the unconThe following bound is a major result of this paper.
ditional ones, and we write RS ∞ (d) / HS ∞ (d) for their
Theorem 9. For the Markov process (62) and a sequence
unconditional versions.
of weight matrices such that
Consider two processes {S̄i }, and {Si }, related as
1
S̄i = Si + fi (Z i ),
lim (det Wt ) n = w,
(57)
t→∞
where fi (·) is an arbitrary function of common side information. A quantizer for {Si } can be constructed from that
for {S̄i }, without altering the rate-distortion performance,
causal rate-distortion function is bounded below as
n
wN (V )
2
∞
,
RS (d) ≥ log a +
2
d/n
9
(66)
(67)
Theorem 12 below is a decreasing function of d, even if
Wi is singular, or if Vi is supported on a subspace of Rn .
where
1
n
a , | det A| .
(68)
The proof of Theorem 9 uses the classical Shannon lower
bound. Theorem 9 can thereby be viewed as an extension of
Shannon’s lower bound to the causal rate-distortion function.
Remark 4. In lieu of the long-term average constraint in (52),
[2], [26] considered a more stringent constraint in which the
average distortion is bounded at each time instant i. Tanaka
[57] showed recently that for Gauss-Markov sources, both
formulations are equivalent, as optimal rates and distortions
settle down to their steady state.
Remark 5. For scalar Gauss-Markov sources, equality holds
in (67), recovering the known result [2], [26]
wVar [V ]
1
2
∞
.
(69)
RS (d) = log a +
2
d
Theorem 12. Consider the uncontrolled process (62) and
the distortion in (52). Assume that the weight matrices Wi
satisfy limi→∞ Wi = LT L, where V is an m × n matrix,
m ≤ n. Suppose further that Vi = Ki Vi′ , where Vi′ is a
k-dimensional random vector with covariance matrix ΣV ′ ,
where k ≥ m, and Ki are n × k matrices such that
limi→∞ Ki = K. The causal rate-distortion function is
bounded below as
!
k
m
wN (V ′ ) m
2
RS ∞ (d) ≥
,
(73)
log a +
2
d/m
where
Theorem 10, stated next, shows that the converse in
Theorem 9 can be approached by a DPCM quantization
scheme with uniform rate and distortion allocations (in the
limit of infinite time horizon). This implies that nonuniform
rate and distortion allocations permitted by (52) and (53)–
(54) cannot yield significant performance gains.
1
! 2im
det LAi KΣV ′ KT Ai T LT
,
a = inf
i≥1
det ΣV ′ det (LLT ) det (KT K)
1
w = det LLT det KT K m .
(74)
(75)
If k = m = n, the result in Theorem 12 recovers that in
Theorem 9.
IV. C ONTROL , ESTIMATION AND COMMUNICATION
Theorem 10. For the Markov process (62) such that V has
SEPARATED
a regular density and a sequence of weight matrices such
that (66) holds, the causal entropy function is bounded by
1
An early quantization-control separation result for
n
wN (V )
2
HS ∞ (d) ≤ log a +
+ O1 (log n) + O2 d 2 , Gaussian systems was proposed by Fischer [58].
2
d/n
(70) Tatikonda et al. [2] leveraged separation in their analysis
of quantized LQG control and tied it to the concept of
where O1 (log n) ≤ C1 log n and O2 (ξ) ≤ C2 min {ξ, c2 } control’s dual effect. Here, we observe that separated design
for some nonnegative constants C1 , C2 and c2 . Furthermore, is optimal even for non-Gaussian system and observation
(70) is attained by a DPCM scheme with asymptotically noises, and for control over noisy channels, as long as both
uniform temporal rate and distortion allocations.
the encoder and the controller have access to past control
We conclude Section III with two extensions of The- signals.
Denote
orem 9. If the dynamic range of the eigenvalues of A
is large, the bound in Theorem 11 below, obtained by
Li , (R + BT Si+1 B)−1 BT Si+1 ,
(76)
projecting out a subset of smaller eigenvalues of A, can
T
−1 T
Mi , Si+1 B(R + B Si+1 B) B Si+1
(77)
provide an improvement over Theorem 9 at large d. Recall
T
T
= Li (R + B Si+1 B)Li ,
(78)
the definition of matrices J, Πℓ in (23), (24), respectively.
where St is that in (3), and Si , 1 ≤ i ≤ t − 1 is the solution
to the discrete-time Ricatti recursion,
Theorem 11. Consider the uncontrolled process (62) and
the distortion in (52). Consider matrices Vi such that 0 ≤
ViT Vi ≤ JT Wi J, and Vi commutes with Πℓ ΠTℓ . The causal
rate-distortion function is bounded below as,
ℓ
w′ N (ΠTℓ J−1 V )
RS ∞ (d) ≥ log a′ 2 +
,
(71)
2
d/ℓ
Si = Q + AT (Si+1 − Mi ) A.
The optimal control of partially observed system in (1),
(2) in the scenario of Fig. 1, for a fixed dynamical channel
{PGi |Gi−1 ,F i }∞
i=1 can be obtained as follows:
where a′ is defined in (26), and
w′ , lim inf det ΠTℓ VtT Vt Πℓ
t→∞
m1
.
(79)
(72)
In the case of a singular Wi , or Vi not having a density
with respect to the Lebesgue measure on Rn , the bounds in
Theorem 9 and Theorem 11 reduce to n log a and n log a′ ,
respectively, losing the dependence on d. The bound in
10
(i) The encoder computes the optimal state estimate
X̂i , E Xi |Y i , U i−1 ;
(80)
(ii) The encoder maps the optimal control signal Ui⋆ =
−Li AX̂i to channel codeword Fi ;
(iii) Having received the channel output Gi , the decoder
computes Ui and applies it to the system.
where bmin is the minimum cost attainable without communication constraints given in (32), and causal rate- and
entropy-distortion functions are evaluated with weight matrices
Wi = AT Mi A.
(92)
The encoder PFi |Gi−1 ,X̂ i and decoder PUi |Gi are designed
to
minimize the weighted long-term average distortion
Pt−1
i=0 di , where
di , E (Ui − Ui⋆ )T (R + BT Si+1 B)(Ui − Ui⋆ ) . (81)
More precisely, we can write down the resulting LQR
cost as follows.
If Vi and Wi are both Gaussian, then the optimal estimator
(80) admits a particularly simple recursive implementation,
given by the Kalman filter. At time i, having observed Yi ,
the Kalman filter forms an estimate of the system state using
Yi and the prior estimate X̂i−1 as follows (e.g. [59]):
Theorem 13. The LQR cost in the scenario of Fig. 1 for
the partially observed system (1), (2) separates as,
t−1
t−1
t−1
X
X
X
LQR X t , U t−1 =
ci +
ei +
di ,
i=0
i=1
(82)
i=1
P1|0 , ΣX1 .
Eliminating Xi from (85) by substituting Xi = AXi−1 +
BUi−1 + Vi−1 into (85), completing the squares and using
that Vi−1 is zero-mean and independent of Xi−1 , Ui−1 , we
re-write (85) equivalently as
T
(86)
bi = bi−1 + E Vi−1
Si Vi−1 + qi−1 ,
where
h
i
qi , E (Li AXi + Ui )T(R + BT Si+1 B)(Li AXi + Ui ) .(87)
Applying (86) repeatedly, we obtain
(88)
i=1
which is equivalent to (82) in the fully observed scenario,
i.e. when X̂i = Xi .
To show the more general case, observe first that for
random variables X, Y and X̂ forming a Markov chain
X − Y − X̂, it holds with X̂ ⋆ , E [X|Y ] that,
i
i
h
i
h
h
E kX − X̂k2 = E kX − X̂ ⋆ k2 + E kX̂ ⋆ − X̂k2 . (89)
i
i−1
Applying (89) to Li AXi − (Y , U )− Ui , qi separates into
two terms: qi = ei + di , and (88) is thereby re-written as
(82).
An immediate corollary to Theorem 13 is the following.
H(b) = HX̂ ∞ kDU ∞ (b − bmin ),
Cov(Ỹi ) = CPi|i−1 CT + ΣW ,
Cov(Xi − X̂i ) = (I − Ki C)Pi|i−1 ,
and the estimation error is thus given by
ei = tr I − Ki C)Pi|i−1 AT Mi A .
(97)
(98)
(99)
(100)
We can now prove the converse for control over noisy
channels in Proposition 1. Transmitting S1 , S2 , . . . over a
causal feedback channel can be represented pictorially as
follows:
/F t h
/ Ŝ t ,
/ Gt
St
(101)
D
where arrows represent causal dependence, and D represents
delay by one. Joint distributions allowable by the diagram
in (101) take form PS t PF t kS t ,DGt PGt kF t PŜ t kGt . The following converse result holds.
Proposition 3. A necessary condition for the existence of an
encoder and decoder pair (PF t kS t ,DGt , PŜ t kGt ) achieving
distortion d in the communication scenario (101) is,
RS ∞ (d) ≤ C.
(102)
Proof. By data processing of mutual information and [3,
Th.1, Th. 3], we have I(S t → Ŝ t ) ≤ I(S t ; Ŝ t ) ≤
I(S t ; Gt ) ≤ I(F t → Gt ), and thus in the communication
scenario (101), 6
I(S t → Ŝ t ) ≤ I(F t → Gt ).
Corollary 1. The rate-cost and the entropy-cost functions
for control of the partially observed system (1), (2) are given
by, respectively,
R(b) = RX̂ ∞ kDU ∞ (b − bmin ),
(96)
The covariances of the innovation and the estimation error
are given by, respectively,
j=1
i=1
(94)
Pi+1|i = A (I − Ki C) Pi|i−1 AT + ΣV ,
Proof. Denote for 1 ≤ i ≤ t
i−1
X
bi , E
XjT QXj + UjT RUj + E XiT Si Xi . (85)
t−1
X
qi ,
E ViT Si+1 Vi +
Ỹi , Yi − CAX̂i−1 − CBUi−1
is the innovation, Gaussian and independent of X̂i−1 , Ui−1
and Ỹ t−1 , and the Kalman filter gain is found from the
Riccati recursion:
−1
,
(95)
Ki , Pi|i−1 CT CPi|i−1 CT + ΣW
and c0 , tr (ΣX1 S1 ).
t−1
X
(93)
where
where communication distortions di are defined in (81), and
control and estimation costs are given by
ci , E ViT Si+1 Vi = tr (ΣV Si+1 ) ,
(83)
T
(84)
ei , E Xi − X̂i AT Mi A Xi − X̂i .
bt = b1 +
X̂i = AX̂i−1 + BUi−1 + Ki Ỹi ,
(103)
Now, (102) follows by taking a sup of the right side
over PF t kS t ,DGt , and an inf of the left side over PŜ t kS t
satisfying the distortion constraint.
(90)
6 Note that (103) continues to hold even if the encoder and decoder
mappings are noncausal.
(91)
11
chain condition X − LX − X̂, since for such X̂, I(X; X̂) ≤
I(LX; X̂).
Proposition 1 follows by plugging (64) and (90) in (102).
V. C ONVERSE THEOREMS : TOOLS AND PROOFS
We start by introducing a few crucial definitions and tools,
some classical, some novel, that form the basis of our proof
technique.
The (single-shot) distortion-rate function with respect to
the weighted mean-square distortion is defined as follows.
Remark 6. As a consequence of Proposition 4, we may
always assume that X has uncorrelated components when
computing distortion-rate functions. Indeed, let L be the
orthogonal transformation that diagonalizes the covariance
Definition 5 (conditional distortion-rate function). Let X ∈ matrix of X. Since LT L = I, by Proposition 4 the distortionRn be a random variable, and M ≥ 0 be an n × n matrix. rate functions (under mean-square error) of X and LX
The distortion-rate function under weighted mean-square coincide.
error with side information U at encoder and decoder is
Define conditional entropy power as
defined as
h
i
1
2
N (X|U ) ,
exp
h(X|U )
(111)
Dr,M (X|U ) , min E (X − X̂)T M(X − X̂) . (104)
2πe
n
PX̂|XU :
R
I(X;X̂|U)≤r
where h(X|U ) = − Rn fX,U (x, u) log fX|U=u (x)dx is the
If no side information is available, i.e. U ≡ 0, we denote conditional differential entropy of X.
the corresponding unconditional distortion-rate function by
The following tool will be instrumental in our analysis.
Dr,M (X). The distortion-rate function under mean-square
error distortion corresponds to M = I. In that case we denote Theorem 14 (Shannon’s lower bound [37]). The conditional
distortion-rate function is bounded below as
for brevity
Dr (X|U ) ≥ Dr (X|U )
(112)
Dr (X|U ) , Dr,I (X|U )
(105)
, nN (X|U ) exp (−2r/n) .
The following Proposition equates the distortion-rate functions under weighted and non-weighed mean-square error.
If X is a white Gaussian vector when conditioned on
U , (112) holds with equality. Although beyond Gaussian
X, Shannon’s lower bound is rarely attained with equality
[60], it is approached at high rates [61]. The tightness of
Shannon’s lower bound at high rates is key to arguing that
the bound in Theorem 1 can in fact be approached.
Proposition 4. Let X ∈ Rn be a random variable, and let
L be an m × n matrix. The following equality holds.
Dr (LX|U ) = Dr, LT L (X|U )
(113)
(106)
Proof. We show the analog of (106) for Dr (LX); the
conditional version in (106) is analogous. First, we prove
that
h
i
Dr (LX) , min
E (LX − X̂)T (LX − X̂)
(107)
X̂ : I(LX;X̂)≤r
h
i
=
min
E (X − X̂)T LT L(X − X̂) . (108)
If L is square, the entropy power scales as
2
N (LX|U ) = | det L| n N (X|U ),
(114)
providing a convenient expression for Dr (LX|U ) in terms
of the conditional entropy power of X. The following
proposition generalizes the scaling property (114) to the case
where the multiplying matrix is not square. 8
X̂ : I(LX;LX̂)≤r
Restricting the domain of minimization in (107) to X̂ ∈
Im(L),7 we observe that ≤ holds in (108). To show the
reverse inequality, let Π be the orthogonal projection matrix
onto the column space of L, and use kΠxk ≤ kxk, ΠLx =
Lx and data processing for mutual information to write
h
i
Dr (LX) ≥
min
E (LX − ΠX̂)T (LX − ΠX̂) .
Proposition 5. Let X ∈ Rn be a random vector with
covariance ΣX > 0, let m ≤ n, and let L be an m × n
matrix with rank m. Then,
! m1
n
det LΣX LT
(N (X)) m
(115)
N (LX) ≥
det ΣX
X̂ : I(LX;ΠX̂)≤r
It follows that ≥ also holds in (108).
Now, we show that
h
i
Dr,LT L (X) , min
E (X − X̂)T LT L(X − X̂) (109)
X̂ : I(X;X̂)≤r
h
i
= min
E (X − X̂)T LT L(X − X̂) . (110)
Proof. Without loss of generality, assume E [X] = 0. Observe that
1
h(X) = log ((2πe)n det ΣX )−D (PX kN (0, ΣX )), (116)
2
where D(·k·) is the relative entropy. By the data processing
7 The image of a linear transformation described by matrix L is the span
of its column vectors.
8 Proposition 5 is stated for the unconditional case for simplicity only;
naturally, its conditional version also holds.
X̂ : I(LX;LX̂)≤r
By data processing, ≥ holds in (110). The reverse direction
follows by restricting to X̂ ∈ Im(L) satisfying the Markov
Equality holds in (115) if m = n or if X is Gaussian.
12
where ν2 is found from
n
1X
min{ν1 , σi2 }
log
= r.
2 i=1
min{ν1 , ν2 , σi2 }
of relative entropy,
D (PX kN (0, ΣX ))
≥ D PLX kN (0, LΣX LT )
(117)
1
(118)
= log (2πe)m det LΣX LT − h(LX),
2
and (115) follows by substituting (118) into (116) and
applying (13).
(125)
On the other hand,
Dr+s (X) =
n
X
i=1
min{ν, σi2 },
(126)
where ν is determined from
n
Another essential component of our analysis, the condi1X
σi2
log
= r + s.
(127)
tional entropy power inequality (EPI), follows by convexity
2 i=1
min{ν, σi2 }
from the unconditional EPI first stated by Shannon [38] and
On the other hand, adding (123) and (125), we obtain
proved by Stam [39].
n
σi2
1X
Theorem 15 (Conditional entropy power inequality). If X
log
= r + s.
(128)
2 i=1
min{ν1 , ν2 , σi2 }
and Y are conditionally independent given U , then
N (X + Y |U ) ≥ N (X|U ) + N (Y |U ).
(119) Comparing (127) and (127), we conclude that ν =
min{ν1 , ν2 }. It follows that (124) and (126) coincide, as
In causal data compression, the quantized data at cur- desired.
rent step creates the side information for the data to be
compressed at the next step. The following bound to the Remark 7. Since Shannon’s lower bound is equal to the
distortion-rate function minimized over side information will distortion-rate function of a white Gaussian vector with the
same differential entropy as the original vector, Proposibe vital in proving our converse theorems.
tion 6 implies that
Proposition 6. Let X ∈ Rn be a random variable. The
Dr (X|U ) = Dr+s (X).
(129)
min
following inequality holds:
U∈Rn : I(X;U)≤s
min
U∈Rn : I(X;U)≤s
Dr (X|U ) ≥ Dr+s (X).
(120)
We are now equipped to prove our converse theorems.
Q
Proof of Theorem 9. For any code i PŜi |S i ,Ŝ i−1 , denote
the per-stage information rates
If X is Gaussian, equality in (120) holds and is achieved
by a Gaussian U .
ri , I(S i ; Ŝi |Ŝ i−1 ).
Proof. Write
min
U:
I(X;U)≤s
Dr (X|U ) =
≥
min
min
U:
Y:
I(X;U)≤s I(X;Y |U)≤r
min
U,Y :
I(X;U,Y )≤r+s
= Dr+s (X),
E kX + Y k
(130)
i−1
For any code, using ri ≥ I(Si ; Ŝi |Ŝ ) and Shannon’s
lower bound (Theorem 14), we lower-bound the distortion
at step i as
h
i
(131)
E (Si − Ŝi )T Wi (Si − Ŝi ) ≥ Dri , Wi (Si |Ŝ i−1 )
2
E kX + Y k2 (121)
(122)
≥ wi di ,
(132)
where (121) is by the chain rule of mutual information, where we denoted for brevity
I(X; U ) + I(X; Y |U ) = I(X; U, Y ).
(133)
d0 , 2−2r0 nN (S0 ),
To verify that equality holds for Gaussian X, we need to
i−1
(134)
di , Dri (Si |Ŝ ), i = 1, 2, . . . , t − 1,
exhibit a U that achieves equality in (120). By Remark 6, we
1
may restrict our attention to X ∼ N (0, diag(σ12 , . . . , σn2 )).
n
wi , (det Wi ) ,
(135)
Let U be the distribution that achieves the distortion-rate
i−1
function of X at rate s, i.e. X = U + Z, where U = and (131) is due to ri ≥ I(Si ; Ŝi |Ŝ ), while (132) is by
N (0, diag(σi2 − di )), and Z ∼ N (0, diag(d1 , . . . , dn )) is Shannon’s lower bound (Theorem 14), Proposition 4 and
independent of U . Here di = min{ν1 , σi2 } is the reverse (114). To lower-bound di , 1 ≤ i ≤ t, we establish the
waterfilling solution (e.g. [62]), where the water level ν1 is following recursion (1 ≤ i ≤ t):
determined from
(136)
di = Dri (ASi−1 + Vi−1 |Ŝ i−1 )
n
2
X
i−1
1
σi
(137)
≥ Dri (ASi−1 |Ŝ ) + Dri (V )
log
= s.
(123)
2 i=1
min{ν1 , σi2 }
i−2
(138)
≥ Dri−1 +ri (ASi−1 |Ŝ ) + Dri (V )
2
−2r
i
Given U = u, the random variable X = U + Z is
a di−1 + nN (V )
=2
(139)
distributed as N (u, diag(d1 , . . . , dn )), so
where (137) is by the entropy power inequality, (138) is
n
X
due to (129). Note that (136)–(139) holds for an arbitrary
min{ν1 , ν2 , σi2 },
(124)
Dr (X|U ) = Dr (N ) =
encoded sequence Ŝ0 , . . . , Ŝi−1 , including the optimal one.
i=1
13
and Πℓ ΠTℓ commutes with Li , we may apply (148) as
follows:
h
i
E (Si′ − Ŝi′ )T JT Wi J(Si′ − Ŝi′ )
(149)
h
i
≥ E (Si′ − Ŝi′ )T Πℓ ΠTℓ JT Wi JΠℓ ΠTℓ (Si′ − Ŝi′ )
(150)
Now, assumption (66) implies that
∀ǫ > 0 ∃tǫ : ∀i ≥ tǫ , wi ≥ w − ǫ.
(140)
Rewriting (139) as
1
1
ri ≥ log(a2 di−1 + nN (V )) − log di ,
(141)
2
2
and defining d0 to satisfy a2 d0 + nN (V ) = nN (S1 ), we
deduce
t
t
X
X
ri ≥
(142)
ri ≥
i=1
≥ wi′ d′i ,
where
d′i , Dri (ΠTℓ Si′ |Ŝ i−1 ), i = 1, 2, . . . , t,
1
a2 dtǫ −1 + nN (V ) 1
log
+
2
a2 dt + nN (V )
2 i=t
ǫ
(152)
and (151) is by Shannon’s lower bound. Furthermore, observing that Πℓ ΠTℓ commutes with A′ and that ΠTℓ Πℓ ΠTℓ =
ΠTℓ , we conclude that
i=tǫ
t
X
(151)
N (V )
(143)
log a2 +
di /n
ΠTℓ A′ = ΠTℓ A′ Πℓ ΠTℓ .
The distortion constraint in (55) and the bound (132) imply
that
t
t
1X
w−ǫ X
di ≤
(144)
wi di ≤ d.
t i=t
t i=1
(153)
Using (153), the entropy power inequality and (129), we
establish the following recursion:
′
′
+ ΠTℓ Vi−1
|Ŝ i−1 )
d′i = Dri (ΠTℓ A′ Si−1
ǫ
≥
t
, which
In particular, (144) implies that dt−1 ≤ d w−ǫ
2
together with a dtǫ −1 + nN (V ) ≥ n min{N (S1 ), N (V )}
implies that the first term in (143) normalized by t is
bounded below by a quantity
that vanishes as t → ∞.
Since the function x 7→ log a2 + nNx(V ) is convex and
decreasing, by Jensen’s inequality and (144) the sum in
(143) is bounded below as
t−1
X
N (V )
log a2 +
di /n
i=tǫ
(w − ǫ)N (V ) t − tǫ
2
≥ (t − tǫ ) log a +
,
(145)
d/n
t
Diving both sides of (145) by t and taking taking a limt→∞
followed by a limǫ→0 , we obtain (67).
=
=
≥
′
′
|Ŝ i−1 ) + Dri (ΠTℓ Vi−1
)
Dri (ΠTℓ A′ Si−1
T ′
T ′
i−1
T ′
Dri (Πℓ A Πℓ Πℓ Si−1 |Ŝ ) + Dri (Πℓ Vi−1 )
′
′
|Ŝ i−1 ) + Dri (ΠTℓ Vi−1
)
a′ 2 Dri (ΠTℓ Si−1
−2ri
′2 ′
T ′
2
a di−1 + mN (Πℓ V )
(154)
(155)
(156)
(157)
(158)
The rest of the proof follows that of Theorem 9.
Proof of Theorem 12. It is easy to see (along the lines of
(140), (142)) that if Wi → LT L, we may put Wi ≡ LT L
without affecting the (causal) rate-distortion function. Similarly, if Vi = Ki Vi′ and Ki → K, we may put Vi = KVi′ . We
will therefore focus on bounding the distortion with weight
matrix LT L and with Vi = KVi′ .
Adopting the convention S1 ≡ V0 , note that (62) implies,
Si =
Proof of Theorem 11. It is enough to consider the case
w′ > 0. The case w′ = 0 will follow by bounding the
distortion with weight matrices Wi + ǫI and taking ǫ → 0.
We start by making two useful observations. First, putting
Si′ , J−1 Si and Vi′ , J−1 Vi , we write
′
Si+1
= A′ Si′ + Vi′ ,
i−1
X
Ai−j−1 Vj
(159)
j=0
Fixing causal reproduction vector Ŝ i , for 0 ≤ j ≤ i − 1,
consider the random variable
h
i
Ṽj , Vj − E Vj |Ŝ i
(160)
h
i
i
= Vj − E Vj |Ŝj+1
,
(161)
(146)
and we observe that causal rate-distortion functions of S ∞
and S ′∞ are related as
where we explicitly indicated the weight matrices in the
subscript.
Second, observe that if 0 ≤ ΠT Π ≤ I, and Π commutes
with L, then
where (161) holds because Vj is independent of Ŝ j . Note
that different Ṽj ’s are uncorrelated. Indeed, to show that Vi
i+j
t
and Vi+j are uncorrelated, put X , Ŝi+j+1
, Y , Ŝi+1
,
note that Vi+j ⊥⊥ Y , Vi+j ⊥⊥ Vi , and thus
i
h
E Ṽi Ṽi+j
ΠT LT L Π ≤ LT L.
= E [(Vi − E [Vi |X, Y ])E [(Vi+j − E [Vi+j |X])|Vi , X, Y ]]
RS ∞ ,{Wi } (d) = RS ′ ∞ ,{JT Wi J} (d),
(147)
(148)
the rateDue to (147) we may focus on evaluating
2
distortion function for S ′ ∞ . Since Πℓ ΠTℓ
≤ Πℓ ΠTℓ ≤ I
14
= E [(Vi − E [Vi |X, Y ])(Vi+j − E [Vi+j |X])]
(162)
= 0.
(163)
Using (159) and the fact that different Ṽj ’s are uncorre-
covered by Theorem 3. Assume rank B = n. According to
Corollary 1, Theorem 9 and (64), it suffices to lower bound
RS ∞ (d) with weight matrices Wi in (92). Since S > 0 (e.g.
[59]), it follows that AT MA > 0 and thus (66) is satisfied
with w = det(AT MA). Therefore, Theorem 9 applies, and
the result of Theorem 1 is immediate.
lated, we write
i
iT
h
h
T
i
i
L L Si − E Si |Ŝ
E Si − E Si |Ŝ
=
i
X
j=1
≥
=
i
X
i
h
T
E Ṽj−1
Ai−j T LT LAi−j Ṽj−1
(164)
DPi−j ri−ℓ (LAi−j Vj−1 )
(165)
′
DPi−j ri−ℓ (LAi−j KVj−1
)
(166)
Proof of Theorem 3. According to Corollary 1 and (64), it
suffices to lower bound RS ∞ (d) with weight matrices Wi in
(92). Consider first the case rank B = n. To use Theorem 11,
let Λi be a diagonal matrix such that 0 ≤ Λi ≤ JT Mi J.
The weight matrices JT AT Mi AJ = A′ T M′i A′ ≥ A′ T Λi A′ ,
and thus the assumption of Theorem 11 is satisfied with
1
Vi = Λi2 A′ , and (25) follows.
If rank B < n, M is singular, (72) is violated, and
Theorem 11 is not applicable. To circumvent the issue,
we augment B and lower bound the cost by that of the
augmented system. Fix some ǫ > 0. Without loss of
generality, assume rank B = m < n, and augment B by
considering a full rank n × n matrix Bǫ :
Bǫ , B ǫB̃ ,
(176)
ℓ=0
j=1
i
X
ℓ=0
j=1
k
≥ wN (V ′ ) m
, wd′′i ,
i
X
2−2
Pi
ℓ=j
rℓ 2(i−j)
a
(167)
j=1
(168)
where to get (166) we used Proposition 4 to minimize each
term of the sum over PŜ i |Vj−1 subject to the constraint
j
I(Vj−1 ; Ŝji ) ≤
i
X
rℓ ,
(169)
ℓ=j
where (n − m) × n matrix B̃ is chosen so that the columns
of Bǫ span Rn .
Accordingly, we augment the m × m matrix R in (3) as
R
0
.
(177)
R̃ ,
0 In−m
where rℓ is the per-stage rate, as defined before in (130);
to get (167) we applied and Proposition 5. To verify that
constraint (130) implies (169), we apply the independence
of Vj−1 and Ŝ j−1 and the chain rule of mutual information
to write
I(Vj−1 ; Ŝji ) = I(Vj−1 ; Ŝji |Ŝ j−1 )
≤ I(S j ; Ŝji |Ŝ j−1 )
=
i
X
ℓ=j
≤
=
i
X
ℓ=j
Consider the augmented system parameterized by ǫ:
(170)
Xi+1 = AXi + Bǫ Ũi + Vi ,
(171)
I(S j ; Ŝℓ |Ŝ ℓ−1 )
(172)
I(S ℓ ; Ŝℓ |Ŝ ℓ−1 )
(173)
where control inputs Ũi are n-dimensional. The augmented
system in (178) achieves the same or smaller quadratic
cost as
in (1), because we can always let
the system
T
Ũi = Ui 0 to equalize the costs. Therefore,
R(b) ≥ sup Rǫ (b),
rℓ .
where Rǫ (b) denotes the rate-cost function for the system in
(178) with parameter ǫ in (176), (177). In particular, since
(28) holds for the augmented system regardless of ǫ > 0,
due to (179), (28) also must hold for the original system.
(174)
ℓ=j
k
i+1
X
2−2
Pi+1
ℓ=j
rℓ 2(i−j+1)
a
j=1
= 2−2ri+1 N (V ′ )
k
m
(179)
ǫ>0
i
X
Finally, observe that d′′i+1 and d′′i are tied in a recursive
relationship akin to that in (139):
d′′i+1 = N (V ′ ) m
(178)
1 +
i
X
2−2
j=1
k
= 2−2ri+1 a2 d′′i + N (V ′ ) m .
Pi
ℓ=j
rℓ 2(i−j+1)
a
(175)
The remainder of the proof follows along the lines of (141)–
(145).
Proof of Theorem 4. According to Proposition 4, the causal
rate-distortion function of the uncontrolled process {Si }
with weight matrices Wi = AT Mi A is equal to that of {ASi }
with weight matrices {Mi }. Putting Si′′ , ASi and noticing
′′
that Si+1
= ASi′′ + AVi , we apply Theorem 12 to conclude
n
µN (AV ) m
m
2
,
(180)
log a +
RS ′′∞ (d) ≥
2
d/m
and Theorem 4 follows via Corollary 1 and (64).
Proof of Theorem 5. We assume rank B = rank C = n.
The more general case is considered in Theorem 7.
According to (93) and (98), {X̂i } is Gauss-Markov
process, whose additive noise Ki Ỹi has covariance matrix
Proof of Theorem 1. If A is rank-deficient, the right side of
(15) is −∞, and there is nothing to prove. Assume det A 6=
0. Further, the case rank B < n implies det M = 0 and is
15
Ni , Ki CPi|i−1 CT + ΣW KTi . Applying Theorem 9 and
(64) yields
!
1
(det NM) n
n
,
RX̂ ∞ kDU ∞ (d) ≥ log | det A| + log 1 +
2
d/n
where VC the volume of the Voronoi cells of lattice C:
VC , Vol ({x ∈ Rn : qC (x) = c}) ,
(185)
ρC ≥ 1,
(186)
where arbitrary c ∈ C, and BC is the volume of a ball whose
radius is equal to that of the Voronoi cells of C. By definition,
and Theorem 5 follows immediately via Corollary 1.
Proof of Theorem 7. The proof is similar to that of Theorem 3 and uses Theorem 11 to lower bound RX̂ ∞ kDU ∞ (d).
Proof of Theorem 8. The proof follows along the lines of
that of Theorem 4, invoking Theorem 12 to lower bound
RX̂ ∞ kDU ∞ (d).
and the closer ρC is to 1 the more sphere-like the Voronoi
cells of C are and the better lattice C is for covering.
The next result gives an upper bound to the output entropy
of lattice quantizers.
Theorem 16 ( [47]). Suppose that fX is (c0 , c1 )-regular.
There exists a lattice quantizer qC such that
sup kx − qC (x)k2 ≤ d
(187)
x∈Rn
VI. ACHIEVABILITY
and
THEOREMS : TOOLS AND PROOFS
N (X)
n
+ αn + n log ρCn (188)
log
˜
2
d/n
d̃≤d
p
1
1
+ 2d˜2 log e(c1 Var [X] + c0 + c1 d˜2 ) ,
H (qC (X)) ≤ min
In this section, we will prove Theorems 2 and 6. Before
describing our achievability schemes, we introduce the tools
that will be key to their analysis. The first result provides
a means to bound the difference between the differential
entropies of two random variables whose distributions are
close to each other.
where ρCn is the lattice covering efficiency defined in (184),
and
n
n
2e
αn , log
+ log Γ
+1 ,
(189)
2
n
2
where Γ(·) is the Gamma function.
Proposition 7 ( [49]). Let X and Y be random vectors with
finite second moments. If the density of X is (c0 , c1 )-regular,
then
The leading term in (188) is Shannon’s lower bound
1
c1
(the
functional inverse of (112)). The contribution of the
2 2
E kXk
h(Y ) − h(X) ≤ log e
2
remaining terms becomes negligible if n is large and d is
1
small. Indeed, by Stirling’s approximation, as n → ∞,
c1
2 2
E kY k
+
+ c0 W (X, Y ), (181)
1
2
(190)
αn = log n + O (1) .
2
where W (X, Y ) is the Wasserstein distance between the
On the other hand, Rogers [42, Theorem 5.9] showed that
distributions of X and Y :
for
each n ≥ 3, there exists an n-dimensional lattice Cn
1
W (X, Y ) , inf E kX − Y k2 2 ,
(182) with covering efficiency
√
where the infimum is over all joint distributions PXY whose
n log ρCn ≤ log2 2πe (log n + log log n + c) , (191)
marginals are PX and PY .
where c is a constant. Therefore, the terms n log ρCn and αn
The next result will help us establish that the random are logarithmic in n, so in high dimension their contribution
variables that we will encounter at each step of the control becomes negligible compared to the first term in (188).
system operation have regular densities.
In low dimension, the contribution of these terms can be
computed
as follows. The thinnest lattice covering is known
Proposition 8. [49, Proposition 3] If the density of Z is
in
dimensions
1 to 23 is Voronoi’s principal lattice of the
(c0 , c1 )-regular and B ⊥⊥ Z, kBk ≤ b a.s., then that of
∗
first
type
[63]
(A
n ), which has covering efficiency
B + Z is (c0 + c1 b, c1 )-regular.
s
1
1
n(n + 2)
π 2 (n + 1) 2n
Our achievability scheme employs lattice quantization. A
.
(192)
ρA∗n =
1
lattice C in Rn is a discrete set of points that is closed under
Γ n2 + 1 n 12(n + 1)
reflection and addition. The nearest-neighbor quantizer is the A∗ is proven to be the thinnest lattice covering possible in
n
mapping qC : Rn 7→ C defined by
dimensions n = 1, 2, . . . , 5. For A∗ -based lattice quantizer,
qC (x) , argmin kx − ck
n
(183)
we can compute the constant appearing in (188) as
αn
1
2πe(n + 2)
log ρA∗n +
= log
1 .
n
2
12(n + 1)1− n
c∈C
Covering efficiency of lattice C is measured by
1
BC n
ρC ,
,
VC
(193)
Proof of Theorem 10. The proof analyses a DPCM scheme.
First, we describe how the codebook is generated, then we
(184)
16
1
bound on the entropy of qCn (Wi2 S̃i ):
(
1
1
1
n
d˜2
N (Wi2 S̃i )
2
+2
log e
log
H qCn (Wi S̃i ) ≤ min
˜
2
wi
d̃≤d
d/n
)
1
1
1
· c1 (ai d + vi ) 2 + c0 + c1 1 + a 2 d˜2
describe the operation of the encoder and the decoder, and
then we proceed to the analysis of the scheme.
Codebook design. To maximize covering efficiency, we
use the best lattice known in dimension n scaled so that its
covering radius is ≤ d:
sup kz − qCn (z)k2 ≤ d,
i
(194)
+ αn + n log ρCn .
z∈Rn
To estimate the entropy power of Wi S̃i , we use (199) and
(201) to bound the Wasserstein distance in Proposition 7 to
obtain
Encoder. Upon observing St , the encoder computes the
state innovation S̃i recursively using the formula
S̃i , Si − AŜi−1 ,
(195)
1
(ai di ) 2
h(Wi S̃i ) ≤ h(Wi Vi−1 ) + log e
(205)
wi
c 1
1
c1
1 2
vi +
(ai d + vi ) 2 + c0 .
·
2
2
Combining (204) and (205), we conclude that
1
(206)
H qCn (Wi2 S̃i ) ≤
)
(
1
N (Wi2 Vi−1 )
n
˜ + αn + n log ρC ,
+ βi (d)
log
min
n
˜
2
d/n
d̃≤d
1
where βi (d) = O d 2 is given by
1
2
where Ŝi is the decoder’s state estimate at time i (put Ŝ0 ,
1
0). The encoder transmits the index of qCn (Wi2 S̃i ).
Decoder. The decoder recovers the lattice cell identified
by the encoder, forms its state estimate as
Ŝi = AŜi−1 + S̃ˆi ,
(196)
1
1
−
(197)
S̃ˆ , W 2 q (W 2 S̃ ).
i
Cn
i
i
i
Analysis. The distortion at step i is given by
T
T
Si − Ŝi Wi Si − Ŝi = S̃i − S̃ˆi Wi S̃i − S̃ˆi
=
1
2
1
2
Wi S̃i − qCn (Wi S̃i )
≤ d,
βi (d) ,
2
(198)
where (198) is due to (194).
z6=0
1
1
1
1
d2
1
(207)
log e
c1 ai2 vi2 + c0 2 + ai2
wi
2
!
1 !
1
1
1
ai2
2
2
2
(ai d + vi ) + 2c1 1 + ai d .
+ c1 2 +
2
Proof of Theorem 2. Due to Corollary 1 and (65), it suffices
to bound the entropy-distortion function of the process (62),
which given in Theorem 2.
Proof of Theorem 6. Due to Corollary 1, it suffices to
bound the conditional entropy-distortion function of the
Kalman filter estimates process in (93), which follows from
(65) and Theorem 2.
where ai is the following operator norm of A:
z T AT Wi Az
.
z T Wi z
1
2
Applying (114) and taking i → ∞, we obtain the statement
of Theorem 10.
It remains to compute the entropy of the quantized
1
1
representation of Wi2 S̃i . First, we establish that Wi2 S̃i
has a regular density. Using the assumption that V has
1
a (c0 , c1 )-regular
density, it’s easy to see that Wi2 V has
wi−1 c0 , wi−1 c1 -regular density, where wi is the minimum
eigenvalue of Wi . Furthermore, similar to (198),
T
(199)
Si − Ŝi AT Wi A Si − Ŝi ≤ ai d,
ai , sup
(204)
1
2
(200)
VII. C ONCLUSION
From (195) and (62),
We studied the fundamental tradeoff between the communication requirements and the attainable quadratic cost in
1
fully and partially observed linear stochastic control systems.
and it follows via Proposition 8 that Wi2 S̃i has (wi−1 (c0 +
1
We
introduced the rate-cost function in Definition 1, and
1
−1
ai2 d 2 c1 ), wi c1 )-regular density.
showed sharp lower bounds to it in Theorems 1, 3, 4
(fully observed system) and Theorems 5, 7, 8 (partially
Observe that combining (199) and (201) yields
observed system). The achievability results in Theorem 2
h 1 i
(202) (fully observed system) and Theorem 6 (partially observed
Var Wi2 S̃i ≤ ai d + vi ,
system) show that the converse can be approached, in the
where we denoted for brevity
high rate / low cost regime, by a simple variable-rate latticevi , tr (ΣV Wi ) .
(203) based scheme in which only the quantized value of the innovation is transmitted. Via the separation principle, the same
Now we can apply Theorem 16 to obtain the following conclusions hold for causal compression of Markov sources:
S̃i = A(Si−1 − Ŝi−1 ) + Vi−1 ,
(201)
17
a converse, which may be viewed as a causal counterpart
of Shannon’s lower bound, is stated in Theorem 9, and a
matching achievability in Theorem 10.
Extending the analysis of the partially observed case to
non-Gaussian noises would be of interest. It also remains
an open question whether the converse bound in Theorem 1
can be approached by fixed-rate quantization, or over noisy
channels. Finally, it would be interesting to see whether
using non-lattice quantizers can help to narrow down the
gap in Fig. 2.
VIII. ACKNOWLEDGEMENT
The authors acknowledge many stimulating discussions
with Dr. Anatoly Khina and his helpful comments on the
earlier versions of the manuscript. The authors are also
grateful to Ayush Pandey, who generated the plot in Fig. 2.
R EFERENCES
[1] V. Kostina and B. Hassibi, “Rate-cost tradeoffs in control,” in Proceedings 54th Annual Allerton Conference on Communication, Control
and Computing, Monticello, IL, Oct. 2016.
[2] S. Tatikonda, A. Sahai, and S. Mitter, “Stochastic linear control over
a communication channel,” IEEE Transactions on Automatic Control,
vol. 49, no. 9, pp. 1549–1561, 2004.
[3] J. Massey, “Causality, feedback and directed information,” in Proc.
Int. Symp. Inf. Theory Applic.(ISITA-90), 1990, pp. 303–305.
[4] G. Kramer, “Directed information for channels with feedback,” Ph.D.
dissertation, ETH Zurich, 1998.
[5] M. Gastpar, B. Rimoldi, and M. Vetterli, “To code, or not to code:
lossy source-channel communication revisited,” IEEE Transactions
on Information Theory, vol. 49, no. 5, pp. 1147–1158, May 2003.
[6] J. Baillieul, “Feedback designs for controlling device arrays with
communication channel bandwidth constraints,” in ARO Workshop
on Smart Structures, Pennsylvania State Univ, 1999, pp. 16–18.
[7] W. S. Wong and R. W. Brockett, “Systems with finite communication bandwidth constraints. II. Stabilization with limited information
feedback,” IEEE Transactions on Automatic Control, vol. 44, no. 5,
pp. 1049–1053, 1999.
[8] S. Tatikonda and S. Mitter, “Control under communication constraints,” IEEE Transactions on Automatic Control, vol. 49, no. 7,
pp. 1056–1068, 2004.
[9] S. Yüksel and T. Başar, “Minimum rate coding for LTI systems over
noiseless channels,” IEEE Transactions on Automatic Control, vol. 51,
no. 12, pp. 1878–1887, 2006.
[10] B. G. N. Nair, F. Fagnani, S. Zampieri, and R. J. Evans, “Feedback
control under data rate constraints: An overview,” Proceedings of the
IEEE, vol. 95, no. 1, pp. 108–137, 2007.
[11] G. N. Nair and R. J. Evans, “Stabilizability of stochastic linear
systems with finite feedback data rates,” SIAM Journal on Control
and Optimization, vol. 43, no. 2, pp. 413–436, 2004.
[12] R. W. Brockett and D. Liberzon, “Quantized feedback stabilization
of linear systems,” IEEE transactions on Automatic Control, vol. 45,
no. 7, pp. 1279–1289, 2000.
[13] S. Yüksel, “Stochastic stabilization of noisy linear systems with fixedrate limited feedback,” IEEE Transactions on Automatic Control,
vol. 55, no. 12, pp. 2847–2853, 2010.
[14] ——, “Jointly optimal LQG quantization and control policies for
multi-dimensional systems,” IEEE Transactions on Automatic Control, vol. 59, no. 6, pp. 1612–1617, 2014.
[15] H. S. Witsenhausen, “On the structure of real-time source coders,”
The Bell System Technical Journal, vol. 58, no. 6, pp. 1437–1451,
1979.
[16] N. Gaarder and D. Slepian, “On optimal finite-state digital transmission systems,” IEEE Transactions on Information Theory, vol. 28,
no. 2, pp. 167–186, 1982.
[17] G. Gabor and Z. Gyorfi, Recursive Source Coding: A theory for the
practice of waveform coding. Springer-Verlag New York, Inc., 1986.
18
[18] J. C. Walrand and P. Varaiya, “Optimal causal coding-decoding
problems,” IEEE Transactions on Information Theory, vol. 29, no. 6,
pp. 814–820, 1983.
[19] V. S. Borkar, S. K. Mitter, and S. Tatikonda, “Optimal sequential
vector quantization of Markov sources,” SIAM journal on control and
optimization, vol. 40, no. 1, pp. 135–148, 2001.
[20] D. Teneketzis, “On the structure of optimal real-time encoders and
decoders in noisy communication,” IEEE Transactions on Information
Theory, vol. 52, no. 9, pp. 4017–4035, 2006.
[21] T. Linder and S. Yüksel, “On optimal zero-delay coding of vector
Markov sources,” IEEE Transactions on Information Theory, vol. 60,
no. 10, pp. 5975–5991, 2014.
[22] R. G. Wood, T. Linder, and S. Yüksel, “Optimality of WalrandVaraiya type policies and approximation results for zero delay coding
of Markov sources,” in Proceedings 2015 IEEE International Symposium on Information Theory, Hong Kong, June 2015.
[23] S. Yüksel, T. Basar, and S. P. Meyn, “Optimal causal quantization of
Markov sources with distortion constraints,” in Information Theory
and Applications Workshop, 2008, 2008, pp. 26–30.
[24] M. Rabi, C. Ramesh, and K. H. Johansson, “Separated design of
encoder and controller for networked linear quadratic optimal control,”
SIAM Journal on Control and Optimization, vol. 54, no. 2, pp. 662–
689, 2016.
[25] N. Elia and S. K. Mitter, “Stabilization of linear systems with limited
information,” IEEE transactions on Automatic Control, vol. 46, no. 9,
pp. 1384–1400, 2001.
[26] A. Gorbunov and M. S. Pinsker, “Prognostic epsilon entropy of
a gaussian message and a gaussian source,” Problemy Peredachi
Informatsii, vol. 10, no. 2, pp. 5–25, 1974.
[27] C. D. Charalambous, P. A. Stavrou, and N. U. Ahmed, “Nonanticipative rate distortion function and relations to filtering theory,” IEEE
Transactions on Automatic Control, vol. 59, no. 4, pp. 937–952, 2014.
[28] C. D. Charalambous and A. Farhadi, “LQG optimality and separation principle for general discrete time partially observed stochastic
systems over finite capacity communication channels,” Automatica,
vol. 44, no. 12, pp. 3181–3188, 2008.
[29] C. D. Charalambous, C. K. Kourtellaris, and C. Hadjicostis, “Optimal
encoder and control strategies in stochastic control subject to rate
constraints for channels with memory and feedback,” in 2011 50th
IEEE Conference on Decision and Control and European Control
Conference, Dec 2011, pp. 4522–4527.
[30] E. Shafieepoorfard, M. Raginsky, and S. P. Meyn, “Rationally inattentive control of Markov processes,” SIAM Journal on Control and
Optimization, vol. 54, no. 2, pp. 987–1016, 2016.
[31] E. I. Silva, M. S. Derpich, and J. Ostergaard, “A framework for
control system design subject to average data-rate constraints,” IEEE
Transactions on Automatic Control, vol. 56, no. 8, pp. 1886–1899,
2011.
[32] E. Silva, M. Derpich, J. Ostergaard, and M. Encina, “A characterization of the minimal average data rate that guarantees a given closedloop performance level,” IEEE Transactions on Automatic Control,
2016.
[33] T. Tanaka, K. H. Johansson, T. Oechtering, H. Sandberg, and
M. Skoglund, “Rate of prefix-free codes in LQG control systems,”
in Proceedings 2016 IEEE International Symposium on Information
Theory, Barcelona, Spain, July 2016, pp. 2399–2403.
[34] M. S. Derpich and J. Ostergaard, “Improved upper bounds to
the causal quadratic rate-distortion function for gaussian stationary
sources,” IEEE Transactions on Information Theory, vol. 58, no. 5,
pp. 3131–3152, May 2012.
[35] T. Tanaka, K.-K. K. Kim, P. A. Parrilo, and S. K. Mitter, “Semidefinite
programming approach to Gaussian sequential rate-distortion tradeoffs,” IEEE Transactions on Automatic Control, vol. 62, no. 4, pp.
1896–1910, 2017.
[36] T. Tanaka, P. M. Esfahani, and S. K. Mitter, “LQG control with
minimum directed information: Semidefinite programming approach,”
IEEE Transactions on Automatic Control, 2017.
[37] C. E. Shannon, “Coding theorems for a discrete source with a fidelity
criterion,” IRE Int. Conv. Rec., vol. 7, no. 1, pp. 142–163, Mar. 1959,
reprinted with changes in Information and Decision Processes, R. E.
Machol, Ed. New York: McGraw-Hill, 1960, pp. 93-126.
[38] ——, “A mathematical theory of communication,” Bell Syst. Tech. J.,
vol. 27, pp. 379–423, 623–656, July and October 1948.
[39] A. J. Stam, “Some inequalities satisfied by the quantities of information of Fisher and Shannon,” Information and Control, vol. 2, no. 2,
pp. 101–112, 1959.
[40] H. Gish and J. Pierce, “Asymptotically efficient quantizing,” IEEE
Transactions on Information Theory, vol. 14, no. 5, pp. 676–683,
1968.
[41] J. Ziv, “On universal quantization,” IEEE Transactions on Information
Theory, vol. 31, no. 3, pp. 344–347, 1985.
[42] C. A. Rogers, Packing and covering. Cambridge University Press,
1964, no. 54.
[43] R. Zamir and M. Feder, “On lattice quantization noise,” IEEE
Transactions on Information Theory, vol. 42, no. 4, pp. 1152–1159,
Jul. 1996.
[44] A. Gersho, “Asymptotically optimal block quantization,” IEEE Transactions on Information Theory, vol. 25, no. 4, pp. 373–380, Jul. 1979.
[45] R. Zamir and M. Feder, “On universal quantization by randomized
uniform/lattice quantizers,” IEEE Transactions on Information Theory, vol. 38, no. 2, pp. 428–436, Mar. 1992.
[46] T. T. Linder and K. K. Zeger, “Asymptotic entropy-constrained performance of tessellating and universal randomized lattice quantization,”
IEEE Transactions on Information Theory, vol. 40, no. 2, pp. 575–
579, Mar. 1994.
[47] V. Kostina, “Data compression with low distortion and finite blocklength,” IEEE Transactions on Information Theory, vol. 63, no. 7, pp.
4268–4285, July 2017.
[48] M. Fu, “Lack of separation principle for quantized linear quadratic
Gaussian control,” IEEE Transactions on Automatic Control, vol. 57,
no. 9, pp. 2385–2390, 2012.
[49] Y. Polyanskiy and Y. Wu, “Wasserstein continuity of entropy and
outer bounds for interference channels,” IEEE Transactions on Information Theory, vol. 62, no. 7, pp. 3992–4002, July 2016.
[50] Y.-H. Kim, “A coding theorem for a class of stationary channels with
feedback,” IEEE Transactions on Information Theory, vol. 54, no. 4,
pp. 1488–1499, 2008.
[51] S. Tatikonda and S. Mitter, “The capacity of channels with feedback,”
IEEE Transactions on Information Theory, vol. 55, no. 1, pp. 323–
349, Jan 2009.
[52] J. S. Freudenberg, R. H. Middleton, and V. Solo, “Stabilization and
disturbance attenuation over a Gaussian communication channel,”
IEEE Transactions on Automatic Control, vol. 55, no. 3, pp. 795–
799, 2010.
[53] A. Khina, G. M. Pettersson, V. Kostina, and B. Hassibi, “Multi-rate
control over AWGN channels: An analog joint source–channel coding
perspective,” in 55th IEEE Conference on Decision and Control. Las
Vegas, NV, Dec. 2016.
[54] A. Wyner, “An upper bound on the entropy series,” Information and
Control, vol. 20, no. 2, pp. 176–181, Mar. 1972.
[55] N. Alon and A. Orlitsky, “A lower bound on the expected length of
one-to-one codes,” IEEE Transactions on Information Theory, vol. 40,
no. 5, pp. 1670–1672, Sep. 1994.
[56] W. Szpankowski and S. Verdú, “Minimum expected length of fixed-tovariable lossless compression without prefix constraints: memoryless
sources,” IEEE Transactions on Information Theory, vol. 57, no. 7,
pp. 4017–4025, July 2011.
[57] T. Tanaka, “Semidefinite representation of sequential rate-distortion
function for stationary gauss-markov processes,” in Proceedings 2015
IEEE Conference on Control Applications (CCA), Sep. 2015, pp.
1217–1222.
[58] T. Fischer, “Optimal quantized control,” IEEE Transactions on Automatic Control, vol. 27, no. 4, pp. 996–998, Aug 1982.
[59] D. P. Bertsekas, Dynamic programming and optimal control. Athena
Scientific Belmont, MA, 1995, vol. 1.
[60] A. Gerrish and P. Schultheiss, “Information rates of non-Gaussian
processes,” IEEE Transactions on Information Theory, vol. 10, no. 4,
pp. 265–271, Oct. 1964.
[61] Y. N. Linkov, “Evaluation of ǫ-entropy of random variables for small
ǫ,” Problems of Information Transmission, vol. 1, no. 2, pp. 18–26,
1965.
[62] T. M. Cover and J. A. Thomas, Elements of information theory,
2nd ed. John Wiley & Sons, 2012.
[63] J. H. Conway and N. J. A. Sloane, Sphere packings, lattices and
groups. Springer Science & Business Media, New York, 2013, vol.
290.
19
| 7cs.IT
|
arXiv:1604.01119v2 [math.GR] 26 Aug 2016
Group Structures on Families of Subsets of a
Group
Mario Gómez, Sergio R. López-Permouth , Fernando Mazariegos,
Alejandro José Vargas De León and Rigoberto Zelada Cifuentes
August 29, 2016
Abstract
A binary operation on any set induces a binary operation on its subsets. We explore families of subsets of a group that become a group under
the induced operation and refer to such families as power groups of the
given group. Our results serve to characterize groups in terms of their
power groups. In particular, we consider when the only power groups of a
group are the factor groups of its subgroups and when that is the case up
to isomorphism. We prove that the former are precisely those groups for
which every element has finite order and provide examples to illustrate
that the latter is not always the case. In the process we consider several natural questions such as whether the identity element of the group
must belong to the identity element of a power group or the inverse of an
element in a power group must consist of the inverses of its elements.
1
Introduction
Traditionally, group theory has been built upon understanding the subgroups
and the factor groups of a given group. In other words, our understanding
of groups has largely rested on understanding their subquotients (factors of
subgroups or subgroups of factors; it’s the same.) We propose the power groups
of a group as a larger collection of groups that might give us a better insight into
the structure of the original group. The Spanish philosopher Ortega y Gasset
said I am I plus my circumstance [2]; inspired by that thought, we have been
referring to the family of power groups of a group G as the circumstance of G.
This paper is a first step as we propose to study the way the structure of a
group is determined by its circumstance.
Reading this paper requires very little background; an undergraduate Modern Algebra course using a textbook like [1] will suffice.
A binary operation on any set S induces a binary operation on its subsets.
Let us refer to the operation on the original set as product and the one on its
subsets as subset product Then, for subsets A, B ⊆ S, the subset product AB
equals {ab | a ∈ A, b ∈ B}. Clearly not all families of subsets of a group G will
1
become groups under the induced subset product. In fact, ∅ can only belong
to one such family, namely {∅}. For convenience, we consider therefore only
families of non-empty subsets in what follows. If G ⊆ P(G) − {∅} is a group
with the above defined product, then G is a power group of the group G.
2
Groups whose only power groups are factor
groups
The construction of factor groups, a standard topic in a first course in group
theory for undergraduates, views the elements of such groups in one of two
ways which are equivalent when the identity N of the obtained structure is a
normal subgroup. For a group G and a normal subgroup N of G, one can
G
think of the operation on N
as being the multiplication of the subsets aN
and bN with the binary operation induced on subsets or think of the identity
(aN )(bN ) = (ab)N as being a definition (which can be shown to be well-defined
(independent from the choice of representatives) under the given hypotheses.
We will see that the two approaches are not necessarily equivalent when the
hypothesis for the identity for the power group is eased. Our considerations will
give rise to two separate notions which we will name subquotients and groups of
cosets. Subquotients are the subject of this section while groups of cosets will
make their appearance in the next one.
G
Since the factor group N
of a group G modulo a normal subgroup N may
be viewed as the group structure induced by the subset product on the family
G
is a natural example of a power
of left (equivalently right) cosets of N in G, N
group of the group G. More generally, give any subgroup H < G and a normal
subgroup N of H, the factor group H
N is also a power group of G. To facilitate
the conversation, we give these examples a name.
Definition. A power group G of G is a subquotient group if there exists a
subgroup H of G, and a normal subgroup N of H, such that G consists precisely
of the cosets of N in H (i.e. G = H
N .)
The next theorem shows that subquotients are characterized by many natural properties.
Theorem 1. The following are equivalent for a power group G of G:
(a) G is a subquotient group of G.
(b) E, the identity element of G, is a subgroup of G.
(c) For every A ∈ G, if x ∈ A then x−1 ∈ A−1
S
(d) H = A∈G A is a subgroup of G, and G is a partition of H.
Proof. Clearly (a) → (b), because then G =
of G is N .
2
H
N
for some H, N and the identity
To prove (b) → (c), take x ∈ A and, by virtue of the fact that ∅ ∈
/ G,
choose y ∈ A−1 . It follows that xy ∈ AA−1 = E; since E is a group, there is
z ∈ E such that (xy)z = e. Finally, since yz ∈ A−1 E = A−1 , then we have
x−1 = yz ∈ A−1 , proving (c). S
To prove (c) → (d), let H = A∈G A; it is easy to see that H is always closed
under products and, under the assumption of (c), also under inverses. Therefore,
H is a subgroup as claimed. To see that G is a partition, supposeA, B ∈ G share
an element x ∈ A ∩ B. Then, under our assumption, A−1 and B −1 also share
an element as x−1 ∈ A−1 ∩ B −1 . But then e ∈ A−1 B and e ∈ B −1 A. This
means that A ⊆ AA−1 B = B and B ⊆ BB −1 A = A, so A = B and therefore G
is a partition.
To prove (d) → (a), we aim to show that G = H
E where E is the identity
element of G. Some notation will be convenient: for x ∈ H, let us use Ax to
denote the (unique) element of G containing x. Our first aim is to show that
E = Ae . As e ∈ A2e , considering that G is a partition, A2e = Ae . Cancellation
in the group G yields that Ae = E, as claimed; in particular e ∈ E. Next, we
show that the elements of G are the cosets of E, which then concludes the proof.
Taking Ax ∈ G , the fact that H is a subgroup of G yields that Ax−1 ∈ G.Under
our assumption, noting that e ∈ Ax Ax−1 implies A−1
x = Ax−1 . For an arbitrary
x ∈ H, consider the coset Ex ⊆ EAx = Ax . We want to get the reverse
inclusion to prove that Ax = Ex; for that purpose, pick y ∈ Ax . Now, Ax = Ay
and Ax−1 = Ay−1 . Considering Ay Ax−1 = Ay Ay−1 = E one concludes that
yx−1 ∈ E and, from this, y ∈ Ex.
Armed with an understanding of subquotients, we aim now to show that
not all power groups are in general subquotients; in fact, the following theorem
provides a characterization of those groups for which the only power groups are
the subquotients.
Theorem 2. Let G be a group, then the only power groups of G are the subquotients of G if and only if for all a ∈ G ord(a) < ∞.
Proof. (←) By virtue of Theorem 1, it suffices to show that E, the identity
element of G, is a subgroup. That is indeed the case since EE = E yields
closure and, since by assumption every element of G is of finite order,E is also
closed under inverses and is consequently a subgroup of G.
(→) Suppose there exists an a ∈ G such that ord(a) = ∞ and consider the
set E = {an |n ∈ N} ⊆ P(G), clearly EE = E and therefore G = {E} is a power
group of G. Notice that E is the identity element of G and is not a group, thus
G is not a subquotient of G.
3
What lies ahead
Central to the understanding of power groups is the analysis of their identity
elements. While in the case of subquotients, the identity element must be a
3
subgroup, in general all one needs is a subset E ⊂ G such that E 2 = E.
Assume, for example that one has such a subset E of a group G and say that
E, in addition, satisfies a normalcy property that for all a ∈ G, aE = Ea.
Examples of appropriate choices of E include, for the additive group of integers
G = Z, the set E of non-negative integers and, for the additive group of rational
numbers G = Q, the set E = Q+ of positive rationals.
Following the steps of the usual proof to show that, for a normal subgroup
G
N of G, N
is a group, one obtains the same result here with weaker hypotheses.
In other words, it is straightforward to show that the group in the following
definition is indeed a group:
Definition. For any group G, given E ⊂ G such that E 2 = E and a subgroup
H < G; assume, further, that for every a ∈ H, aH = Ha, then the family
{aE|a ∈ H} is a power group which will be called the group of cosets induced by
E and H and will be denoted H
E . It satisfies that, for any a, b ∈ H, (aE)(bE) =
(ab)E. We will refer to any such power group as a group of cosets of the group
G. Notice that the definition does not require that E be a subset of H.
Remark. Every group of cosets is isomorphic to a subquotient.
Proof. Define a map ϕ : H → H
E via ϕ(a) = aE and verify it is a group
epimorphism. Notice that subquotients are always groups of cosets but, in
general, groups of cosets need not even be partitions.
An interesting feature of this family of power groups is that not only is the
identity element not necessarily a subgroup but actually it does not even need
contain the identity element of the group (as seen in the example E = Q+ mentioned above.) Motivated by the realization that the traditional construction
yields more power groups, it is natural to consider a modification of the original
question from this paper and ask for which groups it is the case that all power
groups are groups of cosets. In fact, in light of Remark 3, another natural question is to characterize those groups G all power groups of G are isomorphic to
subquotients of G. As we finish this project, we do not yet know the answers
to either one of these questions; however, we can show that Z is an example
of a group whose only power groups are its groups of cosets and that Q is an
example of the opposite situation; in fact Q is a group having power groups that
are not isomorphic to subquotients.
Theorem 3. Every power group of Z is a group of cosets.
Proof. One must determine first that the only possible non-empty choices for
an identity element E of a powergroup G of Z. One easily gets that, outside
subgroups of Z, E must be a subset of Z which contains 0 and is such that
either all its other elements are positive or all of them are negative. Focus,
without losing generality, on the case when the elements of E are non-negative;
the other case is handled similarly. One can easily show that every element AG
has a smallest element, say a ∈ A and A = aE. Incidentally, a may also be
4
recognized by the property that it is the one element of A such that −a ∈ −A.
Denote H the set of smallest elements in elements of G; it follows that G = H
E.
Theorem 4. The additive group of rational numbers Q is an example of a group
having a power group which is not a group of cosets.
Proof. Let G = {(r, ∞) ∩ Q|r ∈ R}. Then G ∼
= R, in particular |G| > |Q| hence
G is not isomorphic to a subquotient.
In general, it seems interesting to be able to determine, given two groups
G1 and G2 , when G1 has a power group isomorphic to G2 . For convenience, let
us use the expression G2 underlies G1 when G1 has a power group isomorphic
to G2 . The last example shows an instance of two non-isomorphic groups that
underlie one another. What properties are shared by groups that underlie one
another? Another interesting question that we suggest here is whether underlying is a transitive relation, namely whether, for three groups Gi (i = 1, 2, 3), if
G3 underlies G2 and G2 underlies G1 , can one conclude that G3 underlies G1 ?
Notice that if all elements of G1 have finite order then the same would be true
of G2 and G3 and (since a subquotient of a subquotient is a subquotient) concluding that the result holds would be a straightforward application of Theorem
1; settling the general case seems to be more difficult.
References
[1] J. Gallian, Contemporary Abstract Algebra, 8th Ed. Brooks/Cole (2012).
[2] José Ortega y Gasset. Obras Completas, Vol. I. Ed. Taurus/Fundación
José Ortega y Gasset, Madrid, 2004, p. 757.
5
| 4math.GR
|
Independence Tuples and Deninger’s Problem
Ben Hayes
arXiv:1502.03858v2 [math.DS] 27 Apr 2016
Vanderbilt University
Stevenson Center
Nashville, TN 37240
[email protected]
April 9, 2018
Abstract
We define modified versions of the independence tuples for sofic entropy developed in [22]. Our
first modification uses an ℓq -distance instead of an ℓ∞ -distance. It turns out this produces the same
version of independence tuples (but for nontrivial reasons), and this allows one added flexibility. Our
second modification considers the “action” a sofic approximation gives on {1, . . . , di }, and forces our
independence sets Ji ⊆ {1, . . . , di } to be such that χJi − udi (Ji ) (i.e. the projection of χJi onto mean
zero functions) spans a representation of Γ weakly contained in the left regular representation. This
modification is motivated by the results in [17]. Using both of these modified versions of independence
tuples we prove that if Γ is sofic, and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)) is not invertible in Mn (Z(Γ)),
then detL(Γ) (f ) > 1. This extends a consequence of the work in [15] and [22] where one needed f ∈
Mn (Z(Γ))∩GLn (ℓ1 (Γ)). As a consequence of our work, we show that if f ∈ Mn (Z(Γ))∩GLn (L(Γ)) is
not invertible in Mn (Z(Γ)) then Γ y (Z(Γ)⊕n /Z(Γ)⊕n f )b has completely positive topological entropy
with respect to any sofic approximation.
Keywords: sofic groups, independence tuples, completely positive entropy, Fuglede-Kadsion determinants.
MSC: 37B40, 47A67, 22D25
Contents
1 Introduction
1
2 ℓp -Versions of Independence Tuples
4
3 Independence Tuples with a Weak Containment Condition
9
4 A Generalization of Deninger’s Problem For Sofic Groups
16
1
1
Introduction
This paper is concerned with a modification of independence tuples in the case of sofic topological entropy
due to Kerr-Li in [22]. We remark that the definition of sofic topological entropy is due to Kerr-Li in
[21], following on the seminal work of Bowen on sofic measure-theoretic entropy in [2]. Independence
tuples were first developed in [20] for actions of amenable groups. Positivity of topological entropy is
equivalent to having a nondiagonal independence pair, and this can be viewed as a topological version of
the fact that a measure-preserving action of an amenable group must have a weakly mixing factor. Using
independence tuples, Kerr-Li showed that if a topological action has positive entropy, then the action
must exhibit some chaotic behavior (see e.g. [22] Theorem 8.1). Let us briefly mention the combinatorial
version of independence. We say that a tuple (Ai,1 , . . . , Ai,k )i∈J of subsets of a set A are independent,
if for every c : J → {1, . . . , k} we have
\
Ai,c(i) 6= ∅.
i∈J
The name coming from the case when (Ai,1 , . . . , Ai,k )i∈J are (probabilistically) independent partitions in
a probability space. If Γ is a countable discrete group acting on a set X, and (A1 , . . . , Ak ) are subsets of
X, we call a finite J ⊆ Γ an independence set for (A1 , . . . , Ak ) if {(s−1 A1 , . . . , s−1 Ak )}s∈J is independent.
When Γ is amenable and F ⊆ Γ is finite, we let φA (F ) is the maximal cardinality of a subset of F which
is an independent set for A. We can then define the independence density of A = (A1 , . . . , Ak ), denoted
n)
where Fn is a Følner sequence. In the case X is compact and the action
I(A), to be the limit of φA|F(F
n|
is by homeomorphisms, we say that a tuple x = (x1 , . . . , xk ) is an independence tuple if every tuple
U = (U1 , . . . , Uk ) where Uj is a neighborhood of xj we have I(U ) > 0. This definition is due to Kerr-Li
in [20].
To generalize to the case of sofic groups (defined in the next section), Kerr-Li considered a sofic
approximation (again we define this in the next section)
σi : Γ → Sdi
and abstracted the internal independent subsets of Γ considered in the amenable case to external independent subsets of {1, . . . , di } in the sofic case. In this manner they defined what it means for a tuple
(x1 , · · · , xk ) ∈ X k to be an independence tuple for the action Γ y X with respect to some fixed sofic
approximation σi : Γ → Sdi . Moreover, they showed that Γ y X has positive entropy with respect to
σi : Γ → Sdi if and only if there is a nondiagonal independence pair in X 2 . This can be viewed as a
topological version of the fact that a probability measure-preserving action with positive entropy must
have a weakly mixing factor.
We give two alternate versions of an independence tuple for actions of sofic groups. For the first it is
useful to rephrase the definition in terms of metrics. Let ρ be a compatible metric on X. The condition
\
g −1 Uc(g) 6= ∅
g∈J
can be replaced by the similar condition that there is a x ∈ X so that
max ρ(gx, gxc(s) ) < ε.
g∈J
Equivalently, for 1 ≤ p ≤ ∞, let us define ρp on X J by
ρp,J (x, y)p =
1 X
ρ(x(g), y(g))p if p < ∞
|J|
g∈J
ρ∞,J (x, y) = sup ρ(x(j), y(j)).
j∈J
Then we are considering the condition that
ρ∞,J (O(x), xc(·) ) < ε
where O : X → X J is defined by O(x)(g) = gx. One can rephrase the sofic version of independence sets
in terms of a similar ℓ∞ -product metric. We define an a priori different version of independent set using
an ℓp -product metric. This is a priori weaker than the ℓ∞ -product version, however by an application
1
of the Sauer-Shelah theorem we can show that they are equivalent. While it appears that we have thus
accomplished nothing, this actually gives us an added degree of flexibility as the ℓ2 -product metric will
be more useful to us. The technique of using ℓp -metrics instead of ℓ∞ -metrics was first used by Li in
[23]. We believe this is a very important technique, which often gives one added flexibility needed to
prove results in entropy theory. We mention that we have already exploited this in [16],[15],[17]. It is
quite useful when one wishes to apply Hilbert space techniques as these are phrased better in terms of
the ℓ2 -product metric. This is precisely the purpose of their use in [16],[15],[17] and we believe this is
crucial for those results, as well as the results in this paper.
The second version of independence tuples is one in which we control the translates of an independence
set J by the left regular representation (in a sense to be made more precise in Section 3), and moreover
only consider partitions
c : J → {1, . . . , k}
where each of the pieces c−1 ({l}), 1 ≤ l ≤ k also has its translates controlled by the left regular representation (again this will be made more precise later). To briefly describe the idea, consider a measurepreserving action Γ y (X, µ). Given a set A ⊆ X we can consider the subspace of L2 (X, µ) given
by
HA = Span{g(χA − µ(A)1) : g ∈ Γ}
k·k2
.
One can then ask for sets where
Γ y HA
is related to representations one is more familiar with, and this provides interesting restrictions of the
translates of A by Γ. For example, one could consider A where Γ y HA extends to the reduced group
C ∗ -algebra (this is the completion of the group algebra in the left regular representation). Equivalently,
for all f ∈ cc (Γ) we have
X
g∈Γ
≤
f (g)(χgA − µ(A)1)
2
X
f (g)λ(g) kχA − µ(A)1k2 ,
g∈Γ
where λ is the left regular representation. This says nothing in the amenable case, but in the nonamenable case implies some mixing behavior of A. For example, if every measurable A ⊆ X has this
property and Γ is non-amenable then the action is strongly ergodic. Based on this idea, we give a version
of independence tuples, called independence tuples satisfying the weak containment condition, where the
“representation” (via the sofic approximation) on the independence sets in question is weakly contained
in the left regular representation. Since the sofic approximation is not actually a representation, we
mention for clarity that we will require our independence sets to be sequences (Ji )i≥1 of subsets of
{1, . . . , di } so that for all f ∈ C(Γ), η > 0 we have
kσi (f )(χJi − udi (J1 )1)k2 ≤ kλ(f )kkχJi − udi (Ji )1k2 + η
for all large i. Moreover, we require that the partitions
c : Ji → {1, . . . , di }
are such that the pieces c−1 ({l}) also exhibit similarly controlled behavior by the left regular representation (albeit in a more finitary sense). Theorem 1.1 and Corollary 1.4 in [17] indicate that the left regular
representation plays a crucial role in entropy theory, and from this our strengthening of independence
tuples is natural.
A priori, this different version of an independence tuple bears no relation to independence tuples
developed by Kerr-Li, as we are requiring a stronger condition on the structure of the independent set
but also considering less general partitions. However, using a probabilistic argument and the SauerShelah Lemma we show that every independence tuple satisfying the weak containment condition is an
independence tuple. It turns out (not surprisingly) that in the amenable case, independence tuples are
independence tuples satisfying the weak containment condition. It is possible that independence tuples
are independence tuples satisfying the weak containment condition for sofic groups, but it is not clear how
one would prove this. However, we strongly believe that positivity of topological entropy is equivalent to
the existence of a nondiagonal independence pair satisfying the weak containment condition. This would
be not only an analogue of Proposition 4.16 (3) of [22], but an analogue of our recent results in [17],
2
where it is shown (see Theorem 1.1 of [17]) that the Koopman representation of a probability measurepreserving action with positive entropy must contain a nonzero subrepresentation of the left regular
representation. The major application in our paper of independence tuples is the following question of
Deninger (see [8], question 26).
Question 1. If Γ is a countable discrete group and f ∈ Mn (Z(Γ)) ∩ GLn (ℓ1 (Γ)) which is not invertible
in Mn (Z(Γ)), is it true that detL(Γ) (f ) > 1?
Here L(Γ) denotes the group von Neumann algebra, which is the strong operator topology closure of
C(Γ) in the left regular representation on ℓ2 (Γ) given by
(gξ)(h) = ξ(g −1 h), g, h ∈ Γ.
Also detL(Γ) (f ) is the Fuglede-Kadison determninant of f, a generalization of the usual determinant in
linear algebra to the infinite-dimensional setting of operators in Mn (L(Γ)) see [27] Chapter 3.2 for the
precise definition. Chung-Li answered this affirmatively in Corollary 7.9 of [5] for all amenable groups
using independence tuples. Following on the techniques in [5], David Kerr and Hanfeng Li in [22] were
able to answer this in the affirmative when Γ is residually finite. Both of these proofs use independence
tuples and their previous calculations of topological entropy for algebraic actions of residually finite
groups or amenable groups. This was further exploited by Chung-Li in [5] to describe algebraic actions
of amenable groups with completely positive entropy. Using the main result of [15], and Theorem 6.8
in [22] one immediately affirmatively answers Deninger’s Problem for sofic groups. However, we will
be interested in generalizing this result to a larger class of f. We will weaken the assumption that
f ∈ GLn (ℓ1 (Γ)).
To motivate our generalization, let us consider the case Γ = Z, and f ∈ Z(Z), and view f as a
Laurent polynomial. By Jensen’s Formula, one can show that detL(Z) (f ) > 1 if and only if f has a
leading coefficient of modulus one and does not have all of its roots on the unit circle. Using Fourier
analysis, we see that f is invertible in ℓ1 if and only if f never vanishes on the unit circle. In particular,
if f is invertible in ℓ1 then detL(Z) (f ) > 1. This analysis also generalizes to any abelian group.
We note here that the Gelfand transforms on ℓ1 (Z) and Cλ∗ (Z) of f are both the Fourier transform,
so f is invertible in ℓ1 (Z) if and only if f is invertible in Cλ∗ (Z) (equivalently L(Z)). Consideration of
the abelian case leads us to believe that it is reasonable to expect that if f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)) is
not invertible in Mn (Z(Γ)), then detL(Γ) (f ) > 1. We prove this is true in the sofic case.
Theorem 1.1. Let Γ be a countable discrete sofic group, and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). If f is not
invertible in Mn (Z(Γ)), then detL(Γ) (f ) > 1.
For readers unfamiliar with operator algebras, we note that f ∈ GLn (L(Γ)) is the same as saying
that f is invertible as a left convolution operator
ℓ2 (Γ)⊕n → ℓ2 (Γ)⊕n .
We also mention in Section 4 a wide class of examples of f ∈ Z(Γ) ∩ L(Γ)× as to illustrate that the
above Theorem is a significant generalization of the case f ∈ Z(Γ) ∩ ℓ1 (Γ)× . We actually prove the above
Theorem by using our results in [15]. For notation, if f ∈ Mn (L(Γ)) we define
r(f ) : ℓ2 (Γ)⊕n → ℓ2 (Γ)⊕n
by
(r(f )ξ)(l) =
n X
X
m=1 g∈Γ
if flm =
P
g∈Γ
fd
lm (g)g for 1 ≤ l, m ≤ n. We then set
ξ(l)(g)fd
lm (g)
Xf = (Z(Γ)⊕n /r(f )Z(Γ)⊕n )b.
Where the hat indicates that we are taking the Pontryagin dual, i.e we look at the compact, abelian
group of al continuous homomorphisms from Z(Γ)⊕n /r(f )Z(Γ)⊕n into T = R/Z. Here we are identifying
Z(Γ) inside of ℓ2 (Γ) via
X
X
fb(g)g 7→
fb(g)χ{g} .
g∈Γ
g∈Γ
3
The compact, abelian group Xf inherits a natural action of Γ by
(gθ)(a) = θ(g −1 a), for θ ∈ Xf , a ∈ Z(Γ)⊕n /r(f )Z(Γ)⊕n , g ∈ Γ.
The proof of Theorem 1.1 then follows from the main result of [15], the following Theorem and the results
of [22].
Theorem 1.2. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi , and let
f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). Then every k-tuple of points in Xf is a (σi )i − IE −k-tuple.
By Theorem 1.1 of [15] as well as Theorem 6.8 in [22] the above Theorem implies Theorem 1.1.
Crucial in the proof of this theorem is both the reduction to ℓ2 -independence tuples and independence
tuples satisfying the weak containment condition. If Ji ⊆ {1, . . . , di } is our candidate independent set
and
(k)
(1)
J i = Ji ∪ · · · ∪ Ji
is our candidate partition, we will need to control
(s)
kσi (α)(χJ (s) − udi (Ji )1)k2
i
for α ∈ C(Γ), 1 ≤ s ≤ k. In particular, since we only assume that f ∈ GLn (L(Γ)) we need to control by
the norm of α in the left regular representation. Because of this, our modified notion of independence
will be the key to proving the above theorem. Thus, we will actually show the more general fact that
every k-tuple of points in Xf is a independence tuple satisfying the weak containment condition.
As a consequence of our work we have the following application to completely positive entropy. Recall
that if Γ is a sofic group, with sofic approximation σi : Γ → Sdi , then Γ y X where X is a compact
metrizable space, and Γ acts by homeomorphisms is said to have completely positive entropy if every
nontrivial factor has positive entropy. Similarly, a probability measure-preserving action is said to have
completely positive entropy if every nontrivial (measure-theoretic) factor has positive entropy.
Corollary 1.3. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)) be not invertible in Mn (Z(Γ)). Then Γ y Xf has completely positive
topological entropy with respect to any sofic approximation. If Γ is amenable and λXf is the Haar
measure on Xf , then Γ y (Xf , λXf ) has completely positive entropy as well.
The amenable case uses important results from [5]. For f ∈ Mn (Z(Γ)) ∩ GLn (ℓ1 (Γ)), the case of
topological entropy is a consequence of the results in [22]. The case of amenable groups and measuretheoretic entropy is contained in the results of [5], again in the situation in which f ∈ Mn (Z(Γ)) ∩
GLn (ℓ1 (Γ)). In Section 4, we will list examples of f ∈ Z(Γ) ∩ L(Γ)× which are not ℓ1 (Γ)× when Γ is
amenable. For example, if Γ is elementary amenable then a result of Chou in [4] implies that Z(Γ) ∩
L(Γ)× ⊆ ℓ1 (Γ)× if and only if Γ is virtually nilpotent. This examples reveal that, even in the amenable
case, the generalization from invertbility in ℓ1 (Γ) to invertibility in L(Γ) is significant.
Acknowledgments. I thank the anonymous referee for their comments, which vastly improved the
understandability of the paper.
2
ℓp -Versions of Independence Tuples
Let us first recall the definition of a sofic group. For an n ∈ N, we let un be the uniform measure on
{1, . . . , n}. In general, if A is a finite set, we use uA for the uniform probability measure on A. We use
Sn for the symmetric group on n letters.
Definition 2.1. Let Γ be a countable discrete group. A sofic approximation is a sequence of functions
(not assumed to be homomorphisms) σi : Γ → Sdi so that
udi ({1 ≤ j ≤ di : σi (gh)(j) = σi (g)σi (h)(j)}) → 1, for all g, h ∈ Γ
udi ({1 ≤ j ≤ di : σi (g)(j) 6= j}) → 1, for all g ∈ Γ \ {e}.
We say that Γ is sofic if it has a sofic approximation.
4
It is known that all amenable groups and residually finite groups are sofic. Also, it is known that
soficity is closed under free products with amalgamtion over amenable subgroups (see [13],[29],[11],[10],
[30]). In fact, it is shown in [6] that graph products of sofic groups are sofic. Additonally, residually sofic
groups and locally sofic groups are sofic. By Malcev’s Theorem, this implies all linear groups are sofic.
Finally, if Λ is a subgroup of Γ and Γ y Γ/Λ is amenable (i.e. there is a Γ invariant mean on Γ/Λ), then
Γ is sofic. For a pseudometric space (X, ρ) and A a finite set, and 1 ≤ p ≤ ∞ we define
1 X
ρp,A (x, y)p =
ρ(x(a), y(a))p , if p < ∞
|A|
a∈A
ρ∞,A (x, y) = max ρ(x(a), y(a)).
a∈A
If A = {1, . . . , n} we shall typically use
ρp,n
instead of
ρp,{1,...,n} .
We recall the definition of sofic entropy.
Definition 2.2. Let Γ be a countable discrete group and Γ y X by homeomorphisms. We say that
a continuous pseudometric ρ on X is dynamically generating if for all x 6= y, there is a g ∈ Γ so that
ρ(gx, gy) > 0.
For a pseudometric space (X, ρ), subsets A, B of X and ε > 0 we say that A is ε-contained in B and
write A ⊆ε B if for all a ∈ A, there is a b ∈ B with ρ(a, b) ≤ ε. We say that A ⊆ X is ε-dense if X ⊆ε A.
We use Sε (X, ρ) for the smallest cardinality of an ε-dense subset of X.
Definition 2.3. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Fix
a continuous dynamically generating pseudometric ρ on X. For a finite F ⊆ Γ, and δ > 0, we let
Map(ρ, F, δ, σi ) be all φ ∈ X di so that
max ρ2,di (φ ◦ σi (g), gφ) < δ.
g∈F
Definition 2.4. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
X be a compact metrizable space with Γ y X by homeomorphisms. Fix a continuous dynamically
generating pseudometric ρ on X. Define the entropy of Γ y X with respect to σi by
h(σi )i (ρ, F, δ, ε) = lim sup
i→∞
1
log Sε (Map(ρ, F, δ, σi ), ρ2,di )
di
h(σi )i (ρ, ε) = inf h(σi )i (ρ, F, δ, ε)
F,δ
h(σi )i (X, Γ) = sup h(σi )i (ρ, ε).
ε>0
In [21] Theorem 4.5 and [19] Proposition 2.4, it is shown that this does not depend upon the choice
of continuous dynamically generating pseudometric. In [22], Kerr-Li defined independence tuples as a
topological measure of randomness of the action, and connected it with positivity of topological entropy.
One of the main results of [22] of relevance for us is Proposition 4.16 (3), which shows that positivity of
entropy is equivalent to the existence of a nondiagonal independence pair. For our purposes, it will be
convenient to consider ℓq -versions of independence tuples.
Definition 2.5. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
X be a compact metrizable space with Γ y X by homeomorphisms, and fix a continuous dynamically
generating pseudometric ρ on X, and 1 ≤ q ≤ ∞. Let x = (x1 , . . . , xk ) ∈ X k . For finite F, K ⊆ Γ and
δ, ε > 0 we say that a subset J ⊆ {1, . . . , di } is a ℓq − (ρ, F, δ, σi ; ε, K)-independence set for x if for every
c : J → {1, . . . , k} there is a φ ∈ Map(ρ, F, δ, σi ) so that
max ρq,J (gφ(·), gxc(·) ) < ε.
g∈K
We let Iq (x, ρ, F, δ, σi ; ε, K) be the maximum of udi (J) where J is a ℓq − (ρ, F, δ, σi ; ε, K)-independence
set for x. Additionally, we let
Iq (x, ρ, F, δ, (σi )i ; ε, K) = lim sup Iq (x, ρ, F, δ, σi ; ε, K)
i→∞
5
Iq (x, ρ, (σi )i ; ε, K) =
inf
finite F ⊆Γ,
δ>0
Iq (x, ρ, F, δ, (σi )i ; ε, K).
We say that x is a ℓq − IE −tuple with respect to ρ if for all ε > 0, and finite K ⊆ Γ,
Iq (x, ρ, (σi )i ; ε, K) > 0.
We let
IEk(σi )i ,ρ (X, Γ, q)
be the set of all ℓq − IE −tuples with respect to ρ.
We shall typically denote IEk(σi )i ,ρ (X, Γ, q) by IEk(σi )i ,ρ (q) if X, Γ are clear from the context. Our goal
in this section is to show that IEk(σi )i ,ρ (q) is independent of the choice of ρ, q, and that in fact IEk(σi )i ,ρ (q)
is the set of independence k-tuples as defined by Kerr-Li in [22]. We will first show that IEk(σi )i ,ρ (q) does
not depend upon ρ.
Lemma 2.6. Let Γ be a countable discrete sofic group and X a compact metrizable space with Γ y X
by homeomorphisms. Let ρ, ρ′ be two continuous dynamically generating pseudometrics on X. Then for
any 1 ≤ q ≤ ∞,
IEk(σi )i ,ρ (q) = IEk(σi )i ,ρ′ (q).
Proof. Let M, M ′ be the diameters of ρ, ρ′ . Let x ∈ IEk(σi )i ,ρ′ (q). Let ε > 0 and a finite K ⊆ Γ be given.
Choose a finite K ′ ⊆ Γ and an ε′ > 0 so that
max ρ(gx, gy) < ε
g∈K
whenever
max ρ′ (gx, gy) < ε′ .
g∈K ′
Let η ′ > 0 depend upon ε in a manner to be determined later. Let α′ > 0 be such that
Iq (x, ρ′ , (σi )i ; η ′ , K ′ ) ≥ α′ .
Suppose we are given a finite F ⊆ Γ, and a δ > 0. By Lemma 2.3 in [22], we may choose a finite
F ′ ⊆ Γ and a δ ′ > 0 so that
Map(ρ′ , F ′ , δ ′ , σi ) ⊆ Map(ρ, F, δ, σi ).
Let Ji′ be a ℓq − (ρ′ , F ′ , δ ′ , σi ; η ′ , K ′ ) independence set of maximal cardinality. Suppose we are given
c : Ji′ → {1, . . . , k},
choose a φ ∈ Map(ρ′ , F ′ , δ ′ , σi ) so that
max ρ′q,Ji′ (gφ(·), gxc(·) ) < η ′ .
g∈K ′
Let
Ci =
\
{j ∈ Ji′ : ρ′ (gφ(j), gxc(j) ) < ε′ }.
g∈K ′
If q < ∞, then
uJi (Ji′
′
\ Ci ) ≤ |K |
η′
ε′
q
.
If q = ∞, we force η ′ < ε′ so that Ci = Ji′ . For j ∈ Ci we have by our choice of ε′ , K ′ that
ρ(gφ(j), gxc(j) ) < ε
for all g ∈ K. Thus for all g ∈ K, and q < ∞,
q
q
q
′
ρq,Ji′ (gφ(·), gxc(·) ) < ε + M |K |
η′
ε′
q
and if q = ∞, then
ρ∞,Ji′ (gφ(·), gxc(·) ) < ε′ .
Choosing η ′ > 0 sufficiently small (depending upon K, q) , we see that we have that Ji′ is a (ρ, F, δ, σi ; 2ε, K)independence set. Thus
Iq (ρ, F, δ, (σi ); 2ε, K) ≥ α′ .
As F, δ, ε are arbitrary this completes the proof.
6
Thus for 1 ≤ q ≤ ∞ we will use IEk(σi )i (q) for IEk(σi )i ,ρ (q) for any continuous dynamically generating
pseudometric ρ. If X, Γ are not clear from the context we will use
IEk(σi )i (X, Γ, q)
instead of IEk(σi )i (q). It is not hard to relate our notion of combinatorial independence to that developed
by Kerr-Li in [22]. We use IEk(σi )i for the set of (σi )i − IE −k-tuples as defined by Kerr-Li in [22] (again
we should really use IEk(σi )i (X, Γ) but in most cases Γ y X will be clear from the context).
Corollary 2.7. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let X
be a compact metrizable space with Γ y X by homeomorphisms. Then IEk(σi )i (∞) = IEk(σi )i .
Proof. It is easily seen that
IEk(σi )i = IEk(σi ),ρ (∞)
when ρ is a compatible metric. Now apply the preceding lemma.
We now show that in fact IEk(σi )i (q) does not depend upon q. We remark that the proof is closely
modeled on the proof of Proposition 4.6 in [22]. We will need Karpovsky and Milman’s generalization
of the Sauer-Shelah lemma (see [18],[31],[33]). For convenience we state the Lemma below.
Lemma 2.8. For any integer k ≥ 2 and any real number λ ∈ (k − 1, k) there is a constant β(λ) > 0 so
that for all n ∈ N if S ⊆ {1, 2, . . . , k}n has |S|1/n ≥ λ, then there is an I ⊆ {1, . . . , n} with |I| ≥ β(λ)n
and
S I = {1, 2, . . . , k}I .
Lemma 2.9. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let X be
a compact metrizable space with Γ y X by homeomorphisms. For any 1 ≤ q1 , q2 ≤ ∞ we have
IEk(σi )i (q1 ) = IEk(σi )i (q2 ).
Proof. It is clear that if q1 < q2 , then
IEk(σi )i (q1 ) ⊇ IEk(σi )i (q2 ).
It thus suffices to prove that
IEk(σi )i (1) ⊆ IEk(σi )i (∞).
Fix k − 1 < λ < k, and let β be the function defined in Lemma 2.8. Then we may find a n0 so that if J
is a finite set with |J| ≥ n0 and E ⊆ {1, . . . , k}J has
|E| ≥ λ|J| ,
then there is a J ′ ⊆ J with
|J ′ | ≥ β(λ)|J|
so that
E
′
J′
= {1, . . . , k}J .
Let ρ be a continuous dynamically generating pseudometric on X. Let
x = (x1 , . . . , xk ) ∈ IEk(σi )i (1).
Let ε > 0 and a finite K ⊆ Γ be given. Let ε′ > 0 depend upon ε in manner to be determined later. Set
α = I1 (x, ρ, (σi )i ; ε′ , K).
Suppose we are given a finite F ⊆ Γ and δ > 0. Choose Ji ⊆ {1, . . . , di } a ℓ1 − (ρ, F, δ, σi ; ε′ , K)independence set for x with
udi (Ji ) = I1 (ρ, F, δ, σi ; ε′ , K, x).
For every c : Ji → {1, . . . , k} choose a φc ∈ Map(ρ, F, δ, σi ) so that
max ρ1,Ji (gφc (·), gxc(·) ) < ε′ .
g∈K
7
Let
Ξc =
\
{j ∈ Ji : ρ(gφc (j), gxc(·) ) < ε}.
g∈K
Then
′
ε
uJi (Ji \ Ξc ) ≤ |K|
.
ε
Let
H(t) = −t log(t) − (1 − t) log(1 − t).
By
Stirling’s
formula there is a A > 0 so that the number of subsets of Ji of cardinality at most
ε′
|K| ε |Ji | is at most
′
′
ε
ε
A exp H |K|
|Ji | |K|
|Ji |.
ε
ε
Thus there is subset Ωi ⊆ {1, . . . , k}di with
′ |Ji |
k |Ji | exp H |K| εε
,
|Ωi | ≥
′
A|K| εε |Ji |
so that Ξc is the same, say equal to Θi , for all c ∈ Ωi . If we choose ε′ > 0 sufficiently small then
|Ωi | ≥ λ|Ji |
for all large i. So by our choice of β for all large i, we can find a Ji′ ⊆ Ji with
|Ji′ | ≥ β(λ)|Ji |
and
Ωi
′
Ji′
= {1, . . . , k}Ji .
Thus
lim sup udi (Ji′ ) ≥ β(λ) lim sup udi (Ji ) ≥ β(λ)α.
i→∞
i→∞
′
Choose ε > 0 sufficiently small so that
′
β(λ)
ε
≤
.
|K|
ε
2
As
′
ε
uJi (Ji \ Θi ) ≤
|K|,
ε
we find that
lim inf uJi (Ji′ ∩ Θi ) ≥
i→∞
so
lim sup udi (Ji′ ∩ Θi ) = lim sup
i→∞
i→∞
β(λ)
,
2
β(λ)
|Ji |
uJ (J ′ ∩ Θi ) ≥ α
.
di i i
2
We claim that Ji′ ∩ Θi is a ℓ∞ − (ρ, F, δ, σi ; ε, K) independence set for x for infinitely many i. Let
c′ : Ji′ ∩ Θi → {1, . . . , k}.
Since
Ωi
′
Ji′
= {1, . . . , k}Ji ,
we have
Ωi
So we may find a c ∈ Ωi so that c
Ji ∩Θi
′
Ji′ ∩Θi
= {1, . . . , k}Ji ∩Θi .
= c′ . Since Θi = Ξc we have that
max ρ(gφc (j), gxc(j) ) < ε
g∈K
8
for all j ∈ Ji′ ∩ Θi . As
lim sup udi (Ji′ ∩ Θi ) ≥
i→∞
β(λ)
α,
2
we find that
lim sup I∞ (ρ, F, δ; ε, K) ≥
i→∞
Thus
3
β(λ)
α.
2
x ∈ IEk(σi )i (∞).
Independence Tuples with a Weak Containment Condition
We now proceed to give our strengthening of independence tuples. The basic idea is instead of considering
our sequence of independence sets (Ji )i≥1 to be arbitrary subsets of {1, . . . , di } we require that in the
“representation” Γ has on ℓ2 (di ) we have that the “subrepresentation” generated by (χJi − udi (Ji )1) is
weakly contained in the left regular representation. Moreover, instead of considering arbitrary partitions
c : Ji → {1, . . . , k}
we only consider one so that the pieces Ji,l = c−1 ({l}) also have the property that the “subrepresentation”
generated by (χJi,l − udi (Ji,l )1) is weakly contained in the left regular representation. The results in [17]
indicate that positivity of entropy is related in an essential way to the left regular representation. Our
modified version of independence tuple is more natural from that perspective. An essential difficulty
here is that since σi is not an honest homomorphism, we do not get an honest representation this way.
As we shall see shortly, one can get around this using ultraproducts. For now, we simply give a direct
definition.
First let us introduce some notation. For a countable discrete group Γ, define the left regular representation
λ : Γ → U (ℓ2 (Γ))
by
(λ(g)ξ)(h) = ξ(g −1 h).
Extend λ to a map λ : C(Γ) → B(ℓ2 (Γ)) by
X
X
λ
αg g =
αg λ(g).
g∈Γ
g∈Γ
For α ∈ C(Γ), g ∈ Γ set
α
b(g) = hλ(α)δe , δg i.
Then
α=
X
g∈Γ
α
b(g)g.
If Γ is a sofic group with sofic approximation σi : Γ → Sdi , we define σi : C(Γ) → Mdi (C) by
X
σi (f ) =
fb(g)σi (g).
g∈Γ
Here we are viewing σi (g) as a permutation matrix.
Definition 3.1. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . For a
finite F ⊆ C(Γ), and a δ > 0 we let Λ(F, δ, σi ) be all J ⊆ {1, . . . , di } so that
max kσi (f )(χJ − udi (J)1)k2 ≤ kλ(f )kkχJ − udi (J)1k2 + δ.
f ∈F
We let Λ(σi )i be the set of all sequences (Ji )i≥1 with Ji ⊆ {1, . . . , di } so that for every finite F ⊆ C(Γ),
and for every δ > 0 it is true that for all large i, Ji ∈ Λ(F, δ, σi ).
9
We now mention the formalization via ultrafilters. Fix a free ultrafilter ω ∈ βN \ N. Let
∞
{(fi )∞
i=1 : fi ∈ ℓ (di ), supi kfi k∞ < ∞}
o.
A= n
∞
(fi )∞
i=1 : fi ∈ ℓ (di ), supi kfi k∞ < ∞, limi→ω kfi kℓ2 (di ,udi ) = 0
For a sequence fi ∈ ℓ∞ (di , udi ) we use (fi )i→ω for the image in A of (fi )i≥1 under the quotient map.
There is a well-defined inner product on A given by
di
1 X
h(fi )i→ω , (gi )i→ω i = lim
fi (j)gi (j).
i→ω di
j=1
Let L2 (A, uω ) be the completion of A under this inner-product. Then we have a well-defined unitary
representation
σω : Γ → U(L2 (A, uω ))
defined densely by
σω (g)(fi )i→ω = (fi ◦ σi (g))i→ω
if fi ∈ ℓ∞ (di , udi ) and
sup kfi k∞ < ∞.
i
We then see that Λ(σi )i can be regarded as all sequences Ji of subsets of {1, . . . , di } so that if ω ∈ βN \ N
is any free ultrafilter and we set
χoJω = (χJi − udi (Ji )1)i→ω ,
K = Span{σω (g)χoJω : g ∈ Γ},
then the representation Γ y K is weakly contained in the left regular representation (see [1] Appendix
F for the relevant facts about weak containment of representations).
Definition 3.2. Let Γ be a countable discrete sofic group and σi : Γ → Sdi a sofic approximation. Let
X be a compact metrizable space with Γ y X by homeomorphisms. Let ρ be a continuous dynamically
generating pseudometric on X. Let 1 ≤ q ≤ ∞, and x ∈ X k . Fix finite K, F ⊆ Γ, E ⊆ C(Γ) and ε, δ, η > 0.
We say that a sequence (Ji )i is a Λ − ℓq − (ρ, F, δ, E, η; ε, K)-independence sequence for x if (Ji )i ∈ Λ(σi )i ,
and for all i, for all c : Ji → {1, . . . , k} with c−1 ({l}) ∈ Λ(E, η, σi ), there is a φ ∈ Map(ρ, F, δ, σi ) with
max ρq,Ji (gφ(·), gxc(·) ) < ε.
g∈K
We let IΛ,q (x, ρ, F, δ, E, η, (σi )i ; ε, K) be the supremum of
lim sup udi (Ji )
i→∞
over all sequences (Ji ) which are Λ − ℓq − (ρ, F, δ, E, η; ε, K)-independence sequences. We then set
IΛ,q (x, ρ, F, δ, (σi )i ; ε, K) =
IΛ,q (x, ρ, F, δ, E, η; ε, K),
sup
finite E⊆C(Γ),
η>0
IΛ,q (x, ρ, (σi )i ; ε, K) =
inf
finiteF ⊆Γ,
δ>0
IΛ,q (x, ρ, F, δ; ε, K).
Definition 3.3. Let Γ be a countable discrete group and σi : Γ → Sdi a sofic approximation. Let X
be a compact metrizable space with Γ y X by homeomorphisms. Let ρ be a continuous dynamically
generating pseudometric on X, and 1 ≤ q < ∞. We say that x = (x1 , . . . , xk ) is a ℓq − (σi ) − IE −k-tuple
satisfying the weak containment condition (or a Λ(σi )i − IE −k-tuple) if for every ε > 0 and K ⊆ Γ finite
IΛ,q (x, ρ, (σi )i ; ε, K) > 0.
q
We use IEΛ,k
(σi ),ρ (X, Γ, q) for the set of ℓ − (σi ) − IE − k-tuples satisfying the weak containment condition.
If X, Γ are clear from context we will simply use IEΛ,k
(σi ),ρ (q).
10
In fact, following the proof of Lemma 2.6, one shows that IEΛ,k
(σi ),ρ (q) is independent of ρ, so we simply
Λ,k
use IEΛ,k
(σi )i (q). However, we do not know if IE(σi )i (q) is independent of q. Note that if M is the diameter
of (X, ρ) then by standard Hölder estimates we have for any finite set A and 1 ≤ q1 < q2 < ∞ :
q1
q1
ρq1 ,A (x, y) ≤ ρq2 ,A (x, y) ≤ M 1− q2 ρq1 ,A (x, y) q2 , for any x, y ∈ X A .
From this it is not hard to see that
Λ,k
IEΛ,k
(σi )i (q1 ) = IE(σi )i (q2 )
for all 1 ≤ q1 , q2 < ∞.
The definition may seem a little ad hoc. The following proposition will hopefully make it seem more
natural. Essentially, this proposition will tell us two things: first if we fix a α > 0, and choose a subset
Ji ⊆ {1, . . . , di } of size roughly αdi uniformly at random, then (Ji )i will be in Λ(σi ) with high probability.
Secondly, suppose we are given a sequence (Ji )i in Λ(σi ) a finite E ⊆ Γ and a η > 0, and a probability
measure µ on {1, . . . , k}. If we choose a partition of (Ji )i into sets of size µ({1})udi (Ji ), . . . , µ({k})udi (Ji )
uniformly at random, then with high probability, each of the pieces of the partition will be in Λ(E, η, σi ).
Thus we may view independence tuples satisfying the weak containment condition as simply a randomization of independence tuples as defined by Kerr-Li in [22]. The proof is a simple adaption of Bowen’s
argument for the computation of sofic entropy of Bernoulli shifts in [2].
Proposition 3.4. Let Γ be a countable discrete sofic group, with sofic approximation σi : Γ → Sdi . Let
µ be a probability measure on {1, . . . , k}. Fix a sequence (Ji )i ∈ Λ(σi )i . Then, for any E ⊆ C(Γ) finite,
and any η > 0 and any 1 ≤ l ≤ k we have
µ⊗Ji ({p ∈ {1, . . . , k}Ji : p−1 ({l}) ∈ Λ(E, η, σi )}) → 1,
µ⊗Ji ({p ∈ {1, . . . , k}Ji : |udi (p−1 ({l})) − µ({l})udi (Ji )| > η}) → 0.
Proof. As our claim is probabilistic, we may assume E = {f }. We make the following two claims.
Claim 1: For all F ⊆ Γ \ {e} finite, for every 1 ≤ l ≤ k, for every δ > 0, we have
µ⊗Ji p ∈ {1, . . . , k}Ji : udi (σi (g)p−1 ({l}) ∩ p−1 ({l})) − udi (Ji ∩ σi (g)Ji )µ({l})2 > δ for some g ∈ F → 0.
Claim 2 : For every 1 ≤ l ≤ k, for every δ > 0, we have
µ⊗Ji ({p ∈ {1, . . . , k}Ji : |udi (p−1 ({l})) − µ({l})udi (Ji )| > δ}) → 0.
Suppose we accept the two claims. Then, we may find a subset Pi ⊆ {1, . . . , k}Ji so that for every
sequence pi ∈ Pi ,
udi (p−1
i ({l})) − µ({l})udi (Ji ) → 0,
−1
2
|udi (σi (g)p−1
i ({l}) ∩ pi ({l})) − µ({l}) udi (Ji ∩ σi (g)Ji )| → 0, for every g ∈ Γ \ {e}
and
µ⊗Ji (Pi ) → 1.
Fix a sequence pi ∈ Pi . Let ξl = χp−1 ({l}) − udi (p−1
i ({l}))1, ζ = χJi − udi (Ji )1. We use o(1) for any
i
expression which goes to zero as i → ∞. Then for any f ∈ C(Γ) and i ∈ N
kσi (f )ξl k22 =
X
g,h∈Γ
fb(g)fb(h)hσi (h)−1 σi (g)ξl , ξl i
= o(1) +
X
g,h∈Γ
= o(1) +
X
g∈Γ
(1)
fb(g)fb(h)hσi (h−1 g)ξl , ξl i
∗ f (g)hσ (g)ξ , ξ i
fd
i
l l
We have that
−1
2
2
2
hσi (e)ξl , ξl i = o(1)+kξl k22 = o(1)+udi (p−1
i ({l}))−udi (pi ({l})) = o(1)+µ({l})udi (Ji )−µ({l}) udi (Ji ) .
(2)
11
and for g 6= e we have
−1
−1
2
hσi (g)ξl , ξl i = udi (σi (g)p−1
i ({l}) ∩ pi ({l})) − udi (pi ({l}))
= o(1) + µ({l})2 udi (Ji ∩ σi (g)Ji ) − µ({l})2 udi (Ji )2
= o(1) + µ({l})2 hσi (g)ζ, ζi.
Additionally
kζk22 = udi (Ji ) − udi (Ji )2 .
(3)
Combining (1), (2), (3) we see that
∗ f (e)(µ({l})u (J ) − µ({l})2 u (J )2 )
kσi (f )ξl k22 = o(1) + fd
di i
di i
X
2
∗
d
f f (g)hσi (g)ζ, ζi
+ µ({l})
g∈Γ\{e}
∗ f (e)(µ({l}) − µ({l})2 )u (J )
= o(1) + fd
di i
X
2
∗ f (g)hσ (g)ζ, ζi.
+ µ({l})
fd
i
g∈Γ
By the same logic,
kσi (f )ζk22 = o(1) +
X
g∈Γ
Thus
∗ f (g)hσ (g)ζ, ζi.
fd
i
∗ f (e)(µ({l}) − µ({l})2 )u (J ) + µ({l})2 kσ (f )ζk2 .
kσi (f )ξl k22 = o(1) + fd
i
di i
2
Since Ji ∈ Λ(σi )i ,
kσi (f )ζk22 ≤ kλ(f )k2 (udi (Ji ) − udi (Ji )2 ) + η
for all large i. Thus for all large i,
kσi (f )ξl k22 ≤ o(1) + η + µ({l})2 kλ(f )k2 (udi (Ji ) − udi (Ji )2 )
+ kλ(f )k2 (µ({l}) − µ({l})2 )udi (Ji )
= o(1) + η + kλ(f )k2 (µ({l})udi (Ji ) − µ({l})2 udi (Ji )2 ).
Since
|kξl k22 − (µ({l})udi (Ji ) − µ({l})2 udi (Ji )2 )| → 0,
and η is arbitrary this proves the proposition.
We thus turn to the proof of Claim 1 and Claim 2. For Claim 1, it suffices to assume F = {g}. We
have
Z
udi (σi (g)p−1 ({l}) ∩ p−1 ({l})) dµ⊗Ji (p) =
di Z
1 X
χp−1 ({l}) (j)χp−1 ({l}) (σi (g)−1 (j)) dµ⊗Ji (p).
di j=1
Note that χp−1 ({l}) (j)χp−1 ({l}) (σi (g)−1 (j)) can only be positive if j ∈ σi (g)Ji ∩ Ji . Thus the above sum
is
Z
X
1
χ{l} (p(j))χ{l} (p(σi (g)−1 (j))) dµ⊗Ji (p).
di
j∈σi (g)Ji ∩Ji
Since
udi ({1 ≤ j ≤ di : σi (g)(j) 6= j}) → 1,
we have that
1
di
X
j∈σi (g)Ji ∩Ji
Z
χ{l} (p(j))χ{l} (p(σi (g)−1 (j))) dµ⊗Ji (p) = udi (σi (g)Ji ∩ Ji )µ({l})2 + o(1).
12
By Chebyshev’s inequality, it thus suffices to show that
Z
udi (σi (g)p−1 ({l}) ∩ p−1 ({l}))2 dµ⊗Ji (p) = udi (σi (g)Ji ∩ Ji )2 µ({l})4 + o(1).
For this, we have
1
d2i
X
j,k∈σi (g)Ji ∩Ji
Z
Z
udi (σi (g)p−1 ({l}) ∩ p−1 ({l}))2 dµ⊗Ji (p) =
χ{l} (p(j))χ{l} (p(σi (g)−1 (j)))χ{l} (p(k))χ{l} (p(σi (g)−1 (k))) dµ⊗Ji (p).
We claim that
udi ⊗ udi ({(j, k) : |{j, k, σi (g)−1 (j), σi (g)−1 (k)}| = 4}) → 1, as i → ∞.
(4)
We already know that
udi ⊗ udi ({(j, k) : |{j, k, σi (g)−1 (j) 6= j, σi (g)−1 (k) 6= k}|}) → 1, as i → ∞.
Additionally,
udi ⊗ udi ({(j, k) : j 6= k}) → 1, as i → ∞.
Thus it suffices to show that
udi ⊗ udi ({(j, k) : j 6= k, σi (g)−1 (j) 6= σi (g)−1 (k), |{j, k, σi (g)−1 (j), σi (g)−1 (k)}| < 4}) → 0.
Suppose then that j 6= k, σi (g)−1 (j) 6= j, σi (g)−1 (k) 6= k. Then, σi (g)−1 (j) 6= σi (g)−1 (k). So
|{j, k, σi (g)−1 (j), σi (g)−1 (k)}| < 4
if and only if j = σi (g)−1 (k) or k = σi (g)−1 (j). However the union of {(j, k) : σi (g)−1 (k) = j}, {(j, k) :
k = σi (g)−1 (j)} has cardinality at most 2di . This proves (4). So
Z
udi (σi (g)p−1 ({l}) ∩ p−1 ({l}))2 dµ⊗Ji (p) = o(1) + udi (σi (g)Ji ∩ Ji )2 µ({l})4 .
This proves Claim 1.
The proof of Claim 2 is similar, and in fact has already been done by Bowen in [2] Theorem 8.1, it
can also be seen as a consequence of the law of large numbers.
We now show that the set of ℓq -independence tuples satisfying the weak containment condition is
contained in the set of ℓq -independence tuples.
Proposition 3.5. Let Γ be a countable discrete group with sofic approximation σi : Γ → Sdi . Let X be
a compact metrizable space with Γ y X by homeomorphisms. Then for any 1 ≤ q ≤ ∞,
k
IEΛ,k
(σi )i (q) ⊆ IE(σi )i (q).
Proof. Fix a compatible metric ρ on X. Let x = (x1 , . . . , xk ) ∈ IEΛ,k
(σi )i (q). Let ε > 0, and K ⊆ Γ finite
be given. Set
α = IΛ,q (x, ρ, (σi )i ; ε, K).
Fix k − 1 < λ < k and let β(λ) be as in the Sauer-Shelah Lemma.
Suppose we are given a finite F ⊆ Γ, and δ > 0. Choose a finite E ⊆ C(Γ), and a η > 0 so that
IΛ (x, ρ, F, δ, E, η, (σi )i ; ε, K) ≥
λ
α.
k
q
Let (Ji )∞
i=1 ∈ Λ(σi ) be a ℓ − (ρ, F, δ, E, η)-independence sequence for x with
lim sup udi (Ji ) ≥
i→∞
13
λ
α.
k
Let Λk (E, η, Ji ) be the set of all c : Ji → {1, . . . , k} so that c−1 ({l}) ∈ Λ(E, η, σi ) for 1 ≤ l ≤ k. By
Proposition 3.4 we have
|Λk (E, η, Ji )|
lim
= 1.
i→∞
k |Ji |
So by Lemma 2.8, for all large i we can find Ji′ ⊆ Ji with
Λk (E, η, Ji )
′
Ji′
= {1, . . . , k}Ji
and
udi (Ji′ ) ≥ β(λ)udi (Ji ).
We claim that Ji′ is a ℓq − (ρ, F, δ, σi )-independence set for x for all large i.
For this, let c′ : Ji′ → {1, . . . , k}. Then, there is a c ∈ Λk (E, η, Ji ) so that
c
Ji′
= c′ .
By the definition of (ρ, F, δ, E, η, (σi )i ) independence, there is a φ ∈ Map(ρ, F, δ, σi ) so that
max ρq,Ji (gφ(·), gxc(·) ) < ε.
g∈K
As c
Ji′
= c′ , we find that
max ρq,Ji′ (gφ(·), gxc′ (·) ) <
g∈K
ε
.
β(λ)
As
lim sup udi (Ji′ ) ≥ β(λ)α,
i→∞
we see that
Iq
ε
ρ, F, δ, (σi );
,K
β(λ)
≥ β(λ)α.
Taking the infimum over all F, δ completes the proof.
We now discuss the analogue of Proposition 4.16 from [22] for independence tuples with a weak
containment condition. Recall that if X, Y are compact metrizable spaces and Γ y X, Γ y Y by
homeomoprhisms, then a continuous, Γ-equivariant, surjection π : X → Y is called a factor map. If there
is a factor map π : X → Y, we call Y a factor of X.
Proposition 3.6. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Fix
1 ≤ q ≤ ∞. Let X be a compact metrizable space with Γ y X by homeomorphisms.
(1): If IEΛ,2
(σi )i (q) \ {(x, x) : x ∈ X} is nonempty, then h(σ)i (X, Γ) > 0.
k
k
(2): We have that IEΛ,k
(σi )i (q) is a closed Γ-invariant subset of X , where Γ y X is the product
action.
(3): Let Y be a compact metrizable space with Γ y Y by homeomorphisms and π : X → Y a factor
map. Then
Λ,k
π k (IEΛ,k
(σi )i (q, X, Γ)) ⊆ IE(σi )i (q, Y, Γ).
Λ,k
(4): Suppose that Z is a closed Γ-invariant subset of X, then IEΛ,k
(σi )i (Z, Γ) ⊆ IE(σi )i (X, Γ).
Proof. (1): This is a consequence of the preceding Proposition and Proposition 4.16 (3) in [22].
(2): Fix a compatible metric ρ on X and g ∈ Γ. Let αg (x) = gx. Then for any finite F ⊆ Γ, for any
δ > 0, there is a δ ′ > 0 so that if
φ ∈ Map(ρ, {g −1 } ∪ (g −1 F ) ∪ {g}, δ ′ , σi )
then
αg ◦ φ ◦ σi (g)−1 ∈ Map(ρ, F, δ, σi ),
for all large i. Thus,
IEΛ,k
(σi )i
is Γ-invariant. The fact that it is closed is a trivial consequence of the definitions.
14
(3): Let ρ, ρ′ be compatible metrics on X, Y. Let M, M ′ be the diameter of ρ, ρ′ . Suppose we are
given a ε′ > 0, and let η ′ > 0 depend upon ε to be determined shortly. Choose a ε > 0 so that
ρ(x, y) < ε
implies
ρ′ (π(x), π(y)) < η ′ .
Let η > 0 depend upon ε in a manner to be determined later. Given a finite F ′ ⊆ Γ finite and a δ ′ > 0
we can find a finite F ⊆ Γ and a δ > 0 so that
π di (Map(ρ, F, δ, σi )) ⊆ Map(ρ′ , F ′ , δ ′ , σi ),
(this follows by the same argument in Lemma 2.3 of [22]). Let x ∈ X k , and let Ji be a (x, ρ, F, δ; η, {e})independence set, and suppose we are given
c : Ji → {1, . . . , k}.
Choose φ ∈ Map(ρ, F, δ, σi ) so that
ρq,Ji (φ, xc(·) ) < η.
Then
uJi ({j ∈ Ji : ρ(φ(j), xc(j) ) < ε}) ≥ (1 −
By our choice of ε,
q
′ q
ρq (π ◦ φ, π(xc(·) )) ≤ (η ) + M
ηq
εq
ηq
).
εq
.
Choosing η, η ′ appropriately we have that Ji is a (π(x), ρ, F ′ , δ ′ ; ε′ , {e}) independence set.
(4): This is trivial.
We now proceed to show that ℓq −Λ(σi ) -tuples are the same as ℓq -independence tuples in the amenable
case. For this, we will need the following general fact: if Γ is an amenable group and π : Γ → U(H) is a
unitary representation, π is weakly contained in λ. See [1] Theorem G.3.2 for a proof.
Proposition 3.7. Let Γ be a countable discrete amenable group. Let σi : Γ → Sdi be a sofic approximation. Then every sequence of subsets of {1, . . . , di } is in Λ(σi )i .
Proof. Automatic from the remarks about ultraproducts and weak containment following Definition 3.1.
Proposition 3.8. Let Γ be a countable discrete amenable group. Let σi : Γ → Sdi be a sofic approximation. Let X be a compact metrizable space with Γ y X by homeomorphisms. Then for any 1 ≤ q ≤ ∞,
IEk(σi )i (q) = IEΛ,k
(σi )i (q).
Proof. Fix a compatible metric ρ on X. By Proposition 3.7 we have
IEk(σi )i (q) ⊇ IEΛ,k
(σi )i (q).
Conversely, let (x1 , . . . , xk ) be in IEk(σi )i but not in IEΛ,k
(σi )i . Choose a ε > 0 and a finite K ⊆ Γ so that
IΛ,q (x, ρ, (σi )i ; ε, K) = 0.
Since
Iq (x, ρ, (σi )i ; ε, K) > 0,
we can find a finite F ⊆ Γ and a δ > 0 so that
IΛ,q (x, ρ, F, δ, (σi )i ; ε, K) < Iq (x, ρ, F, δ, (σ)i ; ε, K).
Choose a finite E ⊆ Γ, and a η > 0 so that
IΛ,q (x, ρ, F, δ, E, η, (σi )i ; ε, K) < Iq (x, ρ, F, δ, (σi )i ; ε, K).
15
Choose (Ji )i≥1 ∈ Λ(σi )i so that (Ji )i≥1 is a Λ − ℓq − (ρ, F, δ, E, η; ε, K) independence sequence with
lim sup udi (Ji ) = IΛ (x, ρ, F, δ, σi ; ε, K).
i→∞
Since
IΛ,q (x, ρ, F, δ, E, η, (σi )i ; ε, K) < Iq (x, ρ, F, δ, (σi )i ; ε, K) = lim sup udi (Ji )
i→∞
we can find a subsequence il , and a partition
(1)
(k)
Jil = Jil ∪ · · · ∪ Jil
(pl )
so that there is a 1 ≤ pl ≤ k with Jil
that pl = p is constant. Thus,
∈
/ Λ(E, η, σil ). Passing to a further subsequence, we may assume
(p)
(Jil )l≥1 ∈
/ Λ(σil )l ,
contradicting Proposition 3.7.
4
A Generalization of Deninger’s Problem For Sofic Groups
Let Γ be a countable discrete group. An algebraic action of Γ is an action Γ y X by automorphisms,
where X is a compact, metrizable, abelian group. An equivalent way to describe this family of actions
b = Hom(A, T) where T = R/Z and A
b is given the
is to start with a countable Z(Γ) module A, and let A
b by
topology of pointwise convergence. We then have the algebraic action Γ y A
(gχ)(a) = χ(g −1 a).
By Pontryagin duality, all algebraic actions arise in this manner. We will mainly be interested in the
b is denoted Xf . An interesting
case A = Z(Γ)⊕n /r(f )(Z(Γ)⊕m ), where f ∈ Mm,n (Z(Γ)), in this case A
aspect of the subject, which has seen great mileage in recent years, (see e.g. [32] Lemma 1.2, Theorem
1.6, [5] Theorem 3.1, [7],[9],[26],[25], [24],[16],[15]) is that dynamical properties of algebraic actions (i.e.
b as an action on a compact metrizable space or probability space)
those which only depend upon Γ y A
b are related to functional analytic objects associated
such as entropy and independence tuples of Γ y A
to Γ. One relevant object is the group von Neumann algebra.
WOT
The group von Neumann algebra L(Γ) is defined to be λ(C(Γ))
, where WOT is the weak-operator
topology. Define τ : L(Γ) → C by τ (x) = hxδe , δe i. For A ∈ Mn (L(Γ)) define
Tr ⊗τ (A) =
n
X
τ (Ajj ).
j=1
Since L(Γ) ⊆ B(ℓ2 (Γ)), we can identify Mm,n (L(Γ)) ⊆ B(ℓ2 (Γ)⊕n , ℓ2 (Γ)⊕m ) in the natural way. For
x ∈ Mm,n (L(Γ)), we use kxk∞ for the operator norm of x (as an operator ℓ2 (Γ)⊕n → ℓ2 (Γ)⊕m ). We also
use
kxk22 = Tr ⊗τ (x∗ x).
Since we identify C(Γ) ⊆ L(Γ), we will us the same notation for elements of Mm,n (C(Γ)). We shall also
identify C(Γ)⊕n ∼
= M1,n (C(Γ)) and use the same notation. For f ∈ GLn (L(Γ)), the Fuglede-Kadison
determinant is defined by exp Tr ⊗τ (log |f |) (here the notation is as in [15]). A particular case of Theorem
4.4 in [15] shows that if Γ is sofic, then
h(σi )i (Xf , Γ) = log detL(Γ) (f ), for f ∈ GLn (L(Γ))
(in fact this is true when f is injective as an operator on ℓ2 (Γ)⊕n ). When Γ is sofic, it is known by [12]
that for f ∈ Mn (Z(Γ)) we have detL(Γ) (f ) ≥ 1. By multiplicativity of Fuglede-Kadison determinants (see
[27] Theorem 3.14 (1)) it follows that if f ∈ GLn (Z(Γ)), then detL(Γ) (f ) = 1. In [8], (see question 26)
Deninger asked a partial converse to this result. Namely, if f ∈ Mn (Z(Γ)) is invertible in GLn (ℓ1 (Γ))
but not invertible in Mn (Z(Γ)) is detL(Γ) (f ) > 1?
From Theorem 4.4 of [15], as well as Theorem 6.7 and Proposition 4.16 (3) in [22] we can automatically
answer Deninger’s problem affirmatively for sofic groups. Thus, we automatically have the following.
16
Theorem 4.1. Let Γ be a countable discrete sofic group and f ∈ Mn (Z(Γ)) ∩ GLn (ℓ1 (Γ)). If f is not in
GLn (Z(Γ)), then
detL(Γ) (f ) > 1.
In this section, we show how one can use Λ(σi ) -IE-tuples to generalize Deninger’s conjecture in the
case of sofic groups. In particular, in this section we show the following.
Theorem 4.2. Let Γ be a countable discrete sofic group. If f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)), but is not in
GLn (Z(Γ)), then detL(Γ) (f ) > 1.
To illustrate the significance of our generalization, we should mention examples of elements in
Mn (Z(Γ)) which are in GLn (L(Γ)) but are not in GLn (ℓ1 (Γ)). Let E ⊆ Γ, and let
∆E = 1 −
1 X
g ∈ Q(Γ).
|E|
g∈E
Note that ∆E is never invertible in ℓ1 (Γ). To see this, consider the homomorphism
t : ℓ1 (Γ) → C
given by
t(f ) =
X
f (g).
g∈Γ
Since t(∆E ) = 0, we know that ∆E is not invertible in ℓ1 (Γ).
First suppose that Γ is a nonamenable group. Let E ⊆ Γ be finite and symmetric, i.e. E = E −1 . By
nonamenability of Γ, we may choose E so that
1 X
λ(g) ≤ 1 − ε
|E|
g∈E
for some ε > 0 (see e.g. [3] Theorem 2.6.8 (8)). Thus λ(∆E ) ≥ ε as an operator on ℓ2 (Γ) and thus is
invertible. So |E|∆E ∈ Z(Γ) ∩ L(Γ)× but is not in ℓ1 (Γ)× and thus we always have examples of elements
in Z(Γ) ∩ L(Γ)× which are not invertible in ℓ1 (Γ) if Γ is nonamenable. So Theorem 4.2 applies to these
elements whereas Theorem 4.1 does not.
Theorem 4.2 is also new in the amenable case. Thus we wish to mention examples when Γ is amenable
of elements f ∈ Z(Γ) ∩ L(Γ)× which are not in ℓ1 (Γ)× . We say that Γ has subexponential growth if for
any finite E ⊆ Γ we have that
|{g1 . . . gn : g1 , . . . , gn ∈ E}|1/n →n→∞ 1.
If Γ has subexponential growth then α ∈ C(Γ) is invertible in L(Γ) if and only if it is invertible in ℓ1 (Γ)
by a result of Nica (see [28], page 3309). Recall that a group is virtually nilpotent if it has a finite index
subgroup which is nilpotent. Every virtually nilpotent group has polynomial, and hence subexponential,
growth. So if Γ is virtually nilpotent then α ∈ C(Γ) is invertible in L(Γ) if and only if it is invertible
in ℓ1 (Γ). The situation is very different when Γ does not have subexponential growth. For example, if
Γ contains a free subsemigroup on two letters, then there are elements α ∈ Z(Γ) which are invertible in
L(Γ), but not in ℓ1 (Γ). For example, if g, h generate a free subsemigroup in Γ, then
±3e − (e + g + g 2 )h
is such an element (see Appendix A of [23] for a detailed argument). If Γ is a finitely-generated, elementary amenable, not virtually nilpotent group, then a result of Chou say that Γ contains a nonabelian free
subsemigroup (see [4]). Additionally Frey in [14] showed that if Γ is an amenable group which contains a
nonamenable subsemigroup, then it contains a nonabelian free group. For a concrete instance of Chou’s
result consider the group R ⋊ (R \ {0}) which is R × (R \ {0}) as set but with operation
(a, b)(c, d) = (a + bc, bd).
If 0 ≤ a ≤ 1/2, the subsemigroup generated by (1, a), (1 − a) is a free nonabelian semigroup.
For our purposes, it will be important to use ℓ2 − Λ(σi ) − IE −k-tuples. Following the methods in
our proof of Theorem 4.4 of [15], given an inclusion B ⊆ A of Z(Γ)-modules, we will want a notion of
17
[ ⊆ A.
b The use of Λ(σ ) − independence tuples for
Λ − (σi )i -IE-tuples corresponding to the inclusion A/B
i i
inclusions will ease extending Theorem 4.1 to the case when f is only invertible in Mn (L(Γ)).
We will need to recall some notation from [15], as the perturbative techniques there will remain to
be important here. For x ∈ R, we use
|x + Z| = inf |x − l|.
l∈Z
Thus |θ| makes sense for any θ ∈ R/Z. Let us recall a definition from [16].
Definition 4.3. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
B ⊆ A be countable Z(Γ)-modules, and let ρ be a continuous dynamically generating pseudometric on
b For finite F ⊆ Γ, D ⊆ B and δ > 0 we let Map(ρ|D, F, δ, σi ) be all φ ∈ Map(ρ, F, δ, σi ) so that
A.
max
b∈D
di
1 X
|φ(j)(b)|2 < δ 2 .
di j=1
The main point of this definition is that it is shown in Proposition 4.3 of [16] that an element of
φ ∈ Map(ρ|D, F, δ, σi ) is close to a map
[
φe : {1, . . . , di } → A/B
which is in Map(ρ, F ′ , δ ′ , σi ) with δ ′ → 0 and F ′ increasing to Γ as δ → 0, and F increases to Γ, and
D increases to B. So Map(ρ A/B
[ , . . . ) and Map(ρ|D, . . . ) are asymptotically the same notion. A crucial
defect of the argument in [16] is that the proof of existence of φe is nonconstructive, using a compactness
argument in an essential way. However, due to its nonconstructive nature it allows one to create more
elements in Map(ρ, F, δ, σi ) than one would initially believe exist. This will be precisely the use here.
We need a similar perturbative idea specifically related to the case of Xf for f ∈ Mm,n (Z(Γ)). Fix a
countable discrete sofic group Γ with sofic approximation σi : Γ → Sdi . For x ∈ ℓ2R (di , udi )⊕n , define
1/2
n
X
kx(j) − l(j)k2ℓ2 (di ,ud ) .
kxk2,(Zdi )⊕n = inf
l∈(Zdi )⊕n
i
j=1
For f ∈ Mm,n (Z(Γ)), we let
Ξδ (σi (f )) = {ξ ∈ (Rdi )⊕n : kσi (f )ξk2,(Zdi )⊕m < δ}.
Definition 4.4. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
B ⊆ A be countable Z(Γ)-modules. Let ρ be a continuous dynamically generating pseudometric for
k
[ and 1 ≤ q ≤ ∞. For finite K, F ⊆ Γ, D ⊆ B, E ⊆ C(Γ) and
b and 1 ≤ p ≤ ∞. Fix x ∈ A/B
Γ y A,
η, δ > 0 we say that a sequence Ji ⊆ {1, . . . , di } is a ℓq − Λ − (x, ρ|D, F, δ, E, η, (σi )i ; ε, K)-independence
sequence if (Ji )i≥1 ∈ Λ(σi ) and for all c : Ji → {1, . . . , k} so that c−1 ({l}) ∈ Λ(E, η, σi ) there is a
φ ∈ Map(ρ|D, F, δ, σi ) so that
max ρq,Ji (gφ(·), gxc(·) ) < ε.
g∈K
We let IΛ,q (x, ρ|D, F, δ, E, η, (σi ); ε, K, B ⊆ A) be the supremum of
lim sup udi (Ji )
i→∞
q
over all ℓ − Λ − (x, ρ|D, F, δ, E, η, (σi )i ; ε, K)-independence sequences (Ji )i≥1 . Set
IΛ,q (x, ρ|D, F, δ, (σi )i ; ε, K, B ⊆ A) =
sup
IΛ,q (x, ρ|D, F, δ, E, η, (σi )i ; ε, K),
finiteE⊆C(Γ),
η>0
IΛ,q (x, ρ, (σi )i ; ε, K, B ⊆ A) =
inf
finiteD⊆B,
finiteF ⊆Γ,
δ>0
IΛ,q (x, ρ|D, F, δ(σi ); ε, K, B ⊆ A).
We say that x is a ℓq − Λ(σi ) − IE −k−tuple for B ⊆ A if for all ε > 0 and for all K ⊆ Γ finite we have
Iq (x, ρ, (σi )i ; ε, K, B ⊆ A) > 0.
q
We let IEΛ,k
(σi )i (ρ, q, B ⊆ A) be the set of all ℓ − Λ(σi )i − IE −k-tuples for B ⊆ A.
18
Definition 4.5. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi , let
f ∈ Mm,n (Z(Γ)). Fix x ∈ Xfk and 1 ≤ q ≤ ∞. For finite K ⊆ Γ, E ⊆ C(Γ), and δ, η, ε > 0 we say
that a sequence Ji ⊆ {1, . . . , di } is a ℓq − Λ(σi ) − (x, δ, E, η, (σi ); ε, K) independence sequence for f if
(Ji )i≥1 ∈ Λ(σi )i and for all c : Ji → {1, . . . , k} with c−1 ({l}) ∈ Λ(E, η, σi ) there is a ξ ∈ Ξδ (σi (f )) so that
max
g∈K
n
1 XX
|ξ(σi (g)−1 (j))(l) − xc(j) (l)(g)|2 < ε2 .
|Ji |
j∈Ji l=1
f
We let IΛ,q
(x, δ, E, η, (σi ); ε, K) be the supremum of
lim sup udi (Ji ),
i→∞
where Ji is a ℓq − Λ(σi )i − (x, δ, E, η, (σi ); ε, K) independence sequence. We set
f
IΛ,q
(x, δ, (σi )i ; ε, K) =
sup
finiteE⊆C(Γ),
η>0
f
IΛ,q
(x, δ, E, η, (σi )i ; ε, K)
f
f
IΛ,q
(x, (σi )i ; ε, K) = inf IΛ,q
(x, δ, (σi )i ; ε, K).
δ>0
q
We say that x is a ℓ − Λ(σi )i − IE −k-tuple for f if for all ε > 0 and K ⊆ Γ finite we have
f
IΛ,q
(x, (σi )i ; ε, K) > 0.
q
We use IEΛ,k
(σi ) (q, f ) for the set of all ℓ − Λ(σi )i − IE −k-tuples for f.
Proposition 4.6. Let Γ be a countable discrete sofic group with sofic approximation Σ.
(a): Let B ⊆ A be countable Z(Γ)-modules. Then for 1 ≤ q < ∞, the set of q − Λ(σi ) -IE-tuples for
[
B ⊆ A, is the same as the set of ℓq − Λ(σi )i -IE-tuples for Γ y A/B.
(b): If f ∈ Mm,n (Z(Γ)) then the set of Λ(σi )i -IE-tuples for Γ y Xf , is the same as the set of
Λ(σi )i -IE-tuples for f.
b Use the pseudoProof. (a): Fix k ∈ N, and a continuous dynamically generating pseudometric ρ on A.
[
metric ρ A/B×
[ A/B
[ on A/B. It is clear that
Λ,k
[
IEΛ,k
(σi ) (ρ, q, B ⊆ A) ⊇ IE(σi ) (q, A/B, Γ).
For the reserve inclusion let
x ∈ IEΛ,k
(σi )i (ρ, q, B ⊆ A).
Fix finite K, F ⊆ Γ, E ⊆ C(Γ) and δ, η, ε > 0. Set
α = IΛ,q (x, ρ, (σi )i ; ε, K, B ⊆ A).
Choose finite F ′ ⊆ Γ, D′ ⊆ B, δ ′ > 0 in a manner depending upon F, δ, η to be determined later. Let
(Ji )i≥1 be a ℓq − Λ(σi )i − (x, ρ|D, E, η, F, δ, (σi )i ; ε, K) independence set with
lim sup udi (Ji ) ≥
i→∞
α
.
2
Suppose we are given
c : Ji → {1, . . . , k}
with
c−1 ({l}) ∈ Λ(E, η, σi ).
Choose a φ ∈ Map(ρ|D′ , F ′ , δ ′ , σi ) with
max ρq,Ji (gφ(·), gxc(·) ) < ε.
g∈K
19
[ so that
Arguing as in Proposition 4.3 in [16], we may find a φe : {1, . . . , di } → A/B
e
e
≤ κq (D′ , F ′ , δ ′ ),
ρ2,di (gφ(·), g φ(·)))
max max(ρq,di (gφ(·), g φ(·)),
g∈K∪F
with
lim
(D′ ,F ′ ,δ ′ )
κq (D′ , F ′ , δ ′ ) = 0.
Here (D1 , F1 , δ1 ) ≤ (D2 , F2 , δ2 ) if D1 ⊆ D2 , F1 ⊆ F2 and δ1 ≥ δ2 . Thus
φe ∈ Map(ρ
and
[ A/B
[ , F, δ
A/B×
′
+ κq (D′ , F ′ , δ ′ ), σi )
e gxc(·) ) < ε + ud (Ji )−1 κq (D′ , F ′ , δ ′ ).
max ρq,Ji (g φ(·),
i
g∈K∪F
Choose D′ , F ′ , δ ′ so that
κq (D′ , F ′ , δ ′ ) + δ ′ < δ
1
< ε.
α
Since α does not depend upon D′ , F ′ , δ ′ this is possible. We then see that
κq (D′ , F ′ , δ ′ )
φe ∈ Map(ρ
and since
[ A/B
[ , F, δ, σi )
A/B×
lim sup udi (Ji ) ≥ α/2
i→∞
we have
e gxc(·) ) < 5ε
max ρq,Ji (g φ(·),
g∈K
for all large i.
(b) View Xf ⊆ (TΓ )⊕n , and let ρ be the dynamically generating pseudometric on (TΓ )⊕n defined by
ρ(θ1 , θ2 )2 =
n
X
|θ1 (l)(e) − θ2 (l)(e)|2 ,
l=1
where | · | on T is in the sense defined in the remarks preceding Definition 4.3. Given ζ ∈ (Tdi )⊕n we can
define
φζ : {1, . . . , di } → (TΓ )⊕n
by
φζ (j)(l)(g) = ζ(σi (g)−1 (j))(l).
Note that for any δ ′ > 0 and finite F ′ ⊆ Γ we have
φζ ∈ Map(ρ, F ′ , δ ′ , σi )
for all large i. Indeed for any h ∈ Γ
ρ(hφζ (·), φζ ◦ σi (h))2 =
n di
1 XX
|ζ(σi (h−1 )−1 (j))(l) − ζ(σi (e)−1 σi (h)(j))(l)|2
di
j=1
l=1
≤ nudi ({j : σi (h−1 )−1 (j) 6= σi (e)−1 σi (h)(j)})
→ 0,
the last line following as (σi )i is a sofic approximation. Given D′ ⊆ Z(Γ)⊕m f, and ξ ∈ Ξδ (σi (f )) by the
proof of Proposition 3.6 in [15], we have that
di
1 X
max
|hφξ+(Zdi )⊕n (j), bi|2 < κ(δ)
b∈D di
j=1
20
with
lim κ(δ) = 0.
δ→0
Thus φξ+(Zdi )⊕n ∈ Map(ρ|D′ , F ′ , δ ′ , σi ) if δ is sufficiently small and i is sufficiently large. From this it is
not hard to argue as in (a) that
Λ,k
⊕m
IEΛ,k
f ⊆ Z(Γ)⊕n ).
(σi ) (q, f ) ⊆ IE(σi )i (q, Z(Γ)
Conversely, suppose we have a finite F ′ ⊆ Γ, a δ ′ > 0, and a finite D′ ⊆ Γ. Given φ ∈ Map(ρ|D′ , F ′ , δ ′ , σi )
we may define
ζφ ∈ (Tdi )⊕n
by
ζφ (l)(j) = φ(j)(l)(e).
di ⊕n
Let ξφ ∈ (R )
be any element such that
ξφ + (Zdi )⊕n = ζφ .
Then by the proof of Proposition 3.6 in [15],
ξφ ∈ Ξκ(D′ ,F ′ ,δ′ ) (σi (f ))
with
lim
(D′ ,F ′ ,δ ′ )
κ(D′ , F ′ , δ ′ ) = 0.
Here the triples (D′ , F ′ , δ ′ ) are ordered as in part (a). Again we can use this to argue as in (a) that
⊕m
IEΛ,k
f ⊆ Z(Γ)⊕n ) ⊆ IEΛ,k
(σi )i (q, Z(Γ)
(σi )i (q, f ).
We will use the above Lemma to show that
k
IEΛ,k
(σi )i (2, Xf , Γ) = Xf ,
when f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). By Proposition 3.5, this will prove that every k-tuples of points in
Xf is a (σi )i -IE-tuple. We first need a way of constructing elements of Λ(σi ) whose translates by a given
finite subset of Γ are disjoint. For this we use the following Lemma.
Lemma 4.7. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Fix a finite
symmetric subset E ⊂ Γ containing the identity. Then, there is a sequence (Ji )i≥1 ∈ Λ(σi )i so that
σi (x)Ji ∩ Ji = ∅ for all x ∈ E \ {e}
lim udi (Ji ) =
i→∞
1
|E|
|E|
.
Proof. Consider the Bernoulli shift action Γ y (E, uE )Γ . Let
J = {x ∈ E Γ : x(g) = g for all g ∈ E}.
Suppose
x∈J
and g ∈ E \ {e}, then
(g −1 x)(e) = x(g) = g 6= e
so x ∈
/ gJ ∩ J. Thus gJ, J are disjoint for all g ∈ E \ {e}. We now use the fact that Bernoulli shifts have
positive sofic entropy to model this behavior on {1, . . . , di }.
First note that for every ε > 0, there is a δ > 0 so that if Ei ⊆ {1, . . . , di } has
udi (σi (x)Ei ∩ Ei ) ≤ δ, for all x ∈ E \ {e}
21
then there is a Ei′ ⊆ Ei with
udi (Ei \ Ei′ ) ≤ ε
and
σi (x)Ei′ ∩ Ei′ = ∅, x ∈ E \ {e}.
Indeed, this is simply proved by setting
\
Ei′ =
Ei \ σi (x)−1 (Ei ).
x∈E\{e}
Using the fact that h(σi ),u⊗Γ (E Γ , Γ) ≥ 0, we may find a sequence Ai,g ⊆ {1, . . . , di }, g ∈ E so that
E
!
k
\
−αl
⊗Γ
αk
α1
Γ
udi (σi (g1 ) Ai,h1 ∩ · · · ∩ σi (gk ) Ai,hk ) → uE
{x ∈ E : x(gl ) = hk }
(5)
l=1
for all k ∈ N, h1 , . . . , hk ∈ E, g1 , . . . , gk ∈ Γ and α1 , . . . , αk ∈ {1, −1} (see e.g. Bowen’s original definition
of sofic entropy in [2]). Set
\
Ji′ =
σi (g)−1 Ai,g ,
g∈E
then by (5),
udi (σi (g)Ji′ ∩ Ji′ ) → 0
for all g ∈ E \ {e} and
udi (σi (x)Ji′ ∩ Ji′ ) → u⊗Γ
E (xJ ∩ J)
for all x ∈ Γ. Applying our previous observation we find Ji ⊆ Ji′ so that
udi (Ji′ \ Ji ) → 0
and
σi (g)Ji ∩ Ji = ∅ for g ∈ E \ {e}.
Since
udi (Ji′
\ Ji ) → 0, we have
⊗Γ
2
udi (σi (g)Ji ∩ Ji ) − udi (Ji )2 → u⊗Γ
E (gJ ∩ J) − uE (J)
(6)
for all g ∈ Γ. It is well-known that Γ y (L2 ((E, uE )Γ )⊖C1) can be equivariantly, isometrically embedded
in Γ y ℓ2 (N × Γ) where the action of Γ y ℓ2 (N × Γ) is given by
(gξ)(n, h) = ξ(n, g −1 h).
We will use this to show that (Ji )i≥1 ∈ Λ(σi )i . We again use o(1) for any expression which goes to zero
as i → ∞. Let α : Γ → U(L2 ((E, uE )Γ ) ⊖ C1) be the representation
(α(g)ξ)(ω) = ξ(g −1 ω), ω ∈ E Γ , g ∈ Γ.
Extend by linearity to a ∗-representation
α : C(Γ) → B(L2 (E, uE )Γ ⊖ C1).
Then for any f ∈ C(Γ) and i ∈ N we have
X
kσi (f )(χJi − udi (Ji )1)k22 =
fb(g)fb(h)hσi (g)(χJi − udi (Ji )1), σi (h)(χJi − udi (Ji )1)i
g,h∈Γ
= o(1) +
X
g,h∈Γ
= o(1) +
X
g∈Γ
= o(1) +
X
g∈Γ
= o(1) +
X
g∈Γ
= o(1) +
X
g∈Γ
fb(g)fb(h)hσi (h−1 g)(χJi − udi (Ji )1), χJi − udi (Ji )1i
∗ f (g)hσ (g)(χ − u (J )1), χ − u (J )1i
fd
di i
Ji
di i
i
Ji
∗ f (g)(u (σ (g)J ∩ J ) − u (J )2 )
fd
i
i
di i
di i
∗ f (g)(u⊗Γ (gJ ∩ J) − u⊗Γ (J)2 )
fd
E
E
∗ f (g)hα(g)(χ − u⊗Γ (J)1), χ − u⊗Γ (J)1i
fd
J
E
E
E
2
= o(1) + kα(f )(χJ − u⊗Γ
E (J)1)k2 .
22
(7)
Since α can be embedded into the infinite direct sum of the left regular representation, we have that
⊗Γ
kα(f )(χJ − uE
(J)1)k2 ≤ kλ(f )kkχJ − u⊗Γ
E (J)1k2
=
kλ(f )k(u⊗Γ
E (J)
−
(8)
2 1/2
u⊗Γ
E (J) )
= o(1) + kλ(f )k(udi (Ji ) − udi (Ji )2 )1/2
= o(1) + kλ(f )kkχJi − udi (Ji )1k2 .
By (7), (8) we have that (Ji )i ∈ Λ(σi )i . From our construction it also follows that
udi (Ji ) → u⊗Γ
E (J) =
1
|E|
|E|
.
For the next Lemma, we need some notation. We use T = R/Z. Define t : C(Γ) → C by
X
t(α) =
α
b(g).
g∈Γ
Lemma 4.8. Let Γ be a countable discrete group, and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). Let φ be the inverse
of f in Mn (L(Γ)). Define Q : ℓ2R (Γ)⊕n → (TΓ )⊕n by
Q(ξ)(l)(g) = ξ(l)(g) + Z.
Then Q({αφ∗ : α ∈ Z(Γ)⊕n , t(α(j)) = 0, 1 ≤ j ≤ n}) is dense in Xf .
Proof. As usual, we view Z(Γ)⊕n ⊆ ℓ2 (Γ)⊕n . For α, β ∈ Z(Γ)⊕n we have
hα, βi =
n
X
τ (β(l)∗ α(l))
l=1
where τ is the trace on L(Γ). For θ ∈ (TΓ )⊕n , α ∈ Z(Γ)⊕n we set
hθ, αiT =
n X
X
l=1 g∈Γ
d
θ(l)(g)α(l)(g)
∈ T.
Then the pairing h·, ·iT allows us to identify (TΓ )⊕n ∼
= (Z(Γ)⊕n )b.
By Pontryagin duality, it suffices to show that if β ∈ Z(Γ)⊕n has
hβ, αφ∗ i ∈ Z
for all α ∈ Z(Γ)⊕n with t(α(l)) = 0, 1 ≤ l ≤ n, then β ∈ Z(Γ)⊕n f. For x ∈ L(Γ), and 1 ≤ l ≤ n we use
x ⊗ el ∈ L(Γ)⊕n which is x in the lth coordinate and 0 in every other coordinate. Fix 1 ≤ l ≤ n and
consider α = (g − 1) ⊗ el . Then
\
\
hβ, αφ∗ i = hβφ, αi = (βφ)(l)(g)
− (βφ)(l)(e).
So
\
\
(βφ)(l)(g)
− (βφ)(l)(e)
∈Z
\ ∈ ℓ2 we find that
for all g ∈ Γ. Letting g → ∞, and using that (βφ)(l)
\
(βφ)(l)(e)
∈ Z.
As
\
\
(βφ)(l)(g)
− (βφ)(l)(e)
∈Z
for all g ∈ Γ, 1 ≤ l ≤ n we find that βφ ∈ Z(Γ)⊕n . Thus
β = (βφ)f ∈ Z(Γ)⊕n f.
23
We are now ready to prove our Theorem, but we first recall the notation we introduced at the
beginning of this section. From the identifications
Mm,n (C(Γ)) ⊆ Mm,n (L(Γ)) ⊆ B(ℓ2 (Γ)⊕n , ℓ2 (Γ)⊕m )
we may think of elements of Mm,n (C(Γ)) as bounded, linear operators ℓ2 (Γ)⊕n → ℓ2 (Γ)⊕m . For a fixed
x ∈ Mm,n (C(Γ)) we let kxk∞ be the norm of x as an operator ℓ2 (Γ)⊕n → ℓ2 (Γ)⊕m under the above
identification. We also identify C(Γ)⊕n ∼
= M1,n (C(Γ)) and use the notation above. We thus caution the
reader that for A ∈ Mm,n (C(Γ))
|d
Aij (g)|,
kAk∞ 6=
sup
g∈Γ,
1≤i≤m,1≤j≤n
with similar remarks for elements of C(Γ)⊕n .
Theorem 4.9. Let Γ be a countable discrete group with sofic approximation σi : Γ → Sdi be a sofic
approximation. Let f ∈ Mn (Z(Γ))∩GLn (L(Γ)), then every k-tuple of points in Xf is a ℓ2 −Λ(σi )i −IE −ktuple.
Proof. Let φ be the inverse of f in Mn (L(Γ)). By the preceding lemma and Proposition 3.6, it suffices
to prove the theorem when (x1 , . . . , xk ) = (Q(α1 φ∗ ), . . . , Q(αk φ∗ )) where t(αj ) = 0. For t > 0, let
φt ∈ Mn (R(Γ)) be such that
kφt − φk∞ < t.
Fix ε > 0, and a A ⊆ Γ finite. Suppose we are given a finite F ⊆ Γ, and a δ > 0. Let E ⊆ C(Γ) be
finite and η > 0 to depend upon F, δ in a manner to be determined later. Let
L1 = (supp(φε ) ∪ {e} ∪ supp(φε )−1 ),
L2,s = (supp(αs ) ∪ {e} ∪ supp(αs )−1 ) for 1 ≤ s ≤ k,
(2015)!
[
K1 =
L2,s (supp(f ) ∪ {e} ∪ supp(f )−1 )L1
,
1≤s≤k
K2 =
[
1≤s≤k
(2015)!
((A ∪ {e} ∪ A−1 )(supp(φε ) ∪ {e} ∪ supp(φε )−1 )(supp(αs ) ∪ {e} ∪ supp(αs )−1 ))
,
K = K1 ∪ K2 .
Apply Lemma 4.7 to find a sequence Ji ⊆ {1, . . . , di } so that {σi (x)Ji }x∈K are a disjoint family, and
(Ji )i ∈ Λ(σi ) and
|K|
|Ji |
1
.
=
lim
i→∞ di
|K|
Note that if Ji′ ⊆ Ji satisfies
udi (Ji \ Ji′ ) → 0,
then Ji′ enjoys the conclusions of Lemma 4.7 as well. So by soficity, we may assume
σi (x)(j) 6= σi (y)(j)
for x 6= y ∈ K, j ∈ Ji and that
σi (x1 . . . xl )(j) = σi (x1 ) . . . σi (xl )(j)
for x1 , . . . , xl ∈ K and 1 ≤ l ≤ (2015)!. Let c : Ji → {1, . . . , k} be such that c−1 ({s}) ∈ Λ(E, η, σi ). Set
(s)
Ji
For t ∈ (0, ∞) let
ξt =
= c−1 ({s}).
X
σi (φt α∗s )χJ (s) .
i
1≤s≤k
Note that
24
σi (f )ξδ −
X
X
σi (α∗s )χJ (s) =
i
1≤s≤k
(σi (f )σi (φδ α∗s ) − σi (α∗s ))χJ (s) .
i
1≤s≤k
For β ∈ C(Γ) we have
σi (β)1 = t(β)1.
Because t(αs ) = 0 for 1 ≤ s ≤ k,
X
X
σi (f )ξδ −
σi (α∗s )χJ (s) =
(σi (f )σi (φδ α∗s ) − σi (α∗s ))(χJ (s) − udi (χJ (s) )1).
i
1≤s≤k
i
1≤s≤k
Since
(9)
i
(s)
kχJ (s) − udi (Ji )1k∞ ≤ 2,
i
we have
(s)
(s)
kσi (f )σi (φδ α∗s )(χJ (s) − udi (Ji )1) − σi (f φδ α∗s )(χJ (s) − udi (Ji )1)k2 →i→∞ 0.
i
i
Thus
X
1≤s≤k
(s)
(s)
σi (f )σi (φδ α∗s )(χJ (s) − udi (Ji )1) − σi (f φδ α∗s ))(χJ (s) − udi (Ji )1)
→i→∞ 0.
i
i
We have
(10)
2
(s)
(s)
kσi (f φδ α∗s )(χJ (s) − udi (Ji )1) − σi (α∗s )(χJ (s) − udi (Ji )1)k2 =
i
i
2 1/2
n
n
X
X
(s)
(σi ((f φδ α∗s − α∗s )lp )(χJ (s) − udi (Ji )1) .
l=1
If
i
p=1
2
E ⊇ {(f φδ α∗s − α∗s )lp : 1 ≤ l, p ≤ n}
then as
kf φδ α∗s − α∗s k∞ ≤ δkf k∞ kαs k∞ ,
we have
(s)
(s)
kσi (f φδ α∗s )(χJ (s) − udi (Ji )1) − σi (α∗s )(χJ (s) − udi (Ji )1)k2
i
i
n
X
≤
kαs k∞ kf k∞ nδkχJ (s) −
i
l=1
(s)
udi (Ji 1)k2
+ nη
≤ kαs k∞ kf k∞ n2 δkχJ (s) k2 + n2 η.
2
!1/2
i
Set
M = (kf k∞ + 1)
then
X
1≤s≤k
X
1≤s≤k
1/2
kαs k2∞
(s)
(s)
,
≤ kn2 η
σi (f φδ α∗s )(χJ (s) − udi (Ji )1) − σi (α∗s )(χJ (s) − udi (Ji )1)
(11)
i
i
2
+ n2 δkf k∞
X
kαs k∞ kχJ (s) k2
1≤s≤k
≤ kn2 η + M n2 δudi (Ji )1/2 ,
where in the last step we use the Cauchy-Schwartz inequality and the fact that
X
kχJ (s) k22 = udi (Ji ).
1≤s≤k
i
25
i
If we force η sufficiently small then by (9), (10), (11) we have for all large i,
ξδ ∈ Ξδ(n2 +1)M (σi (f )).
(12)
We will want to force η to be even smaller later.
If E ⊇ {(φε α∗s − φδ α∗s )pl : 1 ≤ l, p ≤ n}, then for all 1 ≤ s ≤ k, for all 1 ≤ p, l ≤ n
(s)
(s)
(s)
kσi ((φε α∗s −φδ α∗s )pl )(χJ (s) −udi (Ji )1)k2 ≤ 2εkαs k∞ kχJ (s) −udi (Ji )1)k2 +η ≤ 2εkαs k∞ udi (Ji )1/2 +η.
i
i
Note that in our defintion of ℓ2 − Λ(σi )i -tuples we are allowed to have E, η depend upon δ. By the same
arguments as before
X
(s)
kξε − ξδ k2,Ji ≤ n2 η + 2εn2
kαs k∞ udi (Ji )1/2 ≤ n2 η + 2εM udi (Ji )1/2
1≤s≤k
where again we have used the Cauchy-Schwartz inequality and the fact that
X
kχJ (s) k22 = udi (Ji ).
1≤s≤k
i
Thus
1 X
1 X
|ξε (j)−ξδ (j)|2 ≤ (udi (Ji )−1 n4 η 2 +2εM n2 ηudi (Ji )−1/2 +4ε2 M 2 ).
|ξε (j)−ξδ (j)|2 = udi (Ji )−1
|Ji |
di
j∈Ji
j∈Ji
For all large i,
|K|
1
1
.
udi (Ji ) ≥
2 |K|
So we can choose η sufficiently small (depending only upon K) so that
kξδ − ξε k2,Ji < ε(2M + 1).
Then
1/2
X
1
2
≤ kσi (g)ξδ − σi (g)ξε k2,Ji < ε(2M + 1).
|[σi (g)ξδ ](j) − [σi (g)ξε ](j) + Z|
|Ji |
(13)
j∈Ji
(s)
Additionally, for j ∈ Ji , g ∈ A
σi (g)ξε (j) =
X X
x∈Γ 1≤s≤k
=
X
x∈K∩g−1 K
∗
φ[
ε αs (x)χσi (g)σi (x)J (s) (j)
i
X
∗
φ[
ε αs (x)χσi (g)σi (x)J (s) (j),
i
1≤s≤k
here we use our choice of Ji as well as the fact that K ∩ g −1 K ⊇ supp(φε α∗s ). As {σi (k)Ji }k∈K are a
disjoint family, we have for x ∈ K ∩ g −1 K that χσi (gx)J (s) (j) = 1 if and only if gx = e, and thus when
i
x = g −1 . Since K ∩ g −1 K ⊇ supp(φε α∗s ), the above sum is
∗
∗ −1 ) = α
[
φ[
ε αs (g
s φε (g).
As
∗
∗
∗
[
[∗
|α
s φε (g) − αs φ (g)| ≤ kαs φε − αs φ k2 ≤ εkαs k2 ≤ εkαs k∞ .
We find that
Combining with (13)
1/2
X
1
2
∗
max
< εM.
|(σi (g)ξε )(j) + Z − α\
c(j) φ (g)|
g∈A
|Ji |
j∈Ji
1/2
X
1
2
∗
max
|(σi (g)ξδ )(j) + Z − α\
< ε(3M + 1).
c(j) φ (g)|
g∈A
|Ji |
j∈Ji
As ε > 0 is arbitrary, the Theorem is now proved using Proposition 4.6.
26
Corollary 4.10. Let Γ be a countable discrete sofic group with sofic approximation σi : Γ → Sdi . Let
f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). Then every k-tuple of points in Xf is a (σi )i − IE −k-tuple.
Proof. Automatic from the preceding Theorem and Proposition 3.5.
Corollary 4.11. Let Γ be a countable discrete sofic group and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). If f is not
in GLn (Z(Γ)), then detL(Γ) (f ) > 1.
Proof. Observe that Xf is not a single point if and only if f ∈
/ GLn (Z(Γ)). The corollary is then automatic
from the preceding Corollary, Proposition 4.16 (3) in [22] and Theorem 4.4 in [15].
In fact, we have the following more general result. Recall that if Γ is sofic, if σi : Γ → Sdi is
a sofic approximation, an action Γ y X on a compact metrizable space is said to have completely
positive topological entropy with respect to (σi )i if whenever Γ y Y is a nontrivial (i.e. not a onepoint space) topological factor of X, we have h(σi )i (Y, Γ) > 0. The following Corollary was known for
f ∈ Mn (Z(Γ)) ∩ GLn (ℓ1 (Γ)), by Proposition 4.16 (3),(5) and Theorem 6.7 of [22].
Corollary 4.12. Let Γ be a countable discrete sofic group and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). Suppose
that f is not in GLn (Z(Γ)). Then for any sofic approximation σi : Γ → Sdi , the action Γ y Xf has
completely positive topological entropy with respect to (σi )i .
Proof. Automatic from Theorem 4.9, Proposition 3.5, Proposition 4.16 (3) in [22] and Proposition 3.6.
Combining with results of Chung-Li we have the following result in the amenable case, which previously only known for f ∈ GLn (ℓ1 (Γ)) (see Corollary 8.4, Theorem 7.8 and Lemma 5.4 of [5]).
Corollary 4.13. Let Γ be a countable amenable group, and f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)). Suppose that
f is not in GLn (Z(Γ)). Then the action Γ y Xf has completely positive measure-theoretic entropy (with
respect to the Haar measure on Xf ).
Proof. This follows from Corollary 8.4 of [5] and Corollary 4.10.
The preceding corollary was known in the amenable case when f ∈ Mn (Z(Γ)) ∩ GLn (ℓ1 (Γ)) by
combining Proposition 4.16 (3),(5) and Theorem 6.7 of [22] with Corollary 8.4 of [5]. As we mentioned,
at the beginning of this section there are interesting examples in the amenable case of f ∈ Z(Γ) ∩ L(Γ)×
but f ∈
/ ℓ1 (Γ)× . When Γ is sofic, it would be interesting to decide if Γ y Xf has completely positive
measure-theoretic entropy with respect to every sofic approximation if f ∈ Mn (Z(Γ)) ∩ GLn (L(Γ)) is
not invertible in Mn (Z(Γ)).
References
[1] B.Bekka, P. D. L. Harpe, and A. Valette. Kazhdan’s Propety (T). Cambridge University Press,
2008.
[2] L. Bowen. Measure conjugacy invariants for actions of countable sofic groups. J. Amer. Math. Soc,
23:217–245, 2010.
[3] N. Brown and N. Ozawa. C ∗ -Algebras and Finite-Dimensional Approximations. American Mathematical Society, 1994.
[4] C. Chou. Elementary amenable groups. Illinois J. Math., 24(3):396–407, 1980.
[5] N. Chung and H. Li. Homoclinc group, IE group, and expansive algebraic actions. Invent. Math to
appear.
[6] L. Ciobanu, D. Holt, and S. Rees. Sofic groups: graph products and graphs of groups. Pacific
Journal of Mathematics, 271(1):53–64, November 2014.
[7] C. Deninger. Fuglede-Kadison determinants and entropy for actions of discrete amenable groups.
J. Amer. Math. Soc., 19:737–758, 2006.
27
[8] C. Deninger. Mahler measures and Fuglede-Kadison determinants. Münster Journal of Mathematic,
2:45–63, 2009.
[9] C. Deninger and K. Schmidt. Expansive algebraic actions of discrete residually finite amenable
groups and their entropy. Ergodic Theory and Dynamically Systems, 27:769–786, 2007.
[10] K. Dykema, D. Kerr, and M. Pichot. Orbit equivalence and sofic approximation. arXiv:1102.2556.
[11] K. Dykema, D. Kerr, and M. Pichot. Sofic dimension for discrete measurable groupoids. Trans.
Amer. Math. Soc, 366(2):707–748, 2013.
[12] G. Elek and G. Lippner. Sofic equivalence relations. Journal of Functional Analysis, 258:1692–1708,
2010.
[13] G. Elek and E. Szabo. On sofic groups. Journal of Group Theory, 9(2):161–171, 2006.
[14] A. Frey. Studies in amenable semigroups. PhD thesis, University of Washington, Seattle, Washington, 1960.
[15] B. Hayes. Fuglede-Kadison determinants and sofic entropy. arXiv:1402.1135.
[16] B. Hayes. Metric mean dimension for algebraic actions of sofic groups. To appear in Tran. of the
Amer. Math. Soc.
[17] B. Hayes. Polish models and sofic entropy. J. Inst. Math. Jussieu to appear.
[18] M. G. Karpovsky and V. D. Milman. Coordinate density of sets of vectors. Discrete Math, 24(2):177–
184, 1978.
[19] D. Kerr and H.Li. Soficity, amenability, and dynamical entropy. Amer. J. Math, 135(3):721–761,
2013.
[20] D. Kerr and H. Li. Independence in topological and C*-dynamics. Math. Ann., 338:869–926, 2007.
[21] D. Kerr and H. Li. Topological entropy and the variational principle for actions of sofic groups.
Invent. Math, 186:501–558, 2011.
[22] D. Kerr and H. Li. Combinatorial independence and sofic entropy. Comm. Math. Stat., 1(2):213–257,
2014.
[23] H. Li. Compact group automorphisms, addition formulas and Fuglede-Kadison determinants. Ann.
of Math., 176(1):303–347, 2012.
[24] H. Li and B.Liang. Mean dimension, mean rank, and von Neumann-Lück rank. to appear in J.
Reine. Angew. Math., 2013.
[25] H. Li and A. Thom. Entropy, determinants, and ℓ2 -torsion. J. Amer. Math. Soc., 27(1):239–292,
2014.
[26] D. Lind, K. Schmidt, and T. Ward. Mahler measure and entropy for commuting automorphisms of
compact groups. Invent. Math., 101:593–629, 1990.
[27] W. Lück. L2 -Invariants: Theory and Applications to Geometry and K-theory. Springer-Verlag,
Berlin, 2002.
[28] B. Nica. Relatively spectral morphisms and applications to k-theory. J. Funct. Anal., 255(12):33303–
3328, 2008.
[29] L. Paunescu. On sofic actions and equivalence relations.
261(9):2461–2485, November 2011.
Journal of Functional Analysis,
[30] S. Popa. Independence properties in sublagebras of ultraproduct II1 factors. Journal of Functional
Analysis, 266(9):5818–5846, 2014.
[31] N. Sauer. On the density of families of sets. J. Combinatorial Theory Ser. A, 13:145–147, 1972.
28
[32] K. Schmidt. Dynamical Systems of Algebraic Origin, volume 128 of Progress in Mathematics.
Birkhäuser Verlag, Basel, 1995.
[33] S. Shelah. A combinatorial problem; stability and order for models and theories in infinitary languages. Pacific J. Math, 41:247–261, 1972.
29
| 4math.GR
|
Grounding Visual Explanations (Extended Abstract)
arXiv:1711.06465v1 [cs.CV] 17 Nov 2017
Lisa Anne Hendricks1 , Ronghang Hu1 , Trevor Darrell1 and Zeynep Akata2
1
EECS, University of California Berkeley
2
Amsterdam Machine Learning Lab, University of Amsterdam
Abstract
Existing models [2] which generate textual explanations enforce task relevance
through a discriminative term loss function, but such mechanisms only weakly
constrain mentioned object parts to actually be present in the image. In this paper, a
new model is proposed for generating explanations by utilizing localized grounding
of constituent phrases in generated explanations to ensure image relevance. Specifically, we introduce a phrase-critic model to refine (re-score/re-rank) generated
candidate explanations and employ a relative-attribute inspired ranking loss using
‘flipped’ phrases as negative examples for training. At test time, our phrase-critic
model takes an image and a candidate explanation as input and outputs a score
indicating how well the candidate explanation is grounded in the image.
1
Introduction
Modern neural networks are good at localizing objects, predicting object categories and describing
scenes with natural language. However, the decision processes of neural networks are often opaque.
Therefore, in order to interpret and monitor neural networks, providing explanations of network
decisions has gained interest. Here, we focus on providing such explanations via natural language, i.e.
textual explanations. A textual explanation system for a classification network ideally both discusses
discriminative features of its predicted classes and names image-relevant attributes. However,
sometimes these two goals are in opposition. For example, if one discriminative attribute frequently
occurs within a class, a model may learn to justify its prediction by mentioning this attribute as
a discriminative feature even if the input does not contain the attribute. In this work, we aim at
resolving such conflicts and design a framework that automatically generates textual explanations that
justify a classification decision and simultaneously ground discriminative object properties both in
the explanation and in the image via a novel a phrase-critic model. Our model significantly improves
the image-relevance of explanations in comparison to prior works.
In our framework, a phrase-critic model is first trained specifically to ground phrases, irrespective
of linguistic fluency. Fluency is ensured by training an LSTM-based explanation model to generate
candidate sentences which discuss class discriminative features. In other words, since our phrasecritic model does not focus on fluency, it should be more reliable when understanding sentence
correctness; meanwhile, our explanation generation mechanism ensures the appearance of class
discriminative information as well as fluency. As a result, we obtain more accurate and linguistically
satisfying explanations than those generated by only enforcing a discriminative training term as done
in prior work [2]. An important side effect of our method is visual explanation–a visualization of the
grounding of discriminative object parts in the image that are mentioned in the explanation.
2
Related Work
In [8], trust is regarded as a primary reason to explore explainable intelligent systems. We argue
a system which outputs discriminative features of an object class without being image relevant is
Interpretable ML Symposium, 31st Conference on Neural Information Processing Systems (NIPS 2017), Long
Beach, CA, USA.
Explanation Sampler
Explanation Grounder
This red bird has a red
beak and a black face.
red bird
Phrase-Critic
red black
beak face
+2.05
attribute
chunker
Explanation Grounder
attribute
chunker
This red bird has a
black beak and a black
face.
red bird
black black
beak face
Phrase-Critic
+1.02
Figure 1: The proposed model incorporates a phrase-critic to ensure that generated explanations are
not only class discriminative, but also image relevant. We first sample a set of explanations using the
generator from [2], then visually ground constituent phrases using a grounder based on [3]. Then,
our phrase-critic model assigns a score to each explanation which reflects whether the explanation is
image relevant. Our model encourages higher ranked sentences to be both class discriminative and
image relevant.
likely to lose the trust of users. Consequently, we seek to explicitly enforce image relevancy with our
model.
Similar to [1], we aim at providing justifications to explain which evidence is important for a decision
as opposed to introspective explanations that explains the intermediate activations of neural networks.
Recently, [2] proposed to generate natural language justifications of a fine-grained object classifier.
However, it does not ground the relevant object parts in the sentence or the image. In [6], similar
explanations are generated for activities and VQA pairs. Although an attention based explanation
system is proposed, there are no constraints to ensure the actual presence of the mentioned attributes
or entities in the image. Consequentially, these related works [2, 6], albeit generating convincing
textual explanations, do not include a process for networks to correct themselves if their textual
explanation is not well-grounded visually. In contrast, we propose a general process to first check
whether explanations are accurately aligned with image input and then improve textually explanations
by selecting a better-aligned candidate.
3
Grounded Visual Explanations
Our model consists of three main components. First, a text generation system (based on [2]) samples
many (100 in our experiments) textual explanations. [2] is trained with a discriminative loss to
encourage sentences to mention class specific attributes. Next, a phrase grounding model (based
on [3]) grounds phrases in the generated textual explanations. Finally, our proposed phrase-critic
model determines which textual explanations are preferred based on how well textual explanations are
grounded in the image. As shown in Figure 1, our system first generates possible explanations (e.g.,
“The red bird has a red beak and a black face”), determines whether constituent phrases (e.g., “red
bird”, “red beak”,“black face”) are present in the image, and then assigns a score to the explanation.
Phrase-Critic. Our phrase-critic model is the core to our framework. It takes a list of {(Ai , Ri , si )},
where Ai is the attribute phrase, Ri is the corresponding region (more precisely, visual features
extracted from the region), and si the region score, and maps them into a single image relevance score
Sr . For a given attribute phrase Ai such as “black beak”, we ground (localize) it into a corresponding
image region Ri and obtain its localization score si , using an out-of-shelf localization model from [3]
pretrained on VisualGenome. It is worth noting that the scores directly produced by the grounding
model are not comparable across images and difficult to be directly combined with other metrics,
such as sentence fluency, because these scores are difficult to normalize across different images and
different visual parts. For example, a correctly grounded phrase “yellow belly” may have a much
smaller score than the correctly grounded phrase “yellow eye” because a bird belly is less well defined
than a bird eye; for another example, an occluded image tend to score lower with all explanations
Henceforth, our phrase-critic model plays an essential role in producing normalized, utilizable and
comparable scores. More specifically, given an image I, the phrase-critic model processes the list
of {(Ai , Ri , si )} by first encoding each (Ai , Ri , si ) into a fixed-dimensional vector xenc with an
2
LSTM and then applying a two-layer neural network to regress the final score Sr which reflects the
overall image relevancy of an explanation.
Sr ({Ai }, I; θ) = fcritic ({(Ai , Ri , si )}; θ)
where fcritic consists of an LSTM encoder flstm and a small two-layer network fnn .
(1)
We construct a few explanation pairs for each image. Each explanation pair consists of a positive
explanation (image-relevant) and a negative explanation (not image-relevant). We then train our
explanation critic using the following margin-based ranking loss Lrank on each pair of positive and
negative explanations, to encourage the model to give higher scores to positive explanations than
negative explanations:
Lrank = max(0, Sr ({Aneg
}, I; θ) − Sr ({Apos
i
i }, I; θ) + M )
{Apos
i }
(2)
{Aneg
}
i
where
are matching attribute phrase whereas
are mismatching attribute phrases
respectively, therefore Srpos and Srneg are the scores of the positive and the negative explanations. We
use M = 1 in our implementation.
Flipped Attribute Training. The simplest way to sample a negative pair is to consider a mismatching
ground truth image and sentence pair. However, due to the fine-grained nature of our dataset, we
empirically found that naively sampling out-of-class negative examples can risk the negative examples
being visually too different (such as a Cardinal and an American Crow). Inspired by a relative
attribute paradigm for recognition and retrieval [5], we create negative examples by flipping attributes
corresponding to color and size in attribute phrases. For example, if a ground truth sentence mentions
a “yellow belly” and “red head” we might change the attribute phrase “red head” to “black head”.
This means the negative explanation still mentions some attributes present in the image, but is not
completely correct.
Ranking Explanations. After generating a set of candidate explanations and extracting an explanation score using our explanation model, we choose the best explanation based on the score for
each explanation. In practice, we find it is important to rank sentences based on both the relevance
score Sr and a fluency score Sf (defined as the log P (w0:T )). Including Sf is important because the
explanation scorer will rank “This is a bird with a long neck, long neck, and red beak” high (if a long
neck and red beak are present) even though mentioning “long neck” twice is clearly ungrammatical.
Grounding Visual Features. The framework for grounding visual features involves three steps:
generating visual explanations, factorizing the sentence into smaller chunks, and localizing each
chunk with a grounding model. Visual explanations are produced using the model of [2]. For each
explanation we extract a list of i attribute phrases (Ai ) using a rule-based attribute phrase chunker.
Once we have extracted attribute phrases Ai , we ground each of them to a visual region Ri in the
original image by using the baseline model proposed in [3] trained on the Visual Genome dataset
[4]. For a given attribute phrase Ai , the grounding model localizes the phrase into an image region,
returning a bounding box Ri and a score si of how likely the returned bounding box matches the
phrase. The attribute phrase, the corresponding region, and the region score form an attribute phrase
grounding (Ai , Ri , si ). This attribute phrase grounding is used as an input to our phrase-critic.
Whereas visual descriptions are encouraged to discuss attributes which are relevant to a specific class,
the grounding model is only trained to determine whether a natural language phrase is in an image.
Being discriminative rather than generative, the critic model does not have to learn to generate fluent,
grammatically correct sentences, and can thus focus on checking whether the mentioned attribute
phrases are image-relevant. Consequently, the models are complementary, allowing one model to
catch the mistakes of the other.
4
Experiments
For our experiments, we use the Caltech UCSD Birds 200-2011 (CUB) dataset [9] and sentences
collected by [7]. We first compare our proposed model with the baseline visual explanation model
of [2]. We present results in Figure 2. As a general observation, our critic model (1) grounds attribute
phrases both in the image and in the sentence, (2) is in favor of accurate and class-specific attribute
phrases and (3) provides the cumulative score of each explanatory sentence. To further emphasize the
importance of grounding attribute phrases in the image and in the sentence in evaluating the accuracy
3
This is a Bank Swallow
because this bird has a
speckled belly and breast
with a short pointy bill.
This is an Anna
Hummingbird because this
bird has a speckled belly
and breast with a long
needle like bill.
This is a Pigeon Guillemot
because this is a black bird
with a white wing and an
orange beak.
This is a Crested Auklet
because this is a black bird
with a white eye and an
orange beak.
Explanation Critic
Score: 12.1060
This is a Brewer Blackbird because this is
a black bird with a white eye and long
pointy black beak.
Baseline Explanation
This is a Pigeon Guillemot because this
is a black bird with a white wing and
red webbed feet and a black beak.
This is a Pigeon Guillemot because this
is a black bird with a white wing and an
orange beak.
Comparison of Explanation
Critic and Baseline
Best 3 Captions:
this is a black bird with a white eye and long pointy black beak.
this is a black bird with a white eye and a pointy black beak.
this all black bird has a short black beak and a lighter colored eye - coloring.
Score
1.46
1.35
1.32
Worst 3 Captions:
a long and containing yellow bird with a black head and long beak.
a stout bird that is completely a shiny completely black.
this is a shiny all blue bird with a yellow eye and black beak.
-2.54
-2.65
-2.88
Sampled Captions and Corresponding Scores from the Proposed
Critic Model
Figure 2: (Left) Compare our explanation critic to the baseline model [2]. For our model, detected
bounding boxes are color coded with their matching attribute phrases in the sentence. The cumulative
matching score of the explanation sentence and the visually grounded bounding box are provided
above the image. (Right) We show examples of sampled sentences and corresponding scores from
the ranker. Sentences which are more image relevant are scored higher by our phrase-critic model.
of the visual explanation model, let us more closely examine Figure 2 left. We note that the baseline
model mentions an “orange beak”. However, the Pigeon Guillemot in the image actually has a black
beak, which is properly localized using our proposed method.
Additionally, we also compare our phrase-critic ranking method to a ranking method based solely
on sentences fluency (Sf ). We sample 100 random images from the test set and find that attributes
mentioned by our critic model reflect the image more accurately than the baseline (85% image
relevant attributes vs. 79%).
Figure 2 shows sampled sentences and their corresponding scores. We see a precise localization of
small regions such as “white eye” for “Brewer Blackbird”. Note that the highest ranked explanations
for “Brewer Blackbird” both correctly mention “white eye” which is the strongest distinguishing
property of this bird from other black birds. The 3rd sentence gets ranked lower likely due to the
explanation “lighter colored eye-coloring” lacking fluency. Additionally, explanations preferred by
our phrase-critic model do not blindly mention class-specific attributes that do not appear in the
image.
References
[1] O. Biran and K. McKeown. Justification narratives for individual classifications. In Proceedings of the
AutoML workshop at ICML, volume 2014, 2014.
[2] L. A. Hendricks, Z. Akata, M. Rohrbach, J. Donahue, B. Schiele, and T. Darrell. Generating visual
explanations. In ECCV, 2016.
[3] R. Hu, M. Rohrbach, J. Andreas, T. Darrell, and K. Saenko. Modeling relationships in referential expressions
with compositional modular networks. In CVPR, 2017.
[4] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma,
et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. arXiv
preprint arXiv:1602.07332, 2016.
[5] D. Parikh and K. Grauman. Relative attributes. In ICCV, 2011.
[6] D. H. Park, L. A. Hendricks, Z. Akata, B. Schiele, T. Darrell, and M. Rohrbach. Attentive explanations:
Justifying decisions and pointing to the evidence. arXiv preprint arXiv:1612.04757, 2016.
[7] S. Reed, Z. Akata, H. Lee, and B. Schiele. Learning deep representations of fine-grained visual descriptions.
In CVPR, 2016.
[8] R. L. Teach and E. H. Shortliffe. An analysis of physician attitudes regarding computer-based clinical
consultation systems. Computers and Biomedical Research, 14(6):542–558, 1981.
[9] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-ucsd birds 200.
Technical report, California Institute of Technology, 2010.
4
| 1cs.CV
|
Optimal Number of Choices in Rating Contexts
arXiv:1605.06588v6 [cs.AI] 30 Jan 2018
Sam Ganzfried and Farzana Yusuf
School of Computing and Information Sciences
Florida International University
{sganzfri,fyusu003}@cis.fiu.edu
Abstract
In many settings people must give numerical scores to entities from a small discrete set. For instance,
rating physical attractiveness from 1–5 on dating sites, or papers from 1–10 for conference reviewing.
We study the problem of understanding when using a different number of options is optimal. For concreteness we assume the true underlying scores are integers from 1–100. We consider the case when
scores are uniform random and Gaussian. We study when using 2, 3, 4, 5, and 10 options is optimal in
these models. One may expect that using more options would always improve performance in this model,
but we show that this is not necessarily the case, and that using fewer choices—even just two—can surprisingly be optimal in certain situations. While in theory for this setting it would be optimal to use all
100 options, in practice this is prohibitive, and it is preferable to utilize a smaller number of options due
to humans’ limited computational resources. Our results suggest that using a smaller number of options
than is typical could be optimal in certain situations. This would have many potential applications, as
settings requiring entities to be ranked by humans are ubiquitous.
1
Introduction
Humans rate items or entities in many important settings. For example, users of dating websites and mobile
applications rate other users’ physical attractiveness, teachers rate scholarly work of students, and reviewers
rate the quality of academic conference submissions. In these settings, the users assign a numerical (integral)
score to each item from a small discrete set. However, the number of options in this set can vary significantly
between applications, and even within different instantiations of the same application. For instance, for
rating attractiveness, three popular sites all use a different number of options. On “Hot or Not,” users rate
the attractiveness of photographs submitted voluntarily by other users on a scale of 1–10. (Figure 11 ) These
scores are aggregated and the average is assigned as the overall “score” for a photograph. On the dating
website OkCupid, users rate other users on a scale of 1–5 (if a user rates another user 4 or 5 then the rated
user receives a notification). 2 (Figure 1 3 ). And on the mobile application Tinder users “swipe right” (green
heart) or “swipe left” (red X) to express interest in other users (two users are allowed to message each other
if they mutually swipe right), which is essentially equivalent to using a binary {1, 2} scale. (Figure 24 ).
Education is another important application area requiring human ratings. For the 2016 International Joint
Conference on Artificial Intelligence, reviewers assigned a “Summary Rating” score from -5–5 (equivalent
to 1–10) for each submitted paper. (Figure 2).5 ). The papers are then discussed and scores aggregated to
produce an acceptance or rejection decision based largely on the average.
1
http://blog.mrmeyer.com/2007/are-you-hot-or-not/
The likelihood of receiving an initial message is actually much more highly correlated with the variance—and particularly the
number of “5” ratings—than with the average rating [7].
3
http://blog.okcupid.com/index.php/the-mathematics-of-beauty/
4
https://tctechcrunch2011.files.wordpress.com/2015/11/tinder-two.jpg
5
https://easychair.org/conferences/?conf=ijcai16
2
1
Figure 1: Hot or Not users rate attractiveness 1–10; OkCupid users rate attractiveness 1–5.
Figure 2: Tinder users rate attractiveness 1–2; IJCAI reviewers rate papers -5–5.
Despite the importance and ubiquity of the problem, there has been little fundamental research done
on the problem of determining the optimal number of options to allow in such settings. We study a model
in which users have a underlying integral ground truth score for each item in {1, . . . , n} and are required
to submit an integral rating in {1, . . . , k}, for k << n. (For ease of presentation we use the equivalent
formulation {0, . . . , n − 1}, {0, . . . , k − 1}.) We use two generative models for the ground truth scores: a
uniform random model in which the fraction of scores for each value from 0 to n − 1 is chosen uniformly at
random (by choosing a random value for each and then normalizing), and a model where scores are chosen
according to a Gaussian distribution with a given mean and variance. We then compute a “compressed”
score distribution by mapping each full score s from {0, . . . , n − 1} to {0, . . . , k − 1} by applying
$
%
s
.
s←
(1)
n
k
We then compute the average “compressed” score ak , and compute its error ek according to
ek = af −
n−1
· ak ,
k−1
2
(2)
where af is the ground truth average. The goal is to pick argmink ek . While there are many possible generative models and cost functions, these seemed like the most natural.
We derive a closed-form expression for ek that depends on only a small number (k) of parameters of the
underlying distribution for an arbitrary distribution.6 This allows us to exactly characterize the performance
of using each number of choices. In computational simulations we repeatedly compute ek and compare the
average values. We focus on n = 100 and k = 2, 3, 4, 5, 10, which we believe are the most natural and
interesting choices for initial study.
One could argue that this model is somewhat “trivial” in the sense that it would be optimal to set k = n
to permit all the possible scores, as this would result in the “compressed” scores agreeing exactly with the
full scores. However, there are several reasons that would lead us to prefer to select k << n in practice (as
all of the examples previously described have done), thus making this “thought experiment” worthwhile.
It is much easier for a human to assign a score from a small set than from a large set, particularly when
rating many items under time constraints. We could have included an additional term into the cost function
ek that explicitly penalizes larger values of k, which would have a significant effect on the optimal value
of k (providing a favoritism for smaller values). However the selection of this function would be somewhat
arbitrary and would make the model more complex, and we leave this for future study. Given that we do not
include such a penalty term, one may expect that increasing k will always decrease ek in our setting. While
the simulations show a clear negative relationship, we show that smaller values of k actually lead to smaller
ek surprisingly often. These smaller values would receive further preference with a penalty term.
The most closely related theoretical work studies the impact of using finely grained numerical grades
(100, 99, 98) vs. coarse letter grades (A, B, C) [5]. They conclude that if students care primarily about their
rank relative to the other students, they are often best motivated to work by assigning them coarse categories
than exact numerical scores. In a setting of “disparate” student abilities they show that the optimal absolute
grading scheme is always coarse. Their model is game-theoretic; each player (student) selects an effort
level, seeking to optimize a utility function that depends on both the relative score and effort level. Their
setting is quite different from ours in many ways. For one, they study a setting where it is assumed that the
underlying “ground truth” score is known, yet may be disguised for strategic reasons. In our setting the goal
is to approximate the ground truth score as closely as possible.
While we are not aware of prior theoretical study of our problem, there have been experimental studies
on the optimal number of options on a “Likert scale” [10, 12, 18, 4, 6]. The general conclusion is that “the
optimal number of scale categories is content specific and a function of the conditions of measurement.” [8]
There has been study of whether including a “mid-point” option (i.e., the middle choice from an odd number)
is beneficial. One experiment demonstrated that the use of the mid-point category decreases as the number
of choices increases: 20% of respondents choose the mid-point for 3 and 5 options while only 7% did for
7, 9, . . . , 19 [13]. They conclude that it is preferable to either not include a mid-point at all or use a large
number of options. Subsequent experiments demonstrated that eliminating a mid-point can reduce social
desirability bias which results from respondents’ desires to please the interviewer or not give a perceived
socially unacceptable answer [8]. There has also been significant research on questionnaire design and the
concept of “feeling thermometers,” particularly from the fields of psychology and sociology [17, 14, 9, 3,
11, 15]. One study concludes: “in the measurement of satisfaction with various domains of life, 11-point
scales clearly are more reliable than comparable 7-point scales” [1]. Another study shows that “people are
more likely to purchase gourmet jams or chocolates or to undertake optional class essay assignments when
offered a limited array of 6 choices rather than a more extensive array of 24 or 30 choices” [16].
6
For theoretical simplicity we theoretically study a continuous version where scores are chosen according to a distribution over
(0, n) (though the simulations are for the discrete version) and the compressed scores are over {0, . . . , k − 1}. In this setting we
use a normalization factor of nk instead of n−1
for the ek term. Continuous approximations for large discrete spaces have been
k
studied in other settings; for instance, they have led to simplified analysis and insight in poker games with continuous distributions
of private information [2].
3
2
Theoretical characterization
Suppose scores are given by continuous pdf f (with cdf F ) on (0, 100), and we wish to compress them to
two options, {0, 1}.
Scores below 50 are mapped to 0, and above 50 to 1. The average of the full distribution
R 100
is af = E[X] = x=0 xf (x)dx. The average of the compressed version is
50
Z
100
Z
1f (x)dx = 1 − F (50).
0f (x)dx +
a2 =
x=50
x=0
So e2 = |af − 100(1 − F (50))| = |E[X] − 100 + 100F (50)|. For three options,
Z
a3 =
100
3
Z
0f (x)dx +
200
3
Z
x= 100
3
x=0
100
1f (x)dx +
2f (x)dx
x= 200
3
= 2 − F (100/3) − F (200/3)
e3 = |af − 50(2 − F (100/3) − F (200/3))|
= |E[X] − 100 + 50F (100/3) + 50F (200/3)|
In general for n total and k compressed options,
ak =
k−1 Z
X
i=0
n(i+1)
k
if (x)dx
x= ni
k
k−1
X
ni
n(i + 1)
−F
=
i F
k
k
i=0
k−1
X
ni
= (k − 1)F (n) −
F
k
i=1
k−1
X
ni
= (k − 1) −
F
k
i=1
ek =
n
af −
k−1
(k − 1) −
F
i=1
k−1
=
k−1
X
n X
E[X] − n +
F
k−1
i=1
ni
k
ni
k
!
(3)
Equation 3 allows us to characterize the relative performance of choices of k for
a given distribution f .
For each k it requires only knowing k statistics of f (the k − 1 values of F ni
k plus E[X]). In practice
these could likely be closely approximated from historical data for small k values.
As an example we see that e2 < e3 iff
200
100
+ 50F
|E[X] − 100 + 100F (50)| < E[X] − 100 + 50F
3
3
Consider a full distribution that has half its mass right around 30 and half its mass right around 60
(Figure 3). Then af = E[X] = 0.5 · 30 + 0.5 · 60 = 45. If we use k = 2, then the mass at 30 will be
4
mapped down to 0 (since 30 < 50) and the mass at 60 will be mapped up to 1 (since 60 > 50) (Figure 4).
So a2 = 0.5 · 0 + 0.5 · 1 = 0.5. Using normalization of nk = 100, e2 = |45 − 100(0.5)| = |45 − 50| = 5.
If we use k = 3, then the mass at 30 will also be mapped down to 0 (since 0 < 100
3 ); but the mass
200
at 60 will be mapped to 1 (not the maximum possible value of 2 in this case), since 100
3 < 60 < 3
(Figure 4). So again a3 = 0.5 · 0 + 0.5 · 1 = 0.5, but now using normalization of nk = 50 we have
e3 = |45 − 50(0.5)| = |45 − 25| = 20. So, surprisingly, in this example allowing more ranking choices
actually significantly increases error.
Figure 3: Example distribution for which compressing with k = 2 produces lower error than k = 3.
Figure 4: Compressed distributions using k = 2 and k = 3.
If we happened to be in the case where both a2 ≤ af and a3 ≤ af , then we could remove the absolute
R 200
R 50
3
f (x)dx. One could perform
values and reduce the expression to see that e2 < e3 iff x= 100 f (x)dx < x=50
3
more comprehensive analysis considering all cases to obtain better characterization and intuition for the
optimal value of k for distributions with different properties.
3
Rounding compression
An alternative model we could have considered is to use rounding to produce the compressed scores as
opposed to using the floor function from Equation 1. For instance, for the case n = 100, k = 2, instead of
dividing s by 50 and taking the floor, we could instead partition the points according to whether they are
closest to t1 = 25 or t2 = 75. In the example above, the mass at 30 would be mapped to t1 and the mass at
60 would be mapped to t2 . This would produce a compressed average score of a2 = 21 · 25 + 12 · 75 = 50.
5
No normalization would be necessary, and this would produce error of e2 = |af − a2 | = |45 − 50| = 5,
as the floor approach did as well. Similarly, for k = 3 the region midpoints will be q1 = 100
6 , q2 = 50,
100
q3 = 500
.
The
mass
at
30
will
be
mapped
to
q
=
and
the
mass
at
60
will
be
mapped
to q2 = 50.
1
6
6
1
100
This produces a compressed average score of a3 = 21 · 100
+
·
50
=
.
This
produces
an error of
6
2
3
100
35
e3 = |af − a3 | = 45 − 3 = 3 = 11.67. Although the error for k = 3 is smaller than for the floor
case, it is still significantly larger than k = 2’s, and using two options still outperforms using three for the
example in this new model.
In general, this approach would create k “midpoints” {mki }: mki = n(2i−1)
2k . For k = 2 we have
Z
50
Z
100
75 = 75 − 50F (50)
25 +
a2 =
x=0
x=50
e2 = |af − (75 − 50F (50))| = |E[X] − 75 + 50F (50)|
One might wonder whether the floor approach would ever outperform the rounding approach (in the
example above the rounding approach produced lower error k = 3 and the same error for k = 2). As a
simple example to see that it can, consider the distribution with all mass on 0. The floor approach would
produce a2 = 0 giving an error of 0, while the rounding approach would produce a2 = 25 giving an error
of 25. Thus, the superiority of the approach is dependent on the distribution. We explore this further in the
experiments.
For three options,
Z 100
Z 200
Z 100
3 100
3
500
a3 =
f (x) +
f (x)
50f (x) +
100
200
6
6
0
3
3
100
100
200
500 100
−
F
−
F
=
6
3
3
3
3
500 100
100
100
200
e3 = E[X] −
+
F
+
F
6
3
3
3
3
For general n and k, analysis as above yields
ak =
ek =
=
k−1 Z
X
n(i+1)
k
k−1
n(2k − 1) n X
=
−
F
ni
2k
k
x=
k
i=0
i=1
"
#
k−1
X
n(2k − 1) n
ni
af −
−
F
2k
k
k
i=1
k−1
n(2k − 1) n X
ni
E[X] −
+
F
2k
k
k
mki+1 f (x)dx
ni
k
(4)
(5)
i=1
Like for the floor model ek requires only knowing k statistics of f . The rounding model has an advantage
over the floor model that there is no need to convert scores between different scales and perform normalization. One drawback is that it requires knowing n (the expression for mki is dependent on n), while the floor
model does not. In our experiments we assume n = 100, but in practice it may not be clear what the agents’
ground truth granularity is and may be easier to just deal with scores from 1 to k. Furthermore, it may seem
200
unnatural to essentially ask people to rate items as “ 100
6 , 50, 6 ” rather than “1, 2, 3” (though the conversion
k
between the score and mi could be done behind the scenes essentially circumventing the potential practical
complication). One can generalize both the floor and rounding model by using a score of s(n, k)i for the
i’th region. For the floor setting we set s(n, k)i = i, and for the rounding setting s(n, k)i = mki = n(2i+1)
2k .
6
4
Computational simulations
The above analysis leads to the immediate question of whether the example for which e2 < e3 was a fluke
or whether using fewer choices can actually reduce error under reasonable assumptions on the generative
model. We study this question using simulations with what we believe are the two most natural models.
While we have studied the continuous setting where the full set of options is continuous over (0, n) and the
compressed set is discrete {0, . . . , k − 1}, we now consider the perhaps more realistic setting where the full
set is the discrete set {0, . . . , n − 1} and the compressed set is the same (though it should be noted that the
two settings are likely quite similar qualitatively).
The first generative model we consider is a uniform model in which the values of the pmf for each of
the n possible values are chosen independently and uniformly at random. The second is a Gaussian model
in which the values are generated according to a normal distribution with specified mean µ and standard
deviation σ (values below 0 are set to 0 and above n − 1 to n − 1). This model also takes as a parameter a
number of samples s to use for generating the scores.
For our simulations we used n = 100, and considered k = 2, 3, 4, 5, 10, which are popular and natural
values. For the Gaussian model we used s = 1000, µ = 50, σ = 50
3 . For each set of simulations we
computed the errors for all considered values of k for m = 100, 000 “items” (each corresponding to a
different distribution generated according to the specified model). The main quantities we are interested in
computing are the number of times that each value of k produces the lowest error over the m items, and the
average value of the errors over all items for each k value.
In the first set of experiments, we compared performance between using k = 2, 3, 4, 5, 10 to see for how
many of the trials each value of k produced the minimal error. The results are in Table 1. Not surprisingly,
we see that the number of victories increases monotonically with the value of k, while the average error
decreased monotonically (recall that we would have zero error if we set k = 100). However, what is perhaps
surprising is that using a smaller number of compressed scores produced the optimal error in a far from
negligible number of the trials. For the uniform model, using 10 scores minimized error only around 53% of
the time, while using 5 scores minimized error 17% of the time, and even using 2 scores minimized it 5.6%
of the time. The results were similar for the Gaussian model, though a bit more in favor of larger values of
k, which is what we would expect because the Gaussian model is less likely to generate “fluke” distributions
that could favor the smaller values.
2
Uniform # victories 5564
Uniform average error 1.32
Gaussian # victories 3025
Gaussian average error 1.14
3
9265
0.86
7336
0.59
4
14870
0.53
14435
0.30
5
16974
0.41
17800
0.22
10
53327
0.19
57404
0.10
Table 1: Number of times each value of k in {2,3,4,5,10} produces minimal error and average error values,
over 100,000 items generated according to both models.
We next explored the number of victories between just k = 2 and k = 3, with results in Table 2. Again
we observed that using a larger value of k generally reduces error, as expected. However, we find it extremely
surprising that using k = 2 produces a lower error 37% of the time. As before, the larger k value performs
relatively better in the Gaussian model. We also looked at results for the most extreme comparison, k = 2
vs. k = 10 (Table 3). Using 2 scores outperformed 10 8.3% of the time in the uniform setting, which was
larger than we expected. In Figures 5–6, we present a distribution for which k = 2 particularly outperformed
k = 10. The full distribution has mean 54.188, while the k = 2 compression has mean 0.548 (54.253 after
normalization) and k = 10 has mean 5.009 (55.009 after normalization). The normalized errors between the
7
means were 0.906 for k = 10 and 0.048 for k = 2, yielding a difference of 0.859 in favor of k = 2.
2
Uniform number of victories 36805
Uniform average error
1.31
Gaussian number of victories 30454
Gaussian average error
1.13
3
63195
0.86
69546
0.58
Table 2: Results for k = 2 vs. 3.
2
Uniform number of victories 8253
Uniform average error
1.32
Gaussian number of victories 4369
Gaussian average error
1.13
10
91747
0.19
95631
0.10
Table 3: Results for k = 2 vs. 10.
Figure 5: Example distribution where compressing with k = 2 produces significantly lower error than
k = 10. The full distribution has mean 54.188, while the k = 2 compression has mean 0.548 (54.253 after
normalization) and the k = 10 compression has mean 5.009 (55.009 after normalization). The normalized
errors between the means were 0.906 for k = 10 and 0.048 for k = 2, yielding a difference of 0.859 in favor
of k = 2.
We next repeated the extreme k = 2 vs. 10 comparison, but we imposed a restriction that the k = 10
option could not give a score below 3 or above 6 (Table 4). (If it selected a score below 3 then we set it to
3, and if above 6 we set it to 6). For some settings, for instance paper reviewing, extreme scores are very
uncommon, and we strongly suspect that the vast majority of scores are in this middle range. Some possible
explanations are that reviewers who give extreme scores may be required to put in additional work to justify
their scores and are more likely to be involved in arguments with other reviewers (or with the authors in the
rebuttal). Reviewers could also experience higher regret or embarrassment for being “wrong” and possibly
off-base in the review by missing an important nuance. In this setting using k = 2 outperforms k = 10
nearly 31 of the time in the uniform model.
We also considered the situation where we restricted the k = 10 scores to fall between 3 and 7 (as
opposed to 3 and 6). Note that the possible scores range from 0–9, so this restriction is asymmetric in that
8
Figure 6: Compressed distribution for k = 2 vs. 10.
the lowest three possible scores are eliminated while only the highest two are. This is motivated by the
intuition that raters may be less inclined to give extremely low scores which may hurt the feelings of an
author (for the case of paper reviewing). In this setting, which is seemingly quite similar to the 3–6 setting,
k = 2 produced lower error 93% of the time in the uniform model!
Uniform number of victories
Uniform average error
Gaussian number of victories
Gaussian average error
2
32250
1.31
10859
1.13
10
67750
0.74
89141
0.20
Table 4: For k = 10 only scores 3–6 permitted.
We next repeated these experiments for the rounding compression function. There are several interesting
observations from Table 6. In this setting, k = 3 is the clear choice, performing best in both models (by a
large margin for the Gaussian model). The smaller values of k perform significantly better with rounding
than flooring (as indicated by lower errors) while the larger values perform significantly worse, and their
errors seem to approach 0.5 for both models. Taking both compressions into account, the optimal overall
approach would still be to use flooring with k = 10, which produced the smallest average errors of 0.19
and 0.1 in the two models, while using k = 3 with rounding produced errors of 0.47 and 0.24. The 2 vs. 3
experiments produced very similar results for the two compressions (Table 7). The 2 vs. 10 results were quite
different, with 2 performing better almost 40% of the time with rounding, vs. less than 10% with flooring
(Table 8). In the 2 vs. 10 truncated 3–6 experiments 2 performed relatively better with rounding for both
models (Table 9), and for the 2 vs. 10 truncated 3–7 experiments k = 2 performed better nearly all the time
(Table 10).
2
Uniform number of victories 93226
Uniform average error
1.31
Gaussian number of victories 54459
Gaussian average error
1.13
10
6774
0.74
45541
1.09
Table 5: For k = 10 only scores 3–7 permitted.
9
Uniform # victories
Uniform average error
Gaussian # victories
Gaussian average error
2
15766
0.78
13262
0.67
3
33175
0.47
64870
0.24
4
21386
0.55
10331
0.50
5
19995
0.52
9689
0.50
10
9678
0.50
1848
0.50
Table 6: Number of times each value of k produces minimal error and average error values, over 100,000
items generated according to both models with rounding compression.
2
Uniform number of victories 33585
Uniform average error
0.78
Gaussian number of victories 18307
Gaussian average error
0.67
3
66415
0.47
81693
0.24
Table 7: k = 2 vs. 3 with rounding compression.
Uniform number of victories
Uniform average error
Gaussian number of victories
Gaussian average error
2
37225
0.78
37897
0.67
10
62775
0.50
62103
0.50
Table 8: k = 2 vs. 10 with rounding compression.
Uniform number of victories
Uniform average error
Gaussian number of victories
Gaussian average error
2
55676
0.79
24128
0.67
10
44324
0.89
75872
0.34
Table 9: k = 2 vs. 10 with rounding compression. For k = 10 only scores permitted between 3 and 6.
Uniform number of victories
Uniform average error
Gaussian number of victories
Gaussian average error
2
99586
0.78
95692
0.67
10
414
3.50
4308
1.45
Table 10: k = 2 vs. 10 with rounding compression. For k = 10 only scores permitted between 3 and 7.
5
Experiments
The empirical analysis of ranking based datasets depends on the availability of large amounts of data depicting different types of real scenarios. For our experimental setup we used two different datasets from “Rating
and Combinatorial Preference Data” of http://www.preflib.org/data/. One of these datasets
contains 675,069 ratings on scale 1-5 of 1,842 hotels from Trip Advisor website. Other consists of 398 approval ballots and subjective ratings on a 20-point scale collected over 15 potential candidates for the 2002
French Presidential election. The rating was provided by students at Institut d’Etudes Politiques de Paris.
10
For each set of experiments, we computed the errors for all considered values of k for these “items”. The
main quantities we are interested in computing are the number of times that each value of k produces the
lowest error over the items, and the average value of the errors over all items for each k value.
In the first set of experiments, the dataset contains different types of rating based on the price, quality of
rooms, proximity of location, cleanliness, availability of service from front desk etc and finally overall rating
provided by the users scraped from Trip advisor. We compared performance between using k = 2, 3, 4, 5
to see for how many of the trials each value of k produced the minimal error using the floor approach.
The results are in Table 11. Surprisingly, we see that the number of victories sometimes decreases with the
increase in value of k, while the average error decreased monotonically (recall that we would have zero error
if we set k to the actual maximum rating point). Like for k = 2 vs k = 4, the number of victories increases
for some cases compared to k = 2 vs k = 3. The number of victories between k = 2 vs k = 3, k = 2 vs
k = 4, and k = 2 vs k = 5 are shown in Table 13.
Average error
Overall
Price
Rooms
Location
Cleanliness
Front Desk
Service
Business Service
k=2
1.04
1.07
1.06
1.47
1.43
1.34
1.24
0.96
3
0.31
0.27
0.32
0.42
0.40
0.33
0.32
0.28
4
0.15
0.14
0.16
0.16
0.16
0.14
0.14
0.18
Table 11: Average error using floor approach for hotel ratings.
Minimal error k = 2
Overall
235
Price
181
Rooms
254
Location
111
Cleanliness
122
Front Desk
120
Service
140
Business Service 316
3
450
518
406
231
302
387
403
499
4
1157
1143
1182
1500
1418
1335
1299
1027
Table 12: Number of times each k produces minimal error.
We next explored rounding to generate the rating. For different value of k ∈ {2,3,4}, all the ratings
provided by users was compressed with the computed k midpoints and the average score was calculated.
Table 14 shows the average error induced by the compression which performs better than the floor approach
for this dataset.
An interesting observation found for rounding is that using k = n = 5 was outperformed by using k = 4
for some of the ratings. Also # of victories was higher in some case as shown in Table 17.
Next level of experiment was done on 2002 French Presidential Election. This dataset had both the
approval ballots and the subjective ratings of the candidates by each of the voters. Voter’s rated the potential
candidates on a scale of 20 where 0.0 is the lowest possible rating and -1.0 indicates missing rating value.
For the experiment, we have only considered ratings of the candidates. The # of victories and minimal error
while using floor were consistent for all comparison of k values resulting in higher error for lower value of k
11
# of victories
Overall
Price
Rooms
Location
Cleanliness
Front Desk
Service
Business Service
k= 2 vs 3
243, 1599
187, 1655
275, 1567
126, 1716
126, 1716
130, 1712
153, 1689
368, 1474
2 vs 4
277, 1565
211, 1631
283, 1559
122, 1720
141, 1701
133, 1709
152, 1690
329, 1513
2 vs 5
5, 1837
4, 1838
10, 1832
11, 1831
5, 1837
8, 1834
11, 1831
22, 1820
Table 13: Number of times k value produces minimal error.
Average error
Overall
Price
Rooms
Location
Cleanliness
Front Desk
Service
Business Service
k=2
0.50
0.48
0.48
0.63
0.6
0.55
0.52
0.39
3
0.28
0.31
0.30
0.41
0.4
0.39
0.36
0.36
4
0.15
0.15
0.16
0.22
0.21
0.21
0.18
0.18
Table 14: Average error using rounding approach.
Minimal error k = 2 3
4
Overall
82 132 1628
Price
92
74 1676
Rooms
152 81 1609
Location
93
52 1697
Cleanliness
79
44 1719
Front Desk
89
50 1703
Service
102 29 1711
Business Service 246 123 1473
Table 15: Number of times k minimizes error with rounding.
# of victories
Overall
Price
Rooms
Location
Cleanliness
Front Desk
Service
Business Service
k= 2 vs 3
161, 1681
270, 1572
344, 1498
275, 1567
210, 1632
380, 1462
358, 1484
870, 972
2 vs 4
113, 1729
101, 1741
173, 1669
109, 1733
90, 1752
95, 1747
109, 1733
278, 1564
2 vs 5
486, 1356
385, 1457
575, 1267
344, 1498
289, 1553
332, 1510
399, 1443
853, 989
Table 16: Number of times k minimizes error with rounding.
12
Average error
# of victories
Average error
Price
# of victories
Average error
Rooms
# of victories
Average error
Location
# of victories
Average error
Cleanliness
# of victories
Average error
Front Desk
# of victories
Average error
Service
# of victories
Average error
Business Service
# of victories
Overall
0.15, 0.21
1007, 835
0.15, 0.17
955, 887
0.15, 0.23
1076, 766
0.22, 0.22
694, 1148
0.21, 0.19
653, 1189
0.21, 0.17
662, 1180
0.18, 0.18
827, 1015
0.18, 0.31
1233, 609
Table 17: # victories and average error for rounding, k in {4,5}.
for each individual candidate. On the other hand, # of victories and minimal error with rounding compression
only produces minimal error for just one of the instance of k = 2, rests of the minimal error were produced
by the highest value of k i.e k = 10.
Average error
Francois Bayrou
Olivier Besancenot
Christine Boutin
Jacques Cheminade
Jean-Pierre Chevenement
Jacques Chirac
Robert Hue
Lionel Jospin
Arlette Laguiller
Brice Lalonde
Corine Lepage
Jean-Marie Le Pen
Alain Madelin
Noel Mamere
Bruno Maigret
2
3.18
1.7
1.15
0.64
3.69
3.48
2.39
5.45
2.2
1.53
2.24
0.4
1.93
3.68
0.31
3
1.5
0.8
0.54
0.3
1.74
1.64
1.12
2.57
1.04
0.72
1.06
0.19
0.91
1.74
0.15
4
0.94
0.5
0.34
0.19
1.09
1.03
0.7
1.61
0.65
0.45
0.67
0.12
0.57
1.09
0.09
5
0.66
0.35
0.24
0.13
0.77
0.72
0.49
1.13
0.46
0.32
0.47
0.08
0.4
0.77
0.06
8
0.3
0.16
0.11
0.06
0.35
0.33
0.22
0.52
0.21
0.14
0.22
0.04
0.18
0.35
0.03
10
0.2
0.1
0.07
0.04
0.23
0.21
0.14
0.33
0.13
0.09
0.14
0.02
0.12
0.23
0.02
Table 18: Average error of floor approach for French election.
6
Conclusion
Settings in which humans must rate items or entities from a small discrete set of options are ubiquitous. We
have singled out several important applications—rating attractiveness for dating websites, assigning grades
13
Average error
Francois Bayrou
Olivier Besancenot
Christine Boutin
Jacques Cheminade
Jean-Pierre Chevenement
Jacques Chirac
Robert Hue
Lionel Jospin
Arlette Laguiller
Brice Lalonde
Corine Lepage
Jean-Marie Le Pen
Alain Madelin
Noel Mamere
Bruno Maigret
2
1.65
3.88
3.87
4.34
1.47
1.64
2.51
0.33
2.62
3.45
2.89
4.92
3.18
2.02
4.88
3
0.73
2.39
2.39
2.72
0.65
1.0
1.27
0.49
1.34
1.9
1.59
3.26
1.8
1.55
3.23
4
0.91
2.14
1.84
2.07
1.2
1.13
1.14
0.87
1.34
1.55
1.56
2.55
1.52
1.77
2.46
5
0.75
1.7
1.5
1.65
0.82
0.88
1.09
0.67
1.02
1.21
1.16
2.06
1.17
1.44
1.99
8
0.48
1.31
0.9
1.02
0.55
0.55
0.67
0.51
0.6
0.66
0.79
1.39
0.72
1.29
1.28
10
0.62
1.25
0.86
0.88
0.61
0.64
0.77
0.63
0.63
0.78
0.87
1.2
0.7
1.41
1.1
Table 19: Average error using rounding approach.
to students, and reviewing academic papers. The number of available options can vary considerably, even
within different instantiations of the same application. For instance, we saw that three popular sites for
attractiveness rating use completely different systems: Hot or Not uses a 1–10 system, OkCupid uses 1–5
“star” system, and Tinder uses a binary 1–2 “swipe” system.
Despite the problem’s importance, we have not seen it studied theoretically previously. Our goal is to
select k to minimize the average (normalized) error between the compressed average score and the ground
truth average. We studied two natural models for generating the scores. The first is a uniform model where
the scores are selected independently and uniformly at random, and the second is a Gaussian model where
they are selected according to a more structured procedure that gives preference for the options near the
center. We provided a closed-form solution for continuous distributions with arbitrary cdf. This allows us to
characterize the relative performance of choices of k for a given distribution. We saw that, counterintuitively,
using a smaller value of k can actually produce lower error: for some distributions (even though we know
that as k approaches n the error approaches 0): we presented specific distributions for which using k = 2
outperforms 3 and 10.
In our simulations, we observed that performance generally improves monotonically with k as expected,
and more so for the Gaussian model than uniform. However, we observe that small k values can be optimal
a non-negligible amount of the time, which is perhaps counterintuitive. Using k = 2 outperformed 10 8.3%
of the time, and when we restricted k = 10 to only assign values between 3 and 7 inclusive, k = 2 actually
produced lower error 93% of the time! This could correspond to a setting where raters are ashamed to assign
extreme scores (particularly extreme low scores). For smaller k rounding leads to significantly lower error
than flooring, with k = 3 the clear optimal choice, while for larger k rounding leads to much larger error.
A future avenue is to extend our analysis to better understand specific distributions for which different k
values are optimal, while our simulations are in aggregate over many distributions. Application domains will
have distributions with different properties, and improved understanding will allow us to determine which k
is optimal for the types of distributions we expect to encounter. This improved understanding can be coupled
with further data exploration.
14
References
[1] Duane F. Alwin. Feeling thermometers versus 7-point scales: Which are better? Sociological Methods
and Research, 25:318–340, February 1997.
[2] Jerrod Ankenman and Bill Chen. The Mathematics of Poker. ConJelCo LLC, Pittsburgh, PA, USA,
2006.
[3] Carolyn C Preston and Andrew Colman. Optimal number of response categories in rating scales:
Reliability, validity, discriminating power, and respondent preferences. Acta psychologica, 104:1–15,
04 2000.
[4] Eli P. Cox III. The optimal number of response alternatives for a scale: A review. Journal of Marketing
Research, 17:407–442, 1980.
[5] Pradeep Dubey and John Geanakoplos. Grading exams: 100, 99, 98, . . . or A, B, C? Games and
Economic Behavior, 69:72–94, May 2010.
[6] Hershey H. Friedman, Yonah Wilamowsky, and Linda W. Friedman. A comparison of balanced and
unbalanced rating scales. The Mid-Atlantic Journal of Business, 19(2):1–7, 1981.
[7] Hannah Fry. The Mathematics of Love. TED Books, New York, NY, 2015.
[8] Ron Garland. The mid-point on a rating scale: is it desirable? Marketing Bulletin, 2:66–70, 1991.
Research Note 3.
[9] James R. Lewis and Oǧuzhan Erdinç. User experience rating scales with 7, 11, or 101 points: Does it
matter? J. Usability Studies, 12(2):73–91, feb 2017.
[10] Rensis A. Likert. A technique for measurement of attitudes. Archives of Psychology, 22:1–, 01 1932.
[11] Luis Lozano, Eduardo Garca-Cueto, and Jos Muiz. Effect of the number of response categories on the
reliability and validity of rating scales. Methodology, 4:73–79, 05 2008.
[12] Michael S. Matell and Jacob Jacoby. Is there an optimal number of alternatives for likert scale items?
Study 1: reliability and validity. Educational and psychological measurements, 31:657–674, 1971.
[13] Michael S. Matell and Jacob Jacoby. Is there an optimal number of alternatives for likert scale items?
Effects of testing time and scale properties. Journal of Applied Psychology, 56(6):506–509, 1972.
[14] Aubrey C. McKennell. Surveying attitude structures: A discussion of principles and procedures. Quality and Quantity, 7(2):203–294, Sep 1974.
[15] Donald R. Lehmann and James Hulbert. Are three-point scales always good enough? Journal of
Marketing Research, 9:444, 11 1972.
[16] Sheena S. Iyengar and Mark Lepper. When choice is demotivating: Can one desire too much of a good
thing? Journal of personality and social psychology, 79:995–1006, 01 2001.
[17] Seymour Sudman, Norman M. Bradburn, and Norbert Schwartz. Thinking About Answers: The Application of Cognitive Processes to Survey Methodology. Jossey-Bass Publishers, San Francisco, California, 1996.
[18] Albert R. Wildt and Michael B. Mazis. Determinants of scale response: label versus position. Journal
of Marketing Research, 15:261–267, 1978.
15
| 2cs.AI
|
QUASI-AUTOMORPHISMS OF THE INFINITE ROOTED
2-EDGE-COLOURED BINARY TREE
arXiv:1412.7715v2 [math.GR] 20 Jul 2017
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Abstract. We study the group QV , the self-maps of the infinite 2-edge
coloured binary tree which preserve the edge and colour relations at cofinitely
many locations. We introduce related groups QF , QT , Q̃T , and Q̃V , prove
that QF , Q̃T , and Q̃V are of type F∞ , and calculate finite presentations for
them. We calculate the normal subgroup structure of all 5 groups, the Bieri–
Neumann–Strebel–Renz invariants of QF , and discuss the relationship of all 5
groups with other generalisations of Thompson’s groups.
1. Introduction
R. Thompson’s groups F , T , and V were originally used to, among other things,
construct finitely presented groups with unsolvable word problem [Tho80, MT73].
Brown and Geoghegan proved F is FP∞ [BG84], giving the first example of a
torsion-free FP∞ group with infinite cohomological dimension, and later Brown
used an important new geometric technique to prove that F , T , and V are all of
type F∞ [Bro87]. For more background on Thompson’s groups, see [CFP96].
Recently many variations on Thompson’s groups have been studied, and using
techniques similar to Brown’s, many of these have been proven to be of type F∞ .
These include Brin’s groups sV [FMWZ13], generalisations of these [MPMN14],
and the braided Thompson’s groups BV and BF [BFS+ 12] which were originally
defined by Brin [Bri07].
Let T2,c denote the infinite binary 2-edge-coloured tree and let QV be the group
of all bijections on the vertices of T2,c which respect the edge and colour relations
except for possibly at finitely many locations. The group QV was studied by
Lehnert [Leh08] who proved that QV is co-context-free, i.e. the co-word problem in
QV is context-free. He also gave an embedding of V into QV . More recently, Bleak,
Matucci, and Neunhöffer gave an embedding of QV into V [BMN13]. The group
QV was called QAut(T2,c ) by Bleak, Matucci, and Neunhöffer, and was called both
QAut(T2,c ) and G by Lehnert.
We write QV instead of QAut(T2,c ) in order to follow the convention of naming
varations of Thompson’s groups using a prefix (e.g. BV for the braided Thompson’s
groups [BS08] and nV for the Brin–Thompson groups [Bri04]). The Q may be
thought of as standing for “quasi-automorphism”.
There is a natural surjection π : QV −→ V (Lemma 2.1), and we denote by QF
the preimage π −1 (F ) and by QT the preimage π −1 (T ). The group QF was studied
by Lehnert, where it is called G0 [Leh08, Definition 2.9]. The surjection π : QV → V
is not split (Proposition 2.9), although it does split when restricted to QF (Lemma
2.5). This situation is similar to that of the braided Thompson groups—BF splits
as an extension of F by the infinite pure braid group P∞ , whereas BV is a non-split
extension of V by P∞ [Zar14, §1.2]. There is no braided analogue of T .
Date: February 18, 2018.
1991 Mathematics Subject Classification. 20J05.
Key words and phrases. Thompson’s Group, Finiteness Properties, Normal Subgroups, Bieri–
Neumann–Strebel–Renz invariants.
1
2
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
The following theorem appeared in Lehnert’s thesis, [Leh08, Satz 2.14, 2.30],
although his proof that QF is of type FP∞ uses Brown’s criterion, a different
approach from that used here.
Theorems 3.6 and 4.10. QF is F∞ and has presentation
2
−1 −1
+
*
σ 2 , [σ, σ α ], (σσ α )3 , σσ α σσ αβ α ,
−1
−1
−1
−2
2
QF = σ, α, β [αβ , α βα], [αβ , α βα ],
[ν, σ] for all ν ∈ X.
where
X=
β, β α , α3 β −1 α−2 , α2 β 2 α−1 β −1 αβ −1 α−2
2 −1 −1
αβ α β αβ −1 α−1 , αβα−1 β 2 α−1 β −1 αβ −1 αβ −1 α−1
.
We write T2,c ∪ {ζ} for the disjoint union of T2,c and a single vertex labelled
ζ. We write Q̃V for the group of all bijections on the vertices of T2,c ∪ {ζ} which
respect the edge and colour relations except for possibly at finitely many locations.
Again, there is a surjection π : Q̃V → V and we write Q̃T for the preimage π −1 (T ).
The groups Q̃V and Q̃T are easier to study than QV and QT because π : Q̃V → V
splits (Lemma 2.3).
In Section 3 we find finite presentations for Q̃V and Q̃T and in Section 4 we
show Q̃V , Q̃T and QF are of type F∞ .
Section 5 contains a complete classification of the normal subgroups of QF , QT ,
QV , Q̃T , and Q̃V . We find that the structure is very similar to that of F , T , and
V.
ε
0
1
00
···
01
···
···
10
···
···
11
···
···
···
Figure 1. T2,c , the rooted 2-edge-coloured infinite binary tree.
We identify the vertices of T2,c with the set {0, 1}∗ of finite length words on the
alphabet {0, 1}, denoting the empty word by ε (see Figure 1), and hence view QV
as acting on {0, 1}∗. Let Sym({0, 1}∗ ) denote the finite support permutation group
on {0, 1}∗ and Alt({0, 1}∗ ) the finite support alternating group on {0, 1}∗. Both
these groups are normal in QV , QT , and QF . Let Z = {0, 1}∗ ∪ {ζ}, so that Z can
be identified with the set of vertices of T2,c ∪ {ζ}, then Sym(Z) and Alt(Z) have the
obvious definition and are normal in Q̃V and Q̃T . Recall that F/[F, F ] ∼
= Z ⊕ Z,
but since T and V are simple and non-abelian, [T, T ] ∼
= T and [V, V ] ∼
= V [CFP96,
Theorems 4.1, 5.8, 6.9].
Theorem 5.1 (Normal subgroup structure).
(1) A non-trivial normal subgroup of QF is either Alt({0, 1}∗), Sym({0, 1}∗),
or contains
[QF, QF ] = Alt({0, 1}∗) ⋊ [F, F ].
QUASI-AUTOMORPHISMS OF T2,c
3
(2) A proper non-trivial normal subgroup of Q̃T is either Alt(Z), Sym(Z), or
[Q̃T, Q̃T ] = Alt(Z) ⋊ T.
(3) A proper non-trivial normal subgroup of Q̃V is either Alt(Z), Sym(Z), or
[Q̃V, Q̃V ] = Alt(Z) ⋊ V.
(4) A proper non-trivial normal subgroup of QT is one of either Alt({0, 1}∗),
Sym({0, 1}∗), or
[QT, QT ] = (Alt(Z) ⋊ T ) ∩ QT.
Moreover, (Alt(Z) ⋊ T ) ∩ QT is an extension of T by Sym({0, 1}∗).
(5) A proper non-trivial normal subgroup of QV is one of either Alt({0, 1}∗),
Sym({0, 1}∗), or
[QV, QV ] = (Alt(Z) ⋊ V ) ∩ QV.
Moreover, (Alt(Z) ⋊ V ) ∩ QV is an extension of V by Sym({0, 1}∗ ).
In Section 6 we calculate the Bieri–Neumann–Strebel–Renz invariants, or Sigma
invariants, of QF (they are uninteresting for the groups QT , QV , Q̃T , and Q̃V
since they all have finite abelianisation). They have previously been computed
for F by Bieri, Geoghegan, and Kochloukova [BGK10], and for BF by Zaremsky
[Zar14, §3]. We also show that QF is an ascending HNN extension
QF ∼
= QF (1)∗θ,t ,
where QF ∼
= QF (1).
We find that Σi (QF ) ∼
= Σi (F ) for all i. More precisely, we define π ∗ χ = χ ◦ π
for any character χ of F and prove that π ∗ induces an isomorphism between the
character spheres S(F ) and S(QF ).
Let χ0 : F → R and χ1 : F → R be the two characters given by χ0 (A) = −1,
χ0 (B) = 1, and χ0 (A) = χ1 (B) = 1, where A and B are the standard generators
of F .
Theorem 6.3 (The Bieri–Neumann–Strebel–Renz invariants of QF ). The character sphere S(QF ) is isomorphic to S 1 and for any ring R,
(1) Σ1 (QF, R) = Σ1 (QF ) = S(QF ) \ {[π ∗ χ0 ], [π ∗ χ1 ]}.
(2) Σi (QF, R) = Σi (QF ) = S(QF ) \ {[aπ ∗ χ0 + bπ ∗ χ1 ] : a, b ≥ 0} for all i ≥ 2.
In Section 7 we study the relationship of the groups studied in this note to previously studied families of groups which generalise Thompson’s groups and are known
to be of type F∞ . We show that the groups studied here cannot be proved to be of
type F∞ using Farley–Hughes’ study of Groups of Local Similarities [FH12, Theorem 1.1], or Martı́nez-Pérez–Nucinkis’ study of Automorphism Groups of Cantor
Algebras [MPN13]. We also show that one cannot use Thumann’s study of Operad
groups [Thu16] to determine whether QV is of type F∞ .
Remark 1.1 (Notation). A tree diagram representing an element of Thompson’s
group V is a triple (L, R, f ) where L and R are rooted subtrees of the infinite
binary rooted tree with equal numbers of leaves, and f is a bijection from the
leaves of L to the leaves of R. To represent an element of Thompson’s group F
we require only the first two elements of the tuple—the bijection is understood to
preserve the left-to-right ordering of the leaves. For further background on tree
diagrams see [CFP96, §2].
Acknowledgement
The authors would like to thank the referees for helpful comments.
4
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
2. Introducing QF , QT , QV , Q̃T , and Q̃V
2.1. The group QV . As mentioned in the introduction the group QV is the group
of all bijections on the vertices of T2,c , the infinite binary 2-edge-coloured tree,
which respect the edge and colour relations except for possibly at finitely many
locations. Given some τ ∈ QV and a ∈ {0, 1}∗, let sa be the maximal suffix such
that a = xa · sa and τ a = ya · sa for some xa , ya ∈ {0, 1}∗ (here the operation · is
concatenation of words). Bleak, Mattuci and Neunhöffer prove that the set
{(xa , ya ) : a ∈ {0, 1}∗}
is finite [BMN13, Claim 8], and hence that the subset
Mτ = {(x, y) : There are infinitely many a ∈ {0, 1}∗ with (xa , ya ) = (x, y)}
is finite also. Moreover, they show that the sets
Lτ = {x : (x, y) ∈ Mτ }
and
Rτ = {y : (x, y) ∈ Mτ }
both form finite complete anti-chains for the poset {0, 1}∗, where x ≤ x′ if x is a
subword of x′ . Write bτ for the bijection
bτ : Lτ −→ Rτ .
Note that restricting τ to Lτ does not necessarily give the bijection bτ . This is
because if σ is any finite permutation of {0, 1}∗ then bσ◦τ = bτ .
Let L and R be the binary trees with leaves Lτ and Rτ respectively, then
(L, R, bτ ) is a tree-pair diagram for an element of V which we denote by vτ .
In fact, vτ can be viewed as the induced action on the space of ends of the
simplicial tree T2,c , the set whose elements are enumerated by the right-infinite
words {0, 1}N. With its natural topology the space of ends is homeomorphic to
a Cantor set with vτ acting as a homeomorphism. Explicitly, the action of vτ on
{0, 1}∗ is
vτ : {0, 1}N −→ {0, 1}N
x · s 7−→ y · s for (x, y) in Mτ .
This is well-defined since Lτ is a complete anti-chain in the poset {0, 1}∗.
Lemma 2.1. The map π : QV −→ V sending τ 7→ vτ is a group homomorphism.
Proof. For any topological space X let Ends(X) denote the space of ends of X.
If τ ∈ QV then τ may not be a simplicial automorphism of T2,c , but there will
always exist a rooted subtree L ⊆ T2,c such that
τ |T2,c \L : T2,c \ L −→ T2,c \ τ (L)
is a simplicial map. Thus τ determines a map
Ends(τ ) : Ends(T2,c ) −→ Ends(T2,c ).
This is well-defined—if we choose to remove a different rooted subtree L′ instead
then, since removing a compact set doesn’t affect the ends of a space, we see that
T2,c \ (L ∪ L′ ) −→ T2,c \ τ (L ∪ L′ )
and hence
T2,c \ (L′ ) −→ T2,c \ τ (L′ )
induce the same map on Ends(T2,c ).
QUASI-AUTOMORPHISMS OF T2,c
5
If σ is some other element of QV then choosing a large enough finite rooted
subtree L, both maps in the composition
τ
σ
T2,c \ L 7−→ T2,c \ τ (L) 7−→ T2,c \ σ ◦ τ (L)
are simplicial maps. Now use that Ends(−) is a functor.
The kernel of π are those permutations of {0, 1}∗ which are the identity on all
but finitely many points, the finite support permutation group Sym({0, 1}∗ ). Thus
there is a group extension
1 −→ Sym({0, 1}∗) −→ QV −→ V −→ 1.
This extension does not split, see Proposition 2.9.
2.2. The group Q̃V . Recall that T2,c ∪ {ζ} denotes the 2-edge-coloured tree T2,c
together with an isolated vertex ζ, so the vertex set of T2,c ∪{ζ} is Z = {0, 1}∗ ∪{ζ}.
There is a map π from Q̃V to V defined as previously, whose kernel is Sym(Z).
Let ≤lex be the total order on {0, 1}∗ defined by following rule: Given a word
x, y ∈ {0, 1} let x̃, ỹ ∈ {0, 21 , 1}N be the words obtained by concatenating x and
y respectively with the infinite word containing only the symbol 21 , now we say
that x ≤lex y if and only if x̃ is smaller that ỹ in the lexicographical order where
1
1. For example, 00 ≤lex 0 but 01 ≥lex 0. We extend the ≤lex order to Z
0
2
so that ζ is strictly larger than all elements of {0, 1}∗ .
Remark 2.2. If T is any finite binary tree, there is an order preserving bijection
bT : nodes(T ) ∪ {ζ} −→ leaves(T ).
Moreover, adding a caret to T at any leaf doesn’t affect the bijection on the remaining leaves.
Lemma 2.3. The short exact sequence
π
1 −→ Sym(Z) −→ Q̃V −→ V −→ 1
is split.
Using the computer algebra package GAP Bleak, Matucci, and Neunhöffer provide
the same splitting [BMN13, Lemma 15].
Proof. Let (L, R, f ) be a tree diagram for an element v of V , we define the element
ι(v) of Q̃V by
ι(v) : {0, 1}∗ ∪ {ζ} −→ {0, 1}∗ ∪ {ζ}
−1
bR ◦ f ◦ bL (x)
x 7−→
f (x)
if x ∈ nodes(L) ∪ {ζ},
if x ∈ leaves(L),
and extend ι(v) onto all {0, 1}∗ by having ι(v) preserve all edge and colour relations
below leaves(L).
If (L′ , R′ , f ′ ) is an expansion of (L, R, f ) then (L′ , R′ , f ′ ) determines the same
element ι(v) (use Remark 2.2). Since any two tree diagrams representing the same
element of V have a common expansion this proves ι(v) is independent of choice of
tree diagram.
We claim that ι : v 7→ ι(v) is a group homomorphism. Let v and w be two
elements of V represented by tree diagrams (L, R, f ) and (R, S, g) (by performing
expansions we may assume any two elements of V are represented by tree pair
6
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
diagrams in this form). Thus the composition of the two elements of V has tree
pair diagram (L, S, g ◦ f ). Now one verifies that for x ∈ nodes(L) ∪ {ζ},
−1
ι(v) ◦ ι(w)(x) = bS−1 ◦ g ◦ bR ◦ bR
◦ f ◦ bL (x)
= bS−1 ◦ g ◦ f ◦ bL (x)
= ι(vw)(x).
Similarly, for x ∈ leaves(L), ι(v) ◦ ι(w) and ι(vw) agree.
Remark 2.4. For an explicit description of this splitting for the generators of V ,
let A, B, C and D denote the generators of V as denoted by A, B, C and π0 in
[CFP96, Figures 16 and 17]. Then ι(A), ι(B), ι(C), and ι(D) are shown in Figures
2–5.
Hereafter we define α = ι(A), β = ι(B), γ = ι(C), and δ = ι(D).
ε
ζ
ε
1
0
10
ζ
1
11
0
11
10
Figure 2. Element α = ι(A).
ε
ε
ζ
1
0
10
ζ
11
0
11
1
111
110 111
10 110
Figure 3. Element β = ι(B).
ε
ζ
10
ζ
10
1
0
ε
1
11
11
0
Figure 4. Element γ = ι(C).
2.3. The group QF . We define the group QF to be the preimage π −1 (F ) of
Thompson’s group F under π : QV → V .
Lemma 2.5. The short exact sequence
π
1 −→ Sym({0, 1}∗ ) −→ QF −→ F −→ 1
is split.
QUASI-AUTOMORPHISMS OF T2,c
ε
ζ
10
ζ
1
ε
10
1
0
7
11
0
11
Figure 5. Element δ = ι(D).
Proof. Restrict the splitting map ι defined in Lemma 2.3 to QF . Since elements
(L, R) of F always map the right-hand leaf of L to the right-hand leaf of R, we see
that ι(f ) fixes ζ for all f ∈ F .
Remark 2.6. If (L, R) is a tree-diagram for an element f of F , where L has nodes
(x1 , . . . , xn ) and R has nodes (y1 , . . . , yn ) then, assuming xi ≤lex xi+1 and yi ≤lex
yi+1 for all 1 ≤ i ≤ n − 1, ι(f ) maps xi 7→ yi . The action on all other elements of
{0, 1}∗ is determined by L and R—if li is the ith leaf of L and ri is the ith leaf of
R then ι(f ) takes the full subtree with root li to the full subtree with root ri .
In particular, for all f ∈ F , the image ι(f ) preserves the ≤lex order on {0, 1}∗ .
2.4. The group QT . We define the group QT to be the preimage π −1 (T ) of
Thompson’s group T under π : QV −→ V .
Restricting the map π to QT gives a short exact sequence
π
1 −→ Sym({0, 1}∗) −→ QT −→ T −→ 1.
In [Leh08, Lemma 2.10], Lehnert proves that the splitting ι : F → QF does not
extend to an embedding of T into QV . Examining the proof, he in fact shows the
stronger statement below. Recall that A and B denote the standard generators of
F.
Lemma 2.7. [Leh08, Lemma 2.10] Let j be an embedding of F into QV such that
j(A) = ι(A), then j does not extend to an embedding of T into QV .
Lemma 2.8. Let f be an element of F and let j : F → QF be a splitting of
π : QF → F , then the actions of ι(f ) and j(f ) on {0, 1}∗ agree on cofinitely many
points.
Proof. Recall from Section 2.1 that given τ ∈ QV , for all but finitely many a ∈
{0, 1}∗ we have a = xa · sa for some xa ∈ Lτ and τ (a) = bτ (xa ) · sa . Since the
tree pair diagram for π(τ ) is (L, R, bτ ), bτ is determined completely by π(τ ). We
conclude that ι(f ) and j(f ) agree on all but finitely many elements of {0, 1}∗, since
by assumption π ◦ ι(f ) = π ◦ j(f ).
Proposition 2.9. The group extension
π
1 −→ Sym({0, 1}∗) −→ QT −→ T −→ 1.
is not split.
Proof. Let j : T → QT be a splitting of π, by Lemma 2.8 we have that j(A) and
ι(A) agree on cofinitely many points. Hence, after possibly extending via adding
carets, we can find a tree-pair diagram (L, R, f ) and positive integer k such that in
L every leaf under 0 has length exactly k, and every leaf under 1 has length exactly
k + 1 and similarly in R every leaf under 0 has length exactly k + 1, and every leaf
under 1 has length exactly k.
Following the argument of the proof of Lemma 2.7, we see that this is incompatible with every possible choice of j(C).
8
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
3. Finite presentations
In this section we compute finite presentations for the groups QF , Q̃T , and Q̃V .
We use a method for finding a finite presentation for a semi-direct product N ⋊ Q,
whereby we find a presentation for N such that Q acts by permutations and with
finitely many orbits on both the generating set and the relating set. Then as long
as Q is finitely presented, and the Q-stabilisers of the generating set are finitely
generated, one can write down a finite presentation for N ⋊ Q [MPMN14, Lemma
A.1]. This technique predates the referenced paper of Martı́nez-Pérez, Matucci,
and Nucinkis, but we are not aware of anywhere else where the proof is given.
We apply this method to QF , Q̃T and Q̃V using their decompositions as semidirect products with kernel an infinite symmetric group and quotient F , T , and V
respectively.
Let N = hSN |RN i and Q = hSQ |RQ i be groups such that Q acts on SN by
permutations and the action induces an action on RN with Q(RN ) = RN . Let S0
be a set of representatives for the Q-orbits in SN and R0 a set of representatives
for the Q-orbits in RN .
For any set X, we write X ∗ for the finite length words of X and we write q −1 sq
∗
to denote the action of q ∈ SQ on s ∈ S0 . Let r ∈ R0 , then r is a word in SN
and
∗
hence can be expressed as a word in (S0 ∪ SQ ) using the conjugation notation just
b0 .
defined. Denote the set of these words, one for each r ∈ R0 , by R
Lemma 3.1 ([MPMN14, Lemma A.1]). With the notation above,
b0 , RQ , [x, s] for all s ∈ S0 and all x ∈ Xs i,
N ⋊Q∼
= hS0 , SQ |R
where Xs is a generating set for StabQ (s) and the semi-direct product is built using
the given action of Q on N .
3.1. A finite presentation for QF .
Lemma 3.2. F acts transitively on
(
n
Y
Σn = (x1 , . . . , xn ) ∈
{0, 1}∗ : x1
lex
x2
lex
···
lex
xn
1
)
for all n ∈ N, where the action is via the splitting ι : F −→ QF , the inclusion
QF −→ QV , and the usual action of QV on {0, 1}∗.
Proof. For i ≥ 1, let 0i be the following element of {0, 1}∗
i-times,
z }| {
0 = 0 · · · 0,
i
and let 00 = ε. For 1 ≤ i ≤ n, let yi = 0n−i .
For any (x1 , . . . , xn ) ∈ Σn , we build an element of γ of F such that
ι(γ) : (y1 , y2 , . . . , yn ) 7→ (x1 , . . . , xn )
Let R be the smallest rooted subtree of T2,c containing all the xi as nodes and let
L0 be the smallest subtree of T2,c containing all the yi as nodes.
For any two nodes z1 lex z2 in a finite rooted subtree T of T2,c we define
dT (z1 , z2 ) = |{z ∈ nodes(T ) : z1
lex
dT (z1 , ∗) = |{z ∈ nodes(T ) : z1
z
lex
lex
z}|
z2 }|
dT (∗, z2 ) = |{z ∈ nodes(T ) : z lex z2 }|.
Thus dL0 (∗, y0 ) = dL0 (yi , yi+1 ) = dL0 (yn , ∗) = 0 for all 0 ≤ i ≤ n − 1.
Expand L0 to L by adding carets until dL (yi , yi+1 ) = dR (xi , xi+1 ) for all i. This
is possible because adding a caret at a leaf between yi and yi+1 increases dL (yi , yi+1 )
QUASI-AUTOMORPHISMS OF T2,c
9
by 1. Now use the description of Remark 2.6 to check that the element (L, R) maps
(y1 , . . . , yn ) onto (x1 , . . . , xn ).
Lemma 3.3.
Qn+1
(1) The F -stabiliser of any element in Σn is isomorphic to the group 1 F .
(2) The F -stabiliser of (0, ε) ∈ Σ2 has generating set
β, β α , α3 β −1 α−2 , α2 β 2 α−1 β −1 αβ −1 α−2
.
StabΣ2 ((0, ε)) =
αβ 2 α−1 β −1 αβ −1 α−1 , αβα−1 β 2 α−1 β −1 αβ −1 αβ −1 α−1
We use the method in [Bel04, §1] to translate from tree diagrams to words in
the generators α and β.
Proof. Recall that yi = 0n−i for 0 ≤ i ≤ n − 1 and y n = ε. For any x ∈ {0, 1}∗, let
x · {0, 1}∗ denote all finite words beginning with x. We calculate the stabiliser of
(y1 , . . . , yn ) ∈ Σn , since F acts transitively on Σn this determines the stabilisers of
all elements of Σn up to conjugacy.
We claim the stabiliser of (y1 , . . . , yn ) is exactly those elements of F for which
ι(F ) preserves setwise y0 · {0, 1}∗ and yi · 1 · {0, 1}∗ for all 1 ≤ i ≤ n.
If f ∈ F and ι(f ) maps some x ∈ y0 · {0, 1}∗ to an element in yi · 1 · {0, 1}∗
for some 1 ≤ i ≤ n then, since ι(f ) preserves the lex ordering, it is impossible
for ι(f ) to map y1 to y1 as required. Similarly, again because ι(f ) respects the
lex ordering, any ι(f ) setwise preserving these sets will necessarily map y1 to y1 .
Repeating this argument for elements in yj · 1 · {0, 1}∗ completes the proof of (1).
Rephrasing the above, the stabiliser of (0, ε) is exactly those elements of F which
preserve setwise the setwise the subtrees under 00, 01, and 1. Those elements
preserving, for example the subtree under 00, and acting trivially elsewhere give
a copy of F on that interval. Since we have three such intervals, the stabiliser is
F × F × F.
Generators of the elements setwise preserving the subtree under 1 and acting
trivially elsewhere are {β, β α } where α = ι(A) and β = ι(B). Generators of the
elements setwise preserving the subtree under 00 and acting trivially elsewhere are
α3 β −1 α−2 and α2 β 2 α−1 β −1 αβ −1 α−2 . Generators of the elements setwise preserving the subtree under 01 and acting trivially elsewhere are αβ 2 α−1 β −1 αβ −1 α−1
and αβα−1 β 2 α−1 β −1 αβ −1 αβ −1 α−1 .
Lemma 3.4. The finite support symmetric group on Sym({0, 1}∗) has presentation
Sym({0, 1}∗) =
*
σx,y
∀x, y ∈ {0, 1}∗
x lex y
2
σx,y
[σx,y , σz,w ]
(σx,y σy,z )3
σx,y σy,z σx,y σx,z
∀(x, y) ∈ Σ2
+
∀(x, y, z, w) ∈ Σ4
∀(x, y, z) ∈ Σ3
∀(x, y, z) ∈ Σ3
where σx,y denotes the transposition of x and y.
Proof. Let x1 , . . . , xn be some collection of elements of {0, 1}∗ with xi lex xi+1
for all 1 ≤ i ≤ n − 1. The finite symmetric group on {x1 , . . . , xn } has a well-known
presentation
σx2i ,xi+1 , (σxi ,xi+1 σxi+1 ,xi+2 )3 ,
Sym({x1 , . . . , xn }) = σx1 ,x2 , . . . , σxn−1 ,xn
[σxi ,xi+1 , σxj ,xj+1 ] ∀ |i − j| ≥ 2,
where σxi ,xj denotes the transposition of xi and xj . We expand this presentation
using Tietze transformations. First add new generators σxi ,xi+2 for each suitable
1 ≤ i ≤ n − 2, each such generator can be added with the relator
σxi ,xi+1 σxi+1 ,xi+2 σxi ,xi+1 σxi ,xi+2
10
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Next add new generators σxi ,xi+3 , each can be added with the relator
σxi ,xi+2 σxi+2 ,xi+3 σxi ,xi+2 σxi ,xi+3
and so on adding σxi ,xi+j for all i and increasing j, until all the required generators
are present. Finally we can add all relators
σxi ,xj σxj ,xk σxi ,xj σxi ,xk ∀i
j
k
which are not already present. We’ve obtained the presentation
Sym({x1 , . . . , xn }) =
*
σxi ,xj ∀ i
j
σx2i ,xj ∀ i j,
(σxi ,xj σxj ,xk )3 ∀ i j k,
[σxi ,xj , σxk ,xl ] ∀ i j k l
σxi ,xj σxj ,xk σxi ,xj σxi ,xk ∀ i j
k.
+
It is now easy to check that this can be extended to the entire finite support
symmetric group Sym({0, 1}∗).
Proposition 3.5 ([CFP96, §3]). Thompson’s group F has presentation
F = hα, β | [αβ −1 , α−1 βα], [αβ −1 , α−2 βα2 ]i.
Theorem 3.6. QF has presentation
2
−1 −1
+
*
σ 2 , [σ, σ α ], (σσ α )3 , σσ α σσ αβ α ,
QF = σ, α, β [αβ −1 , α−1 βα], [αβ −1 , α−2 βα2 ],
[ν, σ] for all ν ∈ X.
where
X=
and σ = σ0,ε .
β, β α , α3 β −1 α−2 , α2 β 2 α−1 β −1 αβ −1 α−2
2 −1 −1
αβ α β αβ −1 α−1 , αβα−1 β 2 α−1 β −1 αβ −1 αβ −1 α−1
Proof. We apply Lemma 3.1 to the description of QF as the semi-direct product
Sym({0, 1}∗) ⋊ F (Lemma 2.5), using the presentation of Sym({0, 1}∗ ) given in
Lemma 3.4, and the presentation of F given in Proposition 3.5.
Let f ∈ F , then necessarily f (x) lex f (y) (see Remark 2.6) and f −1 σx,y f =
σf (x),f (y) . Hence by Lemma 3.2, F acts transitively on the generating set of
Sym({0, 1}∗) and acts with three F -orbits on the relating set.
Let σ = σ0,ε , this is a representative of the generating set. We take the representatives of the relating set,
{σ 2 , [σ, σ1,11 ], (σσε,1 )3 , σσε,1 σσ0,1 }.
Let α and β be the generators of F from Proposition 3.5. We can express the
transpositions appearing in the relating set as
2
σ1,11 = σ α
σε,1 = σ α
σ0,1 = σ αβ α
Since the stabiliser of σx,y for some x lex y is equal to the stabiliser of the pair
(x, y) in Σ2 , Lemma 3.3 gives that the stabiliser of the F -action on the generators
of Sym({0, 1}∗) is a copy of F × F × F , generated by the set
β, β α , α3 β −1 α−2 , α2 β 2 α−1 β −1 αβ −1 α−2
X=
.
αβ 2 α−1 β −1 αβ −1 α−1 , αβα−1 β 2 α−1 β −1 αβ −1 αβ −1 α−1
−1
−1
In summary, with the notation of Lemma 3.1, Xσ = X is as above, S0 = {σ},
SQ = {α, β},
b0 = {σ 2 , [σ, σ α2 ], (σσ α )3 , σσ α σσ αβ −1 α−1 },
R
QUASI-AUTOMORPHISMS OF T2,c
11
and
RQ = {[αβ −1 , α−1 βα], [αβ −1 , α−2 βα2 ]}.
Corollary 3.7.
QF/[QF, QF ] ∼
= Z ⊕ Z ⊕ C2
Proof. Abelianising the relators we find that the only surviving relation is σ 2 .
3.2. A finite presentation for Q̃T . Recall that Q̃T denotes the preimage of T
under π : Q̃V −→ V . In this section we compute a finite presentation for Q̃T .
Let [−, −, −]lex be the cyclic order induced by the ≤lex ordering, so [x, y, z]lex if
and only if x ≤lex y ≤lex z or y ≤lex z ≤lex x or z ≤lex x ≤lex y.
The next lemma is an analogue of Lemma 3.2.
Lemma 3.8. Let Z = {0, 1}∗ ∪ {ζ}. T acts transitively on
(
)
n
Y
Λn = (x1 , . . . , xn ) ∈
Z : [x1 , x2 , . . . , xn ]lex , xi 6= xj for all i 6= j
1
for all n ∈ N, where the action is via the splitting ι : T −→ Q̃T , the inclusion
Q̃T −→ Q̃V , and the usual action of Q̃V on Z.
Proof. As in Lemma 3.2, let yi = 0n−i for all 1 ≤ i ≤ n. We build an element γ of
T such that ι(γ) maps an arbitrary element (x1 , . . . , xn ) ∈ Λn onto (y1 , . . . , yn ).
Let L be the smallest rooted subtree of T2,c such that L contains all the xi as
nodes. Let f be the cyclic permutation of the leaves of L such that the element t
of T represented by (L, L, f ) satisfies ι(t)xi ≤ ι(t)xi+1 for all i.
If ι(t)xn 6= ζ then, via Lemma 3.2, there exists f ∈ F such that ι(f t)xi = yi for
all i.
Assume ι(t)xn = ζ. Let L′ be formed from L by adding a caret on the left-handmost leaf, and let f ′ be the cyclic permutation of the leaves of L′ which sends each
leaf to it’s immediate left-hand neighbour. Let t′ be the element of T represented
by (L′ , L′ , f ′ ), then ι(t′ t)xi lex ι(t′ t)xi+1 , and ι(t′ t)xn 6= ζ. Once again we may
use Lemma 3.2.
Lemma 3.9.
(1) The T -stabiliser of any element in Λn is isomorphic to
(2) The T -stabiliser of (ε, ζ) ∈ Λ2 has generating set
Qn
1
F.
Stab((ε, ζ)) = hβ, β α , α2 β −1 α−1 , αβ 2 α−1 β −1 αβ −1 α−1 i.
Proof. The T -stabiliser Q
of (y1 , . . . , yn−1 , ζ) is exactly the F -stabiliser of (y1 , . . . , yn−1 ),
n
which is isomorphic to 1 F by Lemma 3.3, this proves part (1).
Finally, the stabiliser of the subtree with root 1 is hβ, β α i and that of the subtree
with root 0 is hα2 β −1 α−1 , αβ 2 α−1 β −1 αβ −1 α−1 i.
The proof of the next lemma is identical to that of Lemma 3.4.
Lemma 3.10. The finite support symmetric group on Sym(Z) has presentation
Sym(Z) =
*
σx,y ∀x, y ∈ Λ2
2
σx,y
[σx,y , σz,w ]
(σx,y σy,w )3
σx,y σy,z σx,y σx,z
∀(x, y) ∈ Λ2
+
∀(x, y, z, w) ∈ Λ4
∀(x, y, z) ∈ Λ3
∀(x, y, z) ∈ Λ3
12
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Proposition 3.11 ([CFP96, §5]). Thompson’s group T has presentation
+
*
[αβ −1 , α−1 βα], [αβ −1 , α−2 βα2 ],
−1
−1
−1 −1
−1 −1
−2
2
T = α, β, γ γ βα γβ, α β αβ γ αβα γβ ,
α−1 γ −1 (α−1 γβ)2 , γ 3 ,
Theorem 3.12. Q̃T has presentation
2
Q̃T =
*
σ, α, β, γ
σ 2 , [σ, σ α ], (σσ α )3 , σσ α σσ αβ α ,
+
[αβ −1 , α−1 βα], [αβ −1 , α−2 βα2 ],
−1
−1
−1 −1
−1 −1
−2
2
γ βα γβ, α β αβ γ αβα γβ ,
α−1 γ −1 (α−1 γβ)2 , γ 3 ,
[ν, σ] for all ν ∈ X.
−1
−1
where
X = {β, β α , α2 β −1 α−1 , αβ 2 α−1 β −1 αβ −1 α−1 }
and σ = σ0,ε .
Proof. The proof is similar to that of Theorem 3.6, except using Lemma 3.10 and
Proposition 3.11.
Question 3.13. Is QT finitely presented?
Corollary 3.14.
Q̃T /[Q̃T, Q̃T ] ∼
= C2
Proof. Abelianising the relators kills the generators α, β, and γ, leaving only the
generator σ and the relation σ 2 .
In Corollary 5.3 we show that the abelianisation of QT is also isomorphic to the
cyclic group of order 2.
3.3. A finite presentation for Q̃V . In this section we compute a finite presentation for Q̃V . Recall that Z = {0, 1}∗ ∪ {ζ}. We also define
βn = α−(n−1) βαn−1 for all n ≥ 1,
γn = α−(n−1) γβ n−1 for all n ≥ 1,
δ1 = γ2−1 δγ2 ,
δn = α−(n−1) δ1 αn−1 for all n ≥ 2.
These definitions will allow us to express the presentation of Q̃V in a simpler form.
The same definitions appear in [CFP96, p.13,p.16], where they are called Xn , Cn ,
and πn respectively.
Lemma 3.15. V acts transitively on
(
∆n =
(x1 , . . . , xn ) ∈
n
Y
1
Z : xi 6= xj for all i 6= j
)
,
for all n ∈ N, where the action is via the splitting ι : V −→ Q̃V and the usual
action of Q̃V on Z.
Proof. Let (x1 , . . . , xn ) ∈ ∆n and let L be the smallest subtree of T2,c containing
all the xi as nodes. Choose a bijection f on the leaves of L such that the element
γ of V represented by (L, L, f ) has ι(f )xi ≤lex ι(f )xi+1 . Now use Lemma 3.8.
QUASI-AUTOMORPHISMS OF T2,c
13
Lemma 3.16. The V -stabiliser of (ε, ζ) ∈ ∆2 has generating set
Stab((ε, ζ)) = hα′ , β ′ , γ ′ , δ ′ , λ, µi,
where
α′ = (α2 )(β −1 γ −1 αδα−1 γβ)(α−1 β −1 ),
β ′ = (αβ22 )(β3−1 β2−1 α−1 ),
γ ′ = (αβ2 )(δδ2 δ1 δ)(β2−1 α−1 ),
δ ′ = (αβ2 )(δ1 δ0 δ1 )(β2−1 α−1 ),
λ = α2 β −1 α−1 ,
µ = β.
See Figures 6–11 for representatives of the elements α′ , β ′ , γ ′ , δ ′ , λ, and µ as tree
diagrams.
Between them, the elements α′ , β ′ , γ ′ , and δ ′ generate the subgroup V ′ of V
which fixes the subtrees under 01 and 11.
The method of finding expressions in the generators α, β, γ, δ is adapted from
that in [Bel04, §1], and the applet [Kog08] was used for checking these calculations.
Proof. Let T be a finite rooted subtree of T2,c with at least two leaves and let S
be the rooted subtree with leaves 00, 01, 10, and 11. We denote by T ′ the subtree
obtained by taking the subtree of T with root 0 and glueing 0 to the node 00 of S,
similarly we take the subtree of T with root 1 and glue it to the leaf 10 of S.
If f : leaves(L) → leaves(R) is a bijection between the leaves of finite rooted
subtrees of T2,c then we write f ′ : leaves(L′ ) → leaves(R′ ) for the corresponding
bijection (which fixes the leaves 01 and 11).
The homomorphism
ϕ : V −→ V
(L, R, f ) 7−→ (L′ , R′ , f ′ )
is an injection whose image is V ′ .
Let α′ = ϕ(A), β ′ = ϕ(B), γ ′ = ϕ(C), and δ ′ = ϕ(D) and let λ and µ be the
elements shown in Figures 10 and 11. By construction, α′ , β ′ , γ ′ , and δ ′ generate
V ′.
The elements λ, µ, and the group V ′ stabilise ε and ζ, we claim that they also
generate StabV ((ε, ζ)). Let τ ∈ V be an element of the stabiliser represented by
(L, R, f ), via Lemma 3.17 we may assume that τ fixes the subtrees under 01 and
11, and thus is an element of V ′ .
Finally, we calculate α′ , β ′ , γ ′ , and δ ′ . Each is a product of three elements
of V : the first is an element of F which maps the tree L to a right-vine (a tree
formed from the trivial tree by adding carets to the right-hand most leaf only) , the
second is the necessary permutation of the leaves on the right vine, and the third
is the element of F which maps the right-vine to the tree R. The outcome of these
calculations is as shown in the statement of the lemma. Note that the word length
of these elements may not be minimal.
Let T be a rooted subtree of T2,c and recall the definition of
bT : nodes(T ) ∪ {ζ} −→ leaves(T )
from Section 2.1. We define lε (T ) to be the word length in {0, 1}∗ of the leaf bT (ε)
and define lζ (T ) to be the word length of the leaf bT (ζ).
14
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
ε
00
ε
ζ
01
11
01
100 101
ζ
101
11
00 100
Figure 6. Element α′ = (α2 )(β −1 γ −1 αδα−1 γβ)(α−1 β −1 ) from
Lemma 3.16.
ε
00
ε
ζ
01
11
00
ζ
01
11
100
1011
1010 1011
100 1010
Figure 7. Element β ′ = (αβ22 )(β3−1 β2−1 α−1 ) from Lemma 3.16.
ε
00
01
ε
ζ
11
100
01
100 101
ζ
11
101 00
Figure 8. Element γ ′ = (αβ2 )(δδ2 δ1 δ)(β2−1 α−1 ) from Lemma 3.16.
ε
00
01
11
100 101
ε
ζ
100
01
ζ
11
00 101
Figure 9. Element δ ′ = (αβ2 )(δ1 δ0 δ1 )(β2−1 α−1 ) from Lemma 3.16.
Lemma 3.17. Let v ∈ V satisfy ι(v)(ζ) = ζ and ι(v)(ε) = ε. There exist nonnegative integers a, b, c, and d such that ι(λ−a µ−b vλc µd ) fixes the subtrees under
01 and 11.
QUASI-AUTOMORPHISMS OF T2,c
ε
15
ε
ζ
1
ζ
1
10
011
010 011
10 010
Figure 10. Element λ = α2 β −1 α−1 from Lemma 3.16.
ε
ε
ζ
0
ζ
0
10
111
110 111
10 110
Figure 11. Element µ = β from Lemma 3.16.
Proof. Note that ι(λ)(ε) = ι(µ)(ε) = ε and ι(λ)(ζ) = ι(µ)(ζ) = ζ. To prove the
lemma, it suffices to find non-negative integers a, b, c, and d and tree diagram
representative (L, R, f ) for λ−a µ−b vλc µd such that
lε (L) = lε (R) = lζ (L) = lζ (R) = 2.
Consider the element λn for some non-negative integer n. This element has tree
diagram representative (L, R, f ), where lε (L) = n + 2, lε (R) = 2, lζ (L) = 1, and
lζ (R) = 1. Similarly consider the element µn for some non-negative integer n. This
element has tree diagram representative (L, R, f ), where lζ (L) = n + 2, lζ (R) = 2,
lε (L) = 1, and lε (R) = 1.
Let v ∈ V be such that ι(v)(ζ) = ζ and ι(v)(ε) = ε and let (L, R, f ) be a tree
diagram representative for v. If either lζ (L), lζ (R), lε (L), or lε (R) are strictly less
than 2 than expand L and R by adding carets until they are equal or greater than
2.
Let a = lε (L) − 2, b = lζ (L) − 2, c = lε (R) − 2, and d = lζ (R) − 2. Then, one
calculates that λ−a µ−b vλc µd has tree diagram representative (L, R, f ) with
lε (L) = lε (R) = lζ (L) = lζ (R) = 2,
as required.
Modifying Lemma 3.10 slightly we obtain the following.
Lemma 3.18. The finite support symmetric group on Sym(Z) has presentation
Sym(Z) =
*
σx,y ∀x, y ∈ ∆2
σx,y = σy,x
2
σx,y
[σx,y , σz,w ]
(σx,y σy,w )3
σx,y σy,z σx,y σx,z
∀(x, y) ∈ ∆2
+
∀(x, y) ∈ ∆2
∀(x, y, z, w) ∈ ∆4
∀(x, y, z) ∈ ∆3
∀(x, y, z) ∈ ∆3
16
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Proposition 3.19 ([CFP96, p.18]). Thompson’s group V has presentation
+
*
[αβ −1 , β2 ], [αβ −1 , β3 ], βγ2 γ1−1 , βγ3 (γ2 β2 )−1 , γ22 (γ1 α)−1
3
2
−1
3
−1
.
V = α, β, γ, δ : γ1 , δ1 , δ3 δ1 (δ1 δ3 ) , (δ2 δ1 ) , δ1 β3 (β3 δ1 ) ,
βγ2 γ1 (γ1 β2 )−1 , βδ3 (δ2 β)−1 , γ3 δ2 (δ1 γ3 )−1 , (δ1 γ2 )3
Theorem 3.20. Q̃V has presentation
2
σσ αδα , σ 2 , [σ, σ α ], (σσ α )3 , σσ α σσ αβ α ,
+
[αβ −1 , β2 ], [αβ −1 , β3 ], βγ2 γ1−1 , βγ3 (γ2 β2 )−1 , γ22 (γ1 α)−1
γ13 , δ12 , δ3 δ1 (δ1 δ3 )−1 , (δ2 δ1 )3 , δ1 β3 (β3 δ1 )−1 ,
βγ2 γ1 (γ1 β2 )−1 , βδ3 (δ2 β)−1 , γ3 δ2 (δ1 γ3 )−1 , (δ1 γ2 )3
[ν, σ] for all ν ∈ X.
−1
Q̃V =
*
σ, α, β, γ, δ
where
X=
−1
−1
(α2 )(β −1 γ −1 αδα−1 γβ)(α−1 β −1 ), (αβ22 )(β3−1 β2−1 α−1 ),
(αβ2 )(δδ2 δ1 δ)(β2−1 α−1 ), (αβ2 )(δ1 δ0 δ1 )(β2−1 α−1 ), α2 β −1 α−1 , β
and σ = σ0,ε .
Proof. As in Theorems 3.6 and 3.12, except using the presentations from Lemma
3.18 and 3.19.
Explicitly, we have:
SQ = {α, β, γ, δ}
[αβ −1 , β2 ], [αβ −1 , β3 ], βγ2 γ1−1 , βγ3 (γ2 β2 )−1 , γ22 (γ1 α)−1
RQ =
γ 3 , δ 2 , δ3 δ1 (δ1 δ3 )−1 , (δ2 δ1 )3 , δ1 β3 (β3 δ1 )−1 ,
1 1
βγ2 γ1 (γ1 β2 )−1 , βδ3 (δ2 β)−1 , γ3 δ2 (δ1 γ3 )−1 , (δ1 γ2 )3
S0 = {σ} = {σ0,ε }
2
R̂0 = {σσ αδα , σ 2 , [σ, σ α ], (σσ α )3 , σσ α σσ αβ α }
(α2 )(β −1 γ −1 αδα−1 γβ)(α−1 β −1 ), (αβ22 )(β3−1 β2−1 α−1 ),
Xσ =
(αβ2 )(δδ2 δ1 δ)(β2−1 α−1 ), (αβ2 )(δ1 δ0 δ1 )(β2−1 α−1 ), α2 β −1 α−1 , β.
−1
−1
−1
Compared to the calculation for Q̃T , there is one new element of R0 , namely
−1
σσ αδα which corresponds to the relation
{σx,y = σy,x for all x, y ∈ ∆2 }.
Corollary 3.21.
Q̃V /[Q̃V, Q̃V ] ∼
= C2
Proof. Abelianising the presentation leaves the generator σ and the relator σ 2 .
In Corollary 5.3 we show that the abelianisation of QV is also isomorphic to C2 .
Question 3.22 ([Leh08, p.31]). Is QV finitely presented?
4. Type F∞
In this section we show that the groups QF , Q̃T and Q̃V are of type F∞ . The
idea of the proof for QF is to consider Sym({0, 1}∗) as a countably generated
Coxeter group then to form the Davis complex U (a certain contractible CWcomplex on which Sym({0, 1}∗) acts properly), and then extend this to an action of
QF on U which has stabilisers of type F∞ . For Q̃T and Q̃V we substitute Sym(Z)
for Sym({0, 1}∗ ).
QUASI-AUTOMORPHISMS OF T2,c
17
Let (W, S) be a countably generated Coxeter group, so W is generated by a
countable set of involutions S. We start by giving a quick overview of the construction of the Davis complex of a Coxeter group, for background see [Dav08, §5, §7].
After this we show that if Q is a group acting by automorphisms on W such that
q(S) = S for all q ∈ Q then there is an action of W ⋊ Q on the Davis complex,
where the semi-direct product is formed using the given action of Q on W . This is
already well-known, see for example [Dav08, §9.1], so we only give an overview.
For any subset T of S we denote by WT the subgroup of W generated by T .
Recall that a spherical subset is a finite subset T of S for which WT is finite. The
group WT is known as a spherical subgroup. Let S denote the poset of spherical
subsets in (W, S) and C the poset of cosets of spherical subgroups, thus elements of
C may be written as wWT for w ∈ W and T ∈ S. The poset C admits a left action
by W ,
w′ · wWT = (w′ w)WT .
The Davis Complex U is the geometric realisation of C and thus admits a left action
by W as well. Note that the W -orbits of n-simplices in U are (n+1)-element subsets
of S generating a finite subgroup of W [DL98, p.2].
Proposition 4.1. [DL98, p.3] The Davis complex U is contractible.
Consider a group Q acting by automorphisms on W such that every q ∈ Q
satisfies q(S) = S. We denote by G the semi-direct product W ⋊ Q formed using
this action.
The action of Q on S extends to an action of Q on S, by setting
q{s1 , . . . , sn } = {qs1 , . . . , qsn }.
We use this to define a G-action on C by
(h, q) · wWT = hwq
−1
WqT .
One checks that this is well-defined and preserves the poset structure. The G-action
on C induces a G-action on U.
The next lemma appears in [Dav08, Propostion 9.1.9], as does the first part of
Proposition 4.3.
Lemma 4.2. The W ⋊ Q-isotropy subgroup of wWT ∈ C is (WT )w
QT is the Q-isotropy of T ∈ S.
−1
⋊ QT , where
Let Sn denote the set of unordered n-element subsets of S, equivalently the
spherical subsets of size n. The set Sn admits a Q-action, the restriction of that on
S.
Proposition 4.3. If, for all positive integers n, Q acts on Sn with finitely many
Q-orbits and with stabilisers of type F∞ , then G is of type F∞ .
Proof. Since the set of W -orbits of n-simplices in U is exactly the set Sn+1 , there are
as many G-orbits of n-simplicies in U as Q-orbits in Sn+1 . Thus G acts cocompactly
on U.
By construction, the Q-isotropy of any point in U is the Q-isotropy of some
spherical subgroup WT . Thus assumption (2) implies that the Q-stabiliser of any
point has type F∞ , so combining with Lemma 4.2 and [Bie81, Proposition 2.7], the
G-stabiliser of any point in U is F∞ . Theorem 4.4 below completes the proof.
Theorem 4.4. [Geo08, Theorem 7.3.1] If there exists a contractible G-CW complex
which is finite type mod G and has stabilisers of type F∞ , then G is of type F∞ .
18
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
At this point we specialise to the Coxeter group Sym(Z). Let S be the set of
unordered 2-element subsets of Z. This is the set of generators of Sym(Z) given
in Lemma 3.10. Recall from Section 3.3 that ∆n is the set of ordered n-element
subsets of Z, so there is a surjective V -map ∆2 → S, given by projection. Recall
that Sn denotes the set of unordered n-element subsets of elements of S (note this
is not equivalent to the set of unordered 2n-element subsets of S). There is also a
surjective V -map ∆2n → Sn .
Lemma 4.5. The V -stabiliser of {(x1 , y1 ), . . . , (xn , yn )} ∈ Sn is of type F∞ for all
{(x1 , y1 ), . . . , (xn , yn )} ∈ Sn .
Proof. Since V acts transitively on Sn it is sufficient to check that the stabiliser
StabV ({(02n−2 , 02n−1 ), . . . , (ε, ζ)}) is F∞ .
Let {(02n−2 , 02n−1 ), . . . , (ε, ζ)} ∈ Sn , and let A be the preimage of this element
under the V -map ∆2n → Sn . There is an homomorphism π : StabV (A) → Sym2n
which records the permutation on the elements {02n−2 , . . . , 0, ε, ζ}. The kernel of
π is exactly StabV ({(02n−2 , 02n−1 ), . . . , (ε, ζ)}) which is F∞ by Corollary 4.17 so
combining with [Bie81, Proposition 2.7] and the fact that Symn is finite and hence
F∞ , we deduce that StabV (A) is F∞ .
Observing that StabV (A) = StabV ({(02n−2 , 02n−1 ), . . . , (ε, ζ)}) completes the
proof for the V -stabiliser.
Theorem 4.6. The group Q̃V is of type F∞ .
Proof. We prove the statement for Q̃V first, using Proposition 4.3. Since V acts
transitively on ∆2n , it acts transitively on Sn also. The stabilisers are of type F∞
is Lemma 4.5.
Next, we will prove the group Q̃T is F∞ , for which we require the following
technical lemma.
Lemma 4.7. T acts with finitely many orbits and stabilisers of type F∞ on ∆n .
Proof. Let Cycn denote the subgroup of Symn generated by the cyclic permutation
(1, 2 . . . , n). Let R be a set of representatives of the cosets Symn / Cycn . For any
σ ∈ R, we define a T -map
iσ : Λn −→ ∆n
(x1 , . . . , xn ) 7−→ (xσ(1) , . . . , xσ(n) ).
Taking a product of these gives an T -map,
`
a
iσ
Λn −→ ∆n ,
σ∈R
which one checks is a bijection. Since the T -stabiliser of any element of Λn is F∞
(Lemma 3.9), the T -stabiliser of any element of ∆n is also.
Theorem 4.8. The group Q̃T is of type F∞ .
Proof. Again, we let S be the unordered 2-element subsets of Z. Using the argument
of Lemma 4.5 we deduce from Lemma 4.7 that T acts with finitely many orbits and
with stabilisers of type F∞ on Sn . The statement follows from Proposition 4.3.
To prove that QF is of type F∞ , we will again use Proposition 4.3, but now we
are considering the F -action on the symmetric group Sym({0, 1}∗) so we let S be
the unordered 2-element subsets of {0, 1}∗. Thus the correct analogue of Lemma
4.7 is the following.
QUASI-AUTOMORPHISMS OF T2,c
19
Lemma 4.9. F acts with finitely many orbits and stabilisers of type F∞ on
n
Y
Xn = {(x1 , . . . , xn ) ∈
{0, 1}∗ : xi 6= xj ∀i 6= j}.
1
Proof. The proof is similar to that of Lemma 4.7, we construct a bijection of F -sets
`
a
iσ
Σn −→ Xn ,
σ∈Symn
where iσ is the map
iσ : Σn −→ Xn
(x1 , . . . , xn ) 7−→ (xσ(1) , . . . , xσ(n) ).
Now, since the F -stabiliser of any element of Σn is F∞ (Lemma 3.3), the F -stabiliser
of any element of Xn is also.
Theorem 4.10. The group QF is of type F∞ .
Proof. Using the argument of Lemma 4.5 we deduce from Lemma 4.9 that F acts
with finitely many orbits and stabilisers of type F∞ on Sn . The statement now
follows from Proposition 4.3.
4.1. The group StabV ((0n−2 , . . . , 0, ε, ζ)) is of type F∞ . In this section we study
the action of V on ∆n and prove in Corollary 4.17 that it acts with stabilisers of
type F∞ .
Let T be a rooted subtree of T2,c and recall the definition of
bT : nodes(T ) ∪ {ζ} −→ leaves(T )
from Section 2.1. For x ∈ {0, 1}∗ we define lx (T ) to be the word length in {0, 1}∗
of the leaf bT (x). This definition has already been seen in Section 3.3, just before
Lemma 3.17.
Let v ∈ V satisfy ι(v)(x) = x and let (L, R, f ) be a representative for v. We
define
χ̃x (v) = lx (L) − lx (R).
Lemma 4.11. For any x ∈ Z, the map χ̃x is a group homomorphism StabV (x) →
Z.
Proof. Since χ̃x is defined only on StabV (x), the values of χ̃x are invariant under
adding carets. Thus, since any two tree diagram representatives of an element of
V have a common expansion, values taken by χ̃x don’t depend on the tree diagram
representative. Let v, w be any two elements of StabV (x), let (L, R, f ) be a tree
diagram representative for v and let (R, S, g) be a tree diagram representative for
w. Then,
χ̃x (vw) = lx (L) − lx (S) = lx (L) − lx (R) + lx (R) − lx (S) = χ̃x (f ) + χ̃x (g).
Let T1 be the tree with 2 leaves and let Tn be obtained from Tn−1 by adjoining
carets to the first and second leaves (measuring from smallest to largest using
≤lex ). In particular, Tn has 2n leaves. See Figure 12 for a picture of T3 . Let
Ln be the subgroup of V consisting of elements which fix the full subtrees with
root the 2ith -leaf of Tn for all 1 ≤ i ≤ n. For example, the group L3 is the
subgroup of V fixing the full subtrees under 001, 011, and 11. One checks that
Ln ∈ StabV ((0n−2 , . . . , 0, ε, ζ)).
Lemma 4.12. The group Ln is isomorphic to V for all n ≥ 2.
20
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
ε
ζ
0
00
000
1
01
001 010
10
11
011
Figure 12. The tree T3 from the proof of Lemma 4.12.
Proof. The group Ln is a copy of V2,n , Thompson’s group V but on the disjoint
union of n-trees. These trees are exactly the full subtrees with roots the (2i − 1)th leaves of Tn for each 1 ≤ i ≤ n. There is a result of Higman that V2,n is isomorphic
to V [Hig74].
For 1 ≤ i ≤ n, let λn,i be the element of F with tree diagram representative
(Ln,i , Rn,i ) where Ln,i is obtained from Tn by adjoining a caret to the 2ith -leaf,
and Rn,i is obtained from Tn by adjoining a caret to the (2i − 1)th -leaf of Tn . For
example, λ2,1 and λ2,2 are the elements λ and µ of Section 3.3 and λ3,2 is shown in
Figure 13.
One calculates that:
Lemma 4.13. For all positive integers n, all integers i such that 1 ≤ i ≤ n − 2,
and all integers j with 1 ≤ j ≤ n,
χ̃ζ : Ln −
7 → 0,
χ̃ε : Ln 7−→ 0,
χ̃0i : Ln 7−→ 0,
χ̃ζ : λn,j 7−→
χ̃ε : λn,j 7−→
χ̃0i : λn,j 7−→
1
0
if j = n,
else,
1
0
if j = n − 1,
else,
1
0
if j = n − 1 − i,
else.
Lemma 4.14. Let v ∈ StabV ((0n−2 , . . . , ε, ζ)). For 1 ≤ i ≤ n, there exist non−a1
−an
n
1
· · · λn,n
negative integers ai and bi such that the element λn,1
vλbn,n
is contained
λbn,1
in Ln .
Proof. In the case n = 2 this is Lemma 3.17, for n > 2 the proof is similar.
Lemma 4.15. The group StabV ((0
for i = 1, . . . , n.
n−2
,0
n−1
, . . . , ε, ζ)) is generated by Ln and λn,i
Proof. This follows from Lemma 4.14.
If G is a group, θ : G → G an endomorphism, and λ̃ any symbol, then we denote
by G∗λ̃,θ the corresponding HNN-extension. Recall that an HNN-extension is said
to be ascending if θ is injective.
Proposition 4.16.
StabV ((0n−2 , 0n−2 , . . . , ε, ζ)) ∼
= Ln ∗λ̃1 ,θ1 ∗λ̃2 ,θ2 · · · ∗λ̃n ,θn ,
where the injective homomorphisms θi correspond to conjugation by λ−1
n,i .
QUASI-AUTOMORPHISMS OF T2,c
21
ε
ζ
10
000
11
001 010
0110
0111
ε
ζ
10
000
001
11
0111
010
0110
Figure 13. λ3,2
Proof. Let θi : Ln → Ln be the injective homomorphism v 7→ λn,i vλ−1
n,i , and let
Φ : (V ′ ∗µ̃,θ1 )∗λ̃i ,θ2 −→ StabV ((0n−2 , 0n−1 , . . . , ε, ζ)),
−1
be the map specialising λ̃i to λn,i
. Since Ln together with the λn,i generate
StabV ((0n−2 , . . . , ε, ζ)) (Lemma 4.15), the map Φ is a surjection. Let v ∈ Ker Φ,
since the HNN-extensions are ascending, we can write
v = λ1−a1 λ2−a2 · · · λn−an wλbnn · · · λb11 ,
for some w ∈ Ln and some non-negative integers ai and bi for 1 ≤ i ≤ n.
For any fixed i, the homomorphism
Ln ∗λ̃1 ,θ1 ∗λ̃2 ,θ2 · · · ∗λ̃n ,θn −→ Z
Ln 7−→ 0
1
λ̃j 7−→
0
if i = j,
else,
factors through StabV ((0n−2 , . . . , ε, ζ)) using Lemma 4.13, thus ai = bi for all i.
In particular, w ∈ Ker Φ, thus w = 1, since Φ is the identity when restricted to
Ln .
Corollary 4.17. Given (x1 , . . . , xn ) ∈ ∆n , the group StabV ((x1 , . . . , xn )) is of type
F∞ .
22
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Proof. Since V acts transitively on ∆n it is sufficient to check that the stabiliser
StabV ((0n−2 , . . . , ε, ζ)) is F∞ . This follows since by Lemma 4.12 Ln is isomorphic
to V and hence F∞ and HNN-extensions of groups of type F∞ are themselves F∞
[Bie81, Proposition 2.13(b)].
5. Normal subgroups
In this section we give a complete description of the normal subgroups of QF ,
QT , QV , Q̃T , and Q̃V .
Recall that Sym({0, 1}∗ ) has only one non-trivial proper normal subgroup, this is
the finite support alternating group Alt({0, 1}∗ ), the subgroup of all permutations
with even parity. Moreover, Sym({0, 1}∗) and Alt({0, 1}∗ ) are normal in the infinite
support symmetric group on {0, 1}∗, because conjugating any finite permutation by
any bijection {0, 1}∗ → {0, 1}∗ preserves the cycle type of the finite permutation.
Recall also that F/[F, F ] ∼
= Z⊕Z but, since T and V are simple and non-abelian,
[T, T ] ∼
= T and [V, V ] ∼
= V [CFP96, Theorems 4.1,5.8,6.9].
Theorem 5.1 (Normal subgroups of QF ).
(1) A non-trivial normal subgroup of QF is either Alt({0, 1}∗), Sym({0, 1}∗),
or contains
[QF, QF ] = Alt({0, 1}∗) ⋊ [F, F ].
(2) A proper non-trivial normal subgroup of Q̃T is either Alt(Z), Sym(Z), or
[Q̃T, Q̃T ] = Alt(Z) ⋊ T.
(3) A proper non-trivial normal subgroup of Q̃V is either Alt(Z), Sym(Z), or
[Q̃V, Q̃V ] = Alt(Z) ⋊ V.
(4) A proper non-trivial normal subgroup of QT is one of either Alt({0, 1}∗),
Sym({0, 1}∗), or
[QT, QT ] = (Alt(Z) ⋊ T ) ∩ QT.
Moreover, (Alt(Z) ⋊ T ) ∩ QT is an extension of T by Sym({0, 1}∗).
(5) A proper non-trivial normal subgroup of QV is one of either Alt({0, 1}∗),
Sym({0, 1}∗), or
[QV, QV ] = (Alt(Z) ⋊ V ) ∩ QV.
Moreover, (Alt(Z) ⋊ V ) ∩ QV is an extension of V by Sym({0, 1}∗ ).
This theorem is similar to results for the braided Thompson groups Fbr and Vbr .
The braided Thompson group Fbr is a split extension with kernel the group Pbr ,
a direct limit of finite pure braid groups, and quotient F . The braided Thompson
group Vbr is a non-split extension with kernel Pbr and quotient V . Zaremsky shows
that for any normal subgroup N of Fbr , either N ≤ Pbr or [Fbr , Fbr ] ≤ N [Zar14,
Theorem 2.1], and a proper normal subgroup N of Vbr is necessarily contained in
Pbr [Zar14, Corollary 2.8]. There is no braided version of Thompson’s group T . We
borrow some of the methods in this section from [Zar14, §2].
The next proposition should be compared with [Ore51, Theorem 5], that every
element of the full permutation group on an infinite set is a commutator, so in
particular the commutator subgroup of the full permutation group on {0, 1}∗ is
itself.
We write hhHiiG for the normal closure in G of a subgroup H ≤ G, and for an
element g ∈ G we write hhgiiG for the normal subgroup generated by g.
Proposition 5.2.
(1) hh[ιF, ιF ]iiQF = [QF, QF ] = Alt({0, 1}∗) ⋊ [F, F ].
QUASI-AUTOMORPHISMS OF T2,c
23
(2) hh[ιF, ιF ]iiQ̃T = [Q̃T, Q̃T ] = Alt(Z) ⋊ T.
(3) hh[ιF, ιF ]iiQ̃V = [Q̃V, Q̃V ] = Alt(Z) ⋊ V.
(4) hh[ιF, ιF ]iiQT = [QT, QT ] = (Alt(Z) ⋊ T ) ∩ QT , moreover this subgroup is
an extension with kernel Sym({0, 1}∗ ) and quotient T .
(5) hh[ιF, ιF ]iiQV = [QV, QV ] = (Alt(Z) ⋊ V ) ∩ QV , moreover this subgroup is
an extension with kernel Sym({0, 1}∗ ) and quotient V .
Proof. For part (1), consider [α, β] ∈ [ιF, ιF ] then for σ1,11 ∈ Sym({0, 1}∗), we have
σ1,11 [α, β]−1 σ1,11 [α, β] ∈ hh[ιF, ιF ]ii.
Calculating explicitly,
σ1,11 [α, β]−1 σ1,11 [α, β] = σ1,11,1110 .
The smallest normal subgroup of Sym({0, 1}∗ ) containing a 3-cycle is Alt({0, 1}∗)
so Ker π|hh[ιF,ι,F ]ii = Alt({0, 1}∗ ). Combining this with the fact that [F, F ] ≤
π(hh[ιF, ιF ]ii) gives
Alt({0, 1}∗) ⋊ [F, F ] ≤ hh[ιF, ιF ]ii.
Since hh[ιF, ιF ]iiQF ≤ [QF, QF ], we know that Alt({0, 1}∗) ⋊ [F, F ] ≤ [QF, QF ]
and we now show they are equal.
Consider the following map induced by π,
π ′ : QF/(Alt({0, 1}∗) ⋊ [F, F ]) −→ F/[F, F ].
Applying π ′ to the quotient [QF, QF ]/(Alt({0, 1}∗)⋊ [F, F ]) gives the trivial group.
Since the kernel of π ′ is contained in Sym({0, 1}∗)/ Alt({0, 1}∗ ) = C2 , the quotient
[QF, QF ]/(Alt({0, 1}∗ ) ⋊ [F, F ]) is either trivial or C2 . However, if the quotient
[QF, QF ]/(Alt({0, 1}∗ ) ⋊ [F, F ]) was C2 then [QF, QF ] = Sym({0, 1}∗ ) ⋊ [F, F ]
which would contradict Corollary 3.7. Hence [QF, QF ] = Alt({0, 1}∗ ) ⋊ [F, F ].
This completes the proof of (1).
For parts (2) and (3) substitute Corollary 3.14 or Corollary 3.21 for Corollary
3.7 and use that [T, T ] ∼
= T and [V, V ] ∼
= V [CFP96, Theorems 5.8,6.9].
For part (5), by the argument at the beginning of this proof, Alt({0, 1})∗ is contained in hh[ιF, ιF ]iiQV and hh[ιF, ιF ]iiQV cannot contain any finite cycle with odd
parity, else hh[ιF, ιF ]iiQ̃V would also, and we already know that it doesn’t. Also,
hh[ιF, ιF ]iiQV projects onto a non-trivial normal subgroup of V , so it must project
onto V . We conclude that hh[ιF, ιF ]iiQV is an extension of V by Alt({0, 1}∗).
Observe that,
hh[ιF, ιF ]iiQV ≤ hh[ιF, ιF ]iiQ̃V ∩ QV = (Alt(Z) ⋊ V ) ∩ QV.
Since (Alt(Z) ⋊ V ) ∩ QV is also an extension of V by Alt({0, 1}∗), we necessarily
have
hh[ιF, ιF ]iiQV = (Alt(Z) ⋊ V ) ∩ QV.
The proof of (4) is similar to that of (5).
Corollary 5.3. The abelianisation of QF is isomorphic Z⊕Z⊕C2 and the abeliansations of QT , QV , Q̃T , and Q̃V are all isomorphic to C2 .
Lemma 5.4.
(1) Let σ ∈ Sym({0, 1}∗ ) and 1 6= f ∈ ιF be arbitrary, then there exists h ∈ ιF
such that [h, σ] = 1 but [h, f ] 6= 1.
(2) Let σ ∈ Sym(Z) and 1 6= v ∈ ιV be arbitrary, then there exists h ∈ ιF such
that [h, σ] = 1 by [h, f ] 6= 1.
24
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
Proof. Let T be the full sub-tree under some x ∈ {0, 1}∗, where T is chosen to
intersect the support of f but not the support of σ, this is possible since the
support of f is infinite but the support of σ is finite. Thus any h with supp(h) ⊆ T
satisfies [h, σ] = 1.
Let F ′ denote the copy of F acting on T , with the usual generators denoted α′
and β ′ . If f restricts to an action on T then, since F ′ is centreless, there exists
some h ∈ F ′ which doesn’t commute with f . If f doesn’t restrict to an action on
T then there is necessarily an orbit {xi }i∈Z which is neither contained in T nor its
complement. Since α′ acts non-trivially on all of T , one of the α′ orbits intersects
{xi }i∈Z but is not equal to {xi }i∈Z , thus α′ and h do not commute.
The proof of (2) is analagous.
Lemma 5.5. Let G be one of QF , QT , QV , Q̃T , and Q̃V , and let f ∈ G such
that π(f ) 6= 1, then hhf iiG ∩ ιF 6= {id}.
Proof. Let σf ∈ QF , so that σ ∈ Sym({0, 1}∗) and f ∈ ιF with π(f ) 6= 1. Let
h ∈ ιF be as in Lemma 5.4(1), then
σhf h−1 = hσf h−1 ∈ hhσf ii.
So,
[f, h] = (σf )−1 σhf h−1 ∈ hhσf ii,
since [f, h] is a non-trivial element of ιF , this is sufficient. This proves the statement
for QF .
For G = Q̃T or Q̃V , start with an element σt ∈ Q̃T (respectively σv ∈ Q̃V ), so
that σ ∈ Sym(Z) and t ∈ ιT (resp. t ∈ ιV ). Let h ∈ ιF be as in Lemma 5.4(2),
then the proof is as for QF .
For G = QT or QV , start with σt ∈ QT (respectively σv ∈ QV ), so that
σ ∈ Sym({0, 1}∗) and t ∈ ιT (resp. t ∈ ιV ). Let h ∈ ιF be as in Lemma 5.4(1),
then the proof is as for QF .
Proof of Theorem 5.1. Let N be a normal subgroup of QF with N 6≤ Sym({0, 1}∗),
then using Lemma 5.5 N contains an element f ∈ ιF . For any g ∈ ιF , we have
g −1 f −1 g ∈ N and hence
[g, f ] = g −1 f −1 gf ∈ N.
Without loss of generality, f ∈ [ιF, ιF ].
Since [ιF, ιF ] is simple [CFP96, Theorem 4.5], we deduce that [ιF, ιF ] ≤ N , and
thus hh[ιF, ιF ]ii ≤ N . Proposition 5.2(1) completes the proof.
For the other parts, start with an element f of either Q̃T , Q̃V , QT , or QV and
then use the appropriate parts of Lemmas 5.5 and Proposition 5.2.
6. Bieri–Neumann–Strebel–Renz invariants
In this section we compute the Bieri–Neumann–Strebel–Renz invariants Σi (QF )
and Σi (QF, R) for any commutative ring R. The invariant Σ1 (G) was introduced
by Bieri, Neumann, and Strebel in [BNS87] and the higher invariants Σi (QF ) for
i ≥ 2 were introduced in [BR88]. In general, for any group G, there is a hierachy
of invariants
Σ1 (G) ⊇ Σ2 (G) ⊇ · · · ⊇ Σi (G) ⊇ · · · ,
T∞
furthermore we set Σ∞ (G) = i=1 Σi (G). There are also homological versions
Σi (G, R) for any commutative ring R fitting into a similar hierachy. Furthermore,
Σi (G) ⊆ Σi (G, R) for all i all commutative rings R and also Σ1 (G, R) = Σ1 (G) for
all commutative rings R.
QUASI-AUTOMORPHISMS OF T2,c
25
Some of the interest in the Bieri–Neumann–Strebel–Renz invariants comes from
the following theorem, which classifies the finiteness length of normal subgroups N
of a group G which contain the commutator subgroup [G, G].
Theorem 6.1 ([BR88, Theorem B][Ren89]). Let G be a group of type Fn (respectively FPn over R) and let N be a normal subgroup containing the commutator
[G, G]. Then N is Fn (resp. FPn over R) if and only if every non-zero character
χ of G such that χ(N ) = 0 satisfies [χ] ∈ Σn (G) (resp. [χ] ∈ Σn (G, R)).
Recall that a character of a group G is a group homomorphism G → R, where
R is viewed as a group under addition, and the character sphere S(G) is the set of
equivalence classes of non-zero characters modulo multiplication by a positive real
number. Thus S(G) is isomorphic to a sphere of dimension n − 1 where n is the
torsion-free rank of G/[G, G].
Since F/[F, F ] ∼
= Z ⊕ Z, the character sphere S(F ) is isomorphic to S 1 . We
denote by χ0 and χ1 the two characters of F given by χ0 (A) = −1, χ0 (B) = 0,
χ1 (A) = 1, and χ1 (B) = 1, these two characters are linearly independent and hence
do not represent antipodal points on S(F ).
Theorem 6.2 (The Bieri–Neumann–Strebel–Renz invariants of F [BGK10]). The
character sphere S(F ) is isomorphic to S 1 and, for any commutative ring R,
(1) Σ1 (F, R) = Σ1 (F ) = S(F ) \ {[χ0 ], [χ1 ]}.
(2) Σi (F, R) = Σi (F ) = S(F ) \ {[aχ0 + bχ1 ] : a, b ≥ 0} for all i ≥ 2.
Since the abelianisation of QF also has torsion-free rank 2, the character sphere
S(QF ) is again isomorphic to S 1 . By pre-composing χ0 and χ1 with π : QF → F
we obtain characters π ∗ χ0 and π ∗ χ1 . Once again these are linearly independent.
In this section we prove the following theorem.
Theorem 6.3 (The Bieri–Neumann–Strebel–Renz invariants of QF ). The character sphere S(QF ) is isomorphic to S 1 and, for any ring R,
(1) Σ1 (QF, R) = Σ1 (QF ) = S(QF ) \ {[π ∗ χ0 ], [π ∗ χ1 ]}.
(2) Σi (QF, R) = Σi (QF ) = S(QF ) \ {[aπ ∗ χ0 + bπ ∗ χ1 ] : a, b ≥ 0} for all i ≥ 2.
Remark 6.4 (ν-symmetry). For x ∈ {0, 1}∗, let x̄ be the element obtained by
swapping 0 and 1 in the word x. The map x 7→ x̄ induces an automorphism
ν : QV → QV , which projects to the automorphism of F denoted ν in [BGK10,
§1.4]. Furthermore, ν induces an automorphism ν ∗ of Σi (QF ) for all i, and one
checks that ν ∗ [χ0 ] = [χ1 ]. Following Bieri, Geoghegan, and Kochloukova we will
call this the ν-symmetry of Σi (QF ). Recall that for any group G the invariants
Σi (G) and Σi (G, R) are invariant under automorphisms of G.
Lemma 6.5 ([Mei97, Corollary 2.8][Mei96, Corollary 3.12]). Let π : G −։ Q be
an split surjection and χ a character of Q. Then if [π ∗ χ] ∈ Σi (G) (respectively
Σi (G, R)) then [χ] ∈ Σi (Q) (resp. Σi (Q, R)).
Theorem 6.6 ([BGK10, Theorems 2.1(1), 2.3]). Let H be a group of type F∞ and
let G = H∗θ,t be an ascending HNN-extension such that θ is not surjective.
(1) Let χ : G → R be the character given by χ(H) = 1 and χ(t) = 1, then
[χ] ∈ Σ∞ (G).
(2) Let χ : G → R be a character such that χ|H 6= 0 and [χ|H ] ∈ Σ∞ (H) then
[χ] ∈ Σ∞ (G), for any i ≥ 0.
Recall that the characters π ∗ χi are completely determined by the values they
take on α and β (defined in Remark 2.4) and satisfy π ∗ χi (α) = χi (A) and π ∗ χi (β) =
χi (B) for all i, where A and B are the standard generators of F .
26
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
We now show that QF can be written as an ascending HNN-extension similarly to
those of F [BG84, Proposition 1.7] and BF [Zar14, Lemma 1.4].
Let QF (1) be the subgroup of QF which fixes all of {0, 1}∗ except the subtree
with root 1. It is easy to see that QF (1) is isomorphic to QF . Using Theorem 3.6 we
see that QF is generated by {α, β, σε,1 }. This shows that QF (1) is generated by β,
α
β α and σ1,11 = σε,1
. Moreover, conjugating QF (1) by α maps QF (1) isomorphically
onto the subgroup of QF which fixes all but the subtree with root 11.
Lemma 6.7. The group QF can be written as an ascending HNN-extension QF =
QF (1)∗θ,t, where QF (1) ∼
= QF .
The proof is similar to [Zar14, Lemma 1.4].
Proof. Let θ : QF (1) → QF (1) be the monomorphism τ 7→ α−1 τ α and let ψ :
QF (1)∗θ,t → QF be the map given by setting t to be α. The map ψ is surjective
t−1
since ψ(β) = β, ψ(t) = α, and ψ(σ1,11
) = σε,1 .
Let g ∈ Ker ψ, since QF (1)∗θ,t is an ascending HNN-extension we can write g in
normal form as g = tn ht−m where n, m ≥ 0 and h ∈ QF (1). Let χ : QF (1)∗θ,t −→
Z be the homomorphism sending t to 1 and sending QF (1) to 0. The homomorphism χ factors through ψ—one can check there is a homomorphism QF → Z
sending α to 1 and β, σε,1 to 0. Thus, if g ∈ Ker ψ then g ∈ Ker χ and so n = m. In
particular, αn ψ(h)α−n = idQF so h ∈ Ker ψ. However ψ is the identity on QF (1),
thus h = 1 and ψ in injective.
Proof of Theorem 6.3. Lemma 6.5 implies that for all i, we have Σi (QF, R) ⊆
Σi (F, R). In particular,
Σ2 (QF, R) ⊆ S(QF ) \ {[aπ ∗ χ0 + bπ ∗ χ1 ] : a, b ≥ 0}
The character which takes 1 on α and 0 elsewhere is exactly −π ∗ χ0 so using
Theorem 6.6(1) we find that [−π ∗ χ0 ] ∈ Σ∞ (QF ) and by Remark 6.4, [−π ∗ χ1 ] ∈
Σ∞ (QF ) as well.
Next we claim that
{[χ] ∈ S(QF ) : χ(β) < 0} ⊆ Σ∞ (QF ),
the argument is similar to that of [BGK10, Corollary 2.4]. Let χ : QF → R be a
character with χ(β) < 0, then
ψ ∗ [χ|QF (1) ](α) = χ(β) < 0,
ψ ∗ [χ|QF (1) ](β) = χ(β α ) = χ(β) < 0.
So ψ [χ|QF (1) ] = −[χ1 ] and thus [χ|QF (1) ] ∈ Σ∞ (QF (1)). Since QF (1) is isomorphic to QF and hence of type F∞ (Theorem 4.10), applying Theorem 6.6(2)
completes the claim.
The above implies that [aχ0 + bχ1 ] ∈ Σ∞ (QF ) if b < 0, so by ν symmetry
[aχ0 + bχ1 ] ∈ Σ∞ (QF ) if a < 0 also. Thus,
∗
{[aπ ∗ χ0 + bπ ∗ χ1 ] : a, b ≥ 0} ⊆ Σ∞ (QF )c
which completes the descriptions of Σi (QF ) and Σi (QF, R) for all i ≥ 2.
It remains to complete the description of Σ1 (QF ), we use Theorem 6.1 for this.
Let χ ∈ S(QF ) \ {[π ∗ χ0 ], [π ∗ χ1 ]} and consider the normal subgroup N = Ker π ∗ χ.
We will show that N is finitely generated, so necessarily χ ∈ Σ1 .
From Theorem 6.2 we know that π(N ) is finitely generated. Since Ker π|N =
Sym({0, 1}∗) we have a group extension
1 −→ Sym({0, 1}∗ ) −→ N −→ π(N ) −→ 1.
Thus, using the method of Section 3 it is sufficient to find a generating set of
Sym({0, 1}∗) on which π(N ) acts by permutations and with finitely many orbits,
QUASI-AUTOMORPHISMS OF T2,c
27
this is given by Lemma 6.8 below, since a generating set for Sym({0, 1}∗) is given
by the set Σ2 , which we recall below.
Recall from Section 3.1 that for all n,
(
n
Y
Σn = (x1 , . . . , xn ) ∈
{0, 1}∗ : x1
lex
x2
lex
···
lex
xn
1
)
,
and Σn admits an action of F via the splitting ι : F −→ QF , the inclusion QF −→
QV , and the usual action of QV on {0, 1}∗.
Recall the definition of χ̃ζ from Section 4.1. By comparing the values they take
on α and β, one checks that π ∗ χ1 = χ̃ζ .
Lemma 6.8. Let χ ∈ S(QF ) and N = Ker χ then π(N ) acts transitively on Σ2 .
Proof. Let (x1 , x2 ) ∈ Σ2 and let (L, R) be a representative for an element f of F
such that ι(f )(x1 ) = 0 and ι(f )(x2 ) = ε (use, for example, Lemma 3.2). Applying
Lemma 6.9 below with a = −χ1 (f ) gives an element f1 , and applying the lemma a
second time with a = −χ0 (f ) gives a second element f0 .
We claim that ν(f0 )f1 f is the required element of F . Note that since f0 fixes ε
and 1 by construction, the element ν(f0 ) fixes 1̄ = 0 and ε̄ = ε. Finally,
χ0 (ν(f0 )f1 f ) = χ0 (ν(f0 )) + χ0 (f1 ) + χ0 (f ) = χ1 (f0 ) + χ0 (f ) = 0,
and similarly one shows that χ1 (ν(f0 )f1 f ) = 0.
Let τ = ι(ν(f0 )f1 f ). Since χ1 (ν(f0 )f1 f ) = χ0 (ν(f0 )f1 f ) = 0, also π ∗ χ1 (τ ) =
∗
π χ0 (τ ) = 0, so the element τ is in the kernel of any character in S(QF ).
Lemma 6.9. For any integer a there exists an element f ∈ F such that ι(f )·ε = ε,
ι(f ) · 0 = 0, ι(f ) · 1 = 1, χ0 ι(f ) = 0, and χ1 ι(f ) = a.
Proof. Assume for now that a > 0. Let T be the smallest tree containing ε, 1, and
11 as nodes. Let L be obtained from T by attaching a carets iteratively to the
largest leaf of T (using the ≤lex ordering). Let R be obtained from T by attaching
a carets iteratively to the second largest caret of T . Denote by f the element of F
with tree diagram representative (L, R). For example when a = 2 we obtain the
element of Figure 14. One checks that χ1 (f ) = a and that ι(f ) fixes ε, 0, and 1.
If a < 0 then perform the steps above for −a and then replace f0 by f0−1 .
Figure 14. Element f from Lemma 6.9 when a = 2.
28
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
7. Relationships with other generalisations of Thompson’s groups
In [Thu16] Thumann studies the class of Operad groups, this class contains the
generalised Thompson groups Fn,r and Vn,r , the higher dimensional Thompson
groups nV , and the braided Thompson groups BV . Moreover, the class of diagram
groups appearing in [GS97] may be described as Operad groups, as may the class
of self-similarity groups described by Hughes [Hug09]—for a proof of these facts see
[Thu16, §3.5]. For each of these classes of groups there exist results showing that
certain groups in the class are of type F∞ [Far03, Theorem 1.3][FH12, Theorem
1.1][Thu16, Theorem 4.3]. We show in Section 7.1 that while QV admits an easy
description in all three of these classes, it cannot be proved to be of type F∞ using
these descriptions and the results listed above.
There is a generalisation of Thompson’s groups due to Martı́nez-Pérez and
Nucinkis, describing a family of automorphism groups of Cantor algebras [MPN13].
We show in Proposition 7.4 that none of the groups studied in this paper appear
in this family of groups.
Witzel and Zaremsky have introduced Thompson groups for systems of groups
[WZ14], this class contains the Thompson groups F and V and also the braided
Thompson groups BF and BV . They study the finiteness lengths of some groups
in this class. We do not know of a way to express any of the groups studied here
as a group in this class and we are unsure of the relationship between Witzel–
Zaremsky’s class of Thompson groups for systems of groups and either Thumann’s
class of Operad groups or the class of groups of Martı́nez-Pérez and Nucinkis.
7.1. Operad groups. Following [Thu16] we denote by
End : MON → OP
the endomorphism functor taking a strict monoidal category C to the endomorphism
operad End(C). Let S denote the left adjoint of End. Given a category C and
an object X ∈ C we write π1 (C, X) for the fundamental group of C based at X.
The operad group associated to an operad O and object X ∈ S(O) is the group
π1 (S(O), X).
Let OQV be the operad with colours l and n (one should think of these as
representing a leaf and a node of a finite subtree of T2,c ) and with a single operation
O(l, n, l; l) = {ϕ} (one should think of this operation as adding a caret).
Lemma 7.1. There is an isomorphism
QV ∼
= π1 (S(OQV ), l).
Proof. This is essentially [FH14, Example 4.4], where Farley and Hughes describe
QV as a braided diagram group over the semi-group presentation
P = hl, n : (l, lnl)i.
One can then convert this to a description of QV as a group acting on a compact
ultrametric space via a small similarity structure [FH14, Theorem 4.12] and then
in turn to a description as an operad group [Thu16, §3.5]. Following this method
one obtains the operad OQV given above.
The next lemma demonstrates that one cannot use the result of Thumann
[Thu16, Theorem 4.3]) to show that QV is F∞ because OQV is not colour-tame,
see [Thu16, Definition 4.2]. Recall that a colour word is said to be reduced if no
subword is in the domain of a non-identity operation.
Lemma 7.2. OQV is not colour-tame.
QUASI-AUTOMORPHISMS OF T2,c
29
Proof. The operad OQV is planar and has a finite set of colours, however the colour
word
i times
z }| {
n · n · · · n,
is reduced for all i ≥ 0.
Using [FH14, Theorem 4.12] to convert the description of QV as a braided diagram group over the semi-group presentation P to a description of QV as a selfsimilarity group gives the following. Let TP be the tree obtained from T2,c by
forgetting the colouring and adding to every node a single child. Let X be the
ultrametric space X = Ends(TP ) obtained by setting dX (p, p′ ) = e−i if p and p′ are
any two paths (without backtracking) which contain exactly i edges in common.
Balls in X are all of the form
Bv = {p ∈ Ends(TP ) : v lies on p},
for some vertex in v. If v and w are both leaves or both nodes then SimX (Bv , Bw )
contains a single element and otherwise SimX (Bv , Bw ) is empty.
The next lemma demonstrates that one cannot use the result of Farley and
Hughes [FH12, Theorem 1.1] together with the description of QV as the selfsimilarity group associated to the ultrametric space X and the similarity structure
SimX to prove that QV is F∞ . This is because [FH12, Theorem 1.1] requires that
SimX is rich in simple contractions [FH12, Definition 5.11].
Lemma 7.3. SimX is not rich in simple contractions.
Proof. Given an arbitrary constant k ∈ N>0 it suffices to exhibit a pseudo-vertex
v of height k such that for every pseudo vertex w ⊆ v, either kwk = 1 or there is
no simple contraction of v at w.
Let k ∈ N>0 be an arbitrary constant and choose k leaves n1 , . . . , nk ∈ Ends(TP )
with no children. Each of these leaves is a ball Bni and so
v = {[inclBni , Bni ] : i = 1, . . . , k}
is a pseudo-vertex of height k. Since any simple expansion of any pseudo-vertex
necessarily introduces balls containing leaves with children, there can be no simple
contraction of v.
7.2. Automorphisms of Cantor algebras. In [MPN13], Martı́nez-Pérez and
Nucinkis study certain automorphism groups Gr (Σ), Tr (Σ), and Fr (Σ) of certain
Cantor algebras.
Proposition 7.4. None of QV , Q̃V , QT , Q̃T , and QF are isomorphic to either
Gr (Σ), Tr (Σ), or Fr (Σ) for any Σ and r.
Proof. By [MPN13, Theorems 4.3, 4.8], any group Gr (Σ) or Tr (Σ) has at most
finitely many conjugacy classes of finite subgroups isomorphic to a given finite
subgroup. This is false however for QV and QT as one can find infintely many nonconjugate subgroups isomorphic to the cyclic group of order 2: let {x1 , . . . , xn , . . .}
and {y1 , . . . , yn , . . .} be two disjoint countably infinite subsets of {0, 1}∗ and let Gi
be the subgroup of QF which tranposes xj and yj for all j ≤ i. Clearly Gi ∼
= C2
for all i, but the Gi are all pairwise non-conjugate (this is because conjugation
preserves cycle type in infinite support permutation groups). Thus none of QV ,
Q̃V , QT , Q̃T , and QF may be isomorphic to Gr (Σ) or Tr (Σ).
Since Fr (Σ) is always torsion-free [MPN13, Remark 2.17], none of QV , Q̃V , QT ,
Q̃T , or QF may be isomorphic to Fr (Σ).
30
BRITA E.A. NUCINKIS AND SIMON ST. JOHN-GREEN
References
[Bel04]
[BFS+ 12]
James M. Belk, Thompson’s group F , Ph.D. thesis, 2004.
Kai-Uwe Bux, Martin Fluch, Marco Schwandt, Stefan Witzel, and Matthew C. B.
Zaremsky, The braided Thompson’s groups are of type F∞ , preprint, arXiv:1210.2931
(2012).
[BG84]
Kenneth S. Brown and Ross Geoghegan, An infinite-dimensional torsion-free FP∞
group, Invent. Math. 77 (1984), no. 2, 367–381. MR 752825 (85m:20073)
[BGK10]
Robert Bieri, Ross Geoghegan, and Dessislava H. Kochloukova, The sigma invariants
of Thompson’s group F , Groups Geom. Dyn. 4 (2010), no. 2, 263–273. MR 2595092
(2011h:20113)
[Bie81]
Robert Bieri, Homological dimension of discrete groups, second ed., Queen Mary
College Mathematical Notes, Queen Mary College Department of Pure Mathematics,
London, 1981. MR 715779 (84h:20047)
[BMN13] Collin Bleak, Francesco Matucci, and Max Neunhöffer, Embeddings into Thompson’s
group V and coCF groups, preprint, arXiv:1312.1855 (2013).
[BNS87]
Robert Bieri, Walter D. Neumann, and Ralph Strebel, A geometric invariant of discrete groups, Invent. Math. 90 (1987), no. 3, 451–477. MR 914846 (89b:20108)
[BR88]
Robert Bieri and Burkhardt Renz, Valuations on free resolutions and higher geometric
invariants of groups, Comment. Math. Helv. 63 (1988), no. 3, 464–497. MR 960770
(90a:20106)
[Bri04]
Matthew G. Brin, Higher dimensional Thompson groups, Geom. Dedicata 108 (2004),
163–192. MR 2112673 (2005m:20008)
, The algebra of strand splitting. I. A braided version of Thompson’s group V ,
[Bri07]
J. Group Theory 10 (2007), no. 6, 757–788. MR 2364825 (2009g:20092)
[Bro87]
Kenneth S. Brown, Finiteness properties of groups, Proceedings of the Northwestern
conference on cohomology of groups (Evanston, Ill., 1985), vol. 44, 1987, pp. 45–75.
MR 885095 (88m:20110)
[BS08]
Kai-Uwe Bux and Dmitriy Sonkin, Some remarks on the braided Thompson group
BV , preprint, arXiv:0807.0061 (2008).
[CFP96]
J. W. Cannon, W. J. Floyd, and W. R. Parry, Introductory notes on Richard Thompson’s groups, Enseign. Math. (2) 42 (1996), no. 3-4, 215–256. MR 1426438 (98g:20058)
[Dav08]
Michael W. Davis, The geometry and topology of Coxeter groups, London Mathematical Society Monographs Series, vol. 32, Princeton University Press, Princeton, NJ,
2008. MR 2360474 (2008k:20091)
[DL98]
Warren Dicks and Ian J. Leary, On subgroups of Coxeter groups, Geometry and cohomology in group theory (Durham, 1994), London Math. Soc. Lecture Note Ser.,
vol. 252, Cambridge Univ. Press, Cambridge, 1998, pp. 124–160. MR 1709956
(2000g:20067)
[Far03]
Daniel S. Farley, Finiteness and CAT(0) properties of diagram groups, Topology 42
(2003), no. 5, 1065–1082. MR 1978047 (2004b:20057)
[FH12]
Daniel Farley and Bruce Hughes, Finiteness properties of some groups of local similarities, preprint, arXiv:1206.2692 (2012).
, Braided diagram groups and local similarity groups, preprint,
[FH14]
arXiv:1406.4587 (2014).
[FMWZ13] Martin G. Fluch, Marco Marschler, Stefan Witzel, and Matthew C. B. Zaremsky,
The Brin-Thompson groups sV are of type F∞ , Pacific J. Math. 266 (2013), no. 2,
283–295. MR 3130623
[Geo08]
Ross Geoghegan, Topological methods in group theory, Graduate Texts in Mathematics, vol. 243, Springer, New York, 2008. MR 2365352 (2008j:57002)
[GS97]
Victor Guba and Mark Sapir, Diagram groups, Mem. Amer. Math. Soc. 130 (1997),
no. 620, viii+117. MR 1396957 (98f:20013)
[Hig74]
Graham Higman, Finitely presented infinite simple groups, Department of Pure Mathematics, Department of Mathematics, I.A.S. Australian National University, Canberra, 1974, Notes on Pure Mathematics, No. 8 (1974). MR 0376874 (51 #13049)
[Hug09]
Bruce Hughes, Local similarities and the Haagerup property, Groups Geom. Dyn.
3 (2009), no. 2, 299–315, With an appendix by Daniel S. Farley. MR 2486801
(2009m:20066)
[Kog08]
Roman Kogan, nV Trees Applet, http://www.math.tamu.edu/~ romwell/nvTrees/
(2008).
[Leh08]
Jörg Lehnert, Gruppen von quasi-automorphismen, Ph.D. thesis, 2008.
QUASI-AUTOMORPHISMS OF T2,c
31
[Mei96]
Holger Meinert, The homological invariants for metabelian groups of finite Prüfer
rank: a proof of the Σm -conjecture, Proc. London Math. Soc. (3) 72 (1996), no. 2,
385–424. MR 1367084 (98b:20082)
, Actions on 2-complexes and the homotopical invariant Σ2 of a group, J. Pure
[Mei97]
Appl. Algebra 119 (1997), no. 3, 297–317. MR 1453518 (98g:20084)
[MPMN14] Conchita Martı́nez-Pérez, Francesco Matucci, and Brita E. A. Nucinkis, Presentations
for centralisers in generalisations of Thompson’s group V , preprint (2014).
[MPN13] Conchita Martı́nez-Pérez and Brita E. A. Nucinkis, Bredon cohomological finiteness
conditions for generalisations of Thompson groups, Groups Geom. Dyn. 7 (2013),
no. 4, 931–959. MR 3134031
[MT73]
Ralph McKenzie and Richard J. Thompson, An elementary construction of unsolvable
word problems in group theory, Word problems: decision problems and the Burnside
problem in group theory (Conf., Univ. California, Irvine, Calif. 1969; dedicated to
Hanna Neumann), Studies in Logic and the Foundations of Math., vol. 71, NorthHolland, Amsterdam, 1973, pp. 457–478. MR 0396769 (53 #629)
[Ore51]
Oystein Ore, Some remarks on commutators, Proc. Amer. Math. Soc. 2 (1951), 307–
314. MR 0040298 (12,671e)
[Ren89]
Burkhardt Renz, Geometric invariants and HNN-extensions, Group theory (Singapore, 1987), de Gruyter, Berlin, 1989, pp. 465–484. MR 981863 (90a:20075)
[Tho80]
Richard J. Thompson, Embeddings into finitely generated simple groups which preserve the word problem, Word problems, II (Conf. on Decision Problems in Algebra,
Oxford, 1976), Stud. Logic Foundations Math., vol. 95, North-Holland, AmsterdamNew York, 1980, pp. 401–441. MR 579955 (81k:20050)
[Thu16]
Werner Thumann, Operad groups and their finiteness properties, preprint,
arXiv:1409.1085 (2016).
[WZ14]
Stefan Witzel and Matthew C. B. Zaremsky, Thompson groups for systems of groups,
and their finiteness properties, preprint, arXiv:1405.5491 (2014).
[Zar14]
Matthew Zaremsky, On normal subgroups of the braided Thompson groups, preprint,
arXiv:1403.8132 (2014).
E-mail address: [email protected]
Department of Mathematics, Royal Holloway, University of London, Egham, TW20
0EX
E-mail address: [email protected]
Department of Mathematics, University of Southampton, SO17 1BJ, UK
| 4math.GR
|
arXiv:1801.07829v1 [cs.CV] 24 Jan 2018
Dynamic Graph CNN for Learning on Point Clouds
Yue Wang∗
MIT
Yongbin Sun∗
MIT
Ziwei Liu
UC Berkeley
Sanjay E. Sarma
MIT
[email protected]
yb [email protected]
[email protected]
[email protected]
Michael M. Bronstein
USI / TAU / Intel
Justin M. Solomon
MIT
[email protected]
[email protected]
Abstract
1. Introduction
Point clouds, or scattered collections of points in 2D or
3D, are arguably the simplest shape representation; they
also comprise the output of 3D sensing technology including LiDAR scanners and stereo reconstruction. With the advent of fast 3D point cloud acquisition, recent pipelines for
graphics and vision often process point clouds directly, bypassing expensive mesh reconstruction or denoising due to
efficiency considerations or instability of these techniques
in the presence of noise. A few of the many recent applications of point cloud processing and analysis include
indoor navigation [57], self-driving vehicles [33], robotics
[40], and shape synthesis and modeling [14].
Modern applications demand high-level processing of
point clouds. Rather than identifying salient geometric features like corners and edges, recent algorithms search for
semantic cues and affordances. These features do not fit
cleanly into the frameworks of computational or differential
geometry and typically require learning-based approaches
that derive relevant information through statistical analysis
of labeled or unlabeled datasets.
In this paper, we primarily consider point cloud classification and segmentation, two model tasks in the point cloud
processing world. Traditional methods for solving these
problems employ handcrafted features to capture geometric properties of point clouds [26, 38, 39]. More recently,
the success of deep neural networks for image processing
has motivated a data-driven approach to learning features
on point clouds. Deep point cloud processing and analysis
methods are developing rapidly and outperform traditional
approaches in various tasks [10].
Adaptation of deep learning to point cloud data, however, is far from straightforward. Most critically, standard
deep neural network models take as input data with regular
structure, while point clouds are fundamentally irregular:
Point positions are continuously distributed in the space,
and any permutation of their ordering does not change the
Point clouds provide a flexible and scalable geometric
representation suitable for countless applications in computer graphics; they also comprise the raw output of most
3D data acquisition devices. Hence, the design of intelligent
computational models that act directly on point clouds is
critical, especially when efficiency considerations or noise
preclude the possibility of expensive denoising and meshing procedures. While hand-designed features on point
clouds have long been proposed in graphics and vision,
however, the recent overwhelming success of convolutional
neural networks (CNNs) for image analysis suggests the
value of adapting insight from CNN to the point cloud
world. To this end, we propose a new neural network module dubbed EdgeConv suitable for CNN-based high-level
tasks on point clouds including classification and segmentation. EdgeConv is differentiable and can be plugged into
existing architectures. Compared to existing modules operating largely in extrinsic space or treating each point independently, EdgeConv has several appealing properties:
It incorporates local neighborhood information; it can be
stacked or recurrently applied to learn global shape properties; and in multi-layer systems affinity in feature space
captures semantic characteristics over potentially long distances in the original embedding. Beyond proposing this
module, we provide extensive evaluation and analysis revealing that EdgeConv captures and exploits fine-grained
geometric properties of point clouds. The proposed approach achieves state-of-the-art performance on standard
benchmarks including ModelNet40 and S3DIS.
∗ Equal
Contribution
1
Figure 1. Point cloud segmentation using the proposed neural network. Bottom: schematic neural network architecture. Top: Structure
of the feature spaces produced at different layers of the network, visualized as the distance from the red point to all the rest of the points
(shown left-to-right are the input and layers 1-3; rightmost figure shows the resulting segmentation). Observe how the feature space
structure in deeper layers captures semantically similar structures such as wings, fuselage, or turbines, despite a large distance between
them in the original input space.
is designed to be invariant to the ordering of neighbors, and
thus permutation invariant.
EdgeConv is easy to implement and integrate into existing deep learning models to improve their performance. In
our experiments, we integrate EdgeConv into the basic version of PointNet without using any feature transformation.
We show performance improvement by a large margin; the
resulting network achieves state-of-the-art performance on
several datasets, most notably ModelNet40 and S3DIS for
classification and segmentation.
spatial distribution. One common approach to process point
cloud data using deep learning models is to first convert raw
point cloud data into a volumetric representation, namely a
3D grid [30, 54]. This approach, however, usually introduces quantization artifacts and excessive memory usage,
making it difficult to go to capture high-resolution or finegrained features.
State-of-the-art deep neural networks are designed
specifically to handle the irregularity of point clouds, directly manipulating raw point cloud data rather than passing to an intermediate regular representation. This approach
was pioneered by PointNet [34], which achieves permutation invariance of points by operating on each point independently and subsequently applying a symmetric function to accumulate features. Various extensions of PointNet consider neighborhoods of points rather than acting on
each independently [36, 43]; these allow the network to
exploit local features, improving upon performance of the
basic model. These techniques largely treat points independently at local scale to maintain permutation invariance.
This independence, however, neglects the geometric relationships among points, presenting a fundamental limitation
that leads to local features missing.
To address these drawbacks, we propose a novel simple
operation, called EdgeConv, which captures local geometric
structure while maintaining permutation invariance. Instead
of generating points’ features directly from their embeddings, EdgeConv generates edge features that describe the
relationships between a point and its neighbors. EdgeConv
Key Contributions. We summarize the key contributions
of our work as follows:
• We present a novel operation for point clouds, EdgeConv, to better capture local geometric features of
point clouds while still maintaining permutation invariance.
• We show the model can learn to semantically group
points by dynamically updating the graph.
• We demonstrate that EdgeConv can be integrated into
multiple existing pipelines for point cloud processing.
• We present extensive analysis and testing of EdgeConv
and show that it achieves state-of-the-art performance
on benchmark datasets.
2
Figure 2. Left: An example of computing an edge feature, eij , from a point pair, xi and xj . In this example, hΘ () is instantiated using
a fully connected layer, and the learnable parameters are its associated weights and bias. Right: Visualize the EdgeConv operation. The
output of EdgeConv is calculated by aggregating the edge features associated with all the edges emanating from each connected vertex.
2. Related Work
insight to geometric data like point clouds. Unlike images,
however, geometric data usually are not on an underlying
grid, requiring new definitions for building blocks like convolution and pooling.
Existing 3D deep learning methods can be split into
two classes. View-based and volumetric representations
exemplify techniques that try to “place” geometric data
onto a grid and apply existing deep learning algorithms
to the adapted structure. Other methods replace the standard building blocks of deep neural architectures with special operations suitable for unstructured geometric data
[29, 6, 31, 34, 36]. We provide details about the closest
techniques to ours below.
Hand-Crafted Features Various tasks in geometric data
processing and analysis — including segmentation, classification, and matching — require some notion of local
similarity between shapes. Traditionally, this similarity is
established by constructing feature descriptors that capture
local geometric structure. Countless papers in computer vision and graphics propose local feature descriptors for point
clouds suitable for different problems and data structures. A
comprehensive overview of hand-designed point features is
out of the scope of this paper, but we refer the reader to
[51, 15, 4] for comprehensive discussion.
Broadly speaking, one can distinguish between extrinsic and intrinsic descriptors. Extrinsic descriptors usually
are derived from the coordinates of the shape in 3D space
and includes classical methods like shape context [3], spin
images [17], integral features [27], distance-based descriptors [24], point feature histograms [39, 38], and normal histograms [50], to name a few. Intrinsic descriptors treat the
3D shape as a manifold whose metric structure is discretized
as a mesh or graph; quantities expressed in terms of the metric are by definition intrinsic and invariant to isometric deformation. Representatives of this class include spectral descriptors such as global point signatures [37], the heat and
wave kernel signatures [48, 2], and variants [8]. Most recently, several approaches wrap machine learning schemes
around standard descriptors [15, 42].
View-based Methods View-based techniques represent a
3D object as a collection of 2D views, to which standard
CNNs used in image analysis can be applied. Typically,
a CNN is applied to each view and then the resulting features are aggregated by a view pooling procedure [47].
View-based approaches are also good match for applications where the input comes from a 3D sensor and represented as a range image [53], in which case a single view
can be used.
Volumetric Methods Voxelization is a straightforward
way to convert unstructured geometric data to a regular
3D grid over which standard CNN operations can be applied [30, 54]. These volumetric representations are often
wasteful, since voxelization produces a sparsely-occupied
3D grid. Time and space complexity considerations limit
the resolution of the volumetric grids, yielding quantization
artifacts. Recent space partition methods like k-d trees [20]
or octrees [49] remedy some resolution issues but still rely
on subdivision of a bounding volume rather than local geometric structure. Finally, [35] studied a combination of
view-based and volumetric approaches for 3D shape classification.
Learned Features. In computer vision, approaches relying on ‘hand-crafted’ features have reached a plateau in performance on challenging image analysis problems like image recognition. A breakthrough came with the use of convolutional neural networks (CNNs) [22, 21], leading to an
overwhelming trend to abandon hand-crafted features in favor of models that learn task-specific features from data.
A basic CNN architecture is the deep neural network,
which interleaves convolutional and pooling layers to aggregate local information in images. This success of deep
learning for images suggests the value of adapting related
3
PointNets PointNets [34] comprise a special class of architectures for point sets like 3D point clouds. The key
ingredient is a symmetric function applied to 3D coordinates in a manner invariant to permutation. While they
achieve impressive performance on point cloud analysis
tasks, PointNets treat each point individually, essentially
learning a mapping from 3D to the latent features without leveraging local geometric structure. Furthermore, the
learned mapping is sensitive to the global transformation of
the point cloud; to cope with this issue, PointNet employs a
complex and computationally expensive spatial transformer
network [16] to learn 3D alignment.
Local information is important for feature learning in
two ways. First, as for handcrafted descriptors, local features usually account for geometric relationships among
neighboring points to be robust to various transformations.
Second, local information is critical to the succcess of
image-based deep convolutional architectures. Follow-up
work proposed an improved PointNet++ architecture exploiting geometric features in local point sets and hierarchically aggregating them for inference [36]. A similar approach is proposed in [43], where initial point features are
obtained from a point kernel correlation layer and then aggregated among nearby points. Benefiting from local structure, PointNet++ achieves state-of-the-art results on several
point cloud analysis benchmarks. PointNet++, however,
still treats individual points in local point sets independently
and does not consider relationships between point pairs.
[29]. Its key advantage over spectral approaches is better generalization. Follow-up work proposed different local charting techniques using anisotropic diffusion [6] or
Gaussian mixture models [52, 31]. [25] incorporate a differentiable functional map [32] layer into a geometric deep
neural network, allowing to do intrinsic structured prediction of correspondence between nonrigid shapes.
The last class of geometric deep learning approaches attempt to pull back a convolution operation by embedding
the shape into a domain with shift-invariant structure such
as the sphere [46], torus [28], or plane [13].
a
3. Our approach
We propose an approach inspired by PointNet and convolution operations. Instead of working on individual points
like PointNet, however, we exploit local geometric structures by constructing a local neighborhood graph and applying convolution-like operations on the edges connecting
neighboring pairs of points, in the spirit graph neural networks. We show in the following that such an operation,
dubbed edge convolution (EdgeConv), has the properties of
lying between translation-invariant and non-locality.
Differently from graph CNNs, the graph is not fixed but
rather is dynamically updated after each layer of the network. That is, the k-nearest neighbors of a point changes
from layer to layer of the network and is computed from the
sequence of embeddings. Proximity in feature space differs
from proximity in the input, leading to nonlocal diffusion of
information throughout the point cloud.
Geometric Deep Learning PointNets exemplify a broad
class of deep learning architectures on non-Euclidean structured data termed geometric deep learning [7]. These methods date back to early methods to construct neural networks
on graphs [41]. More recently, [9] proposed a generalization of convolution for graphs via the Laplacian operator
[44]. This foundational approach had a number of drawbacks including the computational complexity of Laplacian
eigendecomposition, the large number of parameters to express the convolutional filters, and a lack of spatial localization. These issues are alleviated in follow-up work using polynomial [11, 19] or rational [23] spectral filters that
avoid the Laplacian eigendecomposition and also guarantee
localization.
Spectral graph CNN models are notable for isometry invariance and hence have applied to non-rigid shape analysis
[5]. A key difficulty, however, is that the Laplacian eigenbasis is domain-dependent; thus, a filter learned on one shape
may not generalize to others. Spectral transformer networks
address this problem to some extent [56].
An alternative definition of non-Euclidean convolution
employs spatial rather than spectral filters. The Geodesic
CNN (GCNN) is a deep CNN on meshes generalizing the
notion of patches using local intrinsic parameterization
3.1. Edge Convolution
Consider a F -dimensional point cloud with n points,
denoted by X = {x1 , . . . , xn } ⊆ RF . In the simplest
setting of F = 3, each point contains 3D coordinates
xi = (xi , yi , zi ); it is also possible to include additional coordinates representing color, surface normal, and so on. In
a deep neural network architecture, each subsequent layer
operates on the output of the previous layer, so more generally the dimension F represents the feature dimensionality
of a given layer.
We further assume to be given a directed graph G =
(V, E) representing the local structure of the point cloud,
where V = {1, . . . , n} and E ⊆ V × V are the vertices and
edges, respectively. In the simplest case, we construct G as
the k-nearest neighbor (k-NN) graph in RF , containing directed edges of the form (i, ji1 ), . . . , (i, jik ) such that points
xji1 , . . . , xjik are the closest to xi . We define edge features
0
as eij = hΘ (xi , xj ), where hΘ : RF × RF → RF is some
parametric non-linear function parameterized by the set of
learnable parameters Θ.
Finally, we define the EdgeConv operation by applyingPa
channel-wise symmetric aggregation operation (e.g.,
4
Figure 3. Model architectures: The model architectures used for classification (top branch) and segmentation (bottom branch). The
classification model takes as input n points, calculates an edge feature set of size k for each point at an EdgeConv layer, and aggregates
features within each set to compute EdgeConv responses for corresponding points. The output features of the last EdgeConv layer are
aggregated globally to form an 1D global descriptor, which is used to generate classification scores for c classes. The segmentation model
extends the classification model by concatenating the 1D global descriptor and all the EdgeConv outputs (serving as local descriptors)
for each point. It outputs per-point classification scores for p semantic labels. For illustration purposes, two arrowed arcs are plotted to
represent feature concatenation. Point cloud transform block: The point cloud transform block is designed to align an input point set to a
canonical space by applying an estimated 3 × 3 matrix. To estimate the 3 × 3 matrix, a tensor concatenating the coordinates of each point
and the coordinate differences between its k neighboring points is used. EdgeConv block: The EdgeConv block takes as input a tensor of
shape n × f , computes edge features for each point by applying a multi-layer perceptron (mlp) with the number of layer neurons defined
as {a1 , a2 , ..., an }, and generates a tensor of shape n × an after pooling among neighboring edge features.
or max) on the edge features associated with all the edges
emanating from each vertex. The output of EdgeConv at the
i-th vertex is thus given by
x0i =
hΘ (xi , xj ).
where the parameters Θ = (θi , . . . , θk ) act as the weights
of the filter.
The second possible choice of h is hΘ (xi , xj ) =
hΘ (xi ), encoding only global shape information oblivious
of the local neighborhood structure. This type of operation
is used in PointNet, which can thus be regarded as a particular choice of our EdgeConv.
A third option is hΘ (xi , xj ) = hΘ (xj − xi ). Note that
such a choice encodes only local information, essentially
treating the shape as a collection of small patches and losing
the global shape structure.
Finally, the fourth option, which we adopt in this paper,
is an asymmetric edge function of the form hΘ (xi , xj ) =
hΘ (xi , xj − xi ). Such a function combines both the global
shape structure (captured by the coordinates of the patch
centers xi ) and local neighborhood information (captured
by xj − xi ).
(1)
j:(i,j)∈E
Making analogy to the classical convolution operation in
images, we can regard xi as the central pixel and {xj :
(i, j) ∈ E} as a patch around it (see Figure 2). Overall,
given an F -dimensional point cloud with n points, EdgeConv produces an F 0 -dimensional point cloud with the
same number of points.
Choice of h and The choice of the edge function and
the aggregation operation has a crucial influence on the
properties of the resulting EdgeConv operation.
First, note that in the setting when x1 , . . . , xn represent
image pixels layed out on a regular grid and the graph has a
local connectivity representing patches of fixed size around
each pixel, the choice hΘ (xi , xj ) = θj xj as the edge function and sum as the aggregation operation yields the classical Euclidean convolution,
X
x0i =
θ j xj ,
3.2. Dynamic Graph CNNs
Similarly to classical CNNs used in computer vision, the
EdgeConv operation can be applied multiple times, possibly interleaved with pooling depending on the task at
hand. When applied without pooling, multiple applications
of EdgeConv produce effectively larger support (‘receptive
(l)
(l)
field’) of the filter. We denote by X (l) = {x1 , . . . , xn } ⊆
j:(i,j)∈E
5
Figure 4. Structure of the feature spaces produced at different stages of our shape classification neural network architecture, visualized as
the distance between the red point to the rest of the points. For each set, Left: Euclidean distance in the input R3 space; Middle: Distance
after the point cloud transform stage, amounting to a global transformation of the shape; Right: Distance in the feature space of the last
layer. Observe how in the feature space of deeper layers semantically similar structures such as shelves of a bookshelf or legs of a table are
brought close together, although they are distant in the original space.
RFl the output of the l-th layer; X (0) = X is the input point
cloud.
Fl+1 -dimensional output of the (l + 1)-st layer is produced
by applying EdgeConv to the Fl -dimensional output of the
l-th layer,
(l+1)
xi
Dynamic graph update Our experiments suggests that it
is possible and actually beneficial to recompute the graph
using nearest neighbors in the features space produces by
each layer. This is a crucial distinction of our method from
graph CNNs working on a fixed input graph. Such a dynamic graph update is the reason for the name of our architecture, the Dynamic Graph CNN (DGCNN). At each layer
we thus have a different graph G (l) = (V (l) , E (l) ), where the
l-th layer edges are of the form (i, ji1 ), . . . , (i, jikl ) such
(l)
(l)
(l)
that xji1 , . . . , xjik are the kl points closest to xi . The
=
(l)
(l)
(l)
hΘ (xi , xj ),
(2)
j:(i,j)∈E (l)
where h(l) : RFl × RFl → RFl+1 .
3.3. Implementation Details
We consider particular instances of Dynamic Graph
CNNs for two prototypical tasks in point cloud analysis:
classification and segmentation. The respective architectures, depicted in Figure 3 (top and bottom branches), have
l
6
a similar structure to PointNet. Both architectures share a
spatial transformer component, computing a global shape
transformation. The classification network includes two
EdgeConv layers, followed by a pooling operation and
three fully-connected layers producing classification output
scores. The segmentation network uses a sequence of three
EdgeConv layers, followed by three fully-connected layers producing, for each point, segmentation output scores.
For each EdgeConv block, we use a shared edge function
(l)
(l)
(l)
(l)
(l)
h(l) (xi , xj ) = h(xi , xj − xi ) across all layers; the
function is implemented as a multi-layer perceptron (MLP)
and = max aggregation operation.
In our classification architecture, the graph is constructed
using k = 20 nearest neighbors, while in our segmentation
architecture, k = 30.
ture in these coordinates:
x0i =
M
X
wm
m=1
X
gΘm (u(xi , xj ))xj ,
(3)
j:(i,j)∈E
where g denotes the Gaussian kernel, {Θ1 , . . . , ΘM } encode the learnable parameters of the Gaussians (mean and
covariance), and {w1 , . . . , wM } are the learnable filter coefficients. We can easily observe that (3) is an instance of
our more general EdgeConv operation (1), with a particular
choice of the edge function
hw1 ,Θ1 ,...,wM ,ΘM (xi , xj ) =
M
X
wm gΘm (u(xi , xj ))xj
m=1
and summation as the aggregation operation.
A crucial difference between EdgeConv and MoNet and
other graph CNN methods is that the latter assume a given
fixed graph on which the convolution-like operations are applied, while we dynamically update the graph for each layer
output. This way, our model not only learns how to extract
local geometric features, but also how to group points in a
point cloud. Figure 4 shows the distance in different feature spaces, exemplifying that the distances in deeper layers
carry semantic information over long distances.
4. Comparison to existing methods
Our DGCNN is related to two classes of approaches,
PointNets and graph CNNs, which we show to be particular settings of our method.
PointNet is a special case of our method with k = 1,
which results in a graph with an empty edge set E = ∅.
The edge function used in PointNet is h(xi , xj ) = h(xi ),
which considers only the global geometry but discards the
local one. The P
aggregation operation used in PointNet is
= max (or , because the aggregation function only
works on a single node).
PointNet++ tries to account for local point cloud structure by applying PointNet in a local manner. In terms of
our notation, PointNet++ first constructs the graph according to the Euclidean distances between the points, and in
each layer, applies a graph coarsening operation. For each
layer, a certain number of points are selected by using farthest point sampling (FPS) algorithm. Only the selected
points are preserved while others are directly discarded after
this layer and in this way, the graph becomes smaller after
the operation applied on each layer. Different from ours,
PointNet++ computes pairwise distances using point input
coordinates. The edge function used by PointNet++ is also
h(xi , xj ) = h(xi ), and the aggregation operation is also a
max.
Among graph CNNs, MoNet [31], ECC [45], and Graph
Attention Networks [52] are the most related approaches.
The common denominator of these methods is the notion of
a local patch on a graph, in which a convolution-type operation can be defined.1 Specifically, [31] use the graph structure to compute a local “pseudo-coordinate system” u in
which the neighborhood vertices are represented; the convolution is then defined as an M -component Gaussian mix-
5. Evaluation
In this section, we evaluate the models constructed using
EdgeConv for different tasks: classification, part segmentation, and semantic segmentation. We also visualize experimental results to illustrate key differences from previous
work.
5.1. Classification
Data We evaluate our model on the ModelNet40 [54]
classification task, consisting in predicting the category of
a previously unseen shape. The dataset contains 12,311
meshed CAD models from 40 categories. 9,843 models are
used for training and 2,468 models are for testing. We follow verbatim the experimental settings of [34]. For each
model, 1,024 points are uniformly sampled from the mesh
faces and normalized to the unit sphere. Only the (x, y, z)
coordinates of the sampled points are used and the original
meshes are discarded. During the training procedure, we
augment the data as in [36] by randomly rotating and scaling objects and perturbing the object and point locations.
Architecture The network architecture used for the classification task is shown in Figure 3 (top branch). We use
a local-aware spatial transformer network to align the point
cloud. It has two shared fully-connected layers (64, 128)2 to
1 The methods of [45] and [52] can be considered as instances of [31],
with the difference that the weights are constructed employing features
from the adjacent nodes instead of the graph structure.
2 Here, we use (64, 128) to denote that the two fully-connected layers
have 64 filters and 128 filters, respectively; we use the same notation for
the remainder of our discussion.
7
construct one EdgeConv layer and after which, one shared
fully-connected layer (1024) is used to transform the pointwise features to higher-dimensional space. After the global
max pooling, two fully-connected layers (512, 256) are
used to compute the transformation matrix.
We use two EdgeConv layers to extract geometric features. The first EdgeConv layer uses three shared fullyconnected layers (64, 64, 64) while the second EdgeConv
layer uses a shared fully-connected layer (128). Shortcut connections are included to extract multi-scale features
and one shared fully-connected layer (1024) to aggregate
multi-scale features. The number k of nearest neighbors is
20. Then, a global max pooling is used to get the point
cloud global feature, after which two multi-layer perceptrons (512, 256) are used to transform the global feature.
Dropout with keep probability of 0.5 is used in the last two
fully-connected layers. All layers include ReLU and batch
normalization.
time), and achieved classification accuracy.
Our baseline model outperforms the previous state-ofthe-art PointNet++ by 0.5% accuracy, at the same time being 5 times faster compared to PointNet++. Our baseline
model does not use a spatial transformer and uses the fixed
k-NN graph. A more advanced version of our model including a spatial transformer block and dynamically graph
computation outperforms PointNet++ by 1.5% while having comparable number of parameters and computational
complexity.
P OINT N ET (BASELINE )
P OINT N ET
P OINT N ET ++
O URS (BASELINE )
O URS
3DS HAPE N ETS [54]
VOX N ET [30]
S UBVOLUME [35]
ECC [45]
P OINT N ET [34]
P OINT N ET ++ [36]
K D -N ET ( DEPTH 10) [20]
K D -N ET ( DEPTH 15) [20]
77.3
83.0
86.0
83.2
86.0
-
84.7
85.9
89.2
87.4
89.2
90.7
90.6
91.8
O URS ( BASELINE )
O URS
88.8
90.2
91.2
92.2
ACCURACY (%)
11.6
25.3
163.2
29.7
94.6
87.1
89.2
90.7
91.2
92.2
5.3. More Experiments on ModelNet40
We also experiment with various settings of our model
on the ModelNet40 [54] dataset. In particular, we analyze
the effectiveness of local-aware transformer network, different distance metrics, and explicit usage of xi − xj .
Table 3 shows the results. “Centralization” denotes using concatenation of xi and xi − xj as the edge features
rather than concatenating xi and xj . “Spatial Transformer”
denotes the local-aware transformer network while “Dynamic graph recomputation” denotes we reconstruct the
graph rather than using a fixed graph. By dynamically updating graph, there is about 0.2%∼0.3% improvement, and
Figure 4 also verifies our hypothesis that the model can extract semantics. In the later layers, certain patterns occur
for recognition tasks. Explicitly centralizing each patch by
using the concatenation of xi and xi − xj makes the operator more robust to translation, leading to about 0.2%∼0.3%
improvement for overall accuracy. The local-aware transformer makes the model invariant to rigid transformation
and leads to approximately 0.7% improvement.
We also experiment with different numbers k of nearest
neighbors as shown in Table 4. While we do not exhaustively experiment with all possible k, we find with large k
that the performance degenerates. This confirms our hypothesis that with large k the Euclidean distance fails to
approximate geodesic distance, destroying the geometry of
each patch.
We further evaluate the robustness of our model (trained
on 1,024 points with k = 20) to point cloud density. We
simulate the environment that random input points drops out
during testing. Figure 5 shows that even half of points is
dropped, the model still achieves reasonable results. With
fewer than 512 points, however, performance degenerates
Results Table 1 shows the results of the classification
task. Our model achieves the best results on this dataset.
Our baseline without transformer network and using fixed
graph is 0.5% better than PointNet++. An advanced version including a local-aware network and dynamical graph
recomputation achieves best results on this dataset.
OVERALL
ACCURACY
F ORWARD TIME ( MS )
9.4
40
12
11
21
Table 2. Complexity, forward time and accuracy of different models
Training We use the same training strategy as [34]. We
use Adam [18] with learning rate 0.001 that is divided by 2
every 20 epochs. The decay rate for batch normalization is
initially 0.5 and 0.99 finally. The batch size is 32 and the
momentum is 0.9.
M EAN
C LASS ACCURACY
M ODEL SIZE (MB)
Table 1. Classification results on ModelNet40.
5.2. Model Complexity
We use the ModelNet40 [54] classification experiment
to compare the complexity of our model to previous stateof-the-art. Table 2 shows that our model achieve the best
tradeoff between the model complexity (number of parameters), computational complexity (measured as forward pass
8
Figure 5. Left: Results of our model tested with random input
dropout. The model is trained with number of points being 1024
and k being 20. Right: Point clouds with different number of
points. The numbers of points are shown below the bottom row.
Figure 7. Our part segmentation testing results for chairs.
Figure 6. Our part segmentation testing results for tables.
Figure 8. Our part segmentation testing results for lamps.
dramatically.
CENT
DYN
XFORM
M EAN C LASS ACCURACY (%)
OVERALL ACCURACY (%)
X
88.8
88.8
89.6
89.8
90.2
91.2
91.5
91.9
91.9
92.2
X
X
X
X
X
X
X
X
X
contains 16,881 3D shapes from 16 object categories, annotated with 50 parts in total. 2,048 points are sampled
from each training shape, and most sampled point sets are
labeled with less than six parts. We follow the official
train/validation/test split scheme as in [10] in our experiment.
Table 3. Effectiveness of different components. CENT denotes
centralization, DYN denotes dynamical graph recomputation, and
XFORM denotes the use of a spatial transformer.
N UMBER OF NEAREST NEIGHBORS ( K )
5
10
20
40
M EAN
C LASS ACCURACY (%)
OVERALL
ACCURACY (%)
88.0
88.8
90.2
89.2
90.5
91.4
92.2
91.7
Architecture The network architecture is illustrated in
Figure 3 (bottom branch). The same spatial transformer
network is used for segmentation task. Nine shared fullyconnected layers (64, 64, 64, 64, 64, 64, 64, 64, 64) are used
to construct three EdgeConv layers; each EdgeConv layer
has three fully-connected layers (64, 64, 64). A shared
fully-connected layer (1024) is used to aggregate information from the previous layers. Shortcut connections are
used to include all the EdgeConv outputs as local feature
descriptors. At last, three shared fully-connected layers
(256, 256, 128) are used to transform the pointwise features. Batch-norm, dropout, and ReLU are included in the
similar fashion to our classification network.
Table 4. Results of our model with different numbers of nearest
neighbors.
5.4. Part Segmentation
Training The same training setting as in our classification
task is adopted, except k is changed from 20 to 30 due to the
increase of point density. A distributed training scheme is
further implemented on two NVIDIA TITAN X GPUs to
maintain the training batch size.
Data We extend our EdgeConv model architectures for
part segmentation task on ShapeNet part dataset [55]. For
this task, each point from a point cloud set is classified into
one of a few predefined part category labels. The dataset
9
PointNet
Ours
Ground truth
Figure 10. Compare part segmentation results. For each set, from
left to right: PointNet, ours and ground truth.
Source points
Other point clouds from the same
category
Figure 9. Visualize the Euclidean distance (yellow: near, blue: far)
between source points (red points in the left column) and multiple
point clouds from the same category in the feature space after the
third EdgeConv layer. Notice source points not only capture semantically similar structures in the point clouds that they belong
to, but also capture semantically similar structures in other point
clouds from the same category.
Figure 11. Left: The mean IoU (%) improves when the ratio of
kept points increases. Points are dropped from one of six sides
(top, bottom, left, right, front and back) randomly during evaluation process. Right: Part segmentation results on partial data.
Points on each row are dropped from the same side. The keep ratio is shown below the bottom row. Note that the segmentation
results of turbines are improved when more points are included.
Results We use Intersection-over-Union (IoU) on points
to evaluate our model and compare with other benchmarks.
We follow the same evaluation scheme as PointNet: The
IoU of a shape is computed by averaging the IoUs of different parts occurring in that shape; The IoU of a category is
obtained by averaging the IoUs of all the shapes belonging
to that category. The mean IoU (mIoU) is finally calculated
by averaging the IoUs of all the testing shapes. We compare our results with PointNet [34], PointNet++ [36], KdNet [20], and LocalFeatureNet [43]. The evaluation results
are shown in Table 5. We also visually compare the results
of our model and PointNet in Figure 10.
points in other point clouds from the same category. An
interesting finding is that although points are from different
sources, they are close to each other if they are from semantically similar parts. We evaluate on the features after the
third layer of our segmentation model for this experiment.
Segmentation on partial data Our model is robust to
partial data. We simulate the environment that part of the
shape is dropped from one of six sides (top, bottom, right,
left, front and back) with different percentages. The results
are shown in Figure 11. On the left, the mean IoU versus
“keep ratio” is shown. On the right, the results for an airplane model are visualized.
Intra-cloud distances We next explore the relationships
between different point clouds captured using our features.
As shown in Figure 9, we take one red point from a source
point cloud and compute its distance in feature space to
10
MEAN
AREO
BAG
CAP
CAR
CHAIR
.
# SHAPES
EAR
GUITAR
KNIFE
LAMP
LAPTOP
MOTOR
MUG
PISTOL
ROCKET
PHONE
SKATE
TABLE
WINNING
CATEGORIES
BOARD
2690
76
55
898
3758
69
787
392
1547
451
202
184
283
66
152
5271
P OINT N ET
P OINT N ET ++
K D -N ET
L OCAL F EATURE N ET
83.7
85.1
82.3
84.3
83.4
82.4
80.1
86.1
78.7
79.0
74.6
73.0
82.5
87.7
74.3
54.9
74.9
77.3
70.3
77.4
89.6
90.8
88.6
88.8
73.0
71.8
73.5
55.0
91.5
91.0
90.2
90.6
85.9
85.9
87.2
86.5
80.8
83.7
81.0
75.2
95.3
95.3
94.9
96.1
65.2
71.6
57.4
57.3
93.0
94.1
86.7
91.7
81.2
81.3
78.1
83.1
57.9
58.7
51.8
53.9
72.8
76.4
69.9
72.5
80.6
82.6
80.3
83.8
1
5
0
5
O URS
85.1
84.2
83.7
84.4
77.1
90.9
78.5
91.5
87.3
82.9
96.0
67.8
93.3
82.6
59.7
75.5
82.0
6
Table 5. Part segmentation results on ShapeNet part dataset. Metric is mIoU(%) on points.
M EAN
IOU
OVERALL
ACCURACY
P OINT N ET (BASELINE ) [34]
P OINT N ET [34]
MS + CU(2) [12]
G + RCU [12]
20.1
47.6
47.8
49.7
53.2
78.5
79.2
81.1
O URS
56.1
84.1
Table 6. 3D semantic segmentation results on S3DIS. MS+CU for
multi-scale block features with consolidation units; G+RCU for
the grid-blocks with recurrent consolidation Units.
Figure 12. Surface normal estimation results. The colors shown in
the figure are RGB-coded surface normals, meaning XYZ components of surface normal vectors are put into RGB color channels.
For each pair: our prediction (left) and ground truth (right).
5.6. Surface Normal Prediction
5.5. Indoor Scene Segmentation
We can also adapt our segmentation model to predict surface normals from point clouds.
Data We evaluate our model on Stanford Large-Scale 3D
Indoor Spaces Dataset (S3DIS) [1] for a semantic scene segmentation task. This dataset includes 3D scan point clouds
for 6 indoor areas including 272 rooms in total. Each point
belongs to one of 13 semantic categories—e.g. board, bookcase, chair, ceiling, and beam—plus clutter. We follow
the same setting as in [34], where each room is split into
blocks with area 1m × 1m, and each point is represented
as a 9D vector (XYZ, RGB, and normalized spatial coordinates). 4,096 points are sampled for each block during
training process, and all points are used for testing. We also
use the same 6-fold cross validation over the 6 areas, and
the average evaluation results are reported.
The model used for this task is similar to part segmentation model, except that a probability distribution over semantic object classes is generated for each input point. We
compare our model with both PointNet [34] and PointNet baseline, where additional point features (local point
density, local curvature and normal) are used to construct
handcrafted features and then fed to an MLP classifier.
We further compare our work with [12], who present network architectures to enlarge the receptive field over the 3D
scene. Two different approaches are proposed in their work:
MS+CU for multi-scale block features with consolidation
units; G+RCU for the grid-blocks with recurrent consolidation Units. We report evaluation results in Table 6, and
visually compare the results of PointNet and our model in
Figure 13.
Data We still use the ModelNet40 dataset. The surface
normals are sampled directly from CAD models. The normal of one point is represented by (nx , ny , nz ). We use
9,843 models for training and 2,468 models for testing.
Architecture We change the last layer of our segmentation model to output 3 continuous values; mean squared error (MSE) is used as the training loss.
Results Qualitative results are shown in Figure 12, compared with ground truth. Our model faithfully captures orientation even in the presence of fairly sharp features.
6. Discussion
In this work we propose a new operator for learning on
point cloud and show its performance on various tasks. The
success of our technique verifies our hypothesis that local
geometric features are crucial to 3D recognition tasks, even
after introducing machinery from deep learning. Furthermore, we show our model can be easily modified for various tasks like normal prediction while continuing to achieve
reasonable results.
While our architectures easily can be incorporated asis into existing pipelines for point cloud-based graphics,
learning, and vision, our experiments also indicate several
11
PointNet
Ours
Ground truth
Real color
Figure 13. Semantic segmentation results. From left to right: PointNet, ours, ground truth and point cloud with original color. Notice our
model outputs smoother segmentation results, for example, wall (cyan) in top two rows, chairs (red) and columns (magenta) in bottom two
rows.
avenues for future research and extension. Primarily, the
success of our model suggests that intrinsic features can
be equally valuable if not more than simply point coordinates; developing a practical and theoretically-justified
framework for balancing intrinsic and extrinsic considerations in a learning pipeline will require insight from theory
and practice in geometry processing. Another possible extension is to design a non-shared transformer network that
works on each local patch differently, adding flexibility to
our model. Finally, we will consider applications of our
techniques to more abstract point clouds coming from ap-
plications like document retrieval rather than 3D geometry;
beyond broadening the applicability of our technique, these
experiments will provide insight into the role of geometry
in abstract data processing.
References
[1] I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis,
M. Fischer, and S. Savarese. 3d semantic parsing of
large-scale indoor spaces. In Proc. CVPR, 2016.
[2] M. Aubry, U. Schlickewei, and D. Cremers. The wave
kernel signature: A quantum mechanical approach to
12
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
shape analysis. In Proc. ICCV Workshops, 2011.
S. Belongie, J. Malik, and J. Puzicha. Shape context:
A new descriptor for shape matching and object recognition. In Proc. NIPS, 2001.
S. Biasotti, A. Cerri, A. Bronstein, and M. Bronstein.
Recent trends, applications, and perspectives in 3d
shape similarity assessment. Computer Graphics Forum, 35(6):87–119, 2016.
D. Boscaini, J. Masci, S. Melzi, M. M. Bronstein,
U. Castellani, and P. Vandergheynst. Learning classspecific descriptors for deformable shapes using localized spectral convolutional networks. Computer
Graphics Forum, 34(5):13–23, 2015.
D. Boscaini, J. Masci, E. Rodolà, and M. Bronstein.
Learning shape correspondence with anisotropic convolutional neural networks. In Proc. NIPS, 2016.
M. M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, and
P. Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.
M. M. Bronstein and I. Kokkinos. Scale-invariant heat
kernel signatures for non-rigid shape recognition. In
Proc. CVPR, 2010.
J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun.
Spectral networks and locally connected networks on
graphs. arXiv:1312.6203, 2013.
A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song,
H. Su, et al. Shapenet: An information-rich 3d model
repository. arXiv:1512.03012, 2015.
M. Defferrard, X. Bresson, and P. Vandergheynst.
Convolutional neural networks on graphs with fast localized spectral filtering. In Proc. NIPS, 2016.
F. Engelmann, T. Kontogianni, A. Hermans, and
B. Leibe. Exploring spatial context for 3d semantic
segmentation of point clouds. In Proc. CVPR, 2017.
D. Ezuz, J. Solomon, V. G. Kim, and M. Ben-Chen.
Gwcnn: A metric alignment layer for deep shape analysis. Computer Graphics Forum, 36(5):49–57, 2017.
A. Golovinskiy, V. G. Kim, and T. Funkhouser. Shapebased recognition of 3d point clouds in urban environments. In Proc. ICCV, 2009.
Y. Guo, M. Bennamoun, F. Sohel, M. Lu, and J. Wan.
3d object recognition in cluttered scenes with local
surface features: a survey. Trans. PAMI, 36(11):2270–
2287, 2014.
M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. In Proc. NIPS, 2015.
A. E. Johnson and M. Hebert. Using spin images
for efficient object recognition in cluttered 3D scenes.
Trans. PAMI, 21(5):433–449, 1999.
[18] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. 2015.
[19] T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. 2017.
[20] R. Klokov and V. Lempitsky. Escape from cells: Deep
kd-networks for the recognition of 3d point cloud
models. 2017.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural
networks. In Proc. NIPS, 2012.
[22] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E.
Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition.
Neural computation, 1(4):541–551, 1989.
[23] R. Levie, F. Monti, X. Bresson, and M. M.
Bronstein. Cayleynets: Graph convolutional neural networks with complex rational spectral filters.
arXiv:1705.07664, 2017.
[24] H. Ling and D. W. Jacobs. Shape classification using the inner-distance. Trans. PAMI, 29(2):286–299,
2007.
[25] O. Litany, T. Remez, E. Rodolà, A. M. Bronstein, and
M. M. Bronstein. Deep functional maps: Structured
prediction for dense shape correspondence. In Proc.
ICCV, 2017.
[26] M. Lu, Y. Guo, J. Zhang, Y. Ma, and Y. Lei. Recognizing objects in 3d point clouds with multi-scale local
features. Sensors, 14(12):24156–24173, 2014.
[27] S. Manay, D. Cremers, B.-W. Hong, A. J. Yezzi, and
S. Soatto. Integral invariants for shape matching.
Trans. PAMI, 28(10):1602–1618, 2006.
[28] H. Maron, M. Galun, N. Aigerman, M. Trope,
N. Dym, E. Yumer, V. G. Kim, and Y. Lipman. Convolutional neural networks on surfaces via seamless toric
covers. In Proc. SIGGRAPH, 2017.
[29] J. Masci, D. Boscaini, M. Bronstein, and P. Vandergheynst. Geodesic convolutional neural networks
on riemannian manifolds. In Proc. 3dRR, 2015.
[30] D. Maturana and S. Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition.
In Proc. IROS, 2015.
[31] F. Monti, D. Boscaini, J. Masci, E. Rodolà, J. Svoboda, and M. M. Bronstein. Geometric deep learning
on graphs and manifolds using mixture model cnns. In
Proc. CVPR, 2017.
[32] M. Ovsjanikov, M. Ben-Chen, J. Solomon,
A. Butscher, and L. Guibas. Functional maps: a
flexible representation of maps between shapes. TOG,
31(4):30, 2012.
13
[33] C. R. Qi, W. Liu, C. Wu, H. Su, and L. J. Guibas.
Frustum pointnets for 3d object detection from rgb-d
data. arXiv:1711.08488, 2017.
[34] C. R. Qi, H. Su, K. Mo, and L. J. Guibas. Pointnet:
Deep learning on point sets for 3d classification and
segmentation. In Proc. CVPR, 2017.
[35] C. R. Qi, H. Su, M. Nießner, A. Dai, M. Yan, and L. J.
Guibas. Volumetric and multi-view cnns for object
classification on 3d data. In Proc. CVPR, 2016.
[36] C. R. Qi, L. Yi, H. Su, and L. J. Guibas. Pointnet++:
Deep hierarchical feature learning on point sets in a
metric space. In Proc. NIPS, 2017.
[37] R. M. Rustamov. Laplace-beltrami eigenfunctions for
deformation invariant shape representation. In Proc.
SGP, 2007.
[38] R. B. Rusu, N. Blodow, and M. Beetz. Fast point feature histograms (fpfh) for 3d registration. In Proc.
ICRA, 2009.
[39] R. B. Rusu, N. Blodow, Z. C. Marton, and M. Beetz.
Aligning point cloud views using persistent feature
histograms. In Proc. IROS, 2008.
[40] R. B. Rusu, Z. C. Marton, N. Blodow, M. Dolha,
and M. Beetz. Towards 3D Point Cloud Based Object Maps for Household Environments. Robotics
and Autonomous Systems Journal, 56(11):927–941,
30 November 2008.
[41] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner,
and G. Monfardini. The graph neural network model.
IEEE Tran. Neural Networks, 20(1):61–80, 2009.
[42] S. A. A. Shah, M. Bennamoun, F. Boussaid, and A. A.
El-Sallam. 3d-div: A novel local surface descriptor
for feature matching and pairwise range image registration. In Proc. ICIP, 2013.
[43] Y. Shen, C. Feng, Y. Yang, and D. Tian. Neighbors
do help: Deeply exploiting local structures of point
clouds. arXiv:1712.06760, 2017.
[44] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega,
and P. Vandergheynst. The emerging field of signal
processing on graphs: Extending high-dimensional
data analysis to networks and other irregular domains. IEEE Signal Processing Magazine, 30(3):83–
98, 2013.
[45] M. Simonovsky and N. Komodakis. Dynamic edgeconditioned filters in convolutional neural networks on
graphs. In Proc. CVPR, 2017.
[46] A. Sinha, J. Bai, and K. Ramani. Deep learning
3d shape surfaces using geometry images. In Proc.
ECCV, 2016.
[47] H. Su, S. Maji, E. Kalogerakis, and E. LearnedMiller. Multi-view convolutional neural networks for
3d shape recognition. In Proc. CVPR, 2015.
[48] J. Sun, M. Ovsjanikov, and L. Guibas. A concise and provably informative multi-scale signature
based on heat diffusion. Computer Graphics Forum,
28(5):1383–1392, 2009.
[49] M. Tatarchenko, A. Dosovitskiy, and T. Brox. Octree
generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proc. ICCV,
2017.
[50] F. Tombari, S. Salti, and L. Di Stefano. A combined texture-shape descriptor for enhanced 3d feature
matching. In Proc. ICIP, 2011.
[51] O. Van Kaick, H. Zhang, G. Hamarneh, and D. CohenOr. A survey on shape correspondence. Computer
Graphics Forum, 30(6):1681–1707, 2011.
[52] P. Veličković, G. Cucurull, A. Casanova, A. Romero,
P. Liò, and Y. Bengio. Graph attention networks.
arXiv:1710.10903, 2017.
[53] L. Wei, Q. Huang, D. Ceylan, E. Vouga, and H. Li.
Dense human body correspondences using convolutional networks. In Proc. CVPR, 2016.
[54] Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang,
and J. Xiao. 3d shapenets: A deep representation for
volumetric shapes. In Proc. CVPR, 2015.
[55] L. Yi, V. G. Kim, D. Ceylan, I. Shen, M. Yan, H. Su,
A. Lu, Q. Huang, A. Sheffer, L. Guibas, et al. A
scalable active framework for region annotation in 3d
shape collections. TOG, 35(6):210, 2016.
[56] L. Yi, H. Su, X. Guo, and L. Guibas. Syncspeccnn:
Synchronized spectral cnn for 3d shape segmentation.
In Proc. CVPR, 2017.
[57] Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta,
L. Fei-Fei, and A. Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement
learning. In Proc. ICRA, 2017.
14
| 1cs.CV
|
Deep Learning in Neural Networks: An Overview
arXiv:1404.7828v4 [cs.NE] 8 Oct 2014
Technical Report IDSIA-03-14 / arXiv:1404.7828 v4 [cs.NE] (88 pages, 888 references)
Jürgen Schmidhuber
The Swiss AI Lab IDSIA
Istituto Dalle Molle di Studi sull’Intelligenza Artificiale
University of Lugano & SUPSI
Galleria 2, 6928 Manno-Lugano
Switzerland
8 October 2014
Abstract
In recent years, deep artificial neural networks (including recurrent ones) have won numerous
contests in pattern recognition and machine learning. This historical survey compactly summarises
relevant work, much of it from the previous millennium. Shallow and deep learners are distinguished by the depth of their credit assignment paths, which are chains of possibly learnable, causal
links between actions and effects. I review deep supervised learning (also recapitulating the history
of backpropagation), unsupervised learning, reinforcement learning & evolutionary computation,
and indirect search for short programs encoding deep and large networks.
LATEX source: http://www.idsia.ch/˜juergen/DeepLearning8Oct2014.tex
Complete BIBTEX file (888 kB): http://www.idsia.ch/˜juergen/deep.bib
Preface
This is the preprint of an invited Deep Learning (DL) overview. One of its goals is to assign credit
to those who contributed to the present state of the art. I acknowledge the limitations of attempting
to achieve this goal. The DL research community itself may be viewed as a continually evolving,
deep network of scientists who have influenced each other in complex ways. Starting from recent DL
results, I tried to trace back the origins of relevant ideas through the past half century and beyond,
sometimes using “local search” to follow citations of citations backwards in time. Since not all DL
publications properly acknowledge earlier relevant work, additional global search strategies were employed, aided by consulting numerous neural network experts. As a result, the present preprint mostly
consists of references. Nevertheless, through an expert selection bias I may have missed important
work. A related bias was surely introduced by my special familiarity with the work of my own DL
research group in the past quarter-century. For these reasons, this work should be viewed as merely a
snapshot of an ongoing credit assignment process. To help improve it, please do not hesitate to send
corrections and suggestions to [email protected].
1
Contents
1
Introduction to Deep Learning (DL) in Neural Networks (NNs)
4
2
Event-Oriented Notation for Activation Spreading in NNs
5
3
Depth of Credit Assignment Paths (CAPs) and of Problems
6
4
Recurring Themes of Deep Learning
4.1 Dynamic Programming for Supervised/Reinforcement Learning (SL/RL)
4.2 Unsupervised Learning (UL) Facilitating SL and RL . . . . . . . . . . .
4.3 Learning Hierarchical Representations Through Deep SL, UL, RL . . . .
4.4 Occam’s Razor: Compression and Minimum Description Length (MDL) .
4.5 Fast Graphics Processing Units (GPUs) for DL in NNs . . . . . . . . . .
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Supervised NNs, Some Helped by Unsupervised NNs
5.1 Early NNs Since the 1940s (and the 1800s) . . . . . . . . . . . . . . . . . . . . . .
5.2 Around 1960: Visual Cortex Provides Inspiration for DL (Sec. 5.4, 5.11) . . . . . . .
5.3 1965: Deep Networks Based on the Group Method of Data Handling . . . . . . . . .
5.4 1979: Convolution + Weight Replication + Subsampling (Neocognitron) . . . . . .
5.5 1960-1981 and Beyond: Development of Backpropagation (BP) for NNs . . . . . . .
5.5.1 BP for Weight-Sharing Feedforward NNs (FNNs) and Recurrent NNs (RNNs)
5.6 Late 1980s-2000 and Beyond: Numerous Improvements of NNs . . . . . . . . . . .
5.6.1 Ideas for Dealing with Long Time Lags and Deep CAPs . . . . . . . . . . .
5.6.2 Better BP Through Advanced Gradient Descent (Compare Sec. 5.24) . . . .
5.6.3 Searching For Simple, Low-Complexity, Problem-Solving NNs (Sec. 5.24) .
5.6.4 Potential Benefits of UL for SL (Compare Sec. 5.7, 5.10, 5.15) . . . . . . . .
5.7 1987: UL Through Autoencoder (AE) Hierarchies (Compare Sec. 5.15) . . . . . . .
5.8 1989: BP for Convolutional NNs (CNNs, Sec. 5.4) . . . . . . . . . . . . . . . . . .
5.9 1991: Fundamental Deep Learning Problem of Gradient Descent . . . . . . . . . . .
5.10 1991: UL-Based History Compression Through a Deep Stack of RNNs . . . . . . .
5.11 1992: Max-Pooling (MP): Towards MPCNNs (Compare Sec. 5.16, 5.19) . . . . . . .
5.12 1994: Early Contest-Winning NNs . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.13 1995: Supervised Recurrent Very Deep Learner (LSTM RNN) . . . . . . . . . . . .
5.14 2003: More Contest-Winning/Record-Setting NNs; Successful Deep NNs . . . . . .
5.15 2006/7: UL For Deep Belief Networks / AE Stacks Fine-Tuned by BP . . . . . . . .
5.16 2006/7: Improved CNNs / GPU-CNNs / BP for MPCNNs / LSTM Stacks . . . . . .
5.17 2009: First Official Competitions Won by RNNs, and with MPCNNs . . . . . . . . .
5.18 2010: Plain Backprop (+ Distortions) on GPU Breaks MNIST Record . . . . . . . .
5.19 2011: MPCNNs on GPU Achieve Superhuman Vision Performance . . . . . . . . .
5.20 2011: Hessian-Free Optimization for RNNs . . . . . . . . . . . . . . . . . . . . . .
5.21 2012: First Contests Won on ImageNet, Object Detection, Segmentation . . . . . . .
5.22 2013-: More Contests and Benchmark Records . . . . . . . . . . . . . . . . . . . .
5.23 Currently Successful Techniques: LSTM RNNs and GPU-MPCNNs . . . . . . . . .
5.24 Recent Tricks for Improving SL Deep NNs (Compare Sec. 5.6.2, 5.6.3) . . . . . . .
5.25 Consequences for Neuroscience . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.26 DL with Spiking Neurons? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
7
7
7
8
8
8
8
9
10
10
10
11
11
12
12
13
14
14
15
16
16
17
18
18
19
21
21
22
22
23
23
24
24
25
26
27
28
28
6
DL in FNNs and RNNs for Reinforcement Learning (RL)
6.1 RL Through NN World Models Yields RNNs With Deep CAPs . . . . . . .
6.2 Deep FNNs for Traditional RL and Markov Decision Processes (MDPs) . .
6.3 Deep RL RNNs for Partially Observable MDPs (POMDPs) . . . . . . . . .
6.4 RL Facilitated by Deep UL in FNNs and RNNs . . . . . . . . . . . . . . .
6.5 Deep Hierarchical RL (HRL) and Subgoal Learning with FNNs and RNNs .
6.6 Deep RL by Direct NN Search / Policy Gradients / Evolution . . . . . . . .
6.7 Deep RL by Indirect Policy Search / Compressed NN Search . . . . . . . .
6.8 Universal RL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
29
29
30
31
31
31
32
33
33
7
Conclusion and Outlook
34
8
Acknowledgments
35
Abbreviations in Alphabetical Order
HTM: Hierarchical Temporal Memory
HMAX: Hierarchical Model “and X”
LSTM: Long Short-Term Memory (RNN)
MDL: Minimum Description Length
MDP: Markov Decision Process
MNIST: Mixed National Institute of Standards
and Technology Database
MP: Max-Pooling
MPCNN: Max-Pooling CNN
NE: NeuroEvolution
NEAT: NE of Augmenting Topologies
NES: Natural Evolution Strategies
NFQ: Neural Fitted Q-Learning
NN: Neural Network
OCR: Optical Character Recognition
PCC: Potential Causal Connection
PDCC: Potential Direct Causal Connection
PM: Predictability Minimization
POMDP: Partially Observable MDP
RAAM: Recursive Auto-Associative Memory
RBM: Restricted Boltzmann Machine
ReLU: Rectified Linear Unit
RL: Reinforcement Learning
RNN: Recurrent Neural Network
R-prop: Resilient Backpropagation
SL: Supervised Learning
SLIM NN: Self-Delimiting Neural Network
SOTA: Self-Organising Tree Algorithm
SVM: Support Vector Machine
TDNN: Time-Delay Neural Network
TIMIT: TI/SRI/MIT Acoustic-Phonetic Continuous Speech Corpus
UL: Unsupervised Learning
WTA: Winner-Take-All
AE: Autoencoder
AI: Artificial Intelligence
ANN: Artificial Neural Network
BFGS: Broyden-Fletcher-Goldfarb-Shanno
BNN: Biological Neural Network
BM: Boltzmann Machine
BP: Backpropagation
BRNN: Bi-directional Recurrent Neural Network
CAP: Credit Assignment Path
CEC: Constant Error Carousel
CFL: Context Free Language
CMA-ES: Covariance Matrix Estimation ES
CNN: Convolutional Neural Network
CoSyNE: Co-Synaptic Neuro-Evolution
CSL: Context Senistive Language
CTC : Connectionist Temporal Classification
DBN: Deep Belief Network
DCT: Discrete Cosine Transform
DL: Deep Learning
DP: Dynamic Programming
DS: Direct Policy Search
EA: Evolutionary Algorithm
EM: Expectation Maximization
ES: Evolution Strategy
FMS: Flat Minimum Search
FNN: Feedforward Neural Network
FSA: Finite State Automaton
GMDH: Group Method of Data Handling
GOFAI: Good Old-Fashioned AI
GP: Genetic Programming
GPU: Graphics Processing Unit
GPU-MPCNN: GPU-Based MPCNN
HMM: Hidden Markov Model
HRL: Hierarchical Reinforcement Learning
3
1
Introduction to Deep Learning (DL) in Neural Networks (NNs)
Which modifiable components of a learning system are responsible for its success or failure? What
changes to them improve performance? This has been called the fundamental credit assignment problem (Minsky, 1963). There are general credit assignment methods for universal problem solvers that
are time-optimal in various theoretical senses (Sec. 6.8). The present survey, however, will focus on
the narrower, but now commercially important, subfield of Deep Learning (DL) in Artificial Neural
Networks (NNs).
A standard neural network (NN) consists of many simple, connected processors called neurons,
each producing a sequence of real-valued activations. Input neurons get activated through sensors perceiving the environment, other neurons get activated through weighted connections from previously
active neurons (details in Sec. 2). Some neurons may influence the environment by triggering actions.
Learning or credit assignment is about finding weights that make the NN exhibit desired behavior,
such as driving a car. Depending on the problem and how the neurons are connected, such behavior
may require long causal chains of computational stages (Sec. 3), where each stage transforms (often in a non-linear way) the aggregate activation of the network. Deep Learning is about accurately
assigning credit across many such stages.
Shallow NN-like models with few such stages have been around for many decades if not centuries
(Sec. 5.1). Models with several successive nonlinear layers of neurons date back at least to the 1960s
(Sec. 5.3) and 1970s (Sec. 5.5). An efficient gradient descent method for teacher-based Supervised
Learning (SL) in discrete, differentiable networks of arbitrary depth called backpropagation (BP) was
developed in the 1960s and 1970s, and applied to NNs in 1981 (Sec. 5.5). BP-based training of deep
NNs with many layers, however, had been found to be difficult in practice by the late 1980s (Sec. 5.6),
and had become an explicit research subject by the early 1990s (Sec. 5.9). DL became practically feasible to some extent through the help of Unsupervised Learning (UL), e.g., Sec. 5.10 (1991), Sec. 5.15
(2006). The 1990s and 2000s also saw many improvements of purely supervised DL (Sec. 5). In the
new millennium, deep NNs have finally attracted wide-spread attention, mainly by outperforming alternative machine learning methods such as kernel machines (Vapnik, 1995; Schölkopf et al., 1998)
in numerous important applications. In fact, since 2009, supervised deep NNs have won many official
international pattern recognition competitions (e.g., Sec. 5.17, 5.19, 5.21, 5.22), achieving the first
superhuman visual pattern recognition results in limited domains (Sec. 5.19, 2011). Deep NNs also
have become relevant for the more general field of Reinforcement Learning (RL) where there is no
supervising teacher (Sec. 6).
Both feedforward (acyclic) NNs (FNNs) and recurrent (cyclic) NNs (RNNs) have won contests
(Sec. 5.12, 5.14, 5.17, 5.19, 5.21, 5.22). In a sense, RNNs are the deepest of all NNs (Sec. 3)—they
are general computers more powerful than FNNs, and can in principle create and process memories
of arbitrary sequences of input patterns (e.g., Siegelmann and Sontag, 1991; Schmidhuber, 1990a).
Unlike traditional methods for automatic sequential program synthesis (e.g., Waldinger and Lee, 1969;
Balzer, 1985; Soloway, 1986; Deville and Lau, 1994), RNNs can learn programs that mix sequential
and parallel information processing in a natural and efficient way, exploiting the massive parallelism
viewed as crucial for sustaining the rapid decline of computation cost observed over the past 75 years.
The rest of this paper is structured as follows. Sec. 2 introduces a compact, event-oriented notation
that is simple yet general enough to accommodate both FNNs and RNNs. Sec. 3 introduces the
concept of Credit Assignment Paths (CAPs) to measure whether learning in a given NN application is
of the deep or shallow type. Sec. 4 lists recurring themes of DL in SL, UL, and RL. Sec. 5 focuses
on SL and UL, and on how UL can facilitate SL, although pure SL has become dominant in recent
competitions (Sec. 5.17–5.23). Sec. 5 is arranged in a historical timeline format with subsections on
important inspirations and technical contributions. Sec. 6 on deep RL discusses traditional Dynamic
Programming (DP)-based RL combined with gradient-based search techniques for SL or UL in deep
4
NNs, as well as general methods for direct and indirect search in the weight space of deep FNNs and
RNNs, including successful policy gradient and evolutionary methods.
2
Event-Oriented Notation for Activation Spreading in NNs
Throughout this paper, let i, j, k, t, p, q, r denote positive integer variables assuming ranges implicit
in the given contexts. Let n, m, T denote positive integer constants.
An NN’s topology may change over time (e.g., Sec. 5.3, 5.6.3). At any given moment, it can
be described as a finite subset of units (or nodes or neurons) N = {u1 , u2 , . . . , } and a finite set
H ⊆ N × N of directed edges or connections between nodes. FNNs are acyclic graphs, RNNs cyclic.
The first (input) layer is the set of input units, a subset of N . In FNNs, the k-th layer (k > 1) is the set
of all nodes u ∈ N such that there is an edge path of length k − 1 (but no longer path) between some
input unit and u. There may be shortcut connections between distant layers. In sequence-processing,
fully connected RNNs, all units have connections to all non-input units.
The NN’s behavior or program is determined by a set of real-valued, possibly modifiable, parameters or weights wi (i = 1, . . . , n). We now focus on a single finite episode or epoch of information
processing and activation spreading, without learning through weight changes. The following slightly
unconventional notation is designed to compactly describe what is happening during the runtime of
the system.
During an episode, there is a partially causal sequence xt (t = 1, . . . , T ) of real values that I call
events. Each xt is either an input set by the environment, or the activation of a unit that may directly
depend on other xk (k < t) through a current NN topology-dependent set int of indices k representing
incoming causal connections or links. Let the function v encode topology information and map such
event index pairs (k, t) to weight indices.
For example, in the non-input case weQmay have xt =
P
ft (nett ) with real-valued nett =
k∈int xk wv(k,t) (additive case) or nett =
k∈int xk wv(k,t)
(multiplicative case), where ft is a typically nonlinear real-valued activation function such as tanh.
In many recent competition-winning NNs (Sec. 5.19, 5.21, 5.22) there also are events of the type
xt = maxk∈int (xk ); some network types may also use complex polynomial activation functions
(Sec. 5.3). xt may directly affect certain xk (k > t) through outgoing connections or links represented
through a current set outt of indices k with t ∈ ink . Some of the non-input events are called output
events.
Note that many of the xt may refer to different, time-varying activations of the same unit in
sequence-processing RNNs (e.g., Williams, 1989, “unfolding in time”), or also in FNNs sequentially
exposed to time-varying input patterns of a large training set encoded as input events. During an
episode, the same weight may get reused over and over again in topology-dependent ways, e.g., in
RNNs, or in convolutional NNs (Sec. 5.4, 5.8). I call this weight sharing across space and/or time.
Weight sharing may greatly reduce the NN’s descriptive complexity, which is the number of bits of
information required to describe the NN (Sec. 4.4).
In Supervised Learning (SL), certain NN output events xt may be associated with teacher-given,
real-valued labels or targets dt yielding errors et , e.g., et = 1/2(xt −dt )2 . A typical goal of supervised
NN training is to find weights that yield episodes with small total error E, the sum of all such et . The
hope is that the NN will generalize well in later episodes, causing only small errors on previously
unseen sequences of input events. Many alternative error functions for SL and UL are possible.
SL assumes that input events are independent of earlier output events (which may affect the environment through actions causing subsequent perceptions). This assumption does not hold in the
broader fields of Sequential Decision Making and Reinforcement Learning (RL) (Kaelbling et al.,
1996; Sutton and Barto, 1998; Hutter, 2005; Wiering and van Otterlo, 2012) (Sec. 6). In RL, some
of the input events may encode real-valued reward signals given by the environment, and a typical
5
goal is to find weights that yield episodes with a high sum of reward signals, through sequences of
appropriate output actions.
Sec. 5.5 will use the notation above to compactly describe a central algorithm of DL, namely,
backpropagation (BP) for supervised weight-sharing FNNs and RNNs. (FNNs may be viewed as
RNNs with certain fixed zero weights.) Sec. 6 will address the more general RL case.
3
Depth of Credit Assignment Paths (CAPs) and of Problems
To measure whether credit assignment in a given NN application is of the deep or shallow type, I
introduce the concept of Credit Assignment Paths or CAPs, which are chains of possibly causal links
between the events of Sec. 2, e.g., from input through hidden to output layers in FNNs, or through
transformations over time in RNNs.
Let us first focus on SL. Consider two events xp and xq (1 ≤ p < q ≤ T ). Depending on the
application, they may have a Potential Direct Causal Connection (PDCC) expressed by the Boolean
predicate pdcc(p, q), which is true if and only if p ∈ inq . Then the 2-element list (p, q) is defined to
be a CAP (a minimal one) from p to q. A learning algorithm may be allowed to change wv(p,q) to
improve performance in future episodes.
More general, possibly indirect, Potential Causal Connections (PCC) are expressed by the recursively defined Boolean predicate pcc(p, q), which in the SL case is true only if pdcc(p, q), or if
pcc(p, k) for some k and pdcc(k, q). In the latter case, appending q to any CAP from p to k yields a
CAP from p to q (this is a recursive definition, too). The set of such CAPs may be large but is finite.
Note that the same weight may affect many different PDCCs between successive events listed by a
given CAP, e.g., in the case of RNNs, or weight-sharing FNNs.
Suppose a CAP has the form (. . . , k, t, . . . , q), where k and t (possibly t = q) are the first successive elements with modifiable wv(k,t) . Then the length of the suffix list (t, . . . , q) is called the CAP’s
depth (which is 0 if there are no modifiable links at all). This depth limits how far backwards credit
assignment can move down the causal chain to find a modifiable weight.1
Suppose an episode and its event sequence x1 , . . . , xT satisfy a computable criterion used to
decide whether a given problem has been solved (e.g., total error E below some threshold). Then
the set of used weights is called a solution to the problem, and the depth of the deepest CAP within
the sequence is called the solution depth. There may be other solutions (yielding different event
sequences) with different depths. Given some fixed NN topology, the smallest depth of any solution
is called the problem depth.
Sometimes we also speak of the depth of an architecture: SL FNNs with fixed topology imply a
problem-independent maximal problem depth bounded by the number of non-input layers. Certain
SL RNNs with fixed weights for all connections except those to output units (Jaeger, 2001; Maass
et al., 2002; Jaeger, 2004; Schrauwen et al., 2007) have a maximal problem depth of 1, because only
the final links in the corresponding CAPs are modifiable. In general, however, RNNs may learn to
solve problems of potentially unlimited depth.
Note that the definitions above are solely based on the depths of causal chains, and agnostic to the
temporal distance between events. For example, shallow FNNs perceiving large “time windows” of
input events may correctly classify long input sequences through appropriate output events, and thus
solve shallow problems involving long time lags between relevant events.
At which problem depth does Shallow Learning end, and Deep Learning begin? Discussions with
DL experts have not yet yielded a conclusive response to this question. Instead of committing myself
1 An alternative would be to count only modifiable links when measuring depth. In many typical NN applications this would
not make a difference, but in some it would, e.g., Sec. 6.1.
6
to a precise answer, let me just define for the purposes of this overview: problems of depth > 10
require Very Deep Learning.
The difficulty of a problem may have little to do with its depth. Some NNs can quickly learn
to solve certain deep problems, e.g., through random weight guessing (Sec. 5.9) or other types of
direct search (Sec. 6.6) or indirect search (Sec. 6.7) in weight space, or through training an NN first
on shallow problems whose solutions may then generalize to deep problems, or through collapsing
sequences of (non)linear operations into a single (non)linear operation (but see an analysis of nontrivial aspects of deep linear networks, Baldi and Hornik, 1994, Section B). In general, however,
finding an NN that precisely models a given training set is an NP-complete problem (Judd, 1990;
Blum and Rivest, 1992), also in the case of deep NNs (Sı́ma, 1994; de Souto et al., 1999; Windisch,
2005); compare a survey of negative results (Sı́ma, 2002, Section 1).
Above we have focused on SL. In the more general case of RL in unknown environments, pcc(p, q)
is also true if xp is an output event and xq any later input event—any action may affect the environment
and thus any later perception. (In the real world, the environment may even influence non-input events
computed on a physical hardware entangled with the entire universe, but this is ignored here.) It is
possible to model and replace such unmodifiable environmental PCCs through a part of the NN that
has already learned to predict (through some of its units) input events (including reward signals) from
former input events and actions (Sec. 6.1). Its weights are frozen, but can help to assign credit to
other, still modifiable weights used to compute actions (Sec. 6.1). This approach may lead to very
deep CAPs though.
Some DL research is about automatically rephrasing problems such that their depth is reduced
(Sec. 4). In particular, sometimes UL is used to make SL problems less deep, e.g., Sec. 5.10. Often
Dynamic Programming (Sec. 4.1) is used to facilitate certain traditional RL problems, e.g., Sec. 6.2.
Sec. 5 focuses on CAPs for SL, Sec. 6 on the more complex case of RL.
4
4.1
Recurring Themes of Deep Learning
Dynamic Programming for Supervised/Reinforcement Learning (SL/RL)
One recurring theme of DL is Dynamic Programming (DP) (Bellman, 1957), which can help to facilitate credit assignment under certain assumptions. For example, in SL NNs, backpropagation itself
can be viewed as a DP-derived method (Sec. 5.5). In traditional RL based on strong Markovian assumptions, DP-derived methods can help to greatly reduce problem depth (Sec. 6.2). DP algorithms
are also essential for systems that combine concepts of NNs and graphical models, such as Hidden
Markov Models (HMMs) (Stratonovich, 1960; Baum and Petrie, 1966) and Expectation Maximization
(EM) (Dempster et al., 1977; Friedman et al., 2001), e.g., (Bottou, 1991; Bengio, 1991; Bourlard and
Morgan, 1994; Baldi and Chauvin, 1996; Jordan and Sejnowski, 2001; Bishop, 2006; Hastie et al.,
2009; Poon and Domingos, 2011; Dahl et al., 2012; Hinton et al., 2012a; Wu and Shao, 2014).
4.2
Unsupervised Learning (UL) Facilitating SL and RL
Another recurring theme is how UL can facilitate both SL (Sec. 5) and RL (Sec. 6). UL (Sec. 5.6.4)
is normally used to encode raw incoming data such as video or speech streams in a form that is more
convenient for subsequent goal-directed learning. In particular, codes that describe the original data in
a less redundant or more compact way can be fed into SL (Sec. 5.10, 5.15) or RL machines (Sec. 6.4),
whose search spaces may thus become smaller (and whose CAPs shallower) than those necessary for
dealing with the raw data. UL is closely connected to the topics of regularization and compression
(Sec. 4.4, 5.6.3).
7
4.3
Learning Hierarchical Representations Through Deep SL, UL, RL
Many methods of Good Old-Fashioned Artificial Intelligence (GOFAI) (Nilsson, 1980) as well as
more recent approaches to AI (Russell et al., 1995) and Machine Learning (Mitchell, 1997) learn
hierarchies of more and more abstract data representations. For example, certain methods of syntactic pattern recognition (Fu, 1977) such as grammar induction discover hierarchies of formal rules
to model observations. The partially (un)supervised Automated Mathematician / EURISKO (Lenat,
1983; Lenat and Brown, 1984) continually learns concepts by combining previously learnt concepts.
Such hierarchical representation learning (Ring, 1994; Bengio et al., 2013; Deng and Yu, 2014) is also
a recurring theme of DL NNs for SL (Sec. 5), UL-aided SL (Sec. 5.7, 5.10, 5.15), and hierarchical RL
(Sec. 6.5). Often, abstract hierarchical representations are natural by-products of data compression
(Sec. 4.4), e.g., Sec. 5.10.
4.4
Occam’s Razor: Compression and Minimum Description Length (MDL)
Occam’s razor favors simple solutions over complex ones. Given some programming language, the
principle of Minimum Description Length (MDL) can be used to measure the complexity of a solution candidate by the length of the shortest program that computes it (e.g., Solomonoff, 1964;
Kolmogorov, 1965b; Chaitin, 1966; Wallace and Boulton, 1968; Levin, 1973a; Solomonoff, 1978;
Rissanen, 1986; Blumer et al., 1987; Li and Vitányi, 1997; Grünwald et al., 2005). Some methods
explicitly take into account program runtime (Allender, 1992; Watanabe, 1992; Schmidhuber, 1997,
2002); many consider only programs with constant runtime, written in non-universal programming
languages (e.g., Rissanen, 1986; Hinton and van Camp, 1993). In the NN case, the MDL principle suggests that low NN weight complexity corresponds to high NN probability in the Bayesian
view (e.g., MacKay, 1992; Buntine and Weigend, 1991; Neal, 1995; De Freitas, 2003), and to high
generalization performance (e.g., Baum and Haussler, 1989), without overfitting the training data.
Many methods have been proposed for regularizing NNs, that is, searching for solution-computing
but simple, low-complexity SL NNs (Sec. 5.6.3) and RL NNs (Sec. 6.7). This is closely related to
certain UL methods (Sec. 4.2, 5.6.4).
4.5
Fast Graphics Processing Units (GPUs) for DL in NNs
While the previous millennium saw several attempts at creating fast NN-specific hardware (e.g., Jackel
et al., 1990; Faggin, 1992; Ramacher et al., 1993; Widrow et al., 1994; Heemskerk, 1995; Korkin et al.,
1997; Urlbe, 1999), and at exploiting standard hardware (e.g., Anguita et al., 1994; Muller et al., 1995;
Anguita and Gomes, 1996), the new millennium brought a DL breakthrough in form of cheap, multiprocessor graphics cards or GPUs. GPUs are widely used for video games, a huge and competitive
market that has driven down hardware prices. GPUs excel at the fast matrix and vector multiplications
required not only for convincing virtual realities but also for NN training, where they can speed up
learning by a factor of 50 and more. Some of the GPU-based FNN implementations (Sec. 5.16–5.19)
have greatly contributed to recent successes in contests for pattern recognition (Sec. 5.19–5.22), image
segmentation (Sec. 5.21), and object detection (Sec. 5.21–5.22).
5
Supervised NNs, Some Helped by Unsupervised NNs
The main focus of current practical applications is on Supervised Learning (SL), which has dominated recent pattern recognition contests (Sec. 5.17–5.23). Several methods, however, use additional
Unsupervised Learning (UL) to facilitate SL (Sec. 5.7, 5.10, 5.15). It does make sense to treat SL and
8
UL in the same section: often gradient-based methods, such as BP (Sec. 5.5.1), are used to optimize
objective functions of both UL and SL, and the boundary between SL and UL may blur, for example,
when it comes to time series prediction and sequence classification, e.g., Sec. 5.10, 5.12.
A historical timeline format will help to arrange subsections on important inspirations and technical contributions (although such a subsection may span a time interval of many years). Sec. 5.1 briefly
mentions early, shallow NN models since the 1940s (and 1800s), Sec. 5.2 additional early neurobiological inspiration relevant for modern Deep Learning (DL). Sec. 5.3 is about GMDH networks (since
1965), to my knowledge the first (feedforward) DL systems. Sec. 5.4 is about the relatively deep
Neocognitron NN (1979) which is very similar to certain modern deep FNN architectures, as it combines convolutional NNs (CNNs), weight pattern replication, and subsampling mechanisms. Sec. 5.5
uses the notation of Sec. 2 to compactly describe a central algorithm of DL, namely, backpropagation
(BP) for supervised weight-sharing FNNs and RNNs. It also summarizes the history of BP 1960-1981
and beyond. Sec. 5.6 describes problems encountered in the late 1980s with BP for deep NNs, and
mentions several ideas from the previous millennium to overcome them. Sec. 5.7 discusses a first hierarchical stack (1987) of coupled UL-based Autoencoders (AEs)—this concept resurfaced in the new
millennium (Sec. 5.15). Sec. 5.8 is about applying BP to CNNs (1989), which is important for today’s
DL applications. Sec. 5.9 explains BP’s Fundamental DL Problem (of vanishing/exploding gradients)
discovered in 1991. Sec. 5.10 explains how a deep RNN stack of 1991 (the History Compressor) pretrained by UL helped to solve previously unlearnable DL benchmarks requiring Credit Assignment
Paths (CAPs, Sec. 3) of depth 1000 and more. Sec. 5.11 discusses a particular winner-take-all (WTA)
method called Max-Pooling (MP, 1992) widely used in today’s deep FNNs. Sec. 5.12 mentions a
first important contest won by SL NNs in 1994. Sec. 5.13 describes a purely supervised DL RNN
(Long Short-Term Memory, LSTM, 1995) for problems of depth 1000 and more. Sec. 5.14 mentions
an early contest of 2003 won by an ensemble of shallow FNNs, as well as good pattern recognition
results with CNNs and deep FNNs and LSTM RNNs (2003). Sec. 5.15 is mostly about Deep Belief
Networks (DBNs, 2006) and related stacks of Autoencoders (AEs, Sec. 5.7), both pre-trained by UL to
facilitate subsequent BP-based SL (compare Sec. 5.6.1, 5.10). Sec. 5.16 mentions the first SL-based
GPU-CNNs (2006), BP-trained MPCNNs (2007), and LSTM stacks (2007). Sec. 5.17–5.22 focus on
official competitions with secret test sets won by (mostly purely supervised) deep NNs since 2009,
in sequence recognition, image classification, image segmentation, and object detection. Many RNN
results depended on LSTM (Sec. 5.13); many FNN results depended on GPU-based FNN code developed since 2004 (Sec. 5.16, 5.17, 5.18, 5.19), in particular, GPU-MPCNNs (Sec. 5.19). Sec. 5.24
mentions recent tricks for improving DL in NNs, many of them closely related to earlier tricks from
the previous millennium (e.g., Sec. 5.6.2, 5.6.3). Sec. 5.25 discusses how artificial NNs can help to
understand biological NNs; Sec. 5.26 addresses the possibility of DL in NNs with spiking neurons.
5.1
Early NNs Since the 1940s (and the 1800s)
Early NN architectures (McCulloch and Pitts, 1943) did not learn. The first ideas about UL were
published a few years later (Hebb, 1949). The following decades brought simple NNs trained by
SL (e.g., Rosenblatt, 1958, 1962; Widrow and Hoff, 1962; Narendra and Thathatchar, 1974) and
UL (e.g., Grossberg, 1969; Kohonen, 1972; von der Malsburg, 1973; Willshaw and von der Malsburg,
1976), as well as closely related associative memories (e.g., Palm, 1980; Hopfield, 1982).
In a sense NNs have been around even longer, since early supervised NNs were essentially variants
of linear regression methods going back at least to the early 1800s (e.g., Legendre, 1805; Gauss, 1809,
1821); Gauss also refers to his work of 1795. Early NNs had a maximal CAP depth of 1 (Sec. 3).
9
5.2
Around 1960: Visual Cortex Provides Inspiration for DL (Sec. 5.4, 5.11)
Simple cells and complex cells were found in the cat’s visual cortex (e.g., Hubel and Wiesel, 1962;
Wiesel and Hubel, 1959). These cells fire in response to certain properties of visual sensory inputs,
such as the orientation of edges. Complex cells exhibit more spatial invariance than simple cells. This
inspired later deep NN architectures (Sec. 5.4, 5.11) used in certain modern award-winning Deep
Learners (Sec. 5.19–5.22).
5.3
1965: Deep Networks Based on the Group Method of Data Handling
Networks trained by the Group Method of Data Handling (GMDH) (Ivakhnenko and Lapa, 1965;
Ivakhnenko et al., 1967; Ivakhnenko, 1968, 1971) were perhaps the first DL systems of the Feedforward Multilayer Perceptron type, although there was earlier work on NNs with a single hidden
layer (e.g., Joseph, 1961; Viglione, 1970). The units of GMDH nets may have polynomial activation
functions implementing Kolmogorov-Gabor polynomials (more general than other widely used NN
activation functions, Sec. 2). Given a training set, layers are incrementally grown and trained by regression analysis (e.g., Legendre, 1805; Gauss, 1809, 1821) (Sec. 5.1), then pruned with the help of
a separate validation set (using today’s terminology), where Decision Regularisation is used to weed
out superfluous units (compare Sec. 5.6.3). The numbers of layers and units per layer can be learned
in problem-dependent fashion. To my knowledge, this was the first example of open-ended, hierarchical representation learning in NNs (Sec. 4.3). A paper of 1971 already described a deep GMDH
network with 8 layers (Ivakhnenko, 1971). There have been numerous applications of GMDH-style
nets, e.g. (Ikeda et al., 1976; Farlow, 1984; Madala and Ivakhnenko, 1994; Ivakhnenko, 1995; Kondo,
1998; Kordı́k et al., 2003; Witczak et al., 2006; Kondo and Ueno, 2008).
5.4
1979: Convolution + Weight Replication + Subsampling (Neocognitron)
Apart from deep GMDH networks (Sec. 5.3), the Neocognitron (Fukushima, 1979, 1980, 2013a)
was perhaps the first artificial NN that deserved the attribute deep, and the first to incorporate the
neurophysiological insights of Sec. 5.2. It introduced convolutional NNs (today often called CNNs or
convnets), where the (typically rectangular) receptive field of a convolutional unit with given weight
vector (a filter) is shifted step by step across a 2-dimensional array of input values, such as the pixels
of an image (usually there are several such filters). The resulting 2D array of subsequent activation
events of this unit can then provide inputs to higher-level units, and so on. Due to massive weight
replication (Sec. 2), relatively few parameters (Sec. 4.4) may be necessary to describe the behavior of
such a convolutional layer.
Subsampling or downsampling layers consist of units whose fixed-weight connections originate
from physical neighbours in the convolutional layers below. Subsampling units become active if at
least one of their inputs is active; their responses are insensitive to certain small image shifts (compare
Sec. 5.2).
The Neocognitron is very similar to the architecture of modern, contest-winning, purely supervised, feedforward, gradient-based Deep Learners with alternating convolutional and downsampling
layers (e.g., Sec. 5.19–5.22). Fukushima, however, did not set the weights by supervised backpropagation (Sec. 5.5, 5.8), but by local, WTA-based unsupervised learning rules (e.g., Fukushima, 2013b),
or by pre-wiring. In that sense he did not care for the DL problem (Sec. 5.9), although his architecture
was comparatively deep indeed. For downsampling purposes he used Spatial Averaging (Fukushima,
1980, 2011) instead of Max-Pooling (MP, Sec. 5.11), currently a particularly convenient and popular
WTA mechanism. Today’s DL combinations of CNNs and MP and BP also profit a lot from later
work (e.g., Sec. 5.8, 5.16, 5.16, 5.19).
10
5.5
1960-1981 and Beyond: Development of Backpropagation (BP) for NNs
The minimisation of errors through gradient descent (Hadamard, 1908) in the parameter space of
complex, nonlinear, differentiable (Leibniz, 1684), multi-stage, NN-related systems has been discussed at least since the early 1960s (e.g., Kelley, 1960; Bryson, 1961; Bryson and Denham, 1961;
Pontryagin et al., 1961; Dreyfus, 1962; Wilkinson, 1965; Amari, 1967; Bryson and Ho, 1969; Director and Rohrer, 1969), initially within the framework of Euler-LaGrange equations in the Calculus of
Variations (e.g., Euler, 1744).
Steepest descent in the weight space of such systems can be performed (Bryson, 1961; Kelley,
1960; Bryson and Ho, 1969) by iterating the chain rule (Leibniz, 1676; L’Hôpital, 1696) à la Dynamic
Programming (DP) (Bellman, 1957). A simplified derivation of this backpropagation method uses the
chain rule only (Dreyfus, 1962).
The systems of the 1960s were already efficient in the DP sense. However, they backpropagated
derivative information through standard Jacobian matrix calculations from one “layer” to the previous
one, without explicitly addressing either direct links across several layers or potential additional efficiency gains due to network sparsity (but perhaps such enhancements seemed obvious to the authors).
Given all the prior work on learning in multilayer NN-like systems (see also Sec. 5.3 on deep nonlinear nets since 1965), it seems surprising in hindsight that a book (Minsky and Papert, 1969) on the
limitations of simple linear perceptrons with a single layer (Sec. 5.1) discouraged some researchers
from further studying NNs.
Explicit, efficient error backpropagation (BP) in arbitrary, discrete, possibly sparsely connected,
NN-like networks apparently was first described in a 1970 master’s thesis (Linnainmaa, 1970, 1976),
albeit without reference to NNs. BP is also known as the reverse mode of automatic differentiation (Griewank, 2012), where the costs of forward activation spreading essentially equal the costs of
backward derivative calculation. See early FORTRAN code (Linnainmaa, 1970) and closely related
work (Ostrovskii et al., 1971).
Efficient BP was soon explicitly used to minimize cost functions by adapting control parameters
(weights) (Dreyfus, 1973). Compare some preliminary, NN-specific discussion (Werbos, 1974, section 5.5.1), a method for multilayer threshold NNs (Bobrowski, 1978), and a computer program for
automatically deriving and implementing BP for given differentiable systems (Speelpenning, 1980).
To my knowledge, the first NN-specific application of efficient BP as above was described in
1981 (Werbos, 1981, 2006). Related work was published several years later (Parker, 1985; LeCun,
1985, 1988). A paper of 1986 significantly contributed to the popularisation of BP for NNs (Rumelhart
et al., 1986), experimentally demonstrating the emergence of useful internal representations in hidden
layers. See generalisations for sequence-processing recurrent NNs (e.g., Williams, 1989; Robinson
and Fallside, 1987; Werbos, 1988; Williams and Zipser, 1988, 1989b,a; Rohwer, 1989; Pearlmutter,
1989; Gherrity, 1989; Williams and Peng, 1990; Schmidhuber, 1992a; Pearlmutter, 1995; Baldi, 1995;
Kremer and Kolen, 2001; Atiya and Parlos, 2000), also for equilibrium RNNs (Almeida, 1987; Pineda,
1987) with stationary inputs.
5.5.1
BP for Weight-Sharing Feedforward NNs (FNNs) and Recurrent NNs (RNNs)
Using the notation of Sec. 2 for weight-sharing FNNs or RNNs, after an episode of activation spreading through differentiable ft , P
a single iteration of gradient descent through BP computes changes of
∂E ∂nett
∂E
=
all wi in proportion to ∂w
t ∂nett ∂wi as in Algorithm 5.5.1 (for the additive case), where each
i
weight wi is associated with a real-valued variable 4i initialized by 0.
The computational costs of the backward (BP) pass are essentially those of the forward pass
(Sec. 2). Forward and backward passes are re-iterated until sufficient performance is reached.
11
Alg. 5.5.1: One iteration of BP for weight-sharing FNNs or RNNs
for t = T, . . . , 1 do
∂E
, inititalize real-valued error signal variable δt by 0;
to compute ∂net
t
if xt is an input event then continue with next iteration;
if there is an error et P
then δt := xt − dt ;
add to δt the value k∈outt wv(t,k) δk ; (this is the elegant and efficient recursive chain rule
application collecting impacts of nett on future events)
multiply δt by ft0 (nett );
for all k ∈ int add to 4wv(k,t) the value xk δt
end for
change each wi in proportion to 4i and a small real-valued learning rate
As of 2014, this simple BP method is still the central learning algorithm for FNNs and RNNs. Notably, most contest-winning NNs up to 2014 (Sec. 5.12, 5.14, 5.17, 5.19, 5.21, 5.22) did not augment
supervised BP by some sort of unsupervised learning as discussed in Sec. 5.7, 5.10, 5.15.
5.6
Late 1980s-2000 and Beyond: Numerous Improvements of NNs
By the late 1980s it seemed clear that BP by itself (Sec. 5.5) was no panacea. Most FNN applications
focused on FNNs with few hidden layers. Additional hidden layers often did not seem to offer empirical benefits. Many practitioners found solace in a theorem (Kolmogorov, 1965a; Hecht-Nielsen, 1989;
Hornik et al., 1989) stating that an NN with a single layer of enough hidden units can approximate
any multivariate continous function with arbitrary accuracy.
Likewise, most RNN applications did not require backpropagating errors far. Many researchers
helped their RNNs by first training them on shallow problems (Sec. 3) whose solutions then generalized to deeper problems. In fact, some popular RNN algorithms restricted credit assignment to a
single step backwards (Elman, 1990; Jordan, 1986, 1997), also in more recent studies (Jaeger, 2001;
Maass et al., 2002; Jaeger, 2004).
Generally speaking, although BP allows for deep problems in principle, it seemed to work only
for shallow problems. The late 1980s and early 1990s saw a few ideas with a potential to overcome
this problem, which was fully understood only in 1991 (Sec. 5.9).
5.6.1
Ideas for Dealing with Long Time Lags and Deep CAPs
To deal with long time lags between relevant events, several sequence processing methods were proposed, including Focused BP based on decay factors for activations of units in RNNs (Mozer, 1989,
1992), Time-Delay Neural Networks (TDNNs) (Lang et al., 1990) and their adaptive extension (Bodenhausen and Waibel, 1991), Nonlinear AutoRegressive with eXogenous inputs (NARX) RNNs (Lin
et al., 1996), certain hierarchical RNNs (Hihi and Bengio, 1996) (compare Sec. 5.10, 1991), RL
economies in RNNs with WTA units and local learning rules (Schmidhuber, 1989b), and other methods (e.g., Ring, 1993, 1994; Plate, 1993; de Vries and Principe, 1991; Sun et al., 1993a; Bengio
et al., 1994). However, these algorithms either worked for shallow CAPs only, could not generalize
to unseen CAP depths, had problems with greatly varying time lags between relevant events, needed
external fine tuning of delay constants, or suffered from other problems. In fact, it turned out that
certain simple but deep benchmark problems used to evaluate such methods are more quickly solved
by randomly guessing RNN weights until a solution is found (Hochreiter and Schmidhuber, 1996).
While the RNN methods above were designed for DL of temporal sequences, the Neural Heat
Exchanger (Schmidhuber, 1990c) consists of two parallel deep FNNs with opposite flow directions.
12
Input patterns enter the first FNN and are propagated “up”. Desired outputs (targets) enter the “opposite” FNN and are propagated “down”. Using a local learning rule, each layer in each net tries to be
similar (in information content) to the preceding layer and to the adjacent layer of the other net. The
input entering the first net slowly “heats up” to become the target. The target entering the opposite net
slowly “cools down” to become the input. The Helmholtz Machine (Dayan et al., 1995; Dayan and
Hinton, 1996) may be viewed as an unsupervised (Sec. 5.6.4) variant thereof (Peter Dayan, personal
communication, 1994).
A hybrid approach (Shavlik and Towell, 1989; Towell and Shavlik, 1994) initializes a potentially deep FNN through a domain theory in propositional logic, which may be acquired through
explanation-based learning (Mitchell et al., 1986; DeJong and Mooney, 1986; Minton et al., 1989).
The NN is then fine-tuned through BP (Sec. 5.5). The NN’s depth reflects the longest chain of
reasoning in the original set of logical rules. An extension of this approach (Maclin and Shavlik,
1993; Shavlik, 1994) initializes an RNN by domain knowledge expressed as a Finite State Automaton (FSA). BP-based fine-tuning has become important for later DL systems pre-trained by UL, e.g.,
Sec. 5.10, 5.15.
5.6.2
Better BP Through Advanced Gradient Descent (Compare Sec. 5.24)
Numerous improvements of steepest descent through BP (Sec. 5.5) have been proposed. Leastsquares methods (Gauss-Newton, Levenberg-Marquardt) (Gauss, 1809; Newton, 1687; Levenberg,
1944; Marquardt, 1963; Schaback and Werner, 1992) and quasi-Newton methods (Broyden-FletcherGoldfarb-Shanno, BFGS) (Broyden et al., 1965; Fletcher and Powell, 1963; Goldfarb, 1970; Shanno,
1970) are computationally too expensive for large NNs. Partial BFGS (Battiti, 1992; Saito and
Nakano, 1997) and conjugate gradient (Hestenes and Stiefel, 1952; Møller, 1993) as well as other
methods (Solla, 1988; Schmidhuber, 1989a; Cauwenberghs, 1993) provide sometimes useful fast alternatives. BP can be treated as a linear least-squares problem (Biegler-König and Bärmann, 1993),
where second-order gradient information is passed back to preceding layers.
To speed up BP, momentum was introduced (Rumelhart et al., 1986), ad-hoc constants were added
to the slope of the linearized activation function (Fahlman, 1988), or the nonlinearity of the slope was
exaggerated (West and Saad, 1995).
Only the signs of the error derivatives are taken into account by the successful and widely used
BP variant R-prop (Riedmiller and Braun, 1993) and the robust variation iRprop+ (Igel and Hüsken,
2003), which was also successfully applied to RNNs.
The local gradient can be normalized based on the NN architecture (Schraudolph and Sejnowski,
1996), through a diagonalized Hessian approach (Becker and Le Cun, 1989), or related efficient methods (Schraudolph, 2002).
Some algorithms for controlling BP step size adapt a global learning rate (Lapedes and Farber,
1986; Vogl et al., 1988; Battiti, 1989; LeCun et al., 1993; Yu et al., 1995), while others compute individual learning rates for each weight (Jacobs, 1988; Silva and Almeida, 1990). In online learning,
where BP is applied after each pattern presentation, the vario-η algorithm (Neuneier and Zimmermann, 1996) sets each weight’s learning rate inversely proportional to the empirical standard deviation of its local gradient, thus normalizing the stochastic weight fluctuations. Compare a local online
step size adaptation method for nonlinear NNs (Almeida et al., 1997).
Many additional tricks for improving NNs have been described (e.g., Orr and Müller, 1998; Montavon et al., 2012). Compare Sec. 5.6.3 and recent developments mentioned in Sec. 5.24.
13
5.6.3
Searching For Simple, Low-Complexity, Problem-Solving NNs (Sec. 5.24)
Many researchers used BP-like methods to search for “simple,” low-complexity NNs (Sec. 4.4)
with high generalization capability. Most approaches address the bias/variance dilemma (Geman
et al., 1992) through strong prior assumptions. For example, weight decay (Hanson and Pratt, 1989;
Weigend et al., 1991; Krogh and Hertz, 1992) encourages near-zero weights, by penalizing large
weights. In a Bayesian framework (Bayes, 1763), weight decay can be derived (Hinton and van
Camp, 1993) from Gaussian or Laplacian weight priors (Gauss, 1809; Laplace, 1774); see also (Murray and Edwards, 1993). An extension of this approach postulates that a distribution of networks with
many similar weights generated by Gaussian mixtures is “better” a priori (Nowlan and Hinton, 1992).
Often weight priors are implicit in additional penalty terms (MacKay, 1992) or in methods based
on validation sets (Mosteller and Tukey, 1968; Stone, 1974; Eubank, 1988; Hastie and Tibshirani,
1990; Craven and Wahba, 1979; Golub et al., 1979), Akaike’s information criterion and final prediction error (Akaike, 1970, 1973, 1974), or generalized prediction error (Moody and Utans, 1994;
Moody, 1992). See also (Holden, 1994; Wang et al., 1994; Amari and Murata, 1993; Wang et al.,
1994; Guyon et al., 1992; Vapnik, 1992; Wolpert, 1994). Similar priors (or biases towards simplicity)
are implicit in constructive and pruning algorithms, e.g., layer-by-layer sequential network construction (e.g., Ivakhnenko, 1968, 1971; Ash, 1989; Moody, 1989; Gallant, 1988; Honavar and Uhr, 1988;
Ring, 1991; Fahlman, 1991; Weng et al., 1992; Honavar and Uhr, 1993; Burgess, 1994; Fritzke, 1994;
Parekh et al., 2000; Utgoff and Stracuzzi, 2002) (see also Sec. 5.3, 5.11), input pruning (Moody, 1992;
Refenes et al., 1994), unit pruning (e.g., Ivakhnenko, 1968, 1971; White, 1989; Mozer and Smolensky, 1989; Levin et al., 1994), weight pruning, e.g., optimal brain damage (LeCun et al., 1990b), and
optimal brain surgeon (Hassibi and Stork, 1993).
A very general but not always practical approach for discovering low-complexity SL NNs or
RL NNs searches among weight matrix-computing programs written in a universal programming
language, with a bias towards fast and short programs (Schmidhuber, 1997) (Sec. 6.7).
Flat Minimum Search (FMS) (Hochreiter and Schmidhuber, 1997a, 1999) searches for a “flat”
minimum of the error function: a large connected region in weight space where error is low and remains approximately constant, that is, few bits of information are required to describe low-precision
weights with high variance. Compare perturbation tolerance conditions (Minai and Williams, 1994;
Murray and Edwards, 1993; Hanson, 1990; Neti et al., 1992; Matsuoka, 1992; Bishop, 1993; Kerlirzin and Vallet, 1993; Carter et al., 1990). An MDL-based, Bayesian argument suggests that flat
minima correspond to “simple” NNs and low expected overfitting. Compare Sec. 5.6.4 and more
recent developments mentioned in Sec. 5.24.
5.6.4
Potential Benefits of UL for SL (Compare Sec. 5.7, 5.10, 5.15)
The notation of Sec. 2 introduced teacher-given labels dt . Many papers of the previous millennium,
however, were about unsupervised learning (UL) without a teacher (e.g., Hebb, 1949; von der Malsburg, 1973; Kohonen, 1972, 1982, 1988; Willshaw and von der Malsburg, 1976; Grossberg, 1976a,b;
Watanabe, 1985; Pearlmutter and Hinton, 1986; Barrow, 1987; Field, 1987; Oja, 1989; Barlow et al.,
1989; Baldi and Hornik, 1989; Sanger, 1989; Ritter and Kohonen, 1989; Rubner and Schulten, 1990;
Földiák, 1990; Martinetz et al., 1990; Kosko, 1990; Mozer, 1991; Palm, 1992; Atick et al., 1992;
Miller, 1994; Saund, 1994; Földiák and Young, 1995; Deco and Parra, 1997); see also post-2000
work (e.g., Carreira-Perpinan, 2001; Wiskott and Sejnowski, 2002; Franzius et al., 2007; Waydo and
Koch, 2008).
Many UL methods are designed to maximize entropy-related, information-theoretic (Boltzmann,
1909; Shannon, 1948; Kullback and Leibler, 1951) objectives (e.g., Linsker, 1988; Barlow et al., 1989;
MacKay and Miller, 1990; Plumbley, 1991; Schmidhuber, 1992b,c; Schraudolph and Sejnowski,
14
1993; Redlich, 1993; Zemel, 1993; Zemel and Hinton, 1994; Field, 1994; Hinton et al., 1995; Dayan
and Zemel, 1995; Amari et al., 1996; Deco and Parra, 1997).
Many do this to uncover and disentangle hidden underlying sources of signals (e.g., Jutten and
Herault, 1991; Schuster, 1992; Andrade et al., 1993; Molgedey and Schuster, 1994; Comon, 1994;
Cardoso, 1994; Bell and Sejnowski, 1995; Karhunen and Joutsensalo, 1995; Belouchrani et al., 1997;
Hyvärinen et al., 2001; Szabó et al., 2006; Shan et al., 2007; Shan and Cottrell, 2014).
Many UL methods automatically and robustly generate distributed, sparse representations of input patterns (Földiák, 1990; Hinton and Ghahramani, 1997; Lewicki and Olshausen, 1998; Hyvärinen
et al., 1999; Hochreiter and Schmidhuber, 1999; Falconbridge et al., 2006) through well-known feature detectors (e.g., Olshausen and Field, 1996; Schmidhuber et al., 1996), such as off-center-onsurround-like structures, as well as orientation sensitive edge detectors and Gabor filters (Gabor,
1946). They extract simple features related to those observed in early visual pre-processing stages
of biological systems (e.g., De Valois et al., 1982; Jones and Palmer, 1987).
UL can also serve to extract invariant features from different data items (e.g., Becker, 1991)
through coupled NNs observing two different inputs (Schmidhuber and Prelinger, 1992), also called
Siamese NNs (e.g., Bromley et al., 1993; Hadsell et al., 2006; Taylor et al., 2011; Chen and Salman,
2011).
UL can help to encode input data in a form advantageous for further processing. In the context
of DL, one important goal of UL is redundancy reduction. Ideally, given an ensemble of input patterns, redundancy reduction through a deep NN will create a factorial code (a code with statistically
independent components) of the ensemble (Barlow et al., 1989; Barlow, 1989), to disentangle the
unknown factors of variation (compare Bengio et al., 2013). Such codes may be sparse and can be
advantageous for (1) data compression, (2) speeding up subsequent BP (Becker, 1991), (3) trivialising
the task of subsequent naive yet optimal Bayes classifiers (Schmidhuber et al., 1996).
Most early UL FNNs had a single layer. Methods for deeper UL FNNs include hierarchical
(Sec. 4.3) self-organizing Kohonen maps (e.g., Koikkalainen and Oja, 1990; Lampinen and Oja, 1992;
Versino and Gambardella, 1996; Dittenbach et al., 2000; Rauber et al., 2002), hierarchical Gaussian
potential function networks (Lee and Kil, 1991), layer-wise UL of feature hierarchies fed into SL
classifiers (Behnke, 1999, 2003a), the Self-Organising Tree Algorithm (SOTA) (Herrero et al., 2001),
and nonlinear Autoencoders (AEs) with more than 3 (e.g., 5) layers (Kramer, 1991; Oja, 1991; DeMers
and Cottrell, 1993). Such AE NNs (Rumelhart et al., 1986) can be trained to map input patterns
to themselves, for example, by compactly encoding them through activations of units of a narrow
bottleneck hidden layer. Certain nonlinear AEs suffer from certain limitations (Baldi, 2012).
L OCOCODE (Hochreiter and Schmidhuber, 1999) uses FMS (Sec. 5.6.3) to find low-complexity
AEs with low-precision weights describable by few bits of information, often producing sparse or
factorial codes. Predictability Minimization (PM) (Schmidhuber, 1992c) searches for factorial codes
through nonlinear feature detectors that fight nonlinear predictors, trying to become both as informative and as unpredictable as possible. PM-based UL was applied not only to FNNs but also to
RNNs (e.g., Schmidhuber, 1993b; Lindstädt, 1993). Compare Sec. 5.10 on UL-based RNN stacks
(1991), as well as later UL RNNs (e.g., Klapper-Rybicka et al., 2001; Steil, 2007).
5.7
1987: UL Through Autoencoder (AE) Hierarchies (Compare Sec. 5.15)
Perhaps the first work to study potential benefits of UL-based pre-training was published in 1987. It
proposed unsupervised AE hierarchies (Ballard, 1987), closely related to certain post-2000 feedforward Deep Learners based on UL (Sec. 5.15). The lowest-level AE NN with a single hidden layer is
trained to map input patterns to themselves. Its hidden layer codes are then fed into a higher-level AE
of the same type, and so on. The hope is that the codes in the hidden AE layers have properties that
15
facilitate subsequent learning. In one experiment, a particular AE-specific learning algorithm (different from traditional BP of Sec. 5.5.1) was used to learn a mapping in an AE stack pre-trained by
this type of UL (Ballard, 1987). This was faster than learning an equivalent mapping by BP through
a single deeper AE without pre-training. On the other hand, the task did not really require a deep
AE, that is, the benefits of UL were not that obvious from this experiment. Compare an early survey (Hinton, 1989) and the somewhat related Recursive Auto-Associative Memory (RAAM) (Pollack,
1988, 1990; Melnik et al., 2000), originally used to encode sequential linguistic structures of arbitrary
size through a fixed number of hidden units. More recently, RAAMs were also used as unsupervised
pre-processors to facilitate deep credit assignment for RL (Gisslen et al., 2011) (Sec. 6.4).
In principle, many UL methods (Sec. 5.6.4) could be stacked like the AEs above, the historycompressing RNNs of Sec. 5.10, the Restricted Boltzmann Machines (RBMs) of Sec. 5.15, or hierarchical Kohonen nets (Sec. 5.6.4), to facilitate subsequent SL. Compare Stacked Generalization (Wolpert, 1992; Ting and Witten, 1997), and FNNs that profit from pre-training by competitive
UL (e.g., Rumelhart and Zipser, 1986) prior to BP-based fine-tuning (Maclin and Shavlik, 1995). See
also more recent methods using UL to improve subsequent SL (e.g., Behnke, 1999, 2003a; EscalanteB. and Wiskott, 2013).
5.8
1989: BP for Convolutional NNs (CNNs, Sec. 5.4)
In 1989, backpropagation (Sec. 5.5) was applied (LeCun et al., 1989, 1990a, 1998) to Neocognitronlike, weight-sharing, convolutional neural layers (Sec. 5.4) with adaptive connections. This combination, augmented by Max-Pooling (MP, Sec. 5.11, 5.16), and sped up on graphics cards (Sec. 5.19),
has become an essential ingredient of many modern, competition-winning, feedforward, visual Deep
Learners (Sec. 5.19–5.23). This work also introduced the MNIST data set of handwritten digits (LeCun et al., 1989), which over time has become perhaps the most famous benchmark of Machine
Learning. CNNs helped to achieve good performance on MNIST (LeCun et al., 1990a) (CAP depth
5) and on fingerprint recognition (Baldi and Chauvin, 1993); similar CNNs were used commercially
in the 1990s.
5.9
1991: Fundamental Deep Learning Problem of Gradient Descent
A diploma thesis (Hochreiter, 1991) represented a milestone of explicit DL research. As mentioned
in Sec. 5.6, by the late 1980s, experiments had indicated that traditional deep feedforward or recurrent networks are hard to train by backpropagation (BP) (Sec. 5.5). Hochreiter’s work formally
identified a major reason: Typical deep NNs suffer from the now famous problem of vanishing or
exploding gradients. With standard activation functions (Sec. 1), cumulative backpropagated error
signals (Sec. 5.5.1) either shrink rapidly, or grow out of bounds. In fact, they decay exponentially in
the number of layers or CAP depth (Sec. 3), or they explode. This is also known as the long time
lag problem. Much subsequent DL research of the 1990s and 2000s was motivated by this insight.
Later work (Bengio et al., 1994) also studied basins of attraction and their stability under noise from a
dynamical systems point of view: either the dynamics are not robust to noise, or the gradients vanish.
See also (Hochreiter et al., 2001a; Tiňo and Hammer, 2004). Over the years, several ways of partially
overcoming the Fundamental Deep Learning Problem were explored:
I A Very Deep Learner of 1991 (the History Compressor, Sec. 5.10) alleviates the problem
through unsupervised pre-training for a hierarchy of RNNs. This greatly facilitates subsequent
supervised credit assignment through BP (Sec. 5.5). In the FNN case, similar effects can be
achieved through conceptually related AE stacks (Sec. 5.7, 5.15) and Deep Belief Networks
(DBNs, Sec. 5.15).
16
II LSTM-like networks (Sec. 5.13, 5.16, 5.17, 5.21–5.23) alleviate the problem through a special
architecture unaffected by it.
III Today’s GPU-based computers have a million times the computational power of desktop machines of the early 1990s. This allows for propagating errors a few layers further down within
reasonable time, even in traditional NNs (Sec. 5.18). That is basically what is winning many of
the image recognition competitions now (Sec. 5.19, 5.21, 5.22). (Although this does not really
overcome the problem in a fundamental way.)
IV Hessian-free optimization (Sec. 5.6.2) can alleviate the problem for FNNs (Møller, 1993;
Pearlmutter, 1994; Schraudolph, 2002; Martens, 2010) (Sec. 5.6.2) and RNNs (Martens and
Sutskever, 2011) (Sec. 5.20).
V The space of NN weight matrices can also be searched without relying on error gradients,
thus avoiding the Fundamental Deep Learning Problem altogether. Random weight guessing
sometimes works better than more sophisticated methods (Hochreiter and Schmidhuber, 1996).
Certain more complex problems are better solved by using Universal Search (Levin, 1973b)
for weight matrix-computing programs written in a universal programming language (Schmidhuber, 1997). Some are better solved by using linear methods to obtain optimal weights for
connections to output events (Sec. 2), and evolving weights of connections to other events—
this is called Evolino (Schmidhuber et al., 2007). Compare also related RNNs pre-trained by
certain UL rules (Steil, 2007), also in the case of spiking neurons (Yin et al., 2012; Klampfl and
Maass, 2013) (Sec. 5.26). Direct search methods are relevant not only for SL but also for more
general RL, and are discussed in more detail in Sec. 6.6.
5.10
1991: UL-Based History Compression Through a Deep Stack of RNNs
A working Very Deep Learner (Sec. 3) of 1991 (Schmidhuber, 1992b, 2013a) could perform credit assignment across hundreds of nonlinear operators or neural layers, by using unsupervised pre-training
for a hierarchy of RNNs.
The basic idea is still relevant today. Each RNN is trained for a while in unsupervised fashion to
predict its next input (e.g., Connor et al., 1994; Dorffner, 1996). From then on, only unexpected inputs
(errors) convey new information and get fed to the next higher RNN which thus ticks on a slower, selforganising time scale. It can easily be shown that no information gets lost. It just gets compressed
(much of machine learning is essentially about compression, e.g., Sec. 4.4, 5.6.3, 6.7). For each
individual input sequence, we get a series of less and less redundant encodings in deeper and deeper
levels of this History Compressor or Neural Sequence Chunker, which can compress data in both
space (like feedforward NNs) and time. This is another good example of hierarchical representation
learning (Sec. 4.3). There also is a continuous variant of the history compressor (Schmidhuber et al.,
1993).
The RNN stack is essentially a deep generative model of the data, which can be reconstructed from
its compressed form. Adding another RNN to the stack improves a bound on the data’s description
length—equivalent to the negative logarithm of its probability (Huffman, 1952; Shannon, 1948)—as
long as there is remaining local learnable predictability in the data representation on the corresponding
level of the hierarchy. Compare a similar observation for feedforward Deep Belief Networks (DBNs,
2006, Sec. 5.15).
The system was able to learn many previously unlearnable DL tasks. One ancient illustrative
DL experiment (Schmidhuber, 1993b) required CAPs (Sec. 3) of depth 1200. The top level code of
the initially unsupervised RNN stack, however, got so compact that (previously infeasible) sequence
classification through additional BP-based SL became possible. Essentially the system used UL to
17
greatly reduce problem depth. Compare earlier BP-based fine-tuning of NNs initialized by rules of
propositional logic (Shavlik and Towell, 1989) (Sec. 5.6.1).
There is a way of compressing higher levels down into lower levels, thus fully or partially collapsing the RNN stack. The trick is to retrain a lower-level RNN to continually imitate (predict) the
hidden units of an already trained, slower, higher-level RNN (the “conscious” chunker), through additional predictive output neurons (Schmidhuber, 1992b). This helps the lower RNN (the automatizer)
to develop appropriate, rarely changing memories that may bridge very long time lags. Again, this
procedure can greatly reduce the required depth of the BP process.
The 1991 system was a working Deep Learner in the modern post-2000 sense, and also a first
Neural Hierarchical Temporal Memory (HTM). It is conceptually similar to earlier AE hierarchies
(1987, Sec. 5.7) and later Deep Belief Networks (2006, Sec. 5.15), but more general in the sense
that it uses sequence-processing RNNs instead of FNNs with unchanging inputs. More recently,
well-known entrepreneurs (Hawkins and George, 2006; Kurzweil, 2012) also got interested in HTMs;
compare also hierarchical HMMs (e.g., Fine et al., 1998), as well as later UL-based recurrent systems (Klapper-Rybicka et al., 2001; Steil, 2007; Klampfl and Maass, 2013; Young et al., 2014).
Clockwork RNNs (Koutnı́k et al., 2014) also consist of interacting RNN modules with different clock
rates, but do not use UL to set those rates. Stacks of RNNs were used in later work on SL with great
success, e.g., Sec. 5.13, 5.16, 5.17, 5.22.
5.11
1992: Max-Pooling (MP): Towards MPCNNs (Compare Sec. 5.16, 5.19)
The Neocognitron (Sec. 5.4) inspired the Cresceptron (Weng et al., 1992), which adapts its topology during training (Sec. 5.6.3); compare the incrementally growing and shrinking GMDH networks
(1965, Sec. 5.3).
Instead of using alternative local subsampling or WTA methods (e.g., Fukushima, 1980; Schmidhuber, 1989b; Maass, 2000; Fukushima, 2013a), the Cresceptron uses Max-Pooling (MP) layers. Here
a 2-dimensional layer or array of unit activations is partitioned into smaller rectangular arrays. Each
is replaced in a downsampling layer by the activation of its maximally active unit. A later, more complex version of the Cresceptron (Weng et al., 1997) also included “blurring” layers to improve object
location tolerance.
The neurophysiologically plausible topology of the feedforward HMAX model (Riesenhuber and
Poggio, 1999) is very similar to the one of the 1992 Cresceptron (and thus to the 1979 Neocognitron).
HMAX does not learn though. Its units have hand-crafted weights; biologically plausible learning
rules were later proposed for similar models (e.g., Serre et al., 2002; Teichmann et al., 2012).
When CNNs or convnets (Sec. 5.4, 5.8) are combined with MP, they become Cresceptron-like
or HMAX-like MPCNNs with alternating convolutional and max-pooling layers. Unlike Cresceptron
and HMAX, however, MPCNNs are trained by BP (Sec. 5.5, 5.16) (Ranzato et al., 2007). Advantages
of doing this were pointed out subsequently (Scherer et al., 2010). BP-trained MPCNNs have become
central to many modern, competition-winning, feedforward, visual Deep Learners (Sec. 5.17, 5.19–
5.23).
5.12
1994: Early Contest-Winning NNs
Back in the 1990s, certain NNs already won certain controlled pattern recognition contests with secret
test sets. Notably, an NN with internal delay lines won the Santa Fe time-series competition on chaotic
intensity pulsations of an NH3 laser (Wan, 1994; Weigend and Gershenfeld, 1993). No very deep
CAPs (Sec. 3) were needed though.
18
5.13
1995: Supervised Recurrent Very Deep Learner (LSTM RNN)
Supervised Long Short-Term Memory (LSTM) RNN (Hochreiter and Schmidhuber, 1997b; Gers et al.,
2000; Pérez-Ortiz et al., 2003) could eventually perform similar feats as the deep RNN hierarchy
of 1991 (Sec. 5.10), overcoming the Fundamental Deep Learning Problem (Sec. 5.9) without any
unsupervised pre-training. LSTM could also learn DL tasks without local sequence predictability
(and thus unlearnable by the partially unsupervised 1991 History Compressor, Sec. 5.10), dealing
with very deep problems (Sec. 3) (e.g., Gers et al., 2002).
The basic LSTM idea is very simple. Some of the units are called Constant Error Carousels
(CECs). Each CEC uses as an activation function f , the identity function, and has a connection to itself
with fixed weight of 1.0. Due to f ’s constant derivative of 1.0, errors backpropagated through a CEC
cannot vanish or explode (Sec. 5.9) but stay as they are (unless they “flow out” of the CEC to other,
typically adaptive parts of the NN). CECs are connected to several nonlinear adaptive units (some
with multiplicative activation functions) needed for learning nonlinear behavior. Weight changes of
these units often profit from error signals propagated far back in time through CECs. CECs are
the main reason why LSTM nets can learn to discover the importance of (and memorize) events that
happened thousands of discrete time steps ago, while previous RNNs already failed in case of minimal
time lags of 10 steps.
Many different LSTM variants and topologies are allowed. It is possible to evolve good problemspecific topologies (Bayer et al., 2009). Some LSTM variants also use modifiable self-connections of
CECs (Gers and Schmidhuber, 2001).
To a certain extent, LSTM is biologically plausible (O’Reilly, 2003). LSTM learned to solve
many previously unlearnable DL tasks involving: Recognition of the temporal order of widely separated events in noisy input streams; Robust storage of high-precision real numbers across extended
time intervals; Arithmetic operations on continuous input streams; Extraction of information conveyed by the temporal distance between events; Recognition of temporally extended patterns in noisy
input sequences (Hochreiter and Schmidhuber, 1997b; Gers et al., 2000); Stable generation of precisely timed rhythms, as well as smooth and non-smooth periodic trajectories (Gers and Schmidhuber,
2000). LSTM clearly outperformed previous RNNs on tasks that require learning the rules of regular languages describable by deterministic Finite State Automata (FSAs) (Watrous and Kuhn, 1992;
Casey, 1996; Siegelmann, 1992; Blair and Pollack, 1997; Kalinke and Lehmann, 1998; Zeng et al.,
1994; Manolios and Fanelli, 1994; Omlin and Giles, 1996; Vahed and Omlin, 2004), both in terms of
reliability and speed.
LSTM also worked on tasks involving context free languages (CFLs) that cannot be represented
by HMMs or similar FSAs discussed in the RNN literature (Sun et al., 1993b; Wiles and Elman, 1995;
Andrews et al., 1995; Steijvers and Grunwald, 1996; Tonkes and Wiles, 1997; Rodriguez et al., 1999;
Rodriguez and Wiles, 1998). CFL recognition (Lee, 1996) requires the functional equivalent of a runtime stack. Some previous RNNs failed to learn small CFL training sets (Rodriguez and Wiles, 1998).
Those that did not (Rodriguez et al., 1999; Bodén and Wiles, 2000) failed to extract the general rules,
and did not generalize well on substantially larger test sets. Similar for context-sensitive languages
(CSLs) (e.g., Chalup and Blair, 2003). LSTM generalized well though, requiring only the 30 shortest
exemplars (n ≤ 10) of the CSL an bn cn to correctly predict the possible continuations of sequence
prefixes for n up to 1000 and more. A combination of a decoupled extended Kalman filter (Kalman,
1960; Williams, 1992b; Puskorius and Feldkamp, 1994; Feldkamp et al., 1998; Haykin, 2001; Feldkamp et al., 2003) and an LSTM RNN (Pérez-Ortiz et al., 2003) learned to deal correctly with values
of n up to 10 million and more. That is, after training the network was able to read sequences of
30,000,000 symbols and more, one symbol at a time, and finally detect the subtle differences between legal strings such as a10,000,000 b10,000,000 c10,000,000 and very similar but illegal strings such
as a10,000,000 b9,999,999 c10,000,000 . Compare also more recent RNN algorithms able to deal with long
19
time lags (Schäfer et al., 2006; Martens and Sutskever, 2011; Zimmermann et al., 2012; Koutnı́k et al.,
2014).
Bi-directional RNNs (BRNNs) (Schuster and Paliwal, 1997; Schuster, 1999) are designed for input sequences whose starts and ends are known in advance, such as spoken sentences to be labeled by
their phonemes; compare (Fukada et al., 1999). To take both past and future context of each sequence
element into account, one RNN processes the sequence from start to end, the other backwards from
end to start. At each time step their combined outputs predict the corresponding label (if there is
any). BRNNs were successfully applied to secondary protein structure prediction (Baldi et al., 1999).
DAG-RNNs (Baldi and Pollastri, 2003; Wu and Baldi, 2008) generalize BRNNs to multiple dimensions. They learned to predict properties of small organic molecules (Lusci et al., 2013) as well as
protein contact maps (Tegge et al., 2009), also in conjunction with a growing deep FNN (Di Lena
et al., 2012) (Sec. 5.21). BRNNs and DAG-RNNs unfold their full potential when combined with the
LSTM concept (Graves and Schmidhuber, 2005, 2009; Graves et al., 2009).
Particularly successful in recent competitions are stacks (Sec. 5.10) of LSTM RNNs (Fernandez et al., 2007; Graves and Schmidhuber, 2009) trained by Connectionist Temporal Classification (CTC) (Graves et al., 2006), a gradient-based method for finding RNN weights that maximize the probability of teacher-given label sequences, given (typically much longer and more highdimensional) streams of real-valued input vectors. CTC-LSTM performs simultaneous segmentation
(alignment) and recognition (Sec. 5.22).
In the early 2000s, speech recognition was dominated by HMMs combined with FNNs (e.g.,
Bourlard and Morgan, 1994). Nevertheless, when trained from scratch on utterances from the TIDIGITS speech database, in 2003 LSTM already obtained results comparable to those of HMM-based
systems (Graves et al., 2003; Beringer et al., 2005; Graves et al., 2006). In 2007, LSTM outperformed
HMMs in keyword spotting tasks (Fernández et al., 2007); compare recent improvements (Indermuhle
et al., 2011; Wöllmer et al., 2013). By 2013, LSTM also achieved best known results on the famous
TIMIT phoneme recognition benchmark (Graves et al., 2013) (Sec. 5.22). Recently, LSTM RNN /
HMM hybrids obtained best known performance on medium-vocabulary (Geiger et al., 2014) and
large-vocabulary speech recognition (Sak et al., 2014a).
LSTM is also applicable to robot localization (Förster et al., 2007), robot control (Mayer et al.,
2008), online driver distraction detection (Wöllmer et al., 2011), and many other tasks. For example,
it helped to improve the state of the art in diverse applications such as protein analysis (Hochreiter
and Obermayer, 2005), handwriting recognition (Graves et al., 2008, 2009; Graves and Schmidhuber,
2009; Bluche et al., 2014), voice activity detection (Eyben et al., 2013), optical character recognition (Breuel et al., 2013), language identification (Gonzalez-Dominguez et al., 2014), prosody contour
prediction (Fernandez et al., 2014), audio onset detection (Marchi et al., 2014), text-to-speech synthesis (Fan et al., 2014), social signal classification (Brueckner and Schulter, 2014), machine translation (Sutskever et al., 2014), and others.
RNNs can also be used for metalearning (Schmidhuber, 1987; Schaul and Schmidhuber, 2010;
Prokhorov et al., 2002), because they can in principle learn to run their own weight change algorithm (Schmidhuber, 1993a). A successful metalearner (Hochreiter et al., 2001b) used an LSTM
RNN to quickly learn a learning algorithm for quadratic functions (compare Sec. 6.8).
Recently, LSTM RNNs won several international pattern recognition competitions and set numerous benchmark records on large and complex data sets, e.g., Sec. 5.17, 5.21, 5.22. Gradientbased LSTM is no panacea though—other methods sometimes outperformed it at least on certain
tasks (Jaeger, 2004; Schmidhuber et al., 2007; Martens and Sutskever, 2011; Pascanu et al., 2013b;
Koutnı́k et al., 2014); compare Sec. 5.20.
20
5.14
2003: More Contest-Winning/Record-Setting NNs; Successful Deep NNs
In the decade around 2000, many practical and commercial pattern recognition applications were
dominated by non-neural machine learning methods such as Support Vector Machines (SVMs) (Vapnik, 1995; Schölkopf et al., 1998). Nevertheless, at least in certain domains, NNs outperformed other
techniques.
A Bayes NN (Neal, 2006) based on an ensemble (Breiman, 1996; Schapire, 1990; Wolpert, 1992;
Hashem and Schmeiser, 1992; Ueda, 2000; Dietterich, 2000a) of NNs won the NIPS 2003 Feature
Selection Challenge with secret test set (Neal and Zhang, 2006). The NN was not very deep though—
it had two hidden layers and thus rather shallow CAPs (Sec. 3) of depth 3.
Important for many present competition-winning pattern recognisers (Sec. 5.19, 5.21, 5.22) were
developments in the CNN department. A BP-trained (LeCun et al., 1989) CNN (Sec. 5.4, Sec. 5.8) set
a new MNIST record of 0.4% (Simard et al., 2003), using training pattern deformations (Baird, 1990)
but no unsupervised pre-training (Sec. 5.7, 5.10, 5.15). A standard BP net achieved 0.7% (Simard
et al., 2003). Again, the corresponding CAP depth was low. Compare further improvements in
Sec. 5.16, 5.18, 5.19.
Good image interpretation results (Behnke, 2003b) were achieved with rather deep NNs trained
by the BP variant R-prop (Riedmiller and Braun, 1993) (Sec. 5.6.2); here feedback through recurrent
connections helped to improve image interpretation. FNNs with CAP depth up to 6 were used to
successfully classify high-dimensional data (Vieira and Barradas, 2003).
Deep LSTM RNNs started to obtain certain first speech recognition results comparable to those
of HMM-based systems (Graves et al., 2003); compare Sec. 5.13, 5.16, 5.21, 5.22.
5.15
2006/7: UL For Deep Belief Networks / AE Stacks Fine-Tuned by BP
While learning networks with numerous non-linear layers date back at least to 1965 (Sec. 5.3), and explicit DL research results have been published at least since 1991 (Sec. 5.9, 5.10), the expression Deep
Learning was actually coined around 2006, when unsupervised pre-training of deep FNNs helped to
accelerate subsequent SL through BP (Hinton and Salakhutdinov, 2006; Hinton et al., 2006). Compare
earlier terminology on loading deep networks (Sı́ma, 1994; Windisch, 2005) and learning deep memories (Gomez and Schmidhuber, 2005). Compare also BP-based (Sec. 5.5) fine-tuning (Sec. 5.6.1) of
(not so deep) FNNs pre-trained by competitive UL (Maclin and Shavlik, 1995).
The Deep Belief Network (DBN) is a stack of Restricted Boltzmann Machines (RBMs) (Smolensky, 1986), which in turn are Boltzmann Machines (BMs) (Hinton and Sejnowski, 1986) with a single
layer of feature-detecting units; compare also Higher-Order BMs (Memisevic and Hinton, 2010).
Each RBM perceives pattern representations from the level below and learns to encode them in unsupervised fashion. At least in theory under certain assumptions, adding more layers improves a
bound on the data’s negative log probability (Hinton et al., 2006) (equivalent to the data’s description
length—compare the corresponding observation for RNN stacks, Sec. 5.10). There are extensions for
Temporal RBMs (Sutskever et al., 2008).
Without any training pattern deformations (Sec. 5.14), a DBN fine-tuned by BP achieved 1.2%
error rate (Hinton and Salakhutdinov, 2006) on the MNIST handwritten digits (Sec. 5.8, 5.14). This
result helped to arouse interest in DBNs. DBNs also achieved good results on phoneme recognition,
with an error rate of 26.7% on the TIMIT core test set (Mohamed and Hinton, 2010); compare further
improvements through FNNs (Hinton et al., 2012a; Deng and Yu, 2014) and LSTM RNNs (Sec. 5.22).
A DBN-based technique called Semantic Hashing (Salakhutdinov and Hinton, 2009) maps semantically similar documents (of variable size) to nearby addresses in a space of document representations. It outperformed previous searchers for similar documents, such as Locality Sensitive
Hashing (Buhler, 2001; Datar et al., 2004). See the RBM/DBN tutorial (Fischer and Igel, 2014).
21
Autoencoder (AE) stacks (Ballard, 1987) (Sec. 5.7) became a popular alternative way of pretraining deep FNNs in unsupervised fashion, before fine-tuning (Sec. 5.6.1) them through BP
(Sec. 5.5) (Bengio et al., 2007; Vincent et al., 2008; Erhan et al., 2010). Sparse coding (Sec. 5.6.4)
was formulated as a combination of convex optimization problems (Lee et al., 2007a). Recent surveys
of stacked RBM and AE methods focus on post-2006 developments (Bengio, 2009; Arel et al., 2010).
Unsupervised DBNs and AE stacks are conceptually similar to, but in a certain sense less general
than, the unsupervised RNN stack-based History Compressor of 1991 (Sec. 5.10), which can process
and re-encode not only stationary input patterns, but entire pattern sequences.
5.16
2006/7: Improved CNNs / GPU-CNNs / BP for MPCNNs / LSTM Stacks
Also in 2006, a BP-trained (LeCun et al., 1989) CNN (Sec. 5.4, Sec. 5.8) set a new MNIST record
of 0.39% (Ranzato et al., 2006), using training pattern deformations (Sec. 5.14) but no unsupervised
pre-training. Compare further improvements in Sec. 5.18, 5.19. Similar CNNs were used for offroad obstacle avoidance (LeCun et al., 2006). A combination of CNNs and TDNNs later learned to
map fixed-size representations of variable-size sentences to features relevant for language processing,
using a combination of SL and UL (Collobert and Weston, 2008).
2006 also saw an early GPU-based CNN implementation (Chellapilla et al., 2006) up to 4 times
faster than CPU-CNNs; compare also earlier GPU implementations of standard FNNs with a reported
speed-up factor of 20 (Oh and Jung, 2004). GPUs or graphics cards have become more and more
important for DL in subsequent years (Sec. 5.18–5.22).
In 2007, BP (Sec. 5.5) was applied for the first time (Ranzato et al., 2007) to Neocognitroninspired (Sec. 5.4), Cresceptron-like (or HMAX-like) MPCNNs (Sec. 5.11) with alternating convolutional and max-pooling layers. BP-trained MPCNNs have become an essential ingredient of many
modern, competition-winning, feedforward, visual Deep Learners (Sec. 5.17, 5.19–5.23).
Also in 2007, hierarchical stacks of LSTM RNNs were introduced (Fernandez et al., 2007). They
can be trained by hierarchical Connectionist Temporal Classification (CTC) (Graves et al., 2006). For
tasks of sequence labelling, every LSTM RNN level (Sec. 5.13) predicts a sequence of labels fed to
the next level. Error signals at every level are back-propagated through all the lower levels. On spoken
digit recognition, LSTM stacks outperformed HMMs, despite making fewer assumptions about the
domain. LSTM stacks do not necessarily require unsupervised pre-training like the earlier UL-based
RNN stacks (Schmidhuber, 1992b) of Sec. 5.10.
5.17
2009: First Official Competitions Won by RNNs, and with MPCNNs
Stacks of LSTM RNNs trained by CTC (Sec. 5.13, 5.16) became the first RNNs to win official international pattern recognition contests (with secret test sets known only to the organisers). More precisely,
three connected handwriting competitions at ICDAR 2009 in three different languages (French, Arab,
Farsi) were won by deep LSTM RNNs without any a priori linguistic knowledge, performing simultaneous segmentation and recognition. Compare (Graves and Schmidhuber, 2005; Graves et al., 2009;
Schmidhuber et al., 2011; Graves et al., 2013; Graves and Jaitly, 2014) (Sec. 5.22).
To detect human actions in surveillance videos, a 3-dimensional CNN (e.g., Jain and Seung, 2009;
Prokhorov, 2010), combined with SVMs, was part of a larger system (Yang et al., 2009) using a bag
of features approach (Nowak et al., 2006) to extract regions of interest. The system won three 2009
TRECVID competitions. These were possibly the first official international contests won with the
help of (MP)CNNs (Sec. 5.16). An improved version of the method was published later (Ji et al.,
2013).
2009 also saw a GPU-DBN implementation (Raina et al., 2009) orders of magnitudes faster than
previous CPU-DBNs (see Sec. 5.15); see also (Coates et al., 2013). The Convolutional DBN (Lee
22
et al., 2009a) (with a probabilistic variant of MP, Sec. 5.11) combines ideas from CNNs and DBNs,
and was successfully applied to audio classification (Lee et al., 2009b).
5.18
2010: Plain Backprop (+ Distortions) on GPU Breaks MNIST Record
In 2010, a new MNIST (Sec. 5.8) record of 0.35% error rate was set by good old BP (Sec. 5.5)
in deep but otherwise standard NNs (Ciresan et al., 2010), using neither unsupervised pre-training
(e.g., Sec. 5.7, 5.10, 5.15) nor convolution (e.g., Sec. 5.4, 5.8, 5.14, 5.16). However, training pattern
deformations (e.g., Sec. 5.14) were important to generate a big training set and avoid overfitting. This
success was made possible mainly through a GPU implementation of BP that was up to 50 times
faster than standard CPU versions. A good value of 0.95% was obtained without distortions except
for small saccadic eye movement-like translations—compare Sec. 5.15.
Since BP was 3-5 decades old by then (Sec. 5.5), and pattern deformations 2 decades (Baird, 1990)
(Sec. 5.14), these results seemed to suggest that advances in exploiting modern computing hardware
were more important than advances in algorithms.
5.19
2011: MPCNNs on GPU Achieve Superhuman Vision Performance
In 2011, a flexible GPU-implementation (Ciresan et al., 2011a) of Max-Pooling (MP) CNNs or Convnets was described (a GPU-MPCNN), building on earlier MP work (Weng et al., 1992) (Sec. 5.11)
CNNs (Fukushima, 1979; LeCun et al., 1989) (Sec. 5.4, 5.8, 5.16), and on early GPU-based CNNs
without MP (Chellapilla et al., 2006) (Sec. 5.16); compare early GPU-NNs (Oh and Jung, 2004) and
GPU-DBNs (Raina et al., 2009) (Sec. 5.17). MPCNNs have alternating convolutional layers (Sec. 5.4)
and max-pooling layers (MP, Sec. 5.11) topped by standard fully connected layers. All weights are
trained by BP (Sec. 5.5, 5.8, 5.16) (Ranzato et al., 2007; Scherer et al., 2010). GPU-MPCNNs have
become essential for many contest-winning FNNs (Sec. 5.21, Sec. 5.22).
Multi-Column GPU-MPCNNs (Ciresan et al., 2011b) are committees (Breiman, 1996; Schapire,
1990; Wolpert, 1992; Hashem and Schmeiser, 1992; Ueda, 2000; Dietterich, 2000a) of GPUMPCNNs with simple democratic output averaging. Several MPCNNs see the same input; their output
vectors are used to assign probabilities to the various possible classes. The class with the on average
highest probability is chosen as the system’s classification of the present input. Compare earlier, more
sophisticated ensemble methods (Schapire, 1990), the contest-winning ensemble Bayes-NN (Neal,
2006) of Sec. 5.14, and recent related work (Shao et al., 2014).
An ensemble of GPU-MPCNNs was the first system to achieve superhuman visual pattern recognition (Ciresan et al., 2011b, 2012b) in a controlled competition, namely, the IJCNN 2011 traffic
sign recognition contest in San Jose (CA) (Stallkamp et al., 2011, 2012). This is of interest for fully
autonomous, self-driving cars in traffic (e.g., Dickmanns et al., 1994). The GPU-MPCNN ensemble obtained 0.56% error rate and was twice better than human test subjects, three times better than
the closest artificial NN competitor (Sermanet and LeCun, 2011), and six times better than the best
non-neural method.
A few months earlier, the qualifying round was won in a 1st stage online competition, albeit by
a much smaller margin: 1.02% (Ciresan et al., 2011b) vs 1.03% for second place (Sermanet and
LeCun, 2011). After the deadline, the organisers revealed that human performance on the test set
was 1.19%. That is, the best methods already seemed human-competitive. However, during the
qualifying it was possible to incrementally gain information about the test set by probing it through
repeated submissions. This is illustrated by better and better results obtained by various teams over
time (Stallkamp et al., 2012) (the organisers eventually imposed a limit of 10 resubmissions). In the
final competition this was not possible.
23
This illustrates a general problem with benchmarks whose test sets are public, or at least can be
probed to some extent: competing teams tend to overfit on the test set even when it cannot be directly
used for training, only for evaluation.
In 1997 many thought it a big deal that human chess world champion Kasparov was beaten by
an IBM computer. But back then computers could not at all compete with little kids in visual pattern recognition, which seems much harder than chess from a computational perspective. Of course,
the traffic sign domain is highly restricted, and kids are still much better general pattern recognisers. Nevertheless, by 2011, deep NNs could already learn to rival them in important limited visual
domains.
An ensemble of GPU-MPCNNs was also the first method to achieve human-competitive performance (around 0.2%) on MNIST (Ciresan et al., 2012c). This represented a dramatic improvement,
since by then the MNIST record had hovered around 0.4% for almost a decade (Sec. 5.14, 5.16, 5.18).
Given all the prior work on (MP)CNNs (Sec. 5.4, 5.8, 5.11, 5.16) and GPU-CNNs (Sec. 5.16),
GPU-MPCNNs are not a breakthrough in the scientific sense. But they are a commercially relevant
breakthrough in efficient coding that has made a difference in several contests since 2011. Today, most
feedforward competition-winning deep NNs are (ensembles of) GPU-MPCNNs (Sec. 5.21–5.23).
5.20
2011: Hessian-Free Optimization for RNNs
Also in 2011 it was shown (Martens and Sutskever, 2011) that Hessian-free optimization (e.g., Møller,
1993; Pearlmutter, 1994; Schraudolph, 2002) (Sec. 5.6.2) can alleviate the Fundamental Deep Learning Problem (Sec. 5.9) in RNNs, outperforming standard gradient-based LSTM RNNs (Sec. 5.13) on
several tasks. Compare other RNN algorithms (Jaeger, 2004; Schmidhuber et al., 2007; Pascanu et al.,
2013b; Koutnı́k et al., 2014) that also at least sometimes yield better results than steepest descent for
LSTM RNNs.
5.21
2012: First Contests Won on ImageNet, Object Detection, Segmentation
In 2012, an ensemble of GPU-MPCNNs (Sec. 5.19) achieved best results on the ImageNet classification benchmark (Krizhevsky et al., 2012), which is popular in the computer vision community. Here
relatively large image sizes of 256x256 pixels were necessary, as opposed to only 48x48 pixels for
the 2011 traffic sign competition (Sec. 5.19). See further improvements in Sec. 5.22.
Also in 2012, the biggest NN so far (109 free parameters) was trained in unsupervised mode
(Sec. 5.7, 5.15) on unlabeled data (Le et al., 2012), then applied to ImageNet. The codes across its top
layer were used to train a simple supervised classifier, which achieved best results so far on 20,000
classes. Instead of relying on efficient GPU programming, this was done by brute force on 1,000
standard machines with 16,000 cores.
So by 2011/2012, excellent results had been achieved by Deep Learners in image recognition and
classification (Sec. 5.19, 5.21). The computer vision community, however, is especially interested in
object detection in large images, for applications such as image-based search engines, or for biomedical diagnosis where the goal may be to automatically detect tumors etc in images of human tissue.
Object detection presents additional challenges. One natural approach is to train a deep NN classifier
on patches of big images, then use it as a feature detector to be shifted across unknown visual scenes,
using various rotations and zoom factors. Image parts that yield highly active output units are likely
to contain objects similar to those the NN was trained on.
2012 finally saw the first DL system (an ensemble of GPU-MPCNNs, Sec. 5.19) to win a contest
on visual object detection (Ciresan et al., 2013) in large images of several million pixels (ICPR 2012
Contest on Mitosis Detection in Breast Cancer Histological Images, 2012; Roux et al., 2013). Such
biomedical applications may turn out to be among the most important applications of DL. The world
24
spends over 10% of GDP on healthcare (> 6 trillion USD per year), much of it on medical diagnosis
through expensive experts. Partial automation of this could not only save lots of money, but also make
expert diagnostics accessible to many who currently cannot afford it. It is gratifying to observe that
today deep NNs may actually help to improve healthcare and perhaps save human lives.
2012 also saw the first pure image segmentation contest won by DL (Ciresan et al., 2012a), again
through an GPU-MPCNN ensemble (Segmentation of Neuronal Structures in EM Stacks Challenge,
2012).2 EM stacks are relevant for the recently approved huge brain projects in Europe and the
US (e.g., Markram, 2012). Given electron microscopy images of stacks of thin slices of animal
brains, the goal is to build a detailed 3D model of the brain’s neurons and dendrites. But human
experts need many hours and days and weeks to annotate the images: Which parts depict neuronal
membranes? Which parts are irrelevant background? This needs to be automated (e.g., Turaga et al.,
2010). Deep Multi-Column GPU-MPCNNs learned to solve this task through experience with many
training images, and won the contest on all three evaluation metrics by a large margin, with superhuman performance in terms of pixel error.
Both object detection (Ciresan et al., 2013) and image segmentation (Ciresan et al., 2012a) profit
from fast MPCNN-based image scans that avoid redundant computations. Recent MPCNN scanners
speed up naive implementations by up to three orders of magnitude (Masci et al., 2013; Giusti et al.,
2013); compare earlier efficient methods for CNNs without MP (Vaillant et al., 1994).
Also in 2012, a system consisting of growing deep FNNs and 2D-BRNNs (Di Lena et al., 2012)
won the CASP 2012 contest on protein contact map prediction. On the IAM-OnDoDB benchmark,
LSTM RNNs (Sec. 5.13) outperformed all other methods (HMMs, SVMs) on online mode detection (Otte et al., 2012; Indermuhle et al., 2012) and keyword spotting (Indermuhle et al., 2011). On the
long time lag problem of language modelling, LSTM RNNs outperformed all statistical approaches
on the IAM-DB benchmark (Frinken et al., 2012); improved results were later obtained through a
combination of NNs and HMMs (Zamora-Martnez et al., 2014). Compare earlier RNNs for object
recognition through iterative image interpretation (Behnke and Rojas, 1998; Behnke, 2002, 2003b);
see also more recent publications (Wyatte et al., 2012; OReilly et al., 2013) extending work on biologically plausible learning rules for RNNs (O’Reilly, 1996).
5.22
2013-: More Contests and Benchmark Records
A stack (Fernandez et al., 2007; Graves and Schmidhuber, 2009) (Sec. 5.10) of bi-directional LSTM
RNNs (Graves and Schmidhuber, 2005) trained by CTC (Sec. 5.13, 5.17) broke a famous TIMIT
speech (phoneme) recognition record, achieving 17.7% test set error rate (Graves et al., 2013), despite
thousands of man years previously spent on Hidden Markov Model (HMMs)-based speech recognition
research. Compare earlier DBN results (Sec. 5.15).
CTC-LSTM also helped to score first at NIST’s OpenHaRT2013 evaluation (Bluche et al., 2014).
For optical character recognition (OCR), LSTM RNNs outperformed commercial recognizers of historical data (Breuel et al., 2013). LSTM-based systems also set benchmark records in language identification (Gonzalez-Dominguez et al., 2014), medium-vocabulary speech recognition (Geiger et al.,
2014), prosody contour prediction (Fernandez et al., 2014), audio onset detection (Marchi et al.,
2014), text-to-speech synthesis (Fan et al., 2014), and social signal classification (Brueckner and
Schulter, 2014).
An LSTM RNN was used to estimate the state posteriors of an HMM; this system beat the previous
state of the art in large vocabulary speech recognition (Sak et al., 2014b,a). Another LSTM RNN with
hundreds of millions of connections was used to rerank hypotheses of a statistical machine translation
2 It should be mentioned, however, that LSTM RNNs already performed simultaneous segmentation and recognition when
they became the first recurrent Deep Learners to win official international pattern recognition contests—see Sec. 5.17.
25
system; this system beat the previous state of the art in English to French translation (Sutskever et al.,
2014).
A new record on the ICDAR Chinese handwriting recognition benchmark (over 3700 classes)
was set on a desktop machine by an ensemble of GPU-MPCNNs (Sec. 5.19) with almost human
performance (Ciresan and Schmidhuber, 2013); compare (Yin et al., 2013).
The MICCAI 2013 Grand Challenge on Mitosis Detection (Veta et al., 2013) also was won by an
object-detecting GPU-MPCNN ensemble (Ciresan et al., 2013). Its data set was even larger and more
challenging than the one of ICPR 2012 (Sec. 5.21): a real-world dataset including many ambiguous
cases and frequently encountered problems such as imperfect slide staining.
Three 2D-CNNs (with mean-pooling instead of MP, Sec. 5.11) observing three orthogonal projections of 3D images outperformed traditional full 3D methods on the task of segmenting tibial cartilage
in low field knee MRI scans (Prasoon et al., 2013).
Deep GPU-MPCNNs (Sec. 5.19) also helped to achieve new best results on important benchmarks of the computer vision community: ImageNet classification (Zeiler and Fergus, 2013; Szegedy
et al., 2014) and—in conjunction with traditional approaches—PASCAL object detection (Girshick
et al., 2013). They also learned to predict bounding box coordinates of objects in the Imagenet
2013 database, and obtained state-of-the-art results on tasks of localization and detection (Sermanet
et al., 2013). GPU-MPCNNs also helped to recognise multi-digit numbers in Google Street View
images (Goodfellow et al., 2014b), where part of the NN was trained to count visible digits; compare
earlier work on detecting “numerosity” through DBNs (Stoianov and Zorzi, 2012). This system also
excelled at recognising distorted synthetic text in reCAPTCHA puzzles. Other successful CNN applications include scene parsing (Farabet et al., 2013), object detection (Szegedy et al., 2013), shadow
detection (Khan et al., 2014), video classification (Karpathy et al., 2014), and Alzheimers disease
neuroimaging (Li et al., 2014).
Additional contests are mentioned in the web pages of the Swiss AI Lab IDSIA, the University of
Toronto, NY University, and the University of Montreal.
5.23
Currently Successful Techniques: LSTM RNNs and GPU-MPCNNs
Most competition-winning or benchmark record-setting Deep Learners actually use one of two supervised techniques: (a) recurrent LSTM (1997) trained by CTC (2006) (Sec. 5.13, 5.17, 5.21, 5.22), or
(b) feedforward GPU-MPCNNs (2011, Sec. 5.19, 5.21, 5.22) based on CNNs (1979, Sec. 5.4) with
MP (1992, Sec. 5.11) trained through BP (1989–2007, Sec. 5.8, 5.16).
Exceptions include two 2011 contests (Goodfellow et al., 2011; Mesnil et al., 2011; Goodfellow et al., 2012) specialised on Transfer Learning from one dataset to another (e.g., Caruana, 1997;
Schmidhuber, 2004; Pan and Yang, 2010). However, deep GPU-MPCNNs do allow for pure SL-based
transfer (Ciresan et al., 2012d), where pre-training on one training set greatly improves performance
on quite different sets, also in more recent studies (Oquab et al., 2013; Donahue et al., 2013). In
fact, deep MPCNNs pre-trained by SL can extract useful features from quite diverse off-training-set
images, yielding better results than traditional, widely used features such as SIFT (Lowe, 1999, 2004)
on many vision tasks (Razavian et al., 2014). To deal with changing datasets, slowly learning deep
NNs were also combined with rapidly adapting “surface” NNs (Kak et al., 2010).
Remarkably, in the 1990s a trend went from partially unsupervised RNN stacks (Sec. 5.10) to
purely supervised LSTM RNNs (Sec. 5.13), just like in the 2000s a trend went from partially unsupervised FNN stacks (Sec. 5.15) to purely supervised MPCNNs (Sec. 5.16–5.22). Nevertheless, in many
applications it can still be advantageous to combine the best of both worlds—supervised learning and
unsupervised pre-training (Sec. 5.10, 5.15).
26
5.24
Recent Tricks for Improving SL Deep NNs (Compare Sec. 5.6.2, 5.6.3)
DBN training (Sec. 5.15) can be improved through gradient enhancements and automatic learning rate
adjustments during stochastic gradient descent (Cho et al., 2013; Cho, 2014), and through Tikhonovtype (Tikhonov et al., 1977) regularization of RBMs (Cho et al., 2012). Contractive AEs (Rifai et al.,
2011) discourage hidden unit perturbations in response to input perturbations, similar to how FMS
(Sec. 5.6.3) for L OCOCODE AEs (Sec. 5.6.4) discourages output perturbations in response to weight
perturbations.
Hierarchical CNNs in a Neural Abstraction Pyramid (e.g., Behnke, 2003b, 2005) were trained to
reconstruct images corrupted by structured noise (Behnke, 2001), thus enforcing increasingly abstract
image representations in deeper and deeper layers. Denoising AEs later used a similar procedure (Vincent et al., 2008).
Dropout (Hinton et al., 2012b; Ba and Frey, 2013) removes units from NNs during training to
improve generalisation. Some view it as an ensemble method that trains multiple data models simultaneously (Baldi and Sadowski, 2014). Under certain circumstances, it could also be viewed as a form
of training set augmentation: effectively, more and more informative complex features are removed
from the training data. Compare dropout for RNNs (Pham et al., 2013; Pachitariu and Sahani, 2013;
Pascanu et al., 2013a). A deterministic approximation coined fast dropout (Wang and Manning, 2013)
can lead to faster learning and evaluation and was adapted for RNNs (Bayer et al., 2013). Dropout is
closely related to older, biologically plausible techniques for adding noise to neurons or synapses during training (e.g., Hanson, 1990; Murray and Edwards, 1993; Schuster, 1992; Nadal and Parga, 1994;
Jim et al., 1995; An, 1996), which in turn are closely related to finding perturbation-resistant lowcomplexity NNs, e.g., through FMS (Sec. 5.6.3). MDL-based stochastic variational methods (Graves,
2011) are also related to FMS. They are useful for RNNs, where classic regularizers such as weight
decay (Sec. 5.6.3) represent a bias towards limited memory capacity (e.g., Pascanu et al., 2013b).
Compare recent work on variational recurrent AEs (Bayer and Osendorfer, 2014).
The activation function f of Rectified Linear Units (ReLUs) is f (x) = x for x > 0, f (x) = 0
otherwise—compare the old concept of half-wave rectified units (Malik and Perona, 1990). ReLU
NNs are useful for RBMs (Nair and Hinton, 2010; Maas et al., 2013), outperformed sigmoidal activation functions in deep NNs (Glorot et al., 2011), and helped to obtain best results on several
benchmark problems across multiple domains (e.g., Krizhevsky et al., 2012; Dahl et al., 2013).
NNs with competing linear units tend to outperform those with non-competing nonlinear units,
and avoid catastrophic forgetting through BP when training sets change over time (Srivastava et al.,
2013). In this context, choosing a learning algorithm may be more important than choosing activation
functions (Goodfellow et al., 2014a). Maxout NNs (Goodfellow et al., 2013) combine competitive
interactions and dropout (see above) to achieve excellent results on certain benchmarks. Compare
early RNNs with competing units for SL and RL (Schmidhuber, 1989b). To address overfitting,
instead of depending on pre-wired regularizers and hyper-parameters (Hertz et al., 1991; Bishop,
2006), self-delimiting RNNs (SLIM NNs) with competing units (Schmidhuber, 2012) can in principle
learn to select their own runtime and their own numbers of effective free parameters, thus learning
their own computable regularisers (Sec. 4.4, 5.6.3), becoming fast and slim when necessary. One may
penalize the task-specific total length of connections (e.g., Legenstein and Maass, 2002; Schmidhuber,
2012, 2013b; Clune et al., 2013) and communication costs of SLIM NNs implemented on the 3dimensional brain-like multi-processor hardware to be expected in the future.
RmsProp (Tieleman and Hinton, 2012; Schaul et al., 2013) can speed up first order gradient descent methods (Sec. 5.5, 5.6.2); compare vario-η (Neuneier and Zimmermann, 1996), Adagrad (Duchi
et al., 2011) and Adadelta (Zeiler, 2012). DL in NNs can also be improved by transforming hidden
unit activations such that they have zero output and slope on average (Raiko et al., 2012). Many additional, older tricks (Sec. 5.6.2, 5.6.3) should also be applicable to today’s deep NNs; compare (Orr
27
and Müller, 1998; Montavon et al., 2012).
5.25
Consequences for Neuroscience
It is ironic that artificial NNs (ANNs) can help to better understand biological NNs (BNNs)—see
the ISBI 2012 results mentioned in Sec. 5.21 (Segmentation of Neuronal Structures in EM Stacks
Challenge, 2012; Ciresan et al., 2012a).
The feature detectors learned by single-layer visual ANNs are similar to those found in early
visual processing stages of BNNs (e.g., Sec. 5.6.4). Likewise, the feature detectors learned in deep
layers of visual ANNs should be highly predictive of what neuroscientists will find in deep layers
of BNNs. While the visual cortex of BNNs may use quite different learning algorithms, its objective
function to be minimised may be quite similar to the one of visual ANNs. In fact, results obtained with
relatively deep artificial DBNs (Lee et al., 2007b) and CNNs (Yamins et al., 2013) seem compatible
with insights about the visual pathway in the primate cerebral cortex, which has been studied for
many decades (e.g., Hubel and Wiesel, 1968; Perrett et al., 1982; Desimone et al., 1984; Felleman
and Van Essen, 1991; Perrett et al., 1992; Kobatake and Tanaka, 1994; Logothetis et al., 1995; Bichot
et al., 2005; Hung et al., 2005; Lennie and Movshon, 2005; Connor et al., 2007; Kriegeskorte et al.,
2008; DiCarlo et al., 2012); compare a computer vision-oriented survey (Kruger et al., 2013).
5.26
DL with Spiking Neurons?
Many recent DL results profit from GPU-based traditional deep NNs, e.g., Sec. 5.16–5.19. Current
GPUs, however, are little ovens, much hungrier for energy than biological brains, whose neurons efficiently communicate by brief spikes (Hodgkin and Huxley, 1952; FitzHugh, 1961; Nagumo et al.,
1962), and often remain quiet. Many computational models of such spiking neurons have been proposed and analyzed (e.g., Gerstner and van Hemmen, 1992; Zipser et al., 1993; Stemmler, 1996;
Tsodyks et al., 1996; Maex and Orban, 1996; Maass, 1996, 1997; Kistler et al., 1997; Amit and
Brunel, 1997; Tsodyks et al., 1998; Kempter et al., 1999; Song et al., 2000; Stoop et al., 2000; Brunel,
2000; Bohte et al., 2002; Gerstner and Kistler, 2002; Izhikevich et al., 2003; Seung, 2003; Deco and
Rolls, 2005; Brette et al., 2007; Brea et al., 2013; Nessler et al., 2013; Kasabov, 2014; Hoerzer et al.,
2014; Rezende and Gerstner, 2014).
Future energy-efficient hardware for DL in NNs may implement aspects of such models (e.g.,
Liu et al., 2001; Roggen et al., 2003; Glackin et al., 2005; Schemmel et al., 2006; Fieres et al., 2008;
Khan et al., 2008; Serrano-Gotarredona et al., 2009; Jin et al., 2010; Indiveri et al., 2011; Neil and Liu,
2014; Merolla et al., 2014). A simulated, event-driven, spiking variant (Neftci et al., 2014) of an RBM
(Sec. 5.15) was trained by a variant of the Contrastive Divergence algorithm (Hinton, 2002). Spiking
nets were evolved to achieve reasonable performance on small face recognition data sets (Wysoski
et al., 2010) and to control simple robots (Floreano and Mattiussi, 2001; Hagras et al., 2004). A
spiking DBN with about 250,000 neurons (as part of a larger NN; Eliasmith et al., 2012; Eliasmith,
2013) achieved 6% error rate on MNIST; compare similar results with a spiking DBN variant of
depth 3 using a neuromorphic event-based sensor (O’Connor et al., 2013). In practical applications,
however, current artificial networks of spiking neurons cannot yet compete with the best traditional
deep NNs (e.g., compare MNIST results of Sec. 5.19).
28
6
DL in FNNs and RNNs for Reinforcement Learning (RL)
So far we have focused on Deep Learning (DL) in supervised or unsupervised NNs. Such NNs learn
to perceive / encode / predict / classify patterns or pattern sequences, but they do not learn to act
in the more general sense of Reinforcement Learning (RL) in unknown environments (see surveys,
e.g., Kaelbling et al., 1996; Sutton and Barto, 1998; Wiering and van Otterlo, 2012). Here we add a
discussion of DL FNNs and RNNs for RL. It will be shorter than the discussion of FNNs and RNNs
for SL and UL (Sec. 5), reflecting the current size of the various fields.
Without a teacher, solely from occasional real-valued pain and pleasure signals, RL agents must
discover how to interact with a dynamic, initially unknown environment to maximize their expected
cumulative reward signals (Sec. 2). There may be arbitrary, a priori unknown delays between actions
and perceivable consequences. The problem is as hard as any problem of computer science, since any
task with a computable description can be formulated in the RL framework (e.g., Hutter, 2005). For
example, an answer to the famous question of whether P = N P (Levin, 1973b; Cook, 1971) would
also set limits for what is achievable by general RL. Compare more specific limitations, e.g., (Blondel
and Tsitsiklis, 2000; Madani et al., 2003; Vlassis et al., 2012). The following subsections mostly focus
on certain obvious intersections between DL and RL—they cannot serve as a general RL survey.
6.1
RL Through NN World Models Yields RNNs With Deep CAPs
In the special case of an RL FNN controller C interacting with a deterministic, predictable environment, a separate FNN called M can learn to become C’s world model through system identification,
predicting C’s inputs from previous actions and inputs (e.g., Werbos, 1981, 1987; Munro, 1987; Jordan, 1988; Werbos, 1989b,a; Robinson and Fallside, 1989; Jordan and Rumelhart, 1990; Schmidhuber, 1990d; Narendra and Parthasarathy, 1990; Werbos, 1992; Gomi and Kawato, 1993; Cochocki and
Unbehauen, 1993; Levin and Narendra, 1995; Miller et al., 1995; Ljung, 1998; Prokhorov et al., 2001;
Ge et al., 2010). Assume M has learned to produce accurate predictions. We can use M to substitute the environment. Then M and C form an RNN where M ’s outputs become inputs of C, whose
outputs (actions) in turn become inputs of M . Now BP for RNNs (Sec. 5.5.1) can be used to achieve
desired input events such as high real-valued reward signals: While M ’s weights remain fixed, gradient information for C’s weights is propagated back through M down into C and back through M etc.
To a certain extent, the approach is also applicable in probabilistic or uncertain environments, as long
as the inner products of M ’s C-based gradient estimates and M ’s “true” gradients tend to be positive.
In general, this approach implies deep CAPs for C, unlike in DP-based traditional RL (Sec. 6.2).
Decades ago, the method was used to learn to back up a model truck (Nguyen and Widrow, 1989).
An RL active vision system used it to learn sequential shifts (saccades) of a fovea, to detect targets in
visual scenes (Schmidhuber and Huber, 1991), thus learning to control selective attention. Compare
RL-based attention learning without NNs (Whitehead, 1992).
To allow for memories of previous events in partially observable worlds (Sec. 6.3), the most general variant of this technique uses RNNs instead of FNNs to implement both M and C (Schmidhuber,
1990d, 1991c; Feldkamp and Puskorius, 1998). This may cause deep CAPs not only for C but also
for M .
M can also be used to optimize expected reward by planning future action sequences (Schmidhuber, 1990d). In fact, the winners of the 2004 RoboCup World Championship in the fast
league (Egorova et al., 2004) trained NNs to predict the effects of steering signals on fast robots
with 4 motors for 4 different wheels. During play, such NN models were used to achieve desirable
subgoals, by optimizing action sequences through quickly planning ahead. The approach also was
used to create self-healing robots able to compensate for faulty motors whose effects do not longer
match the predictions of the NN models (Gloye et al., 2005; Schmidhuber, 2007).
29
Typically M is not given in advance. Then an essential question is: which experiments should C
conduct to quickly improve M ? The Formal Theory of Fun and Creativity (e.g., Schmidhuber, 2006a,
2013b) formalizes driving forces and value functions behind such curious and exploratory behavior:
A measure of the learning progress of M becomes the intrinsic reward of C (Schmidhuber, 1991a);
compare (Singh et al., 2005; Oudeyer et al., 2013). This motivates C to create action sequences
(experiments) such that M makes quick progress.
6.2
Deep FNNs for Traditional RL and Markov Decision Processes (MDPs)
The classical approach to RL (Samuel, 1959; Bertsekas and Tsitsiklis, 1996) makes the simplifying
assumption of Markov Decision Processes (MDPs): the current input of the RL agent conveys all
information necessary to compute an optimal next output event or decision. This allows for greatly
reducing CAP depth in RL NNs (Sec. 3, 6.1) by using the Dynamic Programming (DP) trick (Bellman,
1957). The latter is often explained in a probabilistic framework (e.g., Sutton and Barto, 1998), but
its basic idea can already be conveyed in a deterministic setting. For simplicity, using the notation
of Sec. 2, let input events xt encode the entire current state of the environment, including a realvalued reward rt (no need to introduce additional vector-valued notation, since real values can encode
arbitrary vectors of real values). The original RL goal (find weights that maximize the sum of all
rewards of an episode) is replaced by an equivalent set of alternative goals set by a real-valued value
function V defined on input events. Consider any two subsequent input events xt , xk . Recursively
define V (xt ) = rt + V (xk ), where V (xk ) = rk if xk is the last input event. Now search for weights
that maximize the V of all input events, by causing appropriate output events or actions.
Due to the Markov assumption, an FNN suffices to implement the policy that maps input to output events. Relevant CAPs are not deeper than this FNN. V itself is often modeled by a separate
FNN (also yielding typically short CAPs) learning to approximate V (xt ) only from local information
rt , V (xk ).
Many variants of traditional RL exist (e.g., Barto et al., 1983; Watkins, 1989; Watkins and Dayan,
1992; Moore and Atkeson, 1993; Schwartz, 1993; Rummery and Niranjan, 1994; Singh, 1994; Baird,
1995; Kaelbling et al., 1995; Peng and Williams, 1996; Mahadevan, 1996; Tsitsiklis and van Roy,
1996; Bradtke et al., 1996; Santamarı́a et al., 1997; Prokhorov and Wunsch, 1997; Sutton and Barto,
1998; Wiering and Schmidhuber, 1998b; Baird and Moore, 1999; Meuleau et al., 1999; Morimoto and
Doya, 2000; Bertsekas, 2001; Brafman and Tennenholtz, 2002; Abounadi et al., 2002; Lagoudakis and
Parr, 2003; Sutton et al., 2008; Maei and Sutton, 2010; van Hasselt, 2012). Most are formulated in
a probabilistic framework, and evaluate pairs of input and output (action) events (instead of input
events only). To facilitate certain mathematical derivations, some discount delayed rewards, but such
distortions of the original RL problem are problematic.
Perhaps the most well-known RL NN is the world-class RL backgammon player (Tesauro, 1994),
which achieved the level of human world champions by playing against itself. Its nonlinear, rather
shallow FNN maps a large but finite number of discrete board states to values. More recently, a
rather deep GPU-CNN was used in a traditional RL framework to play several Atari 2600 computer
games directly from 84x84 pixel 60 Hz video input (Mnih et al., 2013), using experience replay (Lin,
1993), extending previous work on Neural Fitted Q-Learning (NFQ) (Riedmiller, 2005). Even better results are achieved by using (slow) Monte Carlo tree planning to train comparatively fast deep
NNs (Guo et al., 2014). Compare RBM-based RL (Sallans and Hinton, 2004) with high-dimensional
inputs (Elfwing et al., 2010), earlier RL Atari players (Grüttner et al., 2010), and an earlier, raw videobased RL NN for computer games (Koutnı́k et al., 2013) trained by Indirect Policy Search (Sec. 6.7).
30
6.3
Deep RL RNNs for Partially Observable MDPs (POMDPs)
The Markov assumption (Sec. 6.2) is often unrealistic. We cannot directly perceive what is behind our
back, let alone the current state of the entire universe. However, memories of previous events can help
to deal with partially observable Markov decision problems (POMDPs) (e.g., Schmidhuber, 1990d,
1991c; Ring, 1991, 1993, 1994; Williams, 1992a; Lin, 1993; Teller, 1994; Kaelbling et al., 1995;
Littman et al., 1995; Boutilier and Poole, 1996; Jaakkola et al., 1995; McCallum, 1996; Kimura et al.,
1997; Wiering and Schmidhuber, 1996, 1998a; Otsuka et al., 2010). A naive way of implementing
memories without leaving the MDP framework (Sec. 6.2) would be to simply consider a possibly huge
state space, namely, the set of all possible observation histories and their prefixes. A more realistic
way is to use function approximators such as RNNs that produce compact state features as a function
of the entire history seen so far. Generally speaking, POMDP RL often uses DL RNNs to learn which
events to memorize and which to ignore. Three basic alternatives are:
1. Use an RNN as a value function mapping arbitrary event histories to values (e.g., Schmidhuber,
1990b, 1991c; Lin, 1993; Bakker, 2002). For example, deep LSTM RNNs were used in this
way for RL robots (Bakker et al., 2003).
2. Use an RNN controller in conjunction with a second RNN as predictive world model, to obtain
a combined RNN with deep CAPs—see Sec. 6.1.
3. Use an RNN for RL by Direct Search (Sec. 6.6) or Indirect Search (Sec. 6.7) in weight space.
In general, however, POMDPs may imply greatly increased CAP depth.
6.4
RL Facilitated by Deep UL in FNNs and RNNs
RL machines may profit from UL for input preprocessing (e.g., Jodogne and Piater, 2007). In particular, an UL NN can learn to compactly encode environmental inputs such as images or videos, e.g.,
Sec. 5.7, 5.10, 5.15. The compact codes (instead of the high-dimensional raw data) can be fed into an
RL machine, whose job thus may become much easier (Legenstein et al., 2010; Cuccu et al., 2011),
just like SL may profit from UL, e.g., Sec. 5.7, 5.10, 5.15. For example, NFQ (Riedmiller, 2005) was
applied to real-world control tasks (Lange and Riedmiller, 2010; Riedmiller et al., 2012) where purely
visual inputs were compactly encoded by deep autoencoders (Sec. 5.7, 5.15). RL combined with UL
based on Slow Feature Analysis (Wiskott and Sejnowski, 2002; Kompella et al., 2012) enabled a real
humanoid robot to learn skills from raw high-dimensional video streams (Luciw et al., 2013). To
deal with POMDPs (Sec. 6.3) involving high-dimensional inputs, RBM-based RL was used (Otsuka,
2010), and a RAAM (Pollack, 1988) (Sec. 5.7) was employed as a deep unsupervised sequence encoder for RL (Gisslen et al., 2011). Certain types of RL and UL also were combined in biologically
plausible RNNs with spiking neurons (Sec. 5.26) (e.g., Yin et al., 2012; Klampfl and Maass, 2013;
Rezende and Gerstner, 2014).
6.5
Deep Hierarchical RL (HRL) and Subgoal Learning with FNNs and RNNs
Multiple learnable levels of abstraction (Fu, 1977; Lenat and Brown, 1984; Ring, 1994; Bengio
et al., 2013; Deng and Yu, 2014) seem as important for RL as for SL. Work on NN-based Hierarchical RL (HRL) has been published since the early 1990s. In particular, gradient-based subgoal
discovery with FNNs or RNNs decomposes RL tasks into subtasks for RL submodules (Schmidhuber, 1991b; Schmidhuber and Wahnsiedler, 1992). Numerous alternative HRL techniques have
been proposed (e.g., Ring, 1991, 1994; Jameson, 1991; Tenenberg et al., 1993; Weiss, 1994; Moore
and Atkeson, 1995; Precup et al., 1998; Dietterich, 2000b; Menache et al., 2002; Doya et al., 2002;
31
Ghavamzadeh and Mahadevan, 2003; Barto and Mahadevan, 2003; Samejima et al., 2003; Bakker and
Schmidhuber, 2004; Whiteson et al., 2005; Simsek and Barto, 2008). While HRL frameworks such as
Feudal RL (Dayan and Hinton, 1993) and options (Sutton et al., 1999b; Barto et al., 2004; Singh et al.,
2005) do not directly address the problem of automatic subgoal discovery, HQ-Learning (Wiering and
Schmidhuber, 1998a) automatically decomposes POMDPs (Sec. 6.3) into sequences of simpler subtasks that can be solved by memoryless policies learnable by reactive sub-agents. Recent HRL organizes potentially deep NN-based RL sub-modules into self-organizing, 2-dimensional motor control
maps (Ring et al., 2011) inspired by neurophysiological findings (Graziano, 2009).
6.6
Deep RL by Direct NN Search / Policy Gradients / Evolution
Not quite as universal as the methods of Sec. 6.8, yet both practical and more general than most
traditional RL algorithms (Sec. 6.2), are methods for Direct Policy Search (DS). Without a need for
value functions or Markovian assumptions (Sec. 6.2, 6.3), the weights of an FNN or RNN are directly
evaluated on the given RL problem. The results of successive trials inform further search for better
weights. Unlike with RL supported by BP (Sec. 5.5, 6.3, 6.1), CAP depth (Sec. 3, 5.9) is not a crucial
issue. DS may solve the credit assignment problem without backtracking through deep causal chains
of modifiable parameters—it neither cares for their existence, nor tries to exploit them.
An important class of DS methods for NNs are Policy Gradient methods (Williams, 1986, 1988,
1992a; Sutton et al., 1999a; Baxter and Bartlett, 2001; Aberdeen, 2003; Ghavamzadeh and Mahadevan, 2003; Kohl and Stone, 2004; Wierstra et al., 2008; Rückstieß et al., 2008; Peters and Schaal,
2008b,a; Sehnke et al., 2010; Grüttner et al., 2010; Wierstra et al., 2010; Peters, 2010; Grondman
et al., 2012; Heess et al., 2012). Gradients of the total reward with respect to policies (NN weights)
are estimated (and then exploited) through repeated NN evaluations.
RL NNs can also be evolved through Evolutionary Algorithms (EAs) (Rechenberg, 1971; Schwefel, 1974; Holland, 1975; Fogel et al., 1966; Goldberg, 1989) in a series of trials. Here several policies
are represented by a population of NNs improved through mutations and/or repeated recombinations
of the population’s fittest individuals (e.g., Montana and Davis, 1989; Fogel et al., 1990; Maniezzo,
1994; Happel and Murre, 1994; Nolfi et al., 1994b). Compare Genetic Programming (GP) (Cramer,
1985) (see also Smith, 1980) which can be used to evolve computer programs of variable size (Dickmanns et al., 1987; Koza, 1992), and Cartesian GP (Miller and Thomson, 2000; Miller and Harding,
2009) for evolving graph-like programs, including NNs (Khan et al., 2010) and their topology (Turner
and Miller, 2013). Related methods include probability distribution-based EAs (Baluja, 1994; Saravanan and Fogel, 1995; Sałustowicz and Schmidhuber, 1997; Larraanaga and Lozano, 2001), Covariance Matrix Estimation Evolution Strategies (CMA-ES) (Hansen and Ostermeier, 2001; Hansen
et al., 2003; Igel, 2003; Heidrich-Meisner and Igel, 2009), and NeuroEvolution of Augmenting Topologies (NEAT) (Stanley and Miikkulainen, 2002). Hybrid methods combine traditional NN-based RL
(Sec. 6.2) and EAs (e.g., Whiteson and Stone, 2006).
Since RNNs are general computers, RNN evolution is like GP in the sense that it can evolve
general programs. Unlike sequential programs learned by traditional GP, however, RNNs can mix
sequential and parallel information processing in a natural and efficient way, as already mentioned in
Sec. 1. Many RNN evolvers have been proposed (e.g., Miller et al., 1989; Wieland, 1991; Cliff et al.,
1993; Yao, 1993; Nolfi et al., 1994a; Sims, 1994; Yamauchi and Beer, 1994; Miglino et al., 1995;
Moriarty, 1997; Pasemann et al., 1999; Juang, 2004; Whiteson, 2012). One particularly effective
family of methods coevolves neurons, combining them into networks, and selecting those neurons
for reproduction that participated in the best-performing networks (Moriarty and Miikkulainen, 1996;
Gomez, 2003; Gomez and Miikkulainen, 2003). This can help to solve deep POMDPs (Gomez and
Schmidhuber, 2005). Co-Synaptic Neuro-Evolution (CoSyNE) does something similar on the level of
32
synapses or weights (Gomez et al., 2008); benefits of this were shown on difficult nonlinear POMDP
benchmarks.
Natural Evolution Strategies (NES) (Wierstra et al., 2008; Glasmachers et al., 2010; Sun et al.,
2009, 2013) link policy gradient methods and evolutionary approaches through the concept of Natural
Gradients (Amari, 1998). RNN evolution may also help to improve SL for deep RNNs through
Evolino (Schmidhuber et al., 2007) (Sec. 5.9).
6.7
Deep RL by Indirect Policy Search / Compressed NN Search
Some DS methods (Sec. 6.6) can evolve NNs with hundreds or thousands of weights, but not millions. How to search for large and deep NNs? Most SL and RL methods mentioned so far somehow
search the space of weights wi . Some profit from a reduction of the search space through shared
wi that get reused over and over again, e.g., in CNNs (Sec. 5.4, 5.8, 5.16, 5.21), or in RNNs for SL
(Sec. 5.5, 5.13, 5.17) and RL (Sec. 6.1, 6.3, 6.6).
It may be possible, however, to exploit additional regularities/compressibilities in the space of solutions, through indirect search in weight space. Instead of evolving large NNs directly (Sec. 6.6), one
can sometimes greatly reduce the search space by evolving compact encodings of NNs, e.g., through
Lindenmeyer Systems (Lindenmayer, 1968; Jacob et al., 1994), graph rewriting (Kitano, 1990), Cellular Encoding (Gruau et al., 1996), HyperNEAT (D’Ambrosio and Stanley, 2007; Stanley et al., 2009;
Clune et al., 2011; van den Berg and Whiteson, 2013) (extending NEAT; Sec. 6.6), and extensions
thereof (e.g., Risi and Stanley, 2012). This helps to avoid overfitting (compare Sec. 5.6.3, 5.24) and is
closely related to the topics of regularisation and MDL (Sec. 4.4).
A general approach (Schmidhuber, 1997) for both SL and RL seeks to compactly encode weights
of large NNs (Schmidhuber, 1997) through programs written in a universal programming language (Gödel, 1931; Church, 1936; Turing, 1936; Post, 1936). Often it is much more efficient to
systematically search the space of such programs with a bias towards short and fast programs (Levin,
1973b; Schmidhuber, 1997, 2004), instead of directly searching the huge space of possible NN weight
matrices. A previous universal language for encoding NNs was assembler-like (Schmidhuber, 1997).
More recent work uses more practical languages based on coefficients of popular transforms (Fourier,
wavelet, etc). In particular, RNN weight matrices may be compressed like images, by encoding them
through the coefficients of a discrete cosine transform (DCT) (Koutnı́k et al., 2010, 2013). Compact
DCT-based descriptions can be evolved through NES or CoSyNE (Sec. 6.6). An RNN with over a
million weights learned (without a teacher) to drive a simulated car in the TORCS driving game (Loiacono et al., 2009, 2011), based on a high-dimensional video-like visual input stream (Koutnı́k et al.,
2013). The RNN learned both control and visual processing from scratch, without being aided by
UL. (Of course, UL might help to generate more compact image codes (Sec. 6.4, 4.2) to be fed into a
smaller RNN, to reduce the overall computational effort.)
6.8
Universal RL
General purpose learning algorithms may improve themselves in open-ended fashion and
environment-specific ways in a lifelong learning context (Schmidhuber, 1987; Schmidhuber et al.,
1997b,a; Schaul and Schmidhuber, 2010). The most general type of RL is constrained only by the
fundamental limitations of computability identified by the founders of theoretical computer science
(Gödel, 1931; Church, 1936; Turing, 1936; Post, 1936). Remarkably, there exist blueprints of universal problem solvers or universal RL machines for unlimited problem depth that are time-optimal in
various theoretical senses (Hutter, 2005, 2002; Schmidhuber, 2002, 2006b). In particular, the Gödel
Machine can be implemented on general computers such as RNNs and may improve any part of its
software (including the learning algorithm itself) in a way that is provably time-optimal in a certain
33
sense (Schmidhuber, 2006b). It can be initialized by an asymptotically optimal meta-method (Hutter, 2002) (also applicable to RNNs) which will solve any well-defined problem as quickly as the
unknown fastest way of solving it, save for an additive constant overhead that becomes negligible as
problem size grows. Note that most problems are large; only few are small. AI and DL researchers are
still in business because many are interested in problems so small that it is worth trying to reduce the
overhead through less general methods, including heuristics. Here I won’t further discuss universal
RL methods, which go beyond what is usually called DL.
7
Conclusion and Outlook
Deep Learning (DL) in Neural Networks (NNs) is relevant for Supervised Learning (SL) (Sec. 5),
Unsupervised Learning (UL) (Sec. 5), and Reinforcement Learning (RL) (Sec. 6). By alleviating
problems with deep Credit Assignment Paths (CAPs, Sec. 3, 5.9), UL (Sec. 5.6.4) can not only facilitate SL of sequences (Sec. 5.10) and stationary patterns (Sec. 5.7, 5.15), but also RL (Sec. 6.4, 4.2).
Dynamic Programming (DP, Sec. 4.1) is important for both deep SL (Sec. 5.5) and traditional RL with
deep NNs (Sec. 6.2). A search for solution-computing, perturbation-resistant (Sec. 5.6.3, 5.15, 5.24),
low-complexity NNs describable by few bits of information (Sec. 4.4) can reduce overfitting and improve deep SL & UL (Sec. 5.6.3, 5.6.4) as well as RL (Sec. 6.7), also in the case of partially observable
environments (Sec. 6.3). Deep SL, UL, RL often create hierarchies of more and more abstract representations of stationary data (Sec. 5.3, 5.7, 5.15), sequential data (Sec. 5.10), or RL policies (Sec. 6.5).
While UL can facilitate SL, pure SL for feedforward NNs (FNNs) (Sec. 5.5, 5.8, 5.16, 5.18) and recurrent NNs (RNNs) (Sec. 5.5, 5.13) did not only win early contests (Sec. 5.12, 5.14) but also most
of the recent ones (Sec. 5.17–5.22). Especially DL in FNNs profited from GPU implementations
(Sec. 5.16–5.19). In particular, GPU-based (Sec. 5.19) Max-Pooling (Sec. 5.11) Convolutional NNs
(Sec. 5.4, 5.8, 5.16) won competitions not only in pattern recognition (Sec. 5.19–5.22) but also image
segmentation (Sec. 5.21) and object detection (Sec. 5.21, 5.22).
Unlike these systems, humans learn to actively perceive patterns by sequentially directing attention to relevant parts of the available data. Near future deep NNs will do so, too, extending previous
work since 1990 on NNs that learn selective attention through RL of (a) motor actions such as saccade
control (Sec. 6.1) and (b) internal actions controlling spotlights of attention within RNNs, thus closing
the general sensorimotor loop through both external and internal feedback (e.g., Sec. 2, 5.21, 6.6, 6.7).
Many future deep NNs will also take into account that it costs energy to activate neurons, and to
send signals between them. Brains seem to minimize such computational costs during problem solving in at least two ways: (1) At a given time, only a small fraction of all neurons is active because local
competition through winner-take-all mechanisms shuts down many neighbouring neurons, and only
winners can activate other neurons through outgoing connections (compare SLIM NNs; Sec. 5.24).
(2) Numerous neurons are sparsely connected in a compact 3D volume by many short-range and
few long-range connections (much like microchips in traditional supercomputers). Often neighbouring neurons are allocated to solve a single task, thus reducing communication costs. Physics seems
to dictate that any efficient computational hardware will in the future also have to be brain-like in
keeping with these two constraints. The most successful current deep RNNs, however, are not. Unlike certain spiking NNs (Sec. 5.26), they usually activate all units at least slightly, and tend to be
strongly connected, ignoring natural constraints of 3D hardware. It should be possible to improve
them by adopting (1) and (2), and by minimizing non-differentiable energy and communication costs
through direct search in program (weight) space (e.g., Sec. 6.6, 6.7). These more brain-like RNNs
will allocate neighboring RNN parts to related behaviors, and distant RNN parts to less related ones,
thus self-modularizing in a way more general than that of traditional self-organizing maps in FNNs
(Sec. 5.6.4). They will also implement Occam’s razor (Sec. 4.4, 5.6.3) as a by-product of energy min34
imization, by finding simple (highly generalizing) problem solutions that require few active neurons
and few, mostly short connections.
The more distant future may belong to general purpose learning algorithms that improve themselves in provably optimal ways (Sec. 6.8), but these are not yet practical or commercially relevant.
8
Acknowledgments
Since 16 April 2014, drafts of this paper have undergone massive open online peer review through
public mailing lists including [email protected], [email protected], [email protected], genetic [email protected], [email protected], [email protected], Google+ machine learning forum. Thanks to numerous NN / DL experts for valuable
comments. Thanks to SNF, DFG, and the European Commission for partially funding my DL research group in the past quarter-century. The contents of this paper may be used for educational and
non-commercial purposes, including articles for Wikipedia and similar sites.
References
Aberdeen, D. (2003). Policy-Gradient Algorithms for Partially Observable Markov Decision Processes. PhD thesis, Australian National University.
Abounadi, J., Bertsekas, D., and Borkar, V. S. (2002). Learning algorithms for Markov decision
processes with average cost. SIAM Journal on Control and Optimization, 40(3):681–698.
Akaike, H. (1970). Statistical predictor identification. Ann. Inst. Statist. Math., 22:203–217.
Akaike, H. (1973). Information theory and an extension of the maximum likelihood principle. In
Second Intl. Symposium on Information Theory, pages 267–281. Akademinai Kiado.
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic
Control, 19(6):716–723.
Allender, A. (1992). Application of time-bounded Kolmogorov complexity in complexity theory. In
Watanabe, O., editor, Kolmogorov complexity and computational complexity, pages 6–22. EATCS
Monographs on Theoretical Computer Science, Springer.
Almeida, L. B. (1987). A learning rule for asynchronous perceptrons with feedback in a combinatorial
environment. In IEEE 1st International Conference on Neural Networks, San Diego, volume 2,
pages 609–618.
Almeida, L. B., Almeida, L. B., Langlois, T., Amaral, J. D., and Redol, R. A. (1997). On-line step
size adaptation. Technical report, INESC, 9 Rua Alves Redol, 1000.
Amari, S. (1967). A theory of adaptive pattern classifiers. IEEE Trans. EC, 16(3):299–307.
Amari, S., Cichocki, A., and Yang, H. (1996). A new learning algorithm for blind signal separation.
In Touretzky, D. S., Mozer, M. C., and Hasselmo, M. E., editors, Advances in Neural Information
Processing Systems (NIPS), volume 8. The MIT Press.
Amari, S. and Murata, N. (1993). Statistical theory of learning curves under entropic loss criterion.
Neural Computation, 5(1):140–153.
35
Amari, S.-I. (1998). Natural gradient works efficiently in learning. Neural Computation, 10(2):251–
276.
Amit, D. J. and Brunel, N. (1997). Dynamics of a recurrent network of spiking neurons before and
following learning. Network: Computation in Neural Systems, 8(4):373–404.
An, G. (1996). The effects of adding noise during backpropagation training on a generalization
performance. Neural Computation, 8(3):643–674.
Andrade, M. A., Chacon, P., Merelo, J. J., and Moran, F. (1993). Evaluation of secondary structure
of proteins from UV circular dichroism spectra using an unsupervised learning neural network.
Protein Engineering, 6(4):383–390.
Andrews, R., Diederich, J., and Tickle, A. B. (1995). Survey and critique of techniques for extracting
rules from trained artificial neural networks. Knowledge-Based Systems, 8(6):373–389.
Anguita, D. and Gomes, B. A. (1996). Mixing floating- and fixed-point formats for neural network
learning on neuroprocessors. Microprocessing and Microprogramming, 41(10):757 – 769.
Anguita, D., Parodi, G., and Zunino, R. (1994). An efficient implementation of BP on RISC-based
workstations. Neurocomputing, 6(1):57 – 65.
Arel, I., Rose, D. C., and Karnowski, T. P. (2010). Deep machine learning – a new frontier in artificial
intelligence research. Computational Intelligence Magazine, IEEE, 5(4):13–18.
Ash, T. (1989). Dynamic node creation in backpropagation neural networks. Connection Science,
1(4):365–375.
Atick, J. J., Li, Z., and Redlich, A. N. (1992). Understanding retinal color coding from first principles.
Neural Computation, 4:559–572.
Atiya, A. F. and Parlos, A. G. (2000). New results on recurrent network training: unifying the algorithms and accelerating convergence. IEEE Transactions on Neural Networks, 11(3):697–709.
Ba, J. and Frey, B. (2013). Adaptive dropout for training deep neural networks. In Advances in Neural
Information Processing Systems (NIPS), pages 3084–3092.
Baird, H. (1990). Document image defect models. In Proceddings, IAPR Workshop on Syntactic and
Structural Pattern Recognition, Murray Hill, NJ.
Baird, L. and Moore, A. W. (1999). Gradient descent for general reinforcement learning. In Advances
in neural information processing systems 12 (NIPS), pages 968–974. MIT Press.
Baird, L. C. (1995). Residual algorithms: Reinforcement learning with function approximation. In
International Conference on Machine Learning, pages 30–37.
Bakker, B. (2002). Reinforcement learning with Long Short-Term Memory. In Dietterich, T. G.,
Becker, S., and Ghahramani, Z., editors, Advances in Neural Information Processing Systems 14,
pages 1475–1482. MIT Press, Cambridge, MA.
Bakker, B. and Schmidhuber, J. (2004). Hierarchical reinforcement learning based on subgoal discovery and subpolicy specialization. In et al., F. G., editor, Proc. 8th Conference on Intelligent
Autonomous Systems IAS-8, pages 438–445, Amsterdam, NL. IOS Press.
36
Bakker, B., Zhumatiy, V., Gruener, G., and Schmidhuber, J. (2003). A robot that reinforcement-learns
to identify and memorize important previous observations. In Proceedings of the 2003 IEEE/RSJ
International Conference on Intelligent Robots and Systems, IROS 2003, pages 430–435.
Baldi, P. (1995). Gradient descent learning algorithms overview: A general dynamical systems perspective. IEEE Transactions on Neural Networks, 6(1):182–195.
Baldi, P. (2012). Autoencoders, unsupervised learning, and deep architectures. Journal of Machine
Learning Research (Proc. 2011 ICML Workshop on Unsupervised and Transfer Learning), 27:37–
50.
Baldi, P., Brunak, S., Frasconi, P., Pollastri, G., and Soda, G. (1999). Exploiting the past and the
future in protein secondary structure prediction. Bioinformatics, 15:937–946.
Baldi, P. and Chauvin, Y. (1993). Neural networks for fingerprint recognition. Neural Computation,
5(3):402–418.
Baldi, P. and Chauvin, Y. (1996). Hybrid modeling, HMM/NN architectures, and protein applications.
Neural Computation, 8(7):1541–1565.
Baldi, P. and Hornik, K. (1989). Neural networks and principal component analysis: Learning from
examples without local minima. Neural Networks, 2:53–58.
Baldi, P. and Hornik, K. (1994). Learning in linear networks: a survey. IEEE Transactions on Neural
Networks, 6(4):837–858. 1995.
Baldi, P. and Pollastri, G. (2003). The principled design of large-scale recursive neural network
architectures – DAG-RNNs and the protein structure prediction problem. J. Mach. Learn. Res.,
4:575–602.
Baldi, P. and Sadowski, P. (2014). The dropout learning algorithm. Artificial Intelligence, 210C:78–
122.
Ballard, D. H. (1987). Modular learning in neural networks. In Proc. AAAI, pages 279–284.
Baluja, S. (1994). Population-based incremental learning: A method for integrating genetic
search based function optimization and competitive learning. Technical Report CMU-CS-94-163,
Carnegie Mellon University.
Balzer, R. (1985). A 15 year perspective on automatic programming. IEEE Transactions on Software
Engineering, 11(11):1257–1268.
Barlow, H. B. (1989). Unsupervised learning. Neural Computation, 1(3):295–311.
Barlow, H. B., Kaushal, T. P., and Mitchison, G. J. (1989). Finding minimum entropy codes. Neural
Computation, 1(3):412–423.
Barrow, H. G. (1987). Learning receptive fields. In Proceedings of the IEEE 1st Annual Conference
on Neural Networks, volume IV, pages 115–121. IEEE.
Barto, A. G. and Mahadevan, S. (2003). Recent advances in hierarchical reinforcement learning.
Discrete Event Dynamic Systems, 13(4):341–379.
37
Barto, A. G., Singh, S., and Chentanez, N. (2004). Intrinsically motivated learning of hierarchical collections of skills. In Proceedings of International Conference on Developmental Learning (ICDL),
pages 112–119. MIT Press, Cambridge, MA.
Barto, A. G., Sutton, R. S., and Anderson, C. W. (1983). Neuronlike adaptive elements that can
solve difficult learning control problems. IEEE Transactions on Systems, Man, and Cybernetics,
SMC-13:834–846.
Battiti, R. (1989). Accelerated backpropagation learning: two optimization methods. Complex Systems, 3(4):331–342.
Battiti, T. (1992). First- and second-order methods for learning: Between steepest descent and Newton’s method. Neural Computation, 4(2):141–166.
Baum, E. B. and Haussler, D. (1989). What size net gives valid generalization? Neural Computation,
1(1):151–160.
Baum, L. E. and Petrie, T. (1966). Statistical inference for probabilistic functions of finite state
Markov chains. The Annals of Mathematical Statistics, pages 1554–1563.
Baxter, J. and Bartlett, P. L. (2001). Infinite-horizon policy-gradient estimation. J. Artif. Int. Res.,
15(1):319–350.
Bayer, J. and Osendorfer, C. (2014). Variational inference of latent state sequences using recurrent
networks. arXiv preprint arXiv:1406.1655.
Bayer, J., Osendorfer, C., Chen, N., Urban, S., and van der Smagt, P. (2013). On fast dropout and its
applicability to recurrent networks. arXiv preprint arXiv:1311.0701.
Bayer, J., Wierstra, D., Togelius, J., and Schmidhuber, J. (2009). Evolving memory cell structures for
sequence learning. In Proc. ICANN (2), pages 755–764.
Bayes, T. (1763). An essay toward solving a problem in the doctrine of chances. Philosophical
Transactions of the Royal Society of London, 53:370–418. Communicated by R. Price, in a letter
to J. Canton.
Becker, S. (1991). Unsupervised learning procedures for neural networks. International Journal of
Neural Systems, 2(1 & 2):17–33.
Becker, S. and Le Cun, Y. (1989). Improving the convergence of back-propagation learning with
second order methods. In Touretzky, D., Hinton, G., and Sejnowski, T., editors, Proc. 1988 Connectionist Models Summer School, pages 29–37, Pittsburg 1988. Morgan Kaufmann, San Mateo.
Behnke, S. (1999). Hebbian learning and competition in the neural abstraction pyramid. In Proceedings of the International Joint Conference on Neural Networks (IJCNN), volume 2, pages
1356–1361.
Behnke, S. (2001). Learning iterative image reconstruction in the neural abstraction pyramid. International Journal of Computational Intelligence and Applications, 1(4):427–438.
Behnke, S. (2002). Learning face localization using hierarchical recurrent networks. In Proceedings
of the 12th International Conference on Artificial Neural Networks (ICANN), Madrid, Spain, pages
1319–1324.
38
Behnke, S. (2003a). Discovering hierarchical speech features using convolutional non-negative matrix
factorization. In Proceedings of the International Joint Conference on Neural Networks (IJCNN),
volume 4, pages 2758–2763.
Behnke, S. (2003b). Hierarchical Neural Networks for Image Interpretation, volume LNCS 2766 of
Lecture Notes in Computer Science. Springer.
Behnke, S. (2005). Face localization and tracking in the Neural Abstraction Pyramid. Neural Computing and Applications, 14(2):97–103.
Behnke, S. and Rojas, R. (1998). Neural abstraction pyramid: A hierarchical image understanding architecture. In Proceedings of International Joint Conference on Neural Networks (IJCNN),
volume 2, pages 820–825.
Bell, A. J. and Sejnowski, T. J. (1995). An information-maximization approach to blind separation
and blind deconvolution. Neural Computation, 7(6):1129–1159.
Bellman, R. (1957). Dynamic Programming. Princeton University Press, Princeton, NJ, USA, 1st
edition.
Belouchrani, A., Abed-Meraim, K., Cardoso, J.-F., and Moulines, E. (1997). A blind source separation
technique using second-order statistics. IEEE Transactions on Signal Processing, 45(2):434–444.
Bengio, Y. (1991). Artificial Neural Networks and their Application to Sequence Recognition. PhD
thesis, McGill University, (Computer Science), Montreal, Qc., Canada.
Bengio, Y. (2009). Learning Deep Architectures for AI. Foundations and Trends in Machine Learning,
V2(1). Now Publishers.
Bengio, Y., Courville, A., and Vincent, P. (2013). Representation learning: A review and new perspectives. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 35(8):1798–1828.
Bengio, Y., Lamblin, P., Popovici, D., and Larochelle, H. (2007). Greedy layer-wise training of deep
networks. In Cowan, J. D., Tesauro, G., and Alspector, J., editors, Advances in Neural Information
Processing Systems 19 (NIPS), pages 153–160. MIT Press.
Bengio, Y., Simard, P., and Frasconi, P. (1994). Learning long-term dependencies with gradient
descent is difficult. IEEE Transactions on Neural Networks, 5(2):157–166.
Beringer, N., Graves, A., Schiel, F., and Schmidhuber, J. (2005). Classifying unprompted speech by
retraining LSTM nets. In Duch, W., Kacprzyk, J., Oja, E., and Zadrozny, S., editors, Artificial
Neural Networks: Biological Inspirations - ICANN 2005, LNCS 3696, pages 575–581. SpringerVerlag Berlin Heidelberg.
Bertsekas, D. P. (2001). Dynamic Programming and Optimal Control. Athena Scientific.
Bertsekas, D. P. and Tsitsiklis, J. N. (1996). Neuro-dynamic Programming. Athena Scientific, Belmont, MA.
Bichot, N. P., Rossi, A. F., and Desimone, R. (2005). Parallel and serial neural mechanisms for visual
search in macaque area V4. Science, 308:529–534.
Biegler-König, F. and Bärmann, F. (1993). A learning algorithm for multilayered neural networks
based on linear least squares problems. Neural Networks, 6(1):127–131.
39
Bishop, C. M. (1993). Curvature-driven smoothing: A learning algorithm for feed-forward networks.
IEEE Transactions on Neural Networks, 4(5):882–884.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
Blair, A. D. and Pollack, J. B. (1997). Analysis of dynamical recognizers. Neural Computation,
9(5):1127–1142.
Blondel, V. D. and Tsitsiklis, J. N. (2000). A survey of computational complexity results in systems
and control. Automatica, 36(9):1249–1274.
Bluche, T., Louradour, J., Knibbe, M., Moysset, B., Benzeghiba, F., and Kermorvant, C. (2014).
The A2iA Arabic Handwritten Text Recognition System at the OpenHaRT2013 Evaluation. In
International Workshop on Document Analysis Systems.
Blum, A. L. and Rivest, R. L. (1992). Training a 3-node neural network is np-complete. Neural
Networks, 5(1):117–127.
Blumer, A., Ehrenfeucht, A., Haussler, D., and Warmuth, M. K. (1987). Occam’s razor. Information
Processing Letters, 24:377–380.
Bobrowski, L. (1978). Learning processes in multilayer threshold nets. Biological Cybernetics, 31:1–
6.
Bodén, M. and Wiles, J. (2000). Context-free and context-sensitive dynamics in recurrent neural
networks. Connection Science, 12(3-4):197–210.
Bodenhausen, U. and Waibel, A. (1991). The Tempo 2 algorithm: Adjusting time-delays by supervised learning. In Lippman, D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural
Information Processing Systems 3, pages 155–161. Morgan Kaufmann.
Bohte, S. M., Kok, J. N., and La Poutre, H. (2002). Error-backpropagation in temporally encoded
networks of spiking neurons. Neurocomputing, 48(1):17–37.
Boltzmann, L. (1909). In Hasenöhrl, F., editor, Wissenschaftliche Abhandlungen (collection of Boltzmann’s articles in scientific journals). Barth, Leipzig.
Bottou, L. (1991). Une approche théorique de l’apprentissage connexioniste; applications à la reconnaissance de la parole. PhD thesis, Université de Paris XI.
Bourlard, H. and Morgan, N. (1994). Connnectionist Speech Recognition: A Hybrid Approach.
Kluwer Academic Publishers.
Boutilier, C. and Poole, D. (1996). Computing optimal policies for partially observable Markov
decision processes using compact representations. In Proceedings of the AAAI, Portland, OR.
Bradtke, S. J., Barto, A. G., and Kaelbling, L. P. (1996). Linear least-squares algorithms for temporal
difference learning. In Machine Learning, pages 22–33.
Brafman, R. I. and Tennenholtz, M. (2002). R-MAX—a general polynomial time algorithm for nearoptimal reinforcement learning. Journal of Machine Learning Research, 3:213–231.
Brea, J., Senn, W., and Pfister, J.-P. (2013). Matching recall and storage in sequence learning with
spiking neural networks. The Journal of Neuroscience, 33(23):9565–9575.
40
Breiman, L. (1996). Bagging predictors. Machine Learning, 24:123–140.
Brette, R., Rudolph, M., Carnevale, T., Hines, M., Beeman, D., Bower, J. M., Diesmann, M., Morrison, A., Goodman, P. H., Harris Jr, F. C., et al. (2007). Simulation of networks of spiking neurons:
a review of tools and strategies. Journal of Computational Neuroscience, 23(3):349–398.
Breuel, T. M., Ul-Hasan, A., Al-Azawi, M. A., and Shafait, F. (2013). High-performance OCR for
printed English and Fraktur using LSTM networks. In 12th International Conference on Document
Analysis and Recognition (ICDAR), pages 683–687. IEEE.
Bromley, J., Bentz, J. W., Bottou, L., Guyon, I., LeCun, Y., Moore, C., Sackinger, E., and Shah, R.
(1993). Signature verification using a Siamese time delay neural network. International Journal of
Pattern Recognition and Artificial Intelligence, 7(4):669–688.
Broyden, C. G. et al. (1965). A class of methods for solving nonlinear simultaneous equations. Math.
Comp, 19(92):577–593.
Brueckner, R. and Schulter, B. (2014). Social signal classification using deep BLSTM recurrent neural
networks. In Proceedings 39th IEEE International Conference on Acoustics, Speech, and Signal
Processing, ICASSP 2014, Florence, Italy, pages 4856–4860.
Brunel, N. (2000). Dynamics of sparsely connected networks of excitatory and inhibitory spiking
neurons. Journal of Computational Neuroscience, 8(3):183–208.
Bryson, A. and Ho, Y. (1969). Applied optimal control: optimization, estimation, and control. Blaisdell Pub. Co.
Bryson, A. E. (1961). A gradient method for optimizing multi-stage allocation processes. In Proc.
Harvard Univ. Symposium on digital computers and their applications.
Bryson, Jr., A. E. and Denham, W. F. (1961). A steepest-ascent method for solving optimum programming problems. Technical Report BR-1303, Raytheon Company, Missle and Space Division.
Buhler, J. (2001). Efficient large-scale sequence comparison by locality-sensitive hashing. Bioinformatics, 17(5):419–428.
Buntine, W. L. and Weigend, A. S. (1991). Bayesian back-propagation. Complex Systems, 5:603–643.
Burgess, N. (1994). A constructive algorithm that converges for real-valued input patterns. International Journal of Neural Systems, 5(1):59–66.
Cardoso, J.-F. (1994). On the performance of orthogonal source separation algorithms. In Proc.
EUSIPCO, pages 776–779.
Carreira-Perpinan, M. A. (2001). Continuous latent variable models for dimensionality reduction and
sequential data reconstruction. PhD thesis, University of Sheffield UK.
Carter, M. J., Rudolph, F. J., and Nucci, A. J. (1990). Operational fault tolerance of CMAC networks.
In Touretzky, D. S., editor, Advances in Neural Information Processing Systems (NIPS) 2, pages
340–347. San Mateo, CA: Morgan Kaufmann.
Caruana, R. (1997). Multitask learning. Machine Learning, 28(1):41–75.
Casey, M. P. (1996). The dynamics of discrete-time computation, with application to recurrent neural
networks and finite state machine extraction. Neural Computation, 8(6):1135–1178.
41
Cauwenberghs, G. (1993). A fast stochastic error-descent algorithm for supervised learning and optimization. In Lippman, D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural
Information Processing Systems 5, pages 244–244. Morgan Kaufmann.
Chaitin, G. J. (1966). On the length of programs for computing finite binary sequences. Journal of
the ACM, 13:547–569.
Chalup, S. K. and Blair, A. D. (2003). Incremental training of first order recurrent neural networks to
predict a context-sensitive language. Neural Networks, 16(7):955–972.
Chellapilla, K., Puri, S., and Simard, P. (2006). High performance convolutional neural networks for
document processing. In International Workshop on Frontiers in Handwriting Recognition.
Chen, K. and Salman, A. (2011). Learning speaker-specific characteristics with a deep neural architecture. IEEE Transactions on Neural Networks, 22(11):1744–1756.
Cho, K. (2014). Foundations and Advances in Deep Learning. PhD thesis, Aalto University School
of Science.
Cho, K., Ilin, A., and Raiko, T. (2012). Tikhonov-type regularization for restricted Boltzmann machines. In Intl. Conf. on Artificial Neural Networks (ICANN) 2012, pages 81–88. Springer.
Cho, K., Raiko, T., and Ilin, A. (2013). Enhanced gradient for training restricted Boltzmann machines.
Neural Computation, 25(3):805–831.
Church, A. (1936). An unsolvable problem of elementary number theory. American Journal of
Mathematics, 58:345–363.
Ciresan, D. C., Giusti, A., Gambardella, L. M., and Schmidhuber, J. (2012a). Deep neural networks
segment neuronal membranes in electron microscopy images. In Advances in Neural Information
Processing Systems (NIPS), pages 2852–2860.
Ciresan, D. C., Giusti, A., Gambardella, L. M., and Schmidhuber, J. (2013). Mitosis detection in
breast cancer histology images with deep neural networks. In Proc. MICCAI, volume 2, pages
411–418.
Ciresan, D. C., Meier, U., Gambardella, L. M., and Schmidhuber, J. (2010). Deep big simple neural
nets for handwritten digit recogntion. Neural Computation, 22(12):3207–3220.
Ciresan, D. C., Meier, U., Masci, J., Gambardella, L. M., and Schmidhuber, J. (2011a). Flexible, high
performance convolutional neural networks for image classification. In Intl. Joint Conference on
Artificial Intelligence IJCAI, pages 1237–1242.
Ciresan, D. C., Meier, U., Masci, J., and Schmidhuber, J. (2011b). A committee of neural networks
for traffic sign classification. In International Joint Conference on Neural Networks (IJCNN), pages
1918–1921.
Ciresan, D. C., Meier, U., Masci, J., and Schmidhuber, J. (2012b). Multi-column deep neural network
for traffic sign classification. Neural Networks, 32:333–338.
Ciresan, D. C., Meier, U., and Schmidhuber, J. (2012c). Multi-column deep neural networks for image
classification. In IEEE Conference on Computer Vision and Pattern Recognition CVPR 2012. Long
preprint arXiv:1202.2745v1 [cs.CV].
42
Ciresan, D. C., Meier, U., and Schmidhuber, J. (2012d). Transfer learning for Latin and Chinese characters with deep neural networks. In International Joint Conference on Neural Networks (IJCNN),
pages 1301–1306.
Ciresan, D. C. and Schmidhuber, J. (2013). Multi-column deep neural networks for offline handwritten Chinese character classification. Technical report, IDSIA. arXiv:1309.0261.
Cliff, D. T., Husbands, P., and Harvey, I. (1993). Evolving recurrent dynamical networks for robot
control. In Artificial Neural Nets and Genetic Algorithms, pages 428–435. Springer.
Clune, J., Mouret, J.-B., and Lipson, H. (2013). The evolutionary origins of modularity. Proceedings
of the Royal Society B: Biological Sciences, 280(1755):20122863.
Clune, J., Stanley, K. O., Pennock, R. T., and Ofria, C. (2011). On the performance of indirect
encoding across the continuum of regularity. Trans. Evol. Comp, 15(3):346–367.
Coates, A., Huval, B., Wang, T., Wu, D. J., Ng, A. Y., and Catanzaro, B. (2013). Deep learning with
COTS HPC systems. In Proc. International Conference on Machine learning (ICML’13).
Cochocki, A. and Unbehauen, R. (1993). Neural networks for optimization and signal processing.
John Wiley & Sons, Inc.
Collobert, R. and Weston, J. (2008). A unified architecture for natural language processing: Deep
neural networks with multitask learning. In Proceedings of the 25th International Conference on
Machine Learning (ICML), pages 160–167. ACM.
Comon, P. (1994). Independent component analysis – a new concept? Signal Processing, 36(3):287–
314.
Connor, C. E., Brincat, S. L., and Pasupathy, A. (2007). Transformation of shape information in the
ventral pathway. Current Opinion in Neurobiology, 17(2):140–147.
Connor, J., Martin, D. R., and Atlas, L. E. (1994). Recurrent neural networks and robust time series
prediction. IEEE Transactions on Neural Networks, 5(2):240–254.
Cook, S. A. (1971). The complexity of theorem-proving procedures. In Proceedings of the 3rd Annual
ACM Symposium on the Theory of Computing (STOC’71), pages 151–158. ACM, New York.
Cramer, N. L. (1985). A representation for the adaptive generation of simple sequential programs.
In Grefenstette, J., editor, Proceedings of an International Conference on Genetic Algorithms and
Their Applications, Carnegie-Mellon University, July 24-26, 1985, Hillsdale NJ. Lawrence Erlbaum Associates.
Craven, P. and Wahba, G. (1979). Smoothing noisy data with spline functions: Estimating the correct
degree of smoothing by the method of generalized cross-validation. Numer. Math., 31:377–403.
Cuccu, G., Luciw, M., Schmidhuber, J., and Gomez, F. (2011). Intrinsically motivated evolutionary
search for vision-based reinforcement learning. In Proceedings of the 2011 IEEE Conference on
Development and Learning and Epigenetic Robotics IEEE-ICDL-EPIROB, volume 2, pages 1–7.
IEEE.
Dahl, G., Yu, D., Deng, L., and Acero, A. (2012). Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition. Audio, Speech, and Language Processing, IEEE
Transactions on, 20(1):30–42.
43
Dahl, G. E., Sainath, T. N., and Hinton, G. E. (2013). Improving deep neural networks for LVCSR
using rectified linear units and dropout. In IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 8609–8613. IEEE.
D’Ambrosio, D. B. and Stanley, K. O. (2007). A novel generative encoding for exploiting neural network sensor and output geometry. In Proceedings of the Conference on Genetic and Evolutionary
Computation (GECCO), pages 974–981.
Datar, M., Immorlica, N., Indyk, P., and Mirrokni, V. S. (2004). Locality-sensitive hashing scheme
based on p-stable distributions. In Proceedings of the 20th Annual Symposium on Computational
Geometry, pages 253–262. ACM.
Dayan, P. and Hinton, G. (1993). Feudal reinforcement learning. In Lippman, D. S., Moody, J. E.,
and Touretzky, D. S., editors, Advances in Neural Information Processing Systems (NIPS) 5, pages
271–278. Morgan Kaufmann.
Dayan, P. and Hinton, G. E. (1996). Varieties of Helmholtz machine. Neural Networks, 9(8):1385–
1403.
Dayan, P., Hinton, G. E., Neal, R. M., and Zemel, R. S. (1995). The Helmholtz machine. Neural
Computation, 7:889–904.
Dayan, P. and Zemel, R. (1995). Competition and multiple cause models. Neural Computation,
7:565–579.
De Freitas, J. F. G. (2003). Bayesian methods for neural networks. PhD thesis, University of Cambridge.
de Souto, M. C., Souto, M. C. P. D., and Oliveira, W. R. D. (1999). The loading problem for pyramidal
neural networks. In Electronic Journal on Mathematics of Computation.
De Valois, R. L., Albrecht, D. G., and Thorell, L. G. (1982). Spatial frequency selectivity of cells in
macaque visual cortex. Vision Research, 22(5):545–559.
de Vries, B. and Principe, J. C. (1991). A theory for neural networks with time delays. In Lippmann,
R. P., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural Information Processing
Systems (NIPS) 3, pages 162–168. Morgan Kaufmann.
Deco, G. and Parra, L. (1997). Non-linear feature extraction by redundancy reduction in an unsupervised stochastic neural network. Neural Networks, 10(4):683–691.
Deco, G. and Rolls, E. T. (2005). Neurodynamics of biased competition and cooperation for attention:
a model with spiking neurons. Journal of Neurophysiology, 94(1):295–313.
DeJong, G. and Mooney, R. (1986). Explanation-based learning: An alternative view. Machine
Learning, 1(2):145–176.
DeMers, D. and Cottrell, G. (1993). Non-linear dimensionality reduction. In Hanson, S. J., Cowan,
J. D., and Giles, C. L., editors, Advances in Neural Information Processing Systems (NIPS) 5, pages
580–587. Morgan Kaufmann.
Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from incomplete data
via the EM algorithm. Journal of the Royal Statistical Society, B, 39.
44
Deng, L. and Yu, D. (2014). Deep Learning: Methods and Applications. NOW Publishers.
Desimone, R., Albright, T. D., Gross, C. G., and Bruce, C. (1984). Stimulus-selective properties of
inferior temporal neurons in the macaque. The Journal of Neuroscience, 4(8):2051–2062.
Deville, Y. and Lau, K. K. (1994). Logic program synthesis. Journal of Logic Programming,
19(20):321–350.
Di Lena, P., Nagata, K., and Baldi, P. (2012). Deep architectures for protein contact map prediction.
Bioinformatics, 28:2449–2457.
DiCarlo, J. J., Zoccolan, D., and Rust, N. C. (2012). How does the brain solve visual object recognition? Neuron, 73(3):415–434.
Dickmanns, D., Schmidhuber, J., and Winklhofer, A. (1987). Der genetische Algorithmus:
Eine Implementierung in Prolog. Technical Report, Inst. of Informatics, Tech. Univ. Munich.
http://www.idsia.ch/˜juergen/geneticprogramming.html.
Dickmanns, E. D., Behringer, R., Dickmanns, D., Hildebrandt, T., Maurer, M., Thomanek, F., and
Schiehlen, J. (1994). The seeing passenger car ’VaMoRs-P’. In Proc. Int. Symp. on Intelligent
Vehicles ’94, Paris, pages 68–73.
Dietterich, T. G. (2000a). Ensemble methods in machine learning. In Multiple classifier systems,
pages 1–15. Springer.
Dietterich, T. G. (2000b). Hierarchical reinforcement learning with the MAXQ value function decomposition. J. Artif. Intell. Res. (JAIR), 13:227–303.
Director, S. W. and Rohrer, R. A. (1969). Automated network design - the frequency-domain case.
IEEE Trans. Circuit Theory, CT-16:330–337.
Dittenbach, M., Merkl, D., and Rauber, A. (2000). The growing hierarchical self-organizing map.
In IEEE-INNS-ENNS International Joint Conference on Neural Networks, volume 6, pages 6015–
6015. IEEE Computer Society.
Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., and Darrell, T. (2013). DeCAF: A deep convolutional activation feature for generic visual recognition. arXiv preprint
arXiv:1310.1531.
Dorffner, G. (1996). Neural networks for time series processing. In Neural Network World.
Doya, K., Samejima, K., ichi Katagiri, K., and Kawato, M. (2002). Multiple model-based reinforcement learning. Neural Computation, 14(6):1347–1369.
Dreyfus, S. E. (1962). The numerical solution of variational problems. Journal of Mathematical
Analysis and Applications, 5(1):30–45.
Dreyfus, S. E. (1973). The computational solution of optimal control problems with time lag. IEEE
Transactions on Automatic Control, 18(4):383–385.
Duchi, J., Hazan, E., and Singer, Y. (2011). Adaptive subgradient methods for online learning and
stochastic optimization. The Journal of Machine Learning, 12:2121–2159.
45
Egorova, A., Gloye, A., Göktekin, C., Liers, A., Luft, M., Rojas, R., Simon, M., Tenchio, O., and
Wiesel, F. (2004). FU-Fighters Small Size 2004, Team Description. RoboCup 2004 Symposium:
Papers and Team Description Papers. CD edition.
Elfwing, S., Otsuka, M., Uchibe, E., and Doya, K. (2010). Free-energy based reinforcement learning
for vision-based navigation with high-dimensional sensory inputs. In Neural Information Processing. Theory and Algorithms (ICONIP), volume 1, pages 215–222. Springer.
Eliasmith, C. (2013). How to build a brain: A neural architecture for biological cognition. Oxford
University Press, New York, NY.
Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., and Rasmussen, D. (2012).
A large-scale model of the functioning brain. Science, 338(6111):1202–1205.
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2):179–211.
Erhan, D., Bengio, Y., Courville, A., Manzagol, P.-A., Vincent, P., and Bengio, S. (2010). Why does
unsupervised pre-training help deep learning? J. Mach. Learn. Res., 11:625–660.
Escalante-B., A. N. and Wiskott, L. (2013). How to solve classification and regression problems on
high-dimensional data with a supervised extension of slow feature analysis. Journal of Machine
Learning Research, 14:3683–3719.
Eubank, R. L. (1988). Spline smoothing and nonparametric regression. In Farlow, S., editor, SelfOrganizing Methods in Modeling. Marcel Dekker, New York.
Euler, L. (1744). Methodus inveniendi.
Eyben, F., Weninger, F., Squartini, S., and Schuller, B. (2013). Real-life voice activity detection with
LSTM recurrent neural networks and an application to Hollywood movies. In Proc. 38th IEEE
International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2013, Vancouver,
Canada, pages 483–487.
Faggin, F. (1992). Neural network hardware. In International Joint Conference on Neural Networks
(IJCNN), volume 1, page 153.
Fahlman, S. E. (1988). An empirical study of learning speed in back-propagation networks. Technical
Report CMU-CS-88-162, Carnegie-Mellon Univ.
Fahlman, S. E. (1991). The recurrent cascade-correlation learning algorithm. In Lippmann, R. P.,
Moody, J. E., and Touretzky, D. S., editors, Advances in Neural Information Processing Systems
(NIPS) 3, pages 190–196. Morgan Kaufmann.
Falconbridge, M. S., Stamps, R. L., and Badcock, D. R. (2006). A simple Hebbian/anti-Hebbian
network learns the sparse, independent components of natural images. Neural Computation,
18(2):415–429.
Fan, Y., Qian, Y., Xie, F., and Soong, F. K. (2014). TTS synthesis with bidirectional LSTM based
recurrent neural networks. In Proc. Interspeech.
Farabet, C., Couprie, C., Najman, L., and LeCun, Y. (2013). Learning hierarchical features for scene
labeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1915–1929.
Farlow, S. J. (1984). Self-organizing methods in modeling: GMDH type algorithms, volume 54. CRC
Press.
46
Feldkamp, L. A., Prokhorov, D. V., Eagen, C. F., and Yuan, F. (1998). Enhanced multi-stream Kalman
filter training for recurrent networks. In Nonlinear Modeling, pages 29–53. Springer.
Feldkamp, L. A., Prokhorov, D. V., and Feldkamp, T. M. (2003). Simple and conditioned adaptive
behavior from Kalman filter trained recurrent networks. Neural Networks, 16(5):683–689.
Feldkamp, L. A. and Puskorius, G. V. (1998). A signal processing framework based on dynamic neural
networks with application to problems in adaptation, filtering, and classification. Proceedings of
the IEEE, 86(11):2259–2277.
Felleman, D. J. and Van Essen, D. C. (1991). Distributed hierarchical processing in the primate
cerebral cortex. Cerebral Cortex, 1(1):1–47.
Fernandez, R., Rendel, A., Ramabhadran, B., and Hoory, R. (2014). Prosody contour prediction with
Long Short-Term Memory, bi-directional, deep recurrent neural networks. In Proc. Interspeech.
Fernández, S., Graves, A., and Schmidhuber, J. (2007). An application of recurrent neural networks
to discriminative keyword spotting. In Proc. ICANN (2), pages 220–229.
Fernandez, S., Graves, A., and Schmidhuber, J. (2007). Sequence labelling in structured domains with
hierarchical recurrent neural networks. In Proceedings of the 20th International Joint Conference
on Artificial Intelligence (IJCAI).
Field, D. J. (1987). Relations between the statistics of natural images and the response properties of
cortical cells. Journal of the Optical Society of America, 4:2379–2394.
Field, D. J. (1994). What is the goal of sensory coding? Neural Computation, 6:559–601.
Fieres, J., Schemmel, J., and Meier, K. (2008). Realizing biological spiking network models in
a configurable wafer-scale hardware system. In IEEE International Joint Conference on Neural
Networks, pages 969–976.
Fine, S., Singer, Y., and Tishby, N. (1998). The hierarchical hidden Markov model: Analysis and
applications. Machine Learning, 32(1):41–62.
Fischer, A. and Igel, C. (2014). Training restricted Boltzmann machines: An introduction. Pattern
Recognition, 47:25–39.
FitzHugh, R. (1961). Impulses and physiological states in theoretical models of nerve membrane.
Biophysical Journal, 1(6):445–466.
Fletcher, R. and Powell, M. J. (1963). A rapidly convergent descent method for minimization. The
Computer Journal, 6(2):163–168.
Floreano, D. and Mattiussi, C. (2001). Evolution of spiking neural controllers for autonomous visionbased robots. In Evolutionary Robotics. From Intelligent Robotics to Artificial Life, pages 38–61.
Springer.
Fogel, D. B., Fogel, L. J., and Porto, V. (1990). Evolving neural networks. Biological Cybernetics,
63(6):487–493.
Fogel, L., Owens, A., and Walsh, M. (1966). Artificial Intelligence through Simulated Evolution.
Wiley, New York.
47
Földiák, P. (1990). Forming sparse representations by local anti-Hebbian learning. Biological Cybernetics, 64:165–170.
Földiák, P. and Young, M. P. (1995). Sparse coding in the primate cortex. In Arbib, M. A., editor, The
Handbook of Brain Theory and Neural Networks, pages 895–898. The MIT Press.
Förster, A., Graves, A., and Schmidhuber, J. (2007). RNN-based Learning of Compact Maps for
Efficient Robot Localization. In 15th European Symposium on Artificial Neural Networks, ESANN,
pages 537–542, Bruges, Belgium.
Franzius, M., Sprekeler, H., and Wiskott, L. (2007). Slowness and sparseness lead to place, headdirection, and spatial-view cells. PLoS Computational Biology, 3(8):166.
Friedman, J., Hastie, T., and Tibshirani, R. (2001). The elements of statistical learning, volume 1.
Springer Series in Statistics, New York.
Frinken, V., Zamora-Martinez, F., Espana-Boquera, S., Castro-Bleda, M. J., Fischer, A., and Bunke,
H. (2012). Long-short term memory neural networks language modeling for handwriting recognition. In Pattern Recognition (ICPR), 2012 21st International Conference on, pages 701–704.
IEEE.
Fritzke, B. (1994). A growing neural gas network learns topologies. In Tesauro, G., Touretzky, D. S.,
and Leen, T. K., editors, NIPS, pages 625–632. MIT Press.
Fu, K. S. (1977). Syntactic Pattern Recognition and Applications. Berlin, Springer.
Fukada, T., Schuster, M., and Sagisaka, Y. (1999). Phoneme boundary estimation using bidirectional
recurrent neural networks and its applications. Systems and Computers in Japan, 30(4):20–30.
Fukushima, K. (1979). Neural network model for a mechanism of pattern recognition unaffected by
shift in position - Neocognitron. Trans. IECE, J62-A(10):658–665.
Fukushima, K. (1980). Neocognitron: A self-organizing neural network for a mechanism of pattern
recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–202.
Fukushima, K. (2011). Increasing robustness against background noise: visual pattern recognition by
a Neocognitron. Neural Networks, 24(7):767–778.
Fukushima, K. (2013a). Artificial vision by multi-layered neural networks: Neocognitron and its
advances. Neural Networks, 37:103–119.
Fukushima, K. (2013b). Training multi-layered neural network Neocognitron. Neural Networks,
40:18–31.
Gabor, D. (1946). Theory of communication. Part 1: The analysis of information. Electrical Engineers-Part III: Journal of the Institution of Radio and Communication Engineering,
93(26):429–441.
Gallant, S. I. (1988). Connectionist expert systems. Communications of the ACM, 31(2):152–169.
Gauss, C. F. (1809). Theoria motus corporum coelestium in sectionibus conicis solem ambientium.
Gauss, C. F. (1821). Theoria combinationis observationum erroribus minimis obnoxiae (Theory of
the combination of observations least subject to error).
48
Ge, S., Hang, C. C., Lee, T. H., and Zhang, T. (2010). Stable adaptive neural network control.
Springer.
Geiger, J. T., Zhang, Z., Weninger, F., Schuller, B., and Rigoll, G. (2014). Robust speech recognition
using long short-term memory recurrent neural networks for hybrid acoustic modelling. In Proc.
Interspeech.
Geman, S., Bienenstock, E., and Doursat, R. (1992). Neural networks and the bias/variance dilemma.
Neural Computation, 4:1–58.
Gers, F. A. and Schmidhuber, J. (2000). Recurrent nets that time and count. In Neural Networks, 2000.
IJCNN 2000, Proceedings of the IEEE-INNS-ENNS International Joint Conference on, volume 3,
pages 189–194. IEEE.
Gers, F. A. and Schmidhuber, J. (2001). LSTM recurrent networks learn simple context free and
context sensitive languages. IEEE Transactions on Neural Networks, 12(6):1333–1340.
Gers, F. A., Schmidhuber, J., and Cummins, F. (2000). Learning to forget: Continual prediction with
LSTM. Neural Computation, 12(10):2451–2471.
Gers, F. A., Schraudolph, N., and Schmidhuber, J. (2002). Learning precise timing with LSTM
recurrent networks. Journal of Machine Learning Research, 3:115–143.
Gerstner, W. and Kistler, W. K. (2002). Spiking Neuron Models. Cambridge University Press.
Gerstner, W. and van Hemmen, J. L. (1992). Associative memory in a network of spiking neurons.
Network: Computation in Neural Systems, 3(2):139–164.
Ghavamzadeh, M. and Mahadevan, S. (2003). Hierarchical policy gradient algorithms. In Proceedings
of the Twentieth Conference on Machine Learning (ICML-2003), pages 226–233.
Gherrity, M. (1989). A learning algorithm for analog fully recurrent neural networks. In IEEE/INNS
International Joint Conference on Neural Networks, San Diego, volume 1, pages 643–644.
Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2013). Rich feature hierarchies for accurate object
detection and semantic segmentation. Technical Report arxiv.org/abs/1311.2524, UC Berkeley and
ICSI.
Gisslen, L., Luciw, M., Graziano, V., and Schmidhuber, J. (2011). Sequential constant size compressor
for reinforcement learning. In Proc. Fourth Conference on Artificial General Intelligence (AGI),
Google, Mountain View, CA, pages 31–40. Springer.
Giusti, A., Ciresan, D. C., Masci, J., Gambardella, L. M., and Schmidhuber, J. (2013). Fast image
scanning with deep max-pooling convolutional neural networks. In Proc. ICIP.
Glackin, B., McGinnity, T. M., Maguire, L. P., Wu, Q., and Belatreche, A. (2005). A novel approach
for the implementation of large scale spiking neural networks on FPGA hardware. In Computational Intelligence and Bioinspired Systems, pages 552–563. Springer.
Glasmachers, T., Schaul, T., Sun, Y., Wierstra, D., and Schmidhuber, J. (2010). Exponential natural
evolution strategies. In Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO), pages 393–400. ACM.
Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse rectifier networks. In AISTATS, volume 15,
pages 315–323.
49
Gloye, A., Wiesel, F., Tenchio, O., and Simon, M. (2005). Reinforcing the driving quality of soccer
playing robots by anticipation. IT - Information Technology, 47(5).
Gödel, K. (1931). Über formal unentscheidbare Sätze der Principia Mathematica und verwandter
Systeme I. Monatshefte für Mathematik und Physik, 38:173–198.
Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. AddisonWesley, Reading, MA.
Goldfarb, D. (1970). A family of variable-metric methods derived by variational means. Mathematics
of computation, 24(109):23–26.
Golub, G., Heath, H., and Wahba, G. (1979). Generalized cross-validation as a method for choosing
a good ridge parameter. Technometrics, 21:215–224.
Gomez, F. J. (2003). Robust Nonlinear Control through Neuroevolution. PhD thesis, Department of
Computer Sciences, University of Texas at Austin.
Gomez, F. J. and Miikkulainen, R. (2003). Active guidance for a finless rocket using neuroevolution.
In Proc. GECCO 2003, Chicago.
Gomez, F. J. and Schmidhuber, J. (2005). Co-evolving recurrent neurons learn deep memory
POMDPs. In Proc. of the 2005 conference on genetic and evolutionary computation (GECCO),
Washington, D. C. ACM Press, New York, NY, USA.
Gomez, F. J., Schmidhuber, J., and Miikkulainen, R. (2008). Accelerated neural evolution through
cooperatively coevolved synapses. Journal of Machine Learning Research, 9(May):937–965.
Gomi, H. and Kawato, M. (1993). Neural network control for a closed-loop system using feedbackerror-learning. Neural Networks, 6(7):933–946.
Gonzalez-Dominguez, J., Lopez-Moreno, I., Sak, H., Gonzalez-Rodriguez, J., and Moreno, P. J.
(2014). Automatic language identification using Long Short-Term Memory recurrent neural networks. In Proc. Interspeech.
Goodfellow, I., Mirza, M., Da, X., Courville, A., and Bengio, Y. (2014a). An Empirical Investigation
of Catastrophic Forgetting in Gradient-Based Neural Networks. TR arXiv:1312.6211v2.
Goodfellow, I. J., Bulatov, Y., Ibarz, J., Arnoud, S., and Shet, V. (2014b). Multi-digit number
recognition from street view imagery using deep convolutional neural networks. arXiv preprint
arXiv:1312.6082 v4.
Goodfellow, I. J., Courville, A., and Bengio, Y. (2011). Spike-and-slab sparse coding for unsupervised
feature discovery. In NIPS Workshop on Challenges in Learning Hierarchical Models.
Goodfellow, I. J., Courville, A. C., and Bengio, Y. (2012). Large-scale feature learning with spikeand-slab sparse coding. In Proceedings of the 29th International Conference on Machine Learning
(ICML).
Goodfellow, I. J., Warde-Farley, D., Mirza, M., Courville, A., and Bengio, Y. (2013). Maxout networks. In International Conference on Machine Learning (ICML).
Graves, A. (2011). Practical variational inference for neural networks. In Advances in Neural Information Processing Systems (NIPS), pages 2348–2356.
50
Graves, A., Eck, D., Beringer, N., and Schmidhuber, J. (2003). Isolated digit recognition with LSTM
recurrent networks. In First International Workshop on Biologically Inspired Approaches to Advanced Information Technology, Lausanne.
Graves, A., Fernandez, S., Gomez, F. J., and Schmidhuber, J. (2006). Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural nets. In ICML’06: Proceedings
of the 23rd International Conference on Machine Learning, pages 369–376.
Graves, A., Fernandez, S., Liwicki, M., Bunke, H., and Schmidhuber, J. (2008). Unconstrained online handwriting recognition with recurrent neural networks. In Platt, J., Koller, D., Singer, Y.,
and Roweis, S., editors, Advances in Neural Information Processing Systems (NIPS) 20, pages
577–584. MIT Press, Cambridge, MA.
Graves, A. and Jaitly, N. (2014). Towards end-to-end speech recognition with recurrent neural networks. In Proc. 31st International Conference on Machine Learning (ICML), pages 1764–1772.
Graves, A., Liwicki, M., Fernandez, S., Bertolami, R., Bunke, H., and Schmidhuber, J. (2009). A
novel connectionist system for improved unconstrained handwriting recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(5).
Graves, A., Mohamed, A.-R., and Hinton, G. E. (2013). Speech recognition with deep recurrent
neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), pages 6645–6649. IEEE.
Graves, A. and Schmidhuber, J. (2005). Framewise phoneme classification with bidirectional LSTM
and other neural network architectures. Neural Networks, 18(5-6):602–610.
Graves, A. and Schmidhuber, J. (2009). Offline handwriting recognition with multidimensional recurrent neural networks. In Advances in Neural Information Processing Systems (NIPS) 21, pages
545–552. MIT Press, Cambridge, MA.
Graziano, M. (2009). The Intelligent Movement Machine: An Ethological Perspective on the Primate
Motor System. Oxford University Press, USA.
Griewank, A. (2012). Documenta Mathematica - Extra Volume ISMP, pages 389–400.
Grondman, I., Busoniu, L., Lopes, G. A. D., and Babuska, R. (2012). A survey of actor-critic reinforcement learning: Standard and natural policy gradients. Systems, Man, and Cybernetics, Part C:
Applications and Reviews, IEEE Transactions on, 42(6):1291–1307.
Grossberg, S. (1969). Some networks that can learn, remember, and reproduce any number of complicated space-time patterns, I. Journal of Mathematics and Mechanics, 19:53–91.
Grossberg, S. (1976a). Adaptive pattern classification and universal recoding, 1: Parallel development
and coding of neural feature detectors. Biological Cybernetics, 23:187–202.
Grossberg, S. (1976b). Adaptive pattern classification and universal recoding, 2: Feedback, expectation, olfaction, and illusions. Biological Cybernetics, 23.
Gruau, F., Whitley, D., and Pyeatt, L. (1996). A comparison between cellular encoding and direct
encoding for genetic neural networks. NeuroCOLT Technical Report NC-TR-96-048, ESPRIT
Working Group in Neural and Computational Learning, NeuroCOLT 8556.
51
Grünwald, P. D., Myung, I. J., and Pitt, M. A. (2005). Advances in minimum description length:
Theory and applications. MIT Press.
Grüttner, M., Sehnke, F., Schaul, T., and Schmidhuber, J. (2010). Multi-Dimensional Deep Memory
Atari-Go Players for Parameter Exploring Policy Gradients. In Proceedings of the International
Conference on Artificial Neural Networks ICANN, pages 114–123. Springer.
Guo, X., Singh, S., Lee, H., Lewis, R., and Wang, X. (2014). Deep learning for real-time Atari game
play using offline Monte-Carlo tree search planning. In Advances in Neural Information Processing
Systems 27 (NIPS).
Guyon, I., Vapnik, V., Boser, B., Bottou, L., and Solla, S. A. (1992). Structural risk minimization for
character recognition. In Lippman, D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in
Neural Information Processing Systems (NIPS) 4, pages 471–479. Morgan Kaufmann.
Hadamard, J. (1908). Mémoire sur le problème d’analyse relatif à l’équilibre des plaques élastiques
encastrées. Mémoires présentés par divers savants à l’Académie des sciences de l’Institut de
France: Éxtrait. Imprimerie nationale.
Hadsell, R., Chopra, S., and LeCun, Y. (2006). Dimensionality reduction by learning an invariant
mapping. In Proc. Computer Vision and Pattern Recognition Conference (CVPR’06). IEEE Press.
Hagras, H., Pounds-Cornish, A., Colley, M., Callaghan, V., and Clarke, G. (2004). Evolving spiking
neural network controllers for autonomous robots. In IEEE International Conference on Robotics
and Automation (ICRA), volume 5, pages 4620–4626.
Hansen, N., Müller, S. D., and Koumoutsakos, P. (2003). Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (CMA-ES). Evolutionary Computation, 11(1):1–18.
Hansen, N. and Ostermeier, A. (2001). Completely derandomized self-adaptation in evolution strategies. Evolutionary Computation, 9(2):159–195.
Hanson, S. J. (1990). A stochastic version of the delta rule. Physica D: Nonlinear Phenomena,
42(1):265–272.
Hanson, S. J. and Pratt, L. Y. (1989). Comparing biases for minimal network construction with
back-propagation. In Touretzky, D. S., editor, Advances in Neural Information Processing Systems
(NIPS) 1, pages 177–185. San Mateo, CA: Morgan Kaufmann.
Happel, B. L. and Murre, J. M. (1994). Design and evolution of modular neural network architectures.
Neural Networks, 7(6):985–1004.
Hashem, S. and Schmeiser, B. (1992). Improving model accuracy using optimal linear combinations
of trained neural networks. IEEE Transactions on Neural Networks, 6:792–794.
Hassibi, B. and Stork, D. G. (1993). Second order derivatives for network pruning: Optimal brain
surgeon. In Lippman, D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural
Information Processing Systems 5, pages 164–171. Morgan Kaufmann.
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The elements of statistical learning. Springer
Series in Statistics.
52
Hastie, T. J. and Tibshirani, R. J. (1990). Generalized additive models. Monographs on Statisics and
Applied Probability, 43.
Hawkins, J. and George, D. (2006). Hierarchical Temporal Memory - Concepts, Theory, and Terminology. Numenta Inc.
Haykin, S. S. (2001). Kalman filtering and neural networks. Wiley Online Library.
Hebb, D. O. (1949). The Organization of Behavior. Wiley, New York.
Hecht-Nielsen, R. (1989). Theory of the backpropagation neural network. In International Joint
Conference on Neural Networks (IJCNN), pages 593–605. IEEE.
Heemskerk, J. N. (1995). Overview of neural hardware. Neurocomputers for Brain-Style Processing.
Design, Implementation and Application.
Heess, N., Silver, D., and Teh, Y. W. (2012). Actor-critic reinforcement learning with energy-based
policies. In Proc. European Workshop on Reinforcement Learning, pages 43–57.
Heidrich-Meisner, V. and Igel, C. (2009). Neuroevolution strategies for episodic reinforcement learning. Journal of Algorithms, 64(4):152–168.
Herrero, J., Valencia, A., and Dopazo, J. (2001). A hierarchical unsupervised growing neural network
for clustering gene expression patterns. Bioinformatics, 17(2):126–136.
Hertz, J., Krogh, A., and Palmer, R. (1991). Introduction to the Theory of Neural Computation.
Addison-Wesley, Redwood City.
Hestenes, M. R. and Stiefel, E. (1952). Methods of conjugate gradients for solving linear systems.
Journal of research of the National Bureau of Standards, 49:409–436.
Hihi, S. E. and Bengio, Y. (1996). Hierarchical recurrent neural networks for long-term dependencies.
In Touretzky, D. S., Mozer, M. C., and Hasselmo, M. E., editors, Advances in Neural Information
Processing Systems 8, pages 493–499. MIT Press.
Hinton, G. and Salakhutdinov, R. (2006). Reducing the dimensionality of data with neural networks.
Science, 313(5786):504–507.
Hinton, G. E. (1989). Connectionist learning procedures. Artificial intelligence, 40(1):185–234.
Hinton, G. E. (2002). Training products of experts by minimizing contrastive divergence. Neural
Comp., 14(8):1771–1800.
Hinton, G. E., Dayan, P., Frey, B. J., and Neal, R. M. (1995). The wake-sleep algorithm for unsupervised neural networks. Science, 268:1158–1160.
Hinton, G. E., Deng, L., Yu, D., Dahl, G. E., Mohamed, A., Jaitly, N., Senior, A., Vanhoucke, V.,
Nguyen, P., Sainath, T. N., and Kingsbury, B. (2012a). Deep neural networks for acoustic modeling
in speech recognition: The shared views of four research groups. IEEE Signal Process. Mag.,
29(6):82–97.
Hinton, G. E. and Ghahramani, Z. (1997). Generative models for discovering sparse distributed
representations. Philosophical Transactions of the Royal Society B, 352:1177–1190.
53
Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A fast learning algorithm for deep belief nets.
Neural Computation, 18(7):1527–1554.
Hinton, G. E. and Sejnowski, T. E. (1986). Learning and relearning in Boltzmann machines. In
Parallel Distributed Processing, volume 1, pages 282–317. MIT Press.
Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. R. (2012b).
Improving neural networks by preventing co-adaptation of feature detectors. Technical Report
arXiv:1207.0580.
Hinton, G. E. and van Camp, D. (1993). Keeping neural networks simple. In Proceedings of the
International Conference on Artificial Neural Networks, Amsterdam, pages 11–18. Springer.
Hochreiter, S. (1991). Untersuchungen zu dynamischen neuronalen Netzen. Diploma thesis, Institut
für Informatik, Lehrstuhl Prof. Brauer, Technische Universität München. Advisor: J. Schmidhuber.
Hochreiter, S., Bengio, Y., Frasconi, P., and Schmidhuber, J. (2001a). Gradient flow in recurrent nets:
the difficulty of learning long-term dependencies. In Kremer, S. C. and Kolen, J. F., editors, A Field
Guide to Dynamical Recurrent Neural Networks. IEEE Press.
Hochreiter, S. and Obermayer, K. (2005). Sequence classification for protein analysis. In Snowbird
Workshop, Snowbird, Utah. Computational and Biological Learning Society.
Hochreiter, S. and Schmidhuber, J. (1996). Bridging long time lags by weight guessing and “Long
Short-Term Memory”. In Silva, F. L., Principe, J. C., and Almeida, L. B., editors, Spatiotemporal
models in biological and artificial systems, pages 65–72. IOS Press, Amsterdam, Netherlands.
Serie: Frontiers in Artificial Intelligence and Applications, Volume 37.
Hochreiter, S. and Schmidhuber, J. (1997a). Flat minima. Neural Computation, 9(1):1–42.
Hochreiter, S. and Schmidhuber, J. (1997b). Long Short-Term Memory. Neural Computation,
9(8):1735–1780. Based on TR FKI-207-95, TUM (1995).
Hochreiter, S. and Schmidhuber, J. (1999). Feature extraction through LOCOCODE. Neural Computation, 11(3):679–714.
Hochreiter, S., Younger, A. S., and Conwell, P. R. (2001b). Learning to learn using gradient descent.
In Lecture Notes on Comp. Sci. 2130, Proc. Intl. Conf. on Artificial Neural Networks (ICANN2001), pages 87–94. Springer: Berlin, Heidelberg.
Hodgkin, A. L. and Huxley, A. F. (1952). A quantitative description of membrane current and its
application to conduction and excitation in nerve. The Journal of Physiology, 117(4):500.
Hoerzer, G. M., Legenstein, R., and Maass, W. (2014). Emergence of complex computational structures from chaotic neural networks through reward-modulated Hebbian learning. Cerebral Cortex,
24:677–690.
Holden, S. B. (1994). On the Theory of Generalization and Self-Structuring in Linearly Weighted
Connectionist Networks. PhD thesis, Cambridge University, Engineering Department.
Holland, J. H. (1975). Adaptation in Natural and Artificial Systems. University of Michigan Press,
Ann Arbor.
Honavar, V. and Uhr, L. (1993). Generative learning structures and processes for generalized connectionist networks. Information Sciences, 70(1):75–108.
54
Honavar, V. and Uhr, L. M. (1988). A network of neuron-like units that learns to perceive by generation as well as reweighting of its links. In Touretzky, D., Hinton, G. E., and Sejnowski, T.,
editors, Proc. of the 1988 Connectionist Models Summer School, pages 472–484, San Mateo. Morgan Kaufman.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational
abilities. Proc. of the National Academy of Sciences, 79:2554–2558.
Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal
approximators. Neural Networks, 2(5):359–366.
Hubel, D. H. and Wiesel, T. (1962). Receptive fields, binocular interaction, and functional architecture
in the cat’s visual cortex. Journal of Physiology (London), 160:106–154.
Hubel, D. H. and Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate
cortex. The Journal of Physiology, 195(1):215–243.
Huffman, D. A. (1952). A method for construction of minimum-redundancy codes. Proceedings IRE,
40:1098–1101.
Hung, C. P., Kreiman, G., Poggio, T., and DiCarlo, J. J. (2005). Fast readout of object identity from
macaque inferior temporal cortex. Science, 310(5749):863–866.
Hutter, M. (2002). The fastest and shortest algorithm for all well-defined problems. International
Journal of Foundations of Computer Science, 13(3):431–443. (On J. Schmidhuber’s SNF grant
20-61847).
Hutter, M. (2005). Universal Artificial Intelligence: Sequential Decisions based on Algorithmic Probability. Springer, Berlin. (On J. Schmidhuber’s SNF grant 20-61847).
Hyvärinen, A., Hoyer, P., and Oja, E. (1999). Sparse code shrinkage: Denoising by maximum likelihood estimation. In Kearns, M., Solla, S. A., and Cohn, D., editors, Advances in Neural Information
Processing Systems (NIPS) 12. MIT Press.
Hyvärinen, A., Karhunen, J., and Oja, E. (2001). Independent component analysis. John Wiley &
Sons.
ICPR 2012 Contest on Mitosis Detection in Breast Cancer Histological Images (2012). IPAL Laboratory and TRIBVN Company and Pitie-Salpetriere Hospital and CIALAB of Ohio State Univ.,
http://ipal.cnrs.fr/ICPR2012/.
Igel, C. (2003). Neuroevolution for reinforcement learning using evolution strategies. In Reynolds, R.,
Abbass, H., Tan, K. C., Mckay, B., Essam, D., and Gedeon, T., editors, Congress on Evolutionary
Computation (CEC 2003), volume 4, pages 2588–2595. IEEE.
Igel, C. and Hüsken, M. (2003). Empirical evaluation of the improved Rprop learning algorithm.
Neurocomputing, 50(C):105–123.
Ikeda, S., Ochiai, M., and Sawaragi, Y. (1976). Sequential GMDH algorithm and its application to
river flow prediction. IEEE Transactions on Systems, Man and Cybernetics, (7):473–479.
Indermuhle, E., Frinken, V., and Bunke, H. (2012). Mode detection in online handwritten documents
using BLSTM neural networks. In Frontiers in Handwriting Recognition (ICFHR), 2012 International Conference on, pages 302–307. IEEE.
55
Indermuhle, E., Frinken, V., Fischer, A., and Bunke, H. (2011). Keyword spotting in online handwritten documents containing text and non-text using BLSTM neural networks. In Document Analysis
and Recognition (ICDAR), 2011 International Conference on, pages 73–77. IEEE.
Indiveri, G., Linares-Barranco, B., Hamilton, T. J., Van Schaik, A., Etienne-Cummings, R., Delbruck,
T., Liu, S.-C., Dudek, P., Häfliger, P., Renaud, S., et al. (2011). Neuromorphic silicon neuron
circuits. Frontiers in Neuroscience, 5(73).
Ivakhnenko, A. G. (1968). The group method of data handling – a rival of the method of stochastic
approximation. Soviet Automatic Control, 13(3):43–55.
Ivakhnenko, A. G. (1971). Polynomial theory of complex systems. IEEE Transactions on Systems,
Man and Cybernetics, (4):364–378.
Ivakhnenko, A. G. (1995). The review of problems solvable by algorithms of the group method of data
handling (GMDH). Pattern Recognition and Image Analysis / Raspoznavaniye Obrazov I Analiz
Izobrazhenii, 5:527–535.
Ivakhnenko, A. G. and Lapa, V. G. (1965). Cybernetic Predicting Devices. CCM Information Corporation.
Ivakhnenko, A. G., Lapa, V. G., and McDonough, R. N. (1967). Cybernetics and forecasting techniques. American Elsevier, NY.
Izhikevich, E. M. et al. (2003). Simple model of spiking neurons. IEEE Transactions on Neural
Networks, 14(6):1569–1572.
Jaakkola, T., Singh, S. P., and Jordan, M. I. (1995). Reinforcement learning algorithm for partially
observable Markov decision problems. In Tesauro, G., Touretzky, D. S., and Leen, T. K., editors,
Advances in Neural Information Processing Systems (NIPS) 7, pages 345–352. MIT Press.
Jackel, L., Boser, B., Graf, H.-P., Denker, J., LeCun, Y., Henderson, D., Matan, O., Howard, R., and
Baird, H. (1990). VLSI implementation of electronic neural networks: and example in character
recognition. In IEEE, editor, IEEE International Conference on Systems, Man, and Cybernetics,
pages 320–322, Los Angeles, CA.
Jacob, C., Lindenmayer, A., and Rozenberg, G. (1994). Genetic L-System Programming. In Parallel
Problem Solving from Nature III, Lecture Notes in Computer Science.
Jacobs, R. A. (1988). Increased rates of convergence through learning rate adaptation. Neural Networks, 1(4):295–307.
Jaeger, H. (2001). The ”echo state” approach to analysing and training recurrent neural networks.
Technical Report GMD Report 148, German National Research Center for Information Technology.
Jaeger, H. (2004). Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless
communication. Science, 304:78–80.
Jain, V. and Seung, S. (2009). Natural image denoising with convolutional networks. In Koller, D.,
Schuurmans, D., Bengio, Y., and Bottou, L., editors, Advances in Neural Information Processing
Systems (NIPS) 21, pages 769–776. Curran Associates, Inc.
Jameson, J. (1991). Delayed reinforcement learning with multiple time scale hierarchical backpropagated adaptive critics. In Neural Networks for Control.
56
Ji, S., Xu, W., Yang, M., and Yu, K. (2013). 3D convolutional neural networks for human action
recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):221–231.
Jim, K., Giles, C. L., and Horne, B. G. (1995). Effects of noise on convergence and generalization
in recurrent networks. In Tesauro, G., Touretzky, D., and Leen, T., editors, Advances in Neural
Information Processing Systems (NIPS) 7, page 649. San Mateo, CA: Morgan Kaufmann.
Jin, X., Lujan, M., Plana, L. A., Davies, S., Temple, S., and Furber, S. B. (2010). Modeling spiking
neural networks on SpiNNaker. Computing in Science & Engineering, 12(5):91–97.
Jodogne, S. R. and Piater, J. H. (2007). Closed-loop learning of visual control policies. J. Artificial
Intelligence Research, 28:349–391.
Jones, J. P. and Palmer, L. A. (1987). An evaluation of the two-dimensional Gabor filter model of
simple receptive fields in cat striate cortex. Journal of Neurophysiology, 58(6):1233–1258.
Jordan, M. I. (1986). Serial order: A parallel distributed processing approach. Technical Report ICS
Report 8604, Institute for Cognitive Science, University of California, San Diego.
Jordan, M. I. (1988). Supervised learning and systems with excess degrees of freedom. Technical
Report COINS TR 88-27, Massachusetts Institute of Technology.
Jordan, M. I. (1997). Serial order: A parallel distributed processing approach. Advances in Psychology, 121:471–495.
Jordan, M. I. and Rumelhart, D. E. (1990). Supervised learning with a distal teacher. Technical Report
Occasional Paper #40, Center for Cog. Sci., Massachusetts Institute of Technology.
Jordan, M. I. and Sejnowski, T. J. (2001). Graphical models: Foundations of neural computation.
MIT Press.
Joseph, R. D. (1961). Contributions to perceptron theory. PhD thesis, Cornell Univ.
Juang, C.-F. (2004). A hybrid of genetic algorithm and particle swarm optimization for recurrent
network design. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on,
34(2):997–1006.
Judd, J. S. (1990). Neural network design and the complexity of learning. Neural network modeling
and connectionism. MIT Press.
Jutten, C. and Herault, J. (1991). Blind separation of sources, part I: An adaptive algorithm based on
neuromimetic architecture. Signal Processing, 24(1):1–10.
Kaelbling, L. P., Littman, M. L., and Cassandra, A. R. (1995). Planning and acting in partially
observable stochastic domains. Technical report, Brown University, Providence RI.
Kaelbling, L. P., Littman, M. L., and Moore, A. W. (1996). Reinforcement learning: a survey. Journal
of AI research, 4:237–285.
Kak, S., Chen, Y., and Wang, L. (2010). Data mining using surface and deep agents based on neural
networks. AMCIS 2010 Proceedings.
Kalinke, Y. and Lehmann, H. (1998). Computation in recurrent neural networks: From counters to
iterated function systems. In Antoniou, G. and Slaney, J., editors, Advanced Topics in Artificial
Intelligence, Proceedings of the 11th Australian Joint Conference on Artificial Intelligence, volume
1502 of LNAI, Berlin, Heidelberg. Springer.
57
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic
Engineering, 82(1):35–45.
Karhunen, J. and Joutsensalo, J. (1995). Generalizations of principal component analysis, optimization problems, and neural networks. Neural Networks, 8(4):549–562.
Karpathy, A., Toderici, G., Shetty, S., Leung, T., Sukthankar, R., and Fei-Fei, L. (2014). Large-scale
video classification with convolutional neural networks. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR).
Kasabov, N. K. (2014). Neucube: A spiking neural network architecture for mapping, learning and
understanding of spatio-temporal brain data. Neural Networks.
Kelley, H. J. (1960). Gradient theory of optimal flight paths. ARS Journal, 30(10):947–954.
Kempter, R., Gerstner, W., and Van Hemmen, J. L. (1999). Hebbian learning and spiking neurons.
Physical Review E, 59(4):4498.
Kerlirzin, P. and Vallet, F. (1993). Robustness in multilayer perceptrons. Neural Computation,
5(1):473–482.
Khan, M. M., Khan, G. M., and Miller, J. F. (2010). Evolution of neural networks using Cartesian
Genetic Programming. In IEEE Congress on Evolutionary Computation (CEC), pages 1–8.
Khan, M. M., Lester, D. R., Plana, L. A., Rast, A., Jin, X., Painkras, E., and Furber, S. B. (2008).
SpiNNaker: mapping neural networks onto a massively-parallel chip multiprocessor. In International Joint Conference on Neural Networks (IJCNN), pages 2849–2856. IEEE.
Khan, S. H., Bennamoun, M., Sohel, F., and Togneri, R. (2014). Automatic feature learning for robust
shadow detection. In IEEE Conference on Computer Vision and Pattern Recognition CVPR.
Kimura, H., Miyazaki, K., and Kobayashi, S. (1997). Reinforcement learning in POMDPs with
function approximation. In ICML, volume 97, pages 152–160.
Kistler, W. M., Gerstner, W., and van Hemmen, J. L. (1997). Reduction of the Hodgkin-Huxley
equations to a single-variable threshold model. Neural Computation, 9(5):1015–1045.
Kitano, H. (1990). Designing neural networks using genetic algorithms with graph generation system.
Complex Systems, 4:461–476.
Klampfl, S. and Maass, W. (2013). Emergence of dynamic memory traces in cortical microcircuit
models through STDP. The Journal of Neuroscience, 33(28):11515–11529.
Klapper-Rybicka, M., Schraudolph, N. N., and Schmidhuber, J. (2001). Unsupervised learning in
LSTM recurrent neural networks. In Lecture Notes on Comp. Sci. 2130, Proc. Intl. Conf. on Artificial Neural Networks (ICANN-2001), pages 684–691. Springer: Berlin, Heidelberg.
Kobatake, E. and Tanaka, K. (1994). Neuronal selectivities to complex object features in the ventral
visual pathway of the macaque cerebral cortex. J. Neurophysiol., 71:856–867.
Kohl, N. and Stone, P. (2004). Policy gradient reinforcement learning for fast quadrupedal locomotion.
In Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International Conference on,
volume 3, pages 2619–2624. IEEE.
58
Kohonen, T. (1972). Correlation matrix memories. Computers, IEEE Transactions on, 100(4):353–
359.
Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological
Cybernetics, 43(1):59–69.
Kohonen, T. (1988). Self-Organization and Associative Memory. Springer, second edition.
Koikkalainen, P. and Oja, E. (1990). Self-organizing hierarchical feature maps. In International Joint
Conference on Neural Networks (IJCNN), pages 279–284. IEEE.
Kolmogorov, A. N. (1965a). On the representation of continuous functions of several variables by superposition of continuous functions of one variable and addition. Doklady Akademii. Nauk USSR,,
114:679–681.
Kolmogorov, A. N. (1965b). Three approaches to the quantitative definition of information. Problems
of Information Transmission, 1:1–11.
Kompella, V. R., Luciw, M. D., and Schmidhuber, J. (2012). Incremental slow feature analysis: Adaptive low-complexity slow feature updating from high-dimensional input streams. Neural Computation, 24(11):2994–3024.
Kondo, T. (1998). GMDH neural network algorithm using the heuristic self-organization method
and its application to the pattern identification problem. In Proceedings of the 37th SICE Annual
Conference SICE’98, pages 1143–1148. IEEE.
Kondo, T. and Ueno, J. (2008). Multi-layered GMDH-type neural network self-selecting optimum
neural network architecture and its application to 3-dimensional medical image recognition of
blood vessels. International Journal of Innovative Computing, Information and Control, 4(1):175–
187.
Kordı́k, P., Náplava, P., Snorek, M., and Genyk-Berezovskyj, M. (2003). Modified GMDH method
and models quality evaluation by visualization. Control Systems and Computers, 2:68–75.
Korkin, M., de Garis, H., Gers, F., and Hemmi, H. (1997). CBM (CAM-Brain Machine) - a hardware
tool which evolves a neural net module in a fraction of a second and runs a million neuron artificial
brain in real time.
Kosko, B. (1990). Unsupervised learning in noise. IEEE Transactions on Neural Networks, 1(1):44–
57.
Koutnı́k, J., Cuccu, G., Schmidhuber, J., and Gomez, F. (July 2013). Evolving large-scale neural
networks for vision-based reinforcement learning. In Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO), pages 1061–1068, Amsterdam. ACM.
Koutnı́k, J., Gomez, F., and Schmidhuber, J. (2010). Evolving neural networks in compressed weight
space. In Proceedings of the 12th Annual Conference on Genetic and Evolutionary Computation,
pages 619–626.
Koutnı́k, J., Greff, K., Gomez, F., and Schmidhuber, J. (2014). A Clockwork RNN. In Proceedings
of the 31th International Conference on Machine Learning (ICML), volume 32, pages 1845–1853.
arXiv:1402.3511 [cs.NE].
59
Koza, J. R. (1992). Genetic Programming – On the Programming of Computers by Means of Natural
Selection. MIT Press.
Kramer, M. (1991). Nonlinear principal component analysis using autoassociative neural networks.
AIChE Journal, 37:233–243.
Kremer, S. C. and Kolen, J. F. (2001). Field guide to dynamical recurrent networks. Wiley-IEEE
Press.
Kriegeskorte, N., Mur, M., Ruff, D. A., Kiani, R., Bodurka, J., Esteky, H., Tanaka, K., and Bandettini,
P. A. (2008). Matching categorical object representations in inferior temporal cortex of man and
monkey. Neuron, 60(6):1126–1141.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NIPS 2012),
page 4.
Krogh, A. and Hertz, J. A. (1992). A simple weight decay can improve generalization. In Lippman,
D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural Information Processing
Systems 4, pages 950–957. Morgan Kaufmann.
Kruger, N., Janssen, P., Kalkan, S., Lappe, M., Leonardis, A., Piater, J., Rodriguez-Sanchez, A., and
Wiskott, L. (2013). Deep hierarchies in the primate visual cortex: What can we learn for computer
vision? IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1847–1871.
Kullback, S. and Leibler, R. A. (1951). On information and sufficiency. The Annals of Mathematical
Statistics, pages 79–86.
Kurzweil, R. (2012). How to Create a Mind: The Secret of Human Thought Revealed.
Lagoudakis, M. G. and Parr, R. (2003). Least-squares policy iteration. JMLR, 4:1107–1149.
Lampinen, J. and Oja, E. (1992). Clustering properties of hierarchical self-organizing maps. Journal
of Mathematical Imaging and Vision, 2(2-3):261–272.
Lang, K., Waibel, A., and Hinton, G. E. (1990). A time-delay neural network architecture for isolated
word recognition. Neural Networks, 3:23–43.
Lange, S. and Riedmiller, M. (2010). Deep auto-encoder neural networks in reinforcement learning.
In Neural Networks (IJCNN), The 2010 International Joint Conference on, pages 1–8.
Lapedes, A. and Farber, R. (1986). A self-optimizing, nonsymmetrical neural net for content addressable memory and pattern recognition. Physica D, 22:247–259.
Laplace, P. (1774). Mémoire sur la probabilité des causes par les évènements.
l’Academie Royale des Sciences Presentés par Divers Savan, 6:621–656.
Mémoires de
Larraanaga, P. and Lozano, J. A. (2001). Estimation of Distribution Algorithms: A New Tool for
Evolutionary Computation. Kluwer Academic Publishers, Norwell, MA, USA.
Le, Q. V., Ranzato, M., Monga, R., Devin, M., Corrado, G., Chen, K., Dean, J., and Ng, A. Y. (2012).
Building high-level features using large scale unsupervised learning. In Proc. ICML’12.
LeCun, Y. (1985). Une procédure d’apprentissage pour réseau à seuil asymétrique. Proceedings of
Cognitiva 85, Paris, pages 599–604.
60
LeCun, Y. (1988). A theoretical framework for back-propagation. In Touretzky, D., Hinton, G.,
and Sejnowski, T., editors, Proceedings of the 1988 Connectionist Models Summer School, pages
21–28, CMU, Pittsburgh, Pa. Morgan Kaufmann.
LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., and Jackel, L. D.
(1989). Back-propagation applied to handwritten zip code recognition. Neural Computation,
1(4):541–551.
LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., and Jackel, L. D.
(1990a). Handwritten digit recognition with a back-propagation network. In Touretzky, D. S.,
editor, Advances in Neural Information Processing Systems 2, pages 396–404. Morgan Kaufmann.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324.
LeCun, Y., Denker, J. S., and Solla, S. A. (1990b). Optimal brain damage. In Touretzky, D. S., editor,
Advances in Neural Information Processing Systems 2, pages 598–605. Morgan Kaufmann.
LeCun, Y., Muller, U., Cosatto, E., and Flepp, B. (2006). Off-road obstacle avoidance through endto-end learning. In Advances in Neural Information Processing Systems (NIPS 2005).
LeCun, Y., Simard, P., and Pearlmutter, B. (1993). Automatic learning rate maximization by on-line
estimation of the Hessian’s eigenvectors. In Hanson, S., Cowan, J., and Giles, L., editors, Advances
in Neural Information Processing Systems (NIPS 1992), volume 5. Morgan Kaufmann Publishers,
San Mateo, CA.
Lee, H., Battle, A., Raina, R., and Ng, A. Y. (2007a). Efficient sparse coding algorithms. In Advances
in Neural Information Processing Systems (NIPS) 19, pages 801–808.
Lee, H., Ekanadham, C., and Ng, A. Y. (2007b). Sparse deep belief net model for visual area V2. In
Advances in Neural Information Processing Systems (NIPS), volume 7, pages 873–880.
Lee, H., Grosse, R., Ranganath, R., and Ng, A. Y. (2009a). Convolutional deep belief networks for
scalable unsupervised learning of hierarchical representations. In Proceedings of the 26th International Conference on Machine Learning (ICML), pages 609–616.
Lee, H., Pham, P. T., Largman, Y., and Ng, A. Y. (2009b). Unsupervised feature learning for audio
classification using convolutional deep belief networks. In Proc. NIPS, volume 9, pages 1096–1104.
Lee, L. (1996). Learning of context-free languages: A survey of the literature. Technical Report
TR-12-96, Center for Research in Computing Technology, Harvard University, Cambridge, Massachusetts.
Lee, S. and Kil, R. M. (1991). A Gaussian potential function network with hierarchically selforganizing learning. Neural Networks, 4(2):207–224.
Legendre, A. M. (1805). Nouvelles méthodes pour la détermination des orbites des cometes. F. Didot.
Legenstein, R., Wilbert, N., and Wiskott, L. (2010). Reinforcement learning on slow features of
high-dimensional input streams. PLoS Computational Biology, 6(8).
Legenstein, R. A. and Maass, W. (2002). Neural circuits for pattern recognition with small total wire
length. Theor. Comput. Sci., 287(1):239–249.
61
Leibniz, G. W. (1676). Memoir using the chain rule (cited in TMME 7:2&3 p 321-332, 2010).
Leibniz, G. W. (1684). Nova methodus pro maximis et minimis, itemque tangentibus, quae nec
fractas, nec irrationales quantitates moratur, et singulare pro illis calculi genus. Acta Eruditorum,
pages 467–473.
Lenat, D. B. (1983). Theory formation by heuristic search. Machine Learning, 21.
Lenat, D. B. and Brown, J. S. (1984). Why AM an EURISKO appear to work. Artificial Intelligence,
23(3):269–294.
Lennie, P. and Movshon, J. A. (2005). Coding of color and form in the geniculostriate visual pathway.
Journal of the Optical Society of America A, 22(10):2013–2033.
Levenberg, K. (1944). A method for the solution of certain problems in least squares. Quarterly of
applied mathematics, 2:164–168.
Levin, A. U., Leen, T. K., and Moody, J. E. (1994). Fast pruning using principal components. In
Advances in Neural Information Processing Systems 6, page 35. Morgan Kaufmann.
Levin, A. U. and Narendra, K. S. (1995). Control of nonlinear dynamical systems using neural
networks. ii. observability, identification, and control. IEEE Transactions on Neural Networks,
7(1):30–42.
Levin, L. A. (1973a). On the notion of a random sequence. Soviet Math. Dokl., 14(5):1413–1416.
Levin, L. A. (1973b). Universal sequential search problems. Problems of Information Transmission,
9(3):265–266.
Lewicki, M. S. and Olshausen, B. A. (1998). Inferring sparse, overcomplete image codes using an
efficient coding framework. In Jordan, M. I., Kearns, M. J., and Solla, S. A., editors, Advances in
Neural Information Processing Systems (NIPS) 10, pages 815–821.
L’Hôpital, G. F. A. (1696). Analyse des infiniment petits, pour l’intelligence des lignes courbes. Paris:
L’Imprimerie Royale.
Li, M. and Vitányi, P. M. B. (1997). An Introduction to Kolmogorov Complexity and its Applications
(2nd edition). Springer.
Li, R., Zhang, W., Suk, H.-I., Wang, L., Li, J., Shen, D., and Ji, S. (2014). Deep learning based
imaging data completion for improved brain disease diagnosis. In Proc. MICCAI. Springer.
Lin, L. (1993). Reinforcement Learning for Robots Using Neural Networks. PhD thesis, Carnegie
Mellon University, Pittsburgh.
Lin, T., Horne, B., Tino, P., and Giles, C. (1996). Learning long-term dependencies in NARX recurrent neural networks. IEEE Transactions on Neural Networks, 7(6):1329–1338.
Lindenmayer, A. (1968). Mathematical models for cellular interaction in development. J. Theoret.
Biology, 18:280–315.
Lindstädt, S. (1993). Comparison of two unsupervised neural network models for redundancy reduction. In Mozer, M. C., Smolensky, P., Touretzky, D. S., Elman, J. L., and Weigend, A. S., editors,
Proc. of the 1993 Connectionist Models Summer School, pages 308–315. Hillsdale, NJ: Erlbaum
Associates.
62
Linnainmaa, S. (1970). The representation of the cumulative rounding error of an algorithm as a
Taylor expansion of the local rounding errors. Master’s thesis, Univ. Helsinki.
Linnainmaa, S. (1976). Taylor expansion of the accumulated rounding error. BIT Numerical Mathematics, 16(2):146–160.
Linsker, R. (1988). Self-organization in a perceptual network. IEEE Computer, 21:105–117.
Littman, M. L., Cassandra, A. R., and Kaelbling, L. P. (1995). Learning policies for partially observable environments: Scaling up. In Prieditis, A. and Russell, S., editors, Machine Learning:
Proceedings of the Twelfth International Conference, pages 362–370. Morgan Kaufmann Publishers, San Francisco, CA.
Liu, S.-C., Kramer, J., Indiveri, G., Delbrück, T., Burg, T., Douglas, R., et al. (2001). Orientationselective aVLSI spiking neurons. Neural Networks, 14(6-7):629–643.
Ljung, L. (1998). System identification. Springer.
Logothetis, N. K., Pauls, J., and Poggio, T. (1995). Shape representation in the inferior temporal
cortex of monkeys. Current Biology, 5(5):552–563.
Loiacono, D., Cardamone, L., and Lanzi, P. L. (2011). Simulated car racing championship competition software manual. Technical report, Dipartimento di Elettronica e Informazione, Politecnico di
Milano, Italy.
Loiacono, D., Lanzi, P. L., Togelius, J., Onieva, E., Pelta, D. A., Butz, M. V., Lönneker, T. D.,
Cardamone, L., Perez, D., Sáez, Y., Preuss, M., and Quadflieg, J. (2009). The 2009 simulated car
racing championship.
Lowe, D. (1999). Object recognition from local scale-invariant features. In The Proceedings of the
Seventh IEEE International Conference on Computer Vision (ICCV), volume 2, pages 1150–1157.
Lowe, D. (2004). Distinctive image features from scale-invariant key-points. Intl. Journal of Computer Vision, 60:91–110.
Luciw, M., Kompella, V. R., Kazerounian, S., and Schmidhuber, J. (2013). An intrinsic value system
for developing multiple invariant representations with incremental slowness learning. Frontiers in
Neurorobotics, 7(9).
Lusci, A., Pollastri, G., and Baldi, P. (2013). Deep architectures and deep learning in chemoinformatics: the prediction of aqueous solubility for drug-like molecules. Journal of Chemical Information
and Modeling, 53(7):1563–1575.
Maas, A. L., Hannun, A. Y., and Ng, A. Y. (2013). Rectifier nonlinearities improve neural network
acoustic models. In International Conference on Machine Learning (ICML).
Maass, W. (1996). Lower bounds for the computational power of networks of spiking neurons. Neural
Computation, 8(1):1–40.
Maass, W. (1997). Networks of spiking neurons: the third generation of neural network models.
Neural Networks, 10(9):1659–1671.
Maass, W. (2000). On the computational power of winner-take-all. Neural Computation, 12:2519–
2535.
63
Maass, W., Natschläger, T., and Markram, H. (2002). Real-time computing without stable states: A
new framework for neural computation based on perturbations. Neural Computation, 14(11):2531–
2560.
MacKay, D. J. C. (1992). A practical Bayesian framework for backprop networks. Neural Computation, 4:448–472.
MacKay, D. J. C. and Miller, K. D. (1990). Analysis of Linsker’s simulation of Hebbian rules. Neural
Computation, 2:173–187.
Maclin, R. and Shavlik, J. W. (1993). Using knowledge-based neural networks to improve algorithms:
Refining the Chou-Fasman algorithm for protein folding. Machine Learning, 11(2-3):195–215.
Maclin, R. and Shavlik, J. W. (1995). Combining the predictions of multiple classifiers: Using competitive learning to initialize neural networks. In Proc. IJCAI, pages 524–531.
Madala, H. R. and Ivakhnenko, A. G. (1994). Inductive learning algorithms for complex systems
modeling. CRC Press, Boca Raton.
Madani, O., Hanks, S., and Condon, A. (2003). On the undecidability of probabilistic planning and
related stochastic optimization problems. Artificial Intelligence, 147(1):5–34.
Maei, H. R. and Sutton, R. S. (2010). GQ(λ): A general gradient algorithm for temporal-difference
prediction learning with eligibility traces. In Proceedings of the Third Conference on Artificial
General Intelligence, volume 1, pages 91–96.
Maex, R. and Orban, G. (1996). Model circuit of spiking neurons generating directional selectivity in
simple cells. Journal of Neurophysiology, 75(4):1515–1545.
Mahadevan, S. (1996). Average reward reinforcement learning: Foundations, algorithms, and empirical results. Machine Learning, 22:159.
Malik, J. and Perona, P. (1990). Preattentive texture discrimination with early vision mechanisms.
Journal of the Optical Society of America A, 7(5):923–932.
Maniezzo, V. (1994). Genetic evolution of the topology and weight distribution of neural networks.
IEEE Transactions on Neural Networks, 5(1):39–53.
Manolios, P. and Fanelli, R. (1994). First-order recurrent neural networks and deterministic finite
state automata. Neural Computation, 6:1155–1173.
Marchi, E., Ferroni, G., Eyben, F., Gabrielli, L., Squartini, S., and Schuller, B. (2014). Multiresolution linear prediction based features for audio onset detection with bidirectional LSTM neural
networks. In Proc. 39th IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2014, Florence, Italy, pages 2183–2187.
Markram, H. (2012). The human brain project. Scientific American, 306(6):50–55.
Marquardt, D. W. (1963). An algorithm for least-squares estimation of nonlinear parameters. Journal
of the Society for Industrial & Applied Mathematics, 11(2):431–441.
Martens, J. (2010). Deep learning via Hessian-free optimization. In Fürnkranz, J. and Joachims, T.,
editors, Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages
735–742, Haifa, Israel. Omnipress.
64
Martens, J. and Sutskever, I. (2011). Learning recurrent neural networks with Hessian-free optimization. In Proceedings of the 28th International Conference on Machine Learning (ICML), pages
1033–1040.
Martinetz, T. M., Ritter, H. J., and Schulten, K. J. (1990). Three-dimensional neural net for learning
visuomotor coordination of a robot arm. IEEE Transactions on Neural Networks, 1(1):131–136.
Masci, J., Giusti, A., Ciresan, D. C., Fricout, G., and Schmidhuber, J. (2013). A fast learning algorithm
for image segmentation with max-pooling convolutional networks. In International Conference on
Image Processing (ICIP13), pages 2713–2717.
Matsuoka, K. (1992). Noise injection into inputs in back-propagation learning. IEEE Transactions
on Systems, Man, and Cybernetics, 22(3):436–440.
Mayer, H., Gomez, F., Wierstra, D., Nagy, I., Knoll, A., and Schmidhuber, J. (2008). A system for
robotic heart surgery that learns to tie knots using recurrent neural networks. Advanced Robotics,
22(13-14):1521–1537.
McCallum, R. A. (1996). Learning to use selective attention and short-term memory in sequential
tasks. In Maes, P., Mataric, M., Meyer, J.-A., Pollack, J., and Wilson, S. W., editors, From Animals to Animats 4: Proceedings of the Fourth International Conference on Simulation of Adaptive
Behavior, Cambridge, MA, pages 315–324. MIT Press, Bradford Books.
McCulloch, W. and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity.
Bulletin of Mathematical Biophysics, 7:115–133.
Melnik, O., Levy, S. D., and Pollack, J. B. (2000). RAAM for infinite context-free languages. In
Proc. IJCNN (5), pages 585–590.
Memisevic, R. and Hinton, G. E. (2010). Learning to represent spatial transformations with factored
higher-order Boltzmann machines. Neural Computation, 22(6):1473–1492.
Menache, I., Mannor, S., and Shimkin, N. (2002). Q-cut – dynamic discovery of sub-goals in reinforcement learning. In Proc. ECML’02, pages 295–306.
Merolla, P. A., Arthur, J. V., Alvarez-Icaza, R., Cassidy, A. S., Sawada, J., Akopyan, F., Jackson,
B. L., Imam, N., Guo, C., Nakamura, Y., Brezzo, B., Vo, I., Esser, S. K., Appuswamy, R., Taba,
B., Amir, A., Flickner, M. D., Risk, W. P., Manohar, R., and Modha, D. S. (2014). A million
spiking-neuron integrated circuit with a scalable communication network and interface. Science,
345(6197):668–673.
Mesnil, G., Dauphin, Y., Glorot, X., Rifai, S., Bengio, Y., Goodfellow, I., Lavoie, E., Muller, X.,
Desjardins, G., Warde-Farley, D., Vincent, P., Courville, A., and Bergstra, J. (2011). Unsupervised
and transfer learning challenge: a deep learning approach. In JMLR W&CP: Proc. Unsupervised
and Transfer Learning, volume 7.
Meuleau, N., Peshkin, L., Kim, K. E., and Kaelbling, L. P. (1999). Learning finite state controllers for
partially observable environments. In 15th International Conference of Uncertainty in AI, pages
427–436.
Miglino, O., Lund, H., and Nolfi, S. (1995). Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417–434.
65
Miller, G., Todd, P., and Hedge, S. (1989). Designing neural networks using genetic algorithms. In
Proceedings of the 3rd International Conference on Genetic Algorithms, pages 379–384. Morgan
Kauffman.
Miller, J. F. and Harding, S. L. (2009). Cartesian genetic programming. In Proceedings of the 11th Annual Conference Companion on Genetic and Evolutionary Computation Conference: Late Breaking Papers, pages 3489–3512. ACM.
Miller, J. F. and Thomson, P. (2000). Cartesian genetic programming. In Genetic Programming, pages
121–132. Springer.
Miller, K. D. (1994). A model for the development of simple cell receptive fields and the ordered
arrangement of orientation columns through activity-dependent competition between on- and offcenter inputs. Journal of Neuroscience, 14(1):409–441.
Miller, W. T., Werbos, P. J., and Sutton, R. S. (1995). Neural networks for control. MIT Press.
Minai, A. A. and Williams, R. D. (1994). Perturbation response in feedforward networks. Neural
Networks, 7(5):783–796.
Minsky, M. (1963). Steps toward artificial intelligence. In Feigenbaum, E. and Feldman, J., editors,
Computers and Thought, pages 406–450. McGraw-Hill, New York.
Minsky, M. and Papert, S. (1969). Perceptrons. Cambridge, MA: MIT Press.
Minton, S., Carbonell, J. G., Knoblock, C. A., Kuokka, D. R., Etzioni, O., and Gil, Y. (1989).
Explanation-based learning: A problem solving perspective. Artificial Intelligence, 40(1):63–118.
Mitchell, T. (1997). Machine Learning. McGraw Hill.
Mitchell, T. M., Keller, R. M., and Kedar-Cabelli, S. T. (1986). Explanation-based generalization: A
unifying view. Machine Learning, 1(1):47–80.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., and Riedmiller, M.
(Dec 2013). Playing Atari with deep reinforcement learning. Technical Report arXiv:1312.5602
[cs.LG], Deepmind Technologies.
Mohamed, A. and Hinton, G. E. (2010). Phone recognition using restricted Boltzmann machines.
In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
4354–4357.
Molgedey, L. and Schuster, H. G. (1994). Separation of independent signals using time-delayed
correlations. Phys. Reviews Letters, 72(23):3634–3637.
Møller, M. F. (1993). Exact calculation of the product of the Hessian matrix of feed-forward network
error functions and a vector in O(N) time. Technical Report PB-432, Computer Science Department, Aarhus University, Denmark.
Montana, D. J. and Davis, L. (1989). Training feedforward neural networks using genetic algorithms.
In Proceedings of the 11th International Joint Conference on Artificial Intelligence (IJCAI) - Volume 1, IJCAI’89, pages 762–767, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc.
Montavon, G., Orr, G., and Müller, K. (2012). Neural Networks: Tricks of the Trade. Number LNCS
7700 in Lecture Notes in Computer Science Series. Springer Verlag.
66
Moody, J. E. (1989). Fast learning in multi-resolution hierarchies. In Touretzky, D. S., editor, Advances in Neural Information Processing Systems (NIPS) 1, pages 29–39. Morgan Kaufmann.
Moody, J. E. (1992). The effective number of parameters: An analysis of generalization and regularization in nonlinear learning systems. In Lippman, D. S., Moody, J. E., and Touretzky, D. S.,
editors, Advances in Neural Information Processing Systems (NIPS) 4, pages 847–854. Morgan
Kaufmann.
Moody, J. E. and Utans, J. (1994). Architecture selection strategies for neural networks: Application
to corporate bond rating prediction. In Refenes, A. N., editor, Neural Networks in the Capital
Markets. John Wiley & Sons.
Moore, A. and Atkeson, C. (1995). The parti-game algorithm for variable resolution reinforcement
learning in multidimensional state-spaces. Machine Learning, 21(3):199–233.
Moore, A. and Atkeson, C. G. (1993). Prioritized sweeping: Reinforcement learning with less data
and less time. Machine Learning, 13:103–130.
Moriarty, D. E. (1997). Symbiotic Evolution of Neural Networks in Sequential Decision Tasks. PhD
thesis, Department of Computer Sciences, The University of Texas at Austin.
Moriarty, D. E. and Miikkulainen, R. (1996). Efficient reinforcement learning through symbiotic
evolution. Machine Learning, 22:11–32.
Morimoto, J. and Doya, K. (2000). Robust reinforcement learning. In Leen, T. K., Dietterich, T. G.,
and Tresp, V., editors, Advances in Neural Information Processing Systems (NIPS) 13, pages 1061–
1067. MIT Press.
Mosteller, F. and Tukey, J. W. (1968). Data analysis, including statistics. In Lindzey, G. and Aronson,
E., editors, Handbook of Social Psychology, Vol. 2. Addison-Wesley.
Mozer, M. C. (1989). A focused back-propagation algorithm for temporal sequence recognition.
Complex Systems, 3:349–381.
Mozer, M. C. (1991). Discovering discrete distributed representations with iterative competitive learning. In Lippmann, R. P., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural Information
Processing Systems 3, pages 627–634. Morgan Kaufmann.
Mozer, M. C. (1992). Induction of multiscale temporal structure. In Lippman, D. S., Moody, J. E.,
and Touretzky, D. S., editors, Advances in Neural Information Processing Systems (NIPS) 4, pages
275–282. Morgan Kaufmann.
Mozer, M. C. and Smolensky, P. (1989). Skeletonization: A technique for trimming the fat from a
network via relevance assessment. In Touretzky, D. S., editor, Advances in Neural Information
Processing Systems (NIPS) 1, pages 107–115. Morgan Kaufmann.
Muller, U. A., Gunzinger, A., and Guggenbühl, W. (1995). Fast neural net simulation with a DSP
processor array. IEEE Transactions on Neural Networks, 6(1):203–213.
Munro, P. W. (1987). A dual back-propagation scheme for scalar reinforcement learning. Proceedings
of the Ninth Annual Conference of the Cognitive Science Society, Seattle, WA, pages 165–176.
67
Murray, A. F. and Edwards, P. J. (1993). Synaptic weight noise during MLP learning enhances faulttolerance, generalisation and learning trajectory. In S. J. Hanson, J. D. C. and Giles, C. L., editors,
Advances in Neural Information Processing Systems (NIPS) 5, pages 491–498. San Mateo, CA:
Morgan Kaufmann.
Nadal, J.-P. and Parga, N. (1994). Non-linear neurons in the low noise limit: a factorial code maximises information transfer. Network, 5:565–581.
Nagumo, J., Arimoto, S., and Yoshizawa, S. (1962). An active pulse transmission line simulating
nerve axon. Proceedings of the IRE, 50(10):2061–2070.
Nair, V. and Hinton, G. E. (2010). Rectified linear units improve restricted Boltzmann machines. In
International Conference on Machine Learning (ICML).
Narendra, K. S. and Parthasarathy, K. (1990). Identification and control of dynamical systems using
neural networks. Neural Networks, IEEE Transactions on, 1(1):4–27.
Narendra, K. S. and Thathatchar, M. A. L. (1974). Learning automata – a survey. IEEE Transactions
on Systems, Man, and Cybernetics, 4:323–334.
Neal, R. M. (1995). Bayesian learning for neural networks. PhD thesis, University of Toronto.
Neal, R. M. (2006). Classification with Bayesian neural networks. In Quinonero-Candela, J., Magnini,
B., Dagan, I., and D’Alche-Buc, F., editors, Machine Learning Challenges. Evaluating Predictive
Uncertainty, Visual Object Classification, and Recognising Textual Entailment, volume 3944 of
Lecture Notes in Computer Science, pages 28–32. Springer.
Neal, R. M. and Zhang, J. (2006). High dimensional classification with Bayesian neural networks and
Dirichlet diffusion trees. In Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. A., editors, Feature
Extraction: Foundations and Applications, Studies in Fuzziness and Soft Computing, pages 265–
295. Springer.
Neftci, E., Das, S., Pedroni, B., Kreutz-Delgado, K., and Cauwenberghs, G. (2014). Event-driven
contrastive divergence for spiking neuromorphic systems. Frontiers in Neuroscience, 7(272).
Neil, D. and Liu, S.-C. (2014). Minitaur, an event-driven FPGA-based spiking network accelerator.
IEEE Transactions on Very Large Scale Integration (VLSI) Systems, PP(99):1–8.
Nessler, B., Pfeiffer, M., Buesing, L., and Maass, W. (2013). Bayesian computation emerges in
generic cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology, 9(4):e1003037.
Neti, C., Schneider, M. H., and Young, E. D. (1992). Maximally fault tolerant neural networks. In
IEEE Transactions on Neural Networks, volume 3, pages 14–23.
Neuneier, R. and Zimmermann, H.-G. (1996). How to train neural networks. In Orr, G. B. and Müller,
K.-R., editors, Neural Networks: Tricks of the Trade, volume 1524 of Lecture Notes in Computer
Science, pages 373–423. Springer.
Newton, I. (1687). Philosophiae naturalis principia mathematica. William Dawson & Sons Ltd.,
London.
Nguyen, N. and Widrow, B. (1989). The truck backer-upper: An example of self learning in neural
networks. In Proceedings of the International Joint Conference on Neural Networks, pages 357–
363. IEEE Press.
68
Nilsson, N. J. (1980). Principles of artificial intelligence. Morgan Kaufmann, San Francisco, CA,
USA.
Nolfi, S., Floreano, D., Miglino, O., and Mondada, F. (1994a). How to evolve autonomous robots:
Different approaches in evolutionary robotics. In Brooks, R. A. and Maes, P., editors, Fourth
International Workshop on the Synthesis and Simulation of Living Systems (Artificial Life IV), pages
190–197. MIT.
Nolfi, S., Parisi, D., and Elman, J. L. (1994b). Learning and evolution in neural networks. Adaptive
Behavior, 3(1):5–28.
Nowak, E., Jurie, F., and Triggs, B. (2006). Sampling strategies for bag-of-features image classification. In Proc. ECCV 2006, pages 490–503. Springer.
Nowlan, S. J. and Hinton, G. E. (1992). Simplifying neural networks by soft weight sharing. Neural
Computation, 4:173–193.
O’Connor, P., Neil, D., Liu, S.-C., Delbruck, T., and Pfeiffer, M. (2013). Real-time classification and
sensor fusion with a spiking deep belief network. Frontiers in Neuroscience, 7(178).
Oh, K.-S. and Jung, K. (2004). GPU implementation of neural networks. Pattern Recognition,
37(6):1311–1314.
Oja, E. (1989). Neural networks, principal components, and subspaces. International Journal of
Neural Systems, 1(1):61–68.
Oja, E. (1991). Data compression, feature extraction, and autoassociation in feedforward neural
networks. In Kohonen, T., Mäkisara, K., Simula, O., and Kangas, J., editors, Artificial Neural
Networks, volume 1, pages 737–745. Elsevier Science Publishers B.V., North-Holland.
Olshausen, B. A. and Field, D. J. (1996). Emergence of simple-cell receptive field properties by
learning a sparse code for natural images. Nature, 381(6583):607–609.
Omlin, C. and Giles, C. L. (1996). Extraction of rules from discrete-time recurrent neural networks.
Neural Networks, 9(1):41–52.
Oquab, M., Bottou, L., Laptev, I., and Sivic, J. (2013). Learning and transferring mid-level image
representations using convolutional neural networks. Technical Report hal-00911179.
O’Reilly, R. (2003). Making working memory work: A computational model of learning in the
prefrontal cortex and basal ganglia. Technical Report ICS-03-03, ICS.
O’Reilly, R. C. (1996). Biologically plausible error-driven learning using local activation differences:
The generalized recirculation algorithm. Neural Computation, 8(5):895–938.
Orr, G. and Müller, K. (1998). Neural Networks: Tricks of the Trade. Number LNCS 1524 in Lecture
Notes in Computer Science Series. Springer Verlag.
Ostrovskii, G. M., Volin, Y. M., and Borisov, W. W. (1971). Über die Berechnung von Ableitungen.
Wiss. Z. Tech. Hochschule für Chemie, 13:382–384.
Otsuka, M. (2010). Goal-Oriented Representation of the External World: A Free-Energy-Based Approach. PhD thesis, Nara Institute of Science and Technology.
69
Otsuka, M., Yoshimoto, J., and Doya, K. (2010). Free-energy-based reinforcement learning in a
partially observable environment. In Proc. ESANN.
Otte, S., Krechel, D., Liwicki, M., and Dengel, A. (2012). Local feature based online mode detection
with recurrent neural networks. In Proceedings of the 2012 International Conference on Frontiers
in Handwriting Recognition, pages 533–537. IEEE Computer Society.
Oudeyer, P.-Y., Baranes, A., and Kaplan, F. (2013). Intrinsically motivated learning of real world
sensorimotor skills with developmental constraints. In Baldassarre, G. and Mirolli, M., editors,
Intrinsically Motivated Learning in Natural and Artificial Systems. Springer.
OReilly, R. C., Wyatte, D., Herd, S., Mingus, B., and Jilk, D. J. (2013). Recurrent processing during
object recognition. Frontiers in Psychology, 4:124.
Pachitariu, M. and Sahani, M. (2013). Regularization and nonlinearities for neural language models:
when are they needed? arXiv preprint arXiv:1301.5650.
Palm, G. (1980). On associative memory. Biological Cybernetics, 36.
Palm, G. (1992). On the information storage capacity of local learning rules. Neural Computation,
4(2):703–711.
Pan, S. J. and Yang, Q. (2010). A survey on transfer learning. IEEE Transactions on Knowledge and
Data Engineering, 22(10):1345–1359.
Parekh, R., Yang, J., and Honavar, V. (2000). Constructive neural network learning algorithms for
multi-category pattern classification. IEEE Transactions on Neural Networks, 11(2):436–451.
Parker, D. B. (1985). Learning-logic. Technical Report TR-47, Center for Comp. Research in Economics and Management Sci., MIT.
Pascanu, R., Gulcehre, C., Cho, K., and Bengio, Y. (2013a). How to construct deep recurrent neural
networks. arXiv preprint arXiv:1312.6026.
Pascanu, R., Mikolov, T., and Bengio, Y. (2013b). On the difficulty of training recurrent neural
networks. In ICML’13: JMLR: W&CP volume 28.
Pasemann, F., Steinmetz, U., and Dieckman, U. (1999). Evolving structure and function of neurocontrollers. In Angeline, P. J., Michalewicz, Z., Schoenauer, M., Yao, X., and Zalzala, A., editors, Proceedings of the Congress on Evolutionary Computation, volume 3, pages 1973–1978, Mayflower
Hotel, Washington D.C., USA. IEEE Press.
Pearlmutter, B. A. (1989). Learning state space trajectories in recurrent neural networks. Neural
Computation, 1(2):263–269.
Pearlmutter, B. A. (1994). Fast exact multiplication by the Hessian. Neural Computation, 6(1):147–
160.
Pearlmutter, B. A. (1995). Gradient calculations for dynamic recurrent neural networks: A survey.
IEEE Transactions on Neural Networks, 6(5):1212–1228.
Pearlmutter, B. A. and Hinton, G. E. (1986). G-maximization: An unsupervised learning procedure
for discovering regularities. In Denker, J. S., editor, Neural Networks for Computing: American
Institute of Physics Conference Proceedings 151, volume 2, pages 333–338.
70
Peng, J. and Williams, R. J. (1996). Incremental multi-step Q-learning. Machine Learning, 22:283–
290.
Pérez-Ortiz, J. A., Gers, F. A., Eck, D., and Schmidhuber, J. (2003). Kalman filters improve LSTM
network performance in problems unsolvable by traditional recurrent nets. Neural Networks,
(16):241–250.
Perrett, D., Hietanen, J., Oram, M., Benson, P., and Rolls, E. (1992). Organization and functions of
cells responsive to faces in the temporal cortex [and discussion]. Philosophical Transactions of the
Royal Society of London. Series B: Biological Sciences, 335(1273):23–30.
Perrett, D., Rolls, E., and Caan, W. (1982). Visual neurones responsive to faces in the monkey
temporal cortex. Experimental Brain Research, 47(3):329–342.
Peters, J. (2010). Policy gradient methods. Scholarpedia, 5(11):3698.
Peters, J. and Schaal, S. (2008a). Natural actor-critic. Neurocomputing, 71:1180–1190.
Peters, J. and Schaal, S. (2008b). Reinforcement learning of motor skills with policy gradients. Neural
Network, 21(4):682–697.
Pham, V., Kermorvant, C., and Louradour, J. (2013). Dropout Improves Recurrent Neural Networks
for Handwriting Recognition. arXiv preprint arXiv:1312.4569.
Pineda, F. J. (1987). Generalization of back-propagation to recurrent neural networks. Physical
Review Letters, 19(59):2229–2232.
Plate, T. A. (1993). Holographic recurrent networks. In S. J. Hanson, J. D. C. and Giles, C. L., editors,
Advances in Neural Information Processing Systems (NIPS) 5, pages 34–41. Morgan Kaufmann.
Plumbley, M. D. (1991). On information theory and unsupervised neural networks. Dissertation,
published as technical report CUED/F-INFENG/TR.78, Engineering Department, Cambridge University.
Pollack, J. B. (1988). Implications of recursive distributed representations. In Proc. NIPS, pages
527–536.
Pollack, J. B. (1990). Recursive distributed representation. Artificial Intelligence, 46:77–105.
Pontryagin, L. S., Boltyanskii, V. G., Gamrelidze, R. V., and Mishchenko, E. F. (1961). The Mathematical Theory of Optimal Processes.
Poon, H. and Domingos, P. (2011). Sum-product networks: A new deep architecture. In IEEE
International Conference on Computer Vision (ICCV) Workshops, pages 689–690. IEEE.
Post, E. L. (1936). Finite combinatory processes-formulation 1. The Journal of Symbolic Logic,
1(3):103–105.
Prasoon, A., Petersen, K., Igel, C., Lauze, F., Dam, E., and Nielsen, M. (2013). Voxel classification
based on triplanar convolutional neural networks applied to cartilage segmentation in knee MRI. In
Medical Image Computing and Computer Assisted Intervention (MICCAI), volume 8150 of LNCS,
pages 246–253. Springer.
Precup, D., Sutton, R. S., and Singh, S. (1998). Multi-time models for temporally abstract planning. In
Advances in Neural Information Processing Systems (NIPS), pages 1050–1056. Morgan Kaufmann.
71
Prokhorov, D. (2010). A convolutional learning system for object classification in 3-D LIDAR data.
IEEE Transactions on Neural Networks, 21(5):858–863.
Prokhorov, D., Puskorius, G., and Feldkamp, L. (2001). Dynamical neural networks for control. In
Kolen, J. and Kremer, S., editors, A field guide to dynamical recurrent networks, pages 23–78.
IEEE Press.
Prokhorov, D. and Wunsch, D. (1997). Adaptive critic design. IEEE Transactions on Neural Networks, 8(5):997–1007.
Prokhorov, D. V., Feldkamp, L. A., and Tyukin, I. Y. (2002). Adaptive behavior with fixed weights in
RNN: an overview. In Proceedings of the IEEE International Joint Conference on Neural Networks
(IJCNN), pages 2018–2023.
Puskorius, G. V. and Feldkamp, L. A. (1994). Neurocontrol of nonlinear dynamical systems with
Kalman filter trained recurrent networks. IEEE Transactions on Neural Networks, 5(2):279–297.
Raiko, T., Valpola, H., and LeCun, Y. (2012). Deep learning made easier by linear transformations in
perceptrons. In International Conference on Artificial Intelligence and Statistics, pages 924–932.
Raina, R., Madhavan, A., and Ng, A. (2009). Large-scale deep unsupervised learning using graphics
processors. In Proceedings of the 26th Annual International Conference on Machine Learning
(ICML), pages 873–880. ACM.
Ramacher, U., Raab, W., Anlauf, J., Hachmann, U., Beichter, J., Bruels, N., Wesseling, M., Sicheneder, E., Maenner, R., Glaess, J., and Wurz, A. (1993). Multiprocessor and memory architecture
of the neurocomputer SYNAPSE-1. International Journal of Neural Systems, 4(4):333–336.
Ranzato, M., Poultney, C., Chopra, S., and LeCun, Y. (2006). Efficient learning of sparse representations with an energy-based model. In et al., J. P., editor, Advances in Neural Information Processing
Systems (NIPS 2006). MIT Press.
Ranzato, M. A., Huang, F., Boureau, Y., and LeCun, Y. (2007). Unsupervised learning of invariant
feature hierarchies with applications to object recognition. In Proc. Computer Vision and Pattern
Recognition Conference (CVPR’07), pages 1–8. IEEE Press.
Rauber, A., Merkl, D., and Dittenbach, M. (2002). The growing hierarchical self-organizing map: exploratory analysis of high-dimensional data. IEEE Transactions on Neural Networks, 13(6):1331–
1341.
Razavian, A. S., Azizpour, H., Sullivan, J., and Carlsson, S. (2014). CNN features off-the-shelf: an
astounding baseline for recognition. arXiv preprint arXiv:1403.6382.
Rechenberg, I. (1971). Evolutionsstrategie - Optimierung technischer Systeme nach Prinzipien der
biologischen Evolution. Dissertation. Published 1973 by Fromman-Holzboog.
Redlich, A. N. (1993). Redundancy reduction as a strategy for unsupervised learning. Neural Computation, 5:289–304.
Refenes, N. A., Zapranis, A., and Francis, G. (1994). Stock performance modeling using neural
networks: a comparative study with regression models. Neural Networks, 7(2):375–388.
Rezende, D. J. and Gerstner, W. (2014). Stochastic variational learning in recurrent spiking networks.
Frontiers in Computational Neuroscience, 8:38.
72
Riedmiller, M. (2005). Neural fitted Q iteration—first experiences with a data efficient neural reinforcement learning method. In Proc. ECML-2005, pages 317–328. Springer-Verlag Berlin Heidelberg.
Riedmiller, M. and Braun, H. (1993). A direct adaptive method for faster backpropagation learning:
The Rprop algorithm. In Proc. IJCNN, pages 586–591. IEEE Press.
Riedmiller, M., Lange, S., and Voigtlaender, A. (2012). Autonomous reinforcement learning on raw
visual input data in a real world application. In International Joint Conference on Neural Networks
(IJCNN), pages 1–8, Brisbane, Australia.
Riesenhuber, M. and Poggio, T. (1999). Hierarchical models of object recognition in cortex. Nat.
Neurosci., 2(11):1019–1025.
Rifai, S., Vincent, P., Muller, X., Glorot, X., and Bengio, Y. (2011). Contractive auto-encoders:
Explicit invariance during feature extraction. In Proceedings of the 28th International Conference
on Machine Learning (ICML-11), pages 833–840.
Ring, M., Schaul, T., and Schmidhuber, J. (2011). The two-dimensional organization of behavior. In
Proceedings of the First Joint Conference on Development Learning and on Epigenetic Robotics
ICDL-EPIROB, Frankfurt.
Ring, M. B. (1991). Incremental development of complex behaviors through automatic construction of sensory-motor hierarchies. In Birnbaum, L. and Collins, G., editors, Machine Learning:
Proceedings of the Eighth International Workshop, pages 343–347. Morgan Kaufmann.
Ring, M. B. (1993). Learning sequential tasks by incrementally adding higher orders. In S. J. Hanson,
J. D. C. and Giles, C. L., editors, Advances in Neural Information Processing Systems 5, pages 115–
122. Morgan Kaufmann.
Ring, M. B. (1994). Continual Learning in Reinforcement Environments. PhD thesis, University of
Texas at Austin, Austin, Texas 78712.
Risi, S. and Stanley, K. O. (2012). A unified approach to evolving plasticity and neural geometry. In
International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE.
Rissanen, J. (1986). Stochastic complexity and modeling. The Annals of Statistics, 14(3):1080–1100.
Ritter, H. and Kohonen, T. (1989).
61(4):241–254.
Self-organizing semantic maps.
Biological Cybernetics,
Robinson, A. J. and Fallside, F. (1987). The utility driven dynamic error propagation network. Technical Report CUED/F-INFENG/TR.1, Cambridge University Engineering Department.
Robinson, T. and Fallside, F. (1989). Dynamic reinforcement driven error propagation networks
with application to game playing. In Proceedings of the 11th Conference of the Cognitive Science
Society, Ann Arbor, pages 836–843.
Rodriguez, P. and Wiles, J. (1998). Recurrent neural networks can learn to implement symbolsensitive counting. In Advances in Neural Information Processing Systems (NIPS), volume 10,
pages 87–93. The MIT Press.
Rodriguez, P., Wiles, J., and Elman, J. (1999). A recurrent neural network that learns to count.
Connection Science, 11(1):5–40.
73
Roggen, D., Hofmann, S., Thoma, Y., and Floreano, D. (2003). Hardware spiking neural network with
run-time reconfigurable connectivity in an autonomous robot. In Proc. NASA/DoD Conference on
Evolvable Hardware, 2003, pages 189–198. IEEE.
Rohwer, R. (1989). The ‘moving targets’ training method. In Kindermann, J. and Linden, A., editors, Proceedings of ‘Distributed Adaptive Neural Information Processing’, St.Augustin, 24.-25.5,.
Oldenbourg.
Rosenblatt, F. (1958). The perceptron: a probabilistic model for information storage and organization
in the brain. Psychological review, 65(6):386.
Rosenblatt, F. (1962). Principles of Neurodynamics. Spartan, New York.
Roux, L., Racoceanu, D., Lomenie, N., Kulikova, M., Irshad, H., Klossa, J., Capron, F., Genestie, C.,
Naour, G. L., and Gurcan, M. N. (2013). Mitosis detection in breast cancer histological images an ICPR 2012 contest. J. Pathol. Inform., 4:8.
Rubner, J. and Schulten, K. (1990). Development of feature detectors by self-organization: A network
model. Biological Cybernetics, 62:193–199.
Rückstieß, T., Felder, M., and Schmidhuber, J. (2008). State-Dependent Exploration for policy gradient methods. In et al., W. D., editor, European Conference on Machine Learning (ECML) and
Principles and Practice of Knowledge Discovery in Databases 2008, Part II, LNAI 5212, pages
234–249.
Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning internal representations by error
propagation. In Rumelhart, D. E. and McClelland, J. L., editors, Parallel Distributed Processing,
volume 1, pages 318–362. MIT Press.
Rumelhart, D. E. and Zipser, D. (1986). Feature discovery by competitive learning. In Parallel
Distributed Processing, pages 151–193. MIT Press.
Rummery, G. and Niranjan, M. (1994). On-line Q-learning using connectionist sytems. Technical
Report CUED/F-INFENG-TR 166, Cambridge University, UK.
Russell, S. J., Norvig, P., Canny, J. F., Malik, J. M., and Edwards, D. D. (1995). Artificial Intelligence:
a Modern Approach, volume 2. Englewood Cliffs: Prentice Hall.
Saito, K. and Nakano, R. (1997). Partial BFGS update and efficient step-length calculation for threelayer neural networks. Neural Computation, 9(1):123–141.
Sak, H., Senior, A., and Beaufays, F. (2014a). Long Short-Term Memory recurrent neural network
architectures for large scale acoustic modeling. In Proc. Interspeech.
Sak, H., Vinyals, O., Heigold, G., Senior, A., McDermott, E., Monga, R., and Mao, M. (2014b). Sequence discriminative distributed training of Long Short-Term Memory recurrent neural networks.
In Proc. Interspeech.
Salakhutdinov, R. and Hinton, G. (2009). Semantic hashing. Int. J. Approx. Reasoning, 50(7):969–
978.
Sallans, B. and Hinton, G. (2004). Reinforcement learning with factored states and actions. Journal
of Machine Learning Research, 5:1063–1088.
74
Sałustowicz, R. P. and Schmidhuber, J. (1997). Probabilistic incremental program evolution. Evolutionary Computation, 5(2):123–141.
Samejima, K., Doya, K., and Kawato, M. (2003). Inter-module credit assignment in modular reinforcement learning. Neural Networks, 16(7):985–994.
Samuel, A. L. (1959). Some studies in machine learning using the game of checkers. IBM Journal on
Research and Development, 3:210–229.
Sanger, T. D. (1989). An optimality principle for unsupervised learning. In Touretzky, D. S., editor,
Advances in Neural Information Processing Systems (NIPS) 1, pages 11–19. Morgan Kaufmann.
Santamarı́a, J. C., Sutton, R. S., and Ram, A. (1997). Experiments with reinforcement learning in
problems with continuous state and action spaces. Adaptive Behavior, 6(2):163–217.
Saravanan, N. and Fogel, D. B. (1995). Evolving neural control systems. IEEE Expert, pages 23–27.
Saund, E. (1994). Unsupervised learning of mixtures of multiple causes in binary data. In Cowan,
J. D., Tesauro, G., and Alspector, J., editors, Advances in Neural Information Processing Systems
(NIPS) 6, pages 27–34. Morgan Kaufmann.
Schaback, R. and Werner, H. (1992). Numerische Mathematik, volume 4. Springer.
Schäfer, A. M., Udluft, S., and Zimmermann, H.-G. (2006). Learning long term dependencies with
recurrent neural networks. In Kollias, S. D., Stafylopatis, A., Duch, W., and Oja, E., editors, ICANN
(1), volume 4131 of Lecture Notes in Computer Science, pages 71–80. Springer.
Schapire, R. E. (1990). The strength of weak learnability. Machine Learning, 5:197–227.
Schaul, T. and Schmidhuber, J. (2010). Metalearning. Scholarpedia, 6(5):4650.
Schaul, T., Zhang, S., and LeCun, Y. (2013). No more pesky learning rates. In Proc. 30th International
Conference on Machine Learning (ICML).
Schemmel, J., Grubl, A., Meier, K., and Mueller, E. (2006). Implementing synaptic plasticity in
a VLSI spiking neural network model. In International Joint Conference on Neural Networks
(IJCNN), pages 1–6. IEEE.
Scherer, D., Müller, A., and Behnke, S. (2010). Evaluation of pooling operations in convolutional architectures for object recognition. In Proc. International Conference on Artificial Neural Networks
(ICANN), pages 92–101.
Schmidhuber, J. (1987). Evolutionary principles in self-referential learning, or on learning
how to learn: the meta-meta-... hook. Diploma thesis, Inst. f. Inf., Tech. Univ. Munich.
http://www.idsia.ch/˜juergen/diploma.html.
Schmidhuber, J. (1989a). Accelerated learning in back-propagation nets. In Pfeifer, R., Schreter, Z.,
Fogelman, Z., and Steels, L., editors, Connectionism in Perspective, pages 429 – 438. Amsterdam:
Elsevier, North-Holland.
Schmidhuber, J. (1989b). A local learning algorithm for dynamic feedforward and recurrent networks.
Connection Science, 1(4):403–412.
75
Schmidhuber, J. (1990a). Dynamische neuronale Netze und das fundamentale raumzeitliche Lernproblem. (Dynamic neural nets and the fundamental spatio-temporal credit assignment problem.)
Dissertation, Inst. f. Inf., Tech. Univ. Munich.
Schmidhuber, J. (1990b). Learning algorithms for networks with internal and external feedback. In
Touretzky, D. S., Elman, J. L., Sejnowski, T. J., and Hinton, G. E., editors, Proc. of the 1990
Connectionist Models Summer School, pages 52–61. Morgan Kaufmann.
Schmidhuber, J. (1990c). The Neural Heat Exchanger. Talks at TU Munich (1990), University
of Colorado at Boulder (1992), and Z. Li’s NIPS*94 workshop on unsupervised learning. Also
published at the Intl. Conference on Neural Information Processing (ICONIP’96), vol. 1, pages
194-197, 1996.
Schmidhuber, J. (1990d). An on-line algorithm for dynamic reinforcement learning and planning in
reactive environments. In Proc. IEEE/INNS International Joint Conference on Neural Networks,
San Diego, volume 2, pages 253–258.
Schmidhuber, J. (1991a). Curious model-building control systems. In Proceedings of the International
Joint Conference on Neural Networks, Singapore, volume 2, pages 1458–1463. IEEE press.
Schmidhuber, J. (1991b). Learning to generate sub-goals for action sequences. In Kohonen, T.,
Mäkisara, K., Simula, O., and Kangas, J., editors, Artificial Neural Networks, pages 967–972.
Elsevier Science Publishers B.V., North-Holland.
Schmidhuber, J. (1991c). Reinforcement learning in Markovian and non-Markovian environments.
In Lippman, D. S., Moody, J. E., and Touretzky, D. S., editors, Advances in Neural Information
Processing Systems 3 (NIPS 3), pages 500–506. Morgan Kaufmann.
Schmidhuber, J. (1992a). A fixed size storage O(n3 ) time complexity learning algorithm for fully
recurrent continually running networks. Neural Computation, 4(2):243–248.
Schmidhuber, J. (1992b). Learning complex, extended sequences using the principle of history compression. Neural Computation, 4(2):234–242. (Based on TR FKI-148-91, TUM, 1991).
Schmidhuber, J. (1992c). Learning factorial codes by predictability minimization. Neural Computation, 4(6):863–879.
Schmidhuber, J. (1993a). An introspective network that can learn to run its own weight change
algorithm. In Proc. of the Intl. Conf. on Artificial Neural Networks, Brighton, pages 191–195. IEE.
Schmidhuber, J. (1993b). Netzwerkarchitekturen, Zielfunktionen und Kettenregel. (Network architectures, objective functions, and chain rule.) Habilitation Thesis, Inst. f. Inf., Tech. Univ. Munich.
Schmidhuber, J. (1997). Discovering neural nets with low Kolmogorov complexity and high generalization capability. Neural Networks, 10(5):857–873.
Schmidhuber, J. (2002). The Speed Prior: a new simplicity measure yielding near-optimal computable
predictions. In Kivinen, J. and Sloan, R. H., editors, Proceedings of the 15th Annual Conference
on Computational Learning Theory (COLT 2002), Lecture Notes in Artificial Intelligence, pages
216–228. Springer, Sydney, Australia.
Schmidhuber, J. (2004). Optimal ordered problem solver. Machine Learning, 54:211–254.
76
Schmidhuber, J. (2006a). Developmental robotics, optimal artificial curiosity, creativity, music, and
the fine arts. Connection Science, 18(2):173–187.
Schmidhuber, J. (2006b). Gödel machines: Fully self-referential optimal universal self-improvers. In
Goertzel, B. and Pennachin, C., editors, Artificial General Intelligence, pages 199–226. Springer
Verlag. Variant available as arXiv:cs.LO/0309048.
Schmidhuber, J. (2007). Prototype resilient, self-modeling robots. Science, 316(5825):688.
Schmidhuber, J. (2012). Self-delimiting neural networks.
arXiv:1210.0118v1 [cs.NE], The Swiss AI Lab IDSIA.
Technical Report IDSIA-08-12,
Schmidhuber, J. (2013a). My first Deep Learning system of 1991 + Deep Learning timeline 19622013. Technical Report arXiv:1312.5548v1 [cs.NE], The Swiss AI Lab IDSIA.
Schmidhuber, J. (2013b). P OWER P LAY: Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem. Frontiers in Psychology.
Schmidhuber, J., Ciresan, D., Meier, U., Masci, J., and Graves, A. (2011). On fast deep nets for AGI
vision. In Proc. Fourth Conference on Artificial General Intelligence (AGI), Google, Mountain
View, CA, pages 243–246.
Schmidhuber, J., Eldracher, M., and Foltin, B. (1996). Semilinear predictability minimization produces well-known feature detectors. Neural Computation, 8(4):773–786.
Schmidhuber, J. and Huber, R. (1991). Learning to generate artificial fovea trajectories for target
detection. International Journal of Neural Systems, 2(1 & 2):135–141.
Schmidhuber, J., Mozer, M. C., and Prelinger, D. (1993). Continuous history compression. In Hüning,
H., Neuhauser, S., Raus, M., and Ritschel, W., editors, Proc. of Intl. Workshop on Neural Networks,
RWTH Aachen, pages 87–95. Augustinus.
Schmidhuber, J. and Prelinger, D. (1992). Discovering predictable classifications. Technical Report
CU-CS-626-92, Dept. of Comp. Sci., University of Colorado at Boulder. Published in Neural
Computation 5(4):625-635 (1993).
Schmidhuber, J. and Wahnsiedler, R. (1992). Planning simple trajectories using neural subgoal generators. In Meyer, J. A., Roitblat, H. L., and Wilson, S. W., editors, Proc. of the 2nd International
Conference on Simulation of Adaptive Behavior, pages 196–202. MIT Press.
Schmidhuber, J., Wierstra, D., Gagliolo, M., and Gomez, F. J. (2007). Training recurrent networks by
Evolino. Neural Computation, 19(3):757–779.
Schmidhuber, J., Zhao, J., and Schraudolph, N. (1997a). Reinforcement learning with self-modifying
policies. In Thrun, S. and Pratt, L., editors, Learning to learn, pages 293–309. Kluwer.
Schmidhuber, J., Zhao, J., and Wiering, M. (1997b). Shifting inductive bias with success-story algorithm, adaptive Levin search, and incremental self-improvement. Machine Learning, 28:105–130.
Schölkopf, B., Burges, C. J. C., and Smola, A. J., editors (1998). Advances in Kernel Methods Support Vector Learning. MIT Press, Cambridge, MA.
77
Schraudolph, N. and Sejnowski, T. J. (1993). Unsupervised discrimination of clustered data via optimization of binary information gain. In Hanson, S. J., Cowan, J. D., and Giles, C. L., editors,
Advances in Neural Information Processing Systems, volume 5, pages 499–506. Morgan Kaufmann, San Mateo.
Schraudolph, N. N. (2002). Fast curvature matrix-vector products for second-order gradient descent.
Neural Computation, 14(7):1723–1738.
Schraudolph, N. N. and Sejnowski, T. J. (1996). Tempering backpropagation networks: Not all
weights are created equal. In Touretzky, D. S., Mozer, M. C., and Hasselmo, M. E., editors, Advances in Neural Information Processing Systems (NIPS), volume 8, pages 563–569. The MIT
Press, Cambridge, MA.
Schrauwen, B., Verstraeten, D., and Van Campenhout, J. (2007). An overview of reservoir computing:
theory, applications and implementations. In Proceedings of the 15th European Symposium on
Artificial Neural Networks. p. 471-482 2007, pages 471–482.
Schuster, H. G. (1992). Learning by maximization the information transfer through nonlinear noisy
neurons and “noise breakdown”. Phys. Rev. A, 46(4):2131–2138.
Schuster, M. (1999). On supervised learning from sequential data with applications for speech recognition. PhD thesis, Nara Institute of Science and Technolog, Kyoto, Japan.
Schuster, M. and Paliwal, K. K. (1997). Bidirectional recurrent neural networks. IEEE Transactions
on Signal Processing, 45:2673–2681.
Schwartz, A. (1993). A reinforcement learning method for maximizing undiscounted rewards. In
Proc. ICML, pages 298–305.
Schwefel, H. P. (1974). Numerische Optimierung von Computer-Modellen. Dissertation. Published
1977 by Birkhäuser, Basel.
Segmentation of Neuronal Structures in EM Stacks Challenge (2012). IEEE International Symposium
on Biomedical Imaging (ISBI), http://tinyurl.com/d2fgh7g.
Sehnke, F., Osendorfer, C., Rückstieß, T., Graves, A., Peters, J., and Schmidhuber, J. (2010).
Parameter-exploring policy gradients. Neural Networks, 23(4):551–559.
Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., and LeCun, Y. (2013). OverFeat:
Integrated recognition, localization and detection using convolutional networks. arXiv preprint
arXiv:1312.6229.
Sermanet, P. and LeCun, Y. (2011). Traffic sign recognition with multi-scale convolutional networks.
In Proceedings of International Joint Conference on Neural Networks (IJCNN’11), pages 2809–
2813.
Serrano-Gotarredona, R., Oster, M., Lichtsteiner, P., Linares-Barranco, A., Paz-Vicente, R., GómezRodrı́guez, F., Camuñas-Mesa, L., Berner, R., Rivas-Pérez, M., Delbruck, T., et al. (2009). Caviar:
A 45k neuron, 5m synapse, 12g connects/s AER hardware sensory–processing–learning–actuating
system for high-speed visual object recognition and tracking. IEEE Transactions on Neural Networks, 20(9):1417–1438.
78
Serre, T., Riesenhuber, M., Louie, J., and Poggio, T. (2002). On the role of object-specific features
for real world object recognition in biological vision. In Biologically Motivated Computer Vision,
pages 387–397.
Seung, H. S. (2003). Learning in spiking neural networks by reinforcement of stochastic synaptic
transmission. Neuron, 40(6):1063–1073.
Shan, H. and Cottrell, G. (2014). Efficient visual coding: From retina to V2. In Proc. International
Conference on Learning Representations (ICLR). arXiv preprint arXiv:1312.6077.
Shan, H., Zhang, L., and Cottrell, G. W. (2007). Recursive ICA. Advances in Neural Information
Processing Systems (NIPS), 19:1273.
Shanno, D. F. (1970). Conditioning of quasi-Newton methods for function minimization. Mathematics of computation, 24(111):647–656.
Shannon, C. E. (1948). A mathematical theory of communication (parts I and II). Bell System
Technical Journal, XXVII:379–423.
Shao, L., Wu, D., and Li, X. (2014). Learning deep and wide: A spectral method for learning deep
networks. IEEE Transactions on Neural Networks and Learning Systems.
Shavlik, J. W. (1994). Combining symbolic and neural learning. Machine Learning, 14(3):321–331.
Shavlik, J. W. and Towell, G. G. (1989). Combining explanation-based and neural learning: An
algorithm and empirical results. Connection Science, 1(3):233–255.
Siegelmann, H. (1992). Theoretical Foundations of Recurrent Neural Networks. PhD thesis, Rutgers,
New Brunswick Rutgers, The State of New Jersey.
Siegelmann, H. T. and Sontag, E. D. (1991). Turing computability with neural nets. Applied Mathematics Letters, 4(6):77–80.
Silva, F. M. and Almeida, L. B. (1990). Speeding up back-propagation. In Eckmiller, R., editor,
Advanced Neural Computers, pages 151–158, Amsterdam. Elsevier.
Sı́ma, J. (1994). Loading deep networks is hard. Neural Computation, 6(5):842–850.
Sı́ma, J. (2002). Training a single sigmoidal neuron is hard. Neural Computation, 14(11):2709–2728.
Simard, P., Steinkraus, D., and Platt, J. (2003). Best practices for convolutional neural networks
applied to visual document analysis. In Seventh International Conference on Document Analysis
and Recognition, pages 958–963.
Sims, K. (1994). Evolving virtual creatures. In Glassner, A., editor, Proceedings of SIGGRAPH ’94
(Orlando, Florida, July 1994), Computer Graphics Proceedings, Annual Conference, pages 15–22.
ACM SIGGRAPH, ACM Press. ISBN 0-89791-667-0.
Simsek, Ö. and Barto, A. G. (2008). Skill characterization based on betweenness. In NIPS’08, pages
1497–1504.
Singh, S., Barto, A. G., and Chentanez, N. (2005). Intrinsically motivated reinforcement learning. In
Advances in Neural Information Processing Systems 17 (NIPS). MIT Press, Cambridge, MA.
79
Singh, S. P. (1994). Reinforcement learning algorithms for average-payoff Markovian decision processes. In National Conference on Artificial Intelligence, pages 700–705.
Smith, S. F. (1980). A Learning System Based on Genetic Adaptive Algorithms,. PhD thesis, Univ.
Pittsburgh.
Smolensky, P. (1986). Parallel distributed processing: Explorations in the microstructure of cognition,
vol. 1. chapter Information Processing in Dynamical Systems: Foundations of Harmony Theory,
pages 194–281. MIT Press, Cambridge, MA, USA.
Solla, S. A. (1988). Accelerated learning in layered neural networks. Complex Systems, 2:625–640.
Solomonoff, R. J. (1964). A formal theory of inductive inference. Part I. Information and Control,
7:1–22.
Solomonoff, R. J. (1978). Complexity-based induction systems. IEEE Transactions on Information
Theory, IT-24(5):422–432.
Soloway, E. (1986). Learning to program = learning to construct mechanisms and explanations.
Communications of the ACM, 29(9):850–858.
Song, S., Miller, K. D., and Abbott, L. F. (2000). Competitive Hebbian learning through spike-timingdependent synaptic plasticity. Nature Neuroscience, 3(9):919–926.
Speelpenning, B. (1980). Compiling Fast Partial Derivatives of Functions Given by Algorithms. PhD
thesis, Department of Computer Science, University of Illinois, Urbana-Champaign.
Srivastava, R. K., Masci, J., Kazerounian, S., Gomez, F., and Schmidhuber, J. (2013). Compete to
compute. In Advances in Neural Information Processing Systems (NIPS), pages 2310–2318.
Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C. (2011). The German traffic sign recognition
benchmark: A multi-class classification competition. In International Joint Conference on Neural
Networks (IJCNN 2011), pages 1453–1460. IEEE Press.
Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C. (2012). Man vs. computer: Benchmarking
machine learning algorithms for traffic sign recognition. Neural Networks, 32:323–332.
Stanley, K. O., D’Ambrosio, D. B., and Gauci, J. (2009). A hypercube-based encoding for evolving
large-scale neural networks. Artificial Life, 15(2):185–212.
Stanley, K. O. and Miikkulainen, R. (2002). Evolving neural networks through augmenting topologies. Evolutionary Computation, 10:99–127.
Steijvers, M. and Grunwald, P. (1996). A recurrent network that performs a contextsensitive prediction
task. In Proceedings of the 18th Annual Conference of the Cognitive Science Society. Erlbaum.
Steil, J. J. (2007). Online reservoir adaptation by intrinsic plasticity for backpropagation–
decorrelation and echo state learning. Neural Networks, 20(3):353–364.
Stemmler, M. (1996). A single spike suffices: the simplest form of stochastic resonance in model
neurons. Network: Computation in Neural Systems, 7(4):687–716.
Stoianov, I. and Zorzi, M. (2012). Emergence of a ’visual number sense’ in hierarchical generative
models. Nature Neuroscience, 15(2):194–6.
80
Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Roy. Stat. Soc.,
36:111–147.
Stoop, R., Schindler, K., and Bunimovich, L. (2000). When pyramidal neurons lock, when they
respond chaotically, and when they like to synchronize. Neuroscience research, 36(1):81–91.
Stratonovich, R. (1960). Conditional Markov processes. Theory of Probability And Its Applications,
5(2):156–178.
Sun, G., Chen, H., and Lee, Y. (1993a). Time warping invariant neural networks. In S. J. Hanson,
J. D. C. and Giles, C. L., editors, Advances in Neural Information Processing Systems (NIPS) 5,
pages 180–187. Morgan Kaufmann.
Sun, G. Z., Giles, C. L., Chen, H. H., and Lee, Y. C. (1993b). The neural network pushdown automaton: Model, stack and learning simulations. Technical Report CS-TR-3118, University of
Maryland, College Park.
Sun, Y., Gomez, F., Schaul, T., and Schmidhuber, J. (2013). A Linear Time Natural Evolution Strategy for Non-Separable Functions. In Proceedings of the Genetic and Evolutionary Computation
Conference, page 61, Amsterdam, NL. ACM.
Sun, Y., Wierstra, D., Schaul, T., and Schmidhuber, J. (2009). Efficient natural evolution strategies.
In Proc. 11th Genetic and Evolutionary Computation Conference (GECCO), pages 539–546.
Sutskever, I., Hinton, G. E., and Taylor, G. W. (2008). The recurrent temporal restricted Boltzmann
machine. In NIPS, volume 21, page 2008.
Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence to sequence learning with neural networks.
Technical Report arXiv:1409.3215 [cs.CL], Google. NIPS’2014.
Sutton, R. and Barto, A. (1998). Reinforcement learning: An introduction. Cambridge, MA, MIT
Press.
Sutton, R. S., McAllester, D. A., Singh, S. P., and Mansour, Y. (1999a). Policy gradient methods for
reinforcement learning with function approximation. In Advances in Neural Information Processing Systems (NIPS) 12, pages 1057–1063.
Sutton, R. S., Precup, D., and Singh, S. P. (1999b). Between MDPs and semi-MDPs: A framework
for temporal abstraction in reinforcement learning. Artif. Intell., 112(1-2):181–211.
Sutton, R. S., Szepesvári, C., and Maei, H. R. (2008). A convergent O(n) algorithm for off-policy
temporal-difference learning with linear function approximation. In Advances in Neural Information Processing Systems (NIPS’08), volume 21, pages 1609–1616.
Szabó, Z., Póczos, B., and Lőrincz, A. (2006). Cross-entropy optimization for independent process analysis. In Independent Component Analysis and Blind Signal Separation, pages 909–916.
Springer.
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. (2014). Going deeper with convolutions. Technical Report arXiv:1409.4842 [cs.CV],
Google.
Szegedy, C., Toshev, A., and Erhan, D. (2013). Deep neural networks for object detection. pages
2553–2561.
81
Taylor, G. W., Spiro, I., Bregler, C., and Fergus, R. (2011). Learning invariance through imitation. In
Conference on Computer Vision and Pattern Recognition (CVPR), pages 2729–2736. IEEE.
Tegge, A. N., Wang, Z., Eickholt, J., and Cheng, J. (2009). NNcon: improved protein contact map
prediction using 2D-recursive neural networks. Nucleic Acids Research, 37(Suppl 2):W515–W518.
Teichmann, M., Wiltschut, J., and Hamker, F. (2012). Learning invariance from natural images inspired by observations in the primary visual cortex. Neural Computation, 24(5):1271–1296.
Teller, A. (1994). The evolution of mental models. In Kenneth E. Kinnear, J., editor, Advances in
Genetic Programming, pages 199–219. MIT Press.
Tenenberg, J., Karlsson, J., and Whitehead, S. (1993). Learning via task decomposition. In Meyer,
J. A., Roitblat, H., and Wilson, S., editors, From Animals to Animats 2: Proceedings of the Second
International Conference on Simulation of Adaptive Behavior, pages 337–343. MIT Press.
Tesauro, G. (1994). TD-gammon, a self-teaching backgammon program, achieves master-level play.
Neural Computation, 6(2):215–219.
Tieleman, T. and Hinton, G. (2012). Lecture 6.5—RmsProp: Divide the gradient by a running average
of its recent magnitude. COURSERA: Neural Networks for Machine Learning.
Tikhonov, A. N., Arsenin, V. I., and John, F. (1977). Solutions of ill-posed problems. Winston.
Ting, K. M. and Witten, I. H. (1997). Stacked generalization: when does it work?
International Joint Conference on Artificial Intelligence (IJCAI).
In in Proc.
Tiňo, P. and Hammer, B. (2004). Architectural bias in recurrent neural networks: Fractal analysis.
Neural Computation, 15(8):1931–1957.
Tonkes, B. and Wiles, J. (1997). Learning a context-free task with a recurrent neural network: An
analysis of stability. In Proceedings of the Fourth Biennial Conference of the Australasian Cognitive
Science Society.
Towell, G. G. and Shavlik, J. W. (1994). Knowledge-based artificial neural networks. Artificial
Intelligence, 70(1):119–165.
Tsitsiklis, J. N. and van Roy, B. (1996). Feature-based methods for large scale dynamic programming.
Machine Learning, 22(1-3):59–94.
Tsodyks, M., Pawelzik, K., and Markram, H. (1998). Neural networks with dynamic synapses. Neural
Computation, 10(4):821–835.
Tsodyks, M. V., Skaggs, W. E., Sejnowski, T. J., and McNaughton, B. L. (1996). Population dynamics and theta rhythm phase precession of hippocampal place cell firing: a spiking neuron model.
Hippocampus, 6(3):271–280.
Turaga, S. C., Murray, J. F., Jain, V., Roth, F., Helmstaedter, M., Briggman, K., Denk, W., and Seung,
H. S. (2010). Convolutional networks can learn to generate affinity graphs for image segmentation.
Neural Computation, 22(2):511–538.
Turing, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem.
Proceedings of the London Mathematical Society, Series 2, 41:230–267.
82
Turner, A. J. and Miller, J. F. (2013). Cartesian Genetic Programming encoded artificial neural networks: A comparison using three benchmarks. In Proceedings of the Conference on Genetic and
Evolutionary Computation (GECCO), pages 1005–1012.
Ueda, N. (2000). Optimal linear combination of neural networks for improving classification performance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(2):207–215.
Urlbe, A. P. (1999). Structure-adaptable digital neural networks. PhD thesis, Universidad del Valle.
Utgoff, P. E. and Stracuzzi, D. J. (2002). Many-layered learning. Neural Computation, 14(10):2497–
2529.
Vahed, A. and Omlin, C. W. (2004). A machine learning method for extracting symbolic knowledge
from recurrent neural networks. Neural Computation, 16(1):59–71.
Vaillant, R., Monrocq, C., and LeCun, Y. (1994). Original approach for the localisation of objects in
images. IEE Proc on Vision, Image, and Signal Processing, 141(4):245–250.
van den Berg, T. and Whiteson, S. (2013). Critical factors in the performance of HyperNEAT. In
GECCO 2013: Proceedings of the Genetic and Evolutionary Computation Conference, pages 759–
766.
van Hasselt, H. (2012). Reinforcement learning in continuous state and action spaces. In Wiering, M.
and van Otterlo, M., editors, Reinforcement Learning, pages 207–251. Springer.
Vapnik, V. (1992). Principles of risk minimization for learning theory. In Lippman, D. S., Moody,
J. E., and Touretzky, D. S., editors, Advances in Neural Information Processing Systems (NIPS) 4,
pages 831–838. Morgan Kaufmann.
Vapnik, V. (1995). The Nature of Statistical Learning Theory. Springer, New York.
Versino, C. and Gambardella, L. M. (1996). Learning fine motion by using the hierarchical extended Kohonen map. In Proc. Intl. Conf. on Artificial Neural Networks (ICANN), pages 221–226.
Springer.
Veta, M., Viergever, M., Pluim, J., Stathonikos, N., and van Diest, P. J. (2013). MICCAI 2013 Grand
Challenge on Mitosis Detection.
Vieira, A. and Barradas, N. (2003). A training algorithm for classification of high-dimensional data.
Neurocomputing, 50:461–472.
Viglione, S. (1970). Applications of pattern recognition technology. In Mendel, J. M. and Fu, K. S.,
editors, Adaptive, Learning, and Pattern Recognition Systems. Academic Press.
Vincent, P., Hugo, L., Bengio, Y., and Manzagol, P.-A. (2008). Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine
learning, ICML ’08, pages 1096–1103, New York, NY, USA. ACM.
Vlassis, N., Littman, M. L., and Barber, D. (2012). On the computational complexity of stochastic
controller optimization in POMDPs. ACM Transactions on Computation Theory, 4(4):12.
Vogl, T., Mangis, J., Rigler, A., Zink, W., and Alkon, D. (1988). Accelerating the convergence of the
back-propagation method. Biological Cybernetics, 59:257–263.
83
von der Malsburg, C. (1973). Self-organization of orientation sensitive cells in the striate cortex.
Kybernetik, 14(2):85–100.
Waldinger, R. J. and Lee, R. C. T. (1969). PROW: a step toward automatic program writing. In
Walker, D. E. and Norton, L. M., editors, Proceedings of the 1st International Joint Conference on
Artificial Intelligence (IJCAI), pages 241–252. Morgan Kaufmann.
Wallace, C. S. and Boulton, D. M. (1968). An information theoretic measure for classification. Computer Journal, 11(2):185–194.
Wan, E. A. (1994). Time series prediction by using a connectionist network with internal delay lines.
In Weigend, A. S. and Gershenfeld, N. A., editors, Time series prediction: Forecasting the future
and understanding the past, pages 265–295. Addison-Wesley.
Wang, C., Venkatesh, S. S., and Judd, J. S. (1994). Optimal stopping and effective machine complexity
in learning. In Advances in Neural Information Processing Systems (NIPS’6), pages 303–310.
Morgan Kaufmann.
Wang, S. and Manning, C. (2013). Fast dropout training. In Proceedings of the 30th International
Conference on Machine Learning (ICML-13), pages 118–126.
Watanabe, O. (1992). Kolmogorov complexity and computational complexity. EATCS Monographs
on Theoretical Computer Science, Springer.
Watanabe, S. (1985). Pattern Recognition: Human and Mechanical. Willey, New York.
Watkins, C. J. C. H. (1989). Learning from Delayed Rewards. PhD thesis, King’s College, Oxford.
Watkins, C. J. C. H. and Dayan, P. (1992). Q-learning. Machine Learning, 8:279–292.
Watrous, R. L. and Kuhn, G. M. (1992). Induction of finite-state automata using second-order recurrent networks. In Moody, J. E., Hanson, S. J., and Lippman, R. P., editors, Advances in Neural
Information Processing Systems 4, pages 309–316. Morgan Kaufmann.
Waydo, S. and Koch, C. (2008). Unsupervised learning of individuals and categories from images.
Neural Computation, 20(5):1165–1178.
Weigend, A. S. and Gershenfeld, N. A. (1993). Results of the time series prediction competition
at the Santa Fe Institute. In Neural Networks, 1993., IEEE International Conference on, pages
1786–1793. IEEE.
Weigend, A. S., Rumelhart, D. E., and Huberman, B. A. (1991). Generalization by weight-elimination
with application to forecasting. In Lippmann, R. P., Moody, J. E., and Touretzky, D. S., editors,
Advances in Neural Information Processing Systems (NIPS) 3, pages 875–882. San Mateo, CA:
Morgan Kaufmann.
Weiss, G. (1994). Hierarchical chunking in classifier systems. In Proceedings of the 12th National
Conference on Artificial Intelligence, volume 2, pages 1335–1340. AAAI Press/The MIT Press.
Weng, J., Ahuja, N., and Huang, T. S. (1992). Cresceptron: a self-organizing neural network which
grows adaptively. In International Joint Conference on Neural Networks (IJCNN), volume 1, pages
576–581. IEEE.
Weng, J. J., Ahuja, N., and Huang, T. S. (1997). Learning recognition and segmentation using the
cresceptron. International Journal of Computer Vision, 25(2):109–143.
84
Werbos, P. J. (1974). Beyond Regression: New Tools for Prediction and Analysis in the Behavioral
Sciences. PhD thesis, Harvard University.
Werbos, P. J. (1981). Applications of advances in nonlinear sensitivity analysis. In Proceedings of the
10th IFIP Conference, 31.8 - 4.9, NYC, pages 762–770.
Werbos, P. J. (1987). Building and understanding adaptive systems: A statistical/numerical approach
to factory automation and brain research. IEEE Transactions on Systems, Man, and Cybernetics,
17.
Werbos, P. J. (1988). Generalization of backpropagation with application to a recurrent gas market
model. Neural Networks, 1.
Werbos, P. J. (1989a). Backpropagation and neurocontrol: A review and prospectus. In IEEE/INNS
International Joint Conference on Neural Networks, Washington, D.C., volume 1, pages 209–216.
Werbos, P. J. (1989b). Neural networks for control and system identification. In Proceedings of
IEEE/CDC Tampa, Florida.
Werbos, P. J. (1992). Neural networks, system identification, and control in the chemical industries.
In D. A. White, D. A. S., editor, Handbook of Intelligent Control: Neural, Fuzzy, and Adaptive
Approaches, pages 283–356. Thomson Learning.
Werbos, P. J. (2006). Backwards differentiation in AD and neural nets: Past links and new opportunities. In Automatic Differentiation: Applications, Theory, and Implementations, pages 15–34.
Springer.
West, A. H. L. and Saad, D. (1995). Adaptive back-propagation in on-line learning of multilayer
networks. In Touretzky, D. S., Mozer, M., and Hasselmo, M. E., editors, NIPS, pages 323–329.
MIT Press.
White, H. (1989). Learning in artificial neural networks: A statistical perspective. Neural Computation, 1(4):425–464.
Whitehead, S. (1992). Reinforcement Learning for the adaptive control of perception and action. PhD
thesis, University of Rochester.
Whiteson, S. (2012). Evolutionary computation for reinforcement learning. In Wiering, M. and van
Otterlo, M., editors, Reinforcement Learning, pages 325–355. Springer, Berlin, Germany.
Whiteson, S., Kohl, N., Miikkulainen, R., and Stone, P. (2005). Evolving keepaway soccer players
through task decomposition. Machine Learning, 59(1):5–30.
Whiteson, S. and Stone, P. (2006). Evolutionary function approximation for reinforcement learning.
Journal of Machine Learning Research, 7:877–917.
Widrow, B. and Hoff, M. (1962). Associative storage and retrieval of digital information in networks
of adaptive neurons. Biological Prototypes and Synthetic Systems, 1:160.
Widrow, B., Rumelhart, D. E., and Lehr, M. A. (1994). Neural networks: Applications in industry,
business and science. Commun. ACM, 37(3):93–105.
Wieland, A. P. (1991). Evolving neural network controllers for unstable systems. In International
Joint Conference on Neural Networks (IJCNN), volume 2, pages 667–673. IEEE.
85
Wiering, M. and Schmidhuber, J. (1996). Solving POMDPs with Levin search and EIRA. In Saitta,
L., editor, Machine Learning: Proceedings of the Thirteenth International Conference, pages 534–
542. Morgan Kaufmann Publishers, San Francisco, CA.
Wiering, M. and Schmidhuber, J. (1998a). HQ-learning. Adaptive Behavior, 6(2):219–246.
Wiering, M. and van Otterlo, M. (2012). Reinforcement Learning. Springer.
Wiering, M. A. and Schmidhuber, J. (1998b). Fast online Q(λ). Machine Learning, 33(1):105–116.
Wierstra, D., Foerster, A., Peters, J., and Schmidhuber, J. (2010). Recurrent policy gradients. Logic
Journal of IGPL, 18(2):620–634.
Wierstra, D., Schaul, T., Peters, J., and Schmidhuber, J. (2008). Natural evolution strategies. In
Congress of Evolutionary Computation (CEC 2008).
Wiesel, D. H. and Hubel, T. N. (1959). Receptive fields of single neurones in the cat’s striate cortex.
J. Physiol., 148:574–591.
Wiles, J. and Elman, J. (1995). Learning to count without a counter: A case study of dynamics
and activation landscapes in recurrent networks. In In Proceedings of the Seventeenth Annual
Conference of the Cognitive Science Society, pages pages 482 – 487, Cambridge, MA. MIT Press.
Wilkinson, J. H., editor (1965). The Algebraic Eigenvalue Problem. Oxford University Press, Inc.,
New York, NY, USA.
Williams, R. J. (1986). Reinforcement-learning in connectionist networks: A mathematical analysis.
Technical Report 8605, Institute for Cognitive Science, University of California, San Diego.
Williams, R. J. (1988). Toward a theory of reinforcement-learning connectionist systems. Technical
Report NU-CCS-88-3, College of Comp. Sci., Northeastern University, Boston, MA.
Williams, R. J. (1989). Complexity of exact gradient computation algorithms for recurrent neural
networks. Technical Report Technical Report NU-CCS-89-27, Boston: Northeastern University,
College of Computer Science.
Williams, R. J. (1992a). Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256.
Williams, R. J. (1992b). Training recurrent networks using the extended Kalman filter. In International Joint Conference on Neural Networks (IJCNN), volume 4, pages 241–246. IEEE.
Williams, R. J. and Peng, J. (1990). An efficient gradient-based algorithm for on-line training of
recurrent network trajectories. Neural Computation, 4:491–501.
Williams, R. J. and Zipser, D. (1988). A learning algorithm for continually running fully recurrent
networks. Technical Report ICS Report 8805, Univ. of California, San Diego, La Jolla.
Williams, R. J. and Zipser, D. (1989a). Experimental analysis of the real-time recurrent learning
algorithm. Connection Science, 1(1):87–111.
Williams, R. J. and Zipser, D. (1989b). A learning algorithm for continually running fully recurrent
networks. Neural Computation, 1(2):270–280.
86
Willshaw, D. J. and von der Malsburg, C. (1976). How patterned neural connections can be set up by
self-organization. Proc. R. Soc. London B, 194:431–445.
Windisch, D. (2005). Loading deep networks is hard: The pyramidal case. Neural Computation,
17(2):487–502.
Wiskott, L. and Sejnowski, T. (2002). Slow feature analysis: Unsupervised learning of invariances.
Neural Computation, 14(4):715–770.
Witczak, M., Korbicz, J., Mrugalski, M., and Patton, R. J. (2006). A GMDH neural network-based
approach to robust fault diagnosis: Application to the DAMADICS benchmark problem. Control
Engineering Practice, 14(6):671–683.
Wöllmer, M., Blaschke, C., Schindl, T., Schuller, B., Färber, B., Mayer, S., and Trefflich, B. (2011).
On-line driver distraction detection using Long Short-Term Memory. IEEE Transactions on Intelligent Transportation Systems (TITS), 12(2):574–582.
Wöllmer, M., Schuller, B., and Rigoll, G. (2013). Keyword spotting exploiting Long Short-Term
Memory. Speech Communication, 55(2):252–265.
Wolpert, D. H. (1992). Stacked generalization. Neural Networks, 5(2):241–259.
Wolpert, D. H. (1994). Bayesian backpropagation over i-o functions rather than weights. In Cowan,
J. D., Tesauro, G., and Alspector, J., editors, Advances in Neural Information Processing Systems
(NIPS) 6, pages 200–207. Morgan Kaufmann.
Wu, D. and Shao, L. (2014). Leveraging hierarchical parametric networks for skeletal joints based
action segmentation and recognition. In Proc. Conference on Computer Vision and Pattern Recognition (CVPR).
Wu, L. and Baldi, P. (2008). Learning to play Go using recursive neural networks. Neural Networks,
21(9):1392–1400.
Wyatte, D., Curran, T., and O’Reilly, R. (2012). The limits of feedforward vision: Recurrent processing promotes robust object recognition when objects are degraded. Journal of Cognitive Neuroscience, 24(11):2248–2261.
Wysoski, S. G., Benuskova, L., and Kasabov, N. (2010). Evolving spiking neural networks for audiovisual information processing. Neural Networks, 23(7):819–835.
Yamauchi, B. M. and Beer, R. D. (1994). Sequential behavior and learning in evolved dynamical
neural networks. Adaptive Behavior, 2(3):219–246.
Yamins, D., Hong, H., Cadieu, C., and DiCarlo, J. J. (2013). Hierarchical modular optimization of
convolutional networks achieves representations similar to macaque IT and human ventral stream.
Advances in Neural Information Processing Systems (NIPS), pages 1–9.
Yang, M., Ji, S., Xu, W., Wang, J., Lv, F., Yu, K., Gong, Y., Dikmen, M., Lin, D. J., and Huang,
T. S. (2009). Detecting human actions in surveillance videos. In TREC Video Retrieval Evaluation
Workshop.
Yao, X. (1993). A review of evolutionary artificial neural networks. International Journal of Intelligent Systems, 4:203–222.
87
Yin, F., Wang, Q.-F., Zhang, X.-Y., and Liu, C.-L. (2013). ICDAR 2013 Chinese handwriting recognition competition. In 12th International Conference on Document Analysis and Recognition (ICDAR), pages 1464–1470.
Yin, J., Meng, Y., and Jin, Y. (2012). A developmental approach to structural self-organization in
reservoir computing. IEEE Transactions on Autonomous Mental Development, 4(4):273–289.
Young, S., Davis, A., Mishtal, A., and Arel, I. (2014). Hierarchical spatiotemporal feature extraction
using recurrent online clustering. Pattern Recognition Letters, 37:115–123.
Yu, X.-H., Chen, G.-A., and Cheng, S.-X. (1995). Dynamic learning rate optimization of the backpropagation algorithm. IEEE Transactions on Neural Networks, 6(3):669–677.
Zamora-Martnez, F., Frinken, V., Espaa-Boquera, S., Castro-Bleda, M., Fischer, A., and Bunke, H.
(2014). Neural network language models for off-line handwriting recognition. Pattern Recognition,
47(4):1642–1652.
Zeiler, M. D. (2012). ADADELTA: An Adaptive Learning Rate Method. CoRR, abs/1212.5701.
Zeiler, M. D. and Fergus, R. (2013). Visualizing and understanding convolutional networks. Technical
Report arXiv:1311.2901 [cs.CV], NYU.
Zemel, R. S. (1993). A minimum description length framework for unsupervised learning. PhD thesis,
University of Toronto.
Zemel, R. S. and Hinton, G. E. (1994). Developing population codes by minimizing description
length. In Cowan, J. D., Tesauro, G., and Alspector, J., editors, Advances in Neural Information
Processing Systems 6, pages 11–18. Morgan Kaufmann.
Zeng, Z., Goodman, R., and Smyth, P. (1994). Discrete recurrent neural networks for grammatical
inference. IEEE Transactions on Neural Networks, 5(2).
Zimmermann, H.-G., Tietz, C., and Grothmann, R. (2012). Forecasting with recurrent neural networks: 12 tricks. In Montavon, G., Orr, G. B., and Müller, K.-R., editors, Neural Networks:
Tricks of the Trade (2nd ed.), volume 7700 of Lecture Notes in Computer Science, pages 687–707.
Springer.
Zipser, D., Kehoe, B., Littlewort, G., and Fuster, J. (1993). A spiking network model of short-term
active memory. The Journal of Neuroscience, 13(8):3406–3420.
88
| 9cs.NE
|
Generalizing the Convolution Operator
to extend CNNs to Irregular Domains
Jean-Charles Vialatte1,2 , Vincent Gripon2 , Grégoire Mercier2
arXiv:1606.01166v4 [cs.LG] 25 Oct 2017
1
Cityzen Data
Guipavas, Brittany, France
2
Telecom Bretagne
Brest, Brittany, France
1
2
[email protected]
{jc.vialatte, vincent.gripon, gregoire.mercier}@telecom-bretagne.eu
Abstract
Convolutional Neural Networks (CNNs) have become the state-of-the-art in supervised learning vision tasks. Their convolutional filters are of paramount importance
for they allow to learn patterns while disregarding their locations in input images.
When facing highly irregular domains, generalized convolutional operators based
on an underlying graph structure have been proposed. However, these operators do
not exactly match standard ones on grid graphs, and introduce unwanted additional
invariance (e.g. with regards to rotations). We propose a novel approach to generalize CNNs to irregular domains using weight sharing and graph-based operators.
Using experiments, we show that these models resemble CNNs on regular domains
and offer better performance than multilayer perceptrons on distorded ones.
1
Introduction
CNNs[1] are state-of-the-art for performing supervised learning on data defined on lattices [2].
Contrary to classical multilayer perceptrons (MLPs) [3], CNNs take advantage of the underlying
structure of the inputs. When facing data defined on irregular domains, CNNs cannot always be
directly applied even if the data may present an exploitable underlying structure.
An example of such data are spatio-temporal time series generated by a set of Internet of Things
devices. They typically consist of datapoints irregularly spaced out on a Euclidean space. As a result,
a graph G can be defined where the vertices are the devices and the edges connect neighbouring ones.
Other examples include signals on graphs including brain imaging, transport networks, bag of words
graphs. . . In all these examples, a signal can generally be seen as a vector in Rd where d is the order
of the graph. As such, each vertex is associated with a coordinate, and the edges weights represent
some association between the corresponding vertices.
Disregarding the graph G, MLPs can be applied on these datasets. We are interested in defining
convolutional operators that are able to exploit G as well as a possible embedding Euclidean space.
Our motivation is to imitate the gain of performance allowed by CNNs over MLPs on irregular
domains.
In graph signal processing, extended convolutional operators using spectral graph theory [4] have
been proposed [5]. These operators have been applied to deep learning [6] and obtain performance
similar to CNNs on regular domains, despite the fact they differ from classical convolutions. However,
for slightly distorted domains, the obtained operators become non-localized, thus failing taking
account of the underlying structure.
1
We propose another approach that generalizes convolutional operators by taking into account G.
Namely, we make sure that the proposed solution has properties inherent to convolutions: linearity,
locality and kernel weight sharing. We then apply it directly to the graph structure. We use it as a
substitution of a convolutional layer in a CNN and stress our resulting technique on comparative
benchmarks, showing a significant improvement compared with a MLP. The obtained operator
happens to exactly be the classical convolutional one when applied on regular domains.
The outline of the paper is as follows. In Section 2, we discuss related work. In Section 3 we introduce
our proposed operator. In Section 4 we explain how to apply it to CNN-like structures. Section 5
contains the experiments. Section 6 is a conclusion.
2
Related Works
For graph-structured data, Bruna et al. [6] have proposed an extension of the CNN using graph signal
processing theory [5]. Their convolution is defined in the spectral domain related to the Laplacian
matrix of the graph. As such, in the case where the graph is a lattice, the construction is analogous to
the regular convolution defined in the Fourier domain. The operation is defined as spectral multipliers
obtained by smooth interpolation of a weight kernel, and they explain how it ensures the localization
property. In that paper, they also define a construction that creates a multi-resolution structure of
a graph, for allowing it to support a deep learning architecture. Henaff et al. [7] have extended
Bruna’s spectral network to large scale classification tasks and have proposed both supervised and
unsupervided methods to find an underlying graph structure when it isn’t already given.
However, when the graph is irregular, they partly lose the localization property of their convolution is
partially lost. As the spectral domain is undirected with respect to the graph domain, some sort of
rotation invariances are also introduced. Hence, the results in the graph domain aren’t ressembling to
those of a convolution. In our case, we want to define a convolution supported locally. Moreover, we
want that every input can be defined on a different underlying graph structure, so that the learnt filters
can be applied on data embedded in the same space regardless of what structure have supported the
convolution during the training.
These properties are also retained by the convolution defined in the ShapeNet paper[8], which defines
a convolution for data living on non-euclidean manifolds. I.e their construction does maintain the
locality and allow for reusing the learnt filters on other manifolds. However, it requires at least
a manifold embedding of the data and a geodesic polar coordinates system. Although being less
specific, our proposed method will be a strict generalization of CNN in the sense that CNNs are a
special case of it.
In the case where the data is sparse, Graham [9] has proposed a framework to implement a spatially
sparse CNN efficiently. If the underlying graph structure is embedded into an Euclidean space, the
convolution we propose in this paper can be seen as spatially sparse too, in the sense that the data has
only non-zero coordinates on the vertices of the graph it is defined on. In our case we want to define
a convolution for which the inputs can have values on any point of the embedding space, whereas in
the regular case inputs can only have values on vertices of an underlying grid.
3
3.1
Proposed convolution operator
Definitions
Let S be a set of points, such that each one defines a set of neighbourhoods.
An entry e of a dataset D is a column vector that can be of any size, for which each dimension
represents a value taken by e at a certain point u ∈ S. u is said to be activated by e. A point u can be
associated to at most one dimension of e. If it is the ith dimension of e, then we denote the value
taken by e at u by either eu or ei . D is said to be embedded in S.
We say that two entries e and e0 are homogeneous if they have the same size and if their dimensions
are always associated to the same points.
2
3.2
Formal description of the generalized convolution
Let’s denote by C a generalized convolution operator.We want it to observe the following conditions:
• Linearity
• Locality
• Kernel weight sharing
As C must be linear, then for any entry e, there is a matrix W e such that C(e) = W e e. Unless the
entries are all homogeneous, W e depends on e. For example, in the case of regular convolutions on
images, W is a Toeplitz matrix and doesn’t depend on e.
In order to meet the locality condition, we first want that the coordinates of C(e) have a local meaning.
To this end, we impose that C(e) lives in the same space as e, and that C(e) and e are homogeneous.
Secondly, for each u, we want C(e)u to be only function of values taken by e at points contained in a
certain neighbourhood of u. It results that rows of W e are generally sparse.
Let’s attribute to C a kernel of n weights in the form of a row vector w = (w1 , w2 , .., wn ), and let’s
define the set of allocation matrices A as the set of binary matrices that have at most one non-zero
coordinate per column. As C must share its weights across the activated points, then for each row
Wie of W e , there is an allocation matrix Aei ∈ A such that Wie = wAei . To maintain locality, the jth
column of Aei must have a non-zero coordinate if and only if the ith and jth activated points are in a
same neighbourhood.
Let’s Ae denotes the block column vector that has the matrices Aei for attributes, and let’s ⊗ denotes
the tensor product. Hence, C is defined by a weight kernel w and an allocation map e 7→ Ae that
maintains locality, such that
C(e) = (w ⊗ Ae ) . e
3.3
(1)
The underlying graph that supports the generalized convolution
As vertices, the set of activated points of an entry e defines a complete oriented graph that we call Ge .
If all the entries are homogeneous, then Ge is said to be static. In this case, we note G. Otherwise,
there can be one different graph per entry so the kernel of weights can be re-used for new entries
defined on different graphs.
Suppose we are given u 7→ Vu which maps each u ∈ S to a neighbourhood Vu . We then define GeV
as the subgraph of Ge such that it contains the edge (u0 , u) if and only if u0 ∈ Vu . Let aV : e 7→ Ae
be an allocation map such that the jth column of Aei have a non-zero coordinate if and only if (ei , ej )
is an edge of GeV .
Then the generalized convolution of e by the couple (w, aV ) is supported by the underlying graph
GeV in the sense that WVe = w ⊗ aV (e) is its adjacency matrix. Note that the underlying graph of a
regular convolution is a lattice.
Also note that the family (Vu )u∈S can be seen as local receptive fields for the generalized convolution,
and that the map aV can be seen as if it were distributing the kernel weights into each Vu .
Remarks
The generalized convolution has been defined here as an operation between a kernel weight and a
vector. If G = (E, V ) denotes a graph, then such operator ∗ with respect to a third rank tensor A
could also have been defined as a bilinear operator between two graph signals f, g ∈ RE :
f ∗ g = (f > ⊗ A) . g
>
∀v ∈ V, (f ∗ g)(v) = f Av g, with Av ∈ A
(2)
(3)
Note that the underlying graph depends on the entry. So the learnt filter w is reusable regardless of
the entry’s underlying graph structure. This is not the case for convolutions defined on fixed graphs
like in [6, 7].
3
3.4
Example of a generalized convolution shaped by a rectangular window
Let E be a two-dimensional Euclidean space and let’s suppose here that S = E.
Let’s denote by CR , a generalized convolution shaped by a rectangular window R. We suppose that
its weight kernel w is of size Nw = (2p + 1)(2q + 1) and that R is of width (2p + 1)µ and height
(2q + 1)µ, where µ is a given unit scale.
Let Gep,q be the subgraph of Ge such that it contains the edge (u0 , u) if and only if |u0x −ux | ≤ (p+ 12 )µ
and |u0y − uy | ≤ (q + 12 )µ. In other terms, Gep,q connects a vertex u to every vertex contained in R
when centered on u.
Then, we define CR as being supported by Gep,q . As such, its adjacency matrix acts as the convolution
operator. At this point, we still need to affect the kernel weights to its non-zero coordinates, via
the edges of Gep,q . This amounts for defining explicitely the map e 7→ Ae for CR . To this end, let’s
consider the grid of same size as that rectangle, which breaks it down into (2p + 1)(2q + 1) squares
of side length µ, and let’s associate a different weight to each square. Then, for each edge (u0 , u), we
affect to it the weight associated with the square within which u0 falls when the grid is centered on u.
This procedure allows for the weights to be shared across the edges. It is illustrated on figure 1.
Note that if the entries are homogeneous and the activated points are vertices of a regular grid, then
the matrix W , independant of e, is a Toepliz matrix which acts as a regular convolution operator on
the entries e. In this case, CR is just a regular convolution. For example, this is the case in section 3.5.
µ
µ
(2p + 1)µ
S=E
S=E
Figure 1: Example of a moving grid. The grid
defines the allocation of the kernel weights.
3.5
Figure 2: On regular domains, the moving grid
allocates the weights similarly to the moving window of a standard convolution.
Link with the standard convolution on image datasets
Let D be an image dataset. Entries of D are homogeneous and their dimensions represent the value
at each pixel. In this case, we can set S = E, of dimension 2, such that each pixel is located at entire
coordinates. More precisely, if the images are of width n and height m, then the pixels are located at
coordinates (i, j) ∈ {0, 1, 2, ..., n}x{0, 1, 2, ..., m}. Hence, the pixels lie on a regular grid and thus
are spaced out by a constant distance µ = 1.
Let’s consider the static underlying graph Gp,q and the generalized convolution by a rectangular
window CR , as defined in the former section. Then, applying the same weight allocation strategy will
lead to affect every weight of the kernel into the moving window R. Except on the border, one and
only one pixel will fall into each square of the moving grid at each position, as depicted in figure 2.
Indeed, R behaves exactly like a moving window of the standard convolution, except that it considers
that the images are padded with zeroes on the borders.
4
4
4.1
Application to CNNs
Neural network interpretation
Let Ld and Ld+1 be two layers of neurons, such that forward-propagation is defined from Ld to
Ld+1 . Let’s define such layers as a set of neurons being located in S. These layers must contain as
many neurons as points that can be activated. In other terms, S ∼
= Ld ∼
= Ld+1 . As such, we will
abusively use the term of neuron instead of point.
Ld
Ld+1
Figure 3: Generalized convolution between two layers of neurons.
The generalized convolution between these two layers can be interpreted as follow. An entry e
activates the same N neurons in each layer. Then, a convolution shape takes N positions onto Ld ,
each position being associated to one of the activated neurons of Ld+1 . At each position, connections
are drawn from the activated neurons located inside the convolution shape in destination to the
associated neuron. And a subset of weights from w are affected to these connections, according to a
weight sharing strategy defined by an allocation map. Figure 3 illustrates a convolution shaped by a
rectangular window.
The forward and backward propagation between Ld and Ld+1 are applied using the described neurons
and connections. After a generalized convolution operation, an activation function is applied on the
output neurons. Then a pooling operation is done spatially: the input layer is divided into patches of
same size, and all activated neurons included in this patch are pooled together. Unlike a standard
pooling operation, the number of activated neurons in a patch may vary.
Generalized convolution layers can be vectorized. They can have multiple input channels and multiple
feature maps. They shall naturally be placed into the same kind of deep neural network structure than
in a CNN. Thus, they are for the irregular input spaces what the standard convolution layers are for
regular input spaces.
4.2
Implementation
There are two main strategies to implement the propagations. The first one is to start from (1), derive
it and vectorize it. It implies handling semi-sparse representations to minimize memory consumption
and to use adequate semi-sparse tensor products.
Instead, we decide to use the neural network interpretation and the underlying graph structure whose
edges amount for neurons connections. By this mean, the sparse part of the computations is included
via the use of this graph. Also, computations on each edge can be parallelized.
5
4.3
Forward and back propagation formulae
Let’s first recall the propagation formulae from a neural network point of view.
Let’s denote by eu the value of a neuron of Ld located at u ∈ S, by fv for a neuron of Ld+1 , and by
gv if this neuron is activated by the activation function σ.
We denote by prev(v) the set neurons from the previous layer connected to v, and by next(u) those of
the next layer connected to u. wuv is the weight affected to the connection between neurons u and v.
b is the bias term associated to Ld+1 .
After the forward propagation, values of neurons of Ld+1 are determined by those of Ld :
X
fv =
eu wuv
(4)
u∈prev(v)
gv = σ(fv + b)
(5)
Thanks to the chain rule, we can express derivatives of a layer with those of the next layer:
δv =
∂E
∂E ∂gv
=
=
∂fv
∂gv ∂fv
X ∂E
∂E
=
∂eu
∂fv
v∈next(u)
∂E 0
σ (fv + b)
∂gv
X
∂fv
=
δv wuv
∂eu
(6)
(7)
v∈next(u)
We call edges(w) the set of edges to which the weight w is affected. If ω ∈ edges(w), fω+ denotes
the value of the destination neuron, and eω− denotes the value of the origin neuron.
The back propagation allows to express the derivative of any weight w:
∂E
=
∂w
X
ω∈edges(w)
∂E ∂fω+
=
∂fω+ ∂w
X
δ ω + eω −
(8)
ω∈edges(w)
X ∂E ∂gv
X
∂E
=
=
δv
∂b
∂gv ∂b
v
v
(9)
The sets prev(v), next(u) and edges(w) are determined by the graph structure, which in turn is
determined beforehand by a procedure like the one described in section 3.4. The particularization of
the propagation formulae with these sets is the main difference with the standard formlulae.
4.4
Vectorization
Computations are done per batch of entries B. Hence, the graph structure used for the computations
must contain the weighted edges of all entries e ∈ B. If necessary, entries of B are made homogeneous: if a neuron u is not activated by an entry e but is activated by another entry of B, then eu is
defined and is set to zero.
The third-rank tensor counterparts of e, f , and g are thus denoted by E, F and G. Their third dimension
indexes the channels (input channels or feature maps). Their submatrix along the neuron located at
x ∈ S are denoted Ex , Fx and Gx , rows are indexing entries and columns are indexing channels. The
counterparts of w and b are W and β. The first being a third-rank tensor and the second being a vector
with one value per feature map. βe denotes the 3D tensor obtained by broadcasting β along the two
other dimensions of F. Ww denotes the submatrix of W along the kernel weight w. Its rows index the
feature maps and its columns index the input channels.
With these notations, the convolution formulae (1) rewrites:
C(E) = (W ⊗ AE ) E
(10)
⊗ and being tensor products. ⊗ is contracted along the dimensions that index the kernel weights,
and is contracted along the dimensions that index the points present in the entries.
6
The vectorized counterparts of the formulae from section 4.3 can be obtained in the same way:
X
(11)
Fv =
Eu W>
wuv
u∈prev(v)
e
G = σ(F + β)
0
∂E
∂E
e
∆=
=
◦ σ (F + β)
∂F
∂G
X
∂E
∆v Wuv
=
∂E u
v∈next(u)
X
∂E
=
∆>
ω + Eω −
∂W w
ω∈edges(w)
!
X X
∂E
=
∆v
∂β
v
j
(12)
(13)
(14)
(15)
(16)
j-th column
Where ◦ and > respectively denotes Hadamard product and transpose.
5
Experiments
In order to measure the gain of performance allowed by the generalized CNN over MLP on irregular
domains, we made a series of benchmarks on distorded versions of the MNIST dataset [10], consisting
of images of 28x28 pixels. To distort the input domain, we plunged the images into a 2-d euclidean
space by giving entire coordinates to pixels. Then, we applied a gaussian displacement on each
pixel, thus making the data irregular and unsuitable for regular convolutions. For multiple values of
standard deviation of the displacement, we trained a generalized CNN and compared it with a MLP
that has the same number of parameters. We choosed a simple yet standard architecture in order to
better see the impact of generalized layers.
The architecture used is the following: a generalized convolution layer with relu [11] and max pooling,
made of 20 feature maps, followed by a dense layer and a softmax output layer. The generalized
convolution is shaped by a rectangular window of width and height 5µ where the unit scale µ is
chosen to be equal to original distance between two pixels. The max-pooling is done with square
patches of side length 2µ. The dense layer is composed of 500 hidden units and is terminated by
relu activation as well. In order to have the same number of parameters, the compared MLP have 2
dense layers of 500 hidden units each and is followed by the same output layer. For training, we used
stochastic gradient descent [12] with Nesterov momentum [13] and a bit of L2 regularization [14].
2.2
2
MLP
G-CNN
Error
1.8
1.6
1.4
1.2
CNN
1
+
0
1
2
3
4
5
6
10−1
100
101
102
Standard deviation used for the domain distortion
Figure 4: Results for Generalized CNN and MLP
The plot drawn on figure 4 illustrates the gain of performance of a generalized convolutional layer
over a dense layer with equal number of parameters. After 15 epochs for both, it shows that the
7
generalized CNN on a distorded domain performs better than the MLP. Indeed, in the case of no
domain distortion, the score is the same than a CNN with zero-padding. The error rate goes up a bit
with distortion. But even at 200µ, the generalized CNN still performs better than the MLP.
6
Conclusion and future work
In this paper, we have defined a generalized convolution operator. This operator makes possible to
transport the CNN paradigm to irregular domains. It retains the proprieties of a regular convolutional
operator. Namely, it is linear, supported locally and uses the same kernel of weights for each local
operation. The generalized convolution operator can then naturally be used instead of convolutional
layers in a deep learning framework. Typically, the created model is well suited for input data that
has an underlying graph structure.
The definition of this operator is flexible enough for it allows to adapt its weight-allocation map to
any input domain, so that depending on the case, the distribution of the kernel weight can be done in
a way that is natural for this domain. However, in some cases, there is no natural way but multiple
acceptable methods to define the weight allocation. In further works, we plan to study these methods.
We also plan to apply the generalized operator on unsupervised learning tasks.
Acknowledgements
I would like to thank my academic mentors, Vincent Gripon and Grégoire Mercier who helped me
in this work, as well as my industrial mentor, Mathias Herberts who gave me insights in view of
applying the designed model to industrial datasets.
This work was partly funded by Cityzen Data, the company behind the Warp10 platform, and by
the ANRT (Agence Nationale de la Recherche et de la Technologie) through a CIFRE (Convention
Industrielle de Formation par la REcherche), and also by the European Research Council under the
European Union’s Seventh Framework Program (FP7/2007-2013) / ERC grant agreement number
290901.
References
[1] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document
recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional
neural networks,” in Advances in neural information processing systems, pp. 1097–1105, 2012.
[3] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal
approximators,” Neural networks, vol. 2, no. 5, pp. 359–366, 1989.
[4] F. R. K. Chung, Spectral Graph Theory (CBMS Regional Conference Series in Mathematics,
No. 92). American Mathematical Society, 1996.
[5] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field
of signal processing on graphs: Extending high-dimensional data analysis to networks and other
irregular domains,” IEEE Signal Processing Magazine, vol. 30, no. EPFL-ARTICLE-189192,
pp. 83–98, 2013.
[6] J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun, “Spectral networks and locally connected
networks on graphs,” arXiv preprint arXiv:1312.6203, 2013.
[7] M. Henaff, J. Bruna, and Y. LeCun, “Deep convolutional networks on graph-structured data,”
arXiv preprint arXiv:1506.05163, 2015.
[8] J. Masci, D. Boscaini, M. Bronstein, and P. Vandergheynst, “Shapenet: Convolutional neural
networks on non-euclidean manifolds,” tech. rep., 2015.
[9] B. Graham, “Spatially-sparse convolutional neural networks,” arXiv preprint arXiv:1409.6070,
2014.
[10] Y. LeCun, C. Cortes, and C. J. Burges, “The mnist database of handwritten digits,” 1998.
[11] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural networks,” in International
Conference on Artificial Intelligence and Statistics, pp. 315–323, 2011.
8
[12] L. Bottou, “Large-scale machine learning with stochastic gradient descent,” in Proceedings of
COMPSTAT’2010, pp. 177–186, Springer, 2010.
[13] I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the importance of initialization and
momentum in deep learning,” in Proceedings of the 30th international conference on machine
learning (ICML-13), pp. 1139–1147, 2013.
[14] A. Y. Ng, “Feature selection, l 1 vs. l 2 regularization, and rotational invariance,” in Proceedings
of the twenty-first international conference on Machine learning, p. 78, ACM, 2004.
9
| 7cs.IT
|
Distributed Average Tracking for Multiple Signals Generated
by Linear Dynamical Systems: An Edge-based Framework
Yu Zhao a,⋆ , Yongfang Liu a, Zhongkui Li b , Zhisheng Duan b
arXiv:1606.09329v1 [cs.SY] 30 Jun 2016
a
b
Department of Traffic and Control Engineering, School of Automation, Northwestern Polytechnical University, Xi’an
Shaanxi, 710129, China
State Key Laboratory for Turbulence and Complex Systems, Department of Mechanics and Engineering Science, College of
Engineering, Peking University, Beijing 100871, China
Abstract
This paper studies the distributed average tracking problem for multiple time-varying signals generated by linear dynamics,
whose reference inputs are nonzero and not available to any agent in the network. In the edge-based framework, a pair of
continuous algorithms with, respectively, static and adaptive coupling strengths are designed. Based on the boundary layer
concept, the proposed continuous algorithm with static coupling strengths can asymptotically track the average of multiple
reference signals without the chattering phenomenon. Furthermore, for the case of algorithms with adaptive coupling strengths,
average tracking errors are uniformly ultimately bounded and exponentially converge to a small adjustable bounded set.
Finally, a simulation example is presented to show the validity of theoretical results.
Key words: Distributed control, average tracking, linear dynamics, continuous algorithm.
1
Introduction
In the past two decades, there have been lots of interests in the distributed cooperative control [1], [2],
[3], [4], [5], [6], [7], [8], [9], [10], [11], [12], and [13], for
multi-agent systems due to its potential applications in
formation flying, path planning and so forth. Besides,
the clock synchronization problems were also discussed
in [31,32,33,34,35], which are very important to design
distributed algorithms. Distributed average tracking,
as a generalization of consensus and cooperative tracking problems, has received increasing attentions and
been applied in many different perspectives, such as
distributed sensor networks [14], [15] and distributed
coordination [16], [17]. For practical applications, distributed average tracking should be investigated for
signals modeled by more and more complex dynamical
systems.
The objective of distributed average tracking problems is
to design a distributed algorithm for multi-agent systems
to track the average of multiple reference signals. The
motivation of this problem comes from the coordinated
tracking for multiple camera systems. Spurred by the pioneering works in [18], and [19] on the distributed aver⋆ Corresponding author
Email address: [email protected] (Yu Zhao ).
Preprint submitted to Automatica
age tracking via linear algorithms, real applications of
related results can be found in distributed sensor fusion
[14], [15], and formation control [16]. In [20], distributed
average tracking problems were investigated by considering the robustness to initial errors in algorithms. The
above-mentioned results are important for scientific researchers to build up a general framework to investigate
this topic. However, a common assumption in the above
works is that the multiple reference signals are constants
[19] or achieving to values [18]. In practical applications,
reference signals may be produced by more general dynamics. For this reason, a class of nonlinear algorithms
were designed in [21] to track multiple reference signals
with bounded deviations. Then, based on non-smooth
control approaches, a couple of distributed algorithms
were proposed in [22] and [23] for agents to track arbitrary time-varying reference signals with bounded deviations and bounded second deviations, respectively. Using discontinuous algorithms, further, [24] studied the
distributed average tracking problems for multiple signals generated by linear dynamics.
Motivated by the above mentioned observations, this
paper is devoted to solving the distributed average
tracking problem with continuous algorithms, for multiple time-varying signals generated by general linear
dynamical systems, whose reference inputs are assumed
to be nonzero and not available to any agent in networks. First of all, based on relative states of neigh-
1 July 2016
notation i ∼ j is used to denote that node i is connected
to node j, or equivalently, (i, j) ∈ E. We make use of
the |V| × |E| incidence matrix, D(G), for a graph with
an arbitrary orientation, i.e., a graph whose edges have
a head (a terminal node) and a tail (an initial node).
The columns of D(G) are then indexed by the edge set,
and the ith row entry takes the value 1 if it is the initial
node of the corresponding edge, −1 if it is the terminal
node, and zero otherwise. The diagonal matrix ∆(G)
of the graph contains the degree of each vertex on its
diagonal. The adjacency matrix, A(G), is the |V| × |V|
symmetric matrix with zero in the diagonal and one in
the (i, j)th position if node i is adjacent to node j. The
graph Laplacian [25] of G, L := 12 D(G)D(G)T = ∆(G) −
A(G), is a rank deficient positive semi-definite matrix.
boring agents, a class of distributed continuous control
algorithms are proposed and analyzed. Then, a novel
class of distributed algorithms with adaptive coupling
strengths are designed by utilizing an adaptive control
technique. Different from [4] and [5], where the nonlinear
signum function was applied to the whole neighborhood
(node-based algorithm), the proposed algorithms in this
paper are designed along the edge-based framework as
in [22], [23] and [24]. Compared with the above existing
results, the contributions of this paper are three-fold.
First, main results of this paper extend the reference
signals which were generated by first and second-order
integrators in [22] and [23], respectively, to signals generated by linear dynamical systems, which can describe
more complex signals. An advantage of edge-based
algorithms designed here is that they have a certain
symmetry, which is very important to get the average
value of multiple signals under an undirected topology.
By utilizing this property, the edge-based algorithms
obtained in this paper successfully solve distributed
average tracking problems for multiple signals generated by general linear systems with bounded inputs.
Second, by using adaptive control approaches, the requirements of all global information are removed, which
greatly reduce the computational complexity for largescale networks. Third, compared with existing results
in [24], new continuous algorithms are redesigned via
the boundary layer concept with clock synchronization
devices. Since there exist differences between the local
times of the agents, which may effect the distributed
average tracking result, the clock synchronization is introduced in this paper. The clock synchronization problem has been solved in many existing papers such as
[31,32,33,34,35]. With the help of the existing results on
clock synchronization in [31,32,33,34,35], the first step
before beginning computation is to set the local clock to
synchronize the local times. Thus, the boundary layer
concept with clock synchronization devices plays a vital
role to reduce the chattering phenomenon. Continuous
algorithms in this paper is more appropriate for real
engineering applications.
n
An undirected path between node i1 and node is on undirected graph means a sequence of ordered undirected
edges with the form (ik ; ik+1 ), k = 1, · · · , s − 1. A graph
G is said to be connected if there exists a path between
each pair of distinct nodes.
Assumption 1 Graph G is undirected and connected.
Lemma 1 [25] Under Assumption 1, zero is a simple
eigenvalue of L with 1 as an eigenvector and all the other
eigenvalues are positive. Moreover, the smallest nonzero
xT Lx
eigenvalue λ2 of L satisfies λ2 =
min
xT x .
x6=0,1T x=0
Define M = IN − N1 11T . Then M satisfies following
properties: Firstly, it is easy to see that 0 is a simple eigenvalue of M with 1 as the corresponding right
eigenvector and 1 is the other eigenvalue with multiplicity N − 1, i.e., M 1 = 1T M = 0. Secondly, since
LT = L, one has LM = L(IN − N1 11T ) = L− N1 L11T =
L = L − N1 11T L = (IN − N1 11T )L = M L. Finally,
M 2 = M (IN − N1 11T ) = M − N1 M 11T = M .
3
n×n
Notations: Let R and R
be sets of real numbers and real matrices, respectively. In represents
the identity matrix of dimension n. Denote by 1
a column vector with all entries equal to one. The
matrix inequality A > (≥)B means that A − B is
positive (semi-) definite. Denote by A ⊗ B the Kronecker product of matrices A and B. For a vector
x = (x1 , x2 , · · · , xn )T ∈ Rn , let kxk denote the 2-norm
1
1
1
1
of x, sig 2 (x) = (sig 2 (x1 ), sig 2 (x2 ), · · · , sig 2 (xn ))T . For
a set V , |V | represents the number of elements in V .
2
2.1
Distributed average tracking for multiple reference signals with general linear dynamics
Suppose that there are N time-varying reference signals,
ri (t) ∈ Rn , i = 1, 2, · · · , N , which generated by the following linear dynamical systems:
ṙi (t) = Ari (t) + Bfi (t),
(1)
where A ∈ Rn×n and B ∈ Rn×p both are constant matrices with compatible dimensions, ri (t) ∈ Rn is the
state of the ith signal, and fi (t) ∈ Rp represents the
reference input of the ith signal. Here, we assume that
fi (t) is continuous and bounded, i.e., kfi (t)k ≤ f0 , for
i = 1, 2, · · · , N , where f0 is a positive constant. Suppose
that there are N agents with xi ∈ Rn being the state
of the ith agent in distributed algorithms. It is assumed
that agent i has access to ri (t), and agent i can obtain
the relative information from its neighbors denoted by
Ni . Besides, let |Ni | represent the number of elements
in the set Ni , i = 1, 2, · · · , N .
Preliminaries
Graph Theory
An undirected (simple) graph G is specified by a vertex
set V and an edge set E whose elements characterize
the incidence relations between distinct pairs of V. The
2
Assumption 2 (A, B) is stabilizable.
function
The main objective of this paper is to design a class of
distributed algorithms for agents to track the average
of multiple signals ri (t) generated by the general linear
dynamics (1) with bounded reference inputs fi (t), i =
1, 2, · · · , N .
b
hi (ω) =
j∈Ni
j∈Ni
+c2 B
(2)
ω
,
kωk + εe−ϕti
j∈Ni
X
j∈Ni
[K(xi (t) − xj (t))]
hi [K(xi (t) − xj (t)), ti ] + Bfi (t). (5)
Before moving on, an important lemma is proposed.
Lemma 2 Under Assumption 1, states xi (t) in (2) with
si (t0 ) = 0 will track the average of multiple signals, i.e.,
PN
kxi (t)− N1
k=1 rk (t)k = 0, if the closed-loop system (5)
PN
achieves consensus, i.e., limt→∞ kxi − N1 k=1 xk k = 0
for i = 1, 2, · · · , N .
(3)
with a finite-time clock synchronization device:
X
1
dti
sig 2 (ti − tj ),
=1+
dt
X
j∈Ni
where si (t), i = 1, 2, · · · , N , are internal states of the
distributed filter (2), c1 , c2 and K are coupling strengths
and feedback gain matrix, respectively, to be determined. Besides, the nonlinear function hi (·) is defined
as follows: for ω ∈ Rn ,
hi (ω, ti ) =
if ω = 0.
ẋi (t) = Axi (t) + c1 B
hi [K(xi (t) − xj (t)), ti ],
xi (t) = si (t) + ri (t),
0
if ω 6= 0,
It follows from (1) and (2) that the closed-loop system
is described by
ṡi (t) = Asi (t) + Bui (t),
X
ui (t) = c1
[K(xi (t) − xj (t))]
+c2
ω
kωk
The item εe−ϕti in (3) defines the size of the boundary
layer. As ti → ∞, the continuous function hi (ω, ti ) approaches the discontinuous function b
hi (ω).
Therefore, a distributed algorithm is proposed as follows:
X
(
Proof : It follows from Assumption 1 and Remark 1 that
i = 1, 2, · · · , N,
(4)
N X
X
i=1 j∈Ni
where ε and ϕ are positive constants, ti is a local time
of the local clock in agent i.
N X
X
i=1 j∈Ni
Remark 1 Besides (4), there exist many algorithms on
designing clock synchronization device in [31,32,33,34,35],
where the robust finite-time clock synchronization device is considered in [33]. Thus, by using the nonlinear
function as given in (3) with (4), the first step before beginning computation is to set the local clock by using the
clock synchronization device (4) such that all local times
ti to be identical in a finite settling time t0 . Then, the
Algorithm (2) can solve the distributed average tracking
problem without errors.
[K(xi (t) − xj (t))] = 0,
hi [K(xi (t) − xj (t)), ti ] = 0, t ≥ t0 .
(6)
P
PN
Let s(t) = N
i=1 ri (t). From (1), (5) and
i=1 xi (t) −
(6), we have
ṡ(t) = As(t), t ≥ t0 ,
(7)
with s(t0 ) = 0. By solving the differential equation (7) with initial condition above, we always have
limt→∞ s(t) = limt→∞ eA(t−t0 ) s(t0 ) = 0. Thus, we
obtain
Remark 2 In practice, there exists external disturbance, which may result the failure of the clock synchronization device (4). For this case, we can use the nonlinear function (3) with ϕ = 0. Then, the Algorithm (2)
can solve the distributed average tracking problems with
a uniformly ultimately bounded error. As well known,
the bounded result is significant in real application.
lim
t→∞
N
X
i=1
xi (t) = lim
t→∞
N
X
ri (t).
(8)
i=1
According to Assumption 1, if xi (t) in (5) achieves conPN
sensus, i.e., limt→∞ kxi (t) − N1
k=1 xk (t)k = 0 for i =
1, 2, · · · , N , then, it follows from (8) that limt→∞ kxi −
1 PN
k=1 rk (t)k = 0, for i = 1, 2, · · · , N . This completes
N
the proof.
Note that the nonlinear function hi (ω, ti ) in (3) is continuous, which is actually continuous approximations,
via the boundary layer concept [26], of the discontinuous
3
Remark 3 In the proof of Lemma 2, it requires that
si (t0 ) = 0, which is a necessary condition to draw conclusions, when A is not asymptotically stable. In the case
that A is asymptotically stable, without requiring the initial condition si (t0 ) = 0, we can still reach the same conclusions as shown in Lemma 2, since the solution of (7)
will converge to the origin for any initial condition.
Proof : Consider the Lyapunov function candidate
Let x(t) = (xT1 (t), xT2 (t), · · · , xTN (t))T , and F (t) =
T
(f1T (t), f2T (t), · · · , fN
(t))T . Define ξ(t) = (M ⊗ I)x(t),
T
T
where ξ(t) = (ξ1 (t), ξ2T (t), · · · , ξN
(t))T . Then, it follows
that ξ(t) = 0 if and only if x1 (t) = x2 (t) = · · · = xN (t).
Therefore, the consensus problem of (5) is solved if and
only if ξ(t) asymptotically converges to zero. Hereafter,
we refer to ξ(t) as the consensus error. By noting that
LM = L and M D(G) = D(G), it is not difficult to
obtain from (5) that the consensus error ξ(t) satisfies
V1 (t) ≥ λmin (P )kξk2 ,
˙ = (I ⊗ A)ξ(t) + c1 (L ⊗ BK)ξ(t)
ξ(t)
P
h1 [K(ξ1 (t) − ξj (t)), t1 ]
B
j∈N1
..
+c2
.
P
hN [K(ξN (t) − ξj (t)), tN ]
B
V1 (t) = ξ T (M ⊗ P )ξ.
By the definition of ξ(t), it is easy to see that (1T ⊗
I)ξ = 0. For the connected graph G, it then follows from
Lemma 1 that
V̇1 = ξ˙T (M ⊗ P )ξ + ξ T (M ⊗ P )ξ˙
= ξ T (I ⊗ AT + c1 L ⊗ K T B T )(M ⊗ P )ξ
+ξ T (M ⊗ P )(I ⊗ A + c1 L ⊗ BK)ξ
P
h1 [K(ξ1 (t) − ξj (t)), t1 ]
PB
j∈N1
..
T
+2c2 ξ
.
P
hN [K(ξN (t) − ξj (t)), tN ]
PB
j∈NN
+2ξ T (M ⊗ P B)F (t).
(9)
(13)
Substituting K = −B T P into (13), it follows from the
fact LM = M L = L that
V̇1 = ξ T (M ⊗ (P A + AT P ) − 2c1 L ⊗ P BB T P )ξ
P
h1 [B T P (ξ1 (t) − ξj (t)), t1 ]
PB
j∈N1
..
−2c2 ξ T
.
P
T
hN [B P (ξN (t) − ξj (t)), tN ]
PB
Algorithm 1: Under Assumptions 1 and 2, for multiple
reference signals in (1), the distributed average tracking
algorithm (2) can be constructed as follows
(1) Set the local clock such that the synchronization of
the local time ti in finite time by using the clock
synchronization device (4).
(2) Solve the algebraic Ricatti equation (ARE):
P A + AT P − P BB T P + Q = 0,
(12)
where λmin (P ) is the smallest eigenvalue of the positive
matrix P . The time derivative of V1 along (9) can be
obtained as follows
j∈NN
+(M ⊗ B)F (t).
(11)
j∈NN
T
+2ξ (M ⊗ P B)F (t).
Since kF k ≤
(10)
(14)
√
N f0 , we have
ξ T (M ⊗ P B)F (t)
≤ k(M ⊗ B T P )ξkkF (t)k
N
N
f0 X X
kB T P (ξi − ξj )k
≤√
N i=1 j=1,j6=i
X
N
N
f0 X
≤√
kB T P (ξi − ξj )k
max
N i=1 i
j=1,j6=i
X
N
√
kB T P (ξi − ξj )k
= N f0 max
with Q > 0 to obtain a matrix P > 0. Then, choose
K = −B T P .
(3) Select the first coupling strength c1 ≥ 2λ1 2 , where λ2
is the smallest nonzero eigenvalue of the Laplacian
L of G.
(4) Choose the second coupling strength c2 ≥ f0 (N −
1), where f0 is defined as in (1).
It is worthwhile to mention that the originality of the
Riccati based approach in step (2) in Algorithm 1 for
the design of matrix K can be found in [6] and [7].
i
Theorem 1 Under Assumptions 1 and 2, the state xi (t)
in (2) will track the average of multiple reference signals
ri (t), i = 1, 2, · · · , N , generated by the general linear
dynamics (1) with bounded reference inputs if coupling
strengths c1 , c2 and the feedback gain K are designed by
Algorithm 1.
≤
j=1,j6=i
N X
√ X
f0
kB T P (ξi − ξj )k.
(N − 1) N
2
i=1
Then, because of the fact that ω T hi (ω, ti ) =
4
(15)
j∈Ni
kωk2
kωk+εe−ϕti
,
we get
T
−2c2 ξ
= −c2
P
PB
j∈N1
PB
P
j∈NN
N X
X
i=1 j∈Ni
h1 [B T P (ξ1 (t) − ξj (t)), t1 ]
..
.
hN [B T P (ξN (t) − ξj (t)), tN ]
kB T P (ξi − ξj )k2
.
kB T P (ξi − ξj )k + εe−ϕti
where γ =
λmin (Q)
λmax (P ) .
0≤V1 (t)≤e
−γt
Thus, we obtain that
V1 (t0 )+c2
N
X
i=1
|Ni |
Zt
εe−γ(t−τ )−ϕ(τ +η) dτ,
t0
where η = π(t0 ) is a constant. By noting that
Zt
(16)
εe−γ(t−τ )−ϕτ dτ
t0
(
ε(t − t0 )e−γt
ε
−(ϕ+γ)t
− e−γt−ϕt0 )
γ−ϕ (e
if γ = ϕ,
By combining with (15) and (16), it follows from (14)
that
=
V̇1 ≤ ξ T (M ⊗(P A+AT P )−2c1 L⊗P BB T P )ξ
N X
√ X
kB T P (ξi − ξj )k
+f0 (N − 1) N
we have that V1 (t) will converge to the origin as t → ∞,
which means that states of (5) will achieve consensus.
Then, according to Lemma 2, we have that tracking errors
limt→∞ξi (t) =
ξi , i = 1, 2, · · · , N satisfy
P
N
= limt→∞ xi (t) −
limt→∞ xi (t) − N1
k=1 xk (t)
1 PN
k=1 rk (t) = 0. Therefore, the distributed average
N
i=1 j∈Ni
−c2
N
X
X
i=1 j∈Ni
kB T P (ξi − ξj )k2
.
kB T P (ξi − ξj )k + εe−ϕti
(17)
√
Choose c2 ≥ f0 (N − 1) N . We have
T
T
tracking problem is solved. This completes the proof.
Remark 4 As mentioned in Remark 1, for the case with
external disturbance, let ϕ = 0. Therefore, the nonlinear
ω
. From the (20),
function (3) is reduced to hi (ω) = kωk+ε
PN
one has V̇1 < −γV1 + c2 i=1 |Ni |ε. Then, the tracking
error ξ given in (9) uniformly ultimately bounded. According to (12), ξ will exponentially converge to the fol
21
PN
c2
lowing set Ω0 , ξ : kξk < γλmin
.
|N
|ε
i
i=1
(P )
T
V̇1 ≤ ξ (M ⊗(P A+A P )−2c1 L⊗P BB P )ξ
N X
X
+c2
kB T P (ξi − ξj )k
i=1 j∈Ni
T
kB P (ξi − ξj )k2
kB T P (ξi − ξj )k + εe−ϕti
T
≤ ξ (M ⊗(P A+AT P )−2c1 L⊗P BB T P )ξ
N
X
|Ni |εe−ϕti .
+c2
−
The bounded result is meaningful in real application.
(18)
In distributed algorithm (2), it requires the initial state
of si (t) satisfying xi (t0 ) = ri (t0 ). In order to remove the
initial condition xi (t0 ) = ri (t0 ), a modified algorithm is
proposed as follows:
i=1
By Assumption 1, there exists an unitary matrix U such
that L = U T ΛU , where Λ = diag(λ1 , λ2 , · · · , λN ). Without loss of generality, assume that 0 = λ1 < λ2 ≤ · · · ≤
λN . Thereby, following from the fact that M 2 = M , we
obtain
ξ T (M ⊗(P A+AT P )−2c1 L⊗P BB T P )ξ
= ξ T (M U T ⊗ I)[I ⊗ (P A + AT P )
−2c1 Λ ⊗ P BB T P ](U M ⊗ I)ξ
≤ ξ T (M ⊗ (P A + AT P − 2c1 λ2 P BB T P ))ξ.
ṡi (t) = Asi (t) + Bui (t),
X
ui (t) = Kxi (t) + c2
hi [K(xi (t) − xj (t)), ti ],
j∈Ni
xi (t) = si (t) + ri (t).
N
X
i=1
|Ni |εe−ϕti ,
(21)
Corollary 1 By using the modified robust algorithm
(21) with steps (1), (2) and (4) in Algorithm 1, the
distributed average tracking can be achieved without
requiring the initial condition xi (t0 ) = ri (t0 ).
(19)
Select c1 ≥ 2λ1 2 . It follows from (10) that P A + AT P −
2c1 λ2 P BB T P ≤ −Q. Therefore, we have
V̇1 < −γV1 + c2
if γ 6= ϕ,
Proof : First of all, let xi (t) = si (t) + ri (t). From (21),
one has the closed-loop system of (1) and (21) is described by
(20)
ẋi (t) = (A + BK)xi (t)
5
+c2 B
X
j∈Ni
under an undirected topology. By utilizing the symmetry in the edge-based framework, the algorithm (2) is designed. Different from node-based algorithms in [3]-[7],
which can not solve average tracking problems, the edgebased algorithm in this paper can ensure the state of each
agent to track the average value of multiple signals. Besides, in [5], it studied consensus problems of multiple
linear systems with discontinuous algorithms. The discontinuous algorithm can not be realized in practical applications for its large chattering. In order to reduce the
chattering effect, by using the boundary layer approximation, continuous algorithms are proposed in this paper.
Compared with the result in [5], the main contribution of
this paper lies to the feasibility of continuous algorithms
in practical applications.
hi [K(xi (t) − xj (t)), ti ] + Bfi (t). (22)
Then, in the matrix form, let ξ = (M ⊗ I)x. The error
system is given as follows:
˙ = [I ⊗ (A + BK)]ξ(t)
ξ(t)
P
h1 [K(ξ1 (t) − ξj (t)), t1 ]
B
j∈N1
..
−c2
.
P
hN [K(ξN (t) − ξj (t)), tN ]
B
j∈NN
+(M ⊗ B)F (t).
(23)
Consider the same Lyapunov function candidate in (11).
One has
4
V̇1 = ξ T [M ⊗ (P A + AT P − P BB T P )]ξ
P
h1 [B T P (ξ1 (t) − ξj (t)), t1 ]
PB
j∈N1
..
T
−2c2 ξ
.
P
T
hN [B P (ξN (t) − ξj (t)), tN ]
PB
Note that in the above section, the first coupling strength
c1 , designed as c1 > 2λ1 2 , depends on the communication
topology. The second
coupling strength c2 , designed as
√
c2 > f0 (N − 1) N , requires f0 and N . Generally, the
smallest nonzero eigenvalue λ2 , the number N of vertex
set V and the upper bound f0 of fi (t) all are global information, which are difficult to be obtained by agents when
the scale of the network is very large. Therefore, to overcome these restrictions, a distributed average tracking
algorithm with distributed adaptive coupling strengths
is proposed as follows:
j∈NN
T
+2ξ (M ⊗ P B)F (t).
(24)
Similar to the proof of (15)-(18) in Theorem 1, one has
V̇1 = ξ T [M ⊗ (P A + AT P − P BB T P )]ξ
N
X
|Ni |εe−ϕti .
+c2
ṡi (t) = Asi (t) + Bui (t),
X
ui (t) =
αij (t)[K(xi (t) − xj (t))]
(25)
i=1
j∈Ni
+B
From (10), one obtains the same result in Theorem
P
1. Thus, limt→∞ kxi (t) − N1 N
k=1 xk (t)k = 0 for
i = 1, 2, · · · , N .
Second, let s(t) =
PN
i=1
xi (t) −
ṡ(t) = (A + BK)s(t), t > t0 .
PN
i=1 ri (t).
Distributed average tracking with distributed
adaptive coupling strengths
X
j∈Ni
βij (t)hi [K(xi (t) − xj (t)), ti ],
xi (t) = si (t) + ri (t), si (t0 ) = 0,
(27)
with distributed adaptive laws
We have
α̇ij (t) = µ[−ϑαij (t)
+(xi (t) − xj (t))T Γ(xi (t) − xj (t))],
"
(26)
By solving the differential equation (26) with A+BK being asymptotically stable, we always have limt→∞ s(t) =
PN
0. Thus, we obtain limt→∞ i=1
PN
xi (t) = limt→∞ i=1 ri (t). It follows that limt→∞ kxi (t)−
PN
PN
1
1
k=1 rk (t)k = limt→∞ kxi (t) − N
k=1 xk (t)k = 0,
N
for i = 1, 2, · · · , N . This completes the proof.
β̇ij (t) = ν − χβij (t)
#
kK(xi (t) − xj (t))k2
,
+
kK(xi (t) − xj (t))k + εe−ϕti
(28)
where αij (t) and βij (t) are two adaptive coupling
strengths satisfying αij (t0 ) = 0 and βij (t0 ) = 0,
Γ ∈ Rn×n is a constant gain matrix, µ, ν, ϑ and χ are
positive constants.
Remark 5 It is worth mentioning that, different from
the consensus problem in existing papers [3]-[7], where
algorithms were designed in node-based viewpoints, one
advantage of edge-based algorithms designed here is that
they have a certain symmetry in networks, which are very
important to get the average value of the multiple signals
It follows from (1) and (27) that the closed-loop system
6
is described by
verge to the following bounded set
ẋi (t) = Axi (t) + B
+B
X
X
j∈Ni
αij (t)[K(xi (t) − xj (t))]
Ω1 ,
(29)
where δ ≤ min{γ, µϑ, νχ}, α and β are two constants,
Similarly as in the above section, the following lemma is
firstly given.
V2 = ξ T (M ⊗ P )ξ
N X
X
α
eij (t)2 βeij (t)2
,
+
+
2µ
2ν
i=1
Lemma 3 Under Assumption 1, for the algorithm
PN
(27) with (28), if limt→0 kxi − N1 k=1 xk k = 0, i =
P
N
1, 2, · · · , N , then limt→∞ kxi − N1
k=1 rk k = 0, i =
1, 2, · · · , N .
i=1 j∈Ni
(31)
j∈Ni
α
eij (t) = αij (t)−α, βeij (t) = βij (t)−β, α ≥ 2λ1 2 and
β ≥ f0 (N − 1).
(2) If select ϑ and χ small enough, such that ̺ ,
max{µϑ, νχ} < γ, tracking errors ξ will exponentially converge to the bounded set Ω2 given as
follows:
Proof : Since αij (t0 ) = 0 and βij (t0 ) = 0, it follows
from (28) that αij (t) = αji (t) and βij (t) = βji (t). From
Assumption 1, we have
N X
X
(30)
j∈Ni
where αij (t) and βij (t) are given by (28).
i=1 j∈Ni
ξ, α
eij (t), βeij (t) :
)
2
N
1 X X α2
β
,
V2 <
ϑ
+χ
δ i=1
2
2
βij (t)hi [K(xi (t)−xj (t)), ti ]+Bfi (t),
j∈Ni
N X
X
(
αij (t)[K(xi (t) − xj (t))] = 0,
Ω2 ,
βij (t)hi [K(xi (t) − xj (t)), ti ] = 0, t > t0 .
Similar to the proof of Lemma 2, we can draw the conclusion in (8). This completes the proof.
ξ : kξk ≤
X
N
2
ϑα2 + χβ
|Ni |
2λmin (P )(γ − ̺)
i=1
12
,
(32)
where γ is defined in (20).
Algorithm 2: For multiple reference signals in (1), the
distributed average tracking algorithm (27) with adaptive laws (28) can be constructed as follows
Proof : Consider the Lyapunov function candidate V2
in (31). As shown in the proof of Theorem 1, the time
derivation of V2 along (28) and (29) satisfies
(1) Set the local clock such that the synchronization of
the local time ti in finite time by using the clock
synchronization device (4).
(2) Solve the ARE (10) with Q > 0 to obtain a matrix P > 0. Then, choose Γ = P BB T P and K =
−B T P , respectively.
(3) Select µ and ν small enough, respectively, such that
min (Q)
̺ , max{µϑ, νχ} < γ, where γ = λλmax
(P ) .
V̇2 ≤ ξ T [M ⊗ (P A + AT P )]ξ
N X
X
−
αij (t)(ξi − ξj )T P BB T P (ξi − ξj )
i=1 j∈Ni
N X
X
√
f0 (N − 1) N kB T P (ξi − ξj )k
+
The following theorem shows the ultimate boundedness
of tracking errors and adaptive coupling strengths.
i=1 j∈Ni
kB T P (xi (t)−xj (t))k2
−βij (t) T
kB P (xi (t)−xj (t))k+εe−ϕti
N
1XX
α
eij (t)α̇ij (t)
+
µ i=1
Theorem 2 Under the Assumption 1, the fully distributed average tracking problem is solved by (27) with
(28) if feedback gains Γ and K are designed as given
in Algorithm 2. The tracking error ξ defined in (9) and
adaptive gains αij (t) and βij (t) are uniformly ultimately
bounded and following statements are hold:
j∈Ni
+
(1) For any ϑ and χ, ξ, α
eij and βeij exponentially con7
1
ν
N
X
X
i=1 j∈Ni
βeij (t)β̇ij (t).
(33)
√
1) N , we have
By using Γ = P BB T P , it follows from (28) that
−
+
N X
X
i=1 j∈Ni
T
V̇2 ≤ −ξ T (M ⊗(P A + AT P −P BB T P ))ξ
2
N
N
α2
X
X
β
+β
+χ
|Ni | ϑ
|Ni |εe−ϕti +
2
2
i=1
i=1
!
N
XX
α
eij (t)2
βeij (t)2
ϑ
−
.
+χ
2
2
i=1
T
αij (t)(ξi − ξj ) P BB P (ξi − ξj )
N
1XX
α
eij (t)α̇ij (t)
µ i=1
j∈Ni
= −2αξ T (L⊗P BB T P )ξ−ϑ
T
T
N X
X
(e
αij (t)2 +e
αij (t)α)
Since δ ≤ min{γ, µϑ, νχ}, we obtain that
i=1 j∈Ni
≤ −2αξ (L⊗P BB P )ξ
N X 2
X
eij (t)2
α α
,
−
+ϑ
2
2
i=1
V̇2 ≤ −δV2 +
(34)
j∈Ni
+β
and
N
X
i=1
−
+
N X
X
i=1 j∈Ni
1
ν
= −β
−χ
≤ −β
+χ
N
X
kB T P (xi (t) − xj (t))k2
βij (t) T
kB P (xi (t) − xj (t))k + εe−ϕti
X
i=1 j∈Ni
N X
X
N X
X
i=1 j∈Ni
N X
X
+β
"
εe−δ(t−τ )−ϕ(τ +η) dτ
0
2
α2
1
β
+
|Ni | ϑ
.
+χ
δ i=1
2
2
(35)
(39)
Therefore, V2 (t) exponentially converges to the bounded
set Ω1 as given in (30). It implies that ξ(t), αij (t) and
βij (t) are uniformly ultimately bounded.
Next, if ̺ , max{µϑ, νχ} < γ, we can obtain a smaller
set for ξ by rewriting (37) into
kB T P (xi (t) − xj (t))k2
kB T P (xi (t) − xj (t))k + εe−ϕti
α
α2
eij (t)2
ϑ −
+
2
2
i=1 j∈Ni
#
2
βe (t)2
β
ij
+χ −
.
+
2
2
+
N X
X
t
XZ
N
X
i=1 j∈Ni
i=1 j∈Ni
N
X
i=1 j∈Ni t
V̇2 ≤ ξ T (M ⊗(P A + AT P )−2αL⊗P BB T P )ξ
N X
√ X
kB T P (ξi − ξj )k
+f0 (N − 1) N
N X
X
(38)
j∈Ni
Substituting (34) and (35) into (33), we have
−β
i=1
|Ni |εe−ϕti
2
N
h
1 X X α2
β i
ϑ
+χ
V2 (t) ≤ e−δ(t−t0 ) V2 (t0 ) +
δ i=1
2
2
kB T P (xi (t) − xj (t))k2
kB T P (xi (t) − xj (t))k + εe−ϕti
i=1 j∈Ni
N
X
2
α2
β
+χ
|Ni | ϑ
+
2
2
i=1
N
X
In light of the well-known Comparison lemma in [27], we
can obtain from (38) that
(βeij (t)2 + βeij (t)β)
2
βeij (t)2
β
−
.
+
2
2
|Ni |εe
−ϕti
2
α2
β
|Ni | ϑ
+
.
+χ
2
2
i=1
βeij (t)β̇ij (t)
N X
X
j∈Ni
N
X
kB T P (xi (t) − xj (t))k2
− xj (t))k + εe−ϕti
i=1 j∈Ni
N X
X
(δ−µϑ)e
αij (t)2 (δ−νχ)βeij (t)2
+
2µ
2ν
i=1
≤ −δV2 + β
kB T P (xi (t)
i=1 j∈Ni
(37)
j∈Ni
V̇2 ≤ −̺V2 − λmin (P )(γ − ̺)kξk2
2
N
α2
X
β
−ϕti
|Ni | βεe
+
.
+ ϑ +χ
2
2
i=1
(40)
Obviously, it follows from (40)Pthat V̇2 (t) ≤ −̺V2 (t) +
N
PN
|Ni |
2
2
2
−ϕti
i=1
ϑα
|N
|βεe
.
,
if
kξk
>
+
χβ
i
i=1
2λmin (P )(γ−̺)
(36)
Then, in light of V2 (t) ≥ λmin (P )kξk2 , we can get
that if ̺ < γ then ξ exponentially converges to the
bounded set Ω2 in (32). Therefore, we obtain from
As shown in the proof of Theorem 1, by choosing α and
β sufficiently large such that α ≥ 2λ1 2 and β ≥ f0 (N −
8
Lemma 3 that distributed average tracking errors
PN
ξi = xi − N1 k=1 rk , i = 1, 2, · · · , N , converge to the
bounded set Ω2 as t → ∞. This completes the proof.
Remark 6 The adaptive scheme of the algorithm (28)
for updating coupling gains is partly borrowed from adaptive strategies in [5], [28], [29], and [30]. In Algorithm 1,
it requires the smallest nonzero eigenvalue λ2 of L, the
upper bound f0 of fi (t) and the number N of nodes in
the network. Note that λ2 , f0 and N are global information for each agent in the network and might not be obtained in real applications. By using adaptive strategies
(27) with (28) in Theorem 2, the limitation of all these
global information can be removed.
x1i, i=1,2,⋅⋅⋅,6.
Fig. 1. The communication topology.
0
−5
0
5
10
time
15
20
5
10
time
15
20
5
0
−5
0
ξ1i, i=1,2,⋅⋅⋅,6.
Fig. 2. State trajectories xi of six agents in networks.
ξ2i, i=1,2,⋅⋅⋅,6.
5
x2i, i=1,2,⋅⋅⋅,6.
Remark 7 Note that related works in [22], [23], and
[24] studied the distributed average tracking problem for
integrator-type and linear signals by using non-smooth
algorithms, which inevitably produces the chattering phenomenon. Compared with above results, the contribution
of this paper is three-fold. First, main results of this paper extend the dynamics from integrator-type signals in
[22], [23] to linear signals. The proposed algorithms (2)
and (27) successfully solve the distributed average tracking problem for reference signals generated by the more
general linear dynamics. Second, by using adaptive control approaches, the limitation of all these global information is removed. Third, compared with existing results
in [24], new continuous algorithms are redesigned via the
boundary layer concept, which plays a vital role to reduce
the chattering phenomenon in real applications.
5
Simulations
In this section, we will give an example to verify Theorem 2. The dynamics of multiple
reference signals
are
!
!
r1i
0 1
given by (1) with ri =
, A=
, B=
−1 −2
r2i
!
0
, and fi (t) = i+1
2 sin(t), where i = 1, 2, · · · , 6. The
1
communication topology is shown in Fig. 1. Solving the
ARE (10) with Q = I gives the gain matrices K and!Γ
2.4738 6.8092
as K= −1.5728 −4.3293 , Γ=
.
6.8092 18.7428
The state trajectories xi , i = 1, 2, · · · , 6, of six agents
under Algorithm 2 with µ = 10, ν = 10, ϑ = 0.01, χ =
0.01, ε = 5, ϕ = 0.5, K and Γ given above are depicted in Fig. 2, which shows that states achieve a small
bounded neighborhood of the average value of all signals. It follows from Fig. 3 that tracking errors ξi ,
P6
xi − 16 k=1 rk convergent to a small bounded neighborhood of the origin as t → ∞. The adaptive coupling
gains αij (t) and βij (t) are also drawn in Fig. 4, respectively. As a comparison, the discontinuous algorithm in
[24] and continuous algorithms (27) are also shown with
the same parameters in Fig. 5, where we can see that the
5
0
−5
0
5
10
time
15
20
5
10
time
15
20
4
2
0
−2
−4
0
Fig. 3. Tracking error trajectories ξi = xi − 16
agents in the network.
P6
k=1
rk of six
ij
α (t)
5
0
0
5
10
time
15
20
5
10
time
15
20
2
ij
β (t)
3
1
0
0
Fig. 4. Adaptive coupling strengths αij (t) and βij (t) in (28).
chattering effect with discontinuous algorithm in [24] is
greatly reduced by using the continuous algorithm (27).
6
Conclusions
In this paper, we have studied the distributed average
tracking problem of multiple time-varying signals generated by general linear dynamical systems, whose ref-
9
5
[11] Y. Zhao, Y. F. Liu, Z. S, Duan and G. H. Wen, “Distributed
average computation for multiple time-varying signals with
output measurements,” International Journal of Robust and
Nonlinear Control, DOI: 10.1002/rnc.3486, 2015.
0
−5
−10
0
5
10
time(s)
15
[12] Y. Zhao and Z. S, Duan, “Finite-time containment control
without velocity and acceleration measurements,” Nonlinear
Dynamics, vol. 82, no. 1, pp. 259–268, 2015.
20
10
[13] M. Ji, G. Ferrari-Trecate, M. Egerstedt, and A. Buffa,
“Containment control in mobile networks,” IEEE Trans.
Autom. Control, vol. 53, no. 8, pp. 1972–1975, 2008.
5
i
continuous algorithm u (t), i=1,2,⋅⋅⋅,6.
i
discontinuous algorithm u (t), i=1,2,⋅⋅⋅,6.
[10] Y. Zhao, Z. S, Duan, G. H. Wen, and G. R.
Chen, “Distributed finite-time tracking of multiple nonidentical second-order nonlinear systems with settling time
estimation,” Automatica, vol. 64, no. 2, pp. 86–93, 2016.
10
0
[14] D. Spanos and R. Murray, “Distributed sensor fusion using
dynamic consensus,” in Proc. 16th IFAC World Congress,
2005.
−5
−10
0
5
10
time(s)
15
[15] H. Bai, R. Freeman, and K. Lynch, “Distributed kalman
filtering using the internal model average consensus
estimator,” in Proc. Amer. Control Conf., pp. 1500–1505.
20
Fig. 5. Discontinuous algorithm in [24] and the continuous
algorithm (27).
[16] P. Yang, R. Freeman, and K. Lynch, “Multi-agent
coordination by decentralized estimation and control,” IEEE
Trans. Autom. Control, vol. 53, no. 11, pp. 2480–2496, 2008.
erence inputs are nonzero, bounded and not available
to any agents in networks. In the distributed fashion, a
pair of continuous algorithms with static and adaptive
coupling strengths have been developed in light of the
boundary layer concept. Besides, sufficient conditions for
the existence of distributed algorithms are given if each
agent is stabilizable. The future topic will be focused
on the distributed average tracking problem for the case
with only the relative output information of neighboring
agents.
[18] D. Spanos, R. Olfati-Saber, and R. Murray, “Dynamic
consensus on mobile networks,” in Proc. 16th IFAC World
Congress, 2005.
References
[20] H. Bai and R. F. nd K. Lynch, “Robust dynamic average
consensus of time-varying inputs,” in Proc. 49th IEEE Conf.
Decision Control, pp. 3104–3109, 2010.
[17] Y. Sun and M. Lemmon, “Swarming under perfect consensus
using integral action,” in Proc. Amer. Control Conf., pp.
4594–4599, 2007.
[19] R. Freeman, P. Yang, and K. Lynch, “Stability and
convergence properties of dynamic average consensus
estimators,” in Proc. 45th IEEE Conf. Decision Control, pp.
338–343, 2006.
[1] R. Olfati-Saber, J. Fax, and R. Murray, “Consensus and
cooperation in networked multi-agent systems,” in Proc.
IEEE, vol. 95, no. 1, pp. 215–233, 2007.
[2] W. Ren, R. Beard, and E. Atkins, “Information consensus in
multivehicle cooperative control,” IEEE Control Syst. Mag.,
vol. 27, no. 2, pp. 71–82, 2007.
[3] Y. Hong, G. Chen, and L. Bushnell, “Distributed observers
design for leader-following control of multi-agent networks,”
Automatica, vol. 44, no. 3, pp. 846–850, 2008.
[4] Y. Cao and W. Ren, “Distributed coordinated tracking with
reduced interaction via a variable structure approach,” IEEE
Trans. Autom. Control, vol. 56, no. 1, pp. 33–48, 2012.
[5] Z. Li, X. Liu, W. Ren, and L. Xie, “Distributed tracking
control for linear multiagent systems with a leader of bounded
unknown input,” IEEE Trans. Autom. Control, vol. 58, no. 2,
pp. 518–523, 2013.
[6] S. E. Tuna, “Synchronizing linear systems via partial-state
coupling,” Automatica, vol. 44, no. 8, pp. 2179-2184, 2008.
[7] H. Zhang, F. Lewis, and A. Das, “Optimal design for
synchronization of cooperative systems: State feedback,
observer, and output feedback,” IEEE Trans. Autom.
Control, vol. 56, no. 8, pp. 1948–1952, 2011.
[8] Y. F. Liu, and Z. Y. Geng, “Finite-time formation control for
linear multi-agent systems: A motion planning approach,”
Systems and Control Letters, vol. 85, no. 11, pp. 54–60, 2015.
[9] Y. F. Liu, Y. Zhao, and Z. Y. Geng, “Finite-time formation
tracking control for multiple vehicles: A motion planning
approach,” International Journal of Robust and Nonlinear
Control, DOI: 10.1002/rnc.3496, 2015.
[21] S. Nosrati, M. Shafiee, and M. Menhaj, “Dynamic average
consensus via nonlinear protocols,” Automatica, vol. 48, no. 9,
pp. 2262–2270, 2012.
[22] F. Chen, Y. Cao, and W. Ren, “Distributed average tracking
of multiple time-varying reference signals with bounded
derivatives,” IEEE Trans. Autom. Control, vol. 57, no. 12,
pp. 3169–3174, 2012.
[23] F. Chen, W. Ren, W. Lan, and G. Chen, “Distributed average
tracking for reference signals with bounded accelerations,”
IEEE Trans. Autom. Control, vol. 60, no. 3, pp. 863–869,
2015.
[24] Y. Zhao, Z. Duan and Z. Li, “Distributed average tracking
for multiple signals with linear dynamics: an edge-based
framework,” in Proc. 11th IEEE Inter. Conf. Control Auto.
2014.
[25] C. Godsil and G. Royle, Algebraic Graph Theory. New York:
Springer, 2001.
[26] C. Edwards and S. Spurgeon, Sliding Mode Control: Theory
and Applications. London: Taylor and Francis, 1998.
[27] H. Khalil, Nonlinear Systems.
Prentice Hall, 2002.
Englewood Cliffs, NJ:
[28] H. Su, G. Chen, X. Wang, and Z. Lin, “Adaptive secondorder consensus of networked mobile agents with nonlinear
dynamics,” Automatica, vol. 47, no. 2, pp. 368–375, 2011.
[29] W. Yu, W. Zheng, J. Lü, and G. Chen, “Designing distributed
control gains for consensus in multi-agent systems with
second-order nonlinear dynamics,” Automatica, vol. 49, no. 7,
pp. 2107–2115, 2013.
10
[30] H. Zhang, and L. Frank, “Adaptive cooperative tracking
control of higher-order nonlinear systems with unknown
dynamics,” Automatica, vol. 48, no. 7, pp. 1432–1439, 2012.
[31] D. L. Mills, “Internet time synchronization: The network time
protocol,” IEEE Transactions on Communications, vol. 39,
no. 10, pp. 1482–1493, 1991.
[32] B. Sundararaman, U. Buy, and A.D. Kshemkalyani, “Clock
synchronization for wireless sensor networks: a survey,” Ad
Hoc Networks, vol. 3, no. 3, pp. 281–323, 2005.
[33] M. Franceschelli, A. Pisano, A. Giua, and E. Usai, “Finitetime consensus based clock synchronization by discontinuous
control analysis and design of hybrid systems,” The 4th IFAC
Conference on Analysis and Design of Hybrid Systems, pp.
172–177, 2012.
[34] R. Carli, and S. Zampieri, “Network clock synchronization
based on the second order linear consensus algorithm,” IEEE
Transactions on Automatic Control, vol. 59, pp. 409–422,
2014.
[35] S. Bolognani, R. Carli, E. Lovisari, and S. Zampieri, “A
randomized linear algorithm for clock synchronization in
multi-agent systems,” IEEE Transactions on Automatic
Control, published online, 2016.
11
| 3cs.SY
|
Ascending chains of finitely generated subgroups
arXiv:1601.02135v1 [math.GR] 9 Jan 2016
Mark Shusterman
January 12, 2016
Abstract
We show that a nonempty family of n-generated subgroups of a
pro-p group has a maximal element. This suggests that ’Noetherian
Induction’ can be used to discover new features of finitely generated
subgroups of pro-p groups. To demonstrate this, we show that in various pro-p groups Γ (e.g. free pro-p groups, nonsolvable Demushkin
groups) the commensurator of a finitely generated subgroup H 6= 1
is the greatest subgroup of Γ containing H as an open subgroup. We
also show that an ascending sequence of n-generated subgroups of a
limit group must terminate (this extends the analogous result for free
groups proved by Takahasi, Higman, and Kapovich-Myasnikov).
1
Introduction
Chain conditions play a prominent role in Algebra. A good example is
the variety of results on Noetherian rings and their modules. In this
work we consider chain conditions on profinite groups. All the grouptheoretic notions considered for these groups should be understood in
the topological sense, i.e. subgroups are closed, homomorphisms are
continuous, generators are topological, etc. The ascending chain condition on finitely generated subgroups does not hold for pro-p groups
in general, and our first result is some kind of remedy for this.
Theorem 1.1. Let p be a prime number, let n ∈ N, let Γ be a pro-p
group, and let F =
6 ∅ be a family of n-generated subgroups of Γ. Then
F has a maximal element with respect to inclusion.
It is our hope that this simple result will unveil new properties of
pro-p groups and their finitely generated subgroups. An example is
the following theorem, for which we need a definition. Given a prime
number p, we say that a pro-p group Γ has a Hereditarily Linearly
Increasing Rank if for every finitely generated subgroup H ≤c Γ
there exists an ǫ > 0 such that for any open subgroup U ≤o H we have
d(U ) ≥ max{d(H), ǫ(d(H) − 1)[H : U ]}
(1.1)
where d(K) stands for the smallest cardinality of a generating set for
the pro-p group K. That is, our definition says that the minimal number of generators of finite index subgroups of H grows monotonically,
1
and linearly as a function of the index. Examples of groups with this
property include free pro-p groups, nonsolvable Demushkin groups, and
groups from the class L all of whose abelian subgroups are procyclic.1
Theorem 1.2. Let p be a prime number, let Γ be a pro-p group with
a hereditarily linearly increasing rank, and let H ≤c Γ be a nontrivial
finitely generated subgroup. Then the commensurator of H in Γ
CommΓ (H) ··= {g ∈ Γ | [H : H ∩ gHg −1 ], [gHg −1 : H ∩ gHg −1 ] < ∞}
is the greatest subgroup of Γ containing H as an open subgroup. Also,
the action of any CommΓ (H) c L ≤c Γ on L/H is faithful.
Note that several assertions are encompassed in this statement.
First, we claim that the family of subgroups of Γ containing H as an
open subgroup has a (necessarily unique) greatest element with respect
to inclusion. In particular, there are only finitely many subgroups of
Γ that contain H as an open subgroup. The second assertion is that
the aforementioned greatest element is the commensurator of H in Γ.
Hence H is an open subgroup of CommΓ (H), and thus also an open
subgroup of its normalizer in Γ. Furthermore, given finitely generated
commensurable H1 , H2 ≤c Γ (i.e. H1 ∩ H2 is open in both H1 and H2 )
we can apply Theorem 1.2 to H1 ∩ H2 and conclude that it is an open
subgroup of hH1 ∪ H2 i. At last, if [Γ : H] is infinite, then by taking
L = Γ we find that H contains no nontrivial normal subgroup of Γ.
Results analogous to these assertions are abundant in the literature,
where the group Γ is replaced by:
1. Free groups - [9, Corollary 8.8, Proposition 8.9], [11, Theorem 1],
[18].
2. Fuchsian groups - [4].
3. Hyperbolic groups - [10, Theorems 1,3].
4. Limit groups - [2, Theorem 1], [3, Theorem 4.1], [15, Chapters
4,5], [20, Section 6].
5. Groups with a positive first ℓ2 -Betti number - [16, Corollary 5.13,
Proposition 7.3].
6. Free profinite groups - [8, Main Theorem].
7. Absolute Galois groups of Hilbertian fields - [5].
8. Free pro-p groups and free pro-p products - [14, 3.3, 3.5], [13,
Theorem C].
9. Nonsolvable pro-p Demushkin groups and other pro-p IF -groups
with positive deficiency - [12, 3.12, 3.13], [21], [7].
10. Pro-p groups from the class L - [20, Theorem C (5-7)].
1
By [20, Theorem 4.6] the deficiency is ≥ 2, and by the paragraph following [7, Proposition 1] it grows linearly for open subgroups, so the number of generators grows linearly
as well, and is monotonic by [13, Theorem B].
2
It is our point of view that an assumption on the increase in the
number of generators upon passing to finite index subgroups (e.g.
(1.1)) creates a good framework for proving results like those stated in
Theorem 1.2, the paragraph following it, and the list above. Indeed, all
the groups in the list (excluding some of those in 3), exhibit a linear
growth of the number of generators as a function of the index of a
subgroup (i.e. these groups have rank gradient > 0). As a result, arguments from the proof of Theorem 1.2 can be used to obtain most of the
results in the list above. For instance, [14, pro-p Greenberg Theorem],
[12, Lemma 3.12, Proposition 3.13], [21, Theorem A], and a part of [20,
Theorem C (5-7)] are special cases of Theorem 1.2.
Next, we generalize Takahasi’s theorem (see [22, Theorem 1], [6,
Lemma], and [9, Theorem 14.1]) which is the case of free G in the
following.
Theorem 1.3. Let G be a group for which every subgroup H ≤ G
whose profinite completion is finitely generated, is itself finitely generated. Let n ∈ N, and let F =
6 ∅ be a family of n-generated subgroups of
G. Then F contains a maximal element.
Most notably, by Proposition 3.3, the theorem applies to limit groups,
and also to Fuchsian groups. A consequence of Theorem 1.3 is Corollary 3.4.
2
Profinite Groups
2.1
Directed families
Given a set I, we say that a family of subgroups {Ai }i∈I of a group G is
directed if for every i, j ∈ I there exists a k ∈ I such that Ak ≥ Ai , Aj .
In this case, the abstract subgroup generated by the {Ai }i∈I is just
their union. Furthermore, it follows by induction that for all m ∈ N
and i1 , . . . , im ∈ I there exists some i ∈ I such that Ai ≥ Aik for each
1 ≤ k ≤ m.
Lemma 2.1. Let Γ be a profinite group, let {Ai }i∈I be a directed
family of subgroups of Γ, set A ··= hAi ii∈I , let G be a finite group, and
let τ : A → G be an epimorphism. Then there exists some j ∈ I such
that τ |Aj is a surjection.
Proof. Note that
G = τ (A) = τ (hAi ii∈I ) = τ (
[
i∈I
Ai ) ⊆ τ (
[
i∈I
Ai ) =
[
τ (Ai )
(2.1)
i∈I
so for each g ∈ G there exists some ig ∈ I such that g ∈ τ (Aig ). Since
G is finite, directedness implies that there exists some j ∈ I such that
Aj ≥ Aig for all g ∈ G. It follows that τ (Aj ) = G as required.
Corollary 2.2. Let Γ be a profinite group, let n ∈ N, and let {Ai }i∈I
be a directed family of n-generated subgroups of Γ. Then A ··= hAi ii∈I
is n-generated.
3
Proof. Let τ : A → G be an epimorphism onto a finite group. By
Lemma 2.1, there exists some j ∈ I such that τ (Aj ) = G. Hence,
d(G) ≤ d(Aj ) ≤ n, so d(A) ≤ n since by [17, Lemma 2.5.3] we know
that d(A) is determined by the finite homomorphic images of A.
For the proof of Theorem 1.1 recall that the Frattini subgroup of a
profinite group U , denoted by Φ(U ), is defined to be the intersection
of all maximal subgroups of U .
Theorem 2.3. Let p be a prime number, let n ∈ N, let Γ be a pro-p
group, and let F =
6 ∅ be a family of n-generated subgroups of Γ. Then
F contains a maximal element.
Proof. Let C be a chain in F , and let U be the subgroup of Γ generated
by the subgroups in C. Since C is a chain, it is directed, so d(U ) ≤ n by
Corollary 2.2. As U is a finitely generated pro-p group, [17, Proposition
2.8.10] tells us that U → U/Φ(U ) is an epimorphism onto a finite group.
By Lemma 2.1, there exists some H ∈ C such that HΦ(U ) = U , so in
view of [17, Corollary 2.8.5] we must have H = U . Thus U ∈ F is an
upper bound for C. By Zorn’s Lemma, F has a maximal element.
Note that our assumption that Γ is not merely a profinite group but
a pro-p group, has been used in the proof to conclude that Φ(U ) ≤o U
for any finitely generated U ≤c Γ. By [17, Proposition 2.8.11], this
conclusion holds under the weaker assumption that Γ is a prosupersolvable group with order divisible by only finitely many primes. Such
groups have been studied, for instance, in [1].
2.2
2.2.1
Hereditarily linearly increasing rank
Basic properties
Proposition 2.4. Let p be a prime number, let Γ be a pro-p group
with a hereditarily linearly increasing rank, let H ≤c Γ be a finitely
generated subgroup, and let R ≤c Γ be a subgroup containing H as an
open subgroup. Then d(R) ≤ d(H).
Proof. By taking the union of a finite generating set for H with a
transversal for H in R we get a finite generating set for R. It follows
from (1.1) that d(H) ≥ d(R) as required.
Proposition 2.5. Let p be a prime number, and let Γ be a pro-p group
with a hereditarily linearly increasing rank. Then Γ is torsion-free.
Proof. Let C ≤c Γ be a finite subgroup. Since {1} ≤o C, (1.1) implies
that 0 = d({1}) ≥ d(C) which guarantees that C = {1} as required.
4
Corollary 2.6. Let p be a prime number, let Γ be a pro-p group with
a hereditarily linearly increasing rank, and let H be a finitely generated
subgroup of Γ. Then F ··= {R ≤c Γ | H ≤o R} has a maximal element.
Proof. This is immediate from Proposition 2.4 and Theorem 2.3.
The following simple lemma is useful for constructing small generating sets of profinite groups. It remains true both for abstract groups
and for profinite groups which are not finitely generated, but we do
not need it in this generality.
Lemma 2.7. Let G be a finitely generated profinite group, let K ⊳c G
be a normal subgroup, and let H ≤c G be a finitely generated subgroup
containing K. Then d(G) ≤ d(H) + d(G/K).
Proof. Let π : G → G/K be the quotient map, let X be a finite generating set of H, and let Y be a finite generating set of G/K. Since
π is surjective, there exists some Z ⊆ G which is bijectively mapped
by π onto Y . Set S ··= X ∪ Z, M ··= hSi and note that it suffices to
show that M = G, since this guarantees the existence of a generating
set (i.e. S) for G of the required cardinality.
In order to see that M = G, note that π(M ) ⊇ π(S) ⊇ π(Z) = Y
which generates G/K. Therefore, π|M is a surjection, which means
that M K = G. Furthermore, M ⊇ S ⊇ X which generates H so
K ≤c H ≤c M and thus M = M K = G as required.
2.2.2
Faithful action
We establish Theorem 1.2 in a sequence of claims, the most important of which is the following theorem that makes crucial use of the
assumption on linear growth of generating sets for subgroups (i.e. positive rank gradient). For the proof, recall that if U is an open subgroup
of a finitely generated profinite group Γ, then d(U ) ≤ d(Γ)[Γ : U ] as
can be seen from [17, Corollary 3.6.3].
Theorem 2.8. Let p be a prime number, let Γ be a pro-p group with
a hereditarily linearly increasing rank, and let H ≤c Γ be a finitely
generated subgroup of infinite index. Then the action of Γ on Γ/H is
faithful.
Proof. Let K ⊳c Γ be the kernel of the action, and note that K ≤c H.
Towards a contradiction, suppose that K 6= {1}. By Corollary 2.6, we
can find some M ≤c Γ maximal among those having H as an open
subgroup. Since [M : H] < ∞ and [Γ : H] = ∞ by assumption, we can
pick some x ∈ Γ \ M . Set N ··= hM ∪ {x}i, so that [N : H] = ∞ by
the maximality of M . We see that
2.4
d(N ) ≤ d(M ) + 1 ≤ d(H) + 1
5
(2.2)
so N is finitely generated, and from the fact that K is a nontrivial
nonopen subgroup of N , we infer that d(N ) > 1 (if d(N ) ≤ 1 then all
nontrivial subgroups of the pro-p group N are open).
By (1.1), there exists an ǫ > 0 such that for all V ≤o N we have
d(V ) ≥ ǫ(d(N ) − 1)[N : V ] = δ[N : V ]
(2.3)
where δ ··= ǫ(d(N ) − 1) > 0. By Proposition 2.5, K is infinite, and
this fact (along with the fact that [N : H] = ∞) is seen in the finite
quotients of N . For instance, there exists some U ⊳o N such that
[N : U H], [U K : U ] >
2(d(H) + 1)
.
δ
(2.4)
We find that
2.7
d(U ) ≤ d(U/U ∩ K) + d(U ∩ H)
= d(U K/K) + d(U ∩ H)
≤ d(U K) + d(H)[H : U ∩ H]
≤ d(N )[N : U K] + d(H)[U H : U ]
(2.5)
2.2
[N : U ]
[N : U ]
+ d(H)
≤ (d(H) + 1)
[U K : U ]
[N : U H]
2.4
2(d(H) + 1)[N : U ]
≤
< δ[N : U ]
min{[U K : U ], [N : U H]}
which is a contradiction to (2.3).
For the next corollary, recall that given a subgroup H of a profinite
group Γ, the normalizer of H in Γ (the set of γ ∈ Γ for which γH = Hγ)
is denoted by NΓ (H). The normalizer is easily seen to be a subgroup.
Corollary 2.9. Let p be a prime number, let Γ be a pro-p group with
a hereditarily linearly increasing rank, and let H 6= {1} be a finitely
generated subgroup of Γ. Then [NΓ (H) : H] < ∞.
Proof. Suppose that [NΓ (H) : H] = ∞. Since hereditarily linearly
increasing rank is inherited by subgroups, we can apply Theorem 2.8
to the action of NΓ (H) on its cosets modulo H, and get that
\
gHg −1 = {1}
(2.6)
H=
g∈NΓ (H)
where the first equality stems from the normality of H in NΓ (H),
and the second one from the faithfulness of the action of NΓ (H) on
NΓ (H)/H. Clearly, (2.6) contradicts our assumption that H 6= {1}.
6
2.2.3
Roots and Commensurators
Given a subgroup H of a profinite group Γ, we define the family of
’finite extensions’ of H by F ··= {R ≤c Γ | H ≤o R}.√Following [18],
we say that M ∈ F is the root of H (and write M = H) if M is the
greatest element in F with respect to inclusion. Note that F may fail
to have a greatest element, so H does not necessarily have a root.
Theorem 2.10. Let p be a prime number, and let Γ be a pro-p group
with a hereditarily linearly increasing rank. Then every finitely generated subgroup of Γ has a root.
Proof. Let n ∈ N, and let D be the family of all n-generated subgroups
of Γ which do not have a root. Towards a contradiction, suppose that
D 6= ∅. By Theorem 2.3, there is a maximal M ∈ D. By Corollary 2.6,
there exists some T maximal among the subgroups of Γ having M as
an open subgroup. Since M ∈ D, we know that T is not a root of M ,
so there exists an A ≤c Γ not contained in T , such that M ≤o A. As
M ≤ T and A T , we conclude that M
A. The maximality of T
implies that T A so M
T . By Proposition 2.5, M 6= {1} since
otherwise T would have been a nontrivial finite subgroup of Γ.
Set B ··= T ∩ A and suppose that there exists a root C of B. Since
M ≤o B we see that B ≤o A, T so A, T ≤c C by definition of √
the
root. Since A T , we find that T
C. On the other hand, C = B
implies that B ≤o C, and thus M ≤o C which is a contradiction to
the maximality of T . We conclude that B does not have a root, and by
Proposition 2.4, d(B) ≤ d(M ) ≤ n. Hence, B ∈ D and the maximality
of M implies that B = M .
Let A0 , T0 o M be minimal subgroups of A, T respectively, and set
J ··= hA0 ∪T0 i. Evidently, M is a maximal subgroup of the pro-p groups
A0 , T0 so by [17, Lemma 2.8.7 (a)], M ⊳o A0 , T0 which means that
A0 , T0 ≤c NΓ (M ). Consequently, J ≤c NΓ (M ), so from Corollary 2.9
we get that [J : M ] < ∞. By Proposition 2.4, d(T0 ) ≤ d(M ) ≤ n
so from the maximality of M √we infer that T0 has a root R. Since
T0 ≤o T, J we have T, J ≤c T0 = R. The fact that T ∩ A =√M
implies that A0 T , so T
R as A0 ≤o J ≤o R. Since R = T0
we have M ≤o T0 ≤o R so R contains M as an open subgroup, thus
contradicting the maximality of T .
The following assertion holds whenever roots exist.
Proposition 2.11. Let p be a prime number, let Γ be a pro-p group
with a hereditarily linearly increasing rank, let H be a finitely
√
√ generated
subgroup of Γ, and let K be an open subgroup of H. Then H = K.
√
√
√
Proof. Since K ≤o √
H we have H √
≤o K so K√≤o H.
√ On the other
hand, K ≤o H ≤o H so K ≤o H and thus H ≤o K.
More generally, we have the following.
7
Corollary 2.12. Let p be a prime number, let Γ be a pro-p group
with a hereditarily linearly increasing rank, and let H, K be finitely
generated
of Γ. Then H and K are commensurable if and
√ subgroups
√
only if H = K.
√
√
Proof. Suppose first that H = K and denote by L this common
root. We have
[H : H ∩ K], [K : H ∩ K] ≤ [L : H ∩ K] ≤ [L : H][L : K]
(2.7)
√
√
where the right hand side is finite since L = H = K so we have
established commensurability. Conversely, if H and K are commensurable then
√ 2.11 √
2.11 √
H =
H ∩K =
K.
(2.8)
Recall that given a subgroup H of a profinite group Γ, the commensurator of H in Γ (the set of γ ∈ Γ for which H and γHγ −1
are commensurable) is denoted by CommΓ (H). Commensurability is
an equivalence relation on subgroups, and the commensurator is an
abstract subgroup of Γ.
Corollary 2.13. Let p be a prime number, let Γ be a pro-p group with
a hereditarily linearly increasing rank, and let√H 6= {1} be a finitely
generated subgroup of Γ. Then CommΓ (H) = H.
√
Proof. Let g ∈ H. We have
p
√
√
gHg −1 = g Hg −1 = H
(2.9)
so by Corollary 2.12, H and gHg −1 are commensurable, which means
that g ∈ CommΓ (H). Now, let x ∈ CommΓ (H). Thus, H and xHx−1
are commensurable. By Corollary 2.12,
√
√
√
H = xHx−1 = x Hx−1
(2.10)
√
√
H). We have thus shown that Comm
so x ∈ NΓ ( √
√ Γ (H) ≤ NΓ ( H).
Since H ≤o H Proposition 2.4
us that H is finitely gener√ tells √
H)
:
H] < ∞ which √
implies that
ated,√ so by Corollary
2.9,
[N
(
Γ
√
NΓ ( H) = H
in
view
of
the
maximality
of
the
root
H. Hence,
√
CommΓ (H) ≤ H and we have an equality.
Observe that Theorem 1.2 follows at once from Corollary 2.13 and
Theorem 2.8.
3
Abstract Groups
b the profinite completion of G.
For a group G, we denote by G
Corollary 3.1. Let G be a group, let n ∈ N, let {Ai }i∈I be a directed
b ≤ n.
family of n-generated subgroups, and set A ··= hAi ii∈I . Then d(A)
8
b and the closures {Ai }i∈I in A.
b
Proof. Apply Corollary 2.2 to A,
We are now up to proving Theorem 1.3.
Corollary 3.2. Let G be a group for which every subgroup H whose
profinite completion is finitely generated, is itself finitely generated. Let
n ∈ N, and let F =
6 ∅ be a family of n-generated subgroups of G. Then
F contains a maximal element.
Proof. Let C be a chain in F , and let U be its union. Since C is a chain,
b ) ≤ n. By our assumption on G, there
Corollary 3.1 implies that d(U
exists a finite generating set S ⊆ U . Since C is a chain, we can find an
R ∈ C which contains S, and thus all of U . Therefore, U = R ∈ F is
an upper bound for C. By Zorn’s Lemma, F has a maximal element.
Let us now briefly explain why Theorem 1.2 applies to limit groups.
Proposition 3.3. Let L be a limit group, and let H ≤ L be a subgroup with a finitely generated profinite completion. Then H is finitely
generated.
Proof. Suppose that H is not finitely generated. By [19, Theorem 3.2],
L decomposes as a graph of groups Y with cyclic edge groups. This
induces a decomposition of H as a graph of groups X which must
be infinite since H is not finitely generated. If X has an infinite first
Betti number, then H surjects onto a free group of infinite rank, so in
particular, its profinite completion is not finitely generated.
We may thus assume that the first Betti number of X is finite,
which implies that
X = C ∪ T1 ∪ · · · ∪ Tn
(3.1)
where C is compact, n ∈ N, and T1 , . . . , Tn are infinite trees with a
unique leaf each. Hence, in order to show that the profinite completion
of H is not finitely generated, it is sufficient to show this for the fundamental group of T1 . For that, recall that every abelian subgroup of a
limit group is finitely generated free abelian so infinitely many vertex
groups in T1 are not cyclic. Since edge groups are cyclic, by collapsing edges we can assure that all vertex groups in T1 are not cyclic.
It follows from fully residual freeness that every vertex group surjects
onto Z2 , so the fundamental group of T1 surjects onto the fundamental
group of a tree of groups in which every vertex group is Z2 and every
edge group is either {1} or Z. The latter group surjects onto an infinite
direct sum of Z/2Z so its profinite completion is not finitely generated.
The next corollary follows [6, Corollary].
Corollary 3.4. Let G be a limit group, let α be an automorphism of
G, and let H ≤ G be a finitely generated subgroup which is mapped by
α into itself. Then α(H) = H.
9
Proof. Set
F ··= {K ≤ G | d(K) ≤ d(H), α(K)
K}
(3.2)
and suppose that H ∈ F . By Corollary 3.2 and Proposition 3.3, there
exists a maximal M ∈ F . It is easy to verify that α−1 (M ) ∈ F , so
α−1 (M ) = M by maximality. Hence, α(M ) = M - a contradiction.
Acknowledgments
I would like to sincerely thank Henry Wilton for answering many questions, and for his proof of Proposition 3.3. Special thanks go to Pavel
Zalesskii for all the helpful discussions. This research was partially supported by a grant of the Israel Science Foundation with cooperation of
UGC no. 40/14.
References
[1] K. Auinger, B. Steinberg, Varieties of finite supersolvable groups
with the M. Hall property, Math. Ann. 2006, 335, 4, 853-877.
[2] M. Bridson, J. Howie, Normalisers in limit groups, Math. Ann.
337, 2007, 385-394.
[3] M. Bridson, J. Howie, C. F. Miller, H. Short, Subgroups of direct
products of limit groups, Ann. of Math. 2, 170, 2009, 3, 1447–1467.
[4] H. B. Griffiths, The fundamental group of a surface, and a theorem
of Schreier, Acta Mathematica, 1963, 110, 1, 1-17.
[5] D. Haran, The Bottom Theorem, Arch. Math. 45, 229–231, 1985.
[6] H. Higman, A finitely related group with an isomorphic factor
group, Jorn. Lond. Math. Soc. 26, 1951, 59-61.
[7] J. A. Hillman, A. Schmidt, Pro-p groups of positive deficiency,
Bull. Lond. Math. Soc. 40, 2008, 1065–1069.
[8] M. Jarden, A Karrass-Solitar theorem for profinite groups, Journal
of Group Theory, 9 ,2006, 139-146.
[9] I. Kapovich, A. Myasnikov, Stallings Foldings and Subgroups of
Free Groups, Journ. of Algebra, 248, 2, 2002, 608–668.
[10] I. Kapovich, H. Short, Greenberg’s theorem for quasiconvex subgroups of word hyperbolic groups, Canadian Journ. of Math. 48,
6, 1996, 1224-1244.
[11] A. Karrass, D. Solitar, Note on a theorem of Schreier, Proc. Amer.
Math. Soc. (8) 1957, 696-697.
[12] D. Kochloukova, Subdirect products of free pro-p groups and Demushkin groups, Int. J. Algebra Comput. 23, 2013, 1079–1098.
[13] D. Kochloukova, P. Zalesskii, Subgroups and homology of extensions of centralizers of pro-p groups, Mathematische Nachrichten
288, 5-6, 604–618, 2015.
10
[14] A. Lubotzky, Combinatorial group theory for pro-p groups, J. Pure
Appl. Algebra, 25, 1982, 311–325.
[15] A.V. Nikolaev, D.E. Serbin, Finite index subgroups of fully residually free groups, Int. J. Algebra Comput. 21, 651, 2011.
[16] J. Peterson, A. Thom, Group cocycles and the ring of affiliated
operators, Inventiones Math., 185, 3, 2011, 561-592.
[17] L. Ribes, P. Zalesskii, Profinite Groups, A Series of Modern Surveys in Mathematics, c Springer-Verlag Berlin Heidelberg 2010.
[18] S. Rosset, Finite index and finite codimension, Journal of Pure
and Applied Algebra, 104, 1, 1995, 97–107.
[19] Z. Sela, Diophantine geometry over groups I - Makanin-Razborov
diagrams, Publications Mathematiques de lIHES 93, 1, 31–105.
[20] I. Snopce, P. Zalesskii, Subgroup properties of pro-p extensions of
centralisers, Sel. Math. New Ser. 20, 2014, 465–489.
[21] I. Snopce, P. Zalesskii, Subgroup properties of Demushkin groups,
Math. Proc. Camb. Phil. Soc., 2015.
[22] M. Takahasi, Note on Locally Free Groups, Journ. of the Inst. of
Polytechnics, Osaka City Univ., 1, 1950, 65-70.
Author’s address: Open Space - Room Number 2, Schreiber Building (Mathematics), Tel-Aviv University, Levanon Street, Tel-Aviv, Israel.
Author’s email: [email protected]
Author’s website: markshus.wix.com/math
11
| 0math.AC
|
Mixing properties and central limit theorem for
associated point processes
ARNAUD POINAS1,* BERNARD DELYON1,** FRÉDÉRIC LAVANCIER2,†
1
IRMAR – Campus de Beaulieu - Bat. 22/23 – 263 avenue du Général Leclerc – 35042 Rennes – France
arXiv:1705.02276v3 [math.ST] 19 Feb 2018
E-mail: * [email protected]
E-mail: ** [email protected]
2
Laboratoire de Mathématiques Jean Leray – BP 92208 – 2, Rue de la Houssinière – F-44322 Nantes Cedex 03 –
France. Inria, Centre Rennes Bretagne Atlantique, France.
E-mail: † [email protected]
Positively (resp. negatively) associated point processes are a class of point processes that induce attraction (resp.
inhibition) between the points. As an important example, determinantal point processes (DPPs) are negatively
associated. We prove α-mixing properties for associated spatial point processes by controlling their α-coefficients in
terms of the first two intensity functions. A central limit theorem for functionals of associated point processes is
deduced, using both the association and the α-mixing properties. We discuss in detail the case of DPPs, for which
we obtain the limiting distribution of sums, over subsets of close enough points of the process, of any bounded
function of the DPP. As an application, we get the asymptotic properties of the parametric two-step estimator of
some inhomogeneous DPPs.
Keywords: determinantal point process, parametric estimation, strong mixing, negative association, positive association.
1. Introduction
Positive association (PA) and negative association (NA) [1, 19] are properties that quantify the dependence
between random variables. They have found many applications in limit theorems for random fields [10, 40].
Even if the extension of PA to point processes have been used in analysis of functionals of random measures
[11, 20], there are no general applications of PA or NA to limit theorems for point processes. We contribute
in this paper to this aspect for spatial point processes on Rd . We especially discuss in detail the case of
determinantal point processes (DPPs for short), that are an important example of negatively associated
point processes. DPPs are a type of repulsive point processes that were first introduced by Macchi [32] in
1975 to model systems of fermions in the context of quantum mechanics. They have been extensively studied
in Probability theory with applications ranging from random matrix theory to non-intersecting random
walks, random spanning trees and more (see [26]). From a statistical perspective, DPPs have applications in
machine learning [29], telecommunication [15, 33, 22], biology, forestry [30] and computational statistics [2].
As a first result, we relate the association property of a point process to its α-mixing properties. First
introduced in [36], α-mixing is a measure of dependence between random variables, which is actually more
popular than PA or NA. It has been used extensively to prove central limit theorems for dependent random
variables [7, 16, 23, 27, 36]. More details about mixing can be found in [8, 16]. We derive in Section 2 an
important covariance inequality for associated point processes (Theorem 2.5), that turns out to be very
similar to inequalities established in [18] for weakly dependent continuous random processes. We show that
this inequality implies α-mixing and precisely allows to control the α-mixing coefficients by the first two
intensity functions of the point process. This result for point processes is in contrast with the case of random
fields where it is known that association does not imply α-mixing in general (see Examples 5.10-5.11 in [10]).
However, this implication holds true for integer-valued random fields (see [17] or [10]). As explained in [17],
this is because the σ-algebras generated by countable sets are much poorer than σ-algebras generated by
continuous sets. In fact, by this aspect and some others (for instance our proofs boil down to the control of
the number of points in bounded sets), point processes are very similar to discrete processes.
1
We then establish in Section 3 a general central limit theorem (CLT) for random fields defined as a
function of an associated point process (Theorem 3.1). A standard method for proving this kind of theorem
is to rely on sufficiently fast decaying α-mixing coefficients along with some moment assumptions. We use
an alternative procedure that exploits both the mixing properties and the association property. This results
in weaker assumptions on the underlying point process, that can have slower decaying mixing coefficients.
This improvement allows in particular to include all standard DPPs, some of them being otherwise excluded
with the first approach (like for instance DPPs associated to the Bessel-type kernels [4]).
Section 4 discusses in detail the case of DPPs, where we derive a tight explicit bound for their α-mixing
coefficients and prove a central limit theorem for certain functionals of a DPP (Theorem 4.4). Specifically,
these functionals write as a sum of a bounded function of the DPP, over subsets of close-enough points
of the DPP. A particular case concerns sums over p-tuple of close enough points of the DPP, which are
frequently encountered in asymptotic inference. Limit theorems in this setting have been established in [38]
when p “ 1, and in [3] for stationary DPPs and p ě 1. We thus extend these studies to sums over any
subsets and without the stationary assumption. As a statistical application, we consider the parametric
estimation of second-order intensity reweighted stationary DPPs. These DPPs have an inhomogeneous first
order intensity, but translation-invariant higher order (reweighted) intensities. We prove that the two-step
estimator introduced in [39], designed for this kind of inhomogeneous point process models, is consistent and
asymptotically normal when applied to DPPs.
2. Associated point processes and α-mixing
2.1. Notation
In this paper, we consider locally finite simple point processes on Rd , for a fixed d P N. Some theoretical
background on point processes can be found in [12, 34]. We denote by Ω the set of locally finite point
configurations in Rd . For X P Ω and A Ă Rd , we write
N pAq :“ cardpX ∩ Aq
for the random variable representing the number of points of X that fall in A. We also denote by BpAq the
Borel σ-algebra of A and by EpAq the σ-algebra generated by X ∩ A, defined by
EpAq :“ σptX P Ω : N pBq “ mu, B P BpAq, m P Nq.
The notation |.| will have a different meaning depending
on the object it is applied. For x P Rd , |x| stands
ş
d
for the euclidean norm. For a set A Ă R , |A| :“ A dx is the euclidean volume of A, and for a set I Ă Zd we
write |I| for the cardinal of I. For A, B two subsets of Rd (resp. Zd ) we define distpA, Bq as inf xPA,yPB |y ´ x|
and diampAq as supx,yPA |y ´ x| where |.| is the associated norm on Rd (resp. Zd ). For i P Zd , |i|1 denotes the
`1 -norm. Finally, we write Bpx, rq for the euclidean ball centred at x with radius r and }.}p for the p-norm
of random variables and functions where 1 ď p ď 8.
We recall that the intensity functions of a point process (when they exist), with respect to the Lebesgue
measure, are defined as follows.
Definition 2.1. Let X P Ω and n ě 1 be an integer. If there exists a non-negative function ρn : pRd qn Ñ R
such that
«
ff ż
‰
ÿ
E
f px1 , ¨ ¨ ¨ , xn q “
f px1 , ¨ ¨ ¨ , xn qρn px1 , ¨ ¨ ¨ , xn qdx1 ¨ ¨ ¨ dxn
x1 ,¨¨¨ ,xn PX
pRd qn
for all locally integrable functions f : pRd qn Ñ R then ρn is called the nth order intensity function of X.
In particular, ρn px1 , ¨ ¨ ¨ , xn qdx1 ¨ ¨ ¨ dxn can be viewed as the probability that X has a point in each of
the infinitesimally small sets around x1 , ¨ ¨ ¨ , xn with volumes dx1 , ¨ ¨ ¨ , dxn respectively.
2
We further introduce the notation
Dpx, yq :“ ρ2 px, yq ´ ρ1 pxqρ1 pyq.
(2.1)
This quantity is involved in the following equality, deduced from the previous definition and used several
times throughout the paper:
ż
CovpN pAq, N pBqq “
Dpx, yqdxdy.
(2.2)
AˆB
2.2. Negative and positive association
Our goal in this section is to prove a crucial covariance inequality and to deduce an α-mixing property
for associated point processes. We recall that associated point processes are defined the following way (see
Definitions 2.11-2.12 in [6] for example).
Definition 2.2. A point process X is said to be negatively associated (NA for short) if, for all families of
pairwise disjoint Borel sets pAi q1ďiďk and pBi q1ďiďl such that
p∪i Ai q ∩ p∪j Bj q “ H
(2.3)
and for all coordinate-wise increasing functions F : Nk ÞÑ R and G : Nl ÞÑ R it satisfies
ErF pN pA1 q, ¨ ¨ ¨ , N pAk qqGpN pB1 q, ¨ ¨ ¨ , N pBl qqs
ď ErF pN pA1 q, ¨ ¨ ¨ , N pAk qqsErGpN pB1 q, ¨ ¨ ¨ , N pBl qqs. (2.4)
Similarly, a point process is said to be positively associated (PA for short) if it satisfies the reverse inequality
for all families of pairwise disjoint Borel sets pAi q1ďiďk and pBi q1ďiďl (but not necessarily satisfying (2.3)).
If a point process is NA or PA it is said to be associated.
The main difference between the definition of PA and NA is the restriction (2.3) that only affects NA
point processes. Notice that without (2.3), ErN pAqs2 ď ErN pAq2 s contradicts (2.4) hence the need to consider
functions depending on disjoint sets for NA point processes.
These inequalities extend to the more general case of functionals of point processes. The first thing we
need is a more general notion of increasing functions. We associate to Ω the partial order X ď Y iff X Ă Y
and then call a function on Ω increasing if it is increasing respective to this partial order. The association
property can then be extended to these functions. A proof in a general setting can be found in [31, Lemma
3.6] but we give an alternative elementary one in Appendix A.
Theorem 2.3. Let X be a NA point process on Rd and A, B be disjoint subsets of Rd . Let F : Ω ÞÑ R and
G : Ω ÞÑ R be bounded increasing functions, then
ErF pX ∩ AqGpX ∩ Bqs ď ErF pX ∩ AqsErGpX ∩ Bqs.
(2.5)
If X is PA then, for all A, B Ă Rd not necessarily disjoint,
ErF pX ∩ AqGpX ∩ Bqs ě ErF pX ∩ AqsErGpX ∩ Bqs.
(2.6)
Association is a very strong dependence condition. As proved in the following theorem, it implies a strong
covariance inequality that is only controlled by the behaviour of the first two intensity functions of X
(assuming their existence). To state this result, we need to introduce the following seminorm for functionals
over point processes.
Definition 2.4.
For any A Ă Rd , }.}A is the seminorm on the functions f : Ω ÞÑ C defined by
}f }A :“
sup
|f pXq ´ f pX ∪ txuq|.
XPΩ,XĂA
xPA
3
Note that }.}A is a Lipschitz norm in the sense that it controls the way f pXq changes when a point is
added to X ∩ A.
Theorem 2.5. Let X be an associated point process and A, B Ă Rd be two disjoint bounded subsets. Let
f : Ω Ñ R and g : Ω Ñ R be two functions such that f pX ∩ Aq and gpX ∩ Bq are bounded, then
|Covpf pX ∩ Aq, gpX ∩ Bqq| ď }f }A }g}B |CovpN pAq, N pBqq|.
(2.7)
Moreover, if X is PA then it also satisfies the same inequality for all A, B Ă Rd not necessarily disjoint.
Proof. The proof mimics the one from [9] for associated random fields. We only consider the case of NA
point processes but the PA case can be treated in the same way.
Consider the functions f` , f´ : Ω Ñ R, EpAq-measurable, and g` , g´ : Ω Ñ R, EpBq-measurable, defined by
"
f˘ pXq “ f pX ∩ Aq ˘ }f }A N pAq,
g˘ pXq “ gpX ∩ Bq ˘ }g}B N pBq.
For all x P AzX, f` pX∪txuq´f` pXq “ f pX∪txu∩Aq´f pX∩Aq`}f }A which is positive by definition of }f }A .
f` is thus an increasing function. With the same reasoning, g` is also increasing and f´ , g´ are decreasing.
f` is not bounded but it is non-negative and almost surely finite so it can be seen as an increasing limit of the
sequence of functions minpf` , kq when k goes to infinity. These functions are non-negative, increasing and
bounded so for any k and any bounded increasing function g, (2.5) applies where f is replaced by minpf` , kq.
By a limiting argument, the same inequality holds true for f “ f` . We can also treat the other functions
the same way and we get from (2.5)
Covpf` pXq, g` pXqq ď 0 and Covpf´ pXq, g´ pXqq ď 0.
Since these expressions are equal to
Covpf˘ pXq, g˘ pXqq “ Covpf pX ∩ Aq, gpX ∩ Bqq ` }f }A }g}B CovpN pAq, N pBqq
˘ p}g}B Covpf pX ∩ Aq, N pBqq ` }f }A CovpN pAq, gpX ∩ Bqqq,
adding these two expressions together yields the upper bound in (2.7):
Covpf pX ∩ Aq, gpX ∩ Bqq ď ´}f }A }g}B CovpN pAq, N pBqq.
The lower bound is obtained by replacing f by ´f in the previous expression.
A similar inequality as in Theorem 2.5 can also be obtained for complex-valued functions since }<pf q}A
and }=pf q}A are bounded by }f }A , where <pf q and =pf q refer to the real and imaginary part of f respectively.
Corollary 2.6. Let X be an associated point process and A, B Ă Rd be two disjoint bounded subsets. Let
f : Ω Ñ C and g : Ω Ñ C be two functions such that f pX ∩ Aq and gpX ∩ Bq are bounded, then
|Covpf pX ∩ Aq, gpX ∩ Bqq| ď 4}f }A }g}B |CovpN pAq, N pBqq|.
Moreover, if X is PA then it also satisfies the same inequality for all A, B Ă Rd not necessarily disjoint.
If the first two intensity functions of X are well-defined then D in (2.1) is well-defined. As a consequence
of Theorem 2.5 and from (2.2), if |Dpx, yq| vanishes fast enough when |y ´ x| goes to infinity then any
two events respectively in EpAq and EpBq will get closer to independence as distpA, Bq tends to infinity, as
specified by the following corollary.
4
Corollary 2.7. Let X be an associated point process on Rd whose first two intensity functions are welldefined. Let A, B be two bounded disjoint sets of Rd such that distpA, Bq ě r. Then, for all functions
f : Ω Ñ R and g : Ω Ñ R such that f pX ∩ Aq and gpX ∩ Bq are bounded,
ż8
|Covpf pX ∩ Aq, gpX ∩ Bqq| ď sd |A| }f }A }g}B
td´1 sup |Dpx, yq|dt,
(2.8)
r
|x´y|“t
where sd is the pd ´ 1q-dimensional area measure of the unit sphere in Rd . Moreover, if f and/or g are
complex-valued functions, the same inequality holds true with an extra factor 4 on the right hand side.
Proof. Consider A, B to be two bounded disjoint sets of Rd such that distpA, Bq ě r then, from (2.2),
ˇ
ˇż
ˇ
ˇ
Dpx, yqdxdy ˇˇ
|CovpN pAq, N pBqq| “ ˇˇ
AˆB
ż
ď |A| sup
|Dpx, yq|dy
xPA B
ż
ď |A| sup
|Dpx, yq|dy
xPA Bpx,rqc
ż
ď |A| sup
sup
|Dpx, uq|dy
xPA Bpx,rqc
uPRd
|u´x|“|y´x|
ż8
td´1 sup |Dpu, vq|dt.
ď |A|sd
r
|u´v|“t
The final result is then a consequence of Theorem 2.5 and Corollary 2.6.
2.3. Application to α-mixing
Let us first recall some generalities about mixing. Consider a probability space pX , F, Pq and A , B two sub
σ-algebras of F. The α-mixing coefficient is defined as the following measure of dependence between A and
B:
αpA , Bq :“ supt|PpA ∩ Bq ´ PpAqPpBq| : A P A , B P Bu.
In particular, A and B are independent iff αpA , Bq “ 0. This definition leads to the essential covariance
inequality due to Davydov [13] and later generalised by Rio [35]: For all random variables X, Y measurable
with respect to A and B respectively,
|CovpX, Y q| ď 8α1{r pA , Bq}X}p }Y }q , where p, q, r P r1, 8s and
1 1 1
` ` “ 1.
p q
r
(2.9)
This definition is adapted to random fields the following way (see [16] or [23]). Let Y “ pYi qiPZd be a random
fields on Zd and define
αp,q prq :“ suptαpσptYi , i P Auq, σptYi , i P Buqq : |A| ď p, |B| ď q, distpA, Bq ą ru
with the convention αp,8 prq “ supq αp,q prq. The coefficients αp,q prq describe how close two events happening
far enough from each other are from being independent. The parameters p and q play an important role
since, in general, we cannot get any information directly on the behaviour of α8,8 prq.
We can adapt this definition to point processes the following way. For a point process X on Rd , define
αp,q prq :“ suptαpEpAq, EpBqq : |A| ď p, |B| ď q, distpA, Bq ą ru
with the convention αp,8 prq “ supq αp,q prq.
As a consequence of Corollary 2.7, the α-mixing coefficients of an associated point process tend to 0 when
Dpx, yq vanishes fast enough as |y ´ x| goes to infinity. More precisely, we have the following inequalities.
5
Proposition 2.8. Let X be an associated point process on Rd whose first two intensity functions are
well-defined, then for all p, q ą 0,
$
α prq ď pq sup |Dpx, yq|,
’
& p,q
|x´y|ěr
ż8
’
td´1 sup |Dpx, yq|dt.
% αp,8 prq ď psd
r
Proof. We can write
(2.10)
|x´y|“t
αpEpAq, EpBqq “ sup Covp1A pX ∩ Aq, 1B pX ∩ Bqq
A PEpAq
BPEpBq
so Proposition 2.8 is a direct consequence of Theorem 2.5 and Corollary 2.7 applied to indicator functions.
3. Central limit theorem for associated point processes
Consider the lattice pxi qiPZd defined by xi “ R ¨ i, where R ą 0 is a fixed constant. We denote by Ci , i P Zd ,
the d-dimensional cube with centre xi and side length s, where s ě R is another fixed constant. Note that the
union of these cubes forms a covering of Rd . Let X be an associated point process and pfi qiPZd be a family
of real-valued measurable functions defined on Ω. We consider the centred random field pYi qiPZd defined by
Yi :“ fi pX ∩ Ci q ´ Erfi pX ∩ Ci qs, i P Zd ,
(3.1)
ř
and we are interested in this section by the asymptotic behavior of Sn :“ iPIn Yi , where pIn qnPN is a
sequence of strictly increasing finite domains of Zd .
As a consequence of Proposition 2.8, we could directly use one of the different CLT for α-mixing random
fields that already exist in the literature [7, 16, 23] to get the asymptotic distribution of Sn . But, the
coefficients αp,8 decreasing much slower than the coefficients αp,q , this would imply an unnecessary strong
2`δ
assumption on D. Precisely, this would require Dpx, yq to decay at a rate at least op|y ´ x|´2pd`εq δ q, where
ε ą 0 and δ is a positive constant depending on the behaviour of the moments of X. In the next theorem, we
bypass this issue by exploiting both the behaviour of the mixing coefficients αp,q when p ă 8 and q ă 8, and
the association property through inequality (2.8). We show that we can still get a CLT when Dpx, yq decays
2`δ
at a rate op|y ´ x|´pd`εq δ q. This improvement is important to include DPPs with a slow decaying kernel,
thus inducing more repulsiveness, such as Bessel-type kernels, see the applications to DPPs in Section 4.2 and
especially the discussion at the end of the section. Let us also remark that another technique, based on the
convergence of moments, is sometimes used to establish a CLT for point processes. This has been exploited
especially for Brillinger mixing point processes in [28, 25] and other papers. As an example, DPPs have been
proved to be Brillinger mixing in [3, 24]. However, this condition applies to stationary point processes only.
Theorem 3.1. Consider
ř the random field Y given by (3.1), a sequence pIn qnPN of strictly increasing finite
domains of Zd and Sn “ iPIn Yi . Let σn2 :“ VarpSn q. Assume that for some ε, δ ą 0 the following conditions
are satisfied:
(C1) X is an associated point process on Rd whose first two intensity functions are well-defined;
(C2) supiPZd }Yi }2`δ “ M ă 8;
2`δ
(C3) sup|x´y|ěr |Dpx, yq| “ o pr´pd`εq δ q where D is given by (2.1);
rÑ8
(C4) lim inf n |In |´1 σn2 ą 0.
Then
1
L
Sn ÝÑ N p0, 1q.
σn
6
Proof. First, we notice that
S Y inherits its strong mixing coefficients from X. This is due to the fact that we
have σptYi : i P Iuq Ă Ep iPI Ci q for all I Ă Zd as a consequenceSof (3.1). Moreover, we have distpCi , Cj q ě
d
?1 p|i ´ j|1 R ´ sdq as a consequence of Lemma B.2, and since |
iPI Ci | ď s |I|, this gives us the inequality
d
˙
ˆ
sd
1
Y
X
?
, αp,q prq ď αpsd ,qsd
prR ´ sdq ,
@p, q ą 0, @r ą
R
d
where we denote by αX , αY the α-mixing coefficients of X and Y respectively. In particular, conditions (C1),
(C3) and identity (2.10) yields
Y
@p, q ą 0, αp,q
prq “ o pr´pd`εq
rÑ8
2`δ
δ
(3.2)
q.
We deal with the proof in two steps: first, we consider the case of bounded variables and then we extend
the result to the more general case.
The first step of the proof follows the approach used by Bolthausen [7] and Guyon [23], while the second
step exploits elements from [27]. The main difference lies in the way we deal with the term A3 that appears
later on in the proof.
First step: Bounded variables. Without loss of generality, we consider that Erfi pX ∩Ci qs “ 0 for all i P Zd .
Y
Suppose that we have supi }Yi }8 :“ supi }fi p. ∩ Ci q}8 “ M ă 8 instead of Assumption (C2). Since αp,q
prq
´d
is non increasing in r and is a opr q by (3.2), we can choose a sequence prn qnPN such that
a
a
Y
αp,q
prn q |In | Ñ 0 and rn´d |In | Ñ 8.
For i P Zd , define
ÿ
Si,n “
Yj ,
˚
Si,n
“ Sn ´ Si,n ,
an “
jPIn
|i´j|1 ďrn
ÿ
ErYi Si,n s,
iPIn
1
S̄n “ ? Sn ,
an
1
S̄i,n “ ? Si,n .
an
ř
˚
s and, as a consequence of the typical covariance inequality (2.9)
We have σn2 “ VarpSn q “ an ` iPIn ErYi Si,n
for α-mixing random variables, we get
ˇ
ˇ
ˇÿ
ˇ
ÿ
ÿ
ÿ
ˇ
Y
Y
˚ ˇ
ErYi Si,n sˇ ď
|CovpYi , Yj q| ď 8M 2
α1,1
p|i ´ j|1 q ď 8M 2 |In |
|tk P Zd : |k|1 “ ru|α1,1
prq.
ˇ
ˇiPI
ˇ i,jPI
rąr
i,jPI
n
n
n
n
|i´j|1 ąrn
|i´j|1 ąrn
The number of k P Zd satisfying |k|1 “ r is bounded by 2p2r ` 1qd´1 . This is because each of the d ´ 1 first
coordinates of k takes its values in t´r, ¨ ¨ ¨ , ru and the last coordinate is fixed by the other ones, up to the
sign, since |k|1 “ r. Therefore,
ˇ
ˇ
ˇÿ
ˇ
ÿ
ˇ
˚ ˇ
Y
ErYi Si,n sˇ ď 16M 2 |In |
p2r ` 1qd´1 α1,1
prq.
ˇ
ˇiPI
ˇ
rąr
n
n
σn2
By Assumption (3.2), this quantity is op|In |q and thus
„ an as a consequence of Assumption (C4). We
then only need to prove the asymptotic normality of Ssn . Moreover, since supn ErSsn2 s ă 8 then this will be
a consequence of the following condition (see [5, 7])
lim Erpiλ ´ Ssn q exppiλSsn qs “ 0, @λ P R.
nÑ8
We can split this expression into piλ ´ Ssn q exppiλSsn q “ A1 ´ A2 ´ A3 where
¸
˜
$
ÿ
’
1
’
’
Yj Sj,n ,
A1 “ iλ exppiλSsn q 1 ´
’
’
an jPI
’
’
n
’
&
ÿ
`
˘
1
A2 “ ? exppiλSsn q
Yj 1 ´ iλSsj,n ´ expp´iλSsj,n q ,
an
’
’
jPIn
’
’
ÿ
`
˘
’
1
’
’
Yj exp iλpSsn ´ Ssj,n q .
’
% A3 “ ?an
jPI
n
7
ř
Y
It was proved by Bolthausen [7] that ErA21 s and Er|A2 |s vanish when n goes to infinity if rd´1 αp,q
prq ă 8
for p ` q ď 4 which is the case here. We show that ErA3 s vanishes at infinity using (2.8). Notice that we have
ˇ
¨
¨
˛˛ ˇ
ˇ
ˇ
ˇ
ˇ
ÿ
˚
˚
ˇ
ˇ
‹
‹
|In |
iλ
˚
ˇ
‹
‹
f
pX
∩
C
q,
exp
|ErA3 s| ď ? sup ˇˇCov ˚
f
pX
∩
C
q
?
j
j
k
k ‚‚ˇ .
˝
˝
an jPIn ˇ
an kPI
ˇ
n
ˇ
ˇ
|k´j|1 ąrn
Define the function
¨
˛
˚ iλ
gj : X ÞÑ exp ˚
˝ ?an
ÿ
kPIn
|k´j|1 ąrn
‹
fk pX ∩ Ck q‹
‚.
S
This function is bounded by 1 and EpBj q-measurable where Bj :“ kPIn , |k´j|1 ąrn Ck is a bounded set and
?
distpCj , Bj q ě pRrn ´ sdq{ d (see Lemma B.2). We have }fj }Cj ď 2M and for all X P Ω, for all x P Bj , if
we denote by Jx “ tk : x P Ck u the set of cubes that contain x then
ˇ
˜
¸ˇ
ˇ
ˇ 2λM |J |
iλ ÿ
ˇ
ˇ
x
|gj pX ∪ txuq ´ gj pXq| “ ˇ1 ´ exp ?
pfk pX ∩ Ck ∪ txuq ´ fk pX ∩ Ck qq ˇ ď ?
.
ˇ
ˇ
an kPJ
an
x
?
Lemma B.2 gives us the bound |Jx | ď p2sd{R ` 1q and thus }gj }Bj ď 2λM p2sd{R ` 1qd { an . Finally, using
Corollary 2.7 we get
ż8
4|In |sd
|ErA3 s| ď ?
|Cj |}fj }Cj }gj }Bj
td´1 sup |Dpx, yq|dt
an
|x´y|ět
distpBj ,Cj q
˙d
ˆ 2
ż8
2s d
|In |
`s λ
td´1 sup |Dpx, yq|dt.
(3.3)
ď 16sd M 2
R
an ?1 pRrn ´sdq
|x´y|ět
d
d
By assumption (C3) we have that td´1 sup|x´y|ět |Dpx, yq| is integrable and by assumption (C4) we have
|In | “ Opan q which shows that limnÑ8 ErA3 s “ 0 concluding the proof of the theorem for bounded variables.
Second step: General Case. For N ą 0, we define
ř
"
S1,n :“ iPIn pFN pYi q ´ ErFN pYi qsq where FN : x ÞÑ x1|x|ďN ,
ř
S2,n :“ iPIn pFrN pYi q ´ ErFrN pYi qsq where FrN : x ÞÑ x1|x|ąN .
L
Let σn2 pN q :“ VarpS1,n q, from the first step of the proof we have σn pN q´1 S1,n ÝÑ N p0, 1q. Let 1 ą γ ą
p1 ` dε p1 ` 2δ qq´1 and define CN :“ supi }Yi 1|Yi |ąN }2`δγ . By assumption (C2) we have that CN vanishes
when N Ñ 8 and by assumption (C4) we have that |In | ď cσn2 for a sufficiently large n, where c is a positive
constant. By (2.9),
1
1 ÿ
VarpS2,n q “ 2
CovpFrN pYi q, FrN pYj qq
2
σn
σn i,jPI
n
ď
ÿ
δγ
|In | 2
Y
CN sup
8α1,1
p|i ´ j|1 q 2`δγ
2
σn
iPIn jPI
n
2
ď 16 c CN
8
ÿ
δγ
Y
p2r ` 1qd´1 α1,1
prq 2`δγ .
r“0
δγ
Y
By assumption (C3) and the choice of γ we have p2r ` 1qd´1 α1,1
prq 2`δγ ă 8 so σn´1 S2,n converges in mean
square to 0 when N goes to infinity, uniformly in n. With the same reasoning, we also get the inequality
ř
8
ÿ
δγ
1
Y
|CovpS
,
S
q|
ď
16
c
M
C
p2r ` 1qd´1 α1,1
prq 2`δγ ,
1,n
2,n
N
2
σn
r“0
8
where the right hand side tends to 0 when N goes to infinity, uniformly in n. Hence σn2 pN q tends to σn2
uniformly in n as N goes to infinity.
Finally, for all constants ν ą 0 arbitrary small, we can choose N such that Erσn´1 |S2,n |s ď ν and |1 ´
σn pN q{σn | ď ν for all n sufficiently large. By looking at the characteristic function of σn´1 Sn we get
ˇ ˇ ˆ
ˆ ixS ˙ˇ ˇ ˆ ixS ˙
˙ˇ
ˇ ´ ixSn ¯
ˇ ˇˇ ˆ ixS1,n ˙
ixS2,n
ˇ ˇ
ˇ
ˇ
1,n
1,n
1 2ˇ
1 2ˇ
ˇ
´
x
x
´
ˇ
ˇ
ˇ
ˇ
ˇ
σ
pN
q
σ
pN
q
σ
σ
σ
´ e 2 ˇ ` ˇE e n ´ 1 ˇˇ
´E e n
ˇE e n ´ e 2 ˇ ď ˇE e n
ˇ ` ˇE e n
ˇ˙ ˇ
ˇ
ˆˇ
ˇ S1,n ˇ ˇ
σn pN q ˇˇ
ˇ
ˇ
ˇ
ď xE ˇ
1´
` op1q ` xν
σn pN q ˇ ˇ
σn ˇ
ď 2xν ` op1q
concluding the proof.
4. Application to determinantal point processes
In this section, we give a CLT for a wide class of functionals of DPPs. This result is a key tool for the
asymptotic inference of DPPs. As an application treated in Section 4.3, we get the consistency and the
asymptotic normality of the two-step estimation method of [39] for a parametric inhomogeneous DPP.
4.1. Negative association and α-mixing for DPPs
We recall that a DPP X on Rd is defined trough its intensity functions with respect to the Lebesgue measure
that must satisfy
@n P N, @x P pRd qn , ρn px1 , ¨ ¨ ¨ , xn q “ detpKrxsq with Krxs :“ pKpxi , xj qqi,jPt1,¨¨¨ ,nu .
The function K : pRd q2 Ñ C is called the kernel of X and is assumed to satisfy the following standard
general condition ensuring the existence of X.
Condition H: The function K : pRd q2 Ñ C is a locally square integrable hermitian measurable function
such that its associated integral operator K is locally of trace class with eigenvalues in r0, 1s.
This condition is not necessary for existence, in particular there are examples of DPPs having a non-hermitian
kernel. It is nonetheless very general and is assumed in most studies on DPPs. Basic properties of DPPs can
be found in [26, 37, 31]. In particular, from [21, Theorem 1.4] and [31, Theorem 3.7], we know that DPPs
are NA.
Theorem 4.1 ([21, 31]).
Let K satisfy Condition H, then a DPP with kernel K is NA.
By definition, for a DPP with kernel K we have Dpx, yq “ ´|Kpx, yq|2 where D is introduced in (2.1).
Hence, using the last theorem and Proposition 2.8 we get the following strong mixing coefficients of a DPP,
where we define
ωprq :“ sup |Kpx, yq|.
(4.1)
|x´y|ěr
Corollary 4.2.
Let X be a DPP with kernel K satisfying H. Then, for all p, q ą 0,
$
ş
sup
|Kpx, yq|2 ď pqωprq2 ,
’
& αp,q prq ď
AˆB
|A|ăp,|B|ăq
distpA,Bqąr
’
% α prq ď ps ş8 ω 2 ptqtd´1 dt.
p,8
d r
9
(4.2)
It is worth noticing that this result, and so the covariance inequality (2.7), is optimal in the sense that for a
wide class of DPPs, the α-mixing coefficient αp,q prq do not decay faster than sup|A|ăp,|B|ăq |CovpN pAq, N pBqq|
distpA,Bqąr
when r goes to infinity, as stated in the following proposition.
Proposition 4.3. Let X be a DPP with kernel K satisfying H. We further assume that K is bounded,
takes its values in R` and is such that }K} ă 1 where }.} is the operator norm. Then, for all p, q, r ą 0,
ż
ż
pp`qq}K}8
sup
|Kpx, yq|2 ď αp,q prq ď
sup
p1 ´ }K}q }K}
|Kpx, yq|2 .
(4.3)
|A|ăp,|B|ăq AˆB
distpA,Bqąr
|A|ăp,|B|ăq AˆB
distpA,Bqąr
Proof. The upper bound for αp,q prq is just the one in (4.2). The lower bound is obtained through void
probabilities. Let p, q, r ą 0 and A, B Ă Rd such that |A| ă p, |B| ă q and distpA, Bq ą r. By definition, for
any such sets A and B, αp,q prq ě |PpN pAq “ 0qPpN pBq “ 0q ´ PpN pA ∪ Bq “ 0q|. The void probabilities of
DPPs are known (see [37]) and equal to
¸
˜
ÿ TrpKn q
A
PpN pAq “ 0q “ exp ´
n
ně1
where KA is the projection of K on the set of square integrable functions f : A Ñ R. Moreover, PpN pAq “
0qPpN pBq “ 0q ´ PpN pA ∪ Bq “ 0q ě 0 by negative association, and we have
PpN pAq “ 0qPpN pBq “ 0q ´ PpN pA ∪ Bq “ 0q
˜
¸˜
˜
¸
¸
ÿ TrpKn q
ÿ TrpKn q ´ TrpKn q ´ TrpKn q
A∪B
A∪B
A
B
“ exp ´
exp
´1
n
n
ně1
ně1
¸
˜
ÿ TrpKn q ÿ TrpKn q ´ TrpKn q ´ TrpKn q
A∪B
A
B
A∪B
.
ě exp ´
n
n
ně1
ně1
(4.4)
Using the classical trace inequality we get
ż
n
TrpKA∪B
q ď }KA∪B }n´1 TrpKA∪B q ď }K}n´1
Kpx, xqdx ď }K}n´1 pp ` qq}K}8 ,
A∪B
thus
˜
ÿ TrpKn q
A∪B
exp ´
n
ně1
¸
ě p1 ´ }K}q
pp`qq}K}8
}K}
.
(4.5)
Moreover, since A and B are disjoint sets, we can write
ż
n
n
n
TrpKA∪B q ´ TrpKA q ´ TrpKB q “
Kpx1 , x2 q ¨ ¨ ¨ Kpxn´1 , xn qKpxn , x1 qdx1 ¨ ¨ ¨ dxn
pA∪Bqn
ż
´
Kpx1 , x2 q ¨ ¨ ¨ Kpxn´1 , xn qKpxn , x1 qdx1 ¨ ¨ ¨ dxn , (4.6)
An ∪B n
ş
which vanishes when n “ 1, is equal to 2 AˆB |Kpx, yq|2 when n “ 2 and is non-negative for n ě 3 since K is
assumed to be non-negative. Finally, by combining (4.4), (4.5) and (4.6) we get the lower bound in (4.3).
4.2. Central limit theorem for functionals of DPPs
We investigate the asymptotic distribution of functions that can be written as a sum over subsets of close
enough points of X, namely
ÿ
f pXq :“
f0 pSq,
(4.7)
SĂX
10
where f0 is a bounded function vanishing when diampSq ą τ for a certain fixed constant τ ą 0. The typical
example, encountered in asymptotic inference, concerns functions f0 that are supported on sets S having
exactly p elements, in which case (4.7) often takes the form
f pXq “
1
p! x
‰
ÿ
f0 px1 , ¨ ¨ ¨ , xp q,
(4.8)
1 ,¨¨¨ ,xp PX
where the sum is done over ordered p-tuples of X and the symbol ‰ means that we consider distinct points.
The asymptotic distribution of (4.8) has been investigated in [38] when p “ 1 and in [3] for general p and
stationary DPPs.
In the next theorem, we extend these settings to functionals like (4.7) applied to general non stationary
DPPs. Some discussion
and comments are provided after its proof. We use Minkwoski’s notation and write
S
A ‘ r for the set xPA Bpx, rq.
Theorem 4.4. Let X be a DPP associated to a kernel K that satisfies H and that is further bounded. Let
τ ą 0 and f : Ω Ñ R be a function of the form
ÿ
f pXq :“
f0 pSq
SĂX
where f0 is a bounded function vanishing when diampSq ą τ . Let pWn qnPN be a sequence of increasing subsets
of Rd such that |Wn | Ñ 8 and let σn2 :“ Varpf pX ∩ Wn qq. Assume that there exists ε ą 0 and ν ą 0 such
that the following conditions are satisfied:
(H1) |BWn ‘ pτ ` νq| “ op|Wn |q;
(H2) ωprq “ opr´pd`εq{2 q;
(H3) lim inf n |Wn |´1 σn2 ą 0.
Then,
1
L
pf pX ∩ Wn q ´ Erf pX ∩ Wn qsq ÝÑ N p0, 1q.
σn
Proof. In order to apply Theorem 3.1, we would like to rewrite f as a sum over cubes of a lattice. Unfortunately, for disjoint sets A, B Ă Rd , f pX ∩ Aq ` f pX ∩ Bq ‰ f pX ∩ pA ∪ Bqq in general. Instead, we apply
Theorem 3.1 to an auxiliary function, close to f , as follows. Define S 0 as the barycentre of the set S. We
write
ÿ
fW pXq “
f0 pSq1W pS 0 q
(4.9)
SĂX
for the sum over the subsets of points of X with barycentre in W Ă Rd . Now, we split Rd into ?
little cubes
the following way. Let C0 be a given d-dimensional cube with a given side-length 0 ă s ď ν{ d. For all
Ăn “ S
i P Zd , let Ci be the translation of C0 by the vector s ¨ i. Let In :“ ti : Ci ‘řτ Ă Wn u and W
iPIn Ci .
An illustration of these definitions is provided in Figure 1. Since fW
Ăn pXq “
iPIn fCi pXq and each fCi are
EpCi ‘ τ q-measurable then fW
Ăn is the ideal candidate to use Theorem 3.1 on. Thus, we first prove that the
difference between fW
Ăn and f pX ∩Wn q is asymptotically negligible and then that fW
Ăn satisfies the conditions
of Theorem 3.1.
First of all, notice
? that distpCi , BWn q ě τ for all i P In . Therefore, for any point in Wn at a distance
greater than τ ` s d from BWn , the cube Ci of side-length s containing it is at a distance at least τ from
Ăn and we get
BWn , hence it is one of the Ci in W
?
Ăn | ď |BWn ‘ pτ ` s dq|.
|Wn zW
Ăn |. Now,
Hence, by Assumption (H1), |Wn | „ |W
ÿ
f pX ∩ Wn q ´ fW
Ăn pXq “
SĂX∩Wn
11
0
f0 pSq1Wn zW
Ăn pS q.
(4.10)
Ăn . Here, the black border is BWn , the grey area corresponds to
Figure 1. Example of illustration of the definition of W
Ăn .
pBWn ‘ τ q ∩ Wn and the square lattice corresponds to W
Since f0 vanishes when two points of S are at distance further than τ , then the sum in (4.10) only concerns
Ăn q ‘ τ q ∩ Wn q. By Lemma B.6, the variance of f pX ∩ Wn q ´ f Ă pXq is then a
the subsets S of X ∩ ppWn zW
Wn
Ăn q ‘ τ |q, whence a op|Wn |q and finally a opσ 2 q by Assumption (H3). Therefore, σ ´1 pf pX ∩ Wn q ´
Op|pWn zW
n
n
Erf pX ∩ Wn qsq has the same limiting distribution as σn´1 pfW
Ăn pXq ´ ErfW
Ăn pXqsq. Moreover, we have
|CovpfW
Ăn pXqq| ď σn
Ăn pXq, f pX ∩ Wn q ´ fW
b
a
2
Varpf pX ∩ Wn q ´ fW
Ăn pXqq “ σn op |Wn |q “ opσn q
by Assumptions (H1), (H3) and Lemma B.6 proving that σn´1 pf pX ∩ Wn q ´ Erf pX ∩ Wn qsq has the same
´1{2
limiting distribution as VarpfW
pfW
Ăn pXqq
Ăn pXq ´ ErfW
Ăn pXqsq.
We conclude by showing that the random variables Yi “ fCi pXq ´ ErfCi pXqs satisfy the assumptions of
Theorem 3.1. A rough bound on f gives us |fCi pXq| ď }f0 }8 2N pCi ‘τ q so, by Lemma B.5,
@n P N,
sup Er|Yi |n s ă 8.
iPZd
This means that the Yi ’s satisfy Assumption (C2) for all δ ą 0 and thus (C3) as a consequence of (H2).
Ăn | “ Op|Wn |q and Varpf Ă pXqq „ σ 2 , we have
Finally, since |In | “ s´d |W
n
Wn
lim inf |In |´1 VarpfW
Ăn pXqq ą 0
n
by Assumption (H3), which concludes the proof of the theorem.
We highlight some extensions of this result.
i) Since the superposition of independent PA (respectively NA) point processes remains a PA (respectively
NA) point process, then Theorem 4.4 holds true for α-determinantal point processes where α P t´1{m :
m P N˚ u, see [37] for more information about α-DPPs.
ii) Theorem 4.4 also extends to Rq -valued functions f where q ě 2. Let Σn :“ Varpf pX ∩ Wn qq. If we
replace (H3) by
lim inf |Wn |´1 λmin pΣn q ą 0
n
where λmin pΣn q denotes the smallest eigenvalue of Σn , then Theorem 4.4 holds true with the conclusion
L
Σ´1{2
pf pX ∩ Wn q ´ Erf pX ∩ Wn qsq ÝÑ N p0, Idq q
n
where Idq is the q ˆ q identity matrix. Since Σn does not necessary converge, this result is not a direct
application of the Cramér-Wold device. Instead, a detailed proof is given in [5].
12
iii) In (4.7), f0 only depends on finite subsets of Rd and not on the order of the points in each subset.
Nonetheless, we can easily extend (4.7) to functions of the form
f pXq “
‰
ÿ
ÿ 1
n! x
ně0
f0 px1 , ¨ ¨ ¨ , xn q
1 ,¨¨¨ ,xn PX
S
where f0 is a bounded function on ně0 pRd qn that vanishes when two of its coordinates are at a
distance greater than τ . Then f still satisfy Theorem 4.4. This is because we can write
ÿ sym
f0 pSq
f pXq “
SĂX
where f0sym is the symmetrization of f0 defined by
f0sym ptx1 , ¨ ¨ ¨ , xn uq :“
1 ÿ
f0 pxσp1q , ¨ ¨ ¨ , xσpnq q
n! σPS
n
where Sn is the symmetric group on t1, ¨ ¨ ¨ , nu. Since f0sym is also bounded and vanishes when
diampSq ą τ then it satisfies the required assumptions for Theorem 4.4.
Let us comment the assumptions of Theorem 4.4.
• Condition (H1) makes clear the idea that Wn must grow to Rd as n Ñ 8, without being a too irregular
p1q
pdq
set. In the simple case where Wn is the Cartesian product of intervals, i.e. Wn “ ∆n ˆ ¨ ¨ ¨ ˆ ∆n ,
pkq
then (H1) is equivalent to |∆n | Ñ 8 for all k.
• Condition (H2) is not really restrictive and is satisfied by all classical kernel families. For example, the
kernels of the Ginibre ensemble and of the Gaussian unitary ensemble (see [26]) have an exponential
decay. Moreover, all translation-invariant kernels used in spatial statistics (see [30] and [4]) satisfy
ωprq “ Opr´pd`1q{2 q: the Gaussian and the Laguerre-Gaussian covariance functions have an exponential
decay; the Whittle-Matérn and the Cauchy covariance functions satisfy ωprq “ opr´d q; and in the case
of the most repulsive DPP in dimension d (as determined in [30, 4]), which is the slowest decaying
Bessel-type kernel, its kernel is given by
b
?
1
1
´ d`1 ¯
ρΓp d2 ` 1q J d p2 πΓp d2 ` 1q d ρ d ||y ´ x||q
2
Kpx, yq “
ñ ωprq “ O r´ 2 ,
d
d
π4
||y ´ x|| 2
where ρ ą 0 is a constant. While this DPP satisfies Condition (H2), we point out that its α-mixing
coefficients decay too slowly to be able to derive a CLT only from them, see the discussion before
Theorem 3.1. This justifies the importance of Condition (C3) in this theorem, obtained by the NA
property, and which leads to Condition (H2).
• Condition (H3) is harder to control in the broad setting of Theorem 4.4, but we can get sufficient
conditions in some particular cases. For example, if f0 pSq “ 1|S|“1 and K is a translation-invariant
continuous kernel then it was shown in [38] that Condition (H3) holds when K is not the Fourier
transform of an indicator function. In the peculiar case where K is the Fourier transform of an indicator
function, [38] proved that the limiting distribution is still Gaussian but the rate of convergence is
different. As another example extending the previous one, assume that f0 is a non-negative function
supported on the set tS Ă X : |S| “ pu for a given integer p ą 0 and assume that the highest eigenvalue
of the integral operator K associated to K is less than 1. Then, we show in Proposition B.7 that
ż
1
lim inf
f0 pxq detpKrxsqdx ą 0
n
|Wn | Wnp
implies (H3) and is much easier to verify.
13
4.3. Application to the two-step estimation of an inhomogeneous DPP
In this section, we consider DPPs on R2 with kernel of the form
b
b
Kβ,ψ px, yq “ ρβ pxqCψ py ´ xq ρβ pyq,
@x, y P R2 ,
(4.11)
where β P Rp and ψ P Rq are two parameters, Cψ is a correlation function and ρβ is of the form ρβ pxq “
ρpzpxqβ T q where ρ is a known positive strictly increasing function and z is a p-variate bounded function
called covariates. This form implies that the first order intensity, corresponding to ρβ pxq, is inhomogeneous
and depends on the covariates zpxq through the parameter β. But all higher order intensity functions once
normalized, i.e. ρpnq px1 , . . . , xn q{pρβ px1 q . . . ρβ pxn qq, are translation-invariant for n ě 2. In particular, the
pair correlation (the case n “ 2) is invariant by translation. This kind of inhomogeneity is sometimes
named second-order intensity reweighted stationarity and is frequently assumed in the spatial point process
community.
Existence of DPPs with a kernel like above is for instance ensured if ρβ pxq is bounded by ρmax and Cψ
is a continuous, square-integrable correlation function on Rd whose Fourier transform is less than 1{ρmax ,
see [30]. For later use, we call H1 the previous assumptions on Kβ,ψ .
Consider the observation of a DPP X with kernel Kβ ˚ ,ψ˚ , along with the covariates z, within a window
Wn :“ ran, bns ˆ rcn, dns where b ą a and d ą c. Waagepetersen and Guan [39] have proposed the following
two-step estimation procedure of pβ ˚ , ψ ˚ q for second-order intensity reweighted stationary models. First, β̂n
is obtained by solving
ÿ ∇ρβ puq ż
´
∇ρβ puqdu “ 0.
un,1 pβq :“
ρβ puq
Wn
uPX∩W
n
where ∇ρβ denotes the gradient with respect to β. In the second step, ψ̂n is obtained by minimizing mn,β̂n
where
¸c
¸2
ż r ˜˜ ÿ
1t0ă|u´v|ďtu
c
mn,β pψq :“
´ Kψ ptq
dt.
ρβ puqρβ pvq|Wn ∩ Wn,u´v |
rl
u,vPX∩W
n
Here rl , r and c are user-specified non-negative constants, Wn,u´v is Wn translated by u ´ v and Kψ is the
Ripley K-function defined by
ż
Kψ ptq :“
gψ puqdu
}u}ďt
2
2
where gψ puq :“ 1 ´ Cψ puq {Cψ p0q is the pair correlation function of X. If we define
un,2 pβ, ψq :“ ´|Wn |
Bmn,β pψq
,
Bψ
then the two-step procedure amounts to solve
un pβ, ψq :“ pun,1 pβq, un,2 pβ, ψqq “ 0.
The asymptotic properties of this two-step procedure are established in [39], under various moments and
mixing assumptions, with a view to inference for Cox processes. We state hereafter the asymptotic normality
of pβ̂n , ψ̂n q in the case of DPPs with kernel of the form (4.11). This setting allows us to apply Theorem 4.4
and get rid of some restrictive mixing assumptions needed in [39].
The asymptotic covariance matrix of pβ̂n , ψ̂n q depends on two matrices defined in [39, Section 3.1], where
r n and In . We do not reproduce their expression, which is hardly tractable. An assumpthey are denoted by Σ
tion in [39] ensures the asymptotic non-degeneracy of this covariance matrix and we also need this assumption
in our case, see (W4) below. Unfortunately, as discussed in [39], it is hard to check this assumption for a
given model, particularly because it depends on the covariates z. We are confronted by the same limitation
in our setting. On the other hand, the other assumptions of the following theorem are not restrictive. In
particular almost all standard kernels satisfy (W3) below, see the discussion after Theorem 4.4.
14
Theorem 4.5. Let X be a DPP with kernel Kβ ˚ ,ψ˚ given by (4.11) and satisfying H1 . Let pβ̂n , ψ̂n q the
two-step estimator defined above. We assume the following.
(W1)
(W2)
(W3)
(W4)
rl ą 0 if c ă 1; otherwise rl ě 0,
ρβ and Kψ are twice continuously differentiable as functions of β and ψ,
sup}x}ěr Cψ˚ pxq “ Opr´1´ε q,
r n ) is satisfied.
Condition N3 in [39] (concerning the matrices In and Σ
Then, there exists a sequence tpβ̂n , ψ̂n q : n ě 1u for which un pβ̂n , ψ̂n q “ 0 with a probability tending to one
and
L
r ´1{2 ÝÑ
|Wn |1{2 rpβ̂n , ψ̂n q ´ pβ ˚ , ψ ˚ qsIn Σ
N p0, Idq.
n
Proof. Let ρk be the kth intensity function of the DPP with kernel px, yq ÞÑ Cψ˚ py ´ xq. In order to apply
Theorem 1 in [39] we need to show that
ş
2
(i) ρ2 , ρ3 are bounded and there
ş is a constant M such that for all u1 , u2 P R , |ρ3 p0, v, v ` u1 q ´
ρ1 p0qρ2 p0, u1 q|dv ă M and |ρ4 p0, u1 , v, v ` u2 q ´ ρ2 p0, u1 qρ2 p0, u2 q|dv ă M ,
(ii) }ρ4`2δ }8 ă 8 for some δ ą 0,
(iii) αa,8 prq “ Opr´d q for some a ą 8r2 and d ą 2p2 ` δq{δ.
The first property (i) is a consequence of (W3). This is because we can write
|ρ3 p0, v, v ` u1 q ´ ρ1 p0qρ2 p0, u1 q| “ |2Cψ˚ pvqCψ˚ pu1 qCψ˚ pv ` u1 q ´ Cψ˚ p0qpCψ˚ pv ` u1 q2 ` Cψ˚ pvq2 q|
which is bounded by 2|Cψ˚ p0q|pCψ˚ pv ` u1 q2 ` Cψ˚ pvq2 q and
ż
ż8
Cψ˚ pvq2 dv ď 2π
r sup |Cψ˚ pxq|2 dr
R2
0
}x}“r
which is finite by Assumption (W3). The term ρ4 p0, u1 , v, v ` u2 q ´ ρ2 p0, u1 qρ2 p0, u2 q can be treated the
same way. For a DPP, (ii) is satisfied for any δ ą 0. Finally, (iii) is the one that causes an issue since, as
stated before, the α-mixing coefficient we get in Corollary 4.2 decreases slower than what we desire. But,
the only place this assumption is used in [39] is to prove the asymptotic normality of their estimator in their
Lemma 5, which can also be derived as a consequence of our Theorem 4.4 with Assumption (W3).
Appendix A: Proof of Theorem 2.3
We use the following variant of the monotone class theorem (see [14, Theorem 22.1]).
Theorem A.1. Let S be a set of bounded functions stable by bounded monotone convergence and uniform
convergence. Let C be a subspace of S such that C is an algebra containing the constant function r
1. Then, S
contains all bounded functions measurable over σpCq.
Now, let A, B1 , ¨ ¨ ¨ , Bk be pairwise distinct Borel subsets of Rd and g : Nk ÞÑ R be a coordinate-wise
increasing function. We denote by ΩA the set of locally finite point configurations in A and we define S as
the set of functions f : ΩA ÞÑ R such that
ErfˆpX ∩ AqgpN pB1 q, ¨ ¨ ¨ , N pBk qqs ď ErfˆpX ∩ AqsErgpN pB1 q, ¨ ¨ ¨ , N pBk qs,
(A.1)
where fˆpXq :“ supY ďX f pY q. Note that fˆ is an increasing function and that f is increasing iff fˆ “ f . Our
goal is to prove that S contains all bounded functions supported over A. Because of the definition of NA
point processes (2.4), we know that S contains the set C of functions of the form f pN pA1 q, ¨ ¨ ¨ , N pAk qq where
the Ai are pairwise disjoints Borel subsets of A. In particular, since point processes over A are generated by
the set of random vectors tpN pA1 q, ¨ ¨ ¨ , N pAk qq : Ai Ă A disjoints, k P Nu, then we only need to verify that
S and C satisfy the hypothesis of Theorem A.1 to conclude.
15
• Stability of S by bounded monotonic convergence: Since (A.1) is invariant if we add a constant to f
and f is bounded then we can consider f to be positive. Now, notice that for all functions h and k,
h ď k ñ ĥ ď k̂ and h ě k ñ ĥ ě k̂. So, if we take a positive bounded monotonic sequence fn P S
that converges to a bounded function f , then fˆn is also a positive bounded monotonic sequence that
consequently converges to a function g. Suppose that pfn qn is an increasing sequence (the decreasing
case can be treated similarly) and let us show that g “ fˆ. Let X P ΩA , for all Y Ă X, fn pY q ď f pY q.
Taking the supremum then the limit gives us gpXq ď fˆpXq. Moreover, for all Y Ă X, gpXq ě fˆn pXq ě
fn pY q. Taking the limit gives us that gpXq ě f pY q for all Y Ă X so gpXq ě fˆpXq which proves that
g “ fˆ. Using the monotone convergence theorem we conclude that
Erfˆn pX ∩ Aqs Ñ ErfˆpX ∩ Aqs
and Erfˆn pX ∩ AqgpN pB1 q, ¨ ¨ ¨ , N pBk qqs Ñ ErfˆpX ∩ AqgpN pB1 q, ¨ ¨ ¨ , N pBk qqs,
(A.2)
which proves that f P S
• Stability of S by uniform convergence: Let fn be a sequence over S converging uniformly to a function
f then, by Lemma B.1, fˆn also converges uniformly (and therefore in L1 ) to fˆ. As a consequence, (A.2)
is also satisfied in this case so f P S.
1 so we only need to prove that C
• C is an algebra: It is easily shown that C is a linear space containing r
is stable by multiplication. Let A1 , ¨ ¨ ¨ , Ar and A11 ¨ ¨ ¨ , A1s be two sequences of pairwise distinct Borel
subsets of A. Let f “ f pN pA1 q, ¨ ¨ ¨ , N pAr qq P C and h “ hpN pA11 q, ¨ ¨ ¨ , N pA1r qq P C. We can write
ÿ
ÿ
N pAi q “ N pAi z ∪j A1j q ` N pAi ∩ A1j q and N pA1i q “ N pA1i z ∪j Aj q ` N pA1i ∩ Aj q,
j
j
so f ¨ h can be expressed as a function of the number of points in the subsets Ai z ∪j A1j , A1i z ∪j Aj and
Ai ∩ A1j that are all pairwise distinct Borel subsets of A, proving that C is stable by multiplication.
This concludes the proof that S contains all bounded functions supported over ΩA . By doing the same
exact reasoning on the set of bounded functions g satisfying Erf pX ∩ AqgpX ∩ Bqs ď Erf pX ∩ AqsErgpX ∩ Bqs
for a fixed f we obtain the same result which concludes the proof.
Appendix B: Auxiliary results
Lemma B.1.
Let E be a set and f, g : E Ñ R be two functions, then
ˇ
ˇ
ˇ
ˇ
ˇsup f pxq ´ sup gpyqˇ ď }f ´ g}8
ˇ
ˇ
xPE
yPE
Proof. The proposition becomes trivial once we write
f pxq ď gpxq ` }f ´ g}8 ď sup gpyq ` }f ´ g}8
yPE
Taking the supremum yields the first inequality. Moreover, by symmetry of f and g the second one follows
similarly.
řd
Lemma B.2. Let i, j P Zd such that |i ´ j|1 :“ l“1 |il ´ jl | “ r. Let s, R ą 0 and Ci , Cj be the ddimensional cubes with side length s and respective centre xi “ R ¨ i and xj “ R ¨ j. Then,
1
distpCi , Cj q ě ? prR ´ sdq.
d
Moreover, each cube intersects at most p2sd{Rqd other cubes with centers on R ¨ Zd and side length s.
16
?
Proof. Since each point of a d-dimensional
square with side length s is
a
? at distance at most s d{2 from
pRid ´?
Rjd q2 ´ s d which takes its minimum when
its centre, we get distpCi , Cj q ě pRi1 ´ Rj1 q2 ` ¨ ¨ ¨ ` ?
|il ´ jl | “ r{d for all 1 ď l ď d hence distpCi , Cj q ě rR{ d ´ s d.
In particular, if |i ´ j|1 ą sd{R then Ci ∩ Cj “ H, hence for all i P Zd
ˆ
d
|tj P Z : Ci ∩ Cj ‰ H, i ‰ ju| ď |tj : 0 ă |i ´ j|1 ď sd{Ru| ď
˙d
2sd
`1
R
Lemma B.3. Let M and N be two n ˆ n semi-positive definite matrices such that 0 ď M ď N ´1 where ď
denotes the Loewner order. Then,
detpId ´ M N q ě 1 ´ TrpM N q
Proof. First, let us consider the case where N “ Id. If TrpM q ě 1 then detpId ´ M q ě 0 ě 1 ´ TrpM q.
Otherwise, we denote by SppM q the spectrum of M and since all eigenvalues are in [0,1[, we can write
ź
detpId ´ M q “
expplnp1 ´ λqq
λPSppM q
¸
˜
8
ź
ÿ
λn
“
exp ´
n
n“0
λPSppM q
¸
˜
8
ÿ
TrpM n q
“ exp ´
n
n“0
¸
˜
8
ÿ
TrpM qn
ě exp ´
n
n“0
“ expplogp1 ´ TrpM qqq
“ 1 ´ TrpM q.
(B.1)
Getting back to the general case, we can write N as S T S and by Sylvester’s determinant identity we get
that detpId ´ M N q “ detpId ´ SM S T q. Since we assumed that 0 ď M ď N ´1 then 0 ď SM S T ď Id and
by applying (B.1) this concludes the proof:
detpId ´ M N q “ detpId ´ SM S T q ě 1 ´ TrpSM S T q “ 1 ´ TrpM N q.
ˆ
˙
M1 N
where
N T M2
M1 is a k ˆ k semi-definite positive matrix, M2 is a pn ´ kq ˆ pn ´ kq semi-definite positive matrix and N
is a k ˆ pn ´ kq matrix. We define ||A||8 :“ sup |ai,j | for any matrix A. Then,
Proposition B.4.
Let M be a n ˆ n semi-definite positive matrix of the form M “
0 ď detpM1 q detpM2 q ´ detpM q ď kpn ´ kqTrpN T N q||M ||n´2
8 .
Proof. First, we assume that M1 and M2 are invertible. Using Schur’s complement, we can write
detpM q “ detpM1 q detpM2 q detpId ´ M1´1 N T M2´1 N q
where 0 ď N T M2´1 N ď M1 with ď being the Loewner order. N T M2´1 N being semi-definite positive implies
detpM q ď detpM1 q detpM2 q,
17
while the inequality N T M2´1 N ď M1 gives us (see Lemma B.3)
detpM q ě detpM1 q detpM2 qp1 ´ TrpM1´1 N T M2´1 N qq.
Therefore,
0 ď detpM1 q detpM2 q ´ detpM q ď TrpadjpM1 qN T adjpM2 qN q
ď TrpadjpM1 qqTrpadjpM2 qqTrpN T N q “
k
ÿ
∆i pM1 q
i“1
n´k
ÿ
∆j pM2 qTrpN T N q,
j“1
where ∆i pM1 q means the pi, iq minor of the matrix M1 and adjpM1 q is the transpose of the matrix of cofactor
of M1 . But, since all principal sub-matrices of M1 and M2 are positive definiteśmatrices then their determinant
is lower than the product of their diagonal entries, meaning that ∆i pM1 q ď j‰i M1 pj, jq ď ||M ||k´1
8 . Doing
the same thing for the terms ∆j pM2 q gives us the desired result.
If M1 or M2 is not invertible, a limit argument using the continuity of the determinant leads to the same
conclusion.
Lemma B.5.
Let X be a DPP with bounded kernel K satisfying H, s ą 0 and n ą 0, then
Er2nN pAq s ă 8
sup
AĂRd ,|A|“s
Proof. Let n P N and A Ă Rd such that |A| “ s. Since the determinant of a positive semi-definite matrix is
always smaller than the product of its diagonal coefficients we get
ff
«
˙
8 ˆ
ÿ
N pAq
n
k
nN pAq
p2 ´ 1q
Er2
s“E
k
k“0
ż
8
ÿ
p2n ´ 1qk
“
detpKrxsqdx
k!
Ak
k“0
ď ep2
n
´1q}K}8 |A|
ă 8.
d`ε
Lemma B.6. Let X be a DPP on Rd with bounded kernel K satisfying H such that ωprqS“ Opr´ 2 q for
a certain ε ą 0. Then, for all bounded Borel sets W Ă Rd and all bounded functions g : pą0 pRd qp Ñ R
such that gpSq vanishes when diampSq ą τ for a given constant τ ą 0,
˜
¸
ÿ
Var
gpSq “ Op|W |q.
(B.2)
SĂX∩W
Proof. Since W is bounded then N pW q is almost surely finite and we can write
ÿ ÿ
ÿ
gpSq “
gpSq a.s.
SĂX∩W
pě0 SĂX∩W
|S|“p
18
¸2 fi
»˜
ÿ
Looking at the variance of each term individually, we start by developing E –
gpSq1|S|“p
fl as
SĂX∩W
»
p
ÿ
k“0
—
E—
–
fi
ÿ
S,T ĂX∩W
|S|“|T |“p,|S∩T |“k
ffi
gpSqgpT qffi
fl
(B.3)
»
p
ÿ
“
k“0
—
E—
–
p
ÿ
“
fi
ÿ
ÿ
U ĂX∩W
S 1 ĂSĂU
|U |“2p´k |S 1 |“k,|S|“p
1
p2p
´
kq!
k“0
ż
ffi
gpSqgpS 1 ∪ pU zSqqffi
fl
ÿ
W 2p´k
gpSqgpS 1 ∪ pU zSqqρ2p´k px1 , ¨ ¨ ¨ , x2p´k qdx1 ¨ ¨ ¨ dx2p´k
S 1 ĂSĂtx1 ,¨¨¨ ,x2p´k u
|S 1 |“k,|S|“p
p
ÿ
ˆ ˙ˆ
˙ż
p 2p ´ k
1
gpx1 , ¨ ¨ ¨ , xp qgpx1 , ¨ ¨ ¨ , xk , xp`1 , ¨ ¨ ¨ , x2p´k qρ2p´k pxqdx.
“
p2p ´ kq! k
p
W 2p´k
k“0
(B.4)
Since the determinant of a positive semi-definite matrix is smaller than the product of its diagonal terms,
we have |ρ2p´k pxq| ď }K}2p´k
. Moreover, as a consequence of our assumptions on g, each term for k ě 1
8
in (B.4) is bounded by
ˆ ˙ ż
ˆ ˙
1
p
|W | p
2
2p´k
}g}8 }K}8 1t0ď|xi ´x1 |ďτ, @iu dx ď
}g}28 }K}2p´k
|Bp0, τ q|2p´k´1
8
p!pp ´ kq! k
p! k
W 2p´k
ˆ ˙
|W | p
}g}28 p1 ` }K}8 q2p p1 ` |Bp0, τ q|q2p .
ď
p! k
Hence,
»
p
ÿ
k“1
—
E—
–
fi
ÿ
S,T ĂX∩W
|S|“|T |“p,|S∩T |“k
p
ffi
2 C1
gpSqgpT qffi
ď
|W
|}g}
8
fl
p!
(B.5)
where C1 “ 2p1 ` }K}8 q2 p1 ` Bp0, τ qq2 is a constant independent from p and W . However, even if all terms
for k ě 1 in (B.4) are Op|W |q, this is not the case of the term for k “ 0 which is a Op|W |2 q. Instead of
controlling this term alone, we consider its difference with the remaining term in the variance we are looking
at, that is
1
pp!q2
¸ff2
«˜
ż
gpxqgpyqρ2p px, yqdxdy ´ E
W 2p
ÿ
gpSq1|S|“p
SĂX∩W
“
1
pp!q2
ż
gpxqgpyqpρ2p px, yq ´ ρp pxqρp pyqqdxdy.
W 2p
Using Proposition B.4, we get
|ρ2p px, yq ´ ρp pxqρp pyq| ď p2 }K}2p´2
8
ÿ
Kpxi , yj q2 .
1ďi,jďp
Now, notice that for all y P Rd and 1 ď i ď p,
ż
ż
1t0ă|xk ´xj |ďτ, @j,ku |Kpxi , yq|2 dx ď |Bp0, τ q|p´1
Wp
W
19
ż
|Kpxi , yq|2 dxi ď |Bp0, τ q|p´1 sd
rd´1 ωprq2 dr
Rd
which is finite because of our assumption on ωprq. Thus, we obtain the inequality
ż
ż
2
p´1
gpxqgpyq|Kpxi , yj q| dxdy ď }g}8 |Bp0, τ q|
gpxq|Kpxi , y1 q|2 dxdy1
W 2p
W p`1
ż
ď |W |}g}28 |Bp0, τ q|2p´2 sd
rd´1 ωprq2 dr.
(B.6)
Rd
By combining (B.5) and (B.6), we get the bound
¨
˛
˙
ˆ p
ÿ
C2
C1
˚
‹
`
Var ˝
gpSq‚ ď |W |}g}28
p!
p!
SĂX∩W
|S|“p
where
˜
C2 :“
|Bp0, τ q|2p´2
p4 }K}2p´2
8
sup
p!
pě0
is a constant independent from p and W . Finally,
¨
ÿ
˚ ÿ
Var ˝
pě0
¸
ż
rd´1 ωprq2 dr
sd
Rd
˛
‹
gpSq‚ “ Op|W |q
SĂX∩W
|S|“p
and
¨
˛
dˆ
ÿ
pąqě0
˚
Cov ˝
ÿ
gpSq,
SĂX∩W
|S|“p
ÿ
ÿ
‹
gpSq‚ ď |W |}g}28
p,qě0
SĂX∩W
|S|“q
C1p
C2
`
p!
p!
˙ˆ
C1q
C2
`
q!
q!
˙
“ Op|W |q
concluding the proof.
Proposition B.7.
Let p P N, f : Rp Ñ R` be a symmetrical measurable function and define
ÿ
F pXq “
f pSq.
SĂX
|S|“p
Let X be a DPP with kernel K satisfying Condition H such that }K} ă 1 where }K} is the operator norm of
the integral operator associated with K. If, for a given increasing sequence of compact sets Wn Ă Rd ,
ż
1
lim inf
f pxq detpKrxsqdx ą 0,
(B.7)
n
|Wn | Wnp
then
lim inf
n
1
VarpF pX ∩ Wn qq ą 0.
|Wn |
Proof. Let W be a compact subset of Rd . The Cauchy-Schwartz inequality gives us
CovpF pX ∩ W q, N pW qq2 ď VarpF pX ∩ W qqVarpN pW qq.
20
We showed in Lemma B.6 that |W |´1 VarpN pW qq is bounded by a constant C ą 0 so we are only interested
in the behaviour of CovpF pX ∩ W q, N pW qq. We start by developing ErF pX ∩ W qN pW qs:
fi
»
ÿ
ffi
— ÿ
1fl
f pSq
ErF pX ∩ W qN pW qs “ E –
xPX∩W
SĂX∩W
|S|“p
fi
»
ÿ
—
“ E–
ÿ
ÿ
f pSztxuq ` p
ffi
f pSqfl
SĂX∩W
|S|“p
SĂX∩W xPS
|S|“p`1
ż
ż
p`1
ÿ
1
1
“
pf pxq detpKrxsqdx
f pzztzi uq detpKrzsqdz `
pp ` 1q! W p`1 i“1
p! W p
ˆż
ˆ
˙ ˙
ż
1
“
f pxq p detpKrxsq `
detpKrx, asqda dx .
p!
Wp
W
We also have
ErF pX ∩ W qsErN pW qs “
1
p!
ż
ż
f pxq detpKrxsqdx
Wp
Kpa, aqda,
W
hence
CovpF pX ∩ W q, N pW qq “
1
p!
ˆ
ż
f pxq detpKrxsq p ´
ż
Wp
`
˙
˘
Kpa, aq ´ detpKrx, asq detpKrxsq´1 da dx.
W
(B.8)
Using Schur’s complement, we get
T
Kpa, aq ´ detpKrx, asq detpKrxsq´1 “ Kax Krxs´1 Kax
(B.9)
where we define Kax as the vector pKpa, x1 q, ¨ ¨ ¨ , Kpa, xp qq. Moreover, since we look at our point process in
a compact window W , a well-known property of DPPs (see [26]) is that there exists a sequence of eigenvalues
λi in r0, }K}s and an orthonormal basis of L2 pW q of eigenfunctions φi such that
ÿ
Kpx, yq “
λi φi pxqφ̄i pyq @x, y P W.
i
As a consequence, @x, y P W ,
ż
Kpx, aqKpa, yqda “
W
ÿ
λ2i φi pxqφ̄i pyq
i
p
which we define as Lpx, yq. Therefore, for all x P W , Lrxs ď }K}Krxs where ď is the Loewner order for
positive definite symmetric matrices and we get
ˆ
˙
ż
ż
T
T
Kax Krxs´1 Kax
da “ Tr Krxs´1
Kax
Kax da “ TrpKrxs´1 Lrxsq ď p}K}.
(B.10)
W
W
Finally, since f is non negative, by combining (B.8), (B.9) and (B.10) we get the lower bound
p1 ´ }K}q2
CovpF pX ∩ W q, N pW qq2
ě
VarpF pX ∩ W qq ě
VarpN pW qq
Cpp ´ 1q!2 |W |
which proves the proposition.
21
˙2
ˆż
f pxq detpKrxsqdx
Wp
References
[1] K. Alam and K.M.L. Saxena. Positive dependence in multivariate distributions. Comm. Statist. Theory
Methods A, 10:1183–1196, 1981.
[2] R. Bardenet and A. Hardy. Monte Carlo with determinantal point processes. arXiv preprint
arXiv:1605.00361, 2016.
[3] C.A.N. Biscio and F. Lavancier. Brillinger mixing of determinantal point processes and statistical
applications. Electron. J. Statist., 10:582–607, 2016.
[4] C.A.N. Biscio and F. Lavancier. Quantifying repulsiveness of determinantal point processes. Bernoulli,
22:2001–2028, 2016.
[5] C.A.N. Biscio, A. Poinas, and R. Waagepetersen. A note on gaps in proofs of central limit theorems.
Statistics and Probability Letters, 135:7–10, 2018.
[6] B. Błaszczyszyn and D. Yogeshwaran. Clustering comparison of point processes with applications to
random geometric models. In V. Schmidt, editor, Stochastic Geometry, Spatial Statistics and Random
Fields: Models and Algorithms, volume 2120 of Lecture Notes in Mathematics, chapter 2, pages 31–71.
Springer, 2014.
[7] E. Bolthausen. On the central limit theorem for stationnary mixing random fields. The Annals of
Probability, 10:1047–1050, 1982.
[8] R.C. Bradley. Basic properties of strong mixing conditions. a survey and some open questions. Probability
Surveys, 2:107–144, 2005.
[9] A. Bulinski and E. Shabanovich. Asymptotic behaviour of some functionals of positively and negatively
dependent random fields. Fundam. Prikl. Mat., 4:479–492, 1998.
[10] A. Bulinski and A. Shashkin. Limit Theorems for Associated Random Fields and Related Systems.
Springer-Verlag, 2007.
[11] R. Burton and E. Waymire. Scaling limits for associated random measures. Ann. Appl. Probab., 13:1267–
1278, 1985.
[12] D.J. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes, Volume I: Elementary
Theory and Methods. Springer, New York, second edition, 2003.
[13] Yu. A. Davydov. The convergence of distributions which are generated by stationary random processes.
Teor. Verojatnost. i Primenen., 13:730–737, 1968.
[14] C. Dellacherie and P. Meyer. Probabilities and Potential. Elsevier Science Ltd, 1978.
[15] N. Deng, W. Zhou, and M. Haenggi. The ginibre point process as a model for wireless networks with
repulsion. IEEE Transactions on Wireless Communications, 1:479–492, 2015.
[16] P. Doukhan. Mixing: Properties and Examples. Springer-Verlag, 1994.
[17] P. Doukhan, K. Fokianos, and X. Li. On weak dependence conditions: the case of discrete valued
processes. Statist. Probab. Lett., 82(11):1941–1948, 2012.
[18] P. Doukhan and S. Louhichi. A new weak dependence condition and applications to moment inequalities.
Stochastic Process. Appl., 84:313–342, 1999.
[19] J.D. Esary, F. Proschan, and D.W. Walkup. Association of random variables, with applications. Ann.
Math. Statist., 38:1466–1474, 1967.
[20] S.N. Evans. Association and random measures. Probability Theory and Related Fields, 86:1–19, 1990.
[21] S. Ghosh. Determinantal processes and completeness of random exponentials: the critical case. Probability Theory and Related Fields, 163(3):643–665, 2015.
[22] J. S. Gomez, A. Vasseur, A. Vergne, P. Martins, L. Decreusefond, and W. Chen. A Case Study on
Regularity in Cellular Network Deployment. IEEE Wireless Communications Letters, 4(4):421–424,
2015.
[23] X. Guyon. Random Fields on a Network. Springer-Verlag, 1994.
[24] L. Heinrich. On the strong Brillinger-mixing property of α-determinantal point processes and some
applications. Appl. Math., 61(4):443–461, 2016.
[25] L. Heinrich and S. Klein. Central limit theorems for empirical product densities of stationary point
processes. Statistical Inference for Stochastic Processes. An International Journal Devoted to Time
Series Analysis and the Statistics of Continuous Time Processes and Dynamical Systems, 17(2):121–
138, 2014.
22
[26] J.B. Hough, M. Krishnapur, Y. Peres, and B. Virag. Zeros of Gaussian Analytic Functions and Determinantal Point Processes. American Mathematical Society, 2009.
[27] I.A. Ibragimov and Y.V. Linnik. Independant and stationnary sequences of random variables. WoltersNoordhoff, 1971.
[28] E. Jolivet. Central limit theorem and convergence of empirical processes for stationary point processes.
In Point processes and queuing problems (Colloqium, Keszthely, 1978), volume 24 of Colloq. Math. Soc.
János Bolyai, pages 117–161. North-Holland, Amsterdam-New York, 1981.
[29] A. Kulesza and B. Taskar. Determinantal point process models for machine learning. Foundations and
Trends in Machine Learning, 5:123–286, 2012.
[30] F. Lavancier, J. Møller, and E. Rubak. Determinantal point process models and statistical inference.
Journal of Royal Statistical Society: Series B (Statistical Methodology), 77:853–877, 2015.
[31] R. Lyons. Determinantal probability: Basic properties and conjectures. Proceedings of the International
Congress of Mathematicians, Seoul, Korea, IV:137–161, 2014.
[32] O. Macchi. The coincidence approach to stochastic point processes. Advances in Applied Probability,
7:83–122, 1975.
[33] N. Miyoshi and T. Shirai. A cellular network model with ginibre configured base stations. Advances in
Applied Probability, 46:832–845, 2014.
[34] J. Møller and R.P. Waagepetersen. Statistical Inference and Simulation for Spatial Point Processes.
Chapman and Hall/CRC, Boca Raton, 2004.
[35] E. Rio. Covariance inequalities for strongly mixing processes. Ann. Inst. H. Poincaré Probab. Statist.,
29(4):587–597, 1993.
[36] M. Rosenblatt. A central limit theorem and a strong mixing condition. Probc. Nat. Acad. Sci. U.S.A.,
42:43–47, 1956.
[37] T. Shirai and Y. Takahashi. Random point fields associated with certain fredholm determinants i:
fermion, poisson and boson point processes. Journal of Functional Analysis, 205:414–463, 2003.
[38] A. Soshnikov. Determinantal random point fields. Russian Math. Surveys, 55:923–975, 2000.
[39] R. Waagepetersen and Y. Guan. Two-step estimation for inhomogeneous spatial point processes. Journal
of the Royal Statistical Society, 71:685–702, 2009.
[40] M. Yuan, C. Su, and T. Hu. A central limit theorem for random fields of negatively associated processes.
Journal of Theoretical Probability, 16:309–323, 2003.
23
| 10math.ST
|
Toward the Starting Line:
A Systems Engineering Approach to Strong AI
Tansu Alpcan1 , Sarah M. Erfani2, and Christopher Leckie2
arXiv:1707.09095v2 [cs.AI] 18 Oct 2017
1
2
Dept. of Electrical and Electronic Engineering, The University of Melbourne, VIC, Australia.
[email protected]
School of Computing and Information Systems, The University of Melbourne, VIC, Australia.
Abstract. Artificial General Intelligence (AGI) or Strong AI aims to create machines with human-like or human-level intelligence, which is still a very ambitious goal when compared to the existing computing and AI systems. After many
hype cycles and lessons from AI history, it is clear that a big conceptual leap is
needed for crossing the starting line to kick-start mainstream AGI research. This
position paper aims to make a small conceptual contribution toward reaching that
starting line. After a broad analysis of the AGI problem from different perspectives, a system-theoretic and engineering-based research approach is introduced,
which builds upon the existing mainstream AI and systems foundations. Several
promising cross-fertilization opportunities between systems disciplines and AI
research are identified. Specific potential research directions are discussed.
1 Introduction
The terms Strong AI or Artificial General Intelligence (AGI) refer to the type of AI that
has certain capabilities similar to those of human intelligence. The AGI concept differs
significantly from the existing computing systems and mainstream AI algorithms also
called as “weak AI”. Weak AI systems are very successful in solving specific problems
when the problem context is provided directly by a human programmer. In contrast,
AGI is much deeper and broader in its ambition as it aims for human-like flexibility,
understanding, and creativity [26]. Even though it is not well-defined, AGI is one of
those things that “we will know it when we see it” [18].
The fact that we do not have machines with human- or animal-level intelligence today is indisputable. Current computers run manually programmed or trained algorithms
tailored to fulfill specific functions, whereas animals and humans automatically adapt
to and manipulate their environments. Moreover, the view that the Strong (or Hard)
AI problem has not yet been well defined [24] is shared by many experts. Maybe this
should not come as a surprise given that we don’t have a solid grasp on what human
intelligence means. The confusion on what AI means and can achieve has existed since
the beginning of the field in 1950s [12, 15, 30]. The history of AI consists of one hype
cycle following another, in which the pendulum has been swinging between wildly optimistic predictions [26] and deep pessimism [17].
The active research fields on weak AI such as machine and deep learning (DL),
pattern recognition, and data sciences are currently very healthy and growing after a
long AI winter. In contrast, after disappointments of many hype cycles, the dream of
creating AGI is currently sustained only by a small set of idealists [1, 12, 26]. We are
today arguably behind the starting line, which needs to be passed for the AGI research
to become mainstream.
Systems theory and engineering encompassing control, distributed systems, information, and communications theories as well as signal processing disciplines can potentially play a significant role in moving toward the starting line of AGI research. Many
AI research fields and systems theory share the same mathematical underpinnings and
aim to achieve similar goals in different ways. Therefore, a cross-fertilization, which
has been hindered mainly by academic inertia, may become very productive.
The goal of this speculative paper is to make a small conceptual contribution by
advocating a systems approach to AI. The next section analyses the AGI problem from
different perspectives and studies the questions of what do we want from AGI, how
successes of mainstream AI of today lay the foundations for tomorrow, what are the
different types of (artificial) intelligences, and why are we (still) behind the starting
line when it comes to AGI. Section 3 advocates a systems approach to AI incorporating
the biological, information and data, and distributed systems/networking dimensions.
The paper concludes with Section 4, which introduces a specific research vision for
AGI building upon existing work and discusses the potential role of systems theory and
engineering for AGI research.
2 AGI Problem Analysis: Behind the Starting Line
A proper analysis of the AGI problem would span multiple books due to its complex
multifaceted and multi-disciplined nature. This section aims to discuss only certain aspects of the problem as a precursor to the systems approach presented in Section 3.
2.1 What do we want from AGI?
The answer to the question of what we want from (Strong) AI has always been simple to
state: machines with human-like intelligence [26]. It is rare that such a simple statement
can be so ill-defined. Concepts such as creativity, flexibility, autonomy are often used to
clarify what is meant by human-like intelligence and the topic has been the subject of
many philosophical discussions. Two observations are worth making. Firstly, we really
don’t know much about how human (or animal) brain works, so we have a limited
scientific understanding of human-like intelligence [2]. While an engineered AI may
end up looking very different from biological brain, the underlying principles have to
be discovered and digested. The planes and birds fly differently, yet flying is flying
when it comes to aerodynamics. Secondly, even with our limited knowledge, it is easy
to argue that the current computing systems are as far from human-like intelligence as
it gets in terms of these aspects. Therefore, we can argue that we are behind the starting
line of building machines with human-like intelligence.
Another, maybe a more practical answer to the question is: more than what we have
now (in mainstream AI). Despite the recent great successes of mainstream AI research,
there are multiple core problems that are considered AI-complete or AI-hard [42]. We
can count fields of computer vision and natural language understanding among these.
When we ask for human-like conceptual understanding, this is beyond the capability of
current machines [17], even when their closely human-supervised algorithms perform
great in specific tasks [23, 45]. Another interesting AI-complete problem is networked
system security [4], which will remain an open issue as long as computing systems
lack autonomy and certain degree of human-like awareness since a tool can always be
misused by malicious parties who can access it.
Beyond what we want, there are two additional questions that have been hotly debated. The first one is whether we should want AGI. The AI field has never lacked optimists. Recently, there has been exaggerated claims (again) on ‘how thinking machines
will take over the world soon’ [26], which turned into a borderline scare campaign [9]
that influenced many non-experts in the broader society. The arguments in this paper
show that there is no need to panic when it comes to AGI progress.
The second question is whether (human-created) AGI is even feasible. The field
never lacked pessimists either. Maybe the most famous of them has been Hubert Dreyfus, who has contributed to the field of AI with well-thought (but perhaps destructive)
criticism [17]. At the risk of oversimplification, his main arguments center around the
fact that the underpinnings of human intelligence is non-symbolic, which follows the
insights of philosophers such as Heiddeger [16]. Unsurprisingly, his views have met
with hostility when symbolists were dominant in AI history, but have been received
more positively by connectionists (deep, neural, and statistical learning researchers).
On the flip side, the well-known earlier work by Minsky and Papert [31] “showed”
the perceived limits of connectionism and instead has given the symbolic approach an
upper hand in the field for a while. It turned out a decade or so later that their conclusions were quite incomplete and those perceived limits have been addressed through
subsequent discoveries in neural networks (NNs) [37]. Some say that the work even
contributed to the so-called AI winter. Throughout the AI history, the pendulum has
been swinging between exaggerated optimism and deep negativity. This can also be
interpreted as a sign of the field’s immaturity.
In summary, the feasibility question is wide open and will probably remain so until
we have a clear analytical or mathematical understanding of what we really want. For
example, are AGI goals achievable within the Turing Machine formalism? If not what
extensions are needed? Even if it is feasible theoretically, the computational resources
needed may be prohibitively high, especially considering that human brain has 100
billion neurons. Some of the recent theoretical work sidestep the latter issue by (perhaps
optimistically) assuming future availability of unlimited computational resources [24].
2.2 Successes of Mainstream AI
The current state-of-the-art AI research has been experiencing a boom within the last
decade. The impressive successes of mainstream AI arguably lay the foundations of
AGI but may not be sufficient as is. The following developments are noteworthy:
(1) From an algorithmic perspective, most of the early heuristic approaches have now
strong mathematical formulations, which provide solid foundations for future improvements and synthesis [8, 22, 28, 40, 43]. Several classical AI problems, e.g. playing
complex games such as chess, go, and poker have been effectively solved. Significant
progress has been made in various fields, e.g. image/speech recognition.
(2) The mainstream AI research also has a strong experimental culture that helps addressing real world problems. Moreover, the field has numerous practically-relevant
application domains which easily attract popular attention and research funding.
(3) There has been enormous advances in the hardware domain. The advances in computing power (central, graphical, and application processing units) as well as direct
access memory and storage capacities (disk drives) have been exponential.
(4) The exponential improvements in networking (in terms of bandwidth, convenience,
and costs) and maturing Internet have resulted in a data revolution, which has boosted
data-oriented approaches significantly.
2.3 Different types of (artificial) intelligences
A basic definition of intelligence is “the ability to acquire and apply knowledge and
skills” [27]. Based on this definition, humans have been building machines with limited
intelligence for a long time. An adaptive controller such as cruise control in cars or autopilot in planes can be considered “intelligent” in the simplest sense. Similarly, carefully designed rule-based methods, machine learning algorithms, and NNs have been
used to construct intelligent systems that acquire limited knowledge (e.g. parameters)
and achieve well-defined goals using models directly programmed by human.
Purely to further the discussion, let us consider the following arbitrary categories
of intelligence types: (i) controllers in control systems (ii) programmed automata and
Turing machines (iii) statistical and DL algorithms (iv) biological intelligence. The first
three types fit the general definition in varying degrees whereas natural biological systems provide the definition itself. This broad range somewhat explains the conflicting
opinions about AI; it is a bit like the ancient story “blind men and the elephant.” It is
easy for different people to have varying opinions on the topic based on their focus.
When it comes to AGI, what we want is clearly closer to the biological type of
intelligence. However, we still know very little about biological computing systems [2].
For example, we still don’t know the exact mechanisms of how brain stores information
(in short and long terms). We have a very limited understanding of brain’s plasticity or
organization. We don’t know how brain develops from a single cell in an embryo to
enormous complexity in a few years. A solid understanding of the principles behind
biological intelligence is needed to build systems with AGI.
2.4 Why are we (still) behind the AGI starting line?
The AGI problem is still not well-defined even after decades of investigation [24]. The
main issue is not the specific definition or lack of it per se, but it is that we don’t even
know how to make the transformative leap despite many past attempts [12, 30]. The
emergence and successes of weak AI, which is the mainstream AI for now, can be attributed to realizing this issue and scaling down the ambitions of AGI. The resulting
sensible approach has paid significant dividends and great progress has been made in
numerous fields ranging from image recognition to automated translation. The undeniable successes of weak AI in areas which were once considered the domain of AGI
such as the Turing test [29], playing full information games, e.g. chess [23], and go
(DeepMind) as well as limited information ones, e.g. poker [45], certainly support the
collegial joke that “when an AI problem is solved it is not (of interest to) AGI anymore.”
Can the incremental mainstream AI research erode the mountain of AGI? The answer is negative as long as we don’t have a good starting point. The need for the
paradigm shift is due to the fact that AGI we want is very different from the existing computing systems, including weak AI algorithms. Once we reach the starting line,
however, the incremental progress will start. The argument that we are (well) behind
the starting line is supported by the following observations:
(1) We don’t have a proper scientific understanding of AGI objectives beyond vague
allusions. There are no clear analytical or mathematical descriptions of what is meant
by AGI properties such as creativity, flexibility, and autonomy.
(2) The AGI research, as of first two decades of the 21st century, is not part of mainstream AI research. The few conferences and journals on the topic are considered niche
and more philosophical than business-as-usual.
(3) Most of the mainstream AI researchers are very realistic about their results. It is not
claimed that existing recognition or game playing systems have a human-like understanding or flexibility of applying the know-how gained to other problems.
(4) The theoretical and practical knowledge basis we have falls well short for the usual
incremental research progress when it comes to AGI goals. The gap is simply too big
between data-oriented (statistical or deep) learning algorithms or symbolic formalisms
and the AGI goals. Recent great theoretical accomplishments in fields such as algorithmic information theory or decision theory, despite providing pieces of the puzzle, also
fall short in paving an obvious way toward human-like AI systems.
These observations and arguments become clearer when we further analyze next
what we really want from AGI. Without such an analysis, the collegial joke above becomes quite serious!
3 A Systems Approach to AI
3.1 Biological Dimension
Since AGI aims to imitate properties of biological systems, it is appropriate to start
the discussion with the biological dimension. There are several aspects of biological
systems that are arguably directly relevant to AGI system development [3]. Biological
intelligence is always embodied in the physical world. AGI properties such as adaptability (flexibility) and autonomy can only be defined within the context of an environment.
This important point has been recognized by AI researchers, see e.g. embodied cognition, Nouvelle AI, etc. [11]. Even the strip of tape in the Turing Machine abstraction
can be interpreted as an “environment.” One of the main points in Dreyfus’ arguments
against symbolism, which is disembodied, is based on Heidegger’s philosophy where
intelligence is naturally integrated into the environment [16].
Biology strongly suggest that it is necessary to consider an AGI system as part of the
environment it inhabits and interacts with. The conceptual framework used by control
system theory [6] to distinguish between a controller (AI), actuator, dynamical system
(environment), and observer, provides a ready-made formalism as shown in Fig. 1. The
Fig. 1. A control-theoretic formalism for AGI systems.
embodiment of AI can be but does not have to be physical (e.g. robotics). It is conceivable and sometimes desirable to develop AI systems embedded in virtual environments
(e.g. gaming worlds [7], 3D Internet [5]) or other data streams.
Introducing a Universal Continuity Conjecture (UCC), it is possible to consider
inanimate objects, living beings, and intelligence (of all types, including the vague concept of consciousness) within the same continuous spectrum imposed by the underlying
physical laws. In this philosophical view, living organisms as well as different types of
intelligence are natural consequences of underlying physical laws. The fact that we cannot easily draw the boundaries of life and intelligence supports this conjecture. Another
strong argument in favor of the UCC is the fact that intelligence builds upon life (cellular organisms) and life builds upon inanimate materials (biochemistry) [3] in a seamless
way. Despite its purely philosophical nature, UCC can be useful in developing the underlying engineering principles for AGI.
AGI as an embodied system within an environment can be equivalently considered
as an organism with homeostasis (self-regulating) property, consistent with the UCC.
Hence, the system approach goes beyond the computational one and aims to engineer
a type of self-regulated embodied organism or complex system that interacts with the
environment. In this sense, the approach builds upon dual (or adaptive) control [47] as
well as reinforcement learning [38].
An illustration of the UCC can be found in the famous Conway’s Game of Life [13],
which played a significant role in cellular automata research [21]. Conway’s Game of
Life is a cellular automaton that showcases how abstract objects following simple rules
can evolve complex (and stable) structures. While the research on digital organisms and
artificial life are closely related to AGI, the engineering approach adopted here differs
from those efforts. The focus is on conscious design and engineering of self-regulating
systems instead of waiting for them to emerge through purely evolutionary processes.
Is it possible to achieve the level of flexibility desired in AGI through purely humandriven engineering design? Is there a middle ground between classical manual design and hands-off emergence of structures through pure evolutionary processes (as
in game of life)? The answer to these questions may lie in the under-explored field of
L-systems [36]. L-systems are rule-based formal grammars, and hence, are closely related to automata theory, fractals, and iterative and complex systems. They can arguably
be used as a starting point to investigate the concepts discussed.
3.2 Information and Data Dimension
An intelligent system embodied in an environment has to process the input data from its
own sensors, extract the relevant information or knowledge, and store an encoded version of this knowledge for future usage as shown in Fig. 2. An adaptive autonomous system has to incorporate learning into all these functions. Similar problems have been investigated by reinforcement learning as well as adaptive and dual control research [48].
Multiple fundamental system theory fields can potentially play a very useful role
in this direction. Signal processing studies the theory and methodologies of processing
the information from a broad range of signals. Since the AI input data streams (e.g.
provided by specific system’s sensors which can be physical or virtual) are signals,
the approaches developed in the signal processing area directly apply. Similarly, data
science has recently emerged and attracted a lot of interest as an interdisciplinary field
focusing on extracting knowledge from data. Encoding and representation of data has
been studied extensively by coding and information theories [14], which provide an
extensive set of mathematical tools. While these theories have traditionally focused
on communication systems, the fundamental principles can be reused within the model
considered. In conjunction with system-theoretic methods, the database theory provides
a strong foundation for storage of extracted information.
The way information or
knowledge is represented and
processed has caused a lot of
tension in AI research since
the beginning [35]. Connectionist approaches (in the form
of NNs and variants) tend to
encode information using parameters of NNs, which are
trained using input data, while
symbolism adopts a high-level
human-readable approach based
on logic (e.g. rule-based or exFig. 2. Information processing in an AGI system embodied pert systems). Both approaches
obviously have useful properties
in an environment.
and provide important insights
in their own way.
Design and engineering of AGI systems require building a strong connection between symbolism and connectionism, while a purely symbolic approach most probably
will not provide the autonomy and flexibility we want from AGI. This is due to the
fact that such solutions rely on human-provided explicit contextual information through
modeling and programming. However, pure connectionism tends to encode knowledge
in a form not accessible or interpretable by humans. The black-box solutions of NNs
have been criticized for this since their inception. Translation of knowledge to humanreadable formats is necessary at least to facilitate a human-in-the-loop engineering design process. Biological systems strongly suggest that such a translation is possible.
Human brain is a biological neural network yet our abstract thinking is symbolic.
3.3 Distributed System and Networking Dimension
Connectionist approaches to AI are closely related to network science and distributed
systems. NNs and more recent variants under the umbrella of DL are networked systems. Hence, the distributed system and networking dimension has been a part of AI
research since the early days.
The idea that complex solutions emerge from simple interacting components has
been fundamental to artificial life, NNs, and more recently DL. This idea is supported
by observations of biological systems which are almost always cellular. Graph theory
provides strong theoretical foundations to networks. As a related topic, the emergence
of global properties in random graphs have been studied by the percolation theory.
Since the beginning NNs have been
inspired by their biological counterparts.
It turns out that the artificial versions
we have are overly simplistic. Biological variants are multi-cellular, where
each cell is a (complex) subsystem itself
[3]. Hence, the natural next step in connectionism is toward distributed systems
consisting of complex subsystems, similar to multi-core processors, rather than
simple functional components (Fig. 3).
Recent advances such as Long Short
Term Memory (LSTM) [20] can be interFig. 3. A modular, distributed AGI system archipreted as developments in this direction.
tecture inspired by cellular organisms.
System theory provides a strong foundation for analysis and design of distributed
systems building upon know-how from communication networks, distributed optimization, game theory, and control theory.
4 Discussion and Future Research Directions
At the risk of oversimplification, the current mainstream AI research can be crudely
clustered into three main threads that naturally overlap and interleave: (1) symbolic
approaches (good old AI), which rely on formal logic and automata theories [38], (2)
connectionist approaches such as NNs and DL [39], (3) statistical pattern recognition
and machine learning [8]. It is natural to expect that these threads will provide some of
the building blocks of future AGI systems.
The AI field is again climbing a new hype cycle after a long (in Internet time) AI
winter. This cycle is commercially supported by the Internet-age giants (companies)
such as Google, Facebook, Amazon, and Huawei. The successes of pattern recognition
and machine learning in solving specific problems as well as exciting recent developments in the DL field exploring new memory and computational architectures [19, 32,
34, 46] provide a fertile ground for making progress in AGI toward the starting line. In
addition to these data-oriented learning approaches, there has been significant progress
in symbolic AI, especially in formal verification methods. The prevalence of standardized data sets for specific problems and the emergence of frameworks for embodiment
of agents in physical (robotics) and virtual environments (gaming) provide the indispensable experimental setups and verification.
The mathematical understanding supporting mainstream AI has been maturing. Automata theory, formal languages and formal verification methods are directly connected
to well established discrete mathematics and formal logic foundations. Most of the existing pattern recognition methods including kernel-based approaches [40] have been
formalized using optimization theory. Likewise, NNs and DL [39] approaches rely on
variants of well-known (global and stochastic) optimization [10, 25] algorithms.
Certain system-theory methods have also been used in mainstream AI research,
albeit in a more limited sense. In DL, Recurrent NNs [44] and LSTM [20] have been
formalized using nonlinear dynamical system models. Entropy-based learning methods
have been investigated in the learning area [41]. Game theory has been used to model
attacker vs defender interaction in adversarial learning setups [4]. The link between
adaptive control and reinforcement learning has been recently mentioned [24]. Morover,
NNs can be used for identification and control of dynamical systems [33].
System theory encompassing signal processing, communications (information theory) and (distributed) control theories can potentially play a significant role in AGI research. Most of the mathematical foundations are shared by the AI and system research
fields, such as optimization, graph, game, dynamical system theories as well as discrete
event and hybrid systems, and automata theory. The model-based emphasis of system
theory research can sometimes be a restrictive factor yet it can bring the much-needed
contextual leap to purely data-oriented (black-box) approaches of machine and DL. Furthermore, distributed system and game theory can provide a mathematical framework
for analysis and development of novel AI systems by rigorously studying individual
system components, their environment, and the interactions.
Specific open research questions and directions (in no particular order) include:
(1) Combining connectionism and symbolism is necessary to achieve practically useful
hybrid and multi-layer system architectures for AGI. This is closely related to how we
organize (encode and interpret) the data and learned functions inside the system.
(2) Turing machines that can do dual control or reinforcement learning (in an environment) can be developed by combining the lessons from control/system theory, building
upon recent exciting results in DL such as Neural Turing Machines [19] and Neural
Programmer [34]. Better use of fields such as optimization, information, control, and
game theories within the context of AI research may result in faster progress overcoming the historical unnecessary disconnect between these research communities.
(3) Better understanding of biological systems and intelligence is potentially the key
to make the leap towards AGI. Cellular architectures inherent to living intelligent organisms can be modeled using distributed system theory. Specific research subtopics
include: (a) Analysis and design of “developing systems”, which evolve from scratch
only based on the inherent rules and the environment they are embodied, may give
clues for the desired adaptability and flexibility. (b) Creation of engineering architectures based on biological principles of homeostasis and cellular organisms would be
useful for design of distributed intelligent systems.
(4) Addressing the issue of combining explicit human intervention (classical programming) with learning systems would help shorten the evolutionary learning time-scales
and achieve backwards compatibility with existing computers.
Acknowledgements
The authors thank anonymous reviewers and TPC chairs of the AGI 2017 conference
for their valuable feedback and suggestions.
References
1. Conference Series on Artificial General Intelligence (AGI) (2008-2017)
2. The NIH brain initiative. Science 340(6133), 687–688 (2013)
3. Alberts, B., Johnson, A., Lewis, J., Raff, M., Roberts, K., Walter, P.: Molecular biology of
the cell. Garland Science, 6th edn. (2014)
4. Alpcan, T., Başar, T.: Network security: A decision and game-theoretic approach
5. Alpcan, T., Bauckhage, C., Kotsovinos, E.: Towards 3d internet: Why, what, and how? In:
International Conference on Cyberworlds. pp. 95–99 (2007)
6. Anderson, B.D.O., Moore, J.B.: Linear Optimal Control. Prentice-Hall, Inc. (1971)
7. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The Arcade Learning Environment:
An Evaluation Platform for General Agents. J. Artif. Int. Res. 47(1), 253–279 (2013)
8. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer-Verlag (2006)
9. Bostrom, N.: Superintelligence: Paths, Dangers, Strategies. Oxford University Press (2014)
10. Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press (2004)
11. Brooks, R.A.: Intelligence Without Representation. Artificial Intelligence 47, 139–159
(1991)
12. Buchanan, B.G.: A brief history of artificial intelligence. AI Magazine 26(4), 53 (2005)
13. Conway, J.: The game of life. Scientific American 223(4), 4 (1970)
14. Cover, T.M., Thomas, J.A.: Elements of Information Theory. Wiley-Interscience (1991)
15. Crevier, D.: AI: The Tumultuous History of the Search for Artificial Intelligence. Basic
Books, Inc. (1993)
16. Dreyfus, H.: Being-in-the-world: A Commentary on Heidegger’s Being and Time, División
I. MIT Press (1991)
17. Dreyfus, H.L.: What Computers Still Can’t Do: A Critique of Artificial Reason. MIT Press
(1992)
18. Gewirtz, P.: On ”I Know It When I See It” . Yale Law Journal 105, 1023–1047 (1996)
19. Graves, A., Wayne, G., Danihelka, I.: Neural Turing Machines. arXiv:1410.5401 (2014)
20. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735–
1780 (1997)
21. Hopcroft, J.E.: Introduction to Automata Theory, Languages, and Computation. Pearson Addison Wesley, 3rd edn. (2007)
22. Hornik, K., Stinchcombe, M., White, H.: Multilayer Feedforward Networks Are Universal
Approximators. Neural Networks 2(5), 359–366 (1989)
23. Hsu, F.H.: Behind Deep Blue: Building the Computer That Defeated the World Chess Champion. Princeton University Press (2002)
24. Hutter, M.: Universal Artificial Intelligence: Sequential Decisions based on Algorithmic
Probability. Springer (2005)
25. Jones, D.R.: A Taxonomy of Global Optimization Methods Based on Response Surfaces.
Journal of Global Optimization 21(4), 345–383 (2001)
26. Kurzweil, R.: The Singularity Is Near: When Humans Transcend Biology. Penguin (2006)
27. Legg, S.: Machine super intelligence. Ph.D. thesis (2008)
28. MacKay, D.J.C.: Introduction to Gaussian Processes. In: Neural Networks and Machine
Learning, pp. 133–166 (1998)
29. Mauldin, M.L.: Chatterbots, tinymuds, and the turing test: Entering the loebner prize competition. In: AAAI Conference on Artificial Intelligence. vol. 94, pp. 16–21 (1994)
30. McCorduck, P.: Machines Who Think: A Personal Inquiry into the History and Prospects of
Artificial Intelligence. A K Peters/CRC Press (2004)
31. Minsky, M., Papert, S.: Perceptrons: An Introduction to Computational Geometry. MIT Press
(1972)
32. Mnih Volodymyr, Kavukcuoglu Koray, Silver David, Rusu Andrei A., Veness Joel, Bellemare Marc G., Graves Alex, Riedmiller Martin, Fidjeland Andreas K., Ostrovski Georg, Petersen Stig, Beattie Charles, Sadik Amir, Antonoglou Ioannis, King Helen, Kumaran Dharshan, Wierstra Daan, Legg Shane, Hassabis Demis: Human-level control through deep reinforcement learning. Nature 518(7540), 529–533 (2015)
33. Narendra, K.S., Parthasarathy, K.: Identification and Control of Dynamical Systems Using
Neural Networks. Trans. Neur. Netw. 1(1), 4–27 (1990)
34. Neelakantan, A., Le, Q.V., Sutskever, I.: Neural Programmer: Inducing Latent Programs with
Gradient Descent (2015)
35. Papert, S.: One AI or Many? Daedalus 117(1), 1–14 (1988)
36. Rozenberg, G., Salomaa, A.: The mathematical theory of L systems, vol. 90. Academic press
(1980)
37. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error
propagation. Tech. rep., California Univ San Diego Inst for Cognitive Science (1985)
38. Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach. Prentice Hall Press
(2009)
39. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural networks 61, 85–
117 (2015)
40. Scholkopf, B., Smola, A.J.: Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press (2001)
41. Schraudolph, N.N.: Optimization of entropy with neural networks. Ph.D. thesis, University
of California, San Diego (1995)
42. Shapiro, S.C.: Encyclopedia of Artificial Intelligence. John Wiley & Sons, Inc. (1992)
43. Siegelmann, H.T., Horne, B.G., Giles, C.L.: Computational capabilities of recurrent NARX
neural networks. IEEE Trans. on Sys., Man, and Cyber., Part B 27(2), 208–215 (1997)
44. Sutskever, I.: Training recurrent neural networks. Ph.D. thesis, University of Toronto (2013)
45. Tammelin, O., Burch, N., Johanson, M., Bowling, M.: Solving Heads-up Limit Texas
Hold’Em. In: International Conference on Artificial Intelligence. pp. 645–652 (2015)
46. Weston, J., Chopra, S., Bordes, A.: Memory Networks. arXiv:1410.3916 (2014)
47. Wittenmark, B.: Adaptive Dual Control Methods: An Overview. In: Proc. of 5th IFAC symposium on Adaptive Systems in Control and Signal Processing. pp. 67–72 (1995)
48. Wittenmark, B.: Adaptive dual control. In: Control Systems, Robotics and Automation, Encyclopedia of Life Support Systems, Developed under the auspices of the UNESCO (2002)
| 2cs.AI
|
Topology Control of wireless sensor network using Quantum Inspired Genetic
algorithm
Sajid ullah
Institute of space technology, Pakistan
[email protected]
Abstract—In this work, an evolving Linked
Quantum register has been introduced, which are
group vector of binary pair of genes, which in its
local proximity represent those nodes that will
have high connectivity and keep the energy
consumption at low, and which are taken into
account for topology control. The register works
in higher dimension. Here order-2 Quantum
inspired genetic algorithm has been used and also
higher order can be used to achieve greater
versatility in topology control of nodes. Numerical
result has been obtained, analysis is done as how
the result has previously been obtained with
Quantum genetic algorithm and results are
compared too. For future work, factor is hinted
which would exploit the algorithm to work in
more computational intensive problem.
Key Words— relative quantum order, topology control, Linked
quantum register
I. INTRODUCTION
The two algorithms, evolutionary and quantum
computation has made a remarkable success by
solving some problem with relatively less
computational time and memory. From it a new
mataheuristics field has been made which combined
the power of two. The work, Quantum inspired
genetic algorithm (QIGA) was put forward by
Narayan in 1996 [1].Although evolutionary
techniques combined with quantum computing has
been for years but the first work on practical problem
to work out was used for combinatory optimization
[2]. Quantum inspired genetic algorithm used
probability based optimization which include
algorithms from evolutionary and quantum
computing with the likes of parallel computing,
superposition and genetic variation [1].QIGA does
not
require
quantum
computer
for
its
implementation, but there are metrics when reached
certain level, then can quantum computer be
required. Those computing techniques combine with
probabilities achieve such great result which
converge for global optimum. QIGA has been used
Mussarat Wahid
University of Nottingham, England
[email protected]
for combinatory optimization problem, and solved
several computational intensive problem. QIGA has
been used in Power system optimization and
localization of mobile robots [3], Image processing
[4], flow shop Scheduling [5], Optimization of Hot
Extrusion Process [6].Several algorithm are used for
topology control in wireless network (WSN).
LEACH [12], SPAN [7], ASCENT [8], and STEM
[9]. The first use of evolutionary and Quantum
computing used for topology control was in the work
[10], which show better result than simple genetic
algorithm. The topology control was achieved in
fewer generation in Quantum genetic algorithm
(QGA).QGA proves superior to simple genetic
algorithm. Higher order QIGA was presented in
work [11].Order-2 QIGA does need a functional
quantum computer for its implementation as it uses
random features from quantum mechanical system
such as qubits, superposition, and parallel
computing. Quantum factor defines when algorithm
have to be implemented on quantum computer, and
is elaborated in [11]
The optimal point of topology control are
standardized on connectivity of nodes, number of
node exceeding the threshold, energy of nodes, and
interference between the nodes. The interference is
one of key issue in wireless sensor network (WSN).
An optimal point is selected which minimize the
interference in register and the intra node between the
registers. Interference has not been account in
topology control here.
Higher order QIGA have terminology which does
not exist in QGA. Quantum and relative quantum
order, and quantum factor are defined here and for
detailed work, reader should understand the full
algorithm in [11].
This paper is structured as: The section 1 will
describe the various design metrics that influence the
topology of nodes. Section 2 discussed the order -2
QIGA, various definition of the algorithm are
introduced. In section 3 the linked quantum register
has been introduced. In section 4 the algorithm has
been implemented for topology control. In Last
section, Numerical result has been obtained for
topological control of nodes.
II. TOPOLOGICAL INITIALIZATION,
CONSTRUCTION AND MAINTENANCE
The sensor network is initialize with n number of
sensor node at t=0, all the node have equal resources.
So at the start of the process the nodes have identical
probability amplitudes which implies from the above
statement, initially the nodes are randomly selected
in topological space in which they do sensing,
surveillance or other data acquisition. The second
necessary step in topological control is its
construction for which we present a novel higher
dimensional register LQR, which self-varies itself
across search space and Migration status is updated
on each iteration. The Topology maintenance is what
the algorithm does in more optimal way than QIGA,
On each successive iteration the topological point in
space between nodes are varied as to keep the nodes
net energy at minimal level, maintain the
connectivity of node strong enough so that don’t
interfere in one another wireless information signal.
In wireless communication control decision are most
energy consuming, which are kept to minimum
number.
Suppose there are n number of nodes in a network,
1≤ i ≤ n, which in coordinate space are located at (xᵢ,
yᵢ) at the radius of rᵢ. The power consume by each for
any node is Pᵢ.
Rmax –the maximum a node can transmit
Aᵢ - show the connectivity of network, considering it,
the network has to ensure it don’t interfere in
adjacent node’ transmission region.
Rf – the nodes that exceeds the threshold point Rt are
bk in number.
The topological factors mathematically are
∀Cij in C (n), n defines the dimensionality of node
The element of adjacency matrix is one if they are
connected, which should be in cluster for node
connected in a register. For topology control WSN
attain these points.
The total energy of the node should be keep
at minimum
The number of node which exceeds the
threshold are to be sparse.
The interference between the elements of
adjacent register in topological space should
be lowered.
III. ORDER-2 QUANTUM INSPIRED GENETIC
ALGORITHM
The QIGA uses individual independent qubits to
represent binary genes, so they are effectively of
order 1. Higher order QIGA was presented in [11].
We here present introductory definition from their
work. For full understanding of their work reader
should study [11]. QIGA-2 symbolizes classical and
quantum individual population by P and Q. order r is
defined as the largest register used in algorithms.
1≤ r ≤N.
Quantum order r is the size of biggest
quantum register used in the algorithm.
Relative quantum order w shows the ratio of
quantum order to problem size, N. For QIGA2, the 100 individual qubits would have
2
Relative Quantum range of 100 = 0.02, This
is representation based on 2-qubit registers
Quantum factor (λ ϵ [0 1]) which define the
ratio of the ratio of the dimension of space of
current Algorithm to dimension of space of
full quantum register of all individuals.
λ =
2r
w2N
In our topological control problem, there can
scenario in a high dimension space in which there
will be need to deploy different registers size for
nodes in local proximity of topological space. In
QIGA-2, register have a pair of nodes.
(a) Chromosome in order-1 QIGA
(b) Chromosome in Order-2 QIGA
Fig. 1. (a): In order one each individual chromosome
is spanned in two dimensional space. (b): In Order-2
the register has two chromosome, and dimension of
space for the order-2 QIGA is 4
Protocol used are application specific in a
register. The control decision information
in the elements of a register are specific to
the individuals only
LQR have memory because in each update
step ,which is the distance of nodes in each
Register, is to be stored.
a. Algorithm for Topology Control
Fig. 2. The 1st depiction show two node in unlocked
manner while the second show nodes in a Quantum
locked register.
IV. LINKED QUANTUM REGISTER (LQR)
In this work we propose a model which takes the
adjacent 2-qubit quantum register into account to
form a binary paired register. The pair of binary
genes in a register represent a node, while in
register have a pair of nodes. The information
between node in a register is transferred which
achieve the topology optimal point in space. As a
result the LQR has four points in probability in
distribution.|α₀|², |α₁|², |α₂|², and |α₃|².
These are few key points about LQR.
There can be arbitrary number of Quantum
binary genes in register .It can varies
across different register depending upon
the computational complexity a classical
machine can cope with. The network can
have different number of chromosome in
register, and still can achieved the work of
having same number of chromosome
(nodes) with turning off those nodes for the
iterative procedure, but survival of the
fittest results from evolutionary computing
will keep the nodes ahead in the run.
The QIGA-2 starts by searching the adjacent nodes
for which the register identify (RI) is identical. After
selecting the nodes of same RI. The algorithm checks
for the node connectivity in adjacency matrix which
is update by algorithm presented in [11]. The
connectivity of nodes is range of space between the
nodes or discrete set of point in space. A prior
optimal point is set for WSN which is application
dependent. A loop is run over the node and the
distance is incremented or decremented between the
node is achieved which assign node to the same RI.
The algorithms then compute the distance between
the nodes and set it to be connective. The position of
each node is varied by choosing the number of step
in each iterative process, which results in step size of
the increment or decrement in topology control. The
algorithm converges when the given node in WSN
are in connectivity range set prior in the algorithm.
The update of 2x2 dimensional of LQR depends by
choosing one or both of node. the algorithm has
select method that choose one of the node in QLR.
Algorithm 1: Topology Control in QIGA2
1: for i=0,1,2,...,N
2: for j=0,1,2,...,N-1/2
3: distconn(i,i+1)=1
4: select( i,i+1) for j=j+1
5: dist(i,i+1)
6: if dist(i,i+1)=distconn(i,i+1)
7: then Mat[i,i+1]=1
8: else if dist(i,i+1)≤distconn (i.i+1)=1
9: while dist(I,i+1)≠ distconn(i,i+1)
10: do dist(i,i+1)←dist(i,i+1)+Δi
11: else if dist(i,i+1)≥distconn(i,i+1)
12: while dist(I,i+1)≠distconn(i,i+1)
13: do dist(i,i+1)←dist(i,i+1)-Δi
14: end for
15: end for
V. ALGORITHM IMPLEMENTATION FOR
TOPOLOGY CONTROL
Several algorithm has been implemented for
topology control [7], [8], [9], [10], and [12].
The algorithm starts with an arbitrary number of
wireless sensor network (1 ≤ i ≤ n). The initial
population will start with equal resources in term of
energy, processing Capability and Transceiver
power.
In order-2 QIGA the population are represented by
vector of chromosome (register).All amplitude are
𝟏
initialized with magnitude of √𝟐
Q (t) = {q1t, q2t... qnt}
Solution of individual Q-bits in register:
In this step, binary string are produced from qubits
string, this is called observation of states. Q (t) with
n bits represent a superposition of 2n binary states of
gene. The states of individual in register is observed
For any Register, Rid
P (t) = {p1t, p2t ,…, pnt}
Where Ptj represent observational result of jth
individual, Rid will have N/r number of P (t) in each
iterative step
In evaluation step, LQR have the individual nodes,
and they are evaluated on basis of how strongly they
are connected. The individual use minimum energy
and minimum number of nodes that exceeds the
power transmission threshold radius, Rf.. The Binary
solution is stored.
B (t) = { b1t,b2t,…,bnt}
Then the observation of state of Q (t-1) produce
binary solution in P (t).
The variation operator is model specific, and the user
can pick any of the operator that suits the algorithm
the best.
In order-2, genetic operator are taken 4x4 quantum
gates. The observation function from step 2 of
algorithm returns strings of binary genes 00, 01, 10
and 11 with a probability of |α0|,| α1|,| α2 |, and |α3|
respectively.
The distance update can be done in unidirectional
or bi-directional. In unidirectional the position of one
of the node is varied keeping the other in its current
status, while in bi-directional the position of both
node is updated according to the algorithm 1.
VI. NUMERICAL RESULTS
The algorithm has been implemented for 16 nodes.
The WSN took 79 generations to accomplish the
condition specified in algorithm to converge. Roulette
wheel was used for evaluation function of fitness. The
QIGA-2 has been compared with Simple Quantum
Genetic (SGA) algorithm. Solution found in QIGA-2
earlier, which was obtained because there were two
pair of LQR which were in transmission range of one
another.so as a result it takes fewer generation for
finding optimal topology control.
Table 1. Result of algorithm: QGA and QIGA-2
Algorithms Earliest
Solution
Total
generation found at power
finding the the
the radius
solution
generation threshold
QGA
90
47
156
QIGA-2
79
42
140
Rf
7
6
LQR modeling in QIGA-2 results in less generation
for topology control than QGA. The solution
obtained earlier because the two pair of LQR which
have transmission region exceeded than standardized
earlier. The power consumption due to fewer
generation was beneficial in QIGA-2, and results in
less net power than QGA.
Initializatio popoution
Q(0)
Generate P(t) by
searching Q(t)
Evaluation of P(t)
record the optimum
solution for the current
iteration step
Applying Quantum
variation operation
t←t+1
Goal attained
Fig 3. Updating Quantum gene state space
VII. CONCLUSION
In this work, a novel register is introduced based
on QIGA-2 which uses evolutionary computing in
addition to quantum computing to find the optimal
solution for topology control. The attributes inherited
from quantum computing in LQR are the way of
representation of solution of nodes. The order 2
register uses the nodes in register to cooperate for the
information necessary to increase the connectivity to
an optimal level. At the start of algorithm nodes are
paired to make LQR, which transfer the information
in each iteration of algorithm for the topology control.
Future perspective of the QIGA lies in using more
than two number of node in LQR which seem to
obtaine better result. Modeling of LQR for more than
two node and its implementation on Quantum
computer would be challenging.
(10, 10, 10)
(2, 7, 6)
(7, 8, 6)
(4, 8, 10)
(8, 7, 0)
(6, 7, 0)
(3, 5, 4)
(7, 5, 3)
(4, 4, 2)
(1, 5, 5)
(8, 2, 1)
(2, 2, 2)
(6, 2, 1)
(4, 2, 2)
(1, 1, 0)
(9, 1, 1)
Fig 4. Topology attained by QIGA-2
ACKNOWLEDGMENT
We would like to thank our colleagues from institute of
space technology and university of Nottingham for their
valuable discussion about the problem and insights that
greatly assists in our research work.
REFERENCES
[1] A.Narayanan M. Moore, Quantum-inspired genetic
algorithms, Proc. IEEE Evolutionary computation,
61-66 (1996)
[2] K.H Han and J.H. Kim, Genetic quantum algorithm
and its application to combinatorial optimization
problem, Proc. Congress on Evolutionary
Computation, 1354-1360(2000)
[3] S.Je zewski, M. Laski, and R. Nowotniak,
Comparison of Algorithms for simulation
Localization and Mapping Problem for Mobile Robot,
Scientific Bulletin of Academy of Science and
Technology, Automatics 14,439-452 (2010)
[4] H. Talbi, M. Batouche, and A. Draa, A QuantumInspired genetic algorithm for multi-source affine
image registration, Image Analysis and Recognition ,
Springer, 147-154 (2004)
[5] L. Wang, H. Wu,F. Tang, and D.Z. Zheng, A hybrid
quantum-inspired genetic algorithm for flow shop
scheduling, Advances in Intelligent Computing,
Springer, 636-644 (2005)
[6] Rajat Setia, K. Hans Raj, Quantum Inspired
Evolutionary Algorithm for Optimization of Hot
Extrusion Process, International journal of Soft
Computing and Engineering (IJSCE), Volume 2,issue-5, November 2012
[7] Chen, Jamieson, Balkrishnan, and Morris, Span: an
energy efficient coordination algorithms for topology
maintenance in ad hoc wireless network, proc. Of
Mobicom
[8] A. Cerpa, D. Estrin, ASCENT: adaptive selfconfiguring sensor network topologies Transaction
on Mobile Computing, IEEE 3(3), 272-285.
[9] Scrugers et al. STEM : Topology management for
energy efficient sensor network
[10]
SUN Lijuan, GUO Jian, LU Kai, WANG
Ruchuan, Topology Control based on quantum
genetic algorithm in sensor network
[11]
Robert Nowotniak, Jacek Kucharski,” HigherOrder Quantum-Inspired Genetic Algorithms”,
Computer Science and Information Systems
(FedCSIS), 2014 Federated Conference on , vol., no.,
pp.465,470
[12]
Wendi R. Heninzelman, A Chanderakasan, and H
Balakrishan, “ Energy-Efficient Communication
Protocol for Wireless Microsensor Network”,
Proceedings of the 33rd Hawaii International
Conference on System Sciences-2000
| 9cs.NE
|
arXiv:1803.02943v1 [cs.NE] 8 Mar 2018
Multi-objective evolution for 3D RTS Micro
Sushil J. Louis
Siming Liu
Evolutionary Computing Sytems Lab
Dept. of Computer Science and Engineering
University of Nevada
Reno, NV 89557
http://www.cse.unr.edu/∼sushil
Evolutionary Computing Sytems Lab
Dept. of Computer Science and Engineering
University of Nevada
Reno, NV, 89557
http://www.cse.unr.edu/∼simingl
Abstract—We attack the problem of controlling teams of
autonomous units during skirmishes in real-time strategy games.
Earlier work had shown promise in evolving control algorithm
parameters that lead to high performance team behaviors similar
to those favored by good human players in real-time strategy
games like Starcraft. This algorithm specifically encoded parameterized kiting and fleeing behaviors and the genetic algorithm
evolved these parameter values. In this paper we investigate
using influence maps and potential fields alone to compactly
represent and control real-time team behavior for entities that can
maneuver in three dimensions. A two-objective fitness function
that maximizes damage done and minimizes damage taken guides
our multi-objective evolutionary algorithm. Preliminary results
indicate that evolving friend and enemy unit potential field
parameters for distance, weapon characteristics, and entity health
suffice to produce complex, high performing, three-dimensional,
team tactics.
I. I NTRODUCTION
Real-Time Strategy (RTS) games model decision making
in wartime. The game genre has become a popular research
platform for the study of Computational and Artificial Intelligence (CI and AI). In RTS games, players need to establish
bases, collect resources, and train military units with the aim
of eliminating their opponents. To be good, RTS game players
require a variety of decision making skills [1], [2]. First, the
dynamic environment of RTS games requires real-time planning on several levels - strategic, tactical, and reactive. Second,
players have to make decisions with imperfect information
within the “fog of war.” Third, like poker, there is no one
optimal strategy; players must model their opponent and adapt
their strategies and tactics to the opponent’s playing “style.”
Fourth, players must employ spatial and temporal reasoning to
exploit existing terrain and the time-sensitive nature of actions
on a tactical and strategic level. All of the challenges described
above and their impact on decision making are crucial for
winning RTS games and there is a steep learning curve for
novices. In the video game industry, Starcraft (SC) and Starcraft 2 (SC2) players earn significant prize money and points
during a series of championship games worldwide culminating
in a World Series Championship at Blizzcon [3]. Combining
the long time horizon of chess, with the opponent modeling
of poker, and the hand-eye coordination necessary for fast,
fine control of large numbers of game units, RTS games
have garnered significant interest within the computational
and artificial intelligence community. After the success of
AlphaGo, many researchers consider RTS games a significant
next frontier in the computational and articifial intelligence
field [4], [2].
An RTS player builds an economy that provides resources
to generate fighting units to destroy the opponent. The term
“Macro” specifies economy building and choosing the types
and numbers of units to produce. “Micro” refers to the finegrained, nimble, control of units and unit groups during a
skirmish to inflict maximum damage on the enemy while
sustaining minimal damage to friendly units. Most RTS games
only allow maneuvering in two dimensions (2D) with flying
units restricted to a plane and no real three dimensional
(3D) maneuvering (although there are notable exceptions like
Homeworld.1 This paper attacks the problem of generating
good micro for full 3D maneuvering in the context of RTS
games. We believe our approach is extendable to teams of
autonomous real-world entities.
Each RTS game unit type has different properties that
determine the unit’s effectiveness in a skirmish. Depending
on unit type, a unit’s weapons will have a specific range,
a specific amount of armor, and deal different damage to
different opponent unit types. Each unit type has different
movement speed and can take different amounts of damage.
Given this level of complexity, determining the outcome of
a skirmish usually requires a simulation and good RTS micro
can win skirmishes even when outnumbered and unfavourably
positioned.
A. Related work
Good micro for skirmishes in battle or more generally, coordinated group behavior for adaptive agents, has applications in
many fields from wargaming and modeling industrial agents,
to robotics, simulations, and video games [5], [6], [7], [8],
[9], [10], [11], [12], [13], [14]. From flocking, swarming,
and other distributed control algorithms [12], [13] drawn
from observing natural behavior to evolutionary algorithms
that tune the parameters of potential fields that control the
movement of autonomous vehicles [5], much research has
been done in the area of group tactics, adaptive behavior, and
distributed AI. In RTS games, there is strong research interest
in generating good micro since this can lead to your winning
1 http://www.homeworldremastered.com/
skirmishes even when your forces are outnumbered or otherwise disadvantaged. This paper focuses on an evolutionary
computing approach to generating coordinated autonomous
behaviors (micro) for groups of computer game units (or
entities). Specifically, we use an influence map and several
potential fields to represent entity disposition in order to evolve
effective group tactics, or good micro, for controlling a group
of entities battling an opposing group of entities. Good micro
maximizes damage to enemy units while minimizing damage
to friendly units.
Within the games community, Yannakakis [15] evolved
opponent behaviors while Doherty [16] evolved tactical team
behavior for teams of agents. Avery used an evolutionary
computing algorithm to generate influence map parameters
that led to effective group tactics for teams of entities [17],
[18] against a fixed opponent.
In physics, a potential field is usually a distance dependent
vector field generated by a force. For example, the force
of gravity generates a gravitational field, an attractive force,
on objects with mass. This gravitational force depends on
the distance between the objects and on their mass. Well
designed attractive potential fields can keep friendly units close
together while repulsive potential fields enable keeping your
distance from enemies. Combinations (usually linear vector
sums) of several attractive and repulsive potential fields can
lead to fairly complex behavior [19]. Potential fields were
introduced into robotics as a method for real-time obstacle
avoiding navigation and have been used in many video games
for controlling multi-unit group movement [20], [10], [11],
[12]. In this context, Hagelback and Johnson used potential
fields to drive micro in ORTS, a research RTS and partial
clone of Starcraft, a very popular commercial RTS [21], [22].
An influence map structures the world into a 2D or 3D
grid and assigns a value to each grid element or cell. Very
early work used influence maps for spatial reasoning to evolve
a LagoonCraft RTS game player [23]. Sweetser worked on
an AI player designed with an influence map and cellular
automata, where the influence map was used to spatially model
the game world and help the AI player’s decision making in
their RTS game EmerGEnt [24]. Bergsma et al. proposed a
game AI architecture which used influence maps for a turn
based strategy game [25]. Preuss et al. introduced an influence
map based path finding algorithm for group movement in the
RTS game Glest [26], [27]. Su-Hyung et al. used evolutionary
neural networks to evolve non-player characters’ strategies
based on the information provided by a layered influence map
algorithm in the RTS game Conqueror. Uriarte et al. used
influence maps for generating kiting behavior and used this for
their StarCraft (A popular RTS game and research platform)
player Nova [28]. We define potential fields and influence
maps in more detail later in the paper.
To the best of our knowledge, Liu’s work on evolving micro
bots for Starcraft Brood Wars is closest to the work reported in
this paper [29]. Although potential fields and influence maps
guide unit movement, Liu’s approach relies on a hand-coded,
parameterized, control algorithm that defines a space of tar-
geting, firing, and fleeing behaviors. A genetic algorithm then
searches this space for parameters specifying unit behaviors
aiming to maximize damage inflicted on enemy units while
minimizing damage to friendlies against a fixed opponent.
Behaviors like kiting (also known as hit and run) and opponent
encirclement emerge and lead to high performing micro that
can defeat this opponent [30].
Our work differs from Liu’s in that we use a larger number
of potential fields instead of a hand-coded control algorithm
specifying behaviors, and we generalize movement and maneuvering to 3D. Our preliminary results show that we can
quickly evolve unit specific complex behavior including kiting,
fleeing, and englobing. In our experiments with two types
of units called vultures and zealots with different movement
speeds, weapons damage, and health, we can evolve high
performing micro that leads to winning skirmishes. For example, three vultures (fragile, fast, longer ranged units) learn
to disperse and kite a much larger number of baseline zealots
(robust, relatively slow, short ranged units) while suffering
relatively little or no damage. Furthermore, zealots then evolve
against these high performing vultures and learn micro that
increases vulture damage with tactics such as englobing or
dispersing into smaller groups. This kind of manual coevolution points towards our future work where we plan to
co-evolve high performance micro for RTS games.
The next section introduces potential fields, influence maps,
and describes our simulation test-bed. Section III specifies
our representation and describes the genetic algorithm used to
evolve potential field and influence map parameters. We then
provide and discuss our results. The last section summarizes
our results and conclusions and points out directions for future
work.
II. S IMULATION E NVIRONMENT
Starcraft, released in 1998 and Starcraft 2, released in 2010
are the most popular RTS games [22]. In the RTS AI research
community, StarCraft has gained popularity as a research
platform due to the existence of the StarCraft: Brood War
Application Programming Interface (BWAPI) framework, and
the AIIDE and CIG StarCraft AI tournaments [1] which use the
BWAPI to compare Starcraft “Bots,” or Starcraft AI players.
Since Starcraft is essentially 2D and we do not have access to
Starcraft’s source code, we cannot use the BWAPI for our
research. Instead we use FastEcslent, our open source, 3D
capable, modular, RTS game environment. FastEcslent was
developed at the Evolutionary Computing System’s Lab for
computational intelligence research in games, human-robot
and human-computer interaction, and other applications [31].
We note that FastEcslent’s graphics engine runs on a separate
thread and can be turned off, a useful feature for use with
evolutionary computing techniques. We modeled game play
in FastEcslent to be similar to StarCraft.
We modeled two units from Starcraft, Vultures and Zealots.
A Vulture is a somewhat fragile unit with low hit-points but
high movement speed and a ranged weapon. Vultures are
effective when outmaneuvering slower melee units. A Zealot is
TABLE I
U NIT PROPERTIES DEFINED IN FAST E CSLENT
Property
Vulture
Zealot
Hit-points
80
160
MaxSpeed
64
40
MaxDamage
20
16 × 2
Weapon’s Range
256
224
Weapon’s Cooldown
1.1
1.24
where Is is the starting influence, Rh is the entity’s percentage
health, Rc is the entity’s percentage cooldown time, and the
wi for i = [1..3] are evolvable weights. For cells within r
cells of the unit, we then compute the decrease per unit r as
a fraction of the starting influence.
Id = Is If
The influence exerted by a unit at a distance d varying from
0 to r from the unit location is then
I = Is − (dId )
a melee unit with short attack range and low movement speed
but has high hit-points. In the rest of this paper, we use the
term FVulture and FZealot to indicate that our units operate
in 3D and are the Flying versions of the Starcraft units. Since
our research focuses on micro, we disable “fog of war” in our
scenario.
We enabled 3D movement by adding maximum (1000) and
minimum (0) altitudes, as well as a climb rate, rc , of 2. All
units use the RTS physics used in Starcraft and Starcraft2[22].
Noting that a fvulture’s faster speed theoretically makes
it possible for a single expertly microed fvulture to kite a
group of starcraft’s AI controlled fzealots until all fzealots are
eliminated, our experimental scenarios pitted three fvultures
against thirty fzealots. Our baseline AI consisted of hand tuned
potential field parameters that essentially caused fzealots to
move towards and attack the closest enemy unit.
To evaluate a member of the evolving population, we
extracted potential field parameters, created and initialized a
FastEcslent scenerio with these extracted parameters controlling our player’s micro and played against our baseline AI.
At the end of max simulation steps or if all units for one
side were destroyed, the resulting score was returned as this
member’s fitness. The next section describes our representation
and evaluation function in more detail.
III. R EPRESENTATION AND G ENETIC A LGORITHM
When we start our FastEcslent scenario, our player moves
towards a target location defined by the lowest value cell in
our evolving 3D influence map (IM). An IM is a 3D grid
with values assigned to each cell by an IM function. An
IMFunction is usually specified by two parameters, a starting
influence for the location of the entity and a maximum range
(of this influence) in a 3D game world. The influence linearly
decreases to zero as range increases. In this research, we
extend our IMFunction from using two parameters (starting
influence and maximum range) to five parameters in order to
represent more information as shown in Equation 1.
A. Influence maps
To calculate any grid-cell value Vc , we add the influence
from each of the units within the range r from the cell. r is
measured in number of cells. The influence of a unit at the
cell occupied by the unit is computed as the weighted linear
sum below.
Is = w1 Rh + w2 Rc + w3
So the cell at the unit location receives an influence Is and the
eight neighboring cells (at distance 1) receive Is −Id influence,
while cells at distance 2 receive Is − (2Id ). Thus the IM value
at a particular sum given by the sum of all entities’ influence
on that cell is
X
Vc =
(Is − (du Id ))
(1)
u∈r
for all entities u at distance du of the cell. The genetic
algorithm evolves r, If , w1 , w2 and w3 .
This IMFunction not only considers units’ positions in
the game world but also includes the hit-points and weapon
cooldown of each unit. For example, an enemy unit with
low hit-points or its weapon in cooldown can, with the right
wi , generate a low influence map value. Our player always
select the cell with the minimum value as the target attack
location. Thus for example, our units can move towards and
attack a damaged unit (low hitpoints) that cannot fire until
the weapon cooldown period is over. In essence, changes in
the five parameters (w1, w2, w3, r, ∆f ) of the IMFunction
specifies a target location for friendly units. “Move towards
the lowest IM value location attacking enemy units as they
come within range” is our target selection algorithm. This very
simple algorithm can be contrasted with the more complex
target selection and movement algorithms found in [29].
B. Potential Fields
The IM provides a target location to move toward and we
use potential fields to control unit movement to this location.
Potential fields are a type of vector field over space of the
form
f = cde
(2)
where f is the field strength in the direction of the entity
producing the field, c and e are constants and d is distance
(usually).Potential fields in robotics, and in games, have traditionally been used for obstacle avoidance where the target
location manifests an attractive potential field and all obstacles
repel the controlled entity. The vector sum of these potential
fields guides navigation. The direction of the force is in the
direction of the vector difference from the other unit. c and e
are the evolvable parameters for each of the potential fields in
our work.
For specific values of c and e, decentralized flocking and
swarming behaviors emerge when large numbers of units
independently use potential fields to maneuver [32]. Since
good RTS micro takes into account unit health, position, and
weapon state we define and use an attractive and a repulsive
potential field for each of these factors. In addition, we assume
that enemy and friendly units will generate different potential
fields.
Specifically, the potential field controlling a unit’s behavior
is computed from:
• The influence map provided target position ( T)
• The potential fields generated by health (H), distance
(D), and weapon cool-down state (W) of all other units
in the game
Thus the potential field acting on a unit at position p is the
sum of the four potential fields shown below in Equation 3.
F=T+D+H+W
(3)
This potential field controls the unit’s desired heading and
desired speed. The desired heading points in the direction
of the vector and the desired speed is proportional to the
difference between current and desired unit heading. Unit
speed thus increases (up to a maximum) if the unit is already
pointing in the direction of the potential field and decreases
otherwise. In our game, unit speeds vary from 0 to a maximum
(smax ).
The D, H, and W potential fields are composed from
• A direction given by the normalized vector difference
between the current unit’s 3D position p and the other
unit’s 3D position (u)
(u − p)
(4)
n=
|u − p|
where |p − u| denotes the length of the difference vector.
A magnitude given by either the distance to the other
unit, the health of the other unit, or the cool-down time
of the other unit.
• Whether the other unit is a friend or enemy
For example, D is given by
X
X
D=
( nc1 dei 1 − nc2 dei 2 ) +
( nc3 dei 3 − nc4 dei 4 )
•
i∈F
i∈E
where d is distance, n is given by equation 4, and where
odd subscripts for c and d correspond to attraction and even
subscripts correspond to repulsion. The first summation is over
friends (F ) and the second over enemies (E). In the same way,
H is given by
X
X
H=
( nc5 hie5 − nc6 hei 6 ) +
( nc7 hei 7 − nc8 hei 8 )
i∈F
i∈E
where hi denotes the health percentage of the ith other unit,
and W is
X
X
W=
( nc9 wie9 − nc10 wie10 ) +
( nc11 wie11 − nc12 wie12 )
i∈F
i∈E
where wi denotes the percentage time remaining before the
weapon on unit i is ready to fire again. Finally, there is a
single attractive potential field exerted by the target location.
T = mc13 det 13
where dt is the distance to the target location and m is the
direction vector of the target location.
Thus 13 potential fields guide unit movement. The c coefficients have a range of [−10000..10000] and we use 12 bits
for encoding these parameters. The e exponents range from
[−7..8] and we use 4 bits for their encoding. We thus need
12 + 4 = 16 bits per potential field and 16 × 13 = 208 bits
for all potential fields. The influence map parameter r ranges
between 0 to 8 cells while δ, w1 , and w2 range between 0 and
1. Finally w3 also ranges between 0 and 8. This requires 18
bits for a total binary chromosome length of 208 + 18 = 226.
When FastEcslent receives a chromosome, it decodes the
binary string into corresponding parameters. These parameters
then control friendly units as they move in the game world.
The fitness of this chromosome is then computed based on the
amount of damage done to enemy forces and the amount of
damage received by friendly forces during a simulation run in
the game world.
C. Fitness Evaluation
We evaluate individuals in the genetic algorithm’s population in FastEcslent. The decoded chromosome specifies
the parameters needed to compute potential fields and the
influence map within FastEcslent. Every tick of the underlying
simulation clock, friendly units compute potential fields and
update their desired headings and desired speeds. If enemy
units come within weapons range of a friendly unit, the
friendly unit targets the nearest enemy unit. One difference
from Starcraft, is that all units in FastEcslent can fire in any
direction (oriented firing will be investigated as future work).
With the right set of potential fields and influence maps
we expect friendly units using this simple targeting algorithm
to significantly damage enemy units while themselves taking
minimal damage. The simple fitness function below tries to
maximize damage done and minimize damage taken.
X
X
f itness =
De −
Df
enemies
f riends
Where De denotes damage to enemy units and Df , damage
to friendly units. This simple fitness function however, tends
to generate fleeing behavior that simply minimizes damage
to friendly units [29]. One approach to solving this issue
is to wait for the GA to find some way out of this fleeing
behavior, another approach is to fiddle with the fitness function
(for example, give more weight for damage done) to enable
the GA to more easily find a way out [29]. Instead, in this
paper, we re-frame the problem as the problem of exploring
the tradeoff between fleeing and fighting. Evolutionary multiobjective optimization provides an elegant framework for
finding a pareto front of solutions that explore the tradeoff
between damage done and damage received.
Our two objective fitness function is then specified by
X
X
f itness = [(
De ), (1 −
Df )]
enemies
f riends
where both De and Df are normalized to be between 0 and
1 and we are trying to maximize both objectives. We use our
own implementation of Deb’s well known non-dominated sort
elitist genetic algorithm NSGA-II as our evolutionary multiobjective optimization algorithm [33]. The normalized, twoobjective fitness function used within our NSGA-II implementation then produces the results described in the next section.
IV. R ESULTS
We begin by describing experiments conducted to investigate our approach to evolving RTS micro. We drew inspiration
from Starcraft and earlier work by Liu using BWAPI and
created two unit types similar to vultures and zealots. Our
unit equivalents can fly and are hence named fvultures and
fzealots. Unit properties were specified earlier in the paper
in Table I. Because fvultures can kite much larger numbers
of fzealots dealing massive damage while receiving little, our
experiments used 3 fvultures versus 30 fzealots.
In our initial experiments, we had one clump of three
fvultures on the left against a clump of 30 fvultures on the
right side of the map. A clump is defined by a center and a
radius. All units in a clump are distributed randomly within
a sphere defined by this radius (400). Although fvultures
learned to do well against the 30 fzealot clump, their performance degraded significantly when fzealots were not initially
clumped. In addition, when we tried to evolve fzealot micro
against fvultures, we found the same kind of over-specialized
behavior emerging. To reduce overspecialization, we created
three training scenarios (or maps). These scenarios are defined
in terms of clumps and clouds. We have already defined
clumps. A cloud, like a clump is defined by a center and a
radius. However, the units in a cloud are distributed randomly
within 10 units around the sphere boundary defined by the
center and radius (also 400). The first scenario consists of a
clump of 3 fvultures versus a clump of 30 fzealots. The second,
a clump of 3 fvultures surrounded by a spherical cloud of 30
fzealots, and the last scenario consists of a clump of 30 fzealots
surrounded by a (small) cloud of 3 fvultures.
Unit locations within the clump and cloud were generated
randomly. Our evaluation function ran each of these three
scenarios and the objective function values for each objective
were averaged over these three scenarios. Evaluations took
longer but the resulting performance was better and seemed
to generalize well.
A. Pareto front evolution
Figure 1 shows the evolution of the pareto front at intervals
of five (5) generations for one run of our NSGA-II. We used
a population size of 50, run for 75 generations. Simple two
point crossover with a crossover probability of 0.9 and flip
mutation with a probability of 0.05 worked well with our
implementation of NSGA-II. The x-axis plots the first objective, damage done to fzealots, and the y-axis plots 1− damage
received by fvultures. Figure 2 shows the same information for
fzealots evolving against already evolved fvultures. Figure 1
and 2 illustrate how the pareto front progresses for one run
of the multi-objective GA. Broadly speaking, the pareto front
moves towards (1, 1) on the plot making progress towards
Fig. 1. Fvultures evolving against Fzealots
Fig. 2. Fzealots evolving against Fvultures
maximizing damage done and minimizing damage received.
We evolve micro that can destroy all fzealots but this also
involves taking significant damage as shown by the solutions
on the bottom right of Figure 1. We also see that when battling
fvultures from this area of the plot, fzealots can do significant
damage but never evolve micro to kill all the fvultures during
the single run of the NSGA-II shown in Figure 2.
In our experiments, we ran the genetic algorithm ten times
with different random seeds and Figure 3 shows the result.
This figure plots the combined pareto front in the first generation over all ten random seeds versus the combined pareto
front in the last generation over the ten random seeds. That is,
we first did a set union of the pareto fronts in the ten initial
randomly generated populations. The points in this union over
all ten runs are displayed as squares for the initial generation
(generation 0) points and as circles for the points in the
final generation (generation 74). Let us label the union of
the generation 0 pareto fronts as U010 and the union of the
10
generation 74 pareto fronts as U74
. We then compute and plot
the pareto front of this set union - the line on the left. We do
the same for the ten pareto fronts in the last generation - the
line on the right. Let P010 denote the combined ten run pareto
10
front in the first generation and P74
denote the combined ten
run pareto front in the last generation; Figure 3 then shows
progress between first and last generation over all ten runs
against an fzealot from the bottom right of Figure 2. Note that
P010 represents not an average, but the best (non-dominated)
individuals over all ten runs. Individual runs produce initial
pareto fronts that look like the intial front in Figure 1. We also
have videos that show how the fvultures and fzealots behave
at http://www.cse.unr.edu/∼sushil/RTSMicro. Figure 3 shows
and this is reflected in the pareto front moving further to the
right (more damage done to fzealots) and higher (less damage
taken by fvultures). At generation 74, we have a choice of
micro control tactics provided by the points on the pareto front
at generation 74. At one extreme, indicated by the point at
the very top on the generation 74 pareto front (.87, 1.0), all
three fvultures stay alive with no damage while inflicting 87%
damage on the fzealots. 87% damage corresponds to killing
26 out of the 30 fvultures. At the other extreme, denoted
by the point on the bottom right (.99, .32), two fvultures die
while killing 99% of the fzealots. The points in between these
extremes represent other damage done versus damage received
tradeoffs and picking which of these tactics to use depends on
the player’s in-game needs of the moment. That our multiobjective approach naturally provides a range of choices for
micro is an elegant side affect of this approach.
B. Random parameter values
Fig. 3. The initial and final pareto front over ten runs for Fvulture micro.
that even initial generations contain a range of acceptable
micro behaviors. Since the target location is near the opponents
and exerts a purely attractive potential field and units fire when
in range, fzealots and fvultures usually come together and
fight. As the potential field parametes evolve, fvultures learn
to take less damage while doing more damage to fzealots.
The videos support this observation and also show the range
of micro behaviors evolving. Some can be simply described
as “fleeing,” while others, even in early generations, defy
succint explanation. Broadly speaking, the micro behaviors
being evolved tend to keep fvultures outside the weapons range
of the fzealots while looking for an opportunity to dart in
opportunistically and deliver damage. Early on, if fvultures are
surrounded by zealots as in training scenario two, the fvultures
inflict significant damage but have not learned to kite and so
also suffer significant damage. It is worthwhile to note that
the opponent fzealots tend to favor tight circling on a plane,
sticking close together while the fvultures are more dispersed,
tend to change altitude more frequently, and try to keep their
distance from fzealots as they try to balance their attractive
and repulsive potentials.
In later generations the fvultures have learned potential field
parameter values that lead to kiting (or kiting-like) behavior
We also investigated how random potential field parameter
values perform on our three scenarios and on random scenarios. We generated random chromosomes that specified these
parameter values and evaluated them on the three training
scenarios and on a set of on hundred random scenarios. In
these random scenarios, the starting positions and orientations
of the 30 fzealots and 3 fvultures were randomly spread
within a larger clump with radius 500. Figure 4 plots the
average objective values and their standard deviations obtained
from evaluting the 3750 randomly generated chromosomes on
both sets of scenarios. The average objective value on the
three scenarios is average along each objective over 3750
evaluations. For the hundred (100) random scenarios, the
average is along each objective over 3750 × 100 evaluations.
The 3750 comes from a 50 population size multiplied by 75
evolutionary generations equating to the number of evaluations
during one run of the GA. The horizontal and vertical lines
on the plot indicate variation along each objective and are
centered at the average objective values over the 3750 random
chromosomes on the three training scenarios and the hundred
random scenarios. Specifically, the length of a line equals one
standard deviation and we display half a standard deviation
10
on either side of the average. We also plot U010 and U74
for
reference. Figure 4 shows that the average randomly generated
micro behavior on our three training scenarios has fvultures
fleeing fzealots and doing little damage while much little if
any damage. On the other hand, randomly generated behaviors
on random scenarios average out to be near the center of the
pareto plot. They take a little less damage than they mete
out. Both these behaviors can be explained by the positioning
of units in the scenarios. Assuming random parameter values
produce random movement, initial positioning will play a
large part in determing both objectives in the fitness function.
The three training scenarios seperate opposing units and so
fvultures tend to stay seperated from other units and take
little damage. On the random scenarios, randomly positioned
fvultures and fzealots moving randomly will take damage from
Fig. 4. Random chromosome performance on three training scenarios and on
100 random scenarios
each other before their random movement places them out of
range and the clump disperses.
C. Random locations
These results provide evidence that we can use our approach
to evolve a range of good micro against a fixed opponent on a
given set of scenarios. But does the evolved micro generalize
to other scenarios? To investigate this question we also ran the
10
solutions in P010 and P74
on a set of 100 random scenarios
(as described above). The average value of the two objectives
over all 100 random scenarios and their standard deviations
along each objective obtained by evaluating the individuals in
10
P010 and P74
on these random scenarios is shown in Figure 5
10
along with U010 and U74
, again for reference. With arbitrary
10 on a set of 100 random starting
Fig. 5. Average performance of P010 and P74
positions and orientations
starting locations and orientations that do not correspond to
training scenarios, we do not expect fvultures to do as well
on average. We also expect wide variation in performance.
As expected the figure shows that the average performance
of the individuals from P010 on the 100 random scenarios is
dominated by the points on the initial pareto front provided by
U010 . Let us label this average performance R0100 . The average
10
performance of the individuals from P74
on the 100 random
100
scenarios (R74 ) shows improvement over R0100 , but this is
10
still dominated by points on U74
. We also see the existence
of significant variation along both objectives and that the
standard deviations along both objectives decreases from R0100
100
to R74
. It is encouraging to see that although performance
degrades on these 100 random scenarios, individuals from
generation 74 perform better on average than individuals from
the initial generation on the training scenarios. The change in
average performance from the initial generation to the final
generation works out to approximately three (3) more fzealots
eliminated and one more fvulture surviving. Although we
were not specifically after robust solutions, the micro tactics
10
from R74
on average do eliminate the majority of fzealots,
while keeping the majority of fvultures alive on random neverseen scenarios. The figure provides evidence that initial unit
positioning can significantly affect skirmish outcome. This is
true for initial unit positioning in RTS games as well.
V. C ONCLUSION AND F UTURE W ORK
This paper proposed a new approach to evolving micro for
RTS games in three dimensions for two different units. Using
an influence map and an attractive potential field to determine
a target position to attack, we used attractive and repulsive
potential fields for distance, health, and weapon cooldown
to evolve micro behavior for units in our 3D RTS game
arena. Our NSGA-II implementation optimized micro based
on a multi-objective formulation of the fitness function that
maximized damage done and minimized damage taken. The
results show that we can evolve a range of micro behaviors
that go from taking and doing little damage, through micro
that balances damage done and taken, and finally to micro
that does maximal damage while taking some damage. On the
three training scenarios, the evolved micro can eliminate 25 out
of 30 fzealots while keeping all three (3) fvultures alive. Other
evolved micro (on the pareto front) can eliminate all fzealots
but at the cost of two fvultures. The multi-objective problem
formulation and the NSGA-II lead to evolving pareto fronts
that directly and naturally produce this wide range of micro
choices. The fast, longer ranged, but fragile units (fvultures)
learned micro behavior similar to hit and run (kiting) that is
used by human players in RTS games like Starcraft. Videos of
the micro behavior of these fvultures versus fzealots is available at http://www.cseunr.edu/∼sushil/RTSMicro and show the
range of complex behavior evolved.
Our genetic algorithm took 3750 evaluations to evolve the
above micro. We generated 3750 random chromosomes and
evaluated their fitness on our three training scenarios as well as
a hundred random scenarios. As expected, the average micro
behavior of random parameter values on the three training
scenarios was well dominated by the evolved pareto front on
the three training scenarios. Random chromosomes also led to
an average micro performance that balanced damage done and
taken and were explained based on the random dispositions of
units in these scenarios.
Finally, experiments on a hundred random, never before
encountered scenarios show that the potential field values
evolved on three training scenarios used during evolution do
not perform badly on these new random test scenarios. Although we were not aiming for robustness, micro performance
on random scenarios is balanced and shows improvement over
time.
We plan to follow two paths in the future. First, we would
like to investigate how our approach scales with multiple
unit types. That is, we would like to investigate micro for
a group composed from multiple unit types versus another
group also composed of multiple unit types. Second, we
plan to investigate the co-evolution of micro using our new
representation and appoach. Our current results show that we
can evolve good micro against a fixed opponent, we would
like to use a multi-objective, co-evolutionary algorithm to coevolve a range of micro that is robust against a range of
opposition micro.
R EFERENCES
[1] M. Buro, “Real-time strategy games: A new AI research challenge,”
Proceedings of the 18th International Joint Conference on Artificial
Intelligence. International Joint Conferences on Artificial Intelligence,
pp. 1534–1535, 2003.
[2] S. Ontañon, G. Synnaeve, A. Uriarte, F. Richoux, D. Churchill,
M. Preuss et al., “A survey of real-time strategy game AI research
and competition in StarCraft,” IEEE Transactions on Computational
Intelligence and AI in games, vol. 5, no. 4, pp. 1–19, 2013.
[3] (2016) Starcraft world championship series. [Online]. Available:
http://wcs.battle.net/sc2
[4] D. Hassabis. (2016) Official google blog: What we learned in seoul with
alphago. [Online]. Available: https://googleblog.blogspot.com/2016/03/
what-we-learned-in-seoul-with-alphago.html
[5] P. Vadakkepat, K. Tan, and W. Ming-Liang, “Evolutionary artificial
potential fields and their application in real time robot path planning,”
in Evolutionary Computation, 2000. Proceedings of the 2000 Congress
on, vol. 1. IEEE, 2000, pp. 256–263.
[6] J. Ferber and O. Gutknecht, “A meta-model for the analysis and design
of organizations in multi-agent systems,” in Multi Agent Systems, 1998.
Proceedings. International Conference on. IEEE, 1998, pp. 128–135.
[7] M. Barbuceanu and M. Fox, “Cool: A language for describing coordination in multi agent systems,” in Proceedings of the First International
Conference on Multi-Agent Systems (ICMAS-95). Citeseer, 1995, pp.
17–24.
[8] N. Jennings, “Commitments and conventions: The foundation of coordination in multi-agent systems,” Knowledge Engineering Review, vol. 8,
pp. 223–223, 1993.
[9] ——, “Controlling cooperative problem solving in industrial multi-agent
systems using joint intentions,” Artificial intelligence, vol. 75, no. 2, pp.
195–240, 1995.
[10] R. Olfati-Saber, J. Fax, and R. Murray, “Consensus and cooperation in
networked multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1,
pp. 215–233, 2007.
[11] M. Egerstedt and X. Hu, “Formation constrained multi-agent control,”
Robotics and Automation, IEEE Transactions on, vol. 17, no. 6, pp.
947–951, 2001.
[12] C. Reynolds, “Flocks, herds and schools: A distributed behavioral
model,” in ACM SIGGRAPH Computer Graphics, vol. 21, no. 4. ACM,
1987, pp. 25–34.
[13] Y. Chuang, Y. Huang, M. D’Orsogna, and A. Bertozzi, “Multi-vehicle
flocking: scalability of cooperative control algorithms using pairwise
potentials,” in Robotics and Automation, 2007 IEEE International Conference on. IEEE, 2007, pp. 2292–2299.
[14] P. Dasgupta, “A multiagent swarming system for distributed automatic
target recognition using unmanned aerial vehicles,” Systems, Man and
Cybernetics, Part A: Systems and Humans, IEEE Transactions on,
vol. 38, no. 3, pp. 549–563, 2008.
[15] G. Yannakakis and J. Hallam, “Evolving opponents for interesting
interactive computer games,” From Animals to Animats, vol. 8, pp. 499–
508, 2004.
[16] D. Doherty and C. ORiordan, “Evolving tactical behaviours for teams
of agents in single player action games,” in Proceedings of the 9th
International Conference on Computer Games: AI, Animation, Mobile,
Educational & Serious Games, 2006, pp. 121–126.
[17] P. Avery, S. Louis, and B. Avery, “Evolving coordinated spatial tactics
for autonomous entities using influence maps,” in Computational Intelligence and Games, 2009. CIG 2009. IEEE Symposium on. IEEE, 2009,
pp. 341–348.
[18] P. Avery and S. Louis, “Coevolving team tactics for a real-time strategy
game,” in Evolutionary Computation (CEC), 2010 IEEE Congress on.
IEEE, 2010, pp. 1–8.
[19] V. Braitenberg, Vehicles: Experiments in synthetic psychology. MIT
press, 1986.
[20] O. Khatib, “Real-time obstacle avoidance for manipulators and mobile
robots,” The international journal of robotics research, vol. 5.
[21] J. Hagelbäck and S. Johansson, “Using multi-agent potential fields
in real-time strategy games,” in Proceedings of the 7th international
joint conference on Autonomous agents and multiagent systems-Volume
2. International Foundation for Autonomous Agents and Multiagent
Systems, 2008, pp. 631–638.
[22] (2016) Starcraft. [Online]. Available: http://us.battle.net/sc2
[23] C. Miles, J. Quiroz, R. Leigh, and S. Louis, “Co-evolving influence
map tree based strategy game players,” in Computational Intelligence
and Games, 2007. CIG 2007. IEEE Symposium on, april 2007, pp. 88
–95.
[24] P. Sweetser and J. Wiles, “Combining influence maps and cellular
automata for reactive game agents,” Intelligent Data Engineering and
Automated Learning-IDEAL 2005, pp. 209–215, 2005.
[25] M. Bergsma and P. Spronck, “Adaptive spatial reasoning for turn-based
strategy games,” Proceedings of AIIDE, 2008.
[26] M. Preuss, N. Beume, H. Danielsiek, T. Hein, B. Naujoks, N. Piatkowski,
R. Stuer, A. Thom, and S. Wessing, “Towards intelligent team composition and maneuvering in real-time strategy games,” Computational
Intelligence and AI in Games, IEEE Transactions on, vol. 2, no. 2, pp.
82–98, 2010.
[27] H. Danielsiek, R. Stuer, A. Thom, N. Beume, B. Naujoks, and M. Preuss,
“Intelligent moving of groups in real-time strategy games,” in Computational Intelligence and Games, 2008. CIG’08. IEEE Symposium On.
IEEE, 2008, pp. 71–78.
[28] A. Uriarte and S. Ontañón, “Kiting in RTS games using influence maps,”
in Eighth Artificial Intelligence and Interactive Digital Entertainment
Conference, 2012.
[29] S. Liu, S. Louis, and C. Ballinger, “Evolving effective micro behaviors
in real-time strategy games,” IEEE Transactions on Computational
Intelligence and AI in Games, vol. PP, no. 99, pp. 1–1, 2016.
[30] S. Liu, S. J. Louis, and M. Nicolescu, “Comparing heuristic search methods for finding effective group behaviors in RTS game,” in Evolutionary
Computation (CEC), 2013 IEEE Congress on. IEEE, 2013, pp. 1371–
1378.
[31] (2015) Fast evoltionary computing systems lab entity engine. [Online].
Available: http://ecsl.cse.unr.edu/
[32] D. M. Bourg and G. Seemann, AI for Game Developers: Creating
Intelligent Behavior in Games. O’Reilly Media, 2004.
[33] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: Nsga-ii,” IEEE transactions on evolutionary computation, vol. 6, no. 2, pp. 182–197, 2002.
| 9cs.NE
|
1
VIDOSAT: High-dimensional Sparsifying
Transform Learning for Online Video Denoising
arXiv:1710.00947v1 [cs.CV] 3 Oct 2017
Bihan Wen, Student Member, IEEE, Saiprasad Ravishankar, Member, IEEE, and Yoram Bresler, Fellow, IEEE
Abstract—Techniques exploiting the sparsity of images in a
transform domain have been effective for various applications in
image and video processing. Transform learning methods involve
cheap computations and have been demonstrated to perform well
in applications such as image denoising and medical image reconstruction. Recently, we proposed methods for online learning of
sparsifying transforms from streaming signals, which enjoy good
convergence guarantees, and involve lower computational costs
than online synthesis dictionary learning. In this work, we apply
online transform learning to video denoising. We present a novel
framework for online video denoising based on high-dimensional
sparsifying transform learning for spatio-temporal patches. The
patches are constructed either from corresponding 2D patches in
successive frames or using an online block matching technique.
The proposed online video denoising requires little memory, and
offers efficient processing. Numerical experiments compare the
performance to the proposed video denoising scheme but fixing
the transform to be 3D DCT, as well as prior schemes such
as dictionary learning-based schemes, and the state-of-the-art
VBM3D and VBM4D on several video data sets, demonstrating
the promising performance of the proposed methods.
Index Terms—Sparse representations, Sparsifying transforms,
Machine learning, Data-driven techniques, Online learning, Big
data, Video denoising.
I. I NTRODUCTION
Recent techniques in image and video processing make
use of sophisticated models of signals and images. Various
properties such as sparsity, low-rank, etc., have been exploited
in inverse problems such as video denoising, or other dynamic
image reconstruction problems such as magnetic resonance
imaging or positron emission tomography [1], [2]. Adaptive
or data-driven models and approaches are gaining increasing
interest. This work presents novel online data-driven video
denoising techniques based on learning sparsifying transforms for appropriately constructed spatio-temporal patches
of videos. This new framework provides high quality video
restoration from highly corrupted data. In the following, we
briefly review the background on video denoising and sparsifying transform learning, before discussing the contributions
of this work.
This work was supported in part by the National Science Foundation
(NSF) under grant CCF-1320953. Saiprasad Ravishankar was supported in
part by the following grants: ONR grant N00014-15-1-2141, DARPA Young
Faculty Award D14AP00086, ARO MURI grants W911NF-11-1-0391 and
2015-05174-05, and a UM-SJTU seed grant.
B. Wen, and Y. Bresler are with the Department of Electrical and Computer
Engineering and the Coordinated Science Laboratory, University of Illinois,
Urbana-Champaign, IL, 61801 USA e-mail: (bwen3, ybresler)@illinois.edu.
S. Ravishankar is with the Department of Electrical Engineering and
Computer Science, University of Michigan, Ann Arbor, MI 48109, USA email: [email protected].
Input FIFO buffer
(3D Array):
…
𝒑𝒕𝒉 3D tensor
…
𝑛1
−(𝒎−𝟏)/𝟐
−𝒎+𝟏
𝒎
𝑛2
Vectorized 𝒖𝒊
𝒏
Temporal Sliding Window, depth = 𝑚, at 𝜏 = 𝑡
Fig. 1.
Mini-batch 𝑼𝒋 =
[𝒖𝒋𝑴−𝑴+𝟏 | … |𝒖𝒋𝑴 ]
cascade
𝒏 = 𝒏𝟏 × 𝒏𝟐 × 𝒎
𝑴
Video streaming, tensor construction and vectorization.
A. Video Denoising
Denoising is one of the most important problems in video
processing. The ubiquitous use of relatively low-quality smart
phone cameras has also led to the increasing importance of
video denoising. Recovering high-quality video from noisy
footage also improves robustness in high-level vision tasks
[3], [4].
Though image denoising algorithms, such as the popular
BM3D method [5] can be applied to each video frame independently, most of the video denoising techniques (or more
generally, methods for reconstructing dynamic data from measurements [6], [7]) exploit the spatio-temporal correlation in
dynamic image sequences. Natural videos have local structures
that are sparse or compressible in some transform domain, or
in certain dictionaries, e.g., discrete cosince transform (DCT)
[8] and wavelets [9]. Prior works exploited this fact and
proposed video (or high-dimensional data) denoising algorithms based on adaptive sparse approximation [10] or Wiener
filtering [11]. Videos also typically involve various kinds of
motion or dynamics in the scene, e.g., moving objects or
humans, rotations, etc. Recent state-of-the-art video and image
denoising algorithms utilize block matching (BM) to group
local patches over space and time (to account for motion),
and apply denoising jointly for such matched data [5], [11],
[12]. Table I summarizes the key attributes of the popular
and related video denoised methods, as well as the proposed
methods.
B. Sparsifying Transform Learning
Many of the aforementioned video denoising methods exploit sparsity in a fixed transform domain (e.g., DCT) as
part of their framework. Several recent works have shown
that the data-driven adaptation of sparse signal models (e.g.,
based on training signals, or directly from corrupted mea-
2
Methods
fBM3D
3D DCT
sKSVD
VBM3D
VBM4D
VIDOSAT
VIDOSAT
-BM
Sparse Signal Model
Fixed Adaptive Online
3
3
3
3
3
3
3
BM
Temporal
Correlation
3
3
3
Noisy Video
Stream
3
3
3
3
3
𝑌1
𝑌t+1
𝑌t
Input FIFO
Buffer
𝑌t−m+1 | … |𝑌t =
VIDOSAT Mini-batch Denoising
Y~t
accumulate denoised patches
3
3
3
3
C OMPARISON BETWEEN VIDEO DENOISING METHODS , INCLUDING
F BM3D [5], 3D DCT, S KSVD [10], VBM3D [11], VBM4D [12], AS
WELL AS VIDOSAT AND VIDOSAT-BM PRPOPOSED HERE . F BM3D IS
APPLYING BM3D ALGORITHM FOR DENOISING EACH FRAME , AND THE
3D DCT METHOD IS APPLYING THE VIDOSAT FRAMEWORK BUT USING
THE FIXED 3D DCT TRANSFORM .
surements) usually leads to high quality results (e.g., compared to fixed or analytical models) in many applications
[13]–[24]. Synthesis dictionary learning is the best-known
adaptive sparse representation technique [13], [25]. However,
obtaining optimal sparse representations of signals in synthesis
dictionary models, known as synthesis sparse coding, is NPhard (Non-deterministic Polynomial-time hard) in general. The
commonly used approximate sparse coding algorithms [26]–
[29] typically still involve relatively expensive computations
for large-scale problems.
As an alternative, the sparsifying transform model suggests
that the signal u is approximately sparsifiable using a transform W ∈ Rm×n , i.e., Wu = x + e, with x ∈ Rm a sparse
vector called the sparse code and e a modeling error term in
the transform domain. A key advantage of this model over
the synthesis dictionary model, is that for a given transform
W, the optimal sparse code x of sparsity level s minimizing
the modeling error kek2 is obtained exactly and cheaply
by simple thresholding of Wu to its s largest magnitude
components. Another advantage is that with u being given
data, the transform model does not involve a product between
W and unknown data, so learning algorithms for W can be
simpler and more reliable. Recent works proposed learning
sparsifying transforms [19], [30] with cheap algorithms that
alternate between updating the sparse approximations of training signals in a transform domain using simple thresholdingbased transform sparse coding, and efficiently updating the
sparsifying transform. Various properties have been found to
be useful for learned transforms such as double sparsity [20],
union-of-transforms [21], rotation and flip invariance [31], etc.
Transform learning-based techniques have been shown to be
useful in various applications such as sparse data representations, image denoising, inpainting, segmentation, magnetic
resonance imaging (MRI), and computed tomography (CT)
[21], [24], [31]–[38].
In prior works on batch transform learning [19], [21],
[30], [31], the transform was adapted using all the training data, which is efficient and comes with a convergence
guarantee. When processing large-scale streaming data, it is
also important to compute results online, or sequentially over
Output FIFO
Buffer
ഥt−m+1 | … |Y
ഥt =
Y
TABLE I
Normalize & Output the Oldest Frame
t−m+1
Y
Fig. 2.
1
Y
Denoised
Stream
Illustration of online video streaming and denoising framework.
time. Our recent work [39], [40] proposed online transform
learning, which sequentially adapts the sparsifying transform
and transform-sparse coefficients for sequentially processed
signals. This approach involves cheap computation and limited
memory requirement. Compared to popular techniques for
online synthesis dictionary learning [41], the online adaptation
of sparsifying transforms allows for cheaper or exact updates
[39], and is thus well suited for high-dimensional data applications.
C. Methodologies and Contributions
While the data-driven adaptation of synthesis dictionaries
for the purpose of denoising image sequences or volumetric
data [10], [15] has been studied in some recent papers, the
usefulness of learned sparsifying transforms has not been
explored in these applications. Video data typically contain
correlation along the temporal dimension, which will not
be captured by learning sparsifying transforms for the 2D
patches of the video frames. We focus on video denoising
using high-dimensional online transform learning. We refer
to our proposed framework as VIdeo Denoising by Online
SpArsifying Transform learning (VIDOSAT). Spatio-temporal
(3D) patches are constructed using local 2D patches of the
corrupted video, and the sparsifying transform is adapted
to these 3D patches on-the-fly. Fig.1 illustrates one way
of constructing the (vectorized) spatio-temporal patches or
tensors from the streaming video, and Fig.2 is a flow-chart of
the proposed VIDOSAT framework. Though we consider 3D
spatio-temporal tensors formed by 2D patches for gray-scale
video denoising in this work, the proposed denoising methods
readily apply to higer-dimensional data (e.g., color video [42],
hyperspectral images, dynamic 3D MRI, etc) as well.
As far as we know, this is the first online video denoising
method using adaptive sparse signal modeling, and the first
application of high-dimensional sparsifying transform learning
to spatio-temporal data. Our methodology and results are
summarized as follows:
• The proposed video denoising framework processes noisy
frames in an online, sequential fashion to produce streaming denoised video frames. The algorithms require limited
3
storage of a few video frames, and modest computation,
scaling linearly with the number of pixels per frame. As
such, our methods would be able to handle high definition
/ high rate video enabling real-time output with controlled
delay, using modest computational resources.
• The online transform learning technique exploits the
spatio-temporal structure of the video tensors (patches)
using adaptive 3D transform-domain sparsity to process
them sequentially. The denoised tensors are aggregated
to reconstruct the streaming video frames.
• We evaluate the video denoising performance of the
proposed algorithms for several datasets, and demonstrate
their promising performance compared to several prior or
related methods.
This paper is an extension of our previous conference work
[43] that briefly investigated a specific VIDOSAT method.
Compared with this earlier work, here we investigate different
VIDOSAT methodologies such as involving block matching
(referred to as VIDOSAT-BM). Moreover, we provide detailed
experimental results illustrating the properties of the proposed
methods and their performance for several datasets, with extensive evaluation and comparison to prior or related methods.
We also demonstrate the advantages of VIDOSAT-BM over the
VIDOSAT approach of [43].
D. Major Notations
We use the following notations in this work. Vectors (resp.
matrices) are denoted by boldface lowercase (resp. uppercase)
letters such as u (resp. U). We use calligraphic uppercase
letters (e.g., U) to denote tensors. We denote the vectorization
operator for 3D tensors (i.e., for reshaping a 3D array into a
vector) as vec(·) : Rn1 ×n2 ×m → Rn . The vectorized tensor
is u = vec(U) ∈ Rn , with n = n1 n2 m. Correspondingly,
the inverse of the vectorization operator vec−1 (·) : Rn →
Rn1 ×n2 ×m denotes a tensorization operator. The relationship
is summarized as follows:
U ∈ Rn1 ×n2 ×m
vec
vec−1
u ∈ Rn .
The other major notations of the indices and variables that are
used in this work are summarized in Table II. We denote the
underlying signal or variable as ũ, and its noisy measurement
(resp. estimate) is denoted as u (resp. û). The other notations
used in our algorithms are discussed in later sections.
E. Organization
The rest of the paper is organized as follows. Section II
briefly discusses the recently proposed formulations for timesequential signal denoising based on online and mini-batch
sparsifying transform learning [39], [40]. Then, Section III
presents the proposed online video processing framework, and
two online approaches for denoising dynamic data. Section IV
describes efficient algorithms for the proposed formulations.
Section V demonstrates the behavior and promise of the
proposed algorithms for denoising several datasets. Section VI
concludes with proposals for future work.
Indices
Definition
Range
τ
time index
1, 2, 3, etc.
p
spatial index of 3D patches in Yτ
1...P
i
index of patches within mini-batch
1...M
k
local mini-batches index at time τ
1...N
j or Lτk
global mini-batch index
1, 2, 3, etc.
Variables
Definition
Dimension
Wτ
adaptive sparsifying transform
n×n
Yτ
video frames
a×b
Yτ
input FIFO buffer
a×b×m
Ȳτ
output FIFO buffer
a×b×m
Uj
mini-batch of vectorized data
n×M
Xj
sparse codes of the mini-batch
n×M
vp
vectorized 3D patch
n = n1 n2 m
Operators
Definition
Mapping
Rp
extracts 3D patch in A1
RP →
Bp
forms 3D patch in A2 by BM
Rn1 ×n2 ×m
Rp∗
Bp∗
patch deposit operator in A1
Rn1 ×n2 ×m
patch deposit operator in A2
→ RP
TABLE II
N OTATIONS OF THE INDICES AND THE MAIN VARIABLES AND OPERATORS .
II. S IGNAL D ENOISING VIA O NLINE T RANSFORM
L EARNING
The goal in denoising is to recover an estimate of a signal
ũ ∈ Rn from the measurement u = ũ + e, corrupted
by additive noise e. Here, we consider a time sequence of
noisy measurements {ut }, with ut = ũt + et . We assume
noise et ∈ Rn whose entries are independent and identically
distributed (i.i.d.) Gaussian with zero mean and possibly
time-varying but known variance σt2 . Online denoising is to
recover the estimates ût for ũt ∀ t sequentially. Such timesequential denoising with low memory requirements would be
especially useful for streaming data applications. We assume
that the underlying signals {ũt } are approximately sparse in
an (unknown, or to be estimated) transform domain.
A. Online Transform Learning
In prior work [39], we proposed an online signal denoising
methodology based on sparsifying transform learning, where
the transform is adapted based on sequentially processed
data. For time t = 1, 2, 3, etc, the problem of updating the
adaptive sparsifying transform and sparse code (i.e., the sparse
representation in the adaptive transform domain) to account for
the new noisy signal ut ∈ Rn is
t
n
o
o
1 Xn
2
Ŵt , x̂t = arg min
kWuτ − xτ k2 + λτ ν(W)
W, xt t τ =1
t
1X 2
+
α kxτ k0
t τ =1 τ
s.t. xτ = x̂τ , 1 ≤ τ ≤ t − 1 (P1)
where the `0 “norm” counts the number of nonzeros in xτ ,
2
which is the sparse code of uτ . Thus kWuτ − xτ k2 is the
4
sparsification error (i.e., the modeling error in the transform
model) for uτ in the transform W. The term ν(W) =
2
− log |det W| + kWkF is a transform learning regularizer
2
[19], λτ = λ0 kuτ k2 with λ0 > 0 allows the regularizer term
to scale with the first term in the cost, and the weight ατ is
chosen proportional to στ (the standard deviation of noise in
ũτ ). Matrix Ŵt in (P1) is the optimal transform at time t,
and x̂t is the optimal sparse code for ut .
Note that at time t, only the latest optimal sparse code x̂t is
updated in (P1)1 along with the transform Ŵt . The condition
xτ = x̂τ , 1 ≤ τ ≤ t−1, is therefore assumed. For brevity, we
will not explicitly restate this condition (or, its variants) in the
formulations in the rest of this paper. Although at each time t
the transform is updated based on all the past and present
observed data, the online algorithm for (P1) [39] involves
efficient operations based on a few matrices of modest size,
accumulated sequentially over time.
The regularizer ν(W) in (P1) prevents trivial solutions
and controls the condition number and scaling of the learnt
transform [19]. The condition number κ(W) is upper bounded
by a monotonically increasing function of ν(W) [19]. In the
limit λ0 → ∞ (and assuming the uτ , 1 ≤ τ ≤ t, are not all
zero), the condition number of the optimal transform in (P1)
tends to 1. The specific choice of λ0 (and hence the condition
number) depends on the application.
1) Denoising: Given the optimal transform Ŵt and the
sparse code x̂t , a simple estimate of the denoised signal
is obtained as ût = Ŵt−1 x̂t . Online transform learning
can also be used for patch-based denoising of large images
[39]. Overlapping patches of the noisy images are processed
sequentially (e.g., in raster scan order) via (P1), and the
denoised image is obtained by averaging together the denoised
patches at their respective image locations.
2) Forgetting factor: For non-stationary or highly dynamic
data, it may not be desirable to uniformly fit a single transform
W to all the uτ , 1 ≤ τ ≤ t, in (P1). Such data can be handled
by introducing a forgetting factor ρt−τ (with a constant 0 <
ρ < 1) that scales the terms in (P1) [39]. The forgetting factor
diminishes the influence of “old” data. The objective function
in this case is modified as
t
o
1 X t−τ n
2
ρ
kWuτ − xτ k2 + λτ ν(W) + ατ2 kxτ k0 . (1)
Ct τ =1
where Ct =
Pt
τ =1
ρt−τ is the normalization factor.
B. Mini-batch learning
Another useful variation of Problem (P1) involves minibatch learning, where a block (group), or mini-batch of signals
is processed at a time [39]. Assuming a fixed mini-batch
size
M , the Lth (L ≥ 1) mini-batch of signals is UL =
uLM −M +1 | uLM −M +2 | ... | uLM . For L = 1, 2, 3, etc,
1 This is because only the signal ũ is assumed to be stored in memory at
t
time t for the online scheme.
the mini-batch sparsifying transform learning problem is
(P2)
L
n
o
1 X
2
kWUj − Xj kF
ŴL , X̂L = arg min
LM
W, XL
j=1
LM
L
1 X 2
1 X
+
αl kxl k0 +
Λj ν(W)
LM
LM j=1
l=1
2
where the regularizer
weight is Λj = λ0 Uj F , and
the matrix XL = xLM −M +1 | xLM −M +2 | ... | xLM contains
the block of sparse codes corresponding to UL .
Since we only consider a finite number of frames or patches
in practice (e.g., in the proposed VIDOSAT algorithms), the
normalizations by 1/t in (P1), 1/Ct in (1), and 1/LM
in (P2)
correspondingly
have
no
effect
on
the
optimum
Ŵt , X̂t
or ŴL , X̂L . Thus we drop, for clarity2 , normalization
factors from (P3) and all subsequent expressions for the cost
functions.
Once (P2) is solved, a simple denoised estimate of the noisy
−1
block of signals in UL is obtained as ÛL = ŴL
X̂L . The
mini-batch transform learning Problem (P2) is a generalized
version of (P1), with (P2) being equivalent to (P1) for M = 1.
Similar to (1), (P2) can be modified to include a forgetting
factor. Mini-batch learning can provide potential speedups over
the M = 1 case in applications, but this comes at the cost
of higher memory requirements and latency (i.e., delay in
producing output) [39].
III. VIDOSAT F RAMEWORK AND F ORMULATIONS
Prior work on adaptive sparsifying transform-based image
denoising [21], [30], [39] adapted the transform operator to
2D image patches. However, in video denoising, exploiting
the sparsity and redundancy in both the spatial and temporal dimensions typically leads to better performance than
denoising each frame separately [10]. We therefore propose an
online approach to video denoising by learning a sparsifying
transform on appropriate 3D spatio-temporal patches.
A. Video Streaming and Denoising Framework
Fig. 2 illustrates the framework of our proposed online denoising scheme for streaming videos. The frames of the noisy
video (assumed to be corrupted by additive i.i.d. Gaussian
noise) denoted as Yτ ∈ Ra×b arrive at τ = 1, 2, 3, etc.
At time τ = t, the newly arrived frame Yt is added to
a fixed-size FIFO (first in first out) buffer
queue) that
(i.e.,
t
stores a block of m consecutive frames Yi i=t−m+1 . The
oldest (leftmost) frame is dropped from the buffer at each
time instant. We denote the spatio-temporal tensor or 3D
array obtained by stacking
noisy frames along
the temporal
dimension as Yt = Yt−m+1 | ... | Yt ∈ Ra×b×m . We
denoise the noisy array Yt using the proposed VIDOSAT minibatch denoising algorithms (denoted by the red box in Fig. 2)
that are discussed in Sections III-B and IV. These algorithms
denoise groups (mini-batches) of 3D patches sequentially and
2 In practice, such normalizations may still be useful, to control the dynamic
range of various internal variables in the algorithm.
5
is first denoised in
the FIFO buffer at time 𝒕
3D Buffer
Depth = m
ഥ
+(𝒎−𝟏)
+(𝒎−𝟏)
ഥ
3D Buffer
Depth = m
ഥ
3D Patch
Depth = m
−(𝒎−𝟏)
3D Patch
Depth = m
t
m-1 future
adjacent frames
Output the denoised frame
m-1 past
adjacent frames
the mini-batch in both time t and location within the set of P
patches of Yt .
For each t, we solve the following online transform learning
problem for each k = 1, 2, 3, ..., N , to adapt the transform and
sparse codes sequentially to the mini-batches in Yt :
Lk
o
n
X
t
2
(P3) ŴLtk , X̂Ltk = arg min
ρLk −j kWUj − Xj kF
W,XLt j=1
k
at time 𝒕 + (𝒎 − 𝟏)
Fig. 3. Illustration of the output buffer from time t to t + (m − 1) for
generating the denoised frame output Ŷt .
adaptively, by learning sparsifying transforms. Overlapping
patches are used in our framework.
The patches output by the mini-batch denoising algorithms
are deposited at their corresponding spatio-temporal locations
in the fixed-size FIFO output Ȳt = Ȳt−m+1 | ... | Ȳt by
adding them to the contents of Ȳt . We call this process patch
aggregation. The streaming scheme then outputs the oldest
frame Ȳt−m+1 . The denoised estimate Ŷt−m+1 is obtained by
normalizing Ȳt−m+1 pixel-wise by the number of occurrences
of each pixel in the aggregated patches. (see Section IV for
details).
Though any frame could be denoised and output from Ȳt
instantaneously, we observe improved denoising quality by
averaging over multiple denoised estimates at different time.
Fig. 3 illustrates how the output buffer varies from time t to
t + (m − 1), to output the denoised Ŷt . In practice, we set
the length of the output buffer Ȳ to be the same as the 3D
patch depth m, such that each denoised frame Ŷt is output by
averaging over its estimates from all 3D patches that group the
tth frame with m − 1 adjacent frames. We refer to this scheme
as “two-sided” denoising, since the tth frame is denoised
together with both past and future adjacent frames (m − 1
frames on each side), which are highly correlated. Now, data
from frame Yt is contained in 3D patches that also contain
data from frame Yt+m−1 . Once these patches are denoised,
they will contribute (by aggregation into the output buffer)
to the final denoised frame Ŷt . Therefore, we must wait for
frame Yt+m−1 before producing the final estimate Ŷt . Thus
there is a delay of m − 1 frames between the arrival of the
noisy Yt and the generation of its final denoised estimate Ŷt .
B. VIDOSAT Mini-Batch Denoising Formulation
Here, we discuss the mini-batch denoising formulation
that is a core part of the proposed online video denoising
framework. For each time instant t, we denoise P partially
overlapping size n1 × n2 × m 3D
of Yt whose
patches
P
vectorized versions are denoted as vpt p=1 , with vpt ∈ Rn ,
n = mn1 n2 . We sequentially process disjoint groups of
M such patches, and the groups or mini-batches of patches
(total
of NN mini-batches, where P = M N ) are denoted
as ULtk k=1 , with ULtk ∈ Rn×M . Here, k is the local
mini-batch index within the set of P patches of Yt , whereas
Ltk , N ×(t−1)+k is the global mini-batch index, identifiying
t
+
Lk
X
j=1
(
ρ
Ltk −j
Λj ν(W) +
M
X
)
2
αj,i
xj,i
0
.
i=1
Here, the transform is adapted based on patches from all
the
observed Y
τ , 1 ≤ τ ≤ t. The matrix Xj =
xj,1 | ... | xj,M ∈ Rn×M denotes the transform sparse codes
corresponding to the mini-batch Uj . The sparsity penalty
2
weight αj,i
in (P3) controls the number of non-zeros in xj,i .
We set αj,i = α0 σj,i , where α0 is a constant and σj,i is the
noise standard deviation for each patch. We use a forgetting
t
factor ρLk −j in (P3) to diminish the influence of old frames
and old mini-batches.
Once (P3) is solved, the denoised version of the current
noisy mini-batch ÛLtk is computed. The columns of the denoised ÛLtk are tensorized and aggregated at the corresponding
spatial and temporal locations in the output FIFO buffer.
Section IV next discusses the proposed VIDOSAT algorithms
in full detail.
IV. V IDEO D ENOISING A LGORITHMS
We now discuss two video denoising algorithms, namely
VIDOSAT and VIDOSAT-BM. VIDOSAT-BM uses block
matching to generate the 3D patches from Yt . Though these
methods differ in the way they construct the 3D patches, and
the way the denoised patches are aggregated in the output
FIFO, they both denoise groups of 3D patches sequentially
by solving (P3). The VIDOSAT denoising algorithm (without
BM) is summarized in Algorithm A13 . The VIDOSAT-BM
algorithm, a modified version of Algorithm A1, is discussed
in Section IV-B.
A. VIDOSAT
As discussed in Section III-B, the VIDOSAT algorithm
processes each mini-batch Uj in Yτ sequentially. We solve
the mini-batch transform learning problem (P3) using a simple
alternating minimization approach, with one alternation per
mini-batch, which works well and saves computation. Initialized with the most recently estimated transform (warm start),
we perform two steps for (P3): Sparse Coding, and Minibatch Transform Update, which compute X̂j and update Ŵj ,
respectively. Then, we compute the denoised mini-batch Ûj ,
and aggregate the denoised patches into the output buffer Ȳτ .
The major steps of the VIDOSAT algorithm A1 for denoising the kth mini-batch ULtk at time t and further processing
these denoised patches are described below. To facilitate the
3 In practice, we wait for the first m frames to be received, before starting
Algorithm A1, to avoid zero frames in the input FIFO buffer
6
Algorithm A1: VIDOSAT Denoising Algorithm
Input: The noisy frames Yτ (τ = 1, 2, 3, etc.), and the
initial transform W0 (e.g., 3D DCT).
Initialize: Ŵ = W0 , Γ = Θ = 0, β = 0,
and output buffer Ȳ = 0.
For τ = 1, 2, 3, etc., Repeat
The newly arrived frame Yτ → latest frame in the input
FIFO frame buffer Y.
For k = 1, ..., N Repeat
Indices of patches in Y: Sk = {M (k−1)+1, ..., M k}.
1) Noisy Mini-Batch Formation:
a) PatchExtraction: v
p =vec(Rp Y) ∀p ∈ Sk .
b) U = u1 | ... | uM ← vM k−M +1 | ... | vM k .
2) Sparse Coding: x̂i = Hαi (Ŵui ) ∀i ∈ {1, ..., M }.
3) Mini-batch Transform Update:
a) Define Λ , λ0 kUk2F and X̂ , x̂1 | ... | x̂M .
b) Γ ← ρΓ + UUT .
c) Θ ← ρΘ + UX̂T .
d) β ← ρβ + Λ.
e) Matrix square root: Q ← (Γ + βI)1/2 .
T
f) Full SVD: ΦΣΨ
← SVD(Q−1Θ).
1
g) Ŵ ← 0.5Ψ Σ + Σ2 + 2βI 2 ΦT Q−1 .
4) 3D Denoised Patch Reconstruction:
a) Update Sparse Codes: x̂i = Hαi (Ŵui ) ∀i.
−1
b) Denoised mini-batch: Û
= Ŵ X̂.
c) v̂M (k−1)+1 | ... | v̂M k ← Û
d) Tensorization: V̂p = vec−1 (v̂p ) ∀p ∈ Sk .
5) Aggregation: Aggregate patches V̂p at correP
sponding locations: Ȳ ← p∈Sk Rp∗ V̂p .
End
Output: The oldest frame in Ȳ after normalization → the
denoised frame Ŷτ −m+1 .
End
adjacent stacks of frames, we reverse the raster scan order (of
patches) between Yt and Yt+1 .
2) Sparse Coding: Given the sparsifying transform W =
ŴLtk −1 estimated for the most recent mini-batch, we solve
Problem (P3) for the sparse coefficients X̂Ltk :
X̂Ltk = arg min WULtk − X
X
2
F
+
M
X
2
xi
αL
t ,i
k
0
(2)
i=1
A solution for (2) is given in closed-form as x̂Ltk ,i =
HαLt ,i (ŴLtk uLtk ,i ) ∀ i [39]. Here, the hard thresholding
k
operator Hα (·) : Rn → Rn is applied to a vector elementwise, as defined by
(
0 , |dr | < α
(3)
(Hα (d))r =
dr , |dr | ≥ α
This simple hard thresholding operation for transform sparse
coding is similar to traditional techniques involving analytical
sparsifying transforms [45].
3) Mini-batch Transform Update: We solve Problem (P3)
for W with fixed Xj = X̂j , 1 ≤ j ≤ Ltk , as follows:
t
min
W
Lk
X
t
ρLk −j
n
WUj − Xj
2
F
o
+ Λj ν(W) .
(4)
j=1
This problem has a simple solution (similar to Section III-B2
in [39]). P
Set index J = Lkt , and define P
the following quantities:
J
J
J−j
J−j
ΓJ ,
Uj UTj , ΘJ ,
Uj X̂Tj , and
j=1 ρ
j=1 ρ
PJ
J−j
n×n
βJ , j=1 ρ
Λj . Let Q ∈ R
be a square root (e.g.,
Cholesky factor) of (ΓJ +βJ I), i.e., QQT = ΓJ +βJ I. Denoting the full singular value decomposition (SVD) of Q−1 ΘJ
as ΦΣΨT , we then have that the closed-form solution to (4)
is
1
ŴJ = 0.5Ψ Σ + Σ2 + 2βJ I 2 ΦT Q−1
(5)
1
exposition and interpretation in terms of the general online
denoising algorithm described, various quantities (such as
positions of 3D patches in the video stream) are indexed in
the text with respect to absolute time t. On the other hand,
to emphasize the streaming nature of Algorithm A1 and its
finite (and modest) memory requirements, indexing of internal
variables in the statement of the algorithm is local.
1) Noisy Mini-Batch Formation: To construct each minibatch ULtk , partially overlapping size n1 × n2 × m 3D patches
of Yt are extracted sequentially in a spatially contiguous order
(raster scan order with direction reversal on each line)4 . Let
Rp Yt denote the pth vectorized 3D patch of Yt , with Rp
being the patch-extraction
operator. Considering the patch
indices Sk = M (k − 1) + 1, ..., M k for the kth minibatch, we extract vpt = vec(Rp Yt ) p∈S as the patches in the
k
h
i
t
t
|
...
|
v
mini-batch. Thus ULtk = vM
. To impose
M
k
(k−1)+1
spatio-temporal contiguity of 3D patches extracted from two
4 We did not observe any marked improvement in denoising performance,
when using other scan orders such as raster or Peano-Hilbert scan [44].
where I denotes the identity matrix, and (·) 2 denotes the
positive definite square root of a positive definite (diagonal)
matrix. The quantities ΓJ , ΘJ , and βJ are all computed
sequentially over time t and mini-batches k [39].
4) 3D Denoised Patch Reconstruction: We denoise ULtk
using the updated transform. First, we repeat the sparse coding
step using the updated ŴLtk as x̂Ltk ,i = HαLt ,i (ŴLtk uLtk ,i ) ∀
k
i. Then, with fixed ŴLtk and X̂Ltk , the denoised mini-batch is
obtained in the least squares sense under the transform model
as
−1
ÛLtk = ŴL
(6)
t X̂Lt .
k
k
The denoised mini-batch is used to update the denoised
t
t
(vectorized) 3D patches as v̂M
(k−1)+i = ûLk ,i ∀i. All recon t
structed vectors v̂p p∈S from the kth mini-batch denoising
k
result are tensorized as vec−1 (v̂pt ) p∈S .
k
Aggregation:
The
denoised
3D
patches
5) −1
vec (v̂pt ) p∈S from each mini-batch are sequentially
k
aggregated
at
their
corresponding
spatial
and
temporal
locations
in
the
output
FIFO
buffer
as
P
∗
−1
t
a×b×m
R
vec
(v̂
)
→
Ȳ
∈
R
,
where
the
adjoint
t
p
p∈Sk p
7
Direct
Extraction
෩ in
𝑹𝒑 Y
t
VIDOSAT
Yt
Yt 1
Yt 1
Block
Matching
෩ in
𝑩𝒑 Y
t
VIDOSAT-BM
Patch
Reference
Search
Window
Fig. 4.
Illustration of the different 3D patch construction methods in
VIDOSAT (blue) and VIDOSAT-BM (red). The 3D search window used in
VIDOSAT-BM is illustrated in green.
or
in A1
in A2
Vectorization
Reference
Patch
Ascending order
of Euclidean distance in A2
Tensorization
Patch deposit
in A1,
or
in A2.
Fig. 5. Patch deposit Rp∗ vec−1 (v̂p ) (resp. Bp∗ vec−1 (v̂p )) as an adjoint
of patch extraction operator in A1 (resp. an adjoint of BM operator in A2).
Rp∗ is the patch deposit operator. Fig. 5 illustrates the patch
deposit procedure for aggregation.
When all N denoised mini-batches for Yt are generated,
and the patch aggregation in Ȳt completes, the oldest frame
in Ȳt is normalized pixel-wise by the number of occurrences
(which ranges from 2m − 1, for pixels at the corners of a
video frame, to n for pixels away from the borders of a video
frame) of that pixel among patches aggregated into the output
buffer. This normalized result is output as the denoised frame
Ŷt−m+1 .
B. VIDOSAT-BM
For videos with relatively static scenes, each extracted
spatio-temporal tensor Rp Yt in the VIDOSAT Algorithm A1
typically has high temporal correlation, implying high (3D)
transform domain sparsity. However, highly dynamic videos
usually involve various motions, such as translation, rotation,
scaling, etc. Figure 4 demonstrates one example when the
3D patch construction strategy in the VIDOSAT denoising
algorithm A1 fails to capture the properties of the moving
object. Thus, Algorithm A1 could provide sub-optimal denoising performance for highly dynamic videos. We propose an
alternative algorithm, dubbed VIDOSAT-BM, which improves
VIDOSAT denoising by constructing 3D patches using block
matching.
The proposed VIDOSAT-BM solves the online transform
learning problem (P3) with a different methodology for constructing the 3D patches and each mini-batch. The Steps
(2) − (4) in Algorithm A1 remain the same for VIDOSATBM. We now discuss the modified Steps (1) and (5) in the
VIDOSAT-BM denoising algorithm, to which we also refer as
Algorithm A2.
3D Patch and Mini-Batch Formation in VIDOSATBM: Here, we use a small and odd-valued sliding (temporal)
window size m (e.g., we set m = 9 in the video denoising
experiments in Section V, which corresponds to ∼ 0.2s buffer
duration for a video with 40 Hz frame rate). Within the
m-frame input FIFO buffer Yt , we approximate the various
motions in the video using simple (local) translations [46].
We consider the middle frame Yt−(m−1)/2 in the input
FIFO buffer Yt , and sequentially extract all 2D overlapping
patches Ztp ∈ Rn1 ×n2 , 1 ≤ p ≤ P in Yt−(m−1)/2 , in a 2D
spatially contiguous (raster scan) order. For each Ztp , we form
a h1 ×h2 ×m pixel local search window centered at the center
of Ztp (see the illustration in Fig. 4). We apply a spatial BM
operator, denoted Bp , to find (using exhaustive search) the
(m − 1) patches, one for each neighboring frame in the search
window, that are most similar to Ztp in Euclidean distance.
The operator Bp stacks the Ztp , followed by the (m − 1)
matched patches, in an ascending order of their Euclidean
distance to Ztp , to form the pth 3D patch Bp Yt ∈ Rn1 ×n2 ×m .
Similar BM approaches have been used in prior works on
video compression (e.g., MPEG) for motion compensation
[46], and in recent works on spatiotemporal medical imaging
[1]. The coordinates of all selected 2D patches are recorded
to be used later in the denoised patch aggregation step.
Instead of constructing the 3D patches from 2D patches in
corresponding locations in contiguous frames (i.e., Rp Yt in
Algorithm A1), we form the patches using BM and work
with the vectorized vpt = vec(Bp Yt ) ∈ Rn in VIDOSATBM. The
as in Algorithm A1 as
h k-th mini-batch is defined
i
t
t
ULtk = vM (k−1)+1 | ... | vM k .
Each denoised 3D patch (tensor) of
Aggregation:
vec−1 (v̂pt ) p∈S contains the matched (and denoised) 2D
k
patches. They are are sequentially aggregated at their recorded
spatial
and temporal locations in the output FIFO buffer Ȳt as
P
∗
−1
(v̂pt ) → Ȳt ∈ Ra×b×m , where the adjoint
p∈Sk Bp vec
∗
Bp is the patch deposit operator in A2. Fig. 5 illustrates
the patch deposit procedure for aggregation in A2. Once
the aggregation of Ȳt completes, the oldest frame in Ȳt
is normalized pixel-wise by the number of occurrences of
each pixel among patches in the denoising algorithm. Unlike
Algorithm A1 where this number of occurrences is the same
for all frames, in Algorithm A2 this number is data-dependent
and varies from frame to frame and pixel to pixel. We record
the number of occurrences of each pixel which is based on
the recorded locations of the matched patches, and can be
computed online as described. The normalized oldest frame is
output by Algorithm A2 for each time instant.
C. Computational Costs
In Algorithm A1, the computational cost of the sparse
coding step is dominated by the computation of matrix-vector
multiplication Ŵui , which scales as O(M n2 ) [39], [43] for
each mini-batch. The cost of mini-batch transform update step
is O(n3 +M n2 ), which is dominated by full SVD and matrixmatrix multiplications. The cost of the 3D denoised patch
reconstruction step also scales as O(n3 +M n2 ) per mini-batch,
8
σ
fBM3D
∆PSNR
ASU Dataset (26 videos)
Data
5
10
15
20
50
38.78
34.66
32.38
30.82
26.13
5
10
15
20
50
38.05
34.06
31.89
30.42
25.88
3.89
[5]
sKSVD
(std.)
∆PSNR
LASIP Dataset (8 videos)
2.11
(1.41)
(1.03)
1.20
41.27
37.37
35.15
33.59
28.79
1.21
38.87
34.95
32.80
31.33
26.89
(0.34)
[10]
(0.38)
1.69
3D DCT
VBM3D
41.26
37.14
34.73
33.03
27.59
33.60
30.44
28.50
22.31
(1.28)
0.92
37.82
35.78
34.25
28.65
[11]
VBM4D
3.60
38.01
(0.78)
41.10
0.61
39.20
35.75
33.87
32.49
26.51
(0.72)
(0.51)
1.30
41.42
37.59
35.30
33.64
27.76
[12]
0.63
39.37
35.73
33.70
32.24
26.68
(0.86)
(0.49)
0.27
VIDOSAT
VIDOSAT
41.94
38.32
36.13
34.60
29.87
0.55
39.56
35.75
33.54
31.98
27.29
(0.13)
42.22
38.57
36.42
34.88
30.09
0
(std.)
(0.29)
39.95
36.11
34.05
32.60
28.15
0
-BM
TABLE III
C OMPARISON OF VIDEO DENOISING PSNR VALUES ( IN D B), AVERAGED OVER THE ASU DATASET ( LEFT ) AND THE LASIP DATASET ( RIGHT ), FOR THE
PROPOSED VIDOSAT, VIDOSAT-BM, AND OTHER COMPETING METHODS . F OR EACH DATASET AND NOISE LEVEL , THE BEST DENOISING PSNR IS
MARKED IN BOLD . F OR EACH METHOD , WE LIST ∆ PSNR, WHICH DENOTES THE AVERAGE PSNR DIFFERENCE ( WITH ITS STANDARD DEVIATION
INCLUDED IN PARENTHESES ) RELATIVE TO THE PROPOSED VIDOSAT-BM ( HIGHLIGHTED IN BOLD ).
which is dominated by the computation of matrix inverse
Ŵ−1 and multiplications. As all overlapping patches from a
a × b × T video are sequentially processed, the computational
cost of Algorithm A1 scales as O(abT n3 /M + abT n2 ). We
set M = 15n in practice, so that the cost of A1 scales
as O(abT n2 ). The cost of the additional BM step in Algorithm A2 scales as O(abT mh1 h2 ), where h1 × h2 is the
search window size. Therefore, the total cost of A2 scales as
O(abT n2 + abT mh1 h2 ), which is on par with the state-ofthe-art video denoising algorithm VBM3D [11], which is not
an online method.
V. E XPERIMENTS
A. Implementation and Parameters
1) Testing Data: We present experimental results demonstrating the promise of the proposed VIDOSAT and
VIDOSAT-BM online video denoising methods. We evaluate
the proposed algorithms by denoising all 34 videos from 2
public datasets, including 8 videos from the LASIP video
dataset 5 [11], [12], and 26 videos the Arizona State University
(ASU) Video Trace Library 6 [47]. The testing videos contain
50 to 870 frames, with the frame resolution ranging from
176 × 144 to 720 × 576. Each video involves different types of
motion, including translation, rotation, scaling (zooming), etc.
The color videos are all converted to gray-scale. We simulated
5 Available
at http://www.cs.tut.fi/∼lasip/foi wwwstorage/test videos.zip
at http://trace.eas.asu.edu/yuv/. Only videos with less than 1000
frames are selected for our image denoising experiments.
6 Available
i.i.d. zero-mean Gaussian noise at 5 different noise levels (with
standard deviation σ = 5, 10, 15, 20, and 50) for each video.
2) Implementation Details: We include several minor modifications of VIDOSAT and VIDOSAT-BM algorithms for
improved performance. At each time instant t, we perform
multiple passes of denoising for each Yt , by iterating over
Steps (1) to (5) multiple times. In each pass, we denoise
the output from the previous iteration [21], [39]. As the
sparsity penalty weights are set proportional to the noise level,
αj,i = α0 σ, the noise standard deviation σ in each such pass
is set to an empirical estimate [21], [43] of the remaining noise
in the denoised frames from the previous pass. These multiple
passes, although increasing the computation in the algorithm,
do not increase the inherent latency m − 1 of the single pass
algorithm described earlier.
The following details are specifically for VIDOSAT-BM.
First, instead of performing BM over the noisy input buffer
Yt , we pre-clean Yt using the VIDOSAT mini-batch denoising
Algorithm A1, and then perform BM over the VIDOSAT
denoised output. Second, when denoised 3D patches are aggregated to the output buffer, we assign them different weights,
which are proportional to the sparsity level of their optimal
sparse codes [48]. The weights are also accumulated and used
for the output normalization.
3) Hyperparameters: We work with fully overlapping
patches (spatial patch stride of 1 pixel) with spatial size
n1 = n2 = 8, and temporal depth of m = 9 frames, which
also corresponds to the depth of buffer Y. It follows that for
a video with N1 × N2 frames, the buffer Y contains mN1 N2
pixels, and P = (N1 − n1 + 1)(N2 − n2 + 1) 3D patches.
9
(a) Noisy
(b) Original
(c) VBM3D (33.30 dB)
(e) VIDOSAT (35.84 dB)
(g) VIDOSAT-BM (36.11 dB)
40
40
40
20
20
20
(d)
(f)
(h)
Fig. 6. (a) The noisy version (σ = 50) of (b) one frame of the Akiyo (288 × 352 × 300) video. We show the comparison of the denoising results (resp.
the magnitude of error in the denoised frame) using (c) VBM3D (33.30 dB), (e) VIDOSAT (35.84 dB) and (g) VIDOSAT-BM (36.11 dB) (resp. (d), (f) and
(h)). The PSNR of the denoised frame is shown in the parentheses. The zoom-in region is highlighted using red box.
(a) Noisy
(b) Original
(c) VBM4D (33.04 dB)
(e) VIDOSAT (33.43 dB)
(g) VIDOSAT-BM (34.01 dB)
40
40
40
20
20
20
(d)
(f)
(h)
Fig. 7. (a) The noisy version (σ = 20) of (b) one frame of the Salesman (288 × 352 × 50) video. We show the comparison of the denoising results (resp.
the magnitude of error in the denoised frame) using (c) VBM4D (33.04 dB), (e) VIDOSAT (33.43 dB) and (g) VIDOSAT-BM (34.01 dB) (resp. (d), (f) and
(h)). The PSNR of the denoised frame is shown in the parentheses. The zoom-in regions are highlighted using red and green boxes.
We set the sparsity penalty weight parameter α0 = 1.9, the
transform regularizer weight constant λ0 = 10−2 , and the
mini-batch size M = 15 × mn1 n2 . The transform W is
initialized with the 3D DCT W0 . For the other parameters,
we adopt the settings in prior works [21], [39], [43], such as
the forgetting factor ρ = 0.68, 0.72, 0.76, 0.83, 0.89, and the
number of passes Lp = 1, 2, 3, 3, 4 for σ = 5, 10, 15, 20, 50,
respectively. The values of ρ and Lp both increase as the
noise level increases. The larger ρ helps prevent overfitting
to noise, and the larger number of pass improves denoising
performance at higher noise level. For VIDOSAT-BM, we set
the local search window size h1 = h2 = 21.
B. Video Denoising Results
1) Competing Methods: We compare the video denoising
results obtained using the proposed VIDOSAT and VIDOSATBM algorithms to several well-known alternatives, including the frame-wise BM3D denoising method (fBM3D) [5],
the image sequence denoising method using sparse KSVD
(sKSVD) [10], VBM3D [11] and VBM4D methods [12].
We used the publicly available implementations of these
methods. Among these competing methods, fBM3D denoises
each frame independently by applying a popular BM3D image
denoising method; sKSVD exploits adaptive spatio-temporal
sparsity but the dictionary is not learned online; and VBM3D
and VBM4D are popular and state-of-the-art video denoising
methods. Moreover, to better understand the advantages of the
online high-dimensional transform learning, we apply the pro-
10
(a) Noisy
(c) VBM4D (34.00 dB)
(b) Original
(e) VIDOSAT (32.07 dB)
(g) VIDOSAT-BM (35.33 dB)
40
40
40
20
20
20
(d)
(f)
(h)
Fig. 8. (a) The noisy version (σ = 20) of (b) one frame of the Bicycle (576 × 720 × 30) video. We show the comparison of the denoising results (resp.
the magnitude of error in the denoised frame) using (c) VBM4D (34.00 dB), (e) VIDOSAT (32.07 dB) and (g) VIDOSAT-BM (35.33 dB) (resp. (d), (f) and
(h)). The PSNR of the denoised frame is shown in the parentheses. The zoom-in region is highlighted using red box.
37
35
35
33
Denoised PSNR
Denoised PSNR
Denoised PSNR
35
34
33
34
33
VBM3D
VBM4D
VIDOSAT
VIDOSAT-BM
32
31
31
32
50
100
150
200
Frame Number
250
300
10
20
30
Frame Number
(a)
(b)
40
50
30
5
10
15
20
25
30
Frame Number
(c)
Fig. 9. Frame-by-frame PSNR (dB) for (a) Akiyo with σ = 50, (b) Salesman with σ = 20, and (c) Bicycle with σ = 20, denoised by VBM3D, VBM4D,
and the proposed VIDOSAT and VIDOSAT-BM schemes, respectively.
posed video denoising framework, but fixing the sparsifying
transform in VIDOSAT to 3D DCT, which is referred as the
3D DCT method.
2) Denoising Results: We present video denoising results
using the proposed VIDOSAT and VIDOSAT-BM algorithms,
as well as using the other aforementioned competing methods.
To evaluate the performance of the various denoising schemes,
we measure the peak signal-to-noise ratio (PSNR) in decibels
(dB), which is computed between the noiseless reference and
the denoised video.
Table III lists the video denoising PSNRs obtained by
the two proposed VIDOSAT methods, as well as the five
competing methods. It is clear that the proposed VIDOSAT
and VIDOSAT-BM approaches both generate better denoising
results with higher average PSNR values, compared to the
competing methods. The VIDOSAT-BM denoising method
provides average PSNR improvements (averaged over all 34
testing videos from both datasets and all noise levels) of 0.9
dB, 1.1 dB, 1.2 dB, 2.1 dB, and 3.5 dB, over the VBM3D,
VBM4D, sKSVD, 3D DCT, and fBM3D denoising methods.
Importantly, VIDOSAT-BM consistently outperforms all the
competing methods for all testing videos and noise levels.
Among the two proposed VIDOSAT algorithms, the average
video denoising PSNR by VIDOSAT-BM is 0.3 dB higher
than that using the VIDOSAT method, thanks to the use of the
block matching for modeling dynamics and motion in video.
We illustrate the denoising results and improvements provided by VIDOSAT and VIDOSAT-BM with some examples.
a) Fig. 6: shows one denoised frame of the video Akiyo
(σ = 50), which involves static background and a relatively
small moving region (The magnitudes of error in Fig. 6 are
clipped for viewing). The denoising results by VIDOSAT
and VIDOSAT-BM both demonstrate similar visual quality
improvements over the result by VBM3D. Fig. 9(a) shows
the frame-by-frame PSNRs of the denoised Akiyo, in which
VIDOSAT and VIDOSAT-BM provide comparable denoising PSNRs, and both outperform the VBM3D and VBM4D
schemes consistently by a sizable margin.
b) Fig. 7: shows one denoised frame of the video Salesman (σ = 20) that involves occasional but fast movements
(e.g., hand waving) in the foreground. The denoising result
by VIDOSAT improves over the VBM4D result in general,
but also shows some artifacts in regions with strong motion.
Instead, the result by VIDOSAT-BM provides the best visual
11
Initial
DCT
Learned W
t = 10
Initial
DCT
Learned W
t = 10
Learned W
t = 20
Learned W
t = 30
Learned W
t = 40
(a) Ŵt learned using VIDOSAT
Learned W
t = 20
Learned W
t = 30
Learned W
t = 40
(b) Ŵt learned using VIDOSAT-BM
Fig. 10. Example atoms (i.e., 4 rows) of the initial 3D DCT (with depth
m = 9), and the online learned 3D sparsifying transform using (a) VIDOSAT,
and (b) VIDOSAT-BM, at times 10 to 40: the atoms (i.e., rows) of the learned
Ŵ are shown as m = 9 patches in each column. These 9 patches together
form the 8 × 8 × 9 3D atoms.
quality in both the static and the moving parts. Fig. 9(b)
shows the frame-by-frame PSNRs of the denoised Salesman.
VIDOSAT-BM provides large improvements over the other
methods including VIDOSAT for most frames, and the PSNR
is more stable (smaller deviations) over frames. Fig. 10 shows
example atoms (i.e., rows) of the initial 3D DCT transform,
and the online learned transforms using (a) VIDOSAT and (b)
VIDOSAT-BM at different times t. For the learned Ŵt ’s using
both VIDOSAT and VIDOSAT-BM, their atoms are observed
to gradually evolve, in order to adapt to the dynamic video
content. The learned transform atoms using VIDOSAT in Fig.
10(a) demonstrate linear shifting structure along the patch
depth m, which is likely to compensate the video motion
(e.g., translation). On the other hand, since the 3D patches
are formed using BM in VIDOSAT-BM, such structure is not
observed in Fig. 10(b) when Ŵt is learned using VIDOSATBM.
c) Fig. 8: shows one denoised frame of the video Bicycle
(σ = 20), which contains a large area of complex movements
(e.g., rotations) throughout the video. In this case, the denoised
frame using the VIDOSAT is worse than VBM4D. However,
VIDOSAT-BM provides superior quality compared to all the
methods. This example demonstrates the effectiveness of joint
block matching and learning in the proposed VIDOSAT-BM
scheme, especially when processing highly dynamic videos.
Fig. 9(c) shows the frame-by-frame PSNRs of the denoised
Bicycle, in which VIDOSAT-BM significantly improves over
VIDOSAT, and also outperforms both VBM3D and VBM4D
for all frames.
VI. C ONCLUSIONS
We presented a novel framework for online video denoising based on efficient high-dimensional sparsifying transform
learning. The transforms are learned in an online manner
from spatio-temporal patches. These patches are constructed
either from corresponding 2D patches of consecutive frames
or using an online block matching technique. The learned
models effectively capture the dynamic changes in videos.
We demonstrated the promising performance of the proposed
video denoising schemes for several standard datasets. Our
methods outperformed all compared methods, which included
a version of our the proposed video denoising scheme in
which the learning of the sparsifying transform was eliminated
and instead it was fixed to 3D DCT, as well as denoising
using learned synthesis dictionaries, and the state-of-the-art
VBM3D and VBM4D methods. While this work provides an
initial study of the promise of the proposed data-driven online
video denoising methodologies, we plan to study the potential
implementation and acceleration of the proposed schemes for
real-time video processing in future work.
R EFERENCES
[1] H. Yoon, K. S. Kim, D. Kim, Y. Bresler, and J. C. Ye, “Motion adaptive
patch-based low-rank approach for compressed sensing cardiac cine
mri,” IEEE transactions on medical imaging, vol. 33, no. 11, pp. 2069–
2085, 2014.
[2] K. Kim, Y. D. Son, Y. Bresler, Z. H. Cho, J. B. Ra, and J. C.
Ye, “Dynamic pet reconstruction using temporal patch-based low rank
penalty for roi-based brain kinetic analysis,” Physics in medicine and
biology, vol. 60, no. 5, p. 2019, 2015.
[3] J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman, “Non-local
sparse models for image restoration,” in IEEE 12th Int. Conf. Comput.
Vision (ICCV 2009), Sept 2009, pp. 2272–2279.
[4] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol,
“Stacked denoising autoencoders: Learning useful representations in a
deep network with a local denoising criterion,” Journal of Machine
Learning Research, vol. 11, no. Dec, pp. 3371–3408, 2010.
[5] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by
sparse 3-D transform-domain collaborative filtering,” IEEE Transactions
on Image Processing, vol. 16, no. 8, pp. 2080–2095, 2007.
[6] R. Otazo, E. Candès, and D. K. Sodickson, “Low-rank plus sparse
matrix decomposition for accelerated dynamic MRI with separation
of background and dynamic components,” Magnetic Resonance in
Medicine, vol. 73, no. 3, pp. 1125–1136, 2015.
[7] S. Ravishankar, B. Moore, R. Nadakuditi, and J. Fessler, “Low-rank and
adaptive sparse signal (LASSI) models for highly accelerated dynamic
imaging,” IEEE Transactions on Medical Imaging, 2017, to appear.
[8] D. Rusanovskyy and K. Egiazarian, “Video denoising algorithm in
sliding 3D dct domain,” in Proc. Advanced Concepts for Intelligent
Vision Systems, 2005, pp. 618–625.
[9] N. Rajpoot, Z. Yao, and R. Wilson, “Adaptive wavelet restoration of
noisy video sequences,” in Proc. IEEE Int. Conf. Image Proc., ICIP,
vol. 2, 2004, pp. 957–960.
[10] R. Rubinstein, M. Zibulevsky, and M. Elad, “Double sparsity: Learning
sparse dictionaries for sparse signal approximation,” IEEE Transactions
on Signal Processing, vol. 58, no. 3, pp. 1553–1564, 2010.
[11] K. Dabov, A. Foi, and K. Egiazarian, “Video denoising by sparse 3D
transform-domain collaborative filtering,” in Signal Processing Conference, 2007 15th European, Sept 2007, pp. 145–149.
[12] M. Maggioni, G. Boracchi, A. Foi, and K. Egiazarian, “Video
denoising, deblocking, and enhancement through separable 4-d
nonlocal spatiotemporal transforms,” IEEE Transactions on Image
Processing, vol. 21, no. 9, pp. 3952–3966, 2012. [Online]. Available:
http://dx.doi.org/10.1109/TIP.2012.2199324
[13] M. Elad and M. Aharon, “Image denoising via sparse and redundant
representations over learned dictionaries,” IEEE Transactions on Image
Processing, vol. 15, no. 12, pp. 3736–45, Dec. 2006.
[14] J. Mairal, M. Elad, and G. Sapiro, “Sparse representation for color image
restoration,” IEEE Trans. on Image Processing, vol. 17, no. 1, pp. 53–69,
2008.
[15] M. Protter and M. Elad, “Image sequence denoising via sparse and
redundant representations,” IEEE Transactions on Image Processing,
vol. 18, no. 1, pp. 27–35, 2009.
12
[16] I. Ramirez, P. Sprechmann, and G. Sapiro, “Classification and clustering
via dictionary learning with structured incoherence and shared features,”
in Proc. IEEE International Conference on Computer Vision and Pattern
Recognition (CVPR) 2010, 2010, pp. 3501–3508.
[17] S. Ravishankar and Y. Bresler, “MR image reconstruction from highly
undersampled k-space data by dictionary learning,” IEEE Trans. Med.
Imag., vol. 30, no. 5, pp. 1028–1041, 2011.
[18] R. Rubinstein, T. Peleg, and M. Elad, “Analysis K-SVD: A dictionarylearning algorithm for the analysis sparse model,” IEEE Transactions
on Signal Processing, vol. 61, no. 3, pp. 661–677, 2013.
[19] S. Ravishankar and Y. Bresler, “Learning sparsifying transforms,” IEEE
Transactions on Signal Processing, vol. 61, no. 5, pp. 1072–1086, 2013.
[20] ——, “Learning doubly sparse transforms for images,” IEEE Trans.
Image Process., vol. 22, no. 12, pp. 4598–4612, 2013.
[21] B. Wen, S. Ravishankar, and Y. Bresler, “Structured overcomplete sparsifying transform learning with convergence guarantees and applications,”
International Journal of Computer Vision, vol. 114, no. 2-3, pp. 137–
167, 2015.
[22] J.-F. Cai, H. Ji, Z. Shen, and G.-B. Ye, “Data-driven tight frame construction and image denoising,” Applied and Computational Harmonic
Analysis, vol. 37, no. 1, pp. 89 – 105, 2014.
[23] Z. Zhan, J.-F. Cai, D. Guo, Y. Liu, Z. Chen, and X. Qu, “Fast multiclass
dictionaries learning with geometrical directions in mri reconstruction,”
IEEE Transactions on Biomedical Engineering, vol. 63, no. 9, pp. 1850–
1861, 2016.
[24] X. Zheng, Z. Lu, S. Ravishankar, Y. Long, and J. A. Fessier, “Low dose
CT image reconstruction with learned sparsifying transform,” in Proc.
IEEE Wkshp. on Image, Video, Multidim. Signal Proc., Jul. 2016, pp.
1–5.
[25] M. Aharon, M. Elad, and A. Bruckstein, “K-SVD : An algorithm for
designing overcomplete dictionaries for sparse representation,” IEEE
Transactions on Signal Processing, vol. 54, no. 11, pp. 4311–4322, 2006.
[26] Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad, “Orthogonal matching
pursuit : recursive function approximation with applications to wavelet
decomposition,” in Asilomar Conf. on Signals, Systems and Comput.,
1993, pp. 40–44 vol.1.
[27] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani, “Least angle
regression,” Annals of Statistics, vol. 32, pp. 407–499, 2004.
[28] D. Needell and J. Tropp, “Cosamp: Iterative signal recovery from incomplete and inaccurate samples,” Applied and Computational Harmonic
Analysis, vol. 26, no. 3, pp. 301–321, 2009.
[29] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sensing
signal reconstruction,” IEEE Trans. Information Theory, vol. 55, no. 5,
pp. 2230–2249, 2009.
[30] S. Ravishankar and Y. Bresler, “Sparsifying transform learning with efficient optimal updates and convergence guarantees,” IEEE Transactions
on Signal Processing, vol. 63, no. 9, pp. 2389–2404, 2015.
[31] B. Wen, S. Ravishankar, and Y. Bresler, “Learning flipping and rotation
invariant sparsifying transforms,” in Image Processing (ICIP), 2016
IEEE International Conference on. IEEE, 2016, pp. 3857–3861.
[32] S. Ravishankar and Y. Bresler, “Efficient blind compressed sensing using
sparsifying transforms with convergence guarantees and application
to magnetic resonance imaging,” SIAM Journal on Imaging Sciences,
vol. 8, no. 4, pp. 2519–2557, 2015.
[33] L. Pfister and Y. Bresler, “Model-based iterative tomographic reconstruction with adaptive sparsifying transforms,” in Proc. SPIE Computational
Imaging XII, C. A. Bouman and K. D. Sauer, Eds. SPIE, Mar. 2014,
pp. 90 200H–90 200H–11.
[34] ——, “Tomographic reconstruction with adaptive sparsifying transforms,” in Acoustics, Speech and Signal Processing (ICASSP), 2014
IEEE International Conference on. IEEE, 2014, pp. 6914–6918.
[35] ——, “Learning sparsifying filter banks,” in Proc. SPIE Wavelets &
Sparsity XVI, vol. 9597. SPIE, Aug. 2015.
[36] S. Dev, B. Wen, Y. H. Lee, and S. Winkler, “Ground-based image
analysis: A tutorial on machine-learning techniques and applications,”
IEEE Geoscience and Remote Sensing Magazine, vol. 4, no. 2, pp. 79–
93, 2016.
[37] S. Ravishankar and Y. Bresler, “Data-driven learning of a union of
sparsifying transforms model for blind compressed sensing,” IEEE
Transactions on Computational Imaging, vol. 2, no. 3, pp. 294–309,
2016.
[38] X. Zheng, S. Ravishankar, Y. Long, and J. A. Fessler, “Union of learned
sparsifying transforms-based low-dose 3D CT image reconstruction,” in
Proc. Intl. Mtg. on Fully 3D Image Recon. in Rad. and Nuc. Med, 2017,
to appear.
[39] S. Ravishankar, B. Wen, and Y. Bresler, “Online sparsifying transform
learning - part I: Algorithms,” IEEE Journal of Selected Topics in Signal
Processing, vol. 9, no. 4, pp. 625–636, 2015.
[40] S. Ravishankar and Y. Bresler, “Online sparsifying transform learning
- part II: Convergence analysis,” IEEE Journal of Selected Topics in
Signal Processing, vol. 9, no. 4, pp. 637–646, 2015.
[41] J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online learning for matrix
factorization and sparse coding,” Journal of Machine Learning Research,
vol. 11, no. Jan, pp. 19–60, 2010.
[42] B. Wen, Y. Li, L. Pfister, and Y. Bresler, “Joint adaptive sparsity and
low-rankness on the fly: an online tensor reconstruction scheme for
video denoising,” in IEEE International Conference on Computer Vision
(ICCV), 2017.
[43] B. Wen, S. Ravishankar, and Y. Bresler, “Video denoising by online 3D
sparsifying transform learning,” in IEEE International Conference on
Image Processing (ICIP), 2015, pp. 118–122.
[44] T. Ouni and M. Abid, “Scan methods and their application in image
compression,” International Journal of Signal Processing, Image Processing and Pattern Recognition, vol. 5, no. 3, pp. 49–64, 2012.
[45] S. Mallat, A wavelet tour of signal processing. Academic press, 1999.
[46] D. Le Gall, “Mpeg: A video compression standard for multimedia
applications,” Communications of the ACM, vol. 34, no. 4, pp. 46–58,
1991.
[47] P. Seeling and M. Reisslein, “Video traffic characteristics of modern
encoding standards: H. 264/avc with svc and mvc extensions and h.
265/hevc,” The Scientific World Journal, vol. 2014, 2014.
[48] B. Wen, Y. Li, and Y. Bresler, “When sparsity meets low-rankness:
transform learning with non-local low-rank constraint for image restoration,” in IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), 2017, pp. 2297–2301.
| 1cs.CV
|
arXiv:1404.5611v1 [cs.CE] 22 Apr 2014
IMP Science Gateway: from the Portal to the Hub of
Virtual Experimental Labs in Materials Science
Yuri Gordienko*, Lev Bekenev, Olexandra Baskova,
Olexander Gatsenko, Elena Zasimchuk
Sergii Stirenko
G.V.Kurdyumov Institute for Metal Physics,
National Academy of Sciences
Kyiv, Ukraine
Email: *[email protected]
High Performance Computing Center,
National Technical University of Ukraine ”KPI”,
Kyiv, Ukraine
Email: [email protected]
Abstract—The user communities outside of computer science
itself (for example, in materials science, physics, chemistry, biology, nanotechnologies, etc.) would like to access different software
components and various distributed computing infrastructures
(DCIs) (like grids, clouds, clusters) in a such transparent way, that
they do not need to learn their peculiarities. In this context, ”science gateway” (SG) ideology means a user-friendly intuitive interface between scientists (or scientific communities) and different
software components + various DCIs, where researchers can focus
on their scientific goals and less on peculiarities of software/DCI.
”IMP Science Gateway Portal” (http://scigate.imp.kiev.ua) for
complex workflow management and integration of distributed
computing resources (like clusters, service grids, desktop grids,
clouds) is presented. It is created on the basis of WS-PGRADE
and gUSE technologies, where WS-PGRADE is designed for
science workflow operation and gUSE - for smooth integration
of available resources for parallel and distributed computing
in various heterogeneous distributed computing infrastructures
(DCI). The typical scientific workflows with possible scenarios of
its preparation and usage are presented. Several typical use cases
for these science applications (scientific workflows) are considered
for molecular dynamics (MD) simulations of complex behavior
of various nanostructures (nanoindentation of graphene layers,
defect system relaxation in metal nanocrystals, thermal stability
of boron nitride nanotubes, etc.). The user experience is analyzed
in the context of its practical applications for MD simulations in
materials science, physics and nanotechnologies with available
heterogeneous DCIs. The most crucial advantages are as follows:
powerful and flexible workflow construction tools, which allow
users to apply LEGO-style for creation of complex tunable chains
with de facto standard components (like R, LAMMPS, AtomEye,
ffmpeg) and customized components (like debyer, Pizza, etc.);
smooth access to heterogeneous DCI and software; division of
user roles (administrators: operators of portals; power users:
principal scientists; end users: scientists, students); user friendly
and intuitively understandable GUI (with additional modules, ad
hoc changes, etc.); extensive manuals and permanently working
fora and helpdesks; short learning curve for usual scientists
without extensive knowledge in computer science. Some noncritical drawbacks of the approach include: scarce ”stdout”and
”stderr” information, tacit context help for beginners, some
unavoidable upgrade issues, etc. In conclusion, the ”science
gateway” approach - workflow manager (like WS-PGRADE) +
DCI resources manager (like gUSE)- gives opportunity to use
the SG portal (like ”IMP Science Gateway Portal”) in a very
promising way, namely, as a hub of various virtual experimental
labs (different software components + various requirements to
resources) in the context of its practical MD applications in materials science, physics, chemistry, biology, and nanotechnologies.
Keywords—Distributed computing infrastructure (DCI), grid
computing, cluster, service grid, desktop grid, science gateway,
computational physics, molecular dynamics, materials science,
physics, nanotechnologies.
I.
I NTRODUCTION
Recently research boom in nanotechnologies has started to
search and investigate new nanoscale materials, like nanocrystals, graphene, carbon nanotubes (CNT), boron nitride nanotubes (BNNT), and their complexes, because of their unique
properties. Molecular dynamics (MD) simulations of nanoscale
structures, processes, and properties are very promising in
the wide range of physical parameters, because of necessity
to investigate wide range of parameters (so-called “parameter
sweeping”). MD simulations with such “parameter sweeping”
can be carried out in the available heterogeneous distributed
computing infrastructure (DCI) with desktop grids, service
grids, clusters, supercomputers, cloud resources. The recent
advances in computing algorithms and infrastructures, especially in development of DCIs, allow us to solve these
tasks efficiently without expensive scaling-up, especially by
means of the science gateway (SG) technology on the basis
of WS-PGRADE platform for workflow management and
gUSE technology for integration of DCI-resources [1]. Here
we demonstrate the capabilities of these technologies realized
for MD simulations and further data post-processing on the
basis of LAMMPS package for MD simulation [2], and other
packages like R for statistical analysis [3], Pizza.py Toolkit for
manipulations with atomic coordinates files [4], AtomEye for
visualisation of atomic configurations [5], debyer for simulation of X-ray diffraction (XRD) and neutron diffraction (ND)
analysis (https://code.google.com/p/debyer/), etc. Several typical workflows were created for simulation of several physical
processes with various demands for the computing resources:
tension of metal nanocrystals under different physical conditions, tension of ensemble of metal nanocrystals under the
same conditions, manipulations with complex nanostructures
like indentation of graphene membranes, thermal stability of
BNNT, etc.
II.
BACKGROUND AND R ELATED W ORKS
Scientists from user communities outside of computer
science itself (for example, in materials science, physics,
chemistry, biology, nanotechnologies, etc.) would like to access
various DCIs (grids, clouds, clusters) and concentrate on their
scientific applications themselves in a such transparent way,
that they do not need to learn the peculiarities of these
DCIs. In general, “science gateway” (SG) implies a userfriendly intuitive interface between scientists (or scientific
communities), various software (standard and/or proprietary)
and heterogenous DCIs. Usually, SGs are not specialized for
a certain scientific area and scientists from many different
fields can use them. By means of SGs researchers can focus
on their scientific goals and less on assembling the required
DCI components. The most important aspects of SG are: a
simplified intuitive graphical user interface (GUI) that is highly
tailored to the needs of the given scientific community; a
smooth access to national and international computing and
storage resources; collaborative tools for sharing scientific data
on national and international scale. Several SG technologies
and frameworks are known (like ASKALON [6], MOTEUR
[7], WS-PGRADE/gUSE/DCI-bridge [1]), which use different enabling components and technologies: web application
containers (Tomcat, Glassfish, etc.), portal or web application
frameworks (Liferay, Spring, Drupal, etc.), database management systems (MySQL, etc.), workflow management systems
[8]. Here usage of WS-PGRADE/gUSE/DCI-bridge is described with some examples in physics, materials science, nanotechnologies. The architecture of science gateway ideology on
the basis of gUSE+WS-PGRADE+DCI-Bridge technologies
(https://guse.sztaki.hu) is described in details elsewhere [1].
In short, WS-PGRADE portal technology is a web based
front end of the gUSE infrastructure. gUSE (grid and cloud
User Support Environment) is a permanently improving open
source SG framework developed by Laboratory of Parallel and
Distributed Systems (MTA-SZTAKI, Budapest, Hungary), that
provides the convenient and easy access to DCI. It gives the
generic purpose, workflow-oriented GUI to create and manage
workflows on various DCIs. WS-PGRADE supports development and submission of distributed applications executed on
the computational resources of DCIs. The resources of DCIs
(including clusters, service grids, desktop grids, clouds) are
connected to the gUSE by a single point back end, the DCIBridge [1].
Fig. 1.
The graphical user interface of “IMP Science Gateway Portal”.
create (easily in LEGO-style manner) the complex workflows
(Fig. 2a) from separate simple components, which can be
standard or proprietary executables, scripts in various languages, and command-line arguments (Fig. 2b). Application
developers can design and create complex workflows with
taking into account various runtimes and resources needed
for each component (Fig. 2c), and output data type and size
(Fig. 2d). End-users can submit, monitor, save, upload, and
retrieve various workflows with one or many jobs, and obtain
intermediate or final results (Fig. 2d).
“IMP Science Gateway Portal” [10] (Fig. 1) is based
on the WS-PGRADE/gUSE/DC-bridge framework [1], which
is the complex system with many features. It distinguishes
different user communities: administrators, power-users and
end-users. The administrators operate portal itself. The powerusers can develop new workflow applications by means of
various workflow features (graph, abstract workflow, template,
application). The workflows and their components are stored in
built-in repository and published by the application developers,
which can be shared among end-users. It provides scientific
workflow management for both application developers and end
users, and supports various DCIs including clusters, service
grids, desktop grids and cloud.
The whole process of the complex workflow creation and
management can be demonstrated by the MD simulation of
tension of metal nanocrystals (Fig. 2a) with the following
steps: (1) to register/sign in “IMP Science Gateway Portal”;
(2) to create a graph scheme for the basic components of
the workflow (yellow bricks with small input-output portssquares in Fig. 2a); (3) to create a workflow from the graph;
(4) to configure the workflow (links between yellow bricks
in Fig. 2a); (5) to run the workflow; (6) to monitor the
current status of execution. The complex workflow (Fig. 2)
for this application contains several independent components
(LAMMPS package, Pizza.py Toolkit, AtomEye package, debyer, R-package, FFMPEG), which can be combined in various
sets for the purposes of the end users [11], [12]. The endusers can check the intermediate results: after MD simulation
by LAMMPS package (tables in the red rectangles in the
center of Fig. 2d); converted by Pizza.py Toolkit script package
for post-processing (table in the red rectangle in the bottom
of Fig. 2d); visualizations of atomic positions obtained by
AtomEye package (images in the top part of Fig. 2d); postprocessed results of X-ray diffraction (XRD) and neutron
diffraction (ND) analysis obtained by debyer (tables in the
blue rectangles in the bottom part of Fig. 2d); final results
after XRD and ND analysis, coordination number (CN) and
radial distribution function (RDF) analysis, and stress-strain
(SS) analysis obtained by R-package (plots in the green
rectangles at the left and right sides of Fig. 2d); the final postprocessed video of visualizations of atomic positions obtained
by FFMPEG package (video in the violet rectangle in the top
part of Fig. 2d).
A. Workflow Management
B. Integration of Computing Resources
“IMP Science Gateway Portal” leverages easy and simple GUI utilities for power-users (application developers) to
To run jobs gUSE allows to compose various resources
and execute workflows on various DCIs: clusters (PBS, LSF),
III.
IMP S CIENCE G ATEWAY I NTERNALS
side), several clusters (PBS) with different configuration and
policies, tests desktop grids (BOINC) and cluster grids (ARC),
and plans to use supercomputers and clouds. In this context, the
different requirements for the various components were taken
into account as to their possible location, runtime, output file
size (see Table I).
TABLE I.
a)
T HE COMPONENTS OF THE GENERAL WORKFLOW (F IG . 2)
Component
LAMMPS
R-package
Pizza-script
AtomEye
FFMPEG
debyer
Location
Cluster, DCI
Server (Cluster, DCI)
Server (Cluster, DCI)
Cluster, DCI
Cluster, DCI
Cluster, DCI
Runtime
1-10 days
1-10 min
1-10 min
< 1hour
1-10 min
1-10 days
Output Format (Size)
Text (1-10 GB)
Plot (< 1MB)
Text (1-10 GB)
3D Image (< 1MB)
3D Video (< 10MB)
Text (10-100 MB)
C. User Management
b)
The development and execution stages are separated between users with different background and responsibility
(power-users and end-users) in the next two main user groups:
(1) “end users” — the scientists without extensive knowledge
of computer science with limited power and permissions (like
common scientists in materials science, physics, chemistry,
etc.), usually they need only restricted manipulation rights
and obtain the scientific applications in the ready state to
configure and submit them with minimal efforts; (2) “power
users” (“application developers”) — the actual designers of the
scientific processes and related workflows for simulations, usually they need to design and create the application (workflow)
for comfortable work of the end-users.
IV.
c)
U SE C ASES OF P RACTICAL A PPLICATIONS
Several practical applications in materials science are presented below, where the different workflows were designed by
WS-PGRADE workflow manager and used for computations
in DCI by gUSE technology.
A. Thermal Properties of Boron Nitride Nanotubes
d)
Fig. 2. The scheme of workflow created by means of WS-PGRADE workflow
management technology for MD simulations: (a) main components (see their
description in the text); (b) invariant parts like executables, environmental
parameters (green color) and variable parts like input data, scripts (red color);
(c) job runtimes and resources needed: short runtimes on a server side (red
color), medium — on DCI (blue color), long — on cluster (green color);
(d) output data: rough text data (coordinates and velocities of atoms) of the
huge size(red color), post-processed text data of the medium size (blue color),
output image data (plots, 3D images) of the small size (green color), and
output 3D video data of the small size (violet color).
cluster grids (ARC, gLite, GT2, GT4, GT5, UNICORE), supercomputers (e.g. via UNICORE), desktop grids (BOINC) and
clouds (via CloudBroker Platform and GAE). At the moment
“IMP Science Gateway Portal” uses local resources (server
The scientific aim of this use case (Fig. 2) consists in
MD simulations of boron nitride nanotubes (BNNT) with
exceptional physical properties, which are a prerequisite for
their wide practical applications in the future. The current
user community in the field of nanotechnologies and materials
science includes 2 user groups (4 end users) from IMP
and IPMS (Frantsevich Institute for Problems in Materials
Science, Kiev, Ukraine). MD simulation workflow (Fig. 3a)
was used for investigation of thermal stability of BNNT with
different chirality, size and perfection of structure (Fig. 3b).
The actually used workflow (Fig. 3a) was the part of the
general workflow (Fig. 2a). It included the following subset
workflows: 1) LAMMPS + R-package — for MD simulations
and plotting some post-processed values; 2) LAMMPS + Rpackage + AtomEye — for the same operations + visualization
of intermediate states of atoms.
From the physical point of view this work allowed us to
derive some conclusions as to dynamics (Fig. 3c) and temperature dependence of the BNNT collapse for various physical
parameters (chiralities, sizes and perfection of structure) and
the simulation parameters (type of potential, boundary conditions, thermal conditions, etc.). The obtained results extend the
a)
a)
b)
b)
c)
c)
Fig. 3. Use case for MD simulations of thermal properties of BNNTs: (a)
the part of the general workflow (Fig. 2a), which was actually used here;
(b) visualization by AtomEye (collapse of BNNT, nitrogen atoms — red
color, boron atoms — blue color), (c) data post-processing and plot output
by R-package (dynamics of the collapse for BNNT characterized by atom
coordination numbers).
known experimental data on the thermal stability of BNNTs
and described elsewhere [17].
B. Elastic Properties of Graphene
The scientific aim of this use case (Fig. 4) was MD
simulation of nanoindentation of monolayer graphene membrane in an atomic force microscope. Its user community in
the field of nanotechnologies includes 2 user groups (5 end
users) from IMP and SPM/RS-Centre (Centre of scanning
probe microscopy and resonance spectroscopy, Kyiv, Ukraine).
MD simulation workflow (Fig. 4a) was used for investigation of nanoindentation processes with different velocities,
size of membranes and perfection of structure (Fig. 4b).
The actually used workflow (Fig. 4a) was the part of the
general workflow (Fig. 2a). It included the following subset
workflows: 1) LAMMPS + R-package — for MD simulations
and plotting some post-processed values; 2) LAMMPS + Rpackage + AtomEye + FFMPEG — for the same operations +
visualization of intermediate states of atoms, and final video
visualization.
From the physical point of view it was shown that graphene
Fig. 4. Use case for MD simulations of elastic properties of graphene: (a)
the part of the general workflow (Fig. 2a), which was actually used here; (b)
visualization by AtomEye (nanoindentation of the graphene monolayer, mobile
carbon atoms — red color, fixed carbon atoms — yellow color, nanoindentor
atoms — gray color), (c) data post-processing and plot output by R-package
(for different speeds of the nanoindentor).
membrane deforms and collapses within a wide range of
physical parameters (strain, applied force, indentor shape,
indentor velocity, etc.) (Fig. 4c). These results confirm and
enrich the known experimental data of nanoindentation and
described elsewhere [16].
C. Mechanical Properties of Metal Nanocrystals
The scientific aim of this use case (Fig. 5) was MD
simulation of relaxation behavior of nanocrystals after uniaxial
tension (Fig. 5b) [13], [14]. Its user community in the field of
nanotechnologies includes 2 user groups (3 end users) from
IMP and KNU (Kyiv National University, Kyiv, Ukraine).
MD simulation workflow (Fig. 5a) was used for investigation
of relaxation behavior of nanocrystals for different crystals
(Al, Cu, and Si), physical conditions (tensile rate, size and
orientation of the nanocrystals) and methodological parameters
(potential type, boundary conditions, and others). The actually
used workflow (Fig. 5a) includes the biggest parts of the
general workflow (Fig. 2a). The simulations included the
following subset workflows: 1) LAMMPS + R-package —
for MD simulations and plotting some post-processed values,
like stress-strain dependencies (Fig. 5c); 2) LAMMPS + Rpackage + AtomEye + FFMPEG — for the same operations
+ visualization of intermediate states of atoms and final
video visualization; 2) LAMMPS + R-package + AtomEye
+ FFMPEG + debyer — for the same operations + processing
data for X-ray and neutron diffraction and their plotting, like
XRD-plots (Fig. 5d).
From the physical point of view it was shown that the
oscillatory relaxation of internal stresses is associated with the
periodic rearrangement of the metastable defect substructure
(Fig. 5b). These results confirm and illustrate the known
experimental data and described elsewhere [15].
V.
D ISCUSSION
The use cases presented here demonstrate that the general
purpose and application specific workflows can be created in
LEGO-style way by combining separate components-bricks,
which corresponds to some real world operations, physical
methods, and experimental devices even (see Table II). Actually, such workflows can be easily designed and used in
practice for MD simulations of complex natural processes
in materials science, physics, and nanotechnologies. Each
simulation workflow can be considered as a virtual counterpart
(virtual lab) of the real experimental process, and the SG itself
can be used as a hub of such virtual experimental labs in
materials science.
TABLE II.
a)
b)
R ESULTS OBTAINED , VIRTUAL LAB ( SET OF COMPONENTS
USED ), AND CORRESPONDING REAL LAB
Simulation result
(method needed)
Atomic structure
in bulk (TEM)
Stress-strain plot
(AFM)
CN, RDF plots (neutron,
X-ray scattering)
X-ray diffraction
spectrum (XRD)
Neutron diffraction
spectrum (ND)
Virtual lab,
i.e. set of components
LAMMPS [2]
+ AtomEye [5]
LAMMPS [2]
+ R [3]
LAMMPS [2]
+ R [3]
LAMMPS [2]
+ debyer + R [3]
LAMMPS [2]
+ debyer + R [3]
Real lab,
device
JEOL R005
microscope [19]
H-P testing machine,
JEOL AFM
X-ray
diffractometer
SOLEIL [20]
synchrotron facility
Swiss Light Source [21],
Large Hadron Collider
Some conclusions were made as to the possible future
improvements of the workflow and resource management.
The main problem was concerned with different demands for
resources for various job parameters. For example, double
increase of model size leads to nearly proportional double
increase of the required memory and runtime. Sometimes this
was in conflict with the available policies, like queueing rules,
when the job runtime should be lower than the queue walltime.
The typical example for the use case A (Fig. 3) with the
physical simulation time of 1 ns is shown in Table III for one of
the used PBS-clusters, namely NTU HPCC (http://hpcc.kpi.ua)
of NGI-UA. In this connection, the GUI-feature, which was
pre-designed in SG for selection of a queue for PBS-clusters,
was very useful, because it allowed to construct workflows optimized for runtimes lower that the prescribed queue walltime
(or when the walltime is a multiple of the job runtime). This
allowed to increase the overall efficiency, but the more progress
could be reached by additional measures in SG, for example,
by optimal job distribution over various queues (which should
be automatic, and not manual, like here), and by polling all
available queues among all available clusters with selection of
the pool with the empty or more efficient resources.
c)
d)
Fig. 5. Use case for MD simulations of mechanical properties of metal
nanocrystals: (a) the part of the general workflow (Fig. 2a), which was actually
used here; (b) visualization by AtomEye (defect substructure evolution in
Al nanocrystal, after 15 ps, atoms in ideal lattice — green color, stacking
faults — red color), (c) data post-processing and plot output by R-package
(oscillation of the internal stress Pxx for different nanocrystal sizes), (d) data
post-processing and plot output by R-package (X-ray diffraction analysis).
Other means for increase of SG-efficiency, like dual-layer
hardware and software management in cluster systems [18],
will be also of interest and great value and should be taken
TABLE III.
T HE QUEUE RULES AND JOBS IN U SE C ASE A (F IG . 3)
Queue name
ku-small
ku-single
ku-normal
kh-large
Walltime
90 min.
8 days
180 min.
120 min.
Cores/User
32
4
32
128
Max BNNT (atoms)
< 10 × 10 (< 840)
> 40 × 40 (> 3360)
< 30 × 30 (< 2520)
< 20 × 20 (< 1680)
[5]
[6]
[7]
into account. Additional important aspect is related with the
failed jobs, which should be re-run. The problem is that for the
big number of jobs (each for different set of parameters) it was
not very easy to filter the faulty jobs and re-run them again.
At this stage, it was performed manually by means of job
management system, but it would be more efficient, if this will
done automatically by SG after obtaining the corresponding
report.
VI.
[8]
[9]
[10]
C ONCLUSIONS
Science Gateway Portal (like “IMP Science Gateway Portal”) for MD simulations of complex behavior of various
nanostructures can be relatively easily created on the basis of
WS-PGRADE technology for workflow management. These
resourceful MD simulations can be effectively carried out in
the heterogeneous DCI by means of gUSE technology for
parallel and distributed computing in various heterogeneous
DCIs. Several typical scientific applications were considered
as use cases of their porting and practical using. Several advantages should be emphasized: (1) the modular-based (“LEGOstyle”) principle of workflow creation and management; (2)
easy access to heterogeneous DCI and software; (3) division
of user roles (administrators as operators of portals; power
users as principal scientists; end users as scientists, students);
(4) short learning curve for usual scientists without extensive knowledge in computer science. Some ways for further
improvement were briefly outlined in relation to the better
performance and efficiency. Finally, the “science gateway”
approach – workflow manager (like WS-PGRADE) + DCI
resources manager (like gUSE) at the premises of the portal
(like “IMP Science Gateway Portal”) – is very promising in the
context of its practical MD applications in materials science,
physics, chemistry, biology, and nanotechnologies.
[11]
[12]
[13]
[14]
[15]
[16]
ACKNOWLEDGMENTS
The work presented here was partially funded by EU FP7
SCI-BUS (SCIentific gateway Based User Support) project,
No. RI-283481, and partially supported in the framework of
the research theme ”Introduction and Use of Grid Technology
in Scientific Research of IMP NASU” under the State Targeted
Scientific and Technical Program to Implement Grid Technology in 2009-2013.
R EFERENCES
P.Kacsuk,
Z.Farkas,
M.Kozlovszky,
G.Hermann,
A.Balasko,
K.Karoczkai, I.Marton, WS-PGRADE/gUSE generic DCI gateway
framework for a large variety of user communities, Journal of Grid
Computing, 10(4): 601-630, 2012.
[2] S.Plimpton, Fast parallel algorithms for short-range molecular dynamics,
Journal of Computational Physics, 117(1):1-19, 1995.
[3] R Development Core Team, R: A language and environment for statistical
computing, R Foundation for Statistical Computing, Vienna, Austria.
ISBN 3-900051-07-0 (http://www.R-project.org) 2008.
[4] S.Plimpton, M.Jones, Pizza.py Toolkit (http://pizza.sandia.gov/) 2005.
[17]
[18]
[19]
[20]
[1]
[21]
J.Li, AtomEye: an efficient atomistic configuration viewer, Modelling
and Simulation in Materials Science and Engineering, 11(2): 173-177,
2003.
T.Fahringer, A.Jugravu, S.Pllana, R.Prodan, C.Seragiotto, H.L.Truong,
ASKALON: a tool set for cluster and Grid computing, Concurrency and
Computation: Practice and Experience, 17(2-4): 143-169, 2005.
T.Glatard, J.Montagnat, D.Lingrand, X.Pennec, Flexible and efficient
workflow deployment of data-intensive applications on grids with moteur,
International Journal of High Performance Computing Applications,
22(3): 347-360, 2008.
E.Deelman, D.Gannon, M.Shields, I.Taylor, Workflows and e-Science:
An overview of workflow system features and capabilities, Future
Generation Computer Systems, 25(5): 528-540, 2009.
WS-PGRADE
Portal
User
Manual,
2013
(http://sourceforge.net/projects/guse/).
Yu.Gordienko, L.Bekenov, O.Gatsenko, E.Zasimchuk, V.Tatarenko,
Complex Workflow Management and Integration of Distributed Computing Resources by Science Gateway Portal for Molecular Dynamics
Simulations in Materials Science. Proc. Third Int. Conf. “High Performance Computing” HPC-UA 2013, (Ukraine, Kyiv, 2013), 148-155.
Yu.Gordienko, O.Baskova, O.Gatsenko, Enabling High-Performance
Distributed Computing to e-Science by Integration of 4th
Generation Language Environments with Desktop Grid Architecture
and Convergence with Global Computing Grid. Proceedings
of Cracow Grid Workshop’10, Cracow, Poland, pp. 234-243,
(https://www.researchgate.net/publication/261361640) 2011.
O.Gatsenko, L.Bekenev, E.Pavlov, Yu.Gordienko: From Quantity
to Quality: Massive Molecular Dynamics Simulation of Nanostructures Under Plastic Deformation in Desktop and Service Grid Distributed Computing Infrastructure, Computer Science, 14: 27-44;
doi:10.7494/csci.2013.14.1.27, 2013.
Yu.Gordienko, Molecular Dynamics Simulation of Defect Substructure Evolution and Mechanisms of Plastic Deformation in Aluminium
Nanocrystals, Metallofizika i Noveishie Tekhnologii, 33(9): 1217-1247,
2011 (in Ukrainian).
A.Gatsenko, A.Baskova, B.Bandalak, V.Tatarenko, Yu.Gordienko,
Application of Desktop Grid Technology in Material Science,
5th Int. Conf. “Distributed Computing and Grid-technologies in
Science and Education” GRID’2012, Dubna, Russia, pp.123-129,
(http://www.researchgate.net/publication/258669295) 2012.
Yu.Gordienko,
E.E.Zasimchuk,
L.P.Steblenko,
A.N.Krit,
D.V.Kalinichenko, The Oscillatory Nature of Internal Stress
Relaxation due to Self-organization of Defect Substructure
in Nanocrystals, Int. Conf. NANSYS-2013, Kyiv, Ukraine,
(http://www.researchgate.net/publication/259338343) 2013.
A.S.Gatsenko, E.E.Zasimchuk, V.A.Tatarenko, V.L.Karbovsky,
Yu.Gordienko, Elastic Properties and Strength of Graphene
Under Nanoindentation, Int. Conf. NANSYS-2013, Kyiv, Ukraine,
(http://www.researchgate.net/publication/258872193) 2013.
Yu.Gordienko, L.P.Sartinska, Thermal Stability of Boron
Nitride Nanotubes, Int. Conf. NANSYS-2013, Kyiv, Ukraine,
(http://www.researchgate.net/publication/259338325) 2013.
S.Stirenko, O.Zinenko, D.Gribenko, Dual-layer hardware and software
management in cluster systems, Proc. Third Int. Conf. ”High Performance Computing” HPC-UA 2013, (Ukraine, Kyiv, 2013), 380-385.
S.J.Haigh, H.Sawada, K.Takayanagi, A.I.Kirkland, Exceeding conventional resolution limits in high-resolution transmission electron microscopy using tilted illumination and exit-wave restoration, Microscopy
and Microanalysis, 16(04), 409-415, 2010.
F.Baudelet, R.Belkhou, V.Briois, A.Coati, P.Dumas, SOLEIL a new
powerful tool for materials science, Oil Gas Science and Technology,
60(5), 849-874, 2005.
B.D.Patterson, R.Abela, H.Auderset, Q.Chen, F.Fauth, F.Gozzo, et al.,
The materials science beamline at the Swiss Light Source: design and realization, Nuclear Instruments and Methods in Physics Research Section
A: Accelerators, Spectrometers, Detectors and Associated Equipment,
540(1), 42-67, 2005.
| 5cs.CE
|
An Artificial Neural Network Architecture Based on Context
Transformations in Cortical Minicolumns
Vasily Morzhakov, Alexey Redozubov
[email protected], [email protected]
Abstract
Cortical minicolumns are considered a model of cortical organization. Their function
is still a source of research and not reflected properly in modern architecture of nets in
algorithms of Artificial Intelligence. We assume its function and describe it in this
article. Furthermore, we show how this proposal allows to construct a new
architecture, that is not based on convolutional neural networks, test it on MNIST data
and receive close to Convolutional Neural Network accuracy. We also show that the
proposed architecture possesses an ability to train on a small quantity of samples. To
achieve these results, we enable the minicolumns to remember context
transformations.
1. Introduction
Human perception always deals with ambiguity in the brain. Only receptors provide us with clear input. As
the input data from receptors start to be analyzed, explicit treatment disappears. Almost every frame of data
can have a different sense in a different context. This is reflected in all types of perception. There are many
illusions in visual perception that are caused by the presence of different geometric contexts, for example:
“inverted face”. Our language seems to have to exclude the ambiguity, but it doesn't. All humor is based on
unexpected changes of contexts. Even our delusions are possible because of our ability to find context where
they have sense. All these specificities of human perception are not fundamentally reflected in modern
machine-learning approaches.
Artificial neural networks (ANN) are presented by different architectures today. Recurrent ANNs and
Convolution ANNs don't have much in common and are based on different concepts. Meanwhile, the
neocortex is very homogeneous through its surface. This means that a new concept has to be applied for
different types of data: images, video, audio and natural language. Even manipulators' controlling should be
explained in this approach. Now, all these tasks use completely different algorithms and theories.
One more fundamental problem of modern ANNs is that they need more samples for training than humans.
Convolution ANNs (CNN) take into account translation invariants, recurrent ANNs (RNN) – time invariants.
It's impossible to know how rotate objects will be described by a CNN without showing a particular object in
different rotations. RNN can't extract a twice slower process in time if it isn't presented in training data. CNN
and RNN were breakthroughs because they could extract features in different positions and time. Position
and time are contexts too. We also deal with a lot of other contexts all the time: orientation, size, frequency,
tempo, shape and others. We have awareness of how the data would be changed in all of those contexts. It
will dramatically reduce the amount of samples we need for extracting features in the input data we have.
These facts allow us to find new concepts that will allow us to go beyond all limitations described above.
In this article, we propose function of minicolumns in the neocortex and describe a sample implementation
for MNIST data that is based on actual machine-learning framework.
1.1 Brain zone and minicolumns functions
Hubel and Wiesel [2] proposed a concept that is widely accepted in modern implementations of
Convolutional Neural Networks (CNN). They claimed that brain zones in the neocortex are organized into
minicolumns that contain neuron-detectors with similar functions. They found that in the visual zone V1,
minicolumns have orientational selectivity in their experiments. A lot of research shows different types of
selectivity in different brain zones. According to these experiments, a few approaches in computer vision
appeared: self-organizing maps and neocognitron. The neocognitron architecture found its realization in
convolutional neural networks that still provide the best accuracy of recognition in tasks of computer vision.
Self-organizing maps are also well-known in tasks of unsupervised learning.
We propose another interpretation that allows the creation of an original approach for data generalization and
recognition tasks, inspired by more relevant information to today’s research. Roots of our interpretation are
described in detail in [1].
Almost all data processed by the brain is presented unambiguously. Objects are always projected to the retina
in different ways, words are pronounced in different tones, even words in our language have different senses
in different contexts. We assumed that each minicolumn in the neocortex presents the input information in its
own context. The second task of the minicolumn should solve pattern recognition. Activity of minicolumns
is proportional to certainty of recognition in the context of each minicolumn. Close minicolumns represent
similar context transformation. This way we can single out local maximums of activity in the field of
minicolumns and send the following output to the next level: what the minicolumn has recognized and in
what context (transformation). Also, it's important that patterns for recognition are shared among
minicolumns.
Figure 1, Input data received by a zone is translated in different contexts by minicolumns, a few of
minicolumns send what they recognize and its contexts to the following level
Thus, neurons should have two types of memory: memory for transformation and memory for generalization.
Modern vision of neurons' structure has gone far from the vision in times of Hubel and Wiesel. Possible
mechanisms of these types of memory will be described in [1]. The sample architecture in this article is
based on the TensorFlow framework and the classic perceptron conception.
2.
Model for MNIST
MNIST dataset was chosen in order to demonstrate described functionality of minicolumns inside one zone.
We created a model [7], that doesn't contain all features of our approach, but which should illustrate
proposed functions of a brain zone. This sample model was based on well-developed frameworks for deep
learning, thus the architecture and source codes are transparent for deep learning specialists.
A zone receives a compressed vector (code) that describes relevant information for the zone. We trained an
autoencoder with a bottleneck to obtain the input vector. Additionally, we were able to use its decoder part
for visualization of codes in minicolumns. The green layers in figure 2 mark encoder and decoder parts of the
autoencoder.
As in 1.1, described transformation to different contexts should be remembered. We chose geometric
transformations for MNIST images: translation, scale, rotation. 405 minicolumns were set ( 9 translations x
9 scales x 5 rotations = 405 transformations).
Transformation remembering was implemented by ANNs with one hidden layer for each. Each ANN was
trained to transform the input vector to the code of the transformed image.
After input codes were transferred into different contexts, we had codes in each minicolumn. Thus, 405
minicolumns contained codes of images if they were geometrically transformed. These codes are decoded by
the decoder part of the autoencoder for visualization (2.2).
Figure 2, model for MNIST data,
green layers – autoencoder part
red layers – transformation part
yellow – generalization with convolution layers (weights are shared for transformed codes)
blue – codes, that describe images
The next step was feature extracting from the minicolumns. The best way for this was using the convolution
layer. Convolution layers have kernels (or features), the weights for detecting those features are applied for
each position on an image. So, minicolumn codes might be presented as an image that allow the application
of the same weight vector for each minicolumn. In this way, we could extract features despite their context.
It allows for a decrease in the number of weight coefficients in CNNs and an increase in the quality of
prediction. Following the function of a brain zone, the output equals the local maximums of the last
convolution layer. For this sample, we left only one maximum for each channel and applied a max-pooling
layer. The output of the max-pooling layer was connected to a fully-connected layer with 10 outputs and
softmax activation for training recognition on MNIST.
2.1 Autoencoder
First of all, we needed to extract a vector that would be an input for the zone. An autoencoder with two
fully-connected layers allowed to compress input (28,28) images to vectors in 32 dimensions.
Figure 3, An autoencoder with fully-connected layers
Two fully-connected layers were used for encoding an input image and two for decoding it. The network was
trained with geometrically augmented images from MNIST.
Mean squared error was chosen while training. The mean squared average error was 0.011 after 50 epochs.
These results were a bit worse than the error of a convolutional autoencoder in our experiments (0.006), but
it was very important to exclude the CNN for the purity of experiments.
2.2 Сontext transformation training
The next stage was training context transformations. In the article we use the term “context transformation
training” because we used the back-propagation approach, however, this process should be closer to
remembering.
In some cases it's possible to define a set of transformation. For example, for MNIST dataset, we could
define geometric transformations: translation, scale, rotation. There were 405 minicolumns for 405 geometric
transformations. We trained a separate model (figure 4) for each context resulting in 405 nets trained one by
one. While training, we entered two codes into the model: before and after transformation.
Figure 4, context transformation model for a minicolumn
Input images were encoded into short codes. Each code was transformed to 405 vectors by trained nets, that
could be decoded by the second part of the autoencoder (decoder part) and visualized:
Figure 5, a sample of decoded codes after transformations in 405 minicolumns visualized by the decoder
Figure 5 shows that all of minicolumns were trained successfully. After decoding codes in minicolumns, we
could see transformed input images.
Additionally, this approach of training transformation has a remarkable prediction ability. For example, in
figure 6, transformations of a triangle standing on its edge are shown:
figure 6, transformations of a triangle sign
This sign resembling digits was not presented in the training set. This meant that nets could predict
transformations for code they have never seen. It offers hope of a finite training set for remembering
transformations of full a variety of input codes into different contexts.
2.3 Recognition whether contexts
After we received the codes from the minicolums, we could apply the same detector for each minicolumn for
pattern recognition.
Different approaches may be applied for detecting patterns in different contexts, it could be remembering
samples. The only thing that should be defined is context presentation choice for remembering. For this
reason, we decided to use the back-propagation mechanism for training detectors' weights.
The shortest way for sharing detectors through different contexts was to use a convolution layer. Instead of
sharing weights among detectors at different positions, the convolution layer provided weight sharing for
different contexts. This feature is especially important because it allows recognition patterns even if there
were no samples in all representations, therefore, “one shot learning” becomes possible. Then, for each
channel a max-pooling layer is applied that actually chooses the best matching context.
For MNIST data recognition, we needed one additional layer – a fully-connected layer with softmax
activation with 10 output neurons.
While the training procedure weights in the autoencoder and the transformation net was frozen, only the
weights in the convolutional and output fully-connected layers were trained.
3.
Experiments
We tested the model on MNIST data. There were a few experiments:
1)
Training on the first 1,000 images
2)
Training on the whole dataset of 60,000 images
We also compared our results with a convolutional neural network (CNN) in case of augmented data and
original data.
3.1
Training on 1,000 images
It's very important to show that it demands less data for training for the proposed architecture, that's why its
performance was estimated by training on the first 1,000 images from dataset.
Figure 7, training curve for 1,000 images
After 180 epoches, we had 96.0% accuracy on 10,000 test images.
The baseline CNN for comparison had the following layers:
Layer
Parameters
Convolutional Layer
64 channels, 5x5 kernel size, ReLu activation
Max-pooling
2x2 kernel
Convolutional Layer
128 channels, 5x5 kernel size, ReLu activation
Max-pooling
2x2 kernel
Fully-connected layer
256 outputs, ReLu activation
Fully-connected layer
10 outputs, softmax activation
After training the baseline CNN, the accuracy was 95.3%
Once can see that we used augmented data implicitly while training context transformations. For this reason,
the baseline CNN should be trained on augmented data, too. This test was also run resulting in 96.2%
accuracy.
So, we can conclude that the proposed architecture shows comparable performance with the convolutional
neural network, even without internal convolutional layers.
3.2
Training on 60,000 images
The best result was achieved after 1,000 epochs training on 60,000 images – 99.1%
Figure 8, training curve for 60,000 images
This result is worse than what the CNNs achieve. It's supposed to be caused by presence of a not
precise autoencoder that calculates short representation of data for input.
Comparison to other approaches
The idea of image transformation in computer vision tasks is well known and described at least in Spatial
Transformer Networks [5]. The approach authors proposal can be used across different types of
transformations and types of data. The main difference is that it demands for differentiable transformation
functions that can be unattainable in case of discrete data. Additionally, transformation cannot be trained in
Spatial Transformer Networks.
Hinton writes about capsules that are inspired by cortical minicolumns in [3]. The proposed approach
connects to the CapsuleNet in the “visual stream for perception”. Objects can be contexts for recognition of
transformations, but in the CapsuleNet, minicolumns don't remember how the geometric transformation
changes in different contexts (digits in the case of MNIST data).
Two-stream hypothesis
The Two-stream hypothesis is a well-known and accepted model of visual and auditory perception [4,6]. The
ventral is called “what pathway”, the dorsal stream “where pathway”. Minicolumns in the first one are
selective to object types, their shapes and colors. The second stream has minicolumns with position and
orientation selectivity. [6]
Figure 9, Hierarchy of visual processing in ventral and dorsal streams [6]
The interpretations that we propose change the way of perception in the visual stream of the brain
significantly. Minicolumn activity results in the best context or transformation of the input data. Meanwhile,
minicolumns in the “where pathway” contain sufficient information about objects inside to help recognize
them in the context of a particular minicolumn.
It is quite intuitive in the case of the “where pathway”. Contexts are just geometric transformations, as
implemented in the MNIST sample implementation. It is not hard to imagine transformations mentioned in
the upper part of Figure 7, because almost all of them are just changing the coordinate system in time and
space. However, it's a bit counterintuitive when we speak “in context of an object”. For example, in [3]
Hinton has shown a model where the minicolumns' activation was connected with recognition of a particular
digit in MNIST. The problem is regulated through the assumption that codes in minicolumns may
characterize objects' properties. Those properties may be the same for different objects, for example, the font
style for digits in MNIST. Thus codes in minicolumns during context transformation training will designate
style. It is very close to the ideas of CapsuleNet, but styles' code will be shared among minicolumns.
The two-stream hypothesis provides the idea, that the context-transformation described in 2.2 may not
translate input codes in the same space where input vectors were defined, but in a new space that describes
something else.
Conclusions
In this paper, we introduced a new architecture of artificial neural networks that is based on our original
ideas of the cortical minicolumns' functionality. Transformation input data into a context specific to a
minicolumn allows one to choose the best matching pattern while recognition. It also gives an opportunity
for separation patterns and contexts in which they are presented. Moreover, there is no restriction for input
data or for the type of transformation. As we discussed in the chapter “two-stream hypothesis” detected
patterns in one zone may be contexts for other zones, that may give clues for “abstract thinking” of artificial
intelligence.
Acknowledgments
The authors would like to thank Anton Maltsev for his valuable remarks and discussions.
References
1.
2.
3.
4.
5.
6.
7.
Alexey Redozubov. Holographic Memory: A Novel Model of Information Processing by Neuronal
Microcircuits. Springer Series in Cognitive and Neural Systems Volume 11, 2017, 271-295
Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate
cortex. The Journal of physiology, 195(1), 215–243.
Sara Sabour, Nicholas Fross, Geoffrey E. Hinton. Dynamic Routing Between Capsules.
arXiv:1710.09829v2 [cs.CV] 7 Nov 2017
Robert D. McIntosh, Thomas Schenk. Two visual streams for perception and action:
Current trends. Neuropsychologia 47 (2009) 1391–1396
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu. Spatial Transformer
Networks. arXiv:1506.02025v3 [cs.CV] 4 Feb 2016
Perry CJ and Fallah M (2014) Feature integration and object representations along the dorsal stream
visual hierarchy. Front. Comput. Neurosci. 8:84. doi: 10.3389/fncom.2014.00084
Morzhakov Vasily. MNIST sample. https://gitlab.com/Morzhakov/ContextNet
| 9cs.NE
|
ON ASYMPTOTICALLY MINIMAX NONPARAMETRIC
DETECTION
OF SIGNAL IN GAUSSIAN WHITE NOISE
arXiv:1705.07408v1 [math.ST] 21 May 2017
MIKHAIL ERMAKOV
[email protected]
Institute of Problems of Mechanical Engineering, RAS and
St. Petersburg State University, St. Petersburg, RUSSIA
Mechanical Engineering Problems Institute
Russian Academy of Sciences
Bolshoy pr.,V.O., 61
St.Petersburg
Russia
St.Petersburg State University
University pr., 28, Petrodvoretz
198504 St.Petersburg
Russia
St.Petersburg department of
Steklov mathematical institute
Fontanka 27, St.Petersburg 191023
Abstract. For the problem of nonparametric detection of signal in Gaussian
white noise we point out strong aasymptotically minimax tests. The sets of
r
alternatives are a ball in Besov space B2∞
with ”small” balls in L2 removed.
keyword 1 (class=AMS). [Primary ]62G10,62G20,62M02
keyword 2. signal detection, asymptotic minimaxity, asymptotic efficiency, Gaussian white noise
1. Introduction. Main Result.
Let we observe a random process Yǫ (t), t ∈ (0, 1), ǫ > 0, defined by stochastic
differential equation
dYǫ (t) = θ(t) dt + ǫdw(t)
(1.1)
with Gaussian white noise w(t). The signal θ ∈ L2 (0, 1) is unknown.
Our goal is to test the hypothesis
H0 : θ(t) = 0, t ∈ (0, 1)
versus the alternative
Hǫ :
Z
1
θ2 (t) dt = ||θ||2 > ρǫ > 0,
0
if a priori information is provided that
∞
∞
X
X
r
θj φj (t), k −2r
θj2 ≤ P0 , t ∈ (0, 1), 1 ≤ k < ∞
θ ∈ B2∞
(P0 ) = θ : θ(t) =
j=1
j=k
1
2
MIKHAIL ERMAKOV
with P0 > 0. Here φj , 1 ≤ j < ∞, is orthonormal system of functions. For wide
class of orthonormal systems of functions φj , 1 ≤ j < ∞ the space
∞
∞
X
X
θj φj (t), k −2r
θj2 < ∞, t ∈ (0, 1), 1 ≤ k < ∞
θ : θ(t) =
j=1
j=k
r
B2∞
is Besov space
(see [8])
r
Denote Qǫ = {θ : ||θ||2 ≥ ρǫ , θ ∈ B2∞
(P0 )}.
For any test Kǫ denote α(Kǫ ) its type I error probability and denote βθ (Kǫ ) its
type II error probability for the alternative θ ∈ Qǫ .
We put
βǫ (Kǫ ) = sup βθ (Kǫ ).
θ∈Qǫ
We say that family of tests Lǫ is asymptotically minimax if, for any family of tests
Kǫ , α(Kǫ ) ≤ α(Lǫ ), there holds
lim sup βǫ (Kǫ ) − βǫ (Lǫ ) ≥ 0.
ǫ→0
Paper goal is to establish asymptotically minimax families of tests Lǫ for the sets of
alternatives Qǫ . If the sets of alternatives are ellipsoids with ”small balls” removed,
asymptotically minimax families of tests have been found in [2]. For nonparametric
hypothesis testing this result can be considered as a version of Pinsker Theorem
[6, 7, 5] on asymptotically minimax nonparametric estimation. Note that hypothesis
testing with nonparametric sets of alternatives belonging some ball in functional
space is intensively studied (see [4, 1] and references therein).
The proof, in main features, repeats the reasoning in [2]. The main difference
in the proof is the solution of another extremal problem minimizing type II error
probabilities caused another definition of sets of alternatives. Other differences
have technical character and are also caused the differences of definitions of sets of
alternatives.
Define k = kǫ and κ2 = κ2ǫ as a solution of two equations
2rkǫ2r+1 κ2ǫ = P0
(1.2)
kǫ κ2ǫ + kǫ−2r P0 = ρǫ .
(1.3)
and
κ2j
κ2ǫ ,
for 1 ≤ j ≤ kǫ and
Denote
=
Define test statistics
κ2j P0 (2r)−1 j −2r−1 ,
Tǫa (Yǫ ) = ǫ−4
∞
X
for j > kǫ .
κ2j yj2 .
j=1
e
Aǫ = ǫ−4
∞
X
κ4j .
j=1
For type I error probabilities α, 0 < α < 1, define critical regions
Sǫa == {y : (Tǫa (y) − ǫ−2 ρǫ )(2Aǫ )−1/2 > xα }
with xα defined by equation
α = 1 − Φ(xα ) = (2π)−1/2
Z
∞
exp{−t2 /2} dt.
xα
Theorem 1.1. Let
0 < lim inf Aǫ ≤< lim sup Aǫ < ∞.
ǫ→0
ǫ→0
(1.4)
ON ASYMPTOTICALLY MINIMAX NONPARAMETRIC DETECTIONOF SIGNAL IN GAUSSIAN WHITE NOISE
3
Then the tests Laǫ with critical regions Sǫa are asymptotically minimax with α(Laǫ ) =
α(1 + o(1)) and
βǫ (Laǫ ) = Φ(xα − (Aǫ /2)1/2 )(1 + o(1))
(1.5)
as ǫ → 0.
8β
Example. Let ρǫ = Rǫ 4β+1 . Then
1/2r
4r−1
2r
P0
4r + 2
R
.
Aǫ =
2r
4r + 1 2r + 1
In what follows, we shall denote letter C and C with indices different generic constants.
2. Proof of Theorem 1.1
Fix δ, 0 < δ < 1. Denote κ2j (δ) = 0 for j > δ −1 kǫ . Define κ2j (δ), 1 ≤ j < kǫδ =
δ kǫ , the equations (1.2) and (1.3) with P0 and ρǫ replaced with P0 (1 − δ) and
ρǫ (1 + δ) respectively. Similarly to [2], we find Bayes test for a priori distribution
θj = ηj = ηj (δ), 1 ≤ j < ∞, with Gaussian independent random variables ηj , Eηj =
0, Eηj2 = κ2j (δ) and show that this test is asymptotically minimax for some δ = δǫ →
0 as ǫ → 0.
−1
Lemma 2.1. For any δ, 0 < δ < 1, there holds
P (η(δ) = {ηj (δ)}∞
j=1 ∈ Qǫ ) = 1 + o(1)
(2.1)
as ǫ → 0.
Denote
Aǫ,δ = ǫ−4
∞
X
κ4j (δ).
j=1
By straightforward calculations, we get
lim lim Aǫ A−1
ǫ (δ) = 1.
δ→0 ǫ→0
(2.2)
Denote γj2 (δ) = κ2j (δ)(ǫ2 + κ2j (δ))−1 .
By Neymann-Pearson Lemma the Bayes critical region is defined the inequality
Z
kǫδ
kǫδ
X
Y
(2γj2 (δ))−1 (uj − γj2 (δ)yj )2 exp{−Tǫδ (y)}
(2π)−1/2 κj−1 (δ) exp −
C1 <
j=1
j=1
= exp{−Tǫδ (y)}(1 + o(1))
(2.3)
where
Tǫδ (y) = ǫ−2
∞
X
γj2 (δ)yj2 .
j=1
Define critical region
Sǫδ = {y : Rǫδ (y) = (Tǫδ (y) − Cǫδ )(2Aǫ (δ))−1/2 > xα }
with
Cǫδ = E0 Tǫδ (y) = ǫ−2
∞
X
γj2 (δ).
j=1
Denote Lǫδ the tests with critical regions Sǫδ .
Denote γj2 = κ2j (ǫ2 +κ2j )−1 , 1 ≤ j < ∞ Define test statistics Tǫ , Rǫ , critical regions
Sǫ and constants Cǫ by the same way as test statistics Tǫδ , Rǫδ , critical regions Sǫδ
and constants Cǫ,δ respectively with γj2 (δ) replaced with γj2 respectively. Denote
Lǫ the test having critical region Sǫ .
4
MIKHAIL ERMAKOV
Lemma 2.2. Let H0 hold. Then the distributions of tests statistics Rǫa (y) and
Rǫ (y) converge to the standard normal distribution.
For any family θǫ = {θjǫ } ∈ Qǫ there holds
∞
X
2
(2Aǫ )−1/2 < xα = Φ(xα )(1+o(1)). (2.4)
κ2j θjǫ
Pθǫ Tǫa (y) − ǫ−2 ρǫ − ǫ−4
j=1
and
Pθǫ Tǫ (y) − Cǫ − ǫ−4
∞
X
j=1
as ǫ → 0.
2
(2Aǫ )−1/2 < xα = Φ(xα )(1 + o(1)).
κ2j θjǫ
(2.5)
Hence we get the following Lemma.
Lemma 2.3. There holds
βǫ (Lǫ ) = βǫ (Laǫ )(1 + o(1))
(2.6)
as ǫ → 0.
Lemma 2.4. Let H0 hold. Then the distribution of tests statistics (Tǫδ (y) −
Cǫδ )(2Aǫ )−1/2 converge to the standard normal distribution.
There holds
Pη(δ) ((Tǫδ (y) − Cǫδ − Aǫδ )(2Aǫδ )−1/2 < xα ) = Φ(xα )(1 + o(1)).
(2.7)
as ǫ → 0.
Lemma 2.5. There holds
lim lim Eη(δ) βη(δ) (Lǫδ ) = lim Eη0 βη0 (Lǫ )
δ→0 ǫ→0
ǫ→0
(2.8)
2
where η0 = {η0j }∞
j=1 and η0j are i.i.d. Gaussian random variables, Eη0j = 0, η0j =
2
κj , 1 ≤ j < ∞.
Define Bayes a priori distribution Py as a conditional distribution of η given
η ∈ Qǫ . Denote Kǫ = Kǫδ Bayes test with Bayes a priori distribution Py . Denote
Vǫ critical region of Kǫδ .
For any sets A and B denote A△B = (A \ B) ∪ (B \ A).
Lemma 2.6. There holds
lim lim
δ→0 ǫ→0
Z
Pθ (Sǫδ △Vǫδ )d Py = 0
(2.9)
Qǫ
and
lim lim P0 (Sǫδ △Vǫδ ) = 0.
δ→0 ǫ→0
(2.10)
In the proof of Lemma 2.6 we show that the integrals in the right hand-side of
(2.3) with integration domain Qǫ converge to one in probability as ǫ → 0. This
statement is proved both for hypothesis and Bayes alternative (see [2]).
Lemmas 2.1-2.6 implies that, if α(Kǫ ) = α(Lǫ ), then
Z
Z
Z
βθ (Lǫ )d Py (1 + o(1)) = βη0 (Lǫ )d Pη0 (1 + o(1)). (2.11)
βθ (Kǫ )d Py =
Qǫ
Qǫ
Lemma 2.7. There holds
Eη0 βη0 (Lǫ ) = βǫ (Lǫ )(1 + o(1)).
Lemmas 2.2, 2.5, (2.2), (2.11) and Lemma 2.7 imply Theorem 1.1.
(2.12)
ON ASYMPTOTICALLY MINIMAX NONPARAMETRIC DETECTIONOF SIGNAL IN GAUSSIAN WHITE NOISE
5
3. Proof of Lemmas
Proofs of Lemmas 2.2,2.3 and 2.5 are akin to the proofs of similar statements in
[2] and are omitted.
Proof of Lemma 2.1. By straightforward calculations, we get
∞
X
Eηj2 (δ) ≥ ρǫ (1 + δ/2)
(3.1)
j=1
and
Var
∞
X
j=1
ηj2 (δ) < CAǫ ǫ−4 ≍ ρ2ǫ kǫ−1 .
(3.2)
Hence, by Chebyshev inequality, we get
∞
X
P
ηj2 (δ) > ρǫ = 1 + o(1)
(3.3)
j=1
as ǫ → 0. It remains to estimate
r
Pµ (η ∈
/ B2∞
) = P ( max i2r
l1 ≤i≤l2
with
l2
X
ηj2 − P0 (1 − δ1 /2) > P0 δ1 /2) ≤
j=i
Ji = P i2r
l2
X
j=i
l2
X
Ji
(3.4)
i=l1
ηj2 − P0 (1 − δ1 /2) > P0 δ1 /2
To estimate Ji we implement the following Proposition [3]
Proposition 3.1. Let ξ = {ξi }li=1 be Gaussian random vector with i.i.d.r.v.’s ξi ,
E[ξi ] = 0, E[ξi2 ] = 1. Let A ∈ Rl × Rl and Σ = AT A. Then
p
P (||Aξ||2 > tr(Σ) + 2 tr(Σ2 )t + 2||Σ||t) ≤ exp{−t}.
(3.5)
ǫδ
We put Σi = {σlj }kl,j=i
with σjj = j −2r−1 i2r P02r−δ and σlj = 0 if l 6= j.
Let i ≤ kǫ . Then
∞
X
κ4j (δ) < i4r ((kǫ − i)κ4 (δ) + kǫ−4r−1 P0 ) < Ckǫ−1 .
trΣ2i = i4r
(3.6)
j=i
and
Therefore
||Σi || ≤ i2r κ2 < Ckǫ−1 .
(3.7)
q
p
2 tr(Σ2i )t + 2||Σi ||t ≤ C( kǫ−1 t + kǫ−1 t)
(3.8)
1/2
Hence, putting t = kǫ , by Proposition 3.1, we get
kǫ
X
Ji ≤ Ckǫ exp{−Ckǫ1/2 }.
(3.9)
i=1
Let i ≥ kǫ . Then
trΣ2i < Ci−1 ,
Hence, putting t = i
1/2
kǫδ
X
i=kǫ +1
and ||Σi || ≤ Ci−1
(3.10)
, by Proposition 3.1, we get
Ji ≤
kǫδ
X
exp{−Ci1/2 } < exp{−C1 kǫ1/2 }.
i=kǫ +1
Now (3.4), (3.9), (3.11) together implies Lemma 2.1.
(3.11)
6
MIKHAIL ERMAKOV
Proof of Lemma 2.6. By reasoning of the proof of Lemma 4 in [2], Lemma 2.6 will
be proved if we show that
∞
X
(3.12)
P (ηj (δ) + yj γj (δ)ǫ−1 )2 > ρǫ = 1 + o(1)
j=1
and
P sup i2r
i
∞
X
j=i
(ηj (δ) + yj γj (δ)ǫ−1 )2 > ρǫ = 1 + o(1)
(3.13)
where yj , 1 ≤ j < ∞ are distributed by hypothesis or Bayes alternative.
We prove only (3.13) in the case of Bayes alternative. In other cases the reasoning
are similar.
We have
∞
∞
X
X
ηj2 (δ)
(ηj (δ) + yj γj (δ)ǫ−1 )2 = i2r
i2r
j=i
j=i
+i
2r
∞
X
ηj (δ)yj γj (δ)ǫ
−1
+i
2r
∞
X
(3.14)
yj2 γj2 (δ)ǫ−2
= J1i + J2i + J3i .
j=i
j=i
The probability under consideration for the first addendum has been estimated in
Lemma 2.1.
We have
1/2 1/2
(3.15)
J2i ≤ J1i J3i .
Thus it remains to show that, for any C,
∞
X
yj2 γj4 (δ)ǫ−2 > δ/C = o(1)
(3.16)
Pη(δ) sup i2r
i
j=i
as ǫ → 0.
Note that yj = ζj + ǫξj where ζj , yj , 1 ≤ j < ∞ are i.i.d. Gaussian random
variables, Eζj = 0, Eζj2 = κ2j (δ), Eξj = 0, Eξj2 = 1.
Hence, we have
∞
∞
∞
X
X
X
γj4 (δ)ǫ−1 ζj ξj
γj4 (δ)ǫ−2 ζj2 +
yj2 γj4 (δ)ǫ−2 =
+
j=i
j=i
j=i
∞
X
γj4 (δ)ξj2
(3.17)
= I1i + I2i + I3i .
j=i
Since γj2 ǫ−2 = o(1), the estimates for probability of i2r I1i are evident. It suffices
1/2 1/2
to follow the estimates of (3.4). We have I2i ≤ I1i I3i . Thus it remains to show
that, for any C
∞
X
γj4 (δ)ξj2 > δ/C = o(1)
Pη(δ) sup i2r
(3.18)
i
γj2
j=i
κ2j (1+o(1))
as ǫ → 0. Since
=
= o(1), this estimate is also follows from estimates
(3.4).
Proof of Lemma 2.7. By Lemmas 2.2, 2.3 and 2.5, it suffices to show that
∞
∞
X
X
κ2j .
(3.19)
κ2j θj2 =
inf
θ
Denote uk = k
P∞
2r
2
j=k θj .
j=1
j=1
Note that uk ≤ P0 .
ON ASYMPTOTICALLY MINIMAX NONPARAMETRIC DETECTIONOF SIGNAL IN GAUSSIAN WHITE NOISE
7
Then θj2 = uj j −2r − uj+1 (j + 1)−2r . Hence we have
A(θ) =
∞
X
κ2j θj2 = κ2
=κ
2
θj2
θj2 +
j=1
j=1
kǫ
X
kǫ
X
+κ
2
ukǫ kǫ−2r
j=1
∞
X
κ2j (uj j −2r − uj+1 (j + 1)−2r )
j=kǫ
P0
+
r
∞
X
(3.20)
uj (j
−4r−1
−2r−1 −2r
− (j − 1)
j
).
j=kǫ +1
Since j −4r−1 − (j − 1)−2r−1 j −2r is negative, then inf A(θ) is attained for uj = P0
and therefore θj2 = κ2j for j > kǫ .
Thus the problem is reduced to the solution of the following problem
inf κ
θj
if
k
2r
kǫ
X
∞
X
θj2 +
j=1
κ4j
(3.21)
j=kǫ +1
kǫ
X
θj2 +
∞
X
θj2 < P0 ,
j=1
and
2
∞
X
κ2j = ρǫ
j=kǫ +1
1 ≤ j < ∞,
j=k
with θj2 = κ2j for j ≥ kǫ .
It is easy to see that this infimum is attained if κ2j = θj2 = κ2 for j ≤ kǫ .
References
[1] Comminges, L. and Dalalyan, A.S. (2013). Minimax testing of a composite null hypothesis
defined via a quadratic functional in the model of regression. Electronic Journal of Statistics
7 146-190.
[2] Ermakov, M.S. Minimax detection of a signal in a Gaussian white noise. Theory Probab.
Appl., 35 (1990), 667-679.
[3] Hsu D., Kakade S.M., Zang T. 2012 A tail inequality for quadratic forms of subgaussian
random vector. Electronic Commun. Probab. 17 No 52 p.1 - 6.
[4] Ingster, Yu.I., Suslina I.A. (2002). Nonparametric Goodness-of-fit Testing under Gaussian
Models. Lecture Notes in Statistics 169 Springer N.Y.
[5] Johnstone I M 2015 Gaussian estimation. Sequence and wavelet models. Book Draft
http://statweb.stanford.edu/∼imj/
[6] Pinsker M S 1980 Optimal filtering of square integrable signals in Gaussian white noise.
Problems of Information Transmission, 16, 120-133.
[7] Tsybakov A 2009 Introduction to Nonparametric Estimation. Springer Series in Statistics
130 Springer Berlin.
[8] Rivoirard V 2004 Maxisets for linear procedures. Statist. Probab. Lett. 67 267-275
| 10math.ST
|
1
Learning Human Identity from Motion Patterns
arXiv:1511.03908v4 [cs.LG] 21 Apr 2016
Natalia Neverova, Christian Wolf, Griffin Lacey, Lex Fridman, Deepak Chandra,
Brandon Barbello, Graham Taylor
Abstract—We present a large-scale study exploring the capability of temporal deep neural networks to interpret natural
human kinematics and introduce the first method for active
biometric authentication with mobile inertial sensors. At Google,
we have created a first-of-its-kind dataset of human movements,
passively collected by 1500 volunteers using their smartphones
daily over several months. We (1) compare several neural
architectures for efficient learning of temporal multi-modal data
representations, (2) propose an optimized shift-invariant dense
convolutional mechanism (DCWRNN), and (3) incorporate the
discriminatively-trained dynamic features in a probabilistic generative framework taking into account temporal characteristics.
Our results demonstrate that human kinematics convey important information about user identity and can serve as a valuable
component of multi-modal authentication systems. Finally, we
demonstrate that the proposed model can be successfully applied
also in a visual context.
!
y '
x
'y
z
↵y
a
↵x
'z
0
↵z
x
Fig. 1. The accelerometer captures linear acceleration, the gyroscope provides
angular velocity (photo taken from [24]).
Index Terms—Authentication, Biometrics (access control),
Learning, Mobile computing, Recurrent neural networks.
I. I NTRODUCTION
For the billions of smartphone users worldwide, remembering dozens of passwords for all services we need to use
and spending precious seconds on entering pins or drawing
sophisticated swipe patterns on touchscreens becomes a source
of frustration. In recent years, researchers in different fields
have been working on creating fast and secure authentication
alternatives that would make it possible to remove this burden
from the user [29], [6].
Historically, biometrics research has been hindered by the
difficulty of collecting data, both from a practical and legal
perspective. Previous studies have been limited to tightly
constrained lab-scale data collection, poorly representing real
world scenarios: not only due to the limited amount and
variety of data, but also due to essential self consciousness
of participants performing the tasks. In response, we created
an unprecedented dataset of natural prehensile movements
(i.e. those in which an object is seized and held, partly or
wholly, by the hand [20]) collected by 1,500 volunteers over
several months of daily use (Fig. 1).
Apart from data collection, the main challenges in developing a continuous authentication system for smartphones are
(1) efficiently learning task-relevant representations of noisy
Manuscript created: April 22, 2016.
N. Neverova and C. Wolf are with INSA-Lyon, LIRIS, UMR5205, F-69621,
Université de Lyon, CNRS, France. E-mail: [email protected]
G. Lacey and G. Taylor are with the School of Engineering, University of
Guelph, Canada. E-mail: [email protected], [email protected]
L. Fridman is with MIT, Boston, USA. Email: [email protected]
D. Chandra and B. Barbello are with Google, Mountain View, USA. Email:
[email protected], [email protected].
This work was done while N. Neverova, G. Lacey and L. Fridman were at
Google, Mountain View, USA.
inertial data, and (2) incorporating them into a biometrics
setting, characterized by limited resources. Limitations include
low computational power for model adaptation to a new user
and for real-time inference, as well as the absence (or very
limited amount) of “negative” samples.
In response to the above challenges, we propose a noncooperative and non-intrusive method for on-device authentication based on two key components: temporal feature extraction by deep neural networks, and classification via a
probabilistic generative model. We assess several popular deep
architectures including one-dimensional convolutional nets and
recurrent neural networks for feature extraction. However,
apart from the application itself, the main contribution of this
work is in developing a new shift-invariant temporal model
which fixes a deficiency of the recently proposed Clockwork
recurrent neural networks [18] yet retains their ability to
explicitly model multiple temporal scales.
II. R ELATED WORK
Exploiting wearable or mobile inertial sensors for authentication, action recognition or estimating parameters of a
particular activity has been explored in different contexts. Gait
analysis has attracted significant attention from the biometrics community as a non-contact, non-obtrusive authentication
method resistant to spoofing attacks. A detailed overview and
benchmarking of existing state-of-the-art is provided in [23].
Derawi et al. [9], for example, used a smartphone attached to
the human body to extract information about walking cycles,
achieving 20.1% equal error rate.
There exist a number of works which explore the problem of
activity and gesture recognition with motion sensors, including
methods based on deep learning. In [12] and [7], exhaustive
2
overviews of preprocessing techniques and manual feature
extraction from accelerometer data for activity recognition are
given. Perhaps most relevant to this study is [25], the first to
report the effectiveness of RBM-based feature learning from
accelerometer data, and [5], which proposed a data-adaptive
sparse coding framework. Convolutional networks have been
explored in the context of gesture and activity recognition
[11], [34]. Lefebvre et al. [19] applied a bidirectional LSTM
network to a problem of 14-class gesture classification, while
Berlemont et al. [4] proposed a fully-connected Siamese
network for the same task.
We believe that multi-modal frameworks are more likely
to provide meaningful security guarantees. A combination of
face recognition and speech [17], and of gait and voice [32]
have been proposed in this context. Deep learning techniques,
which achieved early success modeling sequential data such
as motion capture [31] and video [16] have shown promise in
multi-modal feature learning [22], [30], [15], [21].
III. A GENERATIVE BIOMETRIC FRAMEWORK
Our goal is to separate a user from an impostor based on a
time series of inertial measurements (Fig. 1). Our method is
based on two components: a feature extraction pipeline which
associates each user’s motion sequence with a collection of
discriminative features, and a biometric model, which accepts
those features as inputs and performs verification. While the
feature extraction component is the most interesting and novel
aspect of our technique, we delay its discussion to Section IV.
We begin by discussing the data format and the biometric
model.
A. Movement data
Each reading (frame) in a synchronized raw input
stream of accelerometer and gyroscope data has the form
{ax , ay , az , ωx , ωy , ωz } ∈ R6 , where a represents linear acceleration, ω angular velocity and x, y, z denote projections
on corresponding axes, aligned with the phone. There are two
important steps we take prior to feature extraction.
Obfuscation-based regularization — it is important to differentiate between the notion of “device” and “user”. In the
dataset we collected (Section VI), each device is assigned
to a single user, thus all data is considered to be authentic.
However, in real-world scenarios such as theft, authentic and
imposter data may originate from the same device.
In a recent study [8], it was shown that under lab conditions
a particular device could be identified by a response of
its motion sensors to a given signal. This happens due to
imperfection in calibration of a sensor resulting in constant
offsets and scaling coefficients (gains) of the output, that can
be estimated by calculating integral statistics from the data.
Formally, the measured output of both the accelerometer and
gyroscope can be expressed as follows [8]:
a = ba + diag(γ a )ã,
ω = bω + diag(γ ω )ω̃,
(1)
where ã and ω̃ are real acceleration and angular velocity
vectors, ba and bω are offset vectors and γ a and γ ω represent
gain errors along each coordinate axes.
To partially obfuscate the inter-device variations and ensure
decorrelation of user identity from device signature in the
learned data representation, we introduce low-level additive
(offset) and multiplicative (gain) noise per training example.
Following [8], the noise vector is obtained by drawing a 12dimensional (3 offset and 3 gain coefficients per sensor) obfuscation vector from a uniform distribution µ ∼ U12 [0.98, 1.02].
Data preprocessing — In addition, we extract a set of
angles α{x,y,z} and ϕ{x,y,z} describing the orientation of
vectors a and ω in the phone’s coordinate system (see Fig. 1),
compute their magnitudes |a| and |ω| and normalize each of
the x, y, z components. Finally, the normalized coordinates,
angles and magnitudes are combined in a 14-dimensional
vector x(t) with t indexing the frames.
B. Biometric model
Relying on cloud computing to authenticate a mobile user
is unfeasible due to privacy and latency. Although this technology is well established for many mobile services, our
application is essentially different from others such as voice
search, as it involves constant background collection of particularly sensitive user data. Streaming this information to the
cloud would create an impermissible threat from a privacy
perspective for users and from a legal perspective for service
providers. Therefore, authentication must be performed on
the device and is constrained by available storage, memory
and processing power. Furthermore, adapting to a new user
should be quick, resulting in a limited amount of training
data for the “positive” class. This data may not be completely
representative of typical usage. For these reasons, a purely
discriminative setting involving learning a separate model per
user, or even fine-tuning a model for each new user would
hardly be feasible.
Therefore, we adapt a generative model, namely a Gaussian
Mixture Model (GMM), to estimate a general data distribution
in the dynamic motion feature space and create a universal background model (UBM). The UBM is learned offline,
i.e. prior to deployment on the phones, using a large amount
of pre-collected training data. For each new user we use a
very small amount of enrollment samples to perform online
(i.e. on-device) adaptation of the UBM to create a client model.
The two models are then used for real time inference of trust
scores allowing continuous authentication.
Universal background model — let y=f ({x(t) })∈RN be
a vector of features extracted from a raw sequence of prehensile movements by one of the deep neural networks described
in Section IV. Probability densities are defined over these feature vectors as a weighted sum of M multi-dimensional Gaussian distributions parameterized by Θ={µi , Σi , πi }, where µi
is a mean vector, Σi a covariance matrix and πi a mixture
coefficient:
p(y|Θ) =
M
X
i=1
Ni (y) = p
πi N (y; µi , Σi ),
1
(2π)N |Σi |
e−
(y−µi )0 Σi −1 (y−µi )
2
(2)
.
(3)
time
synchronized stream of accelerometer and gyroscope data
(set of short sequences)
synchronized stream of accelerometer and gyroscope data
(long sequences)
3
distribution of the extracted
dynamic features
distribution of the extracted
dynamic features
IV. L EARNING EFFECTIVE AND EFFICIENT
REPRESENTATIONS
discriminative
pretraining
temporal
units
discriminative
pretraining
Conv3
Conv3
Conv2
Conv2
Conv1
Conv1
time, t
input {a, !}
input {a, !}
(a)
(b)
Fig. 2. Learning data representations: (a) static convnet directly operating on
sequences, aggregating temporal statistics by temporal pooling; (b) explicitly
modeling temporal transitions with recurrent connections.
The UBM p(y|ΘUBM ) is learned by maximising the likelihood
of feature vectors extracted from the large training set using
the expectation-maximisation (EM) algorithm.
The client model p(y|Θclient ) is adapted from the UBM.
Both models share the same weights and covariance matrices
to avoid overfitting from a limited amount of enrollment data.
Along the lines of [26], maximum a posteriori (MAP) adaptation of mean vectors for a given user is performed. This has
an immediate advantage over creating an independent GMM
for each user, ensuring proper alignment between the welltrained background model and the client model by updating
only a subset of parameters that are specific to the given user.
In particular, given a set of Q enrollment samples {yq } from
the new device, we create a client-specific update to the mean
of each mixture component i as follows:
Ei ({yq }) =
where ni =
Q
1 X
P r(i|yq )yq ,
ni q=1
(4)
Q
X
πi pi (yq )
P r(i|yq ), P r(i|yq ) = PM
. (5)
q=1
j=1 πj pj (yq )
Finally, the means of all Gaussian components are updated
according to the following rule:
µ̂i = αi Ei ({yq }) + (1 − αi )µi ,
where αi =
ni
, (6)
ni + r
where r is a relevance factor balancing the background and
client models. In our experiments, r is held fixed.
Scoring — given a set of samples Y ={ys } from a given
device, authenticity is estimated by scoring the feature vectors
against the UBM and the client model, thresholding the loglikelihood ratio:
Λ(Y ) = log p(Y |Θclient ) − log p(Y |ΘUBM ).
(7)
As a final step, zt-score normalization [2] is performed to
compensate for inter-session and inter-person variations and
reduce the overlap between the distribution of scores from
authentic users and impostors.
Learning effective and efficient data representations is key
to our entire framework since its ability to perform in the realworld is defined by such criteria as latency, representational
power of extracted features and inference speed of the feature
extractor. The first two conditions are known to contradict each
other as performance of a standalone feature typically grows
with integration time [21].
Two paradigms which strike a balance between representational power and speed have dominated the feature learning
landscape in recent years. These are multi-scale temporal
aggregation via 1-dimensional convolutional networks Fig. 2a,
and explicit modeling of temporal dependencies via recurrent
neural networks Fig. 2b.
The former model, popular in speech recognition [13],
involves convolutional learning of integrated temporal statistics from short and long sequences of data (referred to as
“short-term” and “long-term” convnets). Short-term architectures produce outputs at relatively high rate (1 Hz in our
implementation) but fail to model context. Long-term networks
can learn meaningful representations at different scales, but
suffer from a high degree of temporal inertia and do not
generalize to sequences of arbitrary length.
Recurrent models which explicitly model temporal evolutions can generate low-latency feature vectors built in the
context of previously observed user behavior. The dynamic
nature of their representations allow for modeling richer temporal structure and better discrimination among users acting
under different conditions. There have been a sufficiently large
number of neural architectures proposed for modeling temporal dependencies in different contexts: the baseline methods
compared in this work are summarized in Fig. 3c. The rest
of this section provides a brief description of these models.
Then, Section V introduces a new shift-invariant model based
on modified Clockwork RNNs [18].
All feature extractors are first pretrained discriminatively
for a multi-device classification task, then, following removal
of the output layer the activations of the penultimate hidden
layer are provided as input (which we denote, for conciseness1 ,
by y) to the generative model described in Section III.
The final outputs of the background and client models are
integrated over a 30 sec window. Accordingly, after 30 sec the
user is either authenticated or rejected.
A. Classical RNN and Clockwork RNN
The vanilla recurrent neural network (RNN) is governed by
the update equation
h(t) = ψ(Uh(t−1) + Wx(t) ),
(8)
where x is the input, h(t) denotes the network’s hidden state
at time t, W and U are feed-forward and recurrent weight
matrices, respectively, and ψ is a nonlinear activation function,
typically tanh. The output is produced combining the hidden
1 This decision was made to avoid introducing notation to index hidden
layers, as well as simplify and generalize the presentation in the previous
section, where the y are taken as generic temporal features.
·
=
h(t)
(t 1)
output, õ
õ
V
h(t
U
(t)
V
õ
h
(t+1)
1)
h
(t)
h
⨉
(t)
i(t) g
h(t
W
x(t
1)
W
W
1)
f (t) o(t)
(b)
2)
õ(t
V
1)
output, õ(t) õ(t+1)
V
V
k=2
U
U
h(t
2)
W
W
x(t
2)
x(t
U
h(t
1)
1)
W
U
U
h(t)
input, x(t)
W
h(t+1)
W
x(t+1)
U
1 z
(t)
⨉ h̃(t)
(t)
r(t)
h(t
U
⨉
⨉
=
h(t)
õ(t+2)
V
(t)
1)
1
2
4
8
16
x(t)
⨉
⨉
z
·
h(t)
=
z
h(t)
h(t
1)
(a)
õ(t
V
⨉
⨉
c(t
x(t+1)
input, x(t)
1
2
4
8
16
c(t)
(t+1)
U
U
U
(t)
⨉
V
h(t
U
x(t)
U
+
1)
·
4
x(t)
W
+
1)
·
·
x(t)
W
+
·
x(t)
H
W
Fig. 4. Updates made by the Clockwork RNN [18] (top) and our proposed
Dense CWRNN (bottom). Units and weights colored in blue are the ones
updated or read at the example time step t=6.
k=1
k=0
h(t+2)
x(t+2)
(c)
Fig. 3. Temporal models: (a) a basic recurrent unit; (b) an LSTM unit [14];
(c) Clockwork RNN [18] with 3 bands and a base of 2; Increasing k indicates
lower operating frequency. Grey color. indicates inactivity of a unit.
state in a similar way, o(t) = φ(Vh(t) ), where V is a weight
matrix.
One of the main drawbacks of this model is that it operates
at a predefined temporal scale. In the context of free motion
which involves large variability in speed and changing intervals between typical gestures, this may be a serious limitation.
The recently proposed Clockwork RNN (CWRNN) [18] operates at several temporal scales which are incorporated in a
single network and trained jointly. It decomposes a recurrent
layer into several bands of high frequency (“fast”) and low
frequency (“slow”) units (see Fig. 3c). Each band is updated
at its own pace. The size of the step from band to band
typically increases exponentially (which we call exponential
update rule) and is defined as nk , where n is a base and k is
the number of the band.
In the CWRNN, fast units (shown in red) are connected
to all bands, benefitting from the context provided by the
slow bands, while the low frequency units ignore noisy high
frequency oscillations. Equation (8) from classical RNNs is
modified, leading to a new update rule for the k-th band of
output h at iteration t as follows:
(
(t−1)
ψ U(k)hk
+W(k)x(t) if (t mod nk )=0,
(t)
hk =
(t−1)
hk
otherwise.
where U(k) and W(k) denote rows k from matrices U
and W. Matrix U has an upper triangular structure, which
corresponds to the connectivity between frequency bands. This
equation is intuitively explained in the top part of Fig. 4,
inspired from [18]. Each line corresponds to a band. At time
step t=6 for instance, the first two bands k = 0 and k = 1 get
updated. The triangular structure of the matrix results in each
band getting updated from bands of lower (or equal) frequency
only. In Fig. 4, not active rows are also shown as zero (black)
in U and W. In addition to multi-scale dynamics, creating
sparse connections (high-to-low frequency connections are
missing) reduces the number of free parameters and inference
complexity.
B. Long Short-Term Memory
Long Short-Term Memory (LSTM) networks [14], another variant of RNNs, and their recent convolutional extensions [10], [27] have proven to be, so far, the best performing
models for learning long-term temporal dependencies. They
handle information from the past through additional gates,
which regulate how a memory cell is affected by the input
signal. In particular, an input gate allows to add new memory
to the cell’s state, a forget gate resets the memory and an
output gate regulates how gates at the next step will be affected
by the current cell’s state.
The basic unit is composed of input i, output o, forget f ,
and input modulation g gates, and a memory cell c (see
Fig. 3b). Each element is parameterized by corresponding
feed-forward (W) and recurrent (U) weights and bias vectors.
Despite its effectiveness, the high complexity of this architecture may appear computationally wasteful in the mobile
setting. Furthermore, the significance of learning long-term
dependencies in the context of continuous mobile authentication is compromised by the necessity of early detection of
switching between users. Due to absence of annotated ground
truth data for these events, efficient training of forgetting
mechanisms would be problematic.
C. Convolutional learning of RNNs
Given the low correlation of individual frames with user
identity, we found it strongly beneficial to make the input layer
convolutional regardless of model type, thereby forcing earlier
fusion of temporal information. To simplify the presentation,
we have not made convolution explicit in the description of
the methods above, however, it can be absorbed into the inputto-hidden matrix W.
5
õ(t
V
2)
õ(t
V
1)
output, õ(t) õ(t+1)
V
V
õ(t+2)
V
k=2
U
h(t
U
U
2)
W
x
(t 2)
h(t
W
x(t
1)
1)
W
input, x
U
h(t)
(t)
W
x(t+1)
U
h(t+1)
W
U
k=1
k=0
h(t+2)
x(t+2)
Fig. 5. Proposed dense clockwork RNN with the same parameters as the
original clockwork RNN shown in Fig. 3a.
V. D ENSE CONVOLUTIONAL CLOCKWORK RNN S
Among the existing temporal models we considered, the
clockwork mechanisms appear to be the most attractive due
to low computational burden associated with them in combination with their high modeling capacity. However, in practice,
due to inactivity of “slow” units for long periods of time, they
cannot respond to high frequency changes in the input and
produce outputs which, in a sense, are stale. Additionally, in
our setting, where the goal is to learn dynamic data representations serving as an input to a probabilistic framework,
this architecture has one more weakness which stems from
the fact that different bands are active at any given time step.
The network will respond differently to the same input stimuli
applied at different moments in time. This “shift-variance”
convolutes the feature space by introducing a shift-associated
dimension.
In this work, we propose a solution to both issues, namely
“twined” or “dense” clockwork mechanisms (DCWRNN, see
Fig. 5), where during inference at each scale k there exist nk
parallel threads shifted with respect to each other, such that at
each time a unit belonging to one of the threads fires, updating
its own state and providing input to the higher frequency units.
All weights between the threads belonging to the same band
are shared, keeping the overall number of parameters in the
network the same as in the original clockwork architecture.
Without loss of generality, and to keep the notation uncluttered
of unnecessary indices, in the following we will describe
a network with a single hidden unit hk per band k. The
generalization to multiple units per band is straightforward,
and the experiments were of course performed with the more
general case.
The feedforward pass for the whole dense clockwork layer
(i.e. all bands) can be given as follows:
h(t) = ψ Wx(t) + ∆(UH)
(9)
k
K
where H=[ h(t−1) . . . h(t−n ) . . . h(t−n ) ] is a matrix concatenating the history of hidden units and we define ∆(·) as
an operator on matrices returning its diagonal elements in
a column vector. The intuition for this equation is given in
Fig. 4, where we compare the update rules of the original
CWRNN and the proposed DCWRNN using an example of a
network with 5 hidden units each associated with one of K=5
base n=2 bands. To be consistent, we employ the same
matrix form as in the original CWRNN paper [18]) and show
components, which are inactive at time t, in dark gray. As
mentioned in Section IV-A, in the original CWRNN, at time
instant t=6, for instance, only unit h1 and h2 are updated,
i.e. the first two lines in Fig. 4. In the dense network, all
hidden units hk are updated at each moment in time.
In addition, what was vector of previous hidden
states h(t−1) is replaced with a lower triangular “history”
matrix H of size K×K which is obtained by concatenating
several columns from the history of activations h. Here,
K is the number of bands. Time instances are not sampled consecutively, but strided in an exponential range, i.e.
n, n2 , . . . nK . Finally, the diagonal elements of the dot product
of two triangular matrices form the recurrent contribution to
the vector h(t) . The feedforward contribution is calculated in
the same way as in a standard RNN.
The practical implementation of the lower-triangular matrix
containing the history of previous hidden activations in the
DCWRNN requires usage of P
an additional memory buffer
K
whose size can be given as m= k=1 |hk |(nk−1 −1), whereas
here we have stated the general case of |hk |≥1 hidden units
belonging to band k.
During training, updating all bands at a constant rate is
important for preventing simultaneous overfitting of highfrequency and underfitting of low-frequency bands. In practice
it leads to a speedup of the training process and improved
performance. Finally, due to the constant update rate of all
bands in the dense network, the learned representations are
invariant to local shifts in the input signal, which is crucial in
unconstrained settings when the input is unsegmented. This is
demonstrated in Section VII.
VI. DATA COLLECTION
The dataset introduced in this work is a part of a more general multi-modal data collection effort performed by Google
ATAP, known as Project Abacus. To facilitate the research, we
worked with a third party vendor’s panel to recruit and obtain
consent from volunteers and provide them with LG Nexus 5
research phones which had a specialized read only memory
(ROM) for data collection. Volunteers had complete control
of their data throughout its collection, as well as the ability
to review and delete it before sharing for research. Further,
volunteers could opt out after the fact and request that all of
their data be deleted. The third party vendor acted as a privacy
buffer between Google ATAP and the volunteers.
The data corpus consisted of 27.62 TB of smartphone
sensor signals, including images from a front-facing camera,
touchscreen, GPS, bluetooth, wifi, cell antennae, etc. The
motion data was acquired from three sensors: accelerometer,
gyroscope and magnetometer. This study included approximately 1,500 volunteers using the research phones as their
primary devices on a daily basis. The data collection was
completely passive and did not require any action from the
volunteers in order to ensure that the data collected was
representative of their regular usage.
Motion data was recorded from the moment after the phone
was unlocked until the end of a session (i.e., until it is
locked again). For this study, we set the sampling rate for
6
the accelerometer and gyroscope sensors to 200 Hz and for
the magnetometer to 5 Hz. However, to prevent the drain of
a battery, the accelerometer and gyro data were not recorded
when the device was at rest. This was done by defining two
separate thresholds for signal magnitude in each channel. Finally, accelerometer and gyroscope streams were synchronized
on hardware timestamps.
Even though the sampling rate of the accelerometer and
the gyroscope was set to 200 Hz for the study, we noticed
that intervals between readings coming from different devices
varied slightly. To eliminate these differences and decrease
power consumption, for our research we resampled all data to
50 Hz. For the following experiments, data from 587 devices
were used for discriminative feature extraction and training
of the universal background models, 150 devices formed the
validation set for tuning hyperparameters, and another 150
devices represented “clients” for testing.
VII. E XPERIMENTAL RESULTS
In this section, we use an existing but relatively small
inertial dataset to demonstrate the ability of the proposed
DCWRNN to learn shift-invariant representations. We then
describe our study involving a large-scale dataset which was
collected “in the wild”.
A. Visualization: HMOG dataset
To explore the nature of inertial sensor signals, we performed a preliminary analysis on the HMOG dataset [33]
containing similar data, but collected in constrained settings
as a part of a lab study. This data collection was performed
with the help of 100 volunteers, each performing 24 sessions
of predefined tasks, such as reading, typing and navigation,
while sitting or walking.
Unfortunately, direct application of the whole pipeline to
this corpus is not so relevant due to 1) absence of task-totask transitions in a single session and 2) insufficient data
to form separate subsets for feature learning, the background
model, client-specific subsets for enrollment, and still reserve
a separate subset of “impostors” for testing that haven’t been
seen during training.
However, a detailed visual analysis of accelerometer and
gyroscope streams has proven that the inertial data can be
seen as a combination of periodic and quasi-periodic signals
(from walking, typing, natural body rhythms and noise), as
well non-periodic movements. This observation additionally
motivates the clockwork-like architectures allowing for explicit modelling of periodic components.
In this subsection, we describe the use of HMOG data to
explore the shift-invariance of temporal models that do not
have explicit reset gates (i.e. RNN, CWRNN and DCWRNN).
For our experiment, we randomly selected 200 sequences
of normalized accelerometer magnitudes and applied three
different networks each having 8 hidden units and a single
output neuron. All weights of all networks were initialized
randomly from a normal distribution with a fixed seed. For
both clockwork architectures we used a base 2 exponential
setting rule and 8 bands.
TABLE I
P ERFORMANCE AND MODEL COMPLEXITY OF THE FEATURE
EXTRACTORS . T HESE RESULTS ASSUME ONE USER PER DEVICE AND
ACCURACY IS DEFINED BASED ON WHETHER OR NOT THE USER IS IN THE
TOP 5% OF CLASSES ACCORDING TO THE OUTPUT DISTRIBUTION .
Feature learning: evaluation
Model
ST Convnet
LT Convnet
Conv-RNN
Conv-CWRNN
Conv-LSTM
Conv-DCWRNN
Accuracy, %
# parameters
37.13
56.46
64.57
68.83
68.92
69.41
6 102 137
6 102 137
1 960 295
1 964 254
1 965 403
1 964 254
Finally, for each network we performed 128 runs (i.e. 2K−1 )
on a shifted input: for each run x the beginning of the
sequence was padded with x−1 zeros. The resulting hidden
activations were then shifted back to the initial position and
superimposed. Fig. 6 visualizes the hidden unit traces for two
sample sequences from the HMOG dataset, corresponding to
two different activities: reading while walking and writing
while sitting. The figure shows that the RNN and the dense
version of the clockwork network can be considered shiftinvariant (all curves overlap almost everywhere, except for
minor perturbance at the beginning of the sequence and around
narrow peaks), while output of the CWRNN is highly shiftdependent.
For this reason, in spite of their atractiveness in the context
of multi-scale periodic and non-periodic signals, the usage
of the CWRNN for the purpose of feature learning from
unsegmented data may be suboptimal due to high shiftassociated distortion of learned distributions, which is not the
case for DCWRNNs.
B. Large-scale study: Google Abacus dataset
We now evaluate our proposed authentication framework
on the real-world dataset described in Section VI. Table IV in
the Appendix provides architectural hyper-parameters chosen
for two 1-d Convnets (abbreviated ST and LT for short and
log-term) as well as the recurrent models. To make a fair
comparison, we set the number of parameters to be approximately the same for all of the RNNs. The ST Convnet is
trained on sequences of 50 samples (corresponding to a 1 s data
stream), LT Convnets take as input 500 samples (i.e. 10 s). All
RNN architectures are trained on sequences of 20 blocks of
50 samples each with 50% of inter-block overlap to ensure
smooth transitions between blocks (therefore, also a 10 s
duration). For the dense and sparse clockwork architectures
we set the number of bands to 3 with a base of 2. All layers
in all architectures use tanh activations. During training, the
networks produce a softmax output per block in the sequence,
rather than only for the last one. The mean per-block negative
log likelihood loss taken over all blocks is minimized.
The dimensionality of the feature space produced by each of
the networks is PCA-reduced to 100. GMMs with 256 mixture
components are trained for 100 iterations after initialization
with k-means (100 iterations). MAP adaptation for each device
7
Input
RNN
CWRNN
DCWRNN
Fig. 6. On spatial invariance. From left to right: original sequence and traces of RNN, CWRNN and DCWRNN units. The first row: reading while walking,
the second row: typing while sitting.
TABLE II
P ERFORMANCE OF THE GMM- BASED BIOMETRIC MODEL USING
DIFFERENT TYPES OF DEEP NEURAL ARCHITECTURES . EER IS GIVEN ON
THE VALIDATION SET, WHILE HTER IS ESTIMATED ON THE FINAL TEST
SET USING THE SAME THRESHOLD .
User authentication: evaluation
Model
EER, %
HTER, %
Raw features
ST Convnet
LT Convnet
Conv-RNN
Conv-CWRNN
Conv-LSTM
Conv-DCWRNN
Conv-DCWRNN, zt-norm
Conv-DCWRNN (per device)
Conv-DCWRNN (per session)
36.21
32.44
28.15
22.32
21.52
21.13
20.01
18.17
15.84
8.82
42.17
34.89
29.01
22.49
21.92
21.41
20.52
19.29
16.13
9.37
is performed in 5 iterations with a relevance factor of 4
(set empirically). For zt-score normalization, we exploit data
from the same training set and create 200 t-models and
200 z-sequences from non-overlapping subsets. Each t-model
is trained based on UBM and MAP adaptation. All hyperparameters were optimized on the validation set.
The networks are trained using stochastic gradient descent,
dropout in fully connected layers, and negative log likelihood
loss. In the temporal architectures we add a mean pooling
layer before applying the softmax. Each element of the input
is normalized to zero mean and unit variance. All deep nets
were implemented with Theano [3] and trained on 8 Nvidia
Tesla K80 GPUs. UBM-GMMs were trained with the Bob
toolbox [1] and did not employ GPUs.
Feature extraction — we first performed a quantitative
evaluation of the effectiveness of feature extractors alone as
a multi-class classification problem, where one class corresponds to one of 587 devices from the training set. This way,
one class is meant to correspond to one “user”, which is equal
to “device” in the training data (assuming devices do not
change hands). To justify this assumption, we manually annotated periods of non-authentic usage based on input from the
smartphone camera and excluded those sessions from the test
and training sets. Experiments showed that the percentage of
such sessions is insignificant and their presence in the training
data has almost no effect on the classification performance.
Note that for this test, the generative model was not considered and the feature extractor was simply evaluated in terms of
classification accuracy. To define accuracy, we must consider
that human kinematics sensed by a mobile device can be
considered as a weak biometric and used to perform a soft
clustering of users in behavioral groups. To evaluate the quality
of each feature extractor in the classification scenario, for
each session we obtained aggregated probabilities over target
classes and selected the 5% of classes with highest probability
(in the case of 587 classes, the top 5% corresponds to the
29 classes). After that, the user behavior was considered to
be interpreted correctly if the ground truth label was among
them.
The accuracy obtained with each type of deep network
with its corresponding number of parameters is reported in
Table I. These results show that the feed forward convolutional architectures generally perform poorly, while among the
temporal models the proposed dense clockwork mechanism
Conv-DCWRNN appeared to be the most effective, while
the original clockwork network (Conv-CWRNN) was slightly
outperformed by the LSTM.
Authentication evaluation — when moving to the binary
authentication problem, an optimal balance of false rejection and false acceptance rates, which is not captured by
classification accuracy, becomes particularly important. We
use a validation subset to optimize the generative model for
the minimal equal error rate (EER). The obtained threshold
value θEER is then used to evaluate performance on the
test set using the half total error rate (HTER) as a criterion:
HTER = 1/2[FAR(θEER ) + FRR(θEER )], where FAR and
FRR are false acceptance and false rejection rates, respectively.
For the validation set, we also provide an average of per-device
8
TABLE III
P ERFORMANCE OF THE PROPOSED DCWRNN ARCHITECTURE ON THE
Chalearn 2014 Looking at People dataset ( MOCAP MODALITY ). N ETWORK
PARAMETERS : INPUT 183×9, CONV. LAYER 25×3×1, 2 FULLY
CONNECTED LAYERS WITH 700 UNITS , THE RECURRENT LAYER
(RNN-280, CWRNN-300, DCWRNN-300, S MALL LSTM-88, L ARGE
LSTM-300), 21 OUTPUT CLASS .
ChaLearn 2014: sequential learning
Model
Jaccard Index
Accuracy
N parameters
Single network [21]
Ensemble [21]
0.827
0.831
91.62
91.96
1 384 621
4 153 863
Conv-RNN
Small Conv-LSTM
Large Conv-LSTM
Conv-CWRNN
Conv-DCWRNN
0.826
0.815
0.825
0.834
0.841
91.79
91.50
91.89
92.38
93.02
3 974 581
3 976 863
4 900 621
3 972 496
3 972 496
and per-session EERs (obtained by optimizing the threshold
for each device/session separately) to indicate the upper bound
of performance in the case of perfect score normalization (see
italicized rows in Table II).
An EER of 20% means that 80% of the time the correct user
is using the device, s/he is authenticated, only by the way s/he
moves and holds the phone, not necessarily interacting with
it. It also means that 80% of the time the system identifies the
user, it was the correct one. These results align well with the
estimated quality of feature extraction in each case and show
that the context-aware features can be efficiently incorporated
in a generative setting.
To compare the GMM performance with a traditional approach of retraining, or finetuning a separate deep model for
each device (even if not applicable in a mobile setting), we
randomly drew 10 devices from the validation set and replaced
the output layer of the pretrained LSTM feature extractor with
a binary logistic regression. The average performance on this
small subset was 2% inferior with respect to the GMM, due
to overfitting of the enrollment data and poor generalization to
unobserved activities. This is efficiently handled by mean-only
MAP adaptation of a general distribution in the probabilistic
setting.
Another natural question is whether the proposed model
learns something specific to the user “style” of performing
tasks rather than a typical sequence of tasks itself. To explore
this, we performed additional tests by extracting parts of each
session where all users interacted with the same application (a
popular mail client, a messenger and a social network application). We observed that the results were almost identical to
the ones previously obtained on the whole dataset, indicating
low correlation with a particular activity.
VIII. M ODEL ADAPTATION FOR A VISUAL CONTEXT
Finally, we would like to stress that the proposed DCWRNN
framework can also be applied to other sequence modeling
tasks, including the visual context. The described model is
not specific to the data type and there is no particular reason
why it cannot be applied to the general human kinematic
problem (such as, for example, action or gesture recognition
from motion capture).
To support this claim, we have conducted additional tests
of the proposed method within a task of visual gesture
recognition. Namely, we provide results on the motion capture
(mocap) modality of the ChaLearn 2014 Looking at People gesture dataset [28]. This dataset contains about 14000
instances of Italian conversational gestures with the aim to
detect, recognize and localize gestures in continuous noisy
recordings. Generally, this corpus comprises multimodal data
captured with the Kinect and therefore includes RGB video,
depth stream and mocap data. However, only the last channel
is used in this round of experiments. The model evaluation
is performed using the Jaccard index, penalizing for errors in
classification as well as imprecise localization.
Direct application of the GMM to gesture recognition is
suboptimal (as the vocabulary is rather small and defined in
advance), therefore, in this task we perform end-to-end discriminative training of each model to evaluate the effectiveness
of feature extraction with the Dense CWRNN model.
In the spirit of [21] (the method ranked firstst in the
ECCV 2014 ChaLearn competition), we use the same skeleton
descriptor as input. However, as in the described authentication
framework, the input is fed into a convolutional temporal
architecture instead of directly concatenating frames in a
spatio-temporal volume. The final aggregation and localization
step correspond to [21]. Table III reports both the Jaccard
index and per-sequence classification accuracy and shows that
in this application, the proposed DCWRNN also outperforms
the alternative solutions.
IX. C ONCLUSION
From a modeling perspective, this work has demonstrated
that temporal architectures are particularly efficient for learning of dynamic features from a large corpus of noisy temporal
signals, and that the learned representations can be further
incorporated in a generative setting. With respect to the
particular application, we have confirmed that natural human
kinematics convey necessary information about person identity
and therefore can be useful for user authentication on mobile
devices. The obtained results look particularly promising,
given the fact that the system is completely non-intrusive and
non-cooperative, i.e. does not require any effort from the user’s
side.
Non-standard weak biometrics are particularly interesting
for providing the context in, for example, face recognition or
speaker verification scenarios. Further augmentation with data
extracted from keystroke and touch patterns, user location,
connectivity and application statistics (ongoing work) may
be a key to creating the first secure non-obtrusive mobile
authentication framework.
Finally, in the additional round of experiments, we have
demonstrated that the proposed Dense Clockwork RNN can
be successfully applied to other tasks based on analysis of
sequential data, such as gesture recognition from visual input.
A PPENDIX
In this section, we provide additional detail for reproducability that was not provided in the main text.
9
TABLE IV
H YPER - PARAMETERS : VALUES IN PARENTHESES ARE FOR SHORT- TERM (ST) C ONVNETS WHEN DIFFERENT FROM LONG - TERM (LT).
Layer
Filter size / # of units
Pooling
Filter size / # of units
Convolutional feature learning
Input
Conv1
Conv2
Conv3
FCL1
FCL2
Recurrent
500×14 (50 × 14)
25×9×1
25×9×1
25×9×1
2000
1000
-
8×1 (2×1)
4×1 (1×1)
1×1
-
Hyper-parameter selection
Table IV provides the complete set of hyper-parameters that
were chosen based on a held-out validation set. For convolutional nets, we distinguish between convolutional layers (Conv,
which include pooling) and fully-connected layers (FCL). For
recurrent models, we report the total number of units (in the
case of CWRNN and DCWRNN, over all bands).
Details on zt-normalization
Here we provide details on the zt-normalization that were
not given in Section 3. Recall that we estimate authenticity,
given a set of motion features by scoring the features against
a universal background model (UBM) and client model.
Specifically, we threshold the log-likelihood ratio in Eq. 7.
An offline z-step (zero normalization) compensates for intermodel variation by normalizing the scores produced by each
client model to have zero mean and unit variance in order to
use a single global threshold:
Λz (Y |Θclient ) =
Λ(Y ) − µ(Z|Θclient )
,
σ(Z|Θclient )
(10)
where Y is a test session and Z is a set of impostor sessions.
Parameters are defined for a given user once model enrollment
is completed. Then, the T -norm (test normalization) compensates for inter-session differences by scoring a session against
a set of background T -models.
Λzt (Y ) =
Λz (Y |Θclient ) − µz (Y |ΘT )
.
σz (Y |ΘT )
(11)
The T-models are typically obtained through MAP-adaptation
from the universal background model in the same way as
all client models, but using different subsets of the training
corpus. The Z-sequences are taken from a part of the training
data which is not used by the T-models.
ACKNOWLEDGMENT
The authors would like to thank their colleagues Elie
Khoury, Laurent El Shafey, Sébastien Marcel and Anupam
Das for valuable discussions.
R EFERENCES
[1] A. Anjos, L. El Shafey, R. Wallace, M. Günther, C. McCool, and
S. Marcel. Bob: a free signal processing and machine learning toolbox
for researchers. ACMMM, 2012.
Pooling
Sequential feature learning
10×50×14
25×7×1
25×7×1
25×7×1
RNN 894, LSTM 394,
CWRNN and DCWRNN 1000
2×1
2×1
1×1
-
[2] R. Auckenthaler, M. Carey, and H. Lloyd-Thomas. Score normalization for text-independent speaker verification systems. Digital Signal
Processing 10 (1), 42-54, 2000.
[3] J. Bergstra et al. Theano: a CPU and GPU math expression compiler. In
Proceedings of the Python for Scientific Computing Conference (SciPy),
2010.
[4] S. Berlemont, G. Lefebvre, S. Duffner, and C. Garcia. Siamese Neural
Network based Similarity Metric for Inertial Gesture Classification and
Rejection. In FG, 2015.
[5] S. Bhattacharya, P. Nurmi, N. Hammerla, and T. Plotz. Using unlabeled
data in a sparse-coding framework for human activity recognition. In
Pervasive and Mobile Computing, 2014.
[6] C. Bo, L. Zhang, X.-Y. Li, Q. Huang, and Y. Wang. Silentsense: Silent
user identification via touch and movement behavioral biometrics. In
MobiCom, 2013.
[7] A. Bulling, U. Blanke, and B. Schiele. A tutorial on human activity
recognition using body-worn inertial sensors. In ACM Computing
Surveys, 2014.
[8] A. Das, N. Borisov, and M. Caesar. Exploring ways to mitigate sensorbased smartphone fingerprinting. arXiv:1503.01874, 2015.
[9] M. O. Derawi, C. Nickel, P. Bours, and C. Busch. Unobtrusive UserAuthentication on Mobile Phones Using Biometric Gait Recognition. In
IIH-MSP, 2010.
[10] J. Donahue et al. Long-term recurrent convolutional networks for visual
recognition and description. arXiv:1503.01874, 2015.
[11] S. Duffner, S. Berlemont, G. Lefebre, and C. Garcia. 3D gesture
classification with convolutional neural networks. In ICASSP, 2014.
[12] D. Figo, P. Diniz, D. Ferreira, and J. Cardoso. Preprocessing techniques
for context recognition from accelerometer data. In Pervasive and
Ubiquituous Computing, 2010.
[13] A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen,
R. Prenger, S. Satheesh, S. Sengupta, A. Coates, et al. Deepspeech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567,
2014.
[14] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural
Computation, 1997.
[15] S. E. Kahou, C. Pal, X. Bouthillier, P. Froumenty, c. Gülçehre, R. Memisevic, P. Vincent, A. Courville, and Y. Bengio. Combining modality
specific deep neural networks for emotion recognition in video. In ICMI,
2013.
[16] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. FeiFei. Large-scale video classification with convolutional neural networks.
In CVPR, 2014.
[17] E. Khoury, L. El Shafey, C. McCool, M. Gunther, and S. Marcel.
Bi-Modal Biometric Authentication on Mobile Phones in Challenging
Conditions. In Image and Vision Computing, 2014.
[18] J. Koutnik, K. Greff, F. Gomez, and J. Schmidhuber. A clockwork rnn.
ICML, 2014.
[19] G. Lefebvre, S. Berlemont, F. Mamalet, and C. Garcia. BLSTM-RNN
Based 3D Gesture Classification. In ICANN, 2013.
[20] J. Napier. The prehensile movements of the human hand. In Journal of
Bone and Joint Surgery, 1956.
[21] N. Neverova, C. Wolf, G. W. Taylor, and F. Nebout. Moddrop: adaptive
multi-modal gesture recognition. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 2015. To appear.
[22] J. Ngiam, A. Khosla, M. Kin, J. Nam, H. Lee, and A. Y. Ng. Multimodal
deep learning. In ICML, 2011.
[23] C. Nickel, H. Brandt, and C. Busch. Benchmarking the performance
of svms and hmms for accelerometer-based biometric gait recognition.
ISSPIT, 2011.
[24] S. Nilsson. https://flic.kr/p/qndn43, license cc by-sa 2.0.
[25] T. Plotz, N. Y. Hammerla, and P. Olivier. Feature learning for activity
10
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
recognition in ubiquitous computing. In 22nd International Joint
Conference on Artificial Intelligence, 2011.
D. A. Reynolds, T. F. Quatieri, and R. B. Dunn. Speaker verification
using adapted gaussian mixture models. Digital Signal Processing 10,
19-41, 2000.
T. N. Sainath, O. Vinyals, A. Senior, and H. Sak. Convolutional, long
short-term memory, fully connected deep neural networks. ICASSP,
2015.
S. Escalera, X. Baró, J. Gonzàlez, M. Bautista, M. Madadi, M. Reyes, V.
Ponce, H. Escalante, J. Shotton, and I. Guyon. ChaLearn Looking at
People Challenge 2014: Dataset and Results. In ECCVW, 2014.
Z. Sitova, J. Sedenka, Q. Yang, G. Peng, G. Zhou, P. Gasti, and
K. Balagani. HMOG: A New Biometric Modality for Continuous
Authentication of Smartphone Users. In arXiv:1501.01199, 2015.
N. Srivastava and R. Salakhutdinov. Multimodal learning with Deep
Boltzmann Machines. In NIPS, 2013.
G. Taylor, G. Hinton, and S. Roweis. Two distributed-state models for
generating high-dimensional time series. Journal of Machine Learning
Research, 12(Mar):1025–1068, 2011.
E. Vildjiounaite et al. Unobtrusive multimodal biometrics for ensuring
privacy and information security with personal devices. In Pervasive
Computing, 2006.
Q. Yang, G. Peng, D. T. Nguyen, X. Qi, G. Zhou, Z. Sitova, P. Gasti,
and K. S. Balagani. A multimodal data set for evaluating continuous
authentication performance in smartphones. In SenSys. 2014.
M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu, and
J. Zhang. Convolutional neural networks for human activity recognition
using mobile sensors. MobiCASE, 2014.
Natalia Neverova is a PhD candidate at INSA de
Lyon and LIRIS (CNRS, France) working in the area
of gesture and action recognition with emphasis on
multi-modal aspects and deep learning methods. She
is advised by Christian Wolf and Graham Taylor,
and her research is a part of Interabot project in
partnership with Awabot SAS. She was a visiting
researcher at University of Guelph in 2014 and at
Google in 2015.
Christian Wolf is assistant professor at INSA de
Lyon and LIRIS, CNRS, since 2005. He is interested
in computer vision and machine learning, especially
the visual analysis of complex scenes in motion,
structured models, graphical models and deep learning. He received his MSc in computer science from
Vienna University of Technology in 2000, and a
PhD from the National Institute of Applied Science
(INSA de Lyon), in 2003. In 2012 he obtained the
habilitation diploma.
Griffin Lacey is a Masters of Applied Science
student at the University of Guelph. His primary
research interests are in developing tools and techniques to support deep learning on FPGAs. Recently,
he acted as a visiting researcher at Google. Griffin
holds a BENG in Engineering Systems and Computing from the University of Guelph.
Lex Fridman is a postdoctoral associate at the
Massachusetts Institute of Technology (MIT). He received his BS, MS, and PhD from Drexel University.
His research interests include machine learning, decision fusion, and computer vision applied especially
to detection and analysis of human behavior in semiautonomous vehicles.
Deepak Chandra heads authentication at the Machine Intelligence and Research group at Google.
The project aims at completely redefining authentication for digital and physical world. Prior to this he
was the program lead in Google’s Advanced Technology and Projects (ATAP) organization, where he
heads all product, engineering, and design for mobile
authentication projects. Deepak defined company
wide authentication strategy for Motorola prior to
leading the efforts at Google. He has developed
multiple wearable authentication products including
Motorola Skip and Digital Tattoo.
Brandon Barbello is a Product Manager at Google
Research & Machine Intelligence, where he works
on privacy-sensitive on-device machine learning. He
was previously at Google Advanced Technology and
Projects (ATAP) on the Project Abacus team, where
he managed efforts to develop a multimodal continuous authentication system for smartphones. Prior to
Google, Brandon co-founded four companies across
electronics, fintech, and private equity.
Graham Taylor is an assistant professor at University of Guelph. He is interested in statistical
machine learning and biologically-inspired computer
vision, with an emphasis on unsupervised learning
and time series analysis. He completed his PhD at
the University of Toronto in 2009, where his thesis
co-advisors were Geoffrey Hinton and Sam Roweis.
He did a postdoc at NYU with Chris Bregler, Rob
Fergus, and Yann LeCun.
| 9cs.NE
|
1
A 588-Gbps LDPC Decoder Based on
Finite-Alphabet Message Passing
arXiv:1703.05769v2 [cs.AR] 30 Dec 2017
Reza Ghanaatian, Alexios Balatsoukas-Stimming, Christoph Müller, Michael Meidlinger, Gerald Matz,
Adam Teman, and Andreas Burg
Abstract—An ultra-high throughput low-density parity check
(LDPC) decoder with an unrolled full-parallel architecture is
proposed, which achieves the highest decoding throughput compared to previously reported LDPC decoders in the literature.
The decoder benefits from a serial message-transfer approach
between the decoding stages to alleviate the well-known routing
congestion problem in parallel LDPC decoders. Furthermore,
a finite-alphabet message passing algorithm is employed to
replace the variable node update rule of the standard min-sum
decoder with look-up tables, which are designed in a way that
maximizes the mutual information between decoding messages.
The proposed algorithm results in an architecture with reduced
bit-width messages, leading to a significantly higher decoding
throughput and to a lower area as compared to a min-sum
decoder when serial message-transfer is used. The architecture is
placed and routed for the standard min-sum reference decoder
and for the proposed finite-alphabet decoder using a custom
pseudo-hierarchical backend design strategy to further alleviate
routing congestions and to handle the large design. Post-layout
results show that the finite-alphabet decoder with the serial
message-transfer architecture achieves a throughput as large as
588 Gbps with an area of 16.2 mm2 and dissipates an average
power of 22.7 pJ per decoded bit in a 28 nm FD-SOI library.
Compared to the reference min-sum decoder, this corresponds to
3.1 times smaller area and 2 times better energy efficiency.
Index Terms—Low-density parity-check code, min-sum decoding, unrolled architecture, finite-alphabet decoder, 28 nm FD-SOI.
I. I NTRODUCTION
L
OW-DENSITY PARITY-CHECK (LDPC) codes have
become the coding scheme of choice in high datarate communication systems after their re-discovery in the
1990s [1], due to their excellent error correcting performance
along with the availability of efficient high-throughput hardware implementations in modern CMOS technologies. LDPC
codes are commonly decoded using iterative message passing
(MP) algorithms in which the initial estimations of the bits
are improved by a continuous exchange of messages between
decoder computation nodes. Among the various MP decoding algorithms, the min-sum (MS) decoding algorithm [2]
and its variants (e.g., offset MS, scaled MS) are the most
common choices for hardware implementation. LDPC decoder hardware implementations traditionally start from one
of these established algorithms (e.g., MS decoding), where the
R. Ghanaatian, A. Balatsoukas-Stimming, C. Müller, A. Teman, and A.
Burg are with the Telecommunications Circuits Laboratory, EPFL, Lausanne,
Switzerland (email: {reza.ghanaatian, alexios.balatsoukas, christoph.mueller,
adam.teman, andreas.burg}@epfl.ch).
M. Meidlinger and G. Matz are with the Vienna University of Technology,
Vienna, Austria (email: {mmeidlin, gmatz}@nt.tuwien.ac.at).
exchanged messages represent log likelihood ratios (LLRs).
These LLRs are encoded as fixed point numbers in two’scomplement or sign-magnitude representation, using a small
number of uniform quantization levels, in order to realize
the message update rules with low-complexity conventional
arithmetic operations.
Recently, there has been significant interest in the design
of finite-alphabet decoders for LDPC codes [3]–[7]. The main
idea behind finite-alphabet LDPC decoders is to start from
one or multiple arbitrary message alphabets, which can be
encoded with a bit-width that is acceptable from an implementation complexity perspective. The message update rules are
then crafted as generic mapping functions to operate on this
alphabet. The main advantage of such finite-alphabet decoders
is that the message bit-width can be reduced significantly with
respect to a conventional decoder, while maintaining the same
error-correcting performance [5], [6]. The downside of this
approach is that the message update rules of finite-alphabet
decoders usually cannot be described using fast and areaefficient standard arithmetic circuits.
Different hardware architectures for LDPC decoders have
been proposed in the literature in order to fulfill the power and
throughput requirements of various standards. More specifically, various degrees of resource sharing result in flexible decoders with different area requirements. On the one
hand, partial-parallel LDPC decoders [8], [9] and blockparallel LDPC decoders [10], [11] are designed for medium
throughput, with modest silicon area. Full-parallel [12], [13]
and unrolled LDPC decoders [5], [14], on the other hand,
achieve very high throughput (in the order of several tens or
hundreds of Gbps) at the expense of large area requirements.
Even though, in principle, LDPC decoders are massively
parallelizable, the implementation of ultra-high speed LDPC
decoders still remains a challenge, especially for long LDPC
codes with large node degrees [15]. While synthesis results
for such long-length codes, as for example reported in [16],
show the potential for a very high throughput, the actual
implementation requires several further considerations mainly
due to severe routing problems and the impact of parasitic
effects.
Contributions: In this paper, we propose an unrolled fullparallel architecture based on serial transfer of the decoding
messages, which enables an ultra-high throughput implementation of LDPC decoders for codes with large node degrees
by reducing the required interconnect wires for such decoders.
Moreover, we employ a finite-alphabet LDPC decoding algorithm in order to decrease the required quantization bit-
2
width, and thus, to increase the throughput, which is limited
by the serial message-transfer in the proposed architecture.
We also adopt a linear floorplan for the unrolled full-parallel
architecture as well as an efficient pseudo-hierarchical flow
that allow the high-speed physical implementation of the
proposed decoder. To the best of our knowledge, by combining
the aforementioned techniques, we present the fastest fully
placed and routed LDPC decoder in the literature.
Outline: The remainder of this paper is organized as follows: Section II gives an introduction to decoding of LDPC
codes, as well as more details on existing high-throughput
implementations of LDPC decoders. Section III describes
our proposed ultra-high throughout decoder architecture that
employs a serial message-transfer technique. In Section IV,
our algorithm to design a finite-alphabet decoder with nonuniform quantization is explained and applied to the serial
message-transfer decoder of Section III. Section V describes
our proposed approach for the physical implementation and
the timing and area optimization of our serial message-transfer
decoders. Finally, Section VI analyzes the implementation
results, and Section VII concludes the paper.
II. BACKGROUND
In this section, we first briefly summarize the fundamentals
of LDPC codes and the iterative MP algorithm for the decoding. We then review the state-of-the-art in high speed LDPC
decoder architectures to set the stage for the description of our
implementation.
A. LDPC Codes and Decoding Algorithms
A binary LDPC code is a set of codewords which are
defined through an M × N binary-valued sparse parity check
matrix as:
(1)
c ∈ {0, 1}N Hc = 0 ,
where all operations are performed modulo 2. If the parity
check matrix contains exactly dv ones per column and exactly
dc ones per row, the code is called a (dv , dc )-regular LDPC
code. Such codes are usually represented with a Tanner graph,
which contains N variable nodes (VNs) and M check nodes
(CNs) and VN n is connected to CN m if and only if
H mn = 1.
LDPC codes are traditionally decoded using MP algorithms,
where information is exchanged as messages between the VNs
and the CNs over the course of several decoding iterations. At
each iteration the message from VN n to CN m is computed
using a mapping Φv : Rdv → R, which is defined as:
µn→m = Φv Ln , µ̄N (n)\m→n ,
(2)
where N (n) denotes the neighbors of node n in the Tanner
graph, µ̄N (n)\m→n ∈ Rdv −1 is a vector that contains the
incoming messages from all neighboring CNs except m, and
Ln ∈ R denotes the channel LLR corresponding to VN n.
Similarly, the CN-to-VN messages are computed using a
mapping Φc : Rdc −1 → R, which is defined as:
µ̄m→n = Φc µN (m)\n→m .
(3)
n
Ln
Φv
µ̄m1 →n
m1
µn→m
µ̄mdv −1 →n
. . . mdv −1
m
(a)
...
n1
ndc −1
n
µndc −1 →m
µn1 →m
µ̄m→n
Φc
m
(b)
Fig. 1: (a) VN update and (b) CN update for N (n) =
{m, m1 , . . . , mdv −1 } and N (m) = {n, n1 , . . . , ndc −1 }.
Fig. 1 illustrates the message updates in the Tanner graph. In
addition to Φv and Φc , a third mapping Φd : Rdv +1 → {0, 1}
is needed to provide an estimate of the transmitted codeword
bits in the last VN iteration based on the incoming CN
messages and the channel LLR Ln according to:
ĉn = Φd (Ln , µ̄N (n)→n ).
(4)
Messages are exchanged until a valid codeword has been
decoded or until the maximum number of iterations I has
been reached.
For the widely used MS algorithm, the mappings (2) and
(3) are:
X
ΦMS
µ̄i ,
(5)
v (L, µ̄ = L +
i
and,
ΦMS
c (µ = sign µ min |µ|,
(6)
where min |µ| denotes the minimum Q
of the absolute values of
the vector components and sign µ = j sign µj . The decision
mapping Φd is defined as:
!!
X
1
MS
Φd (L, µ̄) =
1 − sign L +
µ̄i
.
(7)
2
i
B. High Throughput LDPC Decoders
Several high throughput LDPC decoders have been developed during the past decade in order to satisfy the high datarate requirements of optical and high-speed Ethernet networks.
These decoders usually rely on a full-parallel isomorphic [17]
architecture and a flooding schedule, which directly maps the
algorithm for one iteration to hardware. More specifically,
the CN and VN update equations are directly mapped to
M CN and N VN processing units and a hard-wired routing
network is responsible for passing the messages between
them. From an implementation perspective, while such an
architecture enables a very high throughput by fully exploiting
the inherent parallelism of each iteration, the complexity of
3
the highly unstructured routing network turns out to be a
severe bottleneck. In addition to this routing problem, such
full-parallel decoders usually require one or two clock cycles
for each iteration and in the worst case as many cycles as the
maximum number of iterations for each codeword, which is
another throughput limitation factor.
Several solutions have been proposed to alleviate the routing
problem in full-parallel decoders, on both architectural and
algorithmic levels. The authors of [18], [19] suggest using
a bit-serial architecture, which only requires a single wire for
each variable-to-check and check-to-variable node connection.
While this approach can reduce the routing congestion, it also
leads to a significant reduction in the decoding throughput. The
decoder in [19], for example, only achieves a throughput of
3.3 Gbps when implemented using a 130 nm CMOS technology. Another architectural technique is reported in [13], where
the long wires of the decoder are partitioned into several short
wires with pipeline registers. As a result, the critical path is
broken down into shorter paths, but the decoder throughput
is also affected since more cycles are required to accomplish
each iteration. Nevertheless, the decoder of [13] is still able
to achieve 13.2 Gbps in 90 nm CMOS with 16 iterations.
On an algorithmic level, the authors of [20] propose a
MP algorithm, called MS split-row threshold, which uses a
column-wise division of the H matrix into Spn partitions.
Each partition contains N/Spn VNs and M CNs, and global
interconnects are minimized by only sharing the minimum
signs between the CNs of each partition. This algorithmic
modification was used to implement a full-parallel decoder
for the challenging (2048, 1723) LDPC code in 65 nm CMOS,
which achieves a throughput of 36.2 Gbps with 11 decoding
iterations. Another decoder, reported in [21], uses a hybrid
hard/soft decoding algorithm, called differential binary MP
algorithm, which reduces the interconnect complexity at the
cost of some error-correcting performance degradation. A
full-parallel (2048, 1723) LDPC decoder using this algorithm
was implemented in 65 nm CMOS technology, achieving a
throughput of up to 126 Gbps [22]. The work of [23] also
proposes another algorithmic level modification, called the
probabilistic MS algorithm, where a probabilistic second minimum value is used instead of the true second minimum value
to simplify the CN operation and to facilitate high-throughput
implementation of full-parallel LDPC decoders. Further, a
mix of tree and butterfly interconnect network is proposed
in the CN unit to balance the interconnect complexity and
the logic overhead and to reduce the routing complexity. The
implementation of a decoder for the (2048, 1723) LDPC code
with the proposed techniques in 90 nm CMOS technology
achieves a throughput of 45.42 Gbps.
Stochastic decoding of LDPC codes [24] was another important improvement based on both algorithm and gate-level
implementation considerations to solve the routing problem of
LDPC decoders, where probabilities are interpreted as binary
Bernoulli sequences. This approach, on the one hand, reduces
the complexity of CNs and the routing overhead, but, on the
other hand, introduces difficulties in VN update rules due
to correlated stochastic streams, which may deteriorate the
error-correcting performance especially in longer-length codes.
To solve this correlation problem by re-randomizing the VN
output streams, the work of [25] proposes to use majoritybased tracking forecast memories in each VN, which results
in a decoder with full-parallel architecture for a (2048, 1723)
LDPC code that achieves a throughput of 61.3 Gbps in 90 nm
CMOS. An alternative method to track the probability values,
called delayed stochastic decoding, is reported in [26] and
the full-parallel decoder for the same code can deliver a
throughput as large as 172.4 Gbps in 90 nm CMOS.
To solve the problem of throughput limitations in fullparallel decoders from potentially using multiple iterations
for decoding, the work of [14] presents an unrolled fullparallel LDPC decoder. In the proposed architecture, each
decoding iteration is mapped to distinct hardware resources,
leading to a decoder with I iterations that can decode one
codeword per clock cycle, at the cost of significantly increased
area requirements with respect to non-unrolled full-parallel
decoders. This unrolled architecture achieves a throughput
of 161 Gbps for a (672, 546) LDPC code with dv = 3 and
dc = 6, when implemented in a 65 nm CMOS technology. It
is noteworthy that an unrolled decoder has 50% reduced wires
between adjacent stages compared to a non-unrolled decoder
since one stage of variable nodes is connected to one stage
of check nodes with uni-directional data flow per decoding
iteration. Even though this measure leads to a lower routing
congestion, it is still challenging to fully place and route
such a decoder. This routing issue becomes more and more
severe when considering longer LDPC codes and especially
with increasing CN and VN degrees to achieve better errorcorrecting performance, as required in wireline applications
such as for the (2048, 1723) code with dv = 6 and dc = 32
used in the IEEE 802.3an standard [15].
III. S ERIAL M ESSAGE -T RANSFER LDPC D ECODER
Unrolled full-parallel LDPC decoders provide the ultimate
throughput with smaller routing congestion than conventional
full-parallel decoders. However, they are still not trivial to
implement for long LDPC codes with high CN and VN
degrees, which suffer from severe routing congestion. Hence,
in this section, we propose an unrolled full-parallel LDPC
decoder architecture that employs a serial message-transfer
technique between CNs and VNs.1 This architecture is similar
to the bit-serial implementations of [18], [19] in the way
the messages are transferred; however, as we shall see later,
it differs in the fact that it is unrolled and in the way the
messages are processed in the CNs and VNs.
A. Decoder Architecture Overview
An overview of the proposed unrolled serial messagetransfer LDPC decoder architecture is shown in Fig. 2. As
with all unrolled LDPC decoders, each decoding iteration is
1 We note that a first implementation of our decoder was based on
parallel (word-level) message-transfer. The place and route tool for this
implementation was hardly able to converge even when the area utilization was
unacceptable. We, therefore, propose serial message-transfer architecture and
we adopt special implementation methodologies for this architecture, which
will be explained in Section V.
4
Check Node
Stage
Variable Node
Stage
Decision Node
Stage
VN Unit-1
channel
N
1
S/P-1
Q
Q
S/P-2
Q
S/P-dc
Q
Q
Q
CN Unit-2
P/S-1
P/S-2
P/S-dc
1
1
1
S/P-2
S/P-dv
CN-VN Routing
1
CN Processor
Input
LLRs
VN-CN Routing
N
1
1
S/P-1
CN Unit-M
Q
Q
Q
DN Unit-1
P/S-ch
Q ch
Q
Q
Q
P/S-1
P/S-2
P/S-dv
1
1
1
1
1
1
1
1
S/P-ch
S/P-1
S/P-2
S/P-dv
Q ch
Q
Q
Q
DN Processor
CN Unit-1
1
S/P-ch
VN Processor
1
messages
N
1
VN Unit-2
DN Unit-2
VN Unit-N
DN Unit-N
Out-1
Fig. 2: Serial message-transfer decoder architecture.
mapped to a distinct set of N VN and M CN units, which
form a processing pipeline. In essence, the unrolled LDPC
decoder is a systolic array, in which a new set of N channel
LLRs is read in every clock cycle and a decoded codeword is
put out in every clock cycle.
Even though both the CNs and VNs can compute their
outgoing messages in a single clock cycle, similar to the
architecture in [5], in the proposed serial message-transfer
architecture each message is transfered one bit at a time
between the consecutive variable node and check node stages
over the course of Qmsg clock cycles, where Qmsg is the
number of bits used for the messages. More specifically,
each CN and VN unit contains a serial-to-parallel (S/P) and
parallel-to-serial (P/S) conversion unit at the input and output,
respectively, which are clocked Qmsg times faster than the
processing clock to collect and transfer messages serially,
while keeping the overall decoding throughput constant. More
details on the architecture of the CN and VN units as well as
the proposed serial message-transfer mechanism are provided
in the sequel.
B. Decoder Stages
The unrolled LDPC decoder, illustrated in Fig. 2, consists
of three types of processing stages, which are described in
more detail below. We note that the CN/VN processors of
this reference decoder are similar to those of a standard MS
decoder, and our modifications for these parts (to realize a
finite-alphabet decoder) are discussed in Section IV.
1) Check Node Stage: Each check node stage consists of
M CN units, each of which contains three components: a CN
processor, which implements (6) similarly to [5], [14], dc S/P
units for the dc input messages, and dc P/S units for the dc
output messages. Moreover, the complete check node stage
contains a register bank that is used to store the channel LLRs,
which are not directly needed by the check node stage, but
nevertheless must be forwarded to the following variable node
stage and thus need to be buffered. Hence, no S/P and P/S units
are required for the channel LLR buffers in the check node
stages as they are simply forwarded serially to the following
variable node stage.
2) Variable Node Stage: Each variable node stage consists
of N VN units, each of which contains a VN processor and S/P
and P/S units at the inputs and outputs, respectively, similar
to the CN unit structure. Each VN processor implements the
update rule (5) similarly to [5].
3) Decision Node Stage: The last variable node stage is
called a decision node stage because it is responsible for taking
the final hard decisions on the decoded codeword bits. The
structure of this stage is similar to a variable node stage, but a
decision node (DN) has a simpler version of the VN processor
that only computes sum of all inputs and put out its sign [5],
and thus no P/S unit is required at its output.
C. Message Transfer Mechanism
One of the modifications, compared to [14] and [5], is the
serial transfer of the channel and message LLRs between the
stages of the decoder, which reduces the required routing
resources by factor of Qmsg . This modification is applied
to make the placement and routing of the decoder feasible,
especially for large values of dv and dc . To this end, as
explained in the previous section, a S/P and a P/S shift register
are added to each input and each output of the CN and VN
units, as illustrated in Fig. 3. We see that the S/P unit consists
of a (Qmsg −1)-bit shift register and Qmsg memory registers,
while the P/S unit has Qmsg registers with multiplexed inputs.
The serial messages are transfered with a fast clock, denoted
by CLKF , that is Qmsg times faster than the slow processing
clock, denoted by CLKS . More specifically, at each CN unit
5
Serial-to-Parallel
Q4 Q3 Q2 Q1 Q0
D Q
D Q
D Q
D Q
D Q
D Q
D Q
D Q
Shift Registers
Fast Serial In
CLK F
D Q
Memory Registers
CLK S
Other inputs
. . .
Slow CN/VN Processor
. . .
Other outputs
Q’4 Q’3 Q’2 Q’1 Q’0
CLK S
CLK F
D Q
D Q
D Q
D Q
D Q
Fast Serial Out
Load/Shift
CLK F
Parallel-to-Serial
Fig. 3: The message receive and transfer mechanism by S/P
and P/S shift registers, enabled by the fast clock (CLKF ), and
the message process enabled by the slow clock (CLKS ) for
Qmsg = 5.
and VN unit input, data is loaded serially into the S/P shift
register using the fast CLKF , and after the Qmsg -th cycle all
message bits are stored in memory registers, clocked by the
slow CLKS . The CN/VN processing can then be performed
in one CLKS cycle and the output messages are saved in the
output P/S shift register and transferred serially to the next
stage using again CLKF . At the same time, a new set of
messages is serially loaded into the input shift register. We note
that for simpler clock tree generation, all registers in Fig. 3
are clocked by CLKF , while CLKS is actually implemented as
a pulsed clock, which is generated using a clock-gating cell
controlled by a finite state machine.
D. Decoder Hardware Complexity and Performance Analysis
In this section, we describe the required memory complexity
of the proposed decoder as well as the decoding latency and
the throughput.
1) Memory Requirement: The memory complexity can
easily be characterized by counting the number of required
registers and can be approximated by:
Rtot ≈ N (dv + 1)Q(6I − 1),
(8)
where I is the number of decoding iterations (which in
unrolled decoders strongly affects the memory requirements)
and Q = Qmsg = Qch , which is often the case for MS LDPC
decoders. From (8), one can easily see that the quantization
bit-width linearly increases the memory requirement for the
proposed architecture.
2) Decoding Latency: Since each stage has a delay of two
CLKS cycles and there are two stages for each decoding iteration, the decoder latency is 4I CLKS cycles or, equivalently,
4IQmax CLKF cycles, where Qmax = max(Qmsg , Qch ).
3) Decoding Throughput: In the proposed unrolled architecture, one decoder codeword is output in each CLKS cycle.
Therefore, the coded throughput of the decoder is:
T = N fSmax ,
(9)
where fSmax is the maximum frequency of CLKS while the
maximum frequency of CLKF or simply maximum frequency
of the decoder is fmax = fFmax = Qmax fSmax . For the proposed
architecture, we have:
−1
fSmax = max (Qmax TCP,route ), (TCP,VN ), (TCP,CN )
,
(10)
where TCP,VN and TCP,CN are the delay of the critical paths
of the CN unit and the VN unit, respectively, and TCP,route
is the critical path delay of the (serial) routing between the
decoding stages. Thus, the decoder throughput will be limited
by the routing, if the VN/CN delay is smaller than Qmax times
the routing delay. Hence, on the one hand, the serial messagetransfer decoder alleviates the routing problem by reducing the
required number of wires, but on the other hand, the decoder
throughput for large quantization bit-widths may be affected,
as the serial message-transfer delay will become the limiting
factor.
IV. F INITE -A LPHABET S ERIAL M ESSAGE -T RANSFER
LDPC D ECODER
Even though the serial message-transfer architecture alleviates the routing congestion of an unrolled full-parallel LDPC
decoder, it has a negative impact on both throughput and
hardware complexity, as discussed in the previous section.
In our previous work [5], [6], we have shown that finitealphabet decoders have the potential to reduce the required
number of message bits while maintaining the same error rate
performance. In this section, we will review the basic idea
and our design method for this new type of decoders and then
show how the bit-width reduction technique of [5], [6] can
be applied verbatim in order to increase the throughput and
reduce the area of the serial message-transfer architecture.
A. Mutual Information Based Finite-Alphabet Decoder
In our approach of [5], [6], the standard message-passing
decoding algorithm update rules are replaced by custom update rules that can be implemented as simple look-up tables
(LUTs). These LUTs take integer-valued input messages and
produce a corresponding output message. Moreover, the inputoutput mapping that is represented by the LUTs is designed
in a way that maximizes the mutual information between the
LUT output messages and the codeword bit that these messages correspond to. We note that a similar approach was also
used in [7], but the corresponding hardware implementation
would have a much higher hardware complexity than the
method of [5], [6]. This happens because, contrary to [7],
in [5], [6] we used LUTs only for the VNs while the CNs
use the standard min-sum update rule.
B. Error-Correcting Performance and Bit-Width Reduction
In Fig. 4, we compare the performance of the IEEE 802.3an
LDPC code under floating-point MS decoding, fixed-point
MS decoding (with Qch = Qmsg ∈ {4, 5}), and LUT-based
decoding (with Qch = 4 and Qmsg = 3) when performing
I = 5 decoding iterations with a flooding schedule. We
also show the performance of a floating-point offset min-sum
6
Frame Error Rate
100
10−2
10−4
10−6
LUT, (Qch = 4, Qmsg = 3)
MS, (Qch = 4, Qmsg = 4)
MS, (Qch = 5, Qmsg = 5)
MS, (floating-point)
10−8
OMS, (Floating-point)2
−10
10
3
3.5
4
4.5
5
5.5
On the negative side, we remark that the VN units for
each variable node stage (decoder iteration) of the LUTbased decoder are different, which slightly complicates the
hierarchical physical implementation as we will see later.
Furthermore, since Qch > Qmsg , we now need to transfer
the channel LLRs with multiple (two) bits per cycle to avoid
the need to artificially limit the number of CLKF cycles per
iteration to Qch rather than to the smaller Qmsg . To reflect this
modification, we redefine (10) as Qmax = max(Qmsg , d Q2ch e).
While this partially parallel transfer of the channel LLRs
impacts routing congestion, we note that the overhead is negligible since the number of channel LLRs is small compared
to the total number of messages.
Eb /N0 [dB]
Fig. 4: Frame error rate (FER) of the IEEE 802.3an LDPC
code under floating-point MS decoding, fixed-point MS decoding with different bit-widths, LUT based decoding, and
floating-point offset min-sum (OMS) decoding (offset=0.5) as
reference, all with I = 5 decoding iterations.
(OMS) decoder as a reference.2 We observe that the fixedpoint decoder with Qch = Qmsg = 5 has almost the same
performance as the floating-point decoder, while the fixedpoint decoder with Qch = Qmsg = 4 shows a significant
loss with respect to the floating-point implementation. Thus,
a standard MS decoder would need to use at least Qch =
Qmsg = 5 quantization bits. The LUT-based decoder, however,
can match the performance of the floating-point decoder with
only Qch = 4 channel quantization bits and Qmsg = 3 message
quantization bits.3 Additionally, with the above quantization
bit choices, no noticeable error floor has been observed for
both the MS decoder and LUT-based decoder when 1010
frames have been transmitted (which corresponds to a BER
≈ 10−12 with the current block length). We note that, for the
LUT-based decoder, the performance in the error floor region
can be traded with the performance in the waterfall region by
an appropriate choice of the design SNR for the LUTs [6].
C. LUT-Based Decoder Hardware Architecture
The LUT-based serial message-transfer decoder hardware
architecture is very similar to the MS decoder architecture,
described in Section III. However, the LUT-based decoder
can take advantage of the significantly fewer message bits
that need to be transferred from one decoding stage to the
next. This reduction reduces the number of CLKF cycles per
iteration, which in turn increases the throughput of the decoder
according to (10) provided that the CN/VN logic is sufficiently
fast. Moreover, the size of the buffers needed for the S/P and
P/S conversions is also reduced significantly, which directly
reduces the memory complexity of the decoder (see (8)).
2 The reference simulation was obtained and matched with our simulation by
using the open-source simulator provided by: Adrien Cassagne; Romain Tajan;
Mathieu Lonardon; Baptiste Petit; Guillaume Delbergue; Thibaud Tonnellier;
Camille Leroux; Olivier Hartmann, AFF3CT: A Fast Forward Error Correction
Tool, 2016. [Online]. Available: https://doi.org/10.5281/zenodo.167837
3 We note that reducing Q
ch further results in a non-negligible loss with
respect to the floating-point decoder.
V. I MPLEMENTATION
Despite the use of a serial message-transfer, the physical
implementation of the decoders proposed in the previous
sections requires special scrutiny since the number of global
wires is still significant and the overall area is particularly
large. Therefore, in this section, we propose and describe
a pseudo-hierarchical design methodology to implement the
serial message-transfer architecture.
A. Physical Design
Due to the large number of identical blocks in the decoder
architecture, a bottom-up flow is expected to provide the best
results. The CN, VN, and DN units are first placed and
routed individually to build hard macros,4 and their timing
and physical information are extracted. These macros are then
instantiated as large cells in the decoder top level. We propose
to treat the macros as custom standard-cells with identical
height to be able to perform the placement using the standardcell placement, rather than the less capable macro placement of
the backend tool, since in our case the number of hard macro
instances is extremely large and the interconnect pattern is
complex and highly irregular.
Fig. 5 illustrates the proposed physical floorplan for the
decoders with the unrolled architecture. In this floorplan, the
CN and VN macros within each stage are constrained to be
placed into dedicated regions (placement regions in Fig. 5a).
This measure enforces the high-level structure of systolic
array pipeline, but it also leaves freedom to the placement
tool to choose the location for the macros in each stage to
minimize routing congestion between stages. Note that the
linear floorplan has also the advantage of being scalable in
the number of iterations since little interaction or interference
exists between stages. Furthermore, the CN and VN macros
are placed in dedicated rows while the area between these rows
is left for repeaters and for the register standard-cells for the
channel LLRs in the check node stages, as shown in Fig. 5b.
We note that the proposed floorplan and the encapsulation of
the VN and CN macros as large standard-cells exploit the
automated algorithm to optimize both custom and conventional
standard-cells placement in order to alleviate the significant
routing congestion.
4 Note that for the LUT-based decoder there are different macros for each
variable node stage as apposed to the MS decoder.
7
Placement region for CN stage 1
Placement region for VN stage 1
TCSTR,CLKS : timing constraint applied to CLKS
TCSTR,CLKF : timing constraint applied to CLKF
Placement region for CN stage 2
Synthesis
Placement region for VN stage 2
TCSTR,CLKS
TCSTR,CLKF
.
.
.
(b)
Implement CN/VN macros with
TCSTR,CLKS and minimum area
Placement region for DN stage
(a)
Implement the decoder toplevel
with TCSTR,CLKF and extract the
minimum achievable CLKF period
TCSTR,CLKS =
Qmax TCSTR,CLKF
Area for macros
Update TCSTR,CLKF
to the minimum
achievable CLKF period
Area for registers
and repeaters
.
.
.
(b)
Fig. 5: The physical floorplan for serial message-transfer
architecture, (a) high level overview of the floorplan with
dedicated placement regions for each decoder stage; and (b)
zoomed in overview showing rows structure for custom macros
(large colored blocks) and conventional standard-cells (small
colored blocks) placement.
no
yes
Stop
Fig. 6: The proposed flowchart to optimize timing and area
for the serial message-transfer architecture.
2.99 mm
B. Timing and Area Optimization Flow
Although the synthesis results can give an approximate
evaluation for timing and area of the physical implementation, several iterations with different constraints are required
to reach an optimal layout. To this end, we propose the
methodology illustrated in the flowchart of Fig. 6 to effectively
implement the serial message-transfer architecture. The main
idea behind this methodology is that three main factors directly
contribute to the decoder throughput and also indirectly to
the decoder area, as discussed in Section III and specifically
summarized in (10). Our goal is to maximize the throughput
at a minimum area.
We define the timing constraint applied to CLKS as
TCSTR,CLKS , and the timing constrain applied to CLKF as
TCSTR,CLKF . The first step is to place and route the CN/VN
macros based on TCSTR,CLKS . This step is followed by the
implementation of the decoder using TCSTR,CLKF . (The initial
constraints for the backend are thereby extracted from synthesis timing results.) The fully placed and routed design can
give an accurate routing delay, which will be used to update
TCSTR,CLKF and then TCSTR,CLKS according to (10). The updated
TCSTR,CLKS will be used to re-implement the CN and VN
macros within the minimum achievable area.
We note that for a long LDPC code with a large
area and long routing delay (such as the one of the
IEEE 802.3an standard), the first implementation starts with
TCSTR,CLKS < Qmax TCSTR,CLKF . After obtaining a realistic value
TCSTR,CLKS ≈
Qmax TCSTR,CLKF
2.45 mm
7.78
mm
6.61
mm
(a)
(b)
Fig. 7: Layouts for (a) the MS decoder and (b) the LUT-based
decoder.
for the CLKF period (and hence for TCSTR,CLKF ) at the end
of the implementation, the TCSTR,CLKS will be updated to
a larger value to approach TCSTR,CLKS ≈ Qmax TCSTR,CLKF .
Consequently, the CN and VN macro area and thus the decoder
area will shrink in the second iteration, which result in larger
achievable CLKF frequency and hence smaller TCSTR,CLKF and
TCSTR,CLKS . The feedback loop will reach the optimum point
after a few iterations.
8
TABLE II: Detailed area for CN/VN unit?
TABLE I: Critical path delays for MS and LUT-based decoder
Path
CN [ns]
VN [ns]
Routing [ns]
15
14.12
MS decoder
2.38
0.96
1.51
LUT-based decoder
1.42
1.24
1.16
LUT-based decoder
MS decoder
Component
CN unit logic [µm2 ]
CN unit register [µm2 ]
CN macro [µm2 ]
MS decoder
1578
1695
3607
LUT-based decoder
485
971
1510
VN unit logic [µm2 ]
VN unit register [µm2 ]
315
381
403†
235†
VN macro [µm2 ]
755
646†
Area (mm2 )
? Logic
10
9.56
7.95
5.41
and register areas are obtained form synthesis, and macro areas
are the final post-layout results.
† Although the VNs are different for each stage of the LUT-based decoder,
their areas are similar and the result for one of them is reported here.
TABLE III: Power and energy efficiency comparison for MS
and LUT-based decoder
5
0.84 1.03
tin
Ro
u
R
C
ro lk t
ut re
in e
g an
bu d
ffe
rs
C
re han
gi ne
ste l
rs LL
C
m N/V
ac N
ro
s
g
0.18 0.15
0
Fig. 8: Detailed area results for the LUT-based and MS decoder with total area of 16.2 mm2 and 23.3 mm2 , respectively.
VI. R ESULTS AND D ISCUSSIONS
To study the impact of the serial message-transfer architecture and the finite-alphabet decoding scheme, we have
implemented the proposed architecture by employing the
methodology explained in Section V and we analyzed the
results for both MS and LUT-based decoding. We used the
parity check matrix of the LDPC code defined in the IEEE
802.3an standard [15], i.e., a (2048, 1723) LDPC code of
R = 0.8413 with dv = 6 and dc = 32. We used I = 5
for both decoders and Qmsg = Qch = 5 for the MS decoder
and Qmsg = 3 and Qch = 4 for the LUT-based decoder to
achieve the same error-correction performance, as described
in Section IV. The decoders were synthesized from a VHDL
description using Synopsys Design Compiler and placed and
routed using Cadence Encounter Digital Implementation. The
layouts are shown in Fig. 7. The results are reported for
a 28 nm FD-SOI library under typical operating conditions
(VDD = 1 V, T = 25◦ C).
A. Delay Analysis
In the serial message-transfer architecture, the critical path
and, hence, the maximum decoding frequency are defined
by (10). To investigate the impact of serially transferring the
messages on the decoder throughput, we consider the delay
of the following register-to-register critical paths for both the
MS and LUT-based decoder.
1) CN Critical Path: The CN critical path (TCP,CN ) is the
path from the S/P memory registers to the P/S shift register
within the CN unit. For both decoders, this path is essentially
comprised of the logic cells for a sorter tree with a depth of
four.
Total power @ fmax [mW]
Total power @ 662 MHz [mW]
Leakage power [mW]
Energy efficiency [pJ/bit]
MS decoder
12248
12248
7.44
45.2
LUT-based decoder
13350
10257
5.27
22.7
2) VN Critical Path: The VN critical path (TCP,VN ) is the
path from S/P memory registers to P/S shift register within
the VN unit. This path is dominated by an adder tree for the
MS decoder and an LUT tree for the LUT-based decoder.
3) Routing Critical Path: The routing critical path
(TCP,route ) comprises mainly the interconnect wires (and
buffers) that connect the CN/VN unit S/P shift register to the
VN/CN unit P/S shift register.
Table I summarizes the critical path delays of the CN/VN
and the routing path. Together with (10), the values in the
table dictate the maximum achievable frequency for CLKS
and CLKF , respectively, for both of the decoders with the
proposed serial message-transfer architecture. We note that the
critical paths are reported after the timing and area constraints
for the CN/VN macros and for the decoder toplevel are
jointly optimized according to the flow shown in Fig. 6.
According to (10), we observe that in both decoders the
message transfer limits the slow clock CLKS to a period
of 5 × 1.51 ns = 7.55 ns and 3 × 1.16 ns = 3.48 ns for the
MS and the LUT-based decoder, respectively, where 1.51 ns
and 1.16 ns are the corresponding minimum CLKF periods.
Consequently, in our flow, the VN and CN units end up as
optimized for minimum area only with relaxed and easy to
meet timing constraints.
B. Area Analysis
Fig. 8 illustrates the area distribution among the various
components after the layout. The area utilization is approximately 67% for both decoders. While almost 62% of the layout
is filled with CN/VN macros and registers, the clock tree and
routing buffers occupy around 5%. Furthermore, we see a 44%
difference in total area between the decoders due to the fact
that the total area for CN and VN macros is 14.12 mm2 in the
MS decoder, as opposed to only 9.56 mm2 in the LUT-based
decoder.
9
TABLE IV: Implementation results for MS and LUT-based decoder and comparison with other works
LUT-based
decoder
MS decoder
[14]
Process technology
28 nm FD-SOI
65 nm CMOS
Supply voltage [V]
1.0
1.2
(2048, 1723)
(6, 32)
(672, 546)
(3, 6)
LDPC code
Node degree (dv , dc )
Algorithm
Imax
Quantization bits
Eb /N0 @ BER= 10−7 [dB]
Architecture
min-sum
finite-alphabet
5
5
4.97
3
4.95
unrolled full-parallel
min-sum
9
4
−
unrolled
full-parallel
[8]
[20]
[23]
[27]
[26]
65 nm CMOS
90 nm CMOS
90 nm CMOS
90 nm CMOS
1.3
0.9
1.2
1.0
(2048, 1723)
(6, 32)
offset min-sum
with post processor
8
4
4.25
(2048, 1723)
(6, 32)
(2048, 1723)
(6, 32)
reduced-complexity
min-sum
30
6
4.32
(2048, 1723)
(6, 32)
11
5
4.55
(2048, 1723)
(6, 32)
normalized probablistic
min-sum
9
4
4.4
65 nm CMOS
low-power
1.2
split-row
delayed stochastic
−
5
4.7
partial-parallel
full-parallel
full-parallel
full-parallel
full-parallel
Core area [mm2 ]
Area utilization [%]
Max. frequency (fmax ) [MHz]
Latency [ns]
Throughput @ Imax [Gbps]
Power @ fmax [mW]
Area eff. [Gbps/mm2 ]
Energy per bit @ Imax [pJ/bit]
23.3
66.4
662
151
271
12248
11.6
45.2
16.2
65.9
862
69.6
588
13350
36.3
22.7
12.9
76
257
105
160.8
5360
12.5
33.3
5.05
84.5
700
137
13.3
2800
2.63
210.5
4.84
97
195
56.4
36.3
1359
7.5
37.4
9.6
91
199.6
45.09
45.42
1110
4.73
24.44
3.84
−
226
−
12.8
1040
3.34
81.2
3.93
93
750
800
172.4?
−
43.86
−
Scaled area eff.† [Gbps/mm2 ]
Scaled energy per bit‡ [pJ/bit]
11.6
45.2
36.3
22.7
156.4
10
32.9
63
93.8
9.5
157.1
9.4
110.9
17.5
1456.8
−
? Maximum
throughput @ Eb /N0 = 5.5 dB (Note that throughput @ Imax is not reported in the original paper)
is done by S 3 where S is the relative dimension to 28 (Note that this is very rough and optimistic since it does not apply to the interconnects)
‡ Scaling is done by 1/SU 2 where U is the relative voltage to 1.0
† Scaling
To understand this fact, we list the area of each CN/VN
macro in Table II. According to this table, the finite-alphabet
message passing algorithm leads to significantly smaller CN
processors because of two important factors: first, the bit-width
reduction of the messages directly affects the data-path area,
and second, the quantized messages in the LUT-based decoder
are processed directly in the sorter tree of the CNs without
the need to compute their absolute values. However, VN
processors are less area-efficient in the LUT-based decoder in
comparison with the ones of the MS decoder. This is caused by
the fact that the LUT-based computations are, in general, less
area-efficient than the conventional arithmetic based update
rules. Thus, the logic area of the VN in the LUT-based decoder
is larger, even though their input/output bit-width is smaller.
Another contributing factor in the Table II is the register area,
which is defined by the number of S/P and P/S registers.
For those, the 40% reduction of bit-width in the LUT-based
decoder is directly noticeable in the register area for both CN
and VN units. Altogether, the CN and VN macros in the
LUT-based decoder are 58% and 14% smaller, respectively,
compared to those of the MS decoder.
C. Power Analysis
The energy which is consumed by each decoder is proportional to the capacitance, which in turn is related to the
decoder area. Also, the number of required CLKF cycles for
the serial message-transfer to decode one codeword, which is
inversely proportional to the decoding throughput at a constant
frequency, directly contributes to the consumed energy for
each decoded bit. Therefore, we analyze both the total power
and the energy efficiency of the decoders using post-layout
vector-based power analysis.5 The results are reported in
Table III. We note that the total powers are calculated at fmax
5 We first extract the parasitic information of both the hard macros and the
top level from the placement and routing tool and then read and link them
using a power computation tool to generate the complete parasitic information.
for both decoders. Also, for comparison purpose, we have
calculated the total powers at a constant CLKF frequency,
here min(fmax,MS , fmax,LUT ) = 662 MHz, for both decoders
and note them in the Table III. According to this table,
the total power consumption of the LUT-based decoder is
16.2% smaller than that of the MS decoder. Furthermore, by
considering the fact that the LUT-based decoder has 66.7%
higher throughput than the MS decoder at a similar CLKF
frequency, the energy efficiency of the LUT-based decoder is
almost 2 times better in comparison with the MS decoder.
D. Summary and Final Comparison
The final post-layout results for our MS and LUT-based decoders and also for some other recently implemented decoders
are summarized in Table IV. Our LUT-based decoder runs
at a maximum CLKF frequency of fmax,LUT = 862 MHz and
delivers a sustained throughput of 588 Gbps, while it occupies
16.2 mm2 area and dissipates 22.7 pJ/bit. Compared to the MS
decoder, the LUT-based decoder is 1.4× smaller, 2.2× faster,
and thus 3.1× more area efficient. It also has 16.2% lower
power dissipation and 2× better energy efficiency, when the
decoding throughout is taken into account.
The work in [14] is the only other unrolled fullparallel decoder in literature, but it is designed for the
IEEE 802.11ad [28] code, which has a shorter block length
and smaller node degrees (dv = 3 and dc = 6 as opposed
to dv = 6 and dc = 32 for the code used in the design
reported in this paper). The work of [8], [20], [23], [26], and
[27] are for the same IEEE 802.3an code considered in this
paper, but with partial-parallel and full-parallel architectures.
The proposed LUT-based decoder has more than an order of
magnitude higher throughput compared to [20] and [23], and
three times higher throughput compared to [26], while the
maximum throughput of the proposed decoder is maintained
for all SNR scenarios as it does not require early termination to
achieve a high throughput. The area efficiency of the proposed
10
unrolled full-parallel architecture, however, is inferior to the
one of the decoders in [20], [23] and [27] with full-parallel
architecture due to the repeated routing overhead between the
decoder stages in our design.
VII. C ONCLUSION
An ultra high throughput LDPC decoder with a serial
message-transfer architecture and based on non-uniform quantization of messages was proposed to achieve the highest
decoding throughput in literature. The proposed decoder architecture is an unrolled full-parallel architecture with serialized
messages for CN/VN units, which was enabled by employing
S/P and P/S shift registers at the inputs and outputs of each
unit. The proposed quantized message passing algorithm replaces conventional MS, resulting in 40% reduction in message
bit-width without any performance penalty. This algorithm was
implemented by using generic LUTs instead of adders for VNs
while the CNs remained unchanged compared to MS decoding. Placement and routing results in 28 nm FD-SOI show
that the LUT-based serial message-transfer decoder delivers
0.588 Tbps throughput and is 3.1 times more area efficient
and 2 times more energy efficient in comparison with the MS
decoder with serial message-transfer architecture.
ACKNOWLEDGMENT
This work was supported by the Swiss National Science
Foundation (SNSF) under the project number 200021-153640.
R EFERENCES
[1] D. J. MacKay, “Good error-correcting codes based on very sparse
matrices,” IEEE Trans. Inf. Theory, vol. 45, no. 2, pp. 399–431, 1999.
[2] M. P. Fossorier, M. Mihaljević, and H. Imai, “Reduced complexity
iterative decoding of low-density parity check codes based on belief
propagation,” IEEE Trans. Commun., vol. 47, no. 5, pp. 673–680, 1999.
[3] S. K. Planjery, D. Declercq, L. Danjean, and B. Vasic, “Finite alphabet
iterative decoderspart i: Decoding beyond belief propagation on the
binary symmetric channel,” IEEE Trans. Commun., vol. 61, no. 10, pp.
4033–4045, 2013.
[4] B. Kurkoski, K. Yamaguchi, and K. Kobayashi, “Noise thresholds for
discrete LDPC decoding mappings,” in IEEE Global Telecommun. Conf.
(GLOBECOM), Nov. 2008, pp. 1–5.
[5] A. Balatsoukas-Stimming, M. Meidlinger, R. Ghanaatian, G. Matz, and
A. Burg, “A fully-unrolled LDPC decoder based on quantized message
passing,” in IEEE Int. Workshop on Signal Process. Syst. (SiPS), Oct
2015, pp. 1–6.
[6] M. Meidlinger, A. Balatsoukas-Stimming, A. Burg, and G. Matz, “Quantized message passing for LDPC codes,” in Asilomar Conf. on Signals,
Syst., and Comput. (ACSSC), Nov 2015, pp. 1606–1610.
[7] F. J. C. Romero and B. M. Kurkoski, “Decoding LDPC codes with
mutual information-maximizing lookup tables,” in IEEE Int. Symp. on
Inf. Theory (ISIT), Jun. 2015, pp. 426–430.
[8] Z. Zhang, V. Anantharam, M. J. Wainwright, and B. Nikolić, “An
efficient 10GBASE-T Ethernet LDPC decoder design with low error
floors,” IEEE J. Solid-State Circuits, vol. 45, no. 4, pp. 843–855, 2010.
[9] A. Cevrero, Y. Leblebici, P. Ienne, and A. Burg, “A 5.35 mm2
10GBASE-T Ethernet LDPC decoder chip in 90 nm CMOS,” in IEEE
Asian Solid-State Circuits Conf. (A-SSCC), 2010, pp. 1–4.
[10] C. Roth, P. Meinerzhagen, C. Studer, and A. Burg, “A 15.8 pJ/bit/iter
quasi-cyclic LDPC decoder for IEEE 802.11n in 90 nm CMOS,” in
IEEE Asian Solid-State Circuits Conf. (A-SSCC), 2010, pp. 1–4.
[11] T.-C. Kuo and A. N. Willson Jr, “A flexible decoder IC for WiMAX
QC-LDPC codes,” in IEEE Custom Integrated Circuits Conf. (CICC),
2008, pp. 527–530.
[12] A. J. Blanksby and C. J. Howland, “A 690-mW 1-Gb/s 1024-b, rate-1/2
low-density parity-check code decoder,” IEEE J. Solid-State Circuits,
vol. 37, no. 3, pp. 404–412, 2002.
[13] N. Onizawa, T. Hanyu, and V. C. Gaudet, “Design of high-throughput
fully parallel LDPC decoders based on wire partitioning,” IEEE Trans.
VLSI Syst., vol. 18, no. 3, pp. 482–489, 2010.
[14] P. Schlafer, N. Wehn, M. Alles, and T. Lehnigk-Emden, “A new
dimension of parallelism in ultra high throughput LDPC decoding,” in
IEEE Int. Workshop on Signal Process. Syst. (SiPS), 2013, pp. 153–158.
[15] “IEEE Standard for Information Technology – Telecommunications and
Information Exchange between Systems – Local and Metropolitan Area
Networks – Specific Requirements Part 3: Carrier Sense Multiple Access
with Collision Detection (CSMA/CD) Access Method and Physical
Layer Specifications,” IEEE Std. 802.3an, Sep. 2006.
[16] K. Cushon, P. Larsson-Edefors, and P. Andrekson, “Low-power 400Gbps soft-decision LDPC FEC for optical transport networks,” J. Lightw.
Technol., vol. 34, no. 18, pp. 4304–4311, Aug. 2016.
[17] H. Kaeslin, Digital integrated circuit design: from VLSI architectures to
CMOS fabrication. Cambridge University Press, 2008.
[18] A. Darabiha, A. C. Carusone, and F. R. Kschischang, “A 3.3-Gbps bitserial block-interlaced min-sum LDPC decoder in 0.13-µm CMOS,” in
IEEE Custom Integrated Circuits Conf. (CICC), 2007, pp. 459–462.
[19] A. Darabiha, A. C. Carusone, and R. Kschischang, “Power reduction
techniques for LDPC decoders,” IEEE J. Solid-State Circuits, vol. 43,
no. 8, pp. 1835–1845, 2008.
[20] T. Mohsenin, D. N. Truong, and B. M. Baas, “A low-complexity
message-passing algorithm for reduced routing congestion in LDPC
decoders,” IEEE Trans. Circuits Syst. I, vol. 57, no. 5, pp. 1048–1061,
2010.
[21] N. Mobini, A. H. Banihashemi, and S. Hemati, “A differential binary
message-passing LDPC decoder,” IEEE Trans. Commun., vol. 57, no. 9,
pp. 2518–2523, 2009.
[22] K. Cushon, S. Hemati, C. Leroux, S. Mannor, and W. J. Gross, “Highthroughput energy-efficient LDPC decoders using differential binary
message passing,” IEEE Trans. Signal Process., vol. 62, no. 3, pp. 619–
631, 2014.
[23] C.-C. Cheng, J.-D. Yang, H.-C. Lee, C.-H. Yang, and Y.-L. Ueng, “A
fully parallel LDPC decoder architecture using probabilistic min-sum
algorithm for high-throughput applications,” IEEE Trans. Circuits Syst.
I, vol. 61, no. 9, pp. 2738–2746, 2014.
[24] S. S. Tehrani, W. Gross, and S. Mannor, “Stochastic decoding of LDPC
codes,” IEEE Commun. Lett., vol. 10, no. 10, pp. 716–718, 2006.
[25] S. S. Tehrani, A. Naderi, G.-A. Kamendje, S. Hemati, S. Mannor, and
W. J. Gross, “Majority-based tracking forecast memories for stochastic
LDPC decoding,” IEEE Trans. Signal Process., vol. 58, no. 9, pp. 4883–
4896, 2010.
[26] A. Naderi, S. Mannor, M. Sawan, and W. J. Gross, “Delayed stochastic
decoding of LDPC codes,” IEEE Trans. Signal Process., vol. 59, no. 11,
pp. 5617–5626, 2011.
[27] F. Angarita, J. Valls, V. Almenar, and V. Torres, “Reduced-complexity
min-sum algorithm for decoding LDPC codes with low error-floor,”
IEEE Trans. Circuits Syst. I, vol. 61, no. 7, pp. 2150–2158, 2014.
[28] “ISO/IEC/IEEE international standard for information technology–
telecommunications and information exchange between systems–local
and metropolitan area networks–specific requirements-part 11: Wireless
lan medium access control (MAC) and physical layer (PHY) specifications amendment 3: Enhancements for very high throughput in the
60 GHz band (adoption of IEEE Std 802.11ad-2012),” ISO/IEC/IEEE
8802-11:2012/Amd.3:2014(E), pp. 1–634, March 2014.
Reza Ghanaatian was born in Jahrom, Iran, in
1988. He received the M.Sc. degree in digital systems from the Department of Electrical Engineering,
Sharif University of Technology (SUT), Tehran,
Iran, in 2012. He was with Advanced Integrated
Circuit Design Laboratory (AICDL), at SUT from
2011 to 2013, working on field-programmable gate
array based systems for wireless and optical communication applications.
In 2014, Mr. Ghanaatian joined Telecommunications Circuits Laboratory (TCL) at EPFL, Lausanne,
Switzerland, working towards his Ph.D. His current research interests include
VLSI circuits for signal processing and communications as well as approximate computing techniques for energy efficient system design.
11
Alexios Balatsoukas-Stimming received the
Diploma and M.Sc. degrees in Electronics
and Computer Engineering from the Technical
University of Crete, Greece, in 2010 and 2012,
respectively. His M.Sc. studies were supported
by a scholarship from the Alexander S. Onassis
foundation. He received the Ph.D degree in
Computer and Communications Sciences from
EPFL, Switzerland, where he performed his research
at the Telecommunications Circuits Laboratory.
He serves as reviewer for several IEEE Journals
and Conferences, and has been recognized as an Exemplary Reviewer
by the IEEE Wireless Communications Letters in 2013 and 2014. His
current research interests include VLSI circuits for signal processing and
communications, as well as error correction coding theory and practice.
Thomas Christoph Müller (S’16) received the
bachelor’s degree in information technology from
the Schmalkalden University of Applied Sciences,
Schmalkalden, Germany, in 2011, and the master’s degree in System-on-Chip from Lund University, Sweden, in 2013. He has been working as
Project/Research Assistant at Lund University and
at Technical University of Denmark, Kongens Lyngby, Denmark. Currently he is pursuing his Ph.D.
degree in the Telecommunication Circuits Laboratory (TCL) at the École polytechnique fédérale de
Lausanne (EPFL), Switzerland with a research focus on digital implementation
and, an emphasis on low power, variation mitigation and design methodology.
Michael Meidlinger was born in Austria in 1989.
He received his Bsc. and Msc. degrees from Technische Universitt (TU) Wien, Vienna, Austria in 2011
and 2013 respectively, both with distinction. From
2011 to 2013, Michael has been working in the
field of mobile communication research and helped
to develop the Vienna LTE-A Simulators. Since
2013, Michael is part of the Communication Theory
group at TU Wien, working towards his Ph.D. His
current research interests include quantizer design
for telecommunication receivers as well as error
correction coding and superposition modulation techniques.
Gerald Matz received the Dipl.-Ing. (1994) and Dr.
techn. (2000) degrees in Electrical Engineering and
the Habilitation degree (2004) for Communication
Systems from Vienna University of Technology,
Austria. He currently holds a tenured Associate
Professor position with the Institute of Telecommunications, Vienna University of Technology. He
has held visiting positions with the Laboratoire des
Signaux et Systèmes at Ecole Supérieure dElectricité
(France, 2004), the Communication Theory Lab at
ETH Zurich (Switzerland, 2007), and with Ecole
Nationale Supérieure d’Electrotechnique, d’Electronique, d’Informatique et
d’Hydraulique de Toulouse (France, 2011).
Prof. Matz has directed or actively participated in several research projects
funded by the Austrian Science Fund (FWF), by the Viennese Science and
Technology Fund (WWTF), and by the European Union. He has published
some 200 scientific articles in international journals, conference proceedings,
and edited books. He is co-editor of the book Wireless Communications over
Rapidly Time-Varying Channels (New York: Academic, 2011). His research
interests include wireless networks, statistical signal processing, information
theory, and big data.
Prof. Matz served as as a member of the IEEE SPS Technical Committee
on Signal Processing Theory and Methods and of the IEEE SPS Technical
Committee on Signal Processing for Communications and Networking. He
was an Associate Editor of the IEEE Transactions on Information Theory
(2013-2015), of the IEEE Transactions on Signal Processing (20062010), of
the EURASIP Journal Signal Processing (20072010), and of the IEEE Signal
Processing Letters (20042008). He was Technical Program Chair of Asilomar
2016, Technical Program Co-Chair of EUSIPCO 2004, Technical Area Chair
for MIMO Communications and Signal Processing at Asilomar 2012, and
Technical Area Chair for Array Processing at Asilomar 2015. He has been
a member of the Technical Program Committee of numerous international
conferences. In 2006 he received the Kardinal Innitzer Most Promising Young
Investigator Award. He is an IEEE Senior Member and a member of the ÖVE.
Adam Teman received the Ph.D., M.Sc., and B.Sc.
degrees in Electrical Engineering from Ben-Gurion
University (BGU), Be’er Sheva, Israel in 2006, 2011,
and 2014, respectively. He worked as a Design
Engineer at Marvell Semiconductors from 2006 to
2007, with an emphasis on Physical Implementation.
Dr. Teman’s research interests include low-voltage
digital design, energy efficient SRAM, NVM, and
eDRAM memory arrays, low power CMOS image
sensors, low power design techniques for digital and
analog VLSI chips, significance-driven approximate
computing, and process tolerant design techniques. He has authored more
than 40 scientific papers and 4 patent applications and is an associate
editor at the Microelectronics Journal and a technical committee member
of several IEEE conferences. In 2010–2012, Dr. Teman was honored with
the Electrical Engineering Department’s Teaching Excellence recognition at
BGU, and in 2011, he was awarded with BGU’s Outstanding Project award.
Dr. Teman received the Yizhak Ben-Yaakov HaCohen Prize in 2010, the
BGU Rector’s Prize for Outstanding Academic Achievement in 2012, the
Wolf Foundation Scholarship for excellence of 2012 and the Intel Prize for
Ph.D. students in 2013. His doctoral studies were conducted under a Kreitman
Foundation Fellowship. Dr. Teman was a post-doctoral researcher at the
Telecommunications Circuits Lab (TCL) at the École Polytechnique Fédérale
de Lausanne (EPFL), Switzerland under a Swiss Government Excellence
Scholarship from 2014–2015. In October 2015, Dr. Teman joined the faculty
of engineering at Bar-Ilan University, Ramat Gan, Israel in 2015, where he is
currently a tenure track researcher in the department of electrical engineering
and a leading member of the Emerging Nanoscaled Integrated Circuits and
Systems (EnICS) Labs.
Andreas Burg (S’97-M’05) was born in Munich,
Germany, in 1975. He received his Dipl.-Ing. degree from the Swiss Federal Institute of Technology
(ETH) Zurich, Zurich, Switzerland, in 2000, and the
Dr. sc. techn. degree from the Integrated Systems
Laboratory of ETH Zurich, in 2006. In 1998, he
worked at Siemens Semiconductors, San Jose, CA.
During his doctoral studies, he worked at Bell Labs
Wireless Research for a total of one year. From
2006 to 2007, he was a postdoctoral researcher
at the Integrated Systems Laboratory and at the
Communication Theory Group of the ETH Zurich. In 2007 he co-founded
Celestrius, an ETH-spinoff in the field of MIMO wireless communication,
where he was responsible for the ASIC development as Director for VLSI.
In January 2009, he joined ETH Zurich as SNF Assistant Professor and as
head of the Signal Processing Circuits and Systems group at the Integrated
Systems Laboratory. Since January 2011, he has been a Tenure Track Assistant
Professor at the Ecole Polytechnique Federale de Lausanne (EPFL) where he
is leading the Telecommunications Circuits Laboratory.
In 2000, Mr. Burg received the Willi Studer Award and the ETH Medal for
his diploma and his diploma thesis, respectively. Mr. Burg was also awarded
an ETH Medal for his Ph.D. dissertation in 2006. In 2008, he received a
4-years grant from the Swiss National Science Foundation (SNF) for an SNF
Assistant Professorship. With his students he received the best paper award
from the EURASIP Journal on Image and Video Processing in 2013 and best
demo/paper awards at ISCAS 2013, ICECS 2013, and at ACSSC 2007.
He has served on the TPC of various conferences on signal processing,
communications, and VLSI. He was a TPC co-chair for VLSI-SoC 2012
and the TCP co-chair for ESSCIRC 2016 and SiPS 2017. He served as an
Editor for the IEEE Transaction of Circuits and Systems in 2013 and is on
the Editorial board of the Springer Microelectronics Journal and the MDPI
Journal on Low Power Electronics and its Applications. He is also a member
of the EURASIP SAT SPCN and of the IEEE TC-DISPS.
| 7cs.IT
|
1
Polar Codes and Polar Lattices for the
Heegard-Berger Problem
Jinwen Shi, Ling Liu, Deniz Gündüz, Senior Member, IEEE,
arXiv:1702.01042v2 [cs.IT] 5 Oct 2017
and Cong Ling, Member, IEEE
Abstract
Explicit coding schemes are proposed to achieve the rate-distortion function of the Heegard-Berger
problem using polar codes. Specifically, a nested polar code construction is employed to achieve the
rate-distortion function for the doubly-symmetric binary sources when the side information may be
absent. The nested structure contains two optimal polar codes for lossy source coding and channel
coding, respectively. Moreover, a similar nested polar lattice construction is employed when the source
and the side information are jointly Gaussian. The proposed polar lattice is constructed by nesting a
quantization polar lattice and a capacity-achieving polar lattice for the additive white Gaussian noise
channel.
Index Terms
Heegard-Berger Problem, source coding, lattices.
I. I NTRODUCTION
The well-known Wyner-Ziv problem is a lossy source coding problem in which a source
sequence is to be reconstructed in the presence of correlated side information at the decoder [1].
An interesting question is whether reconstruction with a non-trivial distortion quality can still
be obtained in the absence of the side information. The equivalent coding system contains two
decoders, one with the side information, and the other without, as shown by Fig. 1.
In 1985, Heegard and Berger [2] characterized the rate-distortion function RHB (D1 , D2 ) for
this scenario, where D1 is the distortion achieved without side information, D2 is the distortion
Jinwen Shi, Ling Liu, Deniz Gündüz and Cong Ling are with Dept. of Electronic and Electrical Engineering, Imperial College
London, SW7 2AZ, UK. (e-mails: {jinwen.shi12, l.liu12, d.gunduz}@imperial.ac.uk, [email protected]).
2
𝑋 1:𝑁
Encoder
2𝑁𝑅 bits
Decoder 1
𝑋11:𝑁
Decoder 2
𝑌1:𝑁
𝑋21:𝑁
Figure 1. Illustration of the Heegard-Berger rate-distortion problem.
achieved with it, and RHB (D1 , D2 ) denotes the minimum rate required to achieve the distortion
pair (D1 , D2 ). They also gave an explicit expression for the quadratic Gaussian case. Kerpez [3]
provided upper and lower bounds on the Heegard-Berger rate-distortion function (HBRDF) for
the binary case. Later, the explicit expression for RHB (D1 , D2 ) in the binary case was derived
in [4] together with the corresponding optimal test channel. Our goal in this paper is to propose
explicit coding schemes that can achieve the HBRDF for binary and Gaussian distributions.
The Heegard-Berger problem is a generalization of the classical Wyner-Ziv problem, in which
a source sequence is to be reproduced at the decoder within a certain distortion target, and the
side information available at the decoder is not available at the encoder. A nested construction
of polar codes is presented in [5] to achieve the binary Wyner-Ziv rate-distortion function. For
Gaussian sources, a polar lattice to achieve both the standard and Wyner-Ziv rate-distortion
functions is proposed in [6]. Different from the solutions of the Wyner-Ziv problem in [5] and
[6], we need to consider the requirements for the two decoders jointly. Therefore, we make use
of the low-fidelity reconstruction at Decoder 1, and combine it with the original source and the
side information to form the nested structure that achieves the optimal distortion for Decoder 2.
The optimality of polar codes for the lossy compression of nonuniform sources is shown in
[7]. We employ this scheme as part of our solution, since the optimal forward test channel may
be asymmetric in the binary Heegard-Berger problem. Furthermore, it is shown in [8] that polar
codes are optimal for general distributed hierarchical source coding problems. The HeegardBerger problem can also be considered as a successive refinement problem. In this paper, we
propose explicit coding schemes using polar codes and polar lattices to achieve the theoretical
performance bound in the Heegard-Berger problem. Practical codes for the Gaussian HeegardBerger problem are also developed in [9] which hybridize trellis and low-density parity-check
codes. However, the optimality of this scheme to achieve the HBRDF is not shown in [9].
The contributions of this paper can be summarized as follows:
•
We propose a nested construction of polar codes for the non-degenerate region of the binary
3
Heegard-Berger problem, and prove that they achieve the HBRDF for doubly symmetric
binary sources (DSBS). We consider the reconstruction of the source sequence at Decoder 1,
i.e., the decoder without side information, denoted by X̂11:N and the original source sequence
X 1:N as a combined source, and further combine this reconstruction X̂11:N with the original
side information Y 1:N to obtain a combined side information. By this argument, we obtain
another nested construction of polar codes, which achieves the HBRDF of the entire nondegenerate region. In addition, we present an explicit coding scheme by using two-level polar
codes to achieve the HBRDF whose forward test channel may be asymmetric. Finally, we
prove that polar codes achieve an exponentially decaying block error probability and excess
distortion at both decoders for the binary Heegard-Berger problem.
•
We then consider the Gaussian Heegard-Berger problem, and propose a polar lattice construction that consists of two nested polar lattices, one of which is additive white Gaussian noise (AWGN) capacity-achieving while the other is Gaussian rate-distortion function
achieving. This construction is similar to the one proposed for the Gaussian Wyner-Ziv
problem in [6]. However, in the Heegard-Berger problem setting, we need to treat the
difference between the original source and its reconstruction at Decoder 1 as a new source,
and the difference between the original side information and the reconstruction at Decoder
1 as a new side information. As a result, we can obtain an optimal test channel that connects
the new source with the new side information by using additive Gaussian noises. According
to this test channel, we can further construct two nested polar lattices that achieve the
Gaussian HBRDF of the entire non-degenerate region.
Organization: The paper is organized as follows: Section II presents the background on binary
and Gaussian Heegard-Berger problems. The construction of polar codes to achieve the HBRDF
for DSBS is investigated in Section III. In Section IV, the polar code construction for the Gaussian
Heegard-Berger problem is addressed. The paper is concluded in Section V.
Notation: All random variables are denoted by capital letters, while sets are denoted by capital
letters in calligraphic font. PX denotes the probability distribution of a random variable X taking
values in set X . For two positive integers i < j, xi:j denotes the vector (xi , . . . , xj ), which
represents the realizations of random variables X i:j . For a set F of positive integers, xF denotes
the subvector {xi }i∈F . For the Gaussian case, we construct polar codes in multiple levels, in
which Xl denotes a random variable at level l, and xil its i-th realization. Then, xli:j denotes
i
c
the vector xil , . . . , xjl , and xF
l denotes the subvector {xl }i∈F at the l-th level. F and |F|
4
denote the complement and cardinality of set F, respectively. For a positive integer N , we
define [N ] , {1, . . . , N }. 1 [x ∈ X ] denotes the indicator function, which equals 1 if x ∈ X
and 0 otherwise. Let I (X; Y ) denote the mutual information between X and Y . In this paper,
all logarithms are base two, and information is measured in bits.
II. P ROBLEM S TATEMENT
A. Heegard-Berger Problem
Let (X , Y, PXY ) be discrete memoryless sources (DMSs) characterized by the random variables X and Y with a generic joint distribution PXY over the finite alphabets X and Y.
Definition 1. An (N, M, D1 , D2 ) Heegard-Berger code for source X with side information Y
(N )
consists of an encoder f (N ) : X 1:N → [IM ] and two decoders g1
(N )
: [IM ] → X̂11:N ; g2
:
[IM ] × Y 1:N → X̂21:N , where X̂1 , X̂2 are finite reconstruction alphabets, such that
#
"
N
1 X j j
d X , X̂i
≤ Di , i = 1, 2,
E
N j=1
where E is the expectation operation, and d (·, ·) < ∞ is a per-letter distortion measure. In this
paper, we set d (·, ·) to be the Hamming distortion for binary sources, and the squared error
distortion for Gaussian sources.
Definition 2. Rate R is said to be {(D1 , D2 ) − achievable}, if for every > 0 and sufficiently
large N there exists an (N, M, D1 + , D2 + ) code with R + ≥
1
N
log M .
The HBRDF, RHB (D1 , D2 ), is defined as the infimum of (D1 , D2 )-achievable rates. A singleletter expression for RHB (D1 , D2 ) is given in the following theorem.
Theorem 3. ( [2, Theorem 1])
RHB (D1 , D2 ) =
min
[I (X; U1 ) + I (X; U2 |U1 , Y )] ,
(U1 ,U2 )∈P(D1 ,D2 )
where P (D1 , D2 ) is the set of all auxiliary random variables (U1 , U2 ) ∈ U1 × U2 jointly
distributed with the generic random variables (X, Y ), such that: i) Y ↔ X ↔ (U1 , U2 ) form
a Markov chain; ii) |U1 | ≤ |X | + 2 and |U2 | ≤ (|X | + 1)2 ; iii) there exist functions ϕ1 and ϕ2
such that E [d (X, ϕ1 (U1 ))] ≤ D1 and E [d (X, ϕ2 (U1 , U2 , Y ))] ≤ D2 .
5
1
D1
Region II
Region IV
0.5
Region I
I-B
I-A
Region III
D1 = D 2
0
dc
p
1
D2
Figure 2. Illustration of the HBRDF regions for DSBS, where dc is the critical distortion.
B. Doubly Symmetric Binary Sources
Let X be a binary DMS, i.e., X = {0, 1}, with uniform distribution. The binary side
information is specified by Y = X ⊕ Z, where Z is an independent Bernoulli random variable
with PZ (z = 1) = p < 0.5, and ⊕ denotes modulo two addition.
The HBRDF for DSBS can be characterized over four regions [3]. Region I (0 ≤ D1 < 0.5
and 0 ≤ D2 < min (D1 , p)) is a non-degenerate region, and RHB (D1 , D2 ) is a function of D1
and D2 ; Region II (D1 ≥ 0.5 and 0 ≤ D2 ≤ p) is a degenerate region as the Heegard-Berger
problem boils down to the Wyner-Ziv problem for the second decoder; Region III (0 ≤ D1 ≤ 0.5
and D2 ≥ min (D1 , p)) is also degenerate since the problem boils down to the standard lossy
compression problem for the first decoder; Region IV (D1 > 0.5 and D2 > p) can be trivially
achieved without coding. These four regions are depicted in Fig. 2. Note that, the HBRDF in
the degenerate Regions II and III can be achieved by using polar codes as described in [5]. Here
we focus on the non-degenerate Region I.
The explicit calculation of HBRDF for DSBS in Region I is given in [4] as follows:
Define the function
SD1 (α, µ, θ, θ1 ) , 1 − h (D1 ∗ p) + (θ − θ1 ) G (α) + θ1 G (µ) + (1 − θ) G (γ) ,
6
(u1 , x) = (0, 0)
u2 = 0
1
θ
2 1
u2 = 1
1
2
u2 = 2
1
2
(u1 , x) = (0, 1)
1
θ µ
2 1
(1 − µ)
(θ − θ1 ) α
(1 − θ) (1 − γ)
1
2
(u1 , x) = (1, 0)
1
2
1
2
(θ − θ1 ) (1 − α)
1
θ µ
2 1
(θ − θ1 ) (1 − α)
1
2
(u1 , x) = (1, 1)
(1 − θ) γ
1
2
(θ − θ1 ) α
1
θ
2 1
1
2
(1 − θ) γ
(1 − µ)
(1 − θ) (1 − γ)
Table I
J OINT DISTRIBUTION PX,U1 ,U2 (x, u1 , u2 ) [4].
where
γ,
D1 −(θ−θ1 )(1−α)−θ1 µ
θ 6= 1
0.5
θ=1
1−θ
,
on the domain 0 ≤ θ1 ≤ θ ≤ 1, 0 ≤ α, µ ≤ p, p ≤ γ ≤ 1 − p.
The following theorem characterizes the HBRDF in Region I.
Theorem 4. [4, Theorem 2] For 0 ≤ D1 < 0.5 and 0 ≤ D2 < min (D1 , p), we have
RHB (D1 , D2 ) = min SD1 (α, µ, θ, θ1 ), where the minimization is over all θ1 , θ, α and µ variables
that satisfy 0 ≤ θ1 ≤ θ ≤ 1, 0 ≤ α, µ ≤ p and (θ − θ1 ) α + θ1 µ + (1 − θ) p = D2 .
The corresponding forward test channel structure is also given in [4], reproduced in Table I. It
constructs random variables with joint distribution PX,U1 ,U2 (x, u1 , u2 ), which satisfy I (X; U1 ) +
I (X; U2 |U1 , Y ) = SD1 (α, µ, θ, θ1 ).
Next, we recall the function G (u) , h (p ∗ u) − h(u) from [1], defined over the domain
0 ≤ u ≤ 1, where h(u) is the binary entropy function h(u) , −u log u − (1 − u) log (1 − u),
and p ∗ u is the binary convolution for 0 ≤ p, u ≤ 1, defined as p ∗ u , p (1 − u) + u (1 − p).
Then, recall the definition of the critical distortion, dc , in the Wyner-Ziv problem for DSBS [1],
for which
G(dc )
dc −p
= G0 (dc ).
The following corollary from [4] specifies an explicit the characterization of HBRDF for
DSBS in Region I-B in Fig. 2 specified by D2 ≤ min (dc , D1 ) and D1 ≤ 0.5.
Corollary 5. ( [4, Corollary 2]) For distortion pairs (D1 , D2 ) satisfying D1 ≤ 0.5 and D2 ≤
min (dc , D1 ) (i.e., Region I-B in Fig. 2), we have
RHB (D1 , D2 ) = 1 − h (D1 ∗ p) + G (D2 ) .
(1)
From [4], the optimal forward test channel for Region I-B is given as a cascade of two binary
symmetric channels (BSCs), as depicted in Fig. 3.
7
Y
BSC(𝑝)
X
BSC(𝐷2 )
𝑈2
BSC(𝜂)
𝑈1
Figure 3. The optimal forward test channel for Region I-B. The crossover probability η for the BSC between U2 and U1 satisfies
D2 ∗ η = D1 .
In Section III, we first propose a polar code design that achieves the HBRDF in Region I-B
for DSBSs. We then provide a general polar code construction achieving the HBRDF in the
entire Region I.
C. Gaussian Sources
Suppose Y = X +Z, where X and Z are independent (zero-mean) Gaussian random variables
2
2
) and Z ∼ N (0, σZ2 ). The explicit
and σZ2 , respectively, i.e., X ∼ N (0, σX
with variances σX
expression for RHB (D1 , D2 ) in this case is given in [2]. The optimal test channels are given
by X = U1 + Z1 and X = U2 + Z2 , where Z, Z1 and Z2 are independent zero-mean Gaussian
random variables. We have Zi ∼ N (0, Di ) , i = 1, 2.
2
and D2 ≥
For D1 ≤ σX
2
D1 σZ
2 ,
D1 +σZ
the problem degenerates into a classical lossy compression
2
σ
2
problem for Decoder 1, and the HBRDF is given by RHB (D1 , D2 ) = 12 log DX1 . For D1 > σX
and D2 ≤
2
D1 σZ
2 ,
D1 +σZ
the problem degenerates into a Wyner-Ziv coding problem for Decoder 2,
2 σ2
σX
1
2
Z
and we have RHB (D1 , D2 ) = 2 log D σ2 +σ
. The region specified by D1 > σX
and
2
)
2( X
Z
2
D σ
D2 ≥ D11+σZ2 requires no coding. Polar lattice codes that meet the classical and Wyner-Ziv rateZ
distortion functions for Gaussian sources, introduced in [6], can be used to achieve the HBRDF
2
in these degenerate regions. The only non-degenerate distortion region is specified by D1 ≤ σX
and D2 ≤
2
D1 σZ
2 ,
D1 +σZ
and the HBRDF in this region is given by [2]:
2 2
1
σX
σZ
RHB (D1 , D2 ) = log
.
2
D2 (D1 + σZ2 )
(2)
We will focus on the construction of polar lattice codes that achieve the HBRDF in (2) in Section
IV.
III. P OLAR C ODES FOR DSBS
In this section, we present a construction of polar codes that achieves RHB (D1 , D2 ) for DSBS
in Region I. First, we give a brief overview of polar codes.
8
Let G2 , [ 11 01 ], and define GN , G⊗n
as the generator matrix of polar codes with length
2
N = 2n , where ‘⊗’ denotes the Kronecker product. A polar code CN (F, uF ) [5] is a linear
n
o
c
code defined by CN (F, uF ) = v 1:N GN : vF = uF , vF c ∈ {0, 1}|F | , for any F ⊆ [N ] and
uF ∈ {0, 1}|F | , where F is referred to as the frozen set. The code CN (F, uF ) is constructed
by fixing uF and varying the values in F c . Moreover, the frozen set can be determined by
the Bhattacharyya parameter [5]. For a binary memoryless asymmetric channel with input X ∈
X = {0, 1} and output Y ∈ Y, the Bhattacharyya parameter Z is defined as Z (X|Y ) ,
P p
2 y PX,Y (0, y) PX,Y (1, y).
A. Polar Code Construction for Region I-B
We observe from the proof of Theorem 4 in [4] that the auxiliary random variable U1 can
be considered as the output of a BSC with crossover probability D1 and input X. Therefore,
as for Region I-B, the minimum rate for Decoder 1 to achieve the target distortion D1 is R1 =
I (U1 ; X) = 1 − h (D1 ) . It is shown in [10, Theorem 3] that polar codes can achieve the ratedistortion function of binary symmetric sources. An explicit code construction is also provided in
[10]. Considering the source sequence X 1:N as N independent and identically distributed (i.i.d)
copies of X, we know from [10, Theorem 3] that Decoder 1 can recover a reconstruction X̂11:N
that is asymptotically close to U11:N as N becomes sufficiently large. Therefore, we assume that
both Decoder 1 and Decoder 2 can obtain U11:N in the following.
Decoder 2 observes the side information Y 1:N , in addition to U11:N that can be reconstructed
using the same method as Decoder 1. Hence, both Y 1:N and U11:N can be considered as side
information for Decoder 2 to achieve distortion D2 . Therefore, the problem at Decoder 2 is very
similar to Wyner-Ziv coding except that the decoder observes extra side information.
Recall that achieving the Wyner-Ziv rate-distortion function using polar codes is based on the
nested code structure proposed in [5]. Consider the Wyner-Ziv problem consisting of compressing
a source X 1:N in the presence of correlated side information Y 1:N using polar codes, where X
and Y are DSBS. The code Cs with corresponding frozen set Fs is designed to be a good source
code for distortion D2 . Further, the code Cc with corresponding frozen set Fc is designed to be
a good channel code for BSC(D2 ∗ p). It has been shown in [5] that Fc ⊇ Fs , because the test
channel BSC(D2 ∗ p) is degraded with respect to BSC(D2 ). In this case, the encoder transmits
to the decoder the sub-vector that belongs to the index set Fc \Fs . The optimality of this scheme
is proven in [5].
9
Similarly, the optimal rate-distortion performance for Decoder 2 in the Heegard-Berger problem can also be achieved by using nested polar codes. For (U1 , U2 ) ∈ P (D1 , D2 ), we have
I (X; U2 |U1 , Y ) = I (U2 ; X, U1 , Y ) − I (U2 ; Y, U1 )
(3)
= I (U2 ; X, U1 ) − I (U2 ; Y, U1 ) .
The second equality holds since Y ↔ X ↔ (U1 , U2 ) form a Markov chain. Motivated by (3),
the code Cs2 with corresponding frozen set Fs2 is designed to be a good source code for the
source pair X 1:N , U11:N with reconstruction U21:N . Ts denotes the test channel for this source
code. Additionally, the code Cc2 with corresponding frozen set Fc2 is designed to be a good
channel code for the test channel Tc with input U21:N and output Y 1:N , U11:N . According to [5,
Lemma 4.7], in order to show the nested structure between Cs2 and Cc2 , it is sufficient to show
that Tc is stochastically degraded with respect to Ts .
Definition 6. (Channel Degradation [5]). Let T1 : U → Y and T2 : U → X be two binary
discrete memoryless channels. We say that T1 is stochastically degraded with respect to T2 , if
there exists a discrete memoryless channel T : X → Y such that
X
PY |U (y|u) =
PX|U (x|u)PY |X (y|x).
x∈X
Proposition 7. Tc : U2 → (Y, U1 ) is stochastically degraded with respect to Ts : U2 → (X, U1 ),
if the random variables (X, Y, U1 , U2 ) agree with the forward test channel as shown in Fig. 3.
Proof: From the test channel structure in Fig. 3, Y ↔ X ↔ U2 ↔ U1 form a Markov
chain. By definition, we have PX,U1 |U2 (x, u1 |u2 ) = PX|U2 (x|u2 ) PU1 |U2 (u1 |u2 ) . We also have
PY,U1 |U2 (y, u1 |u2 ) = PY |U2 (y|u2 ) PU1 |U2 (u1 |u2 )
X
=
PX,Y |U2 (x, y|u2 ) PU1 |U2 (u1 |u2 )
x
=
X
PX|U2 (x|u2 ) PY |X,U2 (y|x, u2 ) PU1 |U2 (u1 |u2 )
x
=
X
PX,U1 |U2 (x, u1 |u2 ) PY |X (y|x) ,
x
completing the proof.
Therefore, we can claim that Fc2 ⊇ Fs2 by [5, Lemma 4.7], and rather than sending the entire
vector that belongs to the index set Fsc2 , the encoder sends only the sub-vector that belongs to
Fc2 \Fs2 to Decoder 2, since Decoder 2 can extract some information on U21:N from the available
10
side information U11:N , Y 1:N . As a result, the polar code construction for the Heegard-Berger
problem in Region I-B is given as follows:
Encoding: The encoder first applies lossy compression to source sequence X 1:N with reconstruction U11:N and corresponding average distortion D1 . We construct the code Cs1 = CN (Fs1 , 0̄) =
o
n
c
w1:N GN : wFs1 = 0̄, wFsc1 ∈ {0, 1}|Fs1 | , and the encoder transmits the compressed sequence
wFsc1 to the decoders. The encoder is also able to recover U11:N from Cs1 . Next, the encoder
applies lossy compression jointly for sources X 1:N , U11:N with reconstruction U21:N and target
distortion D2 and d (U1 , U2 ) = η. We then construct Cs2 = CN (Fs2 , 0̄). Finally, the encoder
applies channel coding to the symmetric test channel Tc with input U21:N and output Y 1:N , U11:N .
We derive Cc2 = CN Fc2 , uFc2 (v̄) , where uFc2 (v̄) is defined by uFs2 = 0̄ and uFc2 \Fs2 = v̄ for
v̄ ∈ {0, 1}|Fc2 \Fs2 | . The encoder sends the sub-vector u
to the decoders.
Fc2 \Fs2
= w1:N GN .
Decoding: Decoder 1 receives wFsc1 and outputs the reconstruction sequence u1:N
1
Decoder 2 receives uFc2 \Fs2 , and hence, it can derive uFc2 . Moreover, Decoder 2 can also recover
U11:N from wFsc1 . Decoder 2 applies the successive cancellation (SC) decoding algorithm to obtain
the codeword U21:N from the realizations of Y 1:N , U11:N .
Next we present the rates that can be achieved by the proposed scheme. From the polarization
theorem for lossy source coding in [10], we know that reliable decoding at Decoder 1 will be
|F c | N →∞
achieved with high probability if Ns1 −−−→ I (U1 ; X) = 1 − h (D1 ).
From the polarization theorems for source and channel coding [5], the code rate required for
reliable decoding at Decoder 2 can be derived by
|Fc2 | − |Fs2 | N →∞
−−−→I (U2 ; X, U1 ) − I (U2 ; Y, U1 )
N
= I (U2 ; X) + I (U1 ; X, U2 ) − I (U1 ; X) − I (U2 ; Y ) − I (U1 ; Y, U2 ) + I (U1 ; Y )
= G (D2 ) − G (D1 ) .
Therefore, the total rate will be asymptotically given by
|Fsc1 |+|Fc2 |−|Fs2 |
N
N →∞
−−−→ 1 − h (D1 ∗ p) +
G (D2 ) , for Region I-B.
Furthermore, according to [5], [10], the expected distortions asymptotically approach the target
values D1 and D2 at Decoders 1 and 2, respectively, as N becomes sufficiently large. The
encoding and decoding complexity of this scheme is O (N log N ).
Note that, in our scheme, the performance of Decoder 2 is more challenging than that of
Decoder 1. Thus, the simulation is conducted by fixing D1 = 0.35, p = 0.4, and varying
D2 ∈ (0, min (dc , D1 )). These settings satisfy the requirements for Region I-B. The performance
11
1
RHB(D1,D2)
0.9
Polar codes N=210
Polar codes N=212
0.8
Polar codes N=214
Polar codes N=216
RHB(D1,D2)
0.7
Polar codes N=218
0.6
0.5
0.4
0.3
0.2
0.1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
D2
Figure 4. Simulation performance of RHB (D1 , D2 ) corresponding to D2 for Region I-B.
curves are shown in Fig. 4 for n = 10, 12, 14, 16, 18. It shows that the performances achieved
by polar codes approaches the HBRDF as n increases.
B. Coding Scheme for Entire Region I
As Theorem 4 defines the RHB (D1 , D2 ) for the entire Region I, we now present a coding
scheme that can achieve the HBRDF for the entire Region I. Note that RHB (D1 , D2 ) of Region
I-B can be explicitly calculated by Corollary 5. Therefore, we can also achieve Region I-B
straightforwardly as shown in Section III-A.
From the optimal test channel structure shown in Table I, U2 is a ternary random variable,
i.e., U2 = {0, 1, 2}. Therefore, we express U2 as two binary random variables Ua and Ub ,
where U2 = 2Ub + Ua , i.e., (Ua , Ub ) ∈ {00, 10, 01}. For Decoder 1, we can apply the same
scheme specified in the previous subsection to achieve D1 . Again, U11:N and Y 1:N can be
considered as side information for Decoder 2. Then, the rate required to transmit U21:N is
evaluated as I (X; U2 |U1 , Y ) = I (X; Ua , Ub |U1 , Y ) = I (X; Ua |U1 , Y ) + I (X; Ub |U1 , Ua , Y ).
Accordingly we can design two separate coding schemes to achieve the rates I (X; Ua |U1 , Y )
and I (X; Ub |U1 , Ua , Y ), respectively.
Since Y ↔ X ↔ (U1 , Ua , Ub ) form a Markov chain, we have I (X; Ua |U1 , Y ) = I (Ua ; X, U1 )−
I (Ua ; Y, U1 ), and the test channel TCa : Ua → (Y, U1 ) is degraded with respect to TSa : Ua →
(X, U1 ). We can observe from Table I that Ua and Ub can be nonuniform.
12
Let K 1:N = Ua1:N GN , and for 0 < β < 0.5, the frozen set FSa (FCa ), the information set
ISa (ICa ), and the shaping set SSa (SCa ) can be identified as
o
n
β
FSa = i ∈ [N ]:Z K i |K 1:i−1 , X 1:N , U11:N ≥ 1 − 2−N
n
o n
o
β
β
ISa = i ∈ [N ]:Z K i |K 1:i−1 , X 1:N , U11:N < 1 − 2−N ∩ i ∈ [N ]:Z K i |K 1:i−1 > 2−N
o
n
β
SSa = i ∈ [N ]:Z K i |K 1:i−1 ≤ 2−N ,
n
o
β
FCa = i ∈ [N ]:Z K i |K 1:i−1 , Y 1:N , U11:N ≥ 1 − 2−N
o
o n
n
β
β
ICa = i ∈ [N ]:Z K i |K 1:i−1 , Y 1:N , U11:N ≤ 2−N ∩ i ∈ [N ]:Z K i |K 1:i−1 ≥ 1 − 2−N
n
o
β
SCa = i ∈ [N ]:Z K i |K 1:i−1 < 1 − 2−N
n
o
−N β
i
1:i−1
1:N
1:N
−N β
∪ i ∈ [N ]: 2
< Z K |K
, Y , U1
<1−2
.
(4)
By [5, Lemma 4.7] and channel degradation, we have FSa ⊆ FCa , ICa ⊆ ISa and SSa ⊆ SCa .
In addition, we observe that SCa \SSa can be written as
n
o
β
β
SCa \SSa = i ∈ [N ]:2−N < Z K i |K 1:i−1 < 1 − 2−N ∪
n
o
β
β
i ∈ [N ]: 2−N < Z K i |K 1:i−1 , Y 1:N , U11:N < 1 − 2−N ,
therefore, the proportion
|SCa \SSa |
N
→ 0, as N → ∞.
Encoding: The encoder first applies lossy compression to X 1:N with target distortion D1 to
obtain U11:N , and treats X 1:N , U11:N as a joint source sequence to evaluate KISa by randomized
rounding with respect to PK i |K 1:i−1 ,X 1:N ,U11:N , i.e.,
0 w.p. P i 1:i−1 1:N 1:N 0|k 1:i−1 , x1:N , u1:N
K |K
,X
,U1
1
i
k =
if i ∈ ISa
1 w.p. PK i |K 1:i−1 ,X 1:N ,U 1:N 1|k 1:i−1 , x1:N , u1:N
1
1
and
ki =
k̃ i
if i ∈ FSa
(5)
(6)
arg maxk PK i |K 1:i−1 (k|k 1:i−1 ) if i ∈ SSa ,
where ‘w.p.’ is an abbreviation of ‘with probability’ in (5), and k̃ i is chosen uniformly from
{0, 1} and shared between the encoder and the decoders before lossy compression. Also note
that the second formula in (6) is in fact the maximum a posteriori (MAP) decision for i ∈ SSa .
The encoder sends KISa \ICa to the decoders.
13
Decoding: Using the pre-shared KFSa and received KISa \ICa , Decoder 2 recovers KICa and
KSSa from the side information sequences Y 1:N and U11:N by SC decoding algorithm and the
MAP rule, respectively. Hence we obtain K 1:N . KICa ∪SSa and KSCa \SSa can be recovered with
vanishing error probability, since their Bhattacharyya parameters are arbitrarily small when N →
∞. Therefore, the reconstruction is given by Ua1:N = K 1:N GN .
Theorem 8. [6, Theorem 2] Let QK 1:N ,X 1:N ,U11:N denote the resulting joint distribution derived
from (5) and (6). Let PK 1:N ,X 1:N ,U11:N denote the joint distribution as a result of another encoder
that uses (5) for i ∈ [N ]. For any β 0 < β < 0.5 satisfying (4) and Ra =
|ISa \ICa |
N
>
I (X; Ua |U1 , Y ), we have
0
−N β
,
V PK 1:N ,X 1:N ,U11:N , QK 1:N ,X 1:N ,U11:N = O 2
where V(PX , QX ) ,
1
2
P
x
(7)
|PX (x) − QX (x)| denotes the variational distance between distribu-
tions PX and QX .
Note that, from [7], KSCa should be covered by a pre-shared random mapping to achieve (7).
However, it is shown in [6, Theorem 2] that replacing the random mapping with MAP decision
for KSSa preserves the optimality. Thus, we utilize MAP decoder if i ∈ SSa in our scheme.
In terms of the second level, the encoder and Decoder 2 first recover Ua1:N . Consequently, the
encoder treats X 1:N , U11:N , Ua1:N as a joint source, and Decoder 2 treats Y 1:N , U11:N , Ua1:N as a
joint side information. Likewise, according to Y ↔ X ↔ (U1 , Ua , Ub ), we have I (X; Ub |U1 , Ua , Y ) =
I (Ub ; X, U1 , Ua ) − I (Ub ; Y, U1 , Ua ) and the test channel TCb : Ub → (Y, U1 , Ua ) is degraded with
respect to TSb : Ub → (X, U1 , Ua ).
Similar to the first level, let W 1:N = Ub1:N GN and the frozen set FSb (FCb ), the information
set ISb (ICb ), and the shaping set SSb (SCb ) can be adopted from (4) by replacing K, (X, U1 )
and (Y, U1 ) with W , (X, U1 , Ua ) and (Y, U1 , Ua ), respectively. As a result, we have FSb ⊆ FCb ,
ICb ⊆ ISb and SSb ⊆ SCb by [5, Lemma 4.7] and channel degradation. The encoder evaluates
WISb by randomized rounding with respect to PW i |W 1:i−1 ,X 1:N ,U11:N ,Ua1:N , WFSb are pre-shared
random bits uniformly chosen from {0, 1}, and WSSb is determined by MAP decoder defined as
arg maxw PW i |W 1:i−1 (w|w1:i−1 ). The encoder sends WISb \ICb to the decoders. Decoder 2 recovers
WICb ∪SSb using the pre-shared WFSb and the side information Y 1:N , U11:N , Ua1:N . Finally, the
reconstruction is given by Ub1:N = W 1:N GN .
14
Let QW 1:N ,X 1:N ,U11:N ,Ua1:N denote the joint distribution when the encoder performs compression,
according to the coding scheme presented in the above paragraph. Let PW 1:N ,X 1:N ,U11:N ,Ua1:N
denote the resulting joint distribution of the encoder using randomized rounding with respect to
PW i |W 1:i−1 ,X 1:N ,U11:N ,Ua1:N for all i ∈ [N ], which means that the encoder dose not perform compres|IS \IC |
sion. Similarly to Theorem 7, for any β 0 < β < 0.5 and Rb = bN b > I (X; Ub |U1 , Ua , Y ),
we have
V PW 1:N ,X 1:N ,U11:N ,Ua1:N , QW 1:N ,X 1:N ,U11:N ,Ua1:N
−N β
0
=O 2
.
(8)
Note that (8) is based on (7), and Ra > I (X; Ua |U1 , Y ) should be satisfied. Thus, we have
Ra + Rb > I (X; U2 |U1 , Y ). With regard to Decoder 2, we can state the following theorem.
Theorem 9. Consider a target distortion 0 ≤ D2 < min (D1 , p) for DSBS X when side
information Y is available only at the Decoder 2. For any 0 < β 0 < β < 0.5, there exists
a two-level polar code with a rate arbitrarily close to I (X; U2 |U1 , Y ), such that the expected
β0
distortion DQ of Decoder 2 satisfies DQ ≤ D2 + O 2−N .
Proof: See Appendix A.
As for Decoder 1, we know that U1 can always be taken as the output of a BSC with crossover
probability D1 and input X. Hence, according to [10, Theorem 3] and Theorem 9, this coding
scheme can achieve the optimal HBRDF, as long as the optimal parameters α, µ, θ, and θ1 that
achieve the minimum value of SD1 (α, µ, θ, θ1 ) can be specified. Finally, we state the achievability
of the HBRDF for DSBS for the entire Region I in the following theorem.
Theorem 10. Consider target distortions 0 ≤ D1 < 0.5 and 0 ≤ D2 < min (D1 , p) for DSBS
X when side information Y is available only at Decoder 2. For any 0 < β 0 < β < 0.5 and
any rate R > min SD1 (α, µ, θ, θ1 ), there exist a polar code C1 with rate R1 < I(X, U1 ) and
a two-level polar code C2 with rate R2 < I (X; U2 |U1 , Y ), with R1 + R2 < R, which together
0
−N β
−N β
achieve the expected distortions D1 + O 2
at Decoder 1 and D2 + O 2
at Decoder
2, respectively, if PX,U1 ,U2 is as given in Table I.
Proof: It has been shown in [10, Theorem 3] that there exists a polar code with a rate
β
arbitrarily close to I(X, U1 ) that achieves an expected distortion D1 + O 2−N . Theorem
9 shows that a two-level polar code with a rate arbitrarily close to I (X; U2 |U1 , Y ) achieves
0
−N β
D2 + O 2
at Decoder 2. Finally, the total rate R1 + R2 < I(X, U1 ) + I (X; U2 |U1 , Y ) =
15
min SD1 (α, µ, θ, θ1 ). The last equality holds if the joint distribution PX,U1 ,U2 is the same as that
given in Table I [4].
Finally, we observe that the encoding and decoding complexity of this coding scheme is
O (N log N ).
IV. P OLAR L ATTICES FOR G AUSSIAN S OURCES
It is shown in [6] that polar lattices achieve the optimal rate-distortion performance for both
the standard and the Wyner-Ziv compression of Gaussian sources under squared-error distortion.
The Wyner-Ziv problem for the Gaussian case can be solved by a nested code structure that
combines AWGN capacity achieving polar lattices [11] and the rate-distortion optimal ones
[6]. Here we show that the HBRDF for the non-degenerate region specified in (2) can also be
achieved by a similar nested code structure.
We start with a basic introduction to polar lattices. An n-dimensional lattice is a discrete
subgroup of Rn which can be described by
Λ = {λ = Bz : z ∈ Zn },
where B is the full rank generator matrix. For σ > 0 and c ∈ Rn , the Gaussian distribution of
variance σ 2 centered at c is defined as
kx−ck2
1
fσ,c (x) = √
e− 2σ2 , x ∈ Rn .
( 2πσ)n
Let fσ,0 (x) = fσ (x) for short. The Λ-periodic function is defined as
fσ,Λ (x) =
X − kx−λk2
1
fσ,λ (x) = √
e 2σ2 .
n
(
2πσ)
λ∈Λ
λ∈Λ
X
Note that, when x is restricted to the fundamental region R (Λ), fσ,Λ (x) is actually a probability
density function (PDF) of the Λ-aliased Gaussian noise [12].
The flatness factor of a lattice Λ is defined as
Λ (σ) , max |V (Λ)fσ,Λ (x) − 1|,
x∈R(Λ)
where V (Λ) = |det (B)| denotes the volume of a fundamental region of Λ [12]. It can be
interpreted as the maximum variation of fσ,Λ (x) with respect to the uniform distribution over a
fundamental region of Λ.
16
We define the discrete Gaussian distribution over Λ centered at c as the discrete distribution
taking values in λ ∈ Λ as
DΛ,σ,c (λ) =
where fσ,c (Λ) =
P
λ∈Λ
fσ,c (λ)
, ∀λ ∈ Λ,
fσ,c (Λ)
fσ,c (λ). For convenience, we write DΛ,σ = DΛ,σ,0 . It has been shown
in [13] that lattice Gaussian distribution preserves many properties of the continuous Gaussian
distribution when the flatness factor is negligible. To keep the notations simple, we always set
c = 0 and n = 1.
A sublattice Λ0 ⊂ Λ induces a partition (denoted by Λ/Λ0 ) of Λ into equivalence groups
modulo Λ0 . The order of the partition equals the number of the cosets. If the order is two, we
call this a binary partition. Let Λ (Λ0 ) /Λ1 / · · · /Λr−1 /Λ0 (Λr ) for r ≥ 1 be an n-dimensional
lattice partition chain. For each partition Λl−1 /Λl (1 ≤ l ≤ r) a code Cl over Λl−1 /Λl selects
a sequence of coset representatives al in a set Al of representatives for the cosets of Λl . This
construction requires a set of nested linear binary codes Cl with block length N and dimension
of information bits kl , and C1 ⊆ C2 · · · ⊆ Cr .
For the Gaussian Heegard-Berger problem, let (X, Y, Z, Z1 , Z2 , U1 , U2 ) be chosen as specified
in Section II-C. Given Y as the side information for Decoder 2, the HBRDF is given by (2). To
achieve the HBRDF at Decoder 1, we can design a quantization polar lattice for source X with
2
variance σX
and target distortion D1 as in [6]. As a result, for a target distortion D1 and any rate
R1 >
1
2
2
log (σX
/D1 ), there exists a multilevel polar lattice with rate R1 , such that the average
distortion is asymptotically close to D1 when the length N → ∞ and the number of levels
r = O (log log N ) [6, Theorem 4]. Therefore, both decoders can recover U11:N and U11:N , Y 1:N
can be regarded as the side information at Decoder 2.
As for Decoder 2, we first need a code that achieves the rate-distortion requirement for source
X 0 , X − U1 with Gaussian reconstruction alphabet U 0 . In fact, X 0 = Z1 ∼ N (0, D1 ) is
Gaussian and independent of U1 and Z. Let
γ,
D1 σZ2
,
D1 σZ2 − D2 (D1 + σZ2 )
and consider an auxiliary Gaussian random variable U 0 defined as U 0 = X 0 + Z4 , where Z4 ∼
N (0, γD2 ). Moreover, we define Y 0 , Y − U1 = X 0 + Z and Y 0 ∼ N (0, D1 + σZ2 ). Then we
can apply the minimum mean square error (MMSE) rescaling parameter α =
D1
2
D1 +σZ
to Y 0 . As
a result, we obtain X 0 = αY 0 + Z3 , where Z3 ∼ N (0, ασZ2 ). We can also write U 0 = αY 0 + Z5 ,
17
U′~N 0, D1 + D′2
⨁
X′~N 0, D1
Z4 ~N 0, D′2
⨁
αY′~N 0,
Z3 ~N 0,
Rate-distortion-achieving code
D12
D1 + σ2Z
D1 σ2Z
D1 + σ2Z
Capacity-achieving code
Figure 5. A test channel for the Gaussian Heegard-Berger problem for Decoder 2 using a continuous Gaussian U 0 .
where Z5 ∼ N (0, γD2 + ασZ2 ), which requires an AWGN capacity-achieving code from αY 0 to
U 0 . This test channel is depicted in Fig. 5.
The final reconstruction at Decoder 2 is given by X̂2 = U1 + αY 0 +
that
1
γ
1
γ
(U 0 − αY 0 ) . Note
(U 0 − αY 0 ) is a scaled version of Z5 , which is independent of Y 0 . Thus, the variance
of αY 0 +
1
γ
(U 0 − αY 0 ) is αD1 +
1
γ2
(γD2 + ασZ2 ) = D1 − D2 . Therefore, we have X − U1 =
X̂2 − U1 + N (0, D2 ) as we desired. Furthermore,
the required
data rate for Decoder 2 is then
2
D σZ
given by R2 > I (U 0 ; X 0 ) − I (U 0 ; αY 0 ) = 21 log D D1 +σ
.
2
2( 1
Z)
Note that U 0 is a continuous Gaussian random variable which is impractical for the design of
polar lattices. Hence, we use the discrete Gaussian distribution DΛ,σ2 0 to replace it. Before that,
U
0
0
we need to perform MMSE rescaling on U for test channels X → U 0 and Y 0 → U 0 with scales
αq and αc , respectively. Consequently, a reversed version of the test channel in Fig. 5 can be
derived, as depicted in Fig. 6, where
αq =
D1 (σZ2 − D2 ) − D2 σZ2
D1
=
D1 + D20
D1 (σZ2 − D2 )
and
αc =
D12
D1 (σZ2 − D2 ) − D2 σZ2
=
.
(D1 + D20 ) (D1 + σZ2 )
(D1 + σZ2 ) (σZ2 − D2 )
The reconstruction of X at Decoder 2 is as given in the following proposition.
Proposition 11. If we use the reversed test channel shown in Fig 6, the reconstruction of X at
Decoder 2 is given by
0
X̂2 = U1 + αq U + η
D2
αq
0
0
αY − αq U , η = 2 .
αc
σZ
Proof: It suffices to prove that X − U1 = X̂2 − U1 + N (0, D2 ). Since we have X 0 =
αq U 0 + N (0, αq D20 ), as illustrated in Fig 6, showing X̂2 − U1 = αq U 0 + N (0, αq D20 − D2 )
18
α𝑞 U′
⨁
Z3′ ~N 0,
Z4′ ~N 0, α𝑞 D′2
𝐴~𝐷Λ,σ2𝑎
α𝑞
αY′
α𝑐
⨁
X′~N 0, D1
⨁
α𝑞
αY′
α𝑐
⨁
X′
α𝑞 2
σ
α𝑐 Z3
Z3′
Z4′
Rate-distortion-achieving 𝐿1
Capacity-achieving 𝐿2
Figure 6. A reversed solution of the test channels in order to construct polar lattices.
would complete the proof. We can see from Fig 6 that
with zero mean and variance αq D20 +
αq 2
σ ,
αc Z3
αq
αY 0
αc
− αq U 0 is Gaussian distributed
and it is independent of U 0 . We also have
D1 D2 σZ2
D1 (σZ2 − D2 ) − D2 σZ2
− D2
D1 (σZ2 − D2 )
D1 σZ2 − D2 (D1 + σZ2 )
D2
= 2 2 ,
σZ − D2
αq D20 − D2 =
and
2
αq 2
D2 σZ2
D1 + σZ2 D1 σZ2
D2
0
η αq D2 + σZ3 =
+
αc
σZ2
(σZ2 − D2 )
D1 D1 + σZ2
D2
= 2 2 ,
σZ − D2
2
as we desired.
Based on the reversed test channel, we can replace the continuous Gaussian random variable
αq U 0 with a discrete Gaussian distributed variable A ∼ DΛ,σa2 , where σa2 = αq2 σU2 0 . Let X̄ 0 ,
A + N (0, αq D20 ) and ααqc αY¯ 0 = X̄ 0 + N 0, ααqc σZ2 3 . We also define B̄ , ααqc αY¯ 0 , whose variance
is σb2 =
α2q 2 2
α σY 0 .
α2c
By [6, Lemma 1], the distributions of X̄ 0 and Y¯ 0 can be made arbitrarily
close to the distributions of X 0 and Y 0 , respectively. Therefore, polar lattices can be designed
for the source X̄ 0 and the side information Y¯ 0 at Decoder 2. Specifically, a rate-distortion boundachieving polar lattice L1 is constructed for the source X̄ 0 with distortion αq D20 , and an AWGN
capacity-achieving polar lattice L2 is constructed for the channel A → ααqc αY¯ 0 , as shown in Fig.
6. In the end, the reconstruction of Decoder 2 is X̌2 = U1 + A + η B̄ − A . Even though
the quantization noise of L1 is not an exact Gaussian distribution, it is shown in [6, Theorem
19
4] that the two distributions can be arbitrarily close when N is sufficiently large. Therefore,
B̄ − A can be treated as Gaussian noise independent of A. By Proposition 11, η scales B̄ − A
to N (0, αq D20 − D2 ) and by [6, Lemma 1], the distributions of X̌2 and X̂2 can be arbitrarily
close, which gives an average distortion close to D2 .
According to [11, Lemma 10], L1 and L2 can be equivalently constructed for the MMSErescaled channel with Gaussian noise variances
σa2 σZ2 D2
σ2
,
σ˜q2 = a αq D20 =
D1
D1 (σZ2 − D2 )
and
σ2
σ˜c2 = a2
σb
αq 2
σa2 σZ4
0
.
αq D2 + σZ3 =
αc
(D1 + σZ2 ) (σZ2 − D2 )
The coding strategy for L1 and L2 can be adapted from [6, Section V]. We briefly describe
it for completeness. First, choose a good constellation DΛ,σa2 such that the flatness factor Λ (σ˜q2 )
is negligible. Let Λ/Λ1 / · · · /Λr−1 /Λr / · · · denote a one-dimensional binary partition chain
labeled by bits A1 /A2 / · · · /Ar−1 /Ar / · · · . Therefore, PA1:r and A1:r approaches DΛ,σa2 and A,
respectively, as r → ∞. Consider N i.i.d copies of A, and let Kl1:N , A1:N
GN for each level
l
1 ≤ l ≤ r. For 0 < β < 0.5, the frozen set FlQ FlC , information set IlQ IlC , and shaping set
SlQ SlC for L1 (L2 ) at level l can be adapted from [6, Equation (35)] and [6, Equation (36)],
respectively, by replacing X̄ with X̄ 0 .
Furthermore, according to [6, Lemma 2], L2 is nested within L1 , i.e., L2 ⊆ L1 . By the fact
σ˜q2 ≤ σ˜c2 and [11, Lemma 3], the partition channel Λl−1 /Λl with noise variance σ˜c2 is degraded
with respect to the one with noise variance σ˜2 . Therefore, we have F Q ⊆ F C , I C ⊆ I Q , and
q
by the definition of shaping set, we observe that
l
SlQ
l
l
l
⊆ SlC .
The encoder can recover the auxiliary codeword U11:N for Decoder 1, and obtains the real
izations x01:N y 01:N of X 01:N = X 1:N − U11:N Y 01:N = Y 1:N − U11:N from given realizations
of variables X 1:N Y 1:N , respectively. The encoder recovers kl1:N from l = [r] successively
according to the random rounding quantization rules given in [6, Equations (13), (14), (17) and
(18)]. Note that x01:N as realization of X̄ 0 1:N is acceptable since the distributions of X 0 and X̄ 0
are arbitrarily close. Also, according to [6, Theorem 2], replacing the random rounding rule with
SQ
MAP decision to obtain kl l will not affect [11, Theorem 5] and [11, Theorem 6]. Consequently,
the coding scheme for Decoder 2 for the Gaussian Heegard-Berger problem can be summarized
as following:
20
Encoding: From the N -dimensional i.i.d. source vector X 1:N , the encoder recovers the aux2
iliary codeword U11:N employing a quantization polar lattice for source X with variance σX
IQ
and distortion D1 , and obtains X 01:N and Y 01:N . Next, the encoder evaluates Kl l by random
I Q \IlC
rounding and sends Kl l
to the decoders.
FlQ
Decoding: By the pre-shared Kl
I Q \IlC
and received Kl l
SQ
IC
, Decoder 2 recovers Kl l and Kl l
from the side information B̄ 1:N with vanishing error probability, by using SC decoding for
Gaussian channels [11]. At each level, Decoder 2 obtains Kl1:N , and A1:N can be recovered
according to [6, Equation (38)]. Finally, the reconstruction of Decoder 2 is
X̌2
1:N
= U11:N + A1:N + η B̄ 1:N − A1:N .
(9)
According to [11, Lemma 8], the encoding and decoding complexities of polar lattices remain
to be O (N log N ).
As for the transmission rate of this scheme, the rate R1 for Decoder 1 can be arbitrarily close
2
σ
1
to 2 log DX1 according to [6, Theorem 4]. By the same argument, the rate RL1 of L1 can be
1
when the flatness factor is negligible. By [11, Theorem 7], the
arbitrarily close to 12 log αD
0
q D2
σb2
1
rate RL2 of the capacity-achieving lattice L2 can be arbitrarily close to 2 log α D0 + αq σ2
with
q
2
αc
Z3
a negligible flatness factor. Since L2 ⊆ L1 , the rate for Decoder 2 after some tedious calculations
is given by
R2 = RL1 − RL2
D1 (σZ2 − D2 )
(D1 + σZ2 ) (σZ2 − D2 )
1
1
− log
→ log
2
D2 σZ2
2
σZ4
D1 σZ2
1
→ log
,
2
D2 (D1 + σZ2 )
and the total rate for the Gaussian Heegard-Berger problem is
2 2
σX
1
σZ
R1 + R2 → log
,
2
D2 (D1 + σZ2 )
which is the same as (2).
Next, we give the main theorem of the Gaussian Heegard-Berger problem for the nondegenerate region.
Theorem 12. Let (X, Y, Z, D1 , D2 ) be as specified in Section II-C. For any rate R1 > 12 log
2
σX
D1
,
there exists a polar lattice code at rate R1 with sufficiently large blocklength, whose expected
distortion is arbitrarily close to D1 and the number of partition levels is O (log log N ). Let
Λ/Λ1 / · · · /Λr−1 /Λr be a one-dimensional binary partition chain of a lattice Λ such that Λ (σ˜2 ) =
q
21
√
− N
and r = O (log N ). For any 0 < β 0 < β < 0.5, there exist nested polar lattices L1
2
D1 σZ
1
and L2 with a rate spread R2 = RL1 − RL2 arbitrarily close to 2 log D D +σ2
such that
2( 1
Z)
β0
the expected distortion DQ2 satisfies DQ2 ≤ D2 + O 2−N .
O 2
Proof: The achievability of the rate-distortion function for Decoder 1 follows from [6,
Theorem 4]. The proof of the achievability for Decoder 2 can be adapted from [6, Theorem 5],
by considering the test channel depicted in Fig. 6 and the reconstruction as given by (9). It is
√
2
˜
worth mentioning that the requirements Λ (σ ) = O 2− N and r = O (log N ) are given by [6,
q
Proposition 2] to guarantee a sub-exponentially decaying error probability for the lattice design
for Decoder 2.
V. C ONCLUSION
We presented nested polar codes and polar lattices that achieve the rate-distortion function
for the binary and Gaussian Heegard-Berger problems, respectively. Different from the code
constructions for the Wyner-Ziv problem [5] and [6], we took advantage of the reconstruction at
Decoder 1 to build the nested structure that achieves the rate-distortion function for Decoder 2.
The proposed schemes achieve the HBRDF in the entire non-degenerate regions for both DSBS
and Gaussian sources.
Finally, the Kaspi problem in [14] is regarded as a generalization of the Heegard-Berger
problem, where the encoder may also have access to the side information. The explicit ratedistortion functions for the Kaspi problem with Gaussian and binary sources have been given in
[15] and [16], respectively. We will study the construction of polar codes and polar lattices for
the Kaspi problem in our future work.
A PPENDIX A
P ROOF OF T HEOREM 9
First, we show that the distortion D2 can be achieved. Since Ub1:N = W 1:N GN gives a oneto-one mapping between W 1:N and Ub1:N , expression (8) is equivalent to
β0
.
V PUa1:N ,Ub1:N ,X 1:N ,U11:N , QUa1:N ,Ub1:N ,X 1:N ,U11:N = O 2−N
(10)
From the coding scheme presented in Section III-B, we assume that KICa and WICb can be correctly decoded by using side information, and KSSa and WSSb can be recovered by the MAP rule.
22
Therefore, Decoder 2 can recover Ua1:N and Ub1:N with the joint distribution QUa1:N ,Ub1:N ,X 1:N ,U11:N .
Denote by QUa1:N ,Ub1:N ,X 1:N ,U11:N ,Y 1:N the resulting distribution when the encoder performs compression at each level, i.e., compresses X 1:N to WISb . Let PUa1:N ,Ub1:N ,X 1:N ,U11:N ,Y 1:N denote the
joint distribution when the encoder does not perform compression. For simplicity, we denote
1:N
random variables U11:N , Ua1:N and Ub1:N by U1,a,b
.
1:N ,X 1:N ,Y 1:N , QU 1:N ,X 1:N ,Y 1:N
2V PU1,a,b
1,a,b
X
1:N
1:N
, y 1:N
=
, y 1:N − Q u1:N
P u1:N
1,a,b , x
1,a,b , x
1:N ,y 1:N
u1:N
1,a,b ,x
X
=
1:N
1:N
1:N
1:N
P u1:N
P y 1:N |u1:N
− Q u1:N
Q y 1:N |u1:N
.
1,a,b , x
1,a,b , x
1,a,b , x
1,a,b , x
1:N ,y 1:N
u1:N
1,a,b ,x
1:N
According to the Markov chain Y ↔ X ↔ U1,a,b
, we have
1:N
1:N
P y 1:N |u1:N
= Q y 1:N |u1:N
= P y 1:N |x1:N .
1,a,b , x
1,a,b , x
Therefore,
1:N ,X 1:N ,Y 1:N , QU 1:N ,X 1:N ,Y 1:N
V PU1,a,b
1,a,b
0
−N β
1:N ,X 1:N , QU 1:N ,X 1:N
=O 2
= V PU1,a,b
.
1,a,b
(11)
1:N
), and the average
The reconstructions of two levels are Ua1:N and Ub1:N (i.e., denoted by Ua,b
1:N ,X 1:N ,Y 1:N is given by
distortion DP achieved by PU1,a,b
DP =
=
=
1
N
1
N
X
1:N
1:N
1:N ,X 1:N ,Y 1:N
, y 1:N d Ua,b
, x1:N
PU1,a,b
u1:N
1,a,b , x
1:N ,y 1:N
u1:N
1,a,b ,x
X
1:N
1:N
1:N ,X 1:N
d Ua,b
, x1:N
PUa,b
u1:N
a,b , x
1:N
u1:N
a,b ,x
X
1
N
PU ,X (ua,b , x) d (Ua,b , x)
N u ,x a,b
a,b
= D2 .
Note that the last equality holds due to the constrains (θ − θ1 ) α + θ1 µ + (1 − θ) p = D2 of
Theorem 4. This is reasonable because DP is achieved when the encoder does not perform
23
1:N ,X 1:N ,Y 1:N
any compression. Combined with (11), the expected distortion DQ achieved by QU1,a,b
satisfies
1
DQ − DP =
N
≤
X
1:N
1:N ,X 1:N ,Y 1:N − PU 1:N ,X 1:N ,Y 1:N
QU1,a,b
d Ua,b
, x1:N
1,a,b
1:N ,y 1:N
u1:N
1,a,b ,x
1
N dmax
N
X
1:N ,X 1:N ,Y 1:N − QU 1:N ,X 1:N ,Y 1:N
PU1,a,b
1,a,b
1:N ,y 1:N
u1:N
1,a,b ,x
β0
.
= O 2−N
i∈ICa
Next we show that the decoder can recover Ua
i∈ICb
and Ub
with a sub-exponentially
decaying block error probability.
1:N ,Y 1:N , QU 1:N ,Y 1:N
=
2V PU1,a,b
1,a,b
X
1:N
1:N
1:N
P u1:N
,
y
−
Q
u
,
y
1,a,b
1,a,b
1:N
u1:N
1,a,b ,y
=
X
X
1:N
1:N
P u1:N
, y 1:N − Q u1:N
, y 1:N
1,a,b , x
1,a,b , x
1:N x1:N
u1:N
1,a,b ,y
≤
X
X
1:N
1:N
P u1:N
, y 1:N − Q u1:N
, y 1:N
1,a,b , x
1,a,b , x
1:N x1:N
u1:N
1,a,b ,y
−N β
=O 2
0
.
Let EPa [P e] and EPb [P e] denote the expectation error probability as a result of the distribution
b
1:N ,Y 1:N at level 1 and 2, respectively. Take E
PU1,a,b
P [P e] as an example to show the decaying error
1:N
1:N
probability. Let Ei denote the set of random variables u1:N
such that the SC decoding
b , u1,a , y
error occurred at the ith bit. Hence the block error event is defined by E b , ∪i∈ICb Ei , and the
24
expectation of decoding block error probability over all random mapping is given by
EPb [P e] =
X
1:N 1:N 1:N
1:N
1:N ,Y 1:N
PU1,a,b
u1:N
1 ub , u1,a , y
∈ Eb
1,a,b , y
1:N
u1:N
1,a,b ,y
≤
X
X
1:N 1:N 1:N
1:N
1:N ,Y 1:N
PU1,a,b
u1:N
1 ub , u1,a , y
∈ Ei
1,a,b , y
i∈ICb ∪SSb u1:N ,y 1:N
1,a,b
≤
X
X
1:N
1:N
P ub1:i−1 , u1:N
P uib |u1:i−1
, u1:N
1,a , y
1,a , y
b
1:N
i∈ICb ∪SSb u1:i ,u1:N
1,a ,y
b
≤
1:N
1:N
, u1:N
≤ P uib ⊕ 1|u1:i−1
· 1 P uib |ub1:i−1 , u1:N
1,a , y
1,a , y
b
X
1:N
1:N
P ub1:i−1 , u1:N
P uib |u1:i−1
, u1:N
1,a , y
1,a , y
b
X
1:N
i∈ICb ∪SSb u1:i ,u1:N
1,a ,y
b
s
1:N
P uib ⊕ 1|u1:i−1
, u1:N
1,a , y
b
·
1:N
1:N
P uib |u1:i−1
,
u
,
y
1,a
b
1:N
≤ N · Z Ubi |Ub1:i−1 , U1,a
, Y 1:N
0
−N β
=O 2
.
β0
Following the same arguments, we also have EPa [P e] = O 2−N . Therefore, by this union
β0
bound, we obtain the two-stage decoding block error probability EP [P e] = O 2−N .
1:N ,Y 1:N , which is an
Let P eHB denote the expectation of error probability caused by QU1,a,b
i∈FCa
average over all choices of Ua
i∈SCa \SSa
, Ua
1:N
the set of random variables u1:N
b , u1,a , y
P eHB − EP [P e] =
X
1:N
i∈FCb
, Ub
i∈SCb \SSb
and Ub
at each level. Let E denote
such that a decoding error occurs. Then we have
1:N 1:N 1:N
1:N
1:N
Q u1:N
− P u1:N
· 1 ub , u1,a , y
∈E
1,a,b , y
1,a,b , y
1:N
u1:N
1,a,b ,y
1:N ,Y 1:N , QU 1:N ,Y 1:N
≤ 2V PU1,a,b
1,a,b
0
β
≤ O 2−N
.
As for the rates, we have
|ISa | N →∞
−−−→
N
I (Ua ; X, U1 ) and
|ICa | N →∞
−−−→
N
level. Therefore, we have
|ISa | − |ICa | N →∞
−−−→ I (X; Ua |U1 , Y ) .
N
I (Ua ; Y, U1 ) at the first
25
For the second level,
|ISb |
N
N →∞
−−−→ I (Ub ; X, U1 , Ua ) and
|ICb |
N
N →∞
−−−→ I (Ub ; Y, U1 , Ua ). Thus,
we have
|ISb | − |ICb | N →∞
−−−→ I (X; Ub |U1 , Ua , Y ) .
N
Finally, the rate of Decoder 2 is
|ISa | − |ICa | + |ISb | − |ICb | N →∞
−−−→ I (X; U2 |U1 , Y ) .
N
R EFERENCES
[1] A. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the decoder,” IEEE Trans.
Inf. Theory, vol. 22, no. 1, pp. 1–10, Jan 1976.
[2] C. Heegard and T. Berger, “Rate distortion when side information may be absent,” IEEE Trans. Inf. Theory, vol. 31, no. 6,
pp. 727–734, 1985.
[3] K. Kerpez, “The rate-distortion function of a binary symmetric source when side information may be absent,” IEEE Trans.
Inf. Theory, vol. 33, no. 3, pp. 448–452, May 1987.
[4] C. Tian and S. N. Diggavi, “A calculation of the Heegard-Berger rate-distortion function for a binary source,” in Proc.
2006 IEEE Inform. Theory Workshop, Oct 2006, pp. 342–346.
[5] S. B. Korada, “Polar codes for channel and source coding,” Ph.D. dissertation, Ecole Polytechnique Fédérale de Lausanne,
Lausanne, Switzerland, 2009.
[6] L. Liu and C. Ling, “Polar lattices for lossy compression,” Jan. 2015. [Online]. Available: http://arxiv.org/abs/1501.05683
[7] J. Honda and H. Yamamoto, “Polar coding without alphabet extension for asymmetric models,” IEEE Trans. Inf. Theory,
vol. 59, no. 12, pp. 7829–7838, Dec. 2013.
[8] M. Ye and A. Barg, “Polar codes for distributed hierarchical source coding,” Advances in Mathematics of Communications,
vol. 9, no. 1, pp. 87–103, 2015.
[9] S. Ramanan and J. M. Walsh, “Practical codes for lossy compression when side information may be absent,” Proc. IEEE
Int. Conf. on Acoustics, Speech and Signal Processing, pp. 3048–3051, May 2011.
[10] S. Korada and R. Urbanke, “Polar codes are optimal for lossy source coding,” IEEE Trans. Inf. Theory, vol. 56, no. 4, pp.
1751–1768, 2010.
[11] Y. Yan, L. Liu, C. Ling, and X. Wu, “Construction of capacity-achieving lattice codes: Polar lattices,” Nov. 2014.
[Online]. Available: http://arxiv.org/abs/1411.0187
[12] G. D. Forney Jr., M. Trott, and S.-Y. Chung, “Sphere-bound-achieving coset codes and multilevel coset codes,” IEEE
Trans. Inf. Theory, vol. 46, no. 3, pp. 820–850, May 2000.
[13] C. Ling and J.-C. Belfiore, “Achieving AWGN channel capacity with lattice Gaussian coding,” IEEE Trans. Inf. Theory,
vol. 60, no. 10, pp. 5918–5929, Oct. 2014.
[14] A. H. Kaspi, “Rate-distortion function when side-information may be present at the decoder,” IEEE Trans. Inf. Theory,
vol. 40, no. 6, pp. 2031–2034, 1994.
[15] E. Perron, S. N. Diggavi, and I. E. Telatar, “The Kaspi rate-distortion problem with encoder side-information: Gaussian
case,” Nov 2005. [Online]. Available: https://infoscience.epfl.ch/record/59938
[16] ——, “The Kaspi rate-distortion problem with encoder side-information: Binary erasure case,” 2006. [Online]. Available:
https://infoscience.epfl.ch/record/96000
| 7cs.IT
|
A Parallel Memetic Algorithm to Solve the
Vehicle Routing Problem with Time Windows
Jakub Nalepa1 and Zbigniew J. Czech1,2
arXiv:1402.6942v1 [cs.DC] 27 Feb 2014
1
Silesian University of Technology, Gliwice, Poland
{jakub.nalepa, zbigniew.czech}@polsl.pl
2
University of Silesia, Sosnowiec, Poland
Abstract. This paper presents a parallel memetic algorithm for solving
the vehicle routing problem with time windows (VRPTW). The VRPTW
is a well-known NP-hard discrete optimization problem with two objectives. The main objective is to minimize the number of vehicles serving
customers scattered on the map, and the second one is to minimize the
total distance traveled by the vehicles. Here, the fleet size is minimized in
the first phase of the proposed method using the parallel heuristic algorithm (PHA), and the traveled distance is minimized in the second phase
by the parallel memetic algorithm (PMA). In both parallel algorithms,
the parallel components co-operate periodically in order to exchange the
best solutions found so far. An extensive experimental study performed
on the Gehring and Homberger’s benchmark proves the high convergence
capabilities and robustness of both PHA and PMA. Also, we present the
speedup analysis of the PMA.
Keywords: parallel memetic algorithm, hybrid genetic algorithm, local search,
vehicle routing problem with time windows.
1
Introduction and Related Work
The vehicle routing problem with time windows (VRPTW) consists in finding
a schedule for a fleet of homogeneous vehicles serving a set of geographically
scattered customers. The capacities of the vehicles cannot be exceeded and the
customers must be visited within their well-defined time windows. We consider
the VRPTW as a hierarchical optimization problem. The primary objective is to
minimize the total fleet size and the second one is to minimize the total distance
traveled by the vehicles. This approach makes it possible to develop and optimize
the algorithms for both phases independently.
The practical applications of the VRPTW include the bus route planning,
post and parcels delivering, food delivering, cash delivering to banks and ATM
terminals and many more. Thus, a number of algorithms including exact and
heuristic methods were introduced for the VRPTW over the years. Due to the
large complexity of the VRPTW and its wide practical applicability the heuristic
and metaheuristic methods capable of producing high-quality feasible solutions
in reasonable time are of main importance here.
Exact algorithms incorporating, among others, dynamic programming, branchand-bound algorithms and greedy approaches, were proposed by Bard et al. [1],
Irnich and Villeneuve [2], Jepsen et al. [3], Kellehauge et al. [4] and Chabrier [5].
Noteworthy, only 9 instances out of 300 belonging to the Gehring and Homberger’s
benchmark [6] have been solved to optimality [7]. An extensive review of the exact methods can be found in Kallehauge [8].
Heuristic algorithms can be divided into two classes, namely the improvement
and the construction techniques. In the construction heuristic algorithms the
customers are iteratively inserted into a partial feasible solution without violating
the time windows and capacity constraints. Several construction heuristics were
proposed by Solomon [9], Potvin and Rousseau [10] and recently by Pang [11].
On the other hand, the improvement heuristics modify an initial solution and
explore the search space by performing local search moves in order to decrease
the fleet size and the total traveled distance. The examples of such heuristics can
be found in Thompson et al. [12], Russell [13] and Potvin and Rousseau [14].
Metaheuristic algorithms incorporate mechanisms to explore the search space
and to exploit its most promising regions. They allow infeasible intermediate
solutions and the solutions deteriorating during the search process in order to
escape the local minima. A number of sequential and parallel algorithms were
introduced during the recent years. The simulated annealing was successfully
applied by Zhong and Pan [15], Debudaj-Grabysz and Czech [16], and recently
by Li [17]. The tabu searches were proposed by Cordeau et al. [18] and Sin et
al. [19]. The ant colony approaches can be found in Xuan et al. [20] and Qi and
Sun [21]. A survey on the metaheuristic algorithms can be found in Bräysy and
Gendreau [22].
Evolutionary algorithms have attracted the scientific attention to solve the
VRPTW due to their high search capabilities. Genetic algorithms were applied
by Cheng and Wang [23], Kamkar et al. [24] and Ursani et al. [25]. The evolution
strategies were proposed by Gehring and Homberger [26], Mester et al. [27]
and Kanoh et al. [28]. The sequential and parallel memetic algorithms (MAs),
combining the evolutionary algorithms for more distant search with the local
optimization and refinements algorithms for local exploitation, were described
by Berger and Barkaoui [29], Labadi et al. [30] and Nagata and Bräysy [7]. The
MA based on the edge-assembly crossover operator (EAX) has been proposed
by Nagata and Bräysy [7].
This paper is organized as follows. Section 2 formulates the VRPTW. The
sequential and the parallel heuristic algorithms for minimizing the number of
routes are discussed in Section 3. The memetic algorithm for the total distance
minimization is described in Section 4. Section 5 describes the experimental
results. Section 6 concludes the paper.
2
Problem Formulation
Let G = (V, E) be a directed graph with a set V of N +1 vertices representing
the customers and the depot, together with a set of edges E = {(vi , vi+1 )|vi , vi+1 ∈
V, vi 6= vi+1 } representing the connections between the travel points. Each vehicle starts and finishes at the depot v0 . Travel costs are given as ci,j , where i 6= j,
i, j ∈ {0, 1, ..., N }. Non-negative customer demands di , i ∈ {0, 1, ..., N }, where
d0 = 0 and the time windows [ei , li ], i ∈ {0, 1, ..., N } are defined. The customers
have their service times si , i ∈ {1, 2, ..., N }. A fleet of K vehicles with a constant
capacity Q is given. The route is defined as a set of customers served by a single
vehicle (v0 , v1 , ..., vn+1 ), where v0 = vn+1 is the depot.
A solution σ is feasible if (i) the total amount of goods delivered to the
customers within each route does not exceed the vehicle capacity Q, (ii) the
service of a customer vi is started before the time window [ei , li ] elapses, (iii)
every customer vi is served in exactly one route, and (iv) every vehicle leaves
and returns to the depot within its time window [e0 , l0 ].
The primary objective of the VRPTW is to minimize the fleet size K (K ≤
PN
Kmin , and Kmin = dD/Qe, where D = i=1 di ). The secondary objective is to
PK
minimize the total distance T = i=1 Ti , where K is the number of routes and
Ti is the distance of the i-th route.
3
Minimizing the Number of Routes by the Parallel
Heuristic Algorithm
In this section we discuss in detail the heuristic algorithm used for minimizing
the number of routes in the VRPTW. First, its sequential version is outlined
along with the proposed enhancements. Then, the parallel heuristic algorithm
(PHA) is presented.
3.1
Sequential Algorithm Outline
The route minimization heuristic algorithm reduces the number of routes in
a feasible solution σ by one at a time. The process of removing routes continues
until the execution time reaches a defined maximum or the number of vehicles
reaches its lower bound Kmin [31]. The initial solution of the VRPTW consists
of N routes, which means that each customer is served by a separate vehicle. A
randomly selected route (Fig. 1, line 1) is excluded from the current solution. The
customers are inserted into the ejection pool (EP) (line 2). The penalty counters
p[i], are set to 1 (line 3). The counters indicate the reinsertion difficulties of the
corresponding customers. The customers are taken from the EP applying the
LIFO strategy.
All feasible insertion positions are determined (line 6). Introducing forward
and backward time window penalty slacks allowed for constant-time verification
of the time window penalties [31] (clearly, the change of the vehicle loads may
be computed in constant time). If the set of feasible insertion positions is not
empty, then a random insertion is performed. The solution σ 0 with the customer
vins becomes a new feasible (possibly partial) solution σ.
f
If the set Nins
(vins , σ) is empty, then it is impossible to insert the customer
vins without violating the constraints. The Squeeze procedure allows for creating
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
Select and remove a random route s from σ;
Put a random permutation of ejected customers of the removed route into EP;
Initialize penalty counters p[i] ← 1, i = 1, 2, . . . , N ;
while (EP6= ∅ and not finished ) do
Select and remove customer vins from EP using LIFO strategy;
f
if Nins
(vins , σ) 6= ∅ then
f
σ ← σ 0 chosen randomly from Nins
(vins , σ);
else
σ ← Squeeze(vins , σ);
end if
if vins ∈
/ σ then
p[vins ] ← p[vins ] + 1;
P
(i)
σ ← σ 0 ∈ N ej (vins , σ) with min( ki=1 p[vout ]);
(1)
(2)
(k)
Insert ejected customers vout , vout , . . . , vout into EP;
σ ← Perturb(σ);
end if
finished ← VerifyStopCondition();
end while
if EP6= ∅ then
Restore σ to the initial solution;
end if
return σ;
Fig. 1. A sequential heuristic algorithm to minimize the number of routes in the
VRPTW (RemoveRoute).
temporarily infeasible solutions (line 9). The infeasible solution with the smallest
value of the penalty function Fp (σ) is chosen [7], and the attempts of restoring
the feasibility are undertaken. If the squeezing fails, then the penalty counter of
the customer is increased (line 12).
The last approach of reinserting vins allows for ejecting other customers from
the solution. A sum of the penalty counters Psum is minimized to eject the customers that will be relatively easy to reinsert later. The customers that were
inserted during the last lmax iterations are not considered for ejections [32]. Noteworthy, the increasing number of the ejected customers k, k ∈ {1, 2, . . . , kmax },
is considered. If at least one feasible ejection is found for a given k, then the other
tests are skipped. If there are more ejections with the same Psum , then one is
chosen randomly [32]. A number of constant-time local moves are performed in
Perturb procedure in order to diversify the search (line 15). The algorithm finishes if the EP is empty, the execution time exceeds the specified limit or the
size of the EP is unacceptably large (line 17). The other breaking conditions are
described in the next section.
3.2
Suggested Modifications
The additional stopping condition of the algorithm introduced in [32] addresses the maximal number of iterations imax . According to that, it should
break even though the size of the EP is small and i > imax , where i is the number of current iteration. It is worth noting that a large number of iterations are
usually performed if the number of routes is close to the optimum. Thus, it may
be proficient to allow for additional loop executions if the EP is small and the
probability of reinserting the customers is still high. The additional parameter
ξ indicates the maximal number of customers allowed to reside in the EP for
which the loop will continue despite of exceeding imax .
However, the size of the EP can stay constant during the execution for a
long time. Here, the maximal number of iterations in the steady state, ψ, is
introduced. It should vary with the maximal number of allowed iterations, thus
ψ is a fraction of imax . If the EP size does not change during the ψ iterations,
then the probability of a feasible customer reinsertion drops and the loop breaks.
In the Squeeze procedure the feasibility restoring attempts are undertaken.
The local search moves that may be calculated in constant time were used during
the construction of Nr (σ). If there are no feasible moves, then the linear-time
moves are considered. This approach is appropriate for smaller instances, since
calculating the moves may become time-consuming for the larger number of
customers. Additionally, it may be worth limiting the number of moves to test,
e.g., until n edge-exchanges are found.
A route to be removed is chosen randomly. The routes may be divided into
two classes – the first containing the routes with the number of customers greater
or equal to the average, and the second class with the other routes. Intuitively, it
should be easier to reinsert customers from the smaller route to a partial solution,
thus choosing a random route from the first class gives a higher probability of
feasible reinsertions. However, it is proficient to get rid of larger routes earlier
and choose a random route from the second class, when the solution size, i.e.,
the number of routes, is still far from the optimum.
If the squeezing fails, then the other ejections are tested. After a successful
reinsertion of vins and ejections of other customers the solution is perturbed. In
many cases perturbing is not necessary for efficient reinsertions. The perturbation may be omitted if the percentage of successful customer insertions without
additional ejections is significant, e.g., 80%, for a given number of subsequent
iterations. However, the algorithm may fail due to exceeding the maximal number of trials. In this case, the perturbations of the partial solutions should be
allowed. The number of moves in the Perturb procedure may depend on the
difficulty of reinserting the customers, i.e., it should increase with the decrease
of the number of routes. The initial number of moves is multiplied by a constant
factor, e.g., 2, every defined number of iterations until it reaches the maximal
value.
3.3
Parallel Algorithm Outline
The p available processors may be used either to achieve a higher accuracy
of a solution in a given time, or to speed up the computations. In the first case,
the goal is to obtain a solution that is closer to the global optimum. The main
goal of the parallel heuristic algorithm (PHA) is to improve the accuracy of final
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
parfor Pi ← P0 to Pp−1 do
σi ← σinit ;
end parfor
while not finished do
parfor Pi ← P0 to Pp−1 do
Call RemoveRoute δ times;
end parfor
Co-operate according to the co-operation scheme;
finished ← VerifyStopCondition();
end while
. See Fig. 1
Fig. 2. A parallel heuristic algorithm (PHA) to minimize the number of routes in the
VRPTW.
solutions. The algorithm consists of p components denoted as P0 , P1 , . . . , Pp−1 1 .
The initial solution is considered as the starting solution for each parallel component (Fig. 2, line 2). Then, each component calls RemoveRoute in parallel
with others δ times. The components co-operate to exchange the best solutions
found up-to-date. The solutions are assessed according to their costs. The solutions with the smaller number of routes K are preferred. If the number of routes
is equal, then the solution with the shorter total travel distance is considered
better. The parallel algorithm guides the search towards the optimal solutions
with respect to the number of vehicles and the travel distance.
3.4
Co-operation of Parallel Components
The co-operation of threads starts from thread P0 . Thread P1 receives the
solution σ0 from P0 , and compares the costs of solution σ1 with the received
one. The solution with the smaller cost replaces the current solution of P1 .
Consequently, thread Pp−1 compares the solution σp−1 with σp−2 received from
Pp−2 . Finally, the best solution is held by thread Pp−1 . If the best solution
is found by P0 , then all the components get σ0 . The co-operation scheme is
presented in Fig. 3.
The second variant of the co-operation scheme is cyclic and includes the
communication between threads Pp−1 and P0 . The better solution will be sent
to thread P0 . The solution σ0 should be updated only if the number of routes is
larger than the number of routes in Pp−1 . Keeping the original solution σ0 with
the same number of vehicles by thread P0 may prevent from having the same
solution held by all threads. Introducing the probability of replacing a worse
solution by the solution received from the neighbor may further decrease the
likelihood of having only one solution in the team. The probability of choosing
the better solution should be large, but not equal to 1. The number of steps that
are executed in parallel before the co-operation must be determined sensitively.
If the co-operation is too frequent, e.g. for large problem instances, the total
parallel overhead becomes more significant and the execution time increases
1
In the OpenMP implementation we consider each thread as a parallel component.
σinit →
σ0t0
σ1t0
/ σ0t1
/ σ0t2
/ ...
/ σ0tn−1
/ σ0tn
/ σ1t1
/ σ1t2
/ ...
/ σ1tn−1
/ σ1tn
...
t0
σp−2
σ t0
p−1
...
...
...
...
...
t1
/ σp−2
t2
/ σp−2
/ ...
tn−1
/ σp−2
tn
/ σp−2
t1
/ σp−1
t2
/ σp−1
/ ...
tn−1
/ σp−1
tn
/ σp−1
Fig. 3. Co-operation scheme
rapidly. However, if the components co-operate rarely for small instances, the
gain from parallelism is hardly noticeable.
The following co-operation frequencies are introduced:
1. Constant – δ is constant.
2. Rare – δ is defined as a ratio of the problem size and a value of rare cooperation factor. Then, after a number of co-operation phases δ is divided
by this factor until it reaches a defined lower limit.
3. Frequent – the scheme is similar to the rare co-operation, but the frequent
co-operation factor is larger than the rare co-operation factor.
4. Adaptive – δ is defined as a ratio of the problem size and a value of adaptive
co-operation factor. The δ is divided by the ratio of the last average time
and the previous average time (if tprev
avg 6= 0, otherwise δ is divided by the
adaptive factor) of RemoveRoute executions.
An extensive study on the co-operation schemes can be found in [33].
4
Minimizing the Total Travel Distance by the Parallel
Memetic Algorithm
Here, we discuss both sequential and parallel versions of the memetic algorithm for minimizing the total distance traveled by the vehicles. The main
goal of the presented parallel algorithm is to reduce the execution time without
decreasing the quality of final solutions.
4.1
Sequential Algorithm Outline
The initial population of size N containing the feasible solutions with K
routes is found by the parallel heuristic algorithm discussed in Section 3. If the
maximal time of generating the initial population is exceeded, then the solutions
already found are copied and perturbed until the population size reaches N .
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Generate an initial population of N feasible solutions;
while not f inished do
Determine N reproduction pairs (pA , pB );
for each pair (pA , pB ) do
for i ← 1 to Nch do
pc ← EAX(pA , pB );
pc ← Repair(pc );
p0c ← LocalSearch(pc );
if η(p0c ) > η(pbc ) then
pbc ← p0c ;
end if
end for
end for
Form the next population;
f inished ← VerifyTerminationCondition();
end while
return best individual in the last population;
. Pre-selection
. Post-selection
Fig. 4. The memetic algorithm for the total distance minimization.
Each individual is chosen once as the parent pA and pB in a random order
to generate the child solutions using the EAX operator (Fig. 4, line 3) using the
AB-selection scheme. It has been shown that the proper pre- and post-selection
schemes have strong impact on the convergence capabilities of a memetic algorithm [34].
Nch defines the number of children generated for each pair pA and pB . The
feasibility of a child is restored by the Repair function if necessary (line 7) using
the concept of local moves utilized while squeezing an infeasible solution in the
route minimization heuristics. If the solution is feasible, then a number of moves
are performed to improve its quality, i.e., to decrease the total travel distance
(line 8). The moves are limited to the customers belonging to the routes modified
by the EAX operator and the repairing procedure [7]. The total distance of a new
solution is compared with the total distance of the best child found up-to-date
(line 9). If a new solution is of higher quality, then the best child is updated.
After generation of child solutions for N pairs of parents the population is
updated, i.e., the best individuals form a new population according to the postselection scheme. It is easy to see that the best child obtained for pA and pB
replaces the first parent, not the worst individual in the population. Removal of
pA is motivated by the fact that the better individual replaces the worse with
the similar characteristics to ensure the population diversity. The additional
termination condition addresses the steady state, i.e., the situation in which for
a large number of subsequent generations the quality of the best individual is
not improved. The algorithm finishes if the number of generations in the steady
state is larger than the defined maximum, maximal number of generations G is
reached, or the maximal execution time is exceeded. The best individual from
the population is finally returned (line 17).
4.2
Parallel Algorithm Outline
The parallel memetic algorithm (PMA) consists of p parallel components
denoted as P0 , P1 , . . . , Pp−1 . The main part of the memetic algorithm, i.e., generating the child solutions, is the most computationally intensive. The iterations
of the loop may be executed in parallel, since Nch children are generated for the
best
parents pA and pB independently. The best child solution is stored as σr(i)
. Each
individual in the population serves once as pA and pB during the combination
best
stage, therefore different σr(i)
solutions are updated in every iteration. The N
iterations are distributed between p threads, where N p. The number of individuals in the population is usually large to avoid the similarities between the
individuals. Once the loop finishes, the best child solutions are found and the
current generation is updated. The cost, i.e., the total travel distance, of each
individual in the current solution is compared with the cost of the best child.
If the cost of the child σibest is smaller, then the child becomes a new individual in the population and replaces the solution σi . Clearly, the N solutions are
compared independently, therefore the iterations may be executed in parallel.
Processing of the next generation of solutions starts with initializing of the set
of the best child solutions. Similarly, the loop iterations are independent and
may be executed in parallel.
5
Experimental Results
The algorithms were implemented in C++ using the OpenMP interface and
were tested on Gehring and Homberger’s problem instances. The code was compiled using Intel C++ Compiler 10.1.015 with -fast and -openmp flags. Calculations were carried out at a single node of Galera supercomputer at the
Academic Computer Center in Gdańsk [35]. The computations were performed
on the nodes with 16 GB RAM (2 GB/core) equipped with Intel Xeon Quad
Core (2.33 GHz) processors with 12 MB of level 3 cache. The parameters used
during the experiments are given in Tab. 1 and Tab. 2. The percentage of the
nearest customers µ is limited for neighborhood calculations to decrease the execution time [31]. The number of additional customers allowed to reside in the EP
has been proposed in [32]. The minimal number of local moves used during the
solution perturbation should allow transforming a current solution to the neighboring, but still not too similar one. If the additional ejections are necessary for
a successful customer insertion, then the number of moves is multiplied by IPf to
increase the probability of getting the new configurations. The maximal number
of moves (for both stages) prevents from a rapid increase of the execution time.
The maximal number of iterations in the steady state corresponds to a decent
fraction of the maximal number of allowed algorithm iterations.
The settings of the co-operation are given in Tab. 3. The EAX strategy [7] is
chosen randomly. If a significant number of consecutive generations, e.g., 50, does
not result in improving the best individual in a population, then the probability
of further improvements drops rapidly. A formula for the maximal execution
time calculation of the memetic algorithm has been proposed in [7].
Table 1. Parameters of the route minimization heuristic algorithm.
Parameter
µ
kmax
lmax
ξ
imax
ψ
IPm
IPM
IPf
IF
τR
τ
Description
Percentage of the nearest customers in the
neighborhood
Maximal number of customers to be ejected
Number of iterations without ejecting a customer
after it is inserted
Additional customers allowed to reside in the EP
Maximal number of iterations in the first phase
algorithm
Maximal number of iterations in the steady state
Minimal number of feasible moves
while perturbing
Maximal number of feasible moves
while perturbing
Update factor for the number of moves
while perturbing
Frequency of updating the number of moves
(in iterations)
Maximal time for reinsertions in RemoveRoute
(in seconds)
Maximal execution time (in seconds)
Value
0.6
3
5
7
1000
imax /5
80
400
2
50
50
1200
Table 2. Parameters of the distance minimization memetic algorithm.
Parameter
Nch
IC
IP
τ
G
Description
Number of child solutions generated for each pair of
parents
Maximal number of moves improving the child solution
Number of moves used during copying and perturbing
Maximal execution time (in minutes)
Maximal number of generations without improvement
Value
20
100
50
Ns N/400
50
Table 3. Co-operation frequency settings; CM – co-operation mode, CF – co-operation
factor, UF – update factor, Ufr – update frequency, Mfr – minimal frequency.
Cust. no.
200
400
600
800
1000
CM
Frequent
Frequent
Adaptive
Rare
Rare
CF
10
10
10
5
5
UF
2
2
–
2
2
Ufr
4
4
1
3
3
Mfr
1
1
1
1
1
A number of possible modifications and improvements have been suggested
in Section 3.2. The exemplary average execution times of the sequential route
minimization heuristic algorithm are given in Fig. 5. If the algorithm gets stuck
in the local minima of the search space (e.g., for RC2 4 1), then the decreased
a)
b)
SHA
SHA-I
5
Average time t (s)
Average time t (s)
50
40
30
20
10
0
C1_8_2 RC2_4_1 R2_2_1
Problem instance
SHA
SHA-I
4
3
2
1
0
R1_10_2 RC1_4_1 C2_4_2
Problem instance
Fig. 5. Average execution time t (in seconds) of a sequential heuristic algorithm (SHA)
and the improved sequential heuristic algorithm (SHA–I) for minimizing the number
of routes for more (a) and less (b) time-consuming tests for 100 experiments.
initial number of local search moves results in the increase of the total number of iterations necessary to leave the local minimum. However, it is not always necessary to explore the vast solution space for large instances (e.g., for
R1 10 2, C1 8 2) and a relatively small number of moves during the perturbation is enough to get satisfactory results. The average execution time has been
decreased for a number of instances that were relatively easy to solve (Fig. 5b)
and for time-consuming ones (a). However, the modifications are less suitable for
the problems with solution spaces containing a large number of local minima.
The cumulative numbers of vehicles (CVNs), i.e. the number of vehicles serving all instances, are presented in Tab. 4 for sequential and parallel algorithms
(SHA–I and PHA, respectively). The number of vehicles was decreased for 16
instances, whereas the world’s best results were obtained in 6 cases using the
parallel heuristics. Therefore, in 271 out of 300 (90%) cases the benchmarking
tests were solved to the current optimum with respect to the number of vehicles using the parallel algorithm. The parallel memetic algorithm significantly
improved the current world’s best result for the problem instance C1 8 22 .
The size of the population N influences the execution time necessary to create a new generation of solutions. However, the probability of ending up with a
set of similar individuals is lower in case of large populations. The problem of
saturating the population is illustrated in Fig. 6. The experiments with the clustered customers have shown that the saturation of the population with similar
individuals may occur relatively fast. The larger population should imply a larger
population diversity. If the number of individuals with similar configurations exceeds a certain threshold, then the population is in the diversity crisis. Similarly,
2
The best-known solutions of the GH tests are published at: http://www.sintef.
no/Projectweb/TOP/VRPTW/Homberger-benchmark/; reference date: September 13,
2011.
Table 4. The percentage of the world’s best CVNs obtained using SHA–I and PHA.
Class
C1
C2
R1
R2
RC1
RC2
Total
SHA–I
82%
70%
94%
100%
100%
84%
88%
a)
b)
7900
R1_2_2
N = 10
N = 25
N = 50
World's best
6900
6400
C2_4_1
N = 10
N = 25
N = 50
World's best
5200
5000
Total distance
7400
Total distance
PHA
84%
78%
94%
100%
100%
86%
90%
4800
5900
4600
5400
4400
4900
4200
4400
4000
3900
0
10
20
30
40
50
Generation
0
10
20
30
40
50
Generation
Fig. 6. Total distance traveled for various population sizes N for tests (a) R1 2 2 and
(b) C2 4 1.
the number of children Nch generated for each pair of parents influences the
quality of final solutions along with the execution time of the algorithm. Clearly,
generating a larger number of children increases the probability of getting a
well-fitted individual. On the other hand, the execution time of the algorithm
increases with larger values of Nch .
The relative speedup obtained for two given problem instances is presented in
Fig. 7. The population size N is usually larger than the number of threads p. The
speedup depends not only on the problem size but also on its internal structure.
If the number of generations required to obtain a minimal travel distance is large,
then the relative speedup is almost ideal. However, if the solution converges to
the minimum relatively fast, then the further improvements become difficult and
the parallel overhead becomes more significant once the steady state is reached.
6
Conclusions and Future Work
The parallel heuristic algorithm for minimizing the fleet size has proven to
be effective and competitive by solving 90% of problem instances to the current
known optimum. The memetic algorithm for the distance minimization turned
b)
8
8
6
6
Speedup
Speedup
a)
4
2
4
2
0
0
1
2
4
Number of threads
8
1
2
4
8
Number of threads
Fig. 7. Speedup S vs. number of threads p for tests (a) RC1 6 3 and (b) C1 2 1
(continuous line shows the ideal speedup).
out to be powerful. A large number of parameters, both for the exploration and
the exploitation of the search space, allow for adjusting the algorithm to the
instance characteristics. The optimal assignment of parameters is to be cleared
up during the further research. The experiments performed for various problem
instances showed that the relative speedup is linear and close to the ideal one
in many cases. The parallel algorithm significantly improved the world’s best
solution of the clustered Gehring and Homberger’s test C1 8 2 containing 800
customers.
A two-stage approach of solving the VRPTW makes it possible to combine
the presented algorithms with other well-known heuristics, e.g., simulated annealing or tabu search. We aim at determining the most effective and scalable
combination of heuristics addressing both objectives of the VRPTW. Also, we
plan to develop more efficient co-operation schemes for both stages of the parallel
algorithm.
Acknowledgments. We thank the following computing centers where the
computations of our project were carried out: Academic Computer Centre in
Gdańsk TASK, Academic Computer Centre CYFRONET AGH, Kraków, Interdisciplinary Centre for Mathematical and Computational Modeling, Warsaw
University, Wroclaw Centre for Networking and Supercomputing.
References
1. Bard, J.F., Kontoravdis, G., Yu, G.: A branch-and-cut procedure for the vehicle
routing problem with time windows. Trans. Sc. 36(2) (2002) 250–269
2. Irnich, S., Villeneuve, D.: The shortest-path problem with resource constraints and
k-cycle elimination. INFORMS J. on Computing 18(3) (2006) 391–406
3. Jepsen, M., Petersen, B., Spoorendonk, S., Pisinger, D.: A non-robust branchand-cut-and-price algorithm for the vehicle routing problem with time windows.
Technical Report 06-03, University of Copenhagen, Denmark (2006)
4. Kallehauge, B., Larsen, J., Madsen, O.B.G.: Lagrangian duality applied to the
vehicle routing problem with time windows. Comput. Oper. Res. 33(5) (2006)
1464–1487
5. Chabrier, A.: Vehicle routing problem with elementary shortest path based column
generation. Comput. Oper. Res. 33(10) (2006) 2972–2990
6. SINTEF: Problems and benchmarks, VRPTW. Website (2011) http://www.
sintef.no/Projectweb/TOP/VRPTW/Homberger-benchmark/.
7. Nagata, Y., Bräysy, O., Dullaert, W.: A penalty-based edge assembly memetic
algorithm for the vehicle routing problem with time windows. Comput. Oper. Res.
37(4) (2010) 724–737
8. Kallehauge, B.: Formulations and exact algorithms for the vehicle routing problem
with time windows. Comput. Oper. Res. 35(7) (2008) 2307–2330
9. Solomon, M.M.: Algorithms for the vehicle routing and scheduling problems with
time window constraints. Oper. Res. 35 (1987) 254–265
10. Potvin, J.Y., Rousseau, J.M.: A parallel route building algorithm for the vehicle
routing and scheduling problem with time windows. European J. of Oper. Res.
66(3) (1993) 331–340
11. Pang, K.W.: An adaptive parallel route construction heuristic for the vehicle
routing problem with time windows constraints. Exp. Syst. Appl. 38(9) (2011)
11939–11946
12. Thompson, P.M., Psaraftis, H.N.: Cyclic transfer algorithms for multivehicle routing and scheduling problems. Oper. Res. 41(5) (1993) 935–946
13. Russell, R.: Hybrid heuristics for the vehicle routing problem with time windows.
Trans. Sc. 29(2) (1995) 156
14. Potvin, J.Y., Rousseau, J.M.: An exchange heuristic for routing problems with
time windows. J. of the Oper. Res. Society 46 (1995) 1433–1446
15. Zhong, Y., Pan, X.: A hybrid optimization solution to VRPTW based on simulated
annealing. In: Automation and Logistics, 2007 IEEE International Conference on.
(2007) 3113 –3117
16. Debudaj-Grabysz, A., Czech, Z.J.: A concurrent implementation of simulated annealing and its application to the VRPTW optimization problem. In: Distributed
and Parallel Systems. Volume 777 of The Kluwer International Series in Engineering and Computer Science. Springer US (2005) 201–209
17. Li, Y.: An improved simulated annealing algorithm and its application in the
logistics distribution center location problem. Applied Mechanics and Materials
389 (2013) 990–994
18. Cordeau, J., Laporte, G., Hautes, E., Commerciales, E., Gerad, L.C.D.: A unified
tabu search heuristic for vehicle routing problems with time windows. J. of the
Oper. Res. Society 52 (2001) 928–936
19. Ho, S.C., Haugland, D.: A tabu search heuristic for the vehicle routing problem
with time windows and split deliveries. Comp. and Oper. Res. 31 (2002) 1947–1964
20. Tan, X., Zhuo, X., Zhang, J.: Ant colony system for optimizing vehicle routing
problem with time windows (VRPTW). In: Proc. of the 2006 int. conf. on Comp.
Intell. and Bioinf. ICIC’06, Berlin, Heidelberg, Springer-Verlag (2006) 33–38
21. Qi, C., Sun, Y.: An improved ant colony algorithm for VRPTW. In: Proc. of the
2008 Int. Conf. on Comp. Sc. and Soft. Eng. (2008) 455–458
22. Bräysy, O., Gendreau, M.: Vehicle Routing Problem with Time Windows, part II:
Metaheuristics. Trans. Sc. 39(1) (2005) 119–139
23. Cheng, C.B., Wang, K.P.: Solving a vehicle routing problem with time windows
by a decomposition technique and a genetic algorithm. Expert Syst. Appl. 36(4)
(2009) 7758–7763
24. Kamkar, I., Poostchi, M., Akbarzadeh Totonchi, M.R.: A cellular genetic algorithm
for solving the vehicle routing problem with time windows. In: Soft Comp. in Industrial App. Volume 75 of Advances in Intelligent and Soft Computing. Springer,
Berlin, Heidelberg (2010) 263–270
25. Ursani, Z., Essam, D., Cornforth, D., Stocker, R.: Localized genetic algorithm
for vehicle routing problem with time windows. Appl. Soft Comput. 11(8) (2011)
5375–5390
26. Gehring, H., Homberger, J.: Parallelization of a two-phase metaheuristic for routing
problems with time windows. Journal of Heuristics 8(3) (2002) 251–276
27. Mester, D., Bräysy, O.: Active guided evolution strategies for large-scale vehicle
routing problems with time windows. Comput. Oper. Res. 32(6) (2005) 1593–1614
28. Kanoh, H., Tsukahara, S.: Solving real-world vehicle routing problems with time
windows using virus evolution strategy. Int. J. Know.-Based Intell. Eng. Syst.
14(3) (2010) 115–126
29. Berger, J., Barkaoui, M.: A parallel hybrid genetic algorithm for the vehicle routing
problem with time windows. Comp. and Oper. Res. (2004) 2037–2053
30. Labadi, N., Prins, C., Reghioui, M.: A memetic algorithm for the vehicle routing
problem with time windows. Oper. Res. 42(3) (2008) 415–431
31. Nagata, Y., Bräysy, O.: A powerful route minimization heuristic for the vehicle
routing problem with time windows. Oper. Res. Lett. 37(5) (2009) 333–338
32. Blocho, M., Czech, Z.J.: An improved route minimization algorithm for the vehicle
routing problem with time windows. Studia Informatica 32(99) (2010) 5–19
33. Nalepa, J., Czech, Z.J.: Adaptive threads co-operation schemes in a parallel heuristic algorithm for the vehicle routing problem with time windows. Theoretical and
Applied Informatics 24(3) (2012) 191–203
34. Kawulok, M., Nalepa, J.: Support vector machines training data selection using a
genetic algorithm. In Gimelfarb, G., Hancock, E., Imiya, A., Kuijper, A., Kudo, M.,
Omachi, S., Windeatt, T., Yamada, K., eds.: Structural, Syntactic, and Statistical
Pattern Recognition. Volume 7626 of Lecture Notes in Computer Science. Springer
Berlin Heidelberg (2012) 557–565
35. : CI TASK. Website (2011) http://www.task.gda.pl/kdm/sprzet/Galera.
| 9cs.NE
|
1
Gaussian 1-2-1 Networks:
Capacity Results for mmWave Communications
Yahya H. Ezzeldin† , Martina Cardone? , Christina Fragouli† , Giuseppe Caire∗
†
UCLA, Los Angeles, CA 90095, USA, Email: {yahya.ezzeldin, christina.fragouli}@ucla.edu
?
University of Minnesota, Minneapolis, MN 55404, USA, Email: [email protected]
arXiv:1801.02553v1 [cs.IT] 8 Jan 2018
∗
Technische Universität Berlin, Berlin, Germany, Email: [email protected]
Abstract
This paper proposes a new model for wireless relay networks referred to as “1-2-1 network”, where two nodes can
communicate only if they point “beams” at each other, while if they do not point beams at each other, no signal can
be exchanged or interference can be generated. This model is motivated by millimeter wave communications where,
due to the high path loss, a link between two nodes can exist only if beamforming gain at both sides is established,
while in the absence of beamforming gain the signal is received well below the thermal noise floor. The main result
in this paper is that the 1-2-1 network capacity can be approximated by routing information along at most 2N + 2
paths, where N is the number of relays connecting a source and a destination through an arbitrary topology.
I. I NTRODUCTION
Millimeter Wave (mmWave) communications are expected to play a vital role in 5G mobile communications,
expanding the available spectrum and enabling multi-gigabit services that range from ultra-high definition video,
to outdoor mesh networks, to autonomous vehicle platoons and drone communication [1]. Although several works
examine channel modeling for mmWave networks [2], the information theoretic capacity of mmWave relay networks
is yet relatively unexplored. In this paper, we present capacity results for a class of networks that we term 1-2-1
networks that offer a simple yet informative model for mmWave networks.
The inherent characteristic of mmWave communications that our model captures is directivity: mmWave requires
beamforming with narrow beams to compensate for high path loss. To establish a communication link, both the
mmWave transmitter and receiver employ antenna arrays that they electronically steer to direct their beams towards
each other - we term this a 1-2-1 link, as both nodes need to focus their beams to face each other for the link to be
active. Thus, in 1-2-1 networks, instead of broadcasting or interference, we have coordinated steering of transmit
and receive beams to activate different links at each time. An example of a diamond network with N = 4 relays
is shown in Fig. 1, where two different states for the configuration of the transmit/receive beams are depicted and
the resulting activated links are highlighted.
Our main results are as follows. We consider a source connected to a destination through an arbitrary topology
of N relay nodes, and derive a min-cut Linear Program (LP) that outer-bounds the capacity within a constant
January 9, 2018
DRAFT
2
R1
S
R2
R1
D
S
R2
R3
R3
R4
R4
D
Fig. 1: 1-2-1 network with N = 4 relays and two states.
gap, which only depends on N ; we then show that its dual is equivalent to a fractional path utilization LP. That
is, we show that routing is a capacity achieving strategy (up to a constant gap). Moreover, out of an exponential
number of paths that potentially connect the source to the destination, we show we need to utilize at most 2N + 2
to approximately achieve the capacity. We also prove tighter results for classes of networks. For example, for the
special case of diamond (or one-layer) networks, where the source is connected to the destination through one
layer of non-interfering relays as in Fig. 1, we prove that we can approximately achieve the network capacity by
routing information along at most two paths, independently of the total number N of relays. As a result, selecting
to operate the best path always achieves half the diamond network capacity.
Although we believe that 1-2-1 networks capture the essence of mmWave networks and enable to build useful
insights on near-optimal information flow algorithms, we recognize that this model makes a number of simplifying
assumptions that include: 1) we assume no interference among communication links (a reasonable assumption for
relays spaced further apart than the beam width), and 2) we do not take into account the overhead of channel
knowledge, and of beam-steering.
Related Work. Several studies examine channel modeling for mmWave networks [2], [3]. However, to the best
of our knowledge, the information theoretic capacity under optimal scheduling has not been analyzed. Recent
studies in networking design communication protocols for mmWave mesh networks [4], [5]. Closer to this work
are perhaps works that examine directional networks in the Gupta&Kumar framework [6], however they only look
at order arguments for multiple unicast sessions [7], and do not consider schedules that arrange for both receiver
and transmitter beams to align.
Paper Organization. Section II describes the N -relay Gaussian 1-2-1 network and derives a constant gap approximation of its capacity; Section III presents our main results; Section IV contains one of the main proofs of this
work.
II. S YSTEM M ODEL AND C APACITY F ORMULATION
With [n1 : n2 ] we denote the set of integers from n1 to n2 ≥ n1 ; Card(S) is the cardinality of the set S; ∅ is
the empty set; 1P is the indicator function; 0N indicates the all-zero vector of length N .
We consider an N -relay Gaussian 1-2-1 network where N relays assist the communication between a source
node (node 0) and a destination node (node N + 1). In particular, in this 1-2-1 network, at any particular time, a
node in the network can only direct (beamform) its transmissions towards at most another node. Similarly, a node
January 9, 2018
DRAFT
3
can only receive transmissions from at most another node (to which its receiving beam points towards). Thus, each
node i ∈ [0 : N + 1] in the network is characterized by two states, namely Si,t and Si,r that represent the node
towards which node i is beamforming its transmissions and the node towards which node i is pointing its receiving
beam, respectively. In particular, ∀i ∈ [0 : N + 1], we have that
Si,t ⊆ [1 : N + 1]\{i}, Card(Si,t ) ≤ 1,
Si,r ⊆ [0 : N ]\{i},
(1a)
Card(Si,r ) ≤ 1,
where S0,r = SN +1,t = ∅ since the source node always transmits and the destination node always receives. We
consider two modes of operation at the relays, namely Full-Duplex (FD) and Half-Duplex (HD). In FD, relay
i ∈ [1 : N ] can be simultaneously receiving and transmitting, i.e., we can have both Si,t 6= ∅ and Si,r 6= ∅. In
HD, relay i ∈ [1 : N ] can either receive or transmit, i.e., if Si,t 6= ∅, then Si,r = ∅ and vice versa. In particular,
∀i ∈ [1 : N ], we have that
2 if relays operate in FD
Card(Si,t )+Card(Si,r ) ≤
.
1 if relays operate in HD
(1b)
We can now write the memoryless channel model for this Gaussian 1-2-1 network. We have that ∀j ∈ [1 : N + 1]
Yj = Zj +
X
hji 1{i∈Sj,r ,
j∈Si,t } Xi ,
(2)
i∈[0:N ]\{j}
where: (i) Si,t and Si,r are defined in (1); (ii) Xi (respectively, Yi ) denotes the channel input (respectively, output)
at node i; (iii) hji ∈ C represents the complex channel coefficient from node i to node j; the channel coefficients
are assumed to be constant for the whole transmission duration and known by the network; (iv) the channel inputs
are subject to an individual power constraint, i.e., E[|Xk |2 ] ≤ P, k ∈ [0 : N ]; (v) Zj , j ∈ [1 : N + 1] indicates
the additive white Gaussian noise at the j-th node; noises across the network are assumed to be independent and
identically distributed as CN (0, 1).
By using a similar approach as the one proposed in [8], the channel model in (2) can be modified to incorporate
bi = (Si , X i ) be the input to the channel at
the state variables in the channel inputs. In particular, let the vector X
node i ∈ [0 : N ], where: (i) Si = (Si,t , Si,r ) with Si,t and Si,r being defined in (1) and (ii) X i ∈ CN +1 , with
elements X i (k) defined as
X i (k) = Xi 1{k∈Si,t } .
(3)
In other words, X i as a vector is a function of Si,t and the input of the original channel Xi . When node i is not
transmitting, i.e., Si,t = ∅, then X i = 0N +1 . It is not hard to see that the power constraint on Xi extends to X i
bi ,
since at most one single index appears in the vector (recall that Card(Si,t ) ≤ 1). Using this new channel input X
we can now equivalently rewrite the channel model in (2) as
hjS X S (j) + Zj if Card(Sj,r ) = 1
j,r
j,r
Yj =
.
0
otherwise
January 9, 2018
(4)
DRAFT
4
The capacity1 C of the network defined in (3) and (4) is not known, but can be approximated to within a constant-gap
as stated in the following theorem which is proved in Appendix A.
Theorem 1. The capacity C of the network defined in (3) and (4) can be lower and upper bounded as
Ccs,iid ≤ C ≤ Ccs,iid + GAP,
(5a)
Ccs,iid = max
min
X
λs :λs ≥0 Ω⊆[1:N ]∪{0}
P
(i,j):i∈Ω,
s λs =1
j∈Ωc
X
λs `j,i ,
s:
(5b)
j∈si,t ,
i∈sj,r
2
`j,i = log 1 + P |hji | ,
(5c)
GAP = G1 + G2 + G3
= (N +1) log e+2 log(N +2)+N log(Card(S1 )),
(5d)
where: (i) Ωc = [0 : N + 1]\Ω; (ii) λs = P(S[0:N +1] = s) is the joint distribution of the states, where s enumerates
the possible network states S[0:N +1] ; (iii) Card(S1 ) is defined as
(N + 1)2 if relays operate in FD
Card(S1 ) =
.
2N + 1
if relays operate in HD
(5e)
The variable GAP in (5d) only depends on the number of relays N and represents the maximum loss incurred
by using independent inputs and deterministic schedules at the nodes. In particular, G1 represents the beamforming
loss due to the use of independent inputs, while G2 (respectively, G3 ) accounts for the loss incurred by using a
fixed schedule at the source and destination (respectively, at the relays), as we explain next. Note that from (3), the
input at the i-th node is also characterized by the random state variable Si,t (which indicates to which node – if any
– node i is transmitting). Therefore, information can be conveyed from the source to the destination by randomly
switching between these states. However, as first highlighted in [8] in the context of the HD relay channel, this
random switch can only improve the capacity by a constant, whose maximum value equals the logarithm of the
cardinality of the support of the state random variable. It therefore follows that the capacity can be approximated to
within this constant by using a fixed/deterministic schedule at the nodes. In particular, for the source and destination
the cardinality of the support of their state random variable equals N + 2 (since the source and the destination can
only be transmitting to and receiving from at most one node, respectively). Differently, the cardinality of the support
of the state variable at the relays depends on the mode of operation (either FD and HD) and is given by (5e).
In other words, Ccs,iid in (5) – which can be achieved using QMF as in [9] or NNC as in [10] – is a constant
gap away from the capacity C of the network defined in (3) and (4). Thus, in the rest of the paper we analyze
Ccs,iid , which we refer to as the approximate capacity for the Gaussian 1-2-1 network.
1 We
use standard definitions for codes, achievable rates and capacity.
January 9, 2018
DRAFT
5
Remark 1. In the rest of the paper, we assume that the point-to-point link capacities are rational numbers. In fact,
as we prove in what follows, when the capacities `j,i ∈ R, we can always further bound the approximate capacity
CR
cs,iid as
Q
R
CQ
cs,iid ≤ Ccs,iid ≤ Ccs,iid + ,
(6)
ˆ
where > 0 and CQ
cs,iid is the approximate capacity of a network with link capacities `j,i such that
`ˆj,i ≤ `j,i ≤ `ˆj,i +
, `ˆj,i ∈ Q.
(N + 1)2
(7)
Note that such an assignment always exists since the set of rationals Q is dense in R. With this, we have
CR
cs,iid = max
X
min
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
X
λs `j,i
s:
j∈si,t ,
i∈sj,r
(7)
≤ max
X
min
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
= max
min
X
ˆ
λs `j,i +
(N + 1)2
s:
j∈si,t ,
i∈sj,r
X
s
min
max
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
≤ max
min
X
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
j∈si,t ,
i∈sj,r
X
λs ≤1
≤
X
λs `ˆj,i +
s:
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
P
X
λs
s:
(N + 1)2
(i,j):i∈Ω, j∈s ,
i,t
j∈Ωc
X
i∈sj,r
X
λs `ˆj,i +
s:
j∈si,t ,
i∈sj,r
(N + 1)2
(i,j):i∈Ω,
j∈Ωc
X
X
λs `ˆj,i + = CQ
cs,iid + .
s:
j∈si,t ,
i∈sj,r
Moreover, we have that
CR
cs,iid = max
min
X
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
X
λs `j,i
s:
j∈si,t ,
i∈sj,r
(7)
≥ max
min
X
Pλ:λ≥0 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
s λs =1
j∈Ωc
X
λs `ˆj,i = CQ
cs,iid .
s:
j∈si,t ,
i∈sj,r
This proves (6) and hence, in the rest of the paper, we will assume that the point-to-point link capacities are rational
numbers.
January 9, 2018
DRAFT
6
III. M AIN R ESULTS
We here present our main results on Gaussian 1-2-1 networks and discuss their implications. Our first main result
is that for FD networks, Ccs,iid in (5b) can be computed as the sum of fractions of the FD capacity of the paths in
the network.
Theorem 2. For any N -relay Gaussian FD 1-2-1 network, we have that
X
P1 : Ccs,iid = max
xp Cp
p∈P
(P1a) xp ≥ 0
X
p
≤1
xp fp.nx(i),i
(P1b)
∀p ∈ P,
(8)
∀i ∈ [0 : N ],
p∈Pi
(P1c)
X
p
xp fi,p.pr(i)
≤1
∀i ∈ [1 : N +1],
p∈Pi
where: (i) P is the collection of all paths from the source to the destination; (ii) Pi ⊆ P is the collection of paths
that pass through node i ∈ [0 : N + 1] (clearly, P0 = PN +1 = P since all paths pass through the source and
the destination); (iii) Cp is the FD capacity of the path p ∈ P, i.e., Cp = min(i,j)∈p `j,i ; (iv) p.nx(i) (respectively,
p.pr(i)) with i ∈ [0 : N + 1] is the node following (respectively, preceding) node i ∈ [0 : N + 1] in path p ∈ P
p
(clearly, p.pr(0) = p.nx(N + 1) = ∅); (v) fj,i
is the optimal activation time for the link of capacity `j,i when the
path p ∈ P, such that (i, j) ∈ p, is operated, i.e.,
p
fj,i
=
Cp
.
`j,i
(9)
Proof. The proof of Theorem 2 is delegated to Section IV.
In the LP in Theorem 2, the variable xp represents the fraction2 of time the path p ∈ P is utilized in the network.
Moreover, each of the constraints in (P1b) (respectively, (P1c)) ensures that a node i ∈ [0 : N + 1] - even though
it can appear in multiple paths in the network - does not transmit (respectively, receive) for more than 100% of the
time.
Lemma 3 follows from P1 in Theorem 2, and states that, although the number of paths P in general is exponential
in the number of relays N , we need to use at most a linear number of paths; this can also be translated to a guarantee
on the rate that can be achieved when only the best path is operated.
Lemma 3. For any N -relay Gaussian FD 1-2-1 relay network, we have the following guarantees:
(L1) For a network with arbitrary topology, the approximate capacity Ccs,iid can always be achieved by activating
at most 2N + 2 paths in the network.
(L2) For a network with arbitrary topology, the best path has an FD capacity C1 such that C1 ≥
1
2N +2 Ccs,iid .
(L3) For a 2-layer relay network with M = N/2 relays per layer, the approximate capacity Ccs,iid can be achieved
by activating at most 2M + 1 paths in the network.
2 Note
p
that xp in P1 implicitly satisfies that xp ≤ 1, ∀p ∈ P. This is due to the fact that for any path p ∈ P, the definition of fj,i
in (9)
p
implies that at least one constraint in (P1b) and (P1c) has fj,i
= 1.
January 9, 2018
DRAFT
7
(L4) For a 2-layer relay network with M = N/2 relays per layer, the best path has an FD capacity C1 such that
C1 ≥
1
2M +1 Ccs,iid .
.
Proof.
Proof of L1: The LP P1 in (8) is bounded and hence there always exists an optimal corner point. In particular, at
any corner point in P1, we have at least P = Card(P) constraints satisfied with equality among (P1a), (P1b) and
(P1c). Therefore, we have at least P − 2N − 2 in (P1a) satisfied with equality (since (P1b) and (P1c) combined
represent 2N + 2 constraints). Thus, at least P − 2N − 2 paths are not operated, which proves the statement in L1.
Proof of L3: Similar to the proof above for L1, a corner point in the LP P1 has at most 2N +2 constraints among
(P1b) and (P1c) satisfied with equality. To prove L3, we need to show that in the case of a 2-layered network
and we have 2N + 2 equality satisfying constraints, then at least one of the equations is redundant.Note that for a
2-layer relay network, any path p in the network has four nodes and is written as 0 − p(1) − p(2) − N + 1 where
p(1) and p(2) represent the node in the path from layer 1 and layer 2, respectively. We assume that the relays in
the first layer are indexed with [1 : N/2] and the second layer relays are indexed with [N/2+ : N ]. Thus, for any
path, p(1) ∈ [1 : N/2] and p(2) ∈ [N/2 + 1 : N ]. Now assume that all constraints (P1b) and (P1c) satisfied with
equality, then by adding all (P1b) constraints for i ∈ [1 : N/2] and subtracting from all constraints from (P1c) for
j ∈ [N/2 + 1 : N ], we get
N/2
LHS :
X
(P1b)i −
i=1
N
X
N/2
(P1c)j =
XX
i=1 p∈Pi
j=N/2+1
=
X
RHS :
X
i=1
(P1b)i −
N
X
p
xp fp(2),p(1)
−
j=N/2+1
X
i=1
p
xp fj,p.pr(j)
X
p
xp fp(2),p(1)
=0
p∈P
N/2
(P1c)j =
X
j=N/2+1 p∈Pj
p∈P
N/2
N
X
p
xp fp.nx(i),i
−
1−
N
X
1 = 0,
j=N/2+1
Thus, when all constraints (P1b) and (P1c) are satisfied with equality, at least one of them redundant, which proves
the statement L3.
Proof of L2 and L4: The proof of L2 follows directly from L1 by considering only the 2N + 2 paths needed
to achieve Ccs,iid and picking the path that has the largest FD capacity among them. In particular, the guarantee in
L2 is true for the selected path due to the fact that for any feasible point in P1, xp ≤ 1, ∀p ∈ P. The proof of L4
from L3 follows the same argument used to prove L2 from L1.
The result L3 in Lemma 3 suggests that for 1-2-1 networks with particular structures, we can further reduce the
number of active paths needed to achieve the approximate capacity. In particular, we explore this observation in the
context of 1-2-1 Gaussian diamond networks operating in FD and HD, through the following two lemmas proved
in Appendix B, Appendix C and Appendix D.
Lemma 4. For the N -relay Gaussian diamond 1-2-1 network, we can calculate the approximate capacity Ccs,iid
January 9, 2018
DRAFT
8
as
P1d : Ccs,iid = max
P
p∈[1:N ]
xp Cp
d
(P1a) 0 ≤ xp ≤ 1
∀p ∈ [1:N ],
P
C
p
(P1b)d
p∈[1:N ] xp `p,0 ≤ 1,
P
Cp
(P1c)d
p∈[1:N ] xp `N+1,p ≤ 1,
(10)
where: (i) P is the collection of all paths from the source to the destination; (ii) Cp is the capacity of the path
0 → p → N + 1 and its value depends on whether the network is operating in FD or HD, namely
min{` , `
p,0 N +1,p } if relays operate in FD
.
Cp =
`p,0 `N +1,p
if relays operate in HD
`p,0 +`N +1,p
Lemma 5. For an N -relay Gaussian FD diamond relay network, we have the following guarantees:
(L1) If the network is operating in FD, then the approximate capacity Ccs,iid can always be achieved by activating
at most 2 relays in the network, independently of N .
(L2) If the network is operating in HD, then the approximate capacity Ccs,iid can always be achieved by activating
at most 3 relays in the network, independently of N .
(L3) In both FD and HD networks, the best path has a capacity C1 such that C1 ≥
1
2 Ccs,iid ;
furthermore, this
guarantee is tight for both the FD and HD cases, i.e., there exists a class of Gaussian diamond 1-2-1 networks
such that C1 ≤ 21 Ccs,iid both for the FD and HD cases.
The results in L1 and L2 in Lemma 5 are surprising as they state that, independently of the total number of
relays in the network, there always exists a subnetwork of 2 (in FD) and 3 (in HD) relays that achieves the full
network approximate capacity. Moreover, the guarantee provided by L3 is tight. To see this, consider N = 2
and `1,0 = `3,2 = 1 and `3,1 = `2,0 = X → ∞. For this network, we have that the approximate capacity is
Ccs,iid = `1,0 + `3,2 = 2, while the capacity of each path (both in FD and HD) is C1 = min {1, X} = 1, hence
C1 /Ccs,iid = 1/2.
IV. P ROOF OF T HEOREM 2
We here prove Theorem 2. We note that for a fixed λs , the inner minimization in (5b) is the standard min-cut
problem over a graph with link capacities given by
(s)
`j,i =
X
s:
j∈si,t , i∈sj,r
λs `j,i .
(11)
Since the min-cut problem is the dual for the standard max-flow problem, then we can replace the inner minimization
in (5b) with the max-flow problem over the graph with link capacities defined in (11) to give that
Ccs,iid = max max
λs :λs ≥0
P
s λs =1
N
+1
X
Fj,0
j=1
(s)
0 ≤ Fj,i ≤ `j,i
X
Fj,i =
j∈[1:N +1]\{i}
January 9, 2018
(i, j) ∈ [0 : N ] × [1 : N +1],
X
Fi,k
(12)
i ∈ [1 : N ],
k∈[0:N ]\{i}
DRAFT
9
where Fj,i represents the flow from node i to node j.
The max-flow problem can be equivalently written as an LP with path flows instead of link flows, thus, (12) can
be written as P2 described next by using the path flows representation of the max-flow problem. A variable Fp is
used for the flow through the path p ∈ P.
P2
: Ccs,iid = max
X
Fp
p∈P
(P2a) Fp ≥ 0
∀p ∈ P,
X
(s)
Fp ≤`j,i =eq.(11) ∀(j, i)∈[1:N +1]×[0:N ],
(P2b)
p∈P,
(i,j)∈p,
j=p.nx(i)
(P2c)
P
s
λs ≤ 1,
(P2d) λs ≥ 0
∀s,
The constraints (P2b) ensure that, for any link from node i to node j, the sum of the flows through the paths that
(s)
use this link does not exceed the link modified capacity `j,i in (11). The constraint (P2c) is the same constraint
on λs as in (5b).
To prove Theorem 2, we first show that P2 above is equivalent to the following LP P3, and then prove that P3
is equivalent to P1 in (8), which completes the proof.
P
P3 : Ccs,iid = max p∈P Fp
(P3a) Fp ≥ 0
X
(P3b)
Fp ≤ λ`j,i `j,i
∀p ∈ P,
∀(i, j)∈[0 : N ]×[1 : N +1],
p∈P,
(i,j)∈p,
j=p.nx(i)
(P3c)
X
λ`j,i ≤ 1
∀i ∈ [0 : N ],
j∈[1:N +1]\{i}
(P3d)
X
λ`j,i ≤ 1
∀j ∈ [1 : N + 1],
i∈[0:N ]\{j}
(P3e) λ`j,i ≥ 0
∀(i, j)∈[0 : N ]×[1 : N +1],
where λ`j,i , (i, j) ∈ [0 : N ] × [1 : N + 1] represents the fraction of time the link of capacity `j,i is active. The
constraints (P3b) are similar to (P2b) except that the capacity of a link is now modified through a multiplication by
λ`j,i . The constraints (P3c) ensure that, for any transmitting node i, the sum of the activation times of its outgoing
links is less than 100% of the time. Similarly, (P3d) ensure the same logic for the incoming edges to a receiving
node.
We delegate the proof of the equivalence between P3 and P1 to Appendix E, and here prove the equivalence
between P2 and P3, which is more involved. To do so, we first show that a feasible point in P2 gives a feasible
point in P3 with the same objective value. We define the following transformation
X
λ`j,i =
λs ,
s:
j∈si,t , i∈sj,r
while the variable Fp in P2 is the same as the variable Fp in P3. By substituting these transformations in the
constraints (P2a)-(P2d), it is not difficult to see that the constructed point (through the transformation) is feasible
January 9, 2018
DRAFT
10
in P3 and has the same objective value in P3 as the objective value in P2. Thus, the solution of P3 is at least as
large as the one of P2.
We now prove the direction from P3 to P2, by showing that every corner point in P3 can be transformed into a
feasible point in P2 with the same objective value. To do this, we introduce a visualization for P2 and P3 in terms
of bipartite graphs. We divide each node i ∈ [0 : N + 1] in the network into two nodes (iT and iR ) representing
the transmitting and receiving functions of the node; note that 0R = (N + 1)T = ∅ since the source (node 0)
is always transmitting and the destination (node N + 1) is always receiving. This gives us the bipartite graph
GB = (T , R, E), where the vertices T (respectively, R) are the transmitting modules of our nodes (respectively, R
collects our receiving modules), and we have an edge (iT , jR ) ∈ E for each link in the network. It is easy to see
that a valid state in P2 represents a matching in the bipartite graph GB . Furthermore, we can write the program P3
in terms of GB by simply renaming all λ`j,i with λjR ,iT , i.e., the activation time of the edge (iT , jR ) in GB .
Starting with a corner point in P3, we can follow the procedure described below to construct a feasible point in P2;
note that since all coefficients (i.e., `jR ,iT ) in P3 are rational (see Remark 1), the corner points λ?`j
R ,iT
are rational. The main intuition is to use the fractions λ?`j
R ,iT
∈ Q, ∀(iT , jR )
and our bipartite graph GB to construct a bipartite
multigraph with edges of unit capacity such that for higher values of λ?`j
R ,iT
, we will have more parallel edges
from the iT -th node to the jR -th node. In particular, our procedure consists of the three following main steps.
Step 1. We multiply all λ?`j
R ,iT
if λ?`j
R ,iT
by the Least Common Multiple (LCM) M of their denominators (for simplicity,
= 0, then the denominator is set to be one). We then calculate nj,i = M λ?`j
and construct the bipartite multigraph
R ,iT
•
GB
, ∀iT , jR ∈ [0 : N + 1],
from GB that has nj,i parallel edges from node iT to node jR .
•
Step 2. We edge color the multigraph GB
. If any of the parallel edges from node iT to node jR are colored with
color ci , we say that ci activates the link (iT , jR ) in GB (or equivalently the link (i, j) in the network). Note that
•
is a bipartite graph, then there is an optimal coloring for the graph that uses ∆ colors, where ∆ is the
since GB
•
can have the same color, every
maximum degree of the nodes. Furthermore, since no two adjacent edges in GB
•
individual color represents a matching in the bipartite graph GB
(and by extension the graph GB ). Each of these
matchings represents a state in the network; since there are ∆ colors in total, then we assign to state si , i ∈ [1 : ∆]
(represented by color ci ) an activation time of wi = 1/∆.
Step 3. From Step 2, we know that each color in the network represents a matching in GB , and hence a state.
However, some colors can correspond to the same matching in GB , i.e., two or more colors activate exactly the
same set of links. We combine these colors together into one state by multiplying 1/∆ by the number of times
this state is repeated. The remaining unique states represent the feasible states in P2 that were obtained from our
optimal λ? from P3.
We now need to prove that the states that we obtain from the previous steps, in addition to the flows through
the paths that we have from P3, indeed give us a feasible point in P2. Without loss of generality, we will prove
that the states generated in Step 2 give a feasible point since combining similar states (Step 3) does not change the
total sum of activation times and does not change the amount of time a link is active (which is what we look for
in the constraint (P2b).
To show the feasibility of the constructed schedule, we first derive a consequence of the fact that λ?`j
R ,iT
January 9, 2018
is feasible
DRAFT
11
in P3. In particular, we can show the following inequality between the maximum degree ∆ and the constant LCM
M
∆ = max max
i∈[0:N ]
X
nj,i ,
j∈[1:N +1]
max
j∈[1:N +1]
X
= M max max
iT ∈[0:N ]
λ?`j
R ,iT
i∈[0:N ]
,
X
jR ∈[1:N +1]
nj,i
jR ∈[1:N +1]
max
X
λ?`j
(P3c,d)
R ,iT
≤
M,
(13)
iT ∈[0:N ]
where the inequality follows from the constraints (P3c) and (P3d) in P3. Using (13), we can now show that the
constructed schedule (that uses coloring arguments) is feasible in P2 (note that since Fp is unchanged then the
constraint in (P2a) is already satisfied due to the constraint in (P3a)). Moreover, we have
X
(P3b) : ∀(j, i),
Fp ≤ λ`j,i `j,i =
p∈P, (i,j)∈p,
j=p..nx(i)
λs
∆
`j,i ≤
M
=
X
(13)
s:(i,j)∈s
By construction:
X
λs =
s
By construction: λs =
nj,i
`j,i
M
X
λs `j,i =⇒ (P2b)
s:(i,j)∈s
X 1
1
= ∆ = 1 =⇒ (P2c)
∆
∆
s
1
≥ 0 =⇒ (P2d).
∆
From the discussion above, it follows that we can map a rational point λ?`j
R ,iT
in P3 to a feasible point in P2 that
has the same objective function value. Note that the variables Fp in P3 and P2 are unchanged and therefore, the
objective values will remain the same. Thus, the solution of P2 is at least as large as that of P3, concluding the
proof that P2 and P3 are equivalent.
January 9, 2018
DRAFT
12
A PPENDIX A
C ONSTANT G AP C APACITY A PPROXIMATION FOR THE G AUSSIAN 1-2-1 N ETWORK
The memoryless model of the channel allows to upper bound the channel capacity C using the cut-set upper
bound Ccs as
Ccs =
max
P{X
=
i ,Si }
min
(·) Ω⊆[1:N ]∪{0}
min
max
P{X
=
i ,Si }
(·) Ω⊆[1:N ]∪{0}
max
P{X
≤
i ,Si }
min
(·) Ω⊆[1:N ]∪{0}
min
max
P{X
i ,Si }
(·) Ω⊆[1:N ]∪{0}
bΩ ; YΩc |X
bΩc )
I(X
I(SΩ , X Ω ; YΩc |SΩc , X Ωc )
I(X Ω ; YΩc |SΩ , SΩc , X Ωc ) + I(SΩ ; YΩc |SΩc , X Ωc )
I(X Ω ; YΩc |S[0:N +1] , X Ωc ) + H(SΩ )
(a)
≤
min
max
P{X
i ,Si }
(·) Ω⊆[1:N ]∪{0}
(b)
= max
P{Si } (·) P{X
X
min
max
i }|{Si }
I(X Ω ; YΩc |S[0:N +1] , X Ωc ) + 2 log(N + 2) + N log(Card(S1 ))
(·) Ω⊆[1:N ]∪{0}
(c)
≤ max
min
P{Si } (·) Ω⊆[1:N ]∪{0} P{X
X
max
i }|{Si }
(·)
λs I(X Ω ; YΩc |S[0:N +1] =s, X Ωc )+2 log(N + 2)+N log(Card(S1 ))
s
λs I(X Ω ; YΩc |S[0:N +1] =s, X Ωc )+2 log(N + 2)+N log(Card(S1 )),
s
(14)
+1
where: (i) Ωc = [0 : N + 1]\Ω; (ii) P{X i ,Si } (·) is the probability distribution of the channel input {(Si , X i )}N
i=0 ;
(iii) SΩ = {Si |i ∈ Ω}; (iv) the inequality in (a) is due to the fact that the state variable at the source and destination
can take N +2 values (since the source can only be transmitting to at most one node and the destination can only be
receiving from at most one node), while at each relay the state variable can take Card(S1 ) values, where Card(S1 )
depends on the mode of operation at the relays, namely
(N + 1)2 if relays operate in FD
Card(S1 ) =
;
2N + 1
if relays operate in HD
(v) in the equality in (b) we use s to enumerate the possible network states S[0:N +1] and we denote with λs =
P(S[0:N +1] = s) the joint distribution of the states; (vi) the inequality in (c) follows from the max-min inequality.
b s , where the element [H
b s ]i,j is defined as
For a network state s, we define the channel matrix H
hij if i ∈ sj,t and j ∈ si,r
b
[Hs ]i,j =
0
otherwise,
(15)
where hij is the channel coefficient of the link from node j to node i. It is not difficult to see that every row (and
b s has at most one non-zero element and thus there exists a permutation matrix Π such that ΠH
b s is a
column) of H
diagonal matrix. Also, let s+ = {i|si,t 6= ∅, ∀i ∈ [0 : N ]} and s− = {i|si,r 6= ∅, ∀i ∈ [1 : N + 1]}.
January 9, 2018
DRAFT
13
With this, we can further simplify the mutual information expression in (14) as follows
X
max
λs I(X Ω ; YΩc |S[0:N +1] =s, X Ωc )
P{X
i }|{Si }
(a)
=
(·)
s
X
max
(·)
λs I(X s+ ,Ω ; Ys− ,Ωc |S[0:N +1] =s, X Ωc )
P{X
i }|{Si }
(b)
=
X
b s,Ω Ks,Ω H
bH
λs log det I + H
s,Ω
=
X
s
s
H b
b s,Ω
λs log det I + H
Hs,Ω Ks,Ω ,
(16)
s
where: (i) we define X s+ ,Ω as X s+ ,Ω =
X i (si,t ) i ∈ Ω ∩ s+
and Y s− ,Ωc as Y s− ,Ωc =
Y i i ∈ Ωc ∩ s− ;
(ii) the equality in (a) follows since, given the state s, all variables X i (j), with j 6= si,t , as well as all Yi with
si,r = ∅ are deterministic; (iii) the equality in (b) follows due to the maximization of the mutual information by
b s,Ω is a submatrix of H
b s (defined in (15)) and is defined as H
b s,Ω = [H
b s ]Ωc ,Ω and
the Gaussian distribution; (iv) H
T
Ks,Ω is the submatrix of the covariance matrix of the random vector X̄0 (s0,t ) X̄1 (s1,t ) . . . X̄N (sN,t ) , where
the rows and columns are indexed by Ω.
We now further upper bound the Right-Hand Side (RHS) of (16) using [10, Lemma 1], for any γ ≥ e − 1 as
follows
H b
H b
b s,Ω
b s,Ω
log det I + H
Hs,Ω Ks,Ω ≤ log det I + γ −1 P H
Hs,Ω + |Ω| log α(Ω, s, γ)
(a)
H
b s,Ω H
b s,Ω
≤ log det I + P H
+ |Ω| log α(Ω, s, γ),
(17)
where the inequality in (a) follows since γ > 1 and by applying Sylvester’s determinant identity and α(Ω, s, γ) is
defined based on [10, Lemma 1] as
eγ/e
α(Ω, s, γ) =
s,Ω )
rank(H
|s+ ∩Ω|
γ |s+ ∩Ω|
rank(Hs,Ω )
rank(H
)
s,Ω
if γ ≤ e trace(Ks,Ω
/P ) = e
rank(Hs,Ω )
|s+ ∩Ω|
(18)
otherwise.
If we select γ = e, then we have that
s,Ω )
rank(H
|s+ ∩Ω|
1
|s+ ∩ Ω|
α(Ω, s, e) = e
≤ max (ex) x = e.
x≥0
rank(Hs,Ω )
(19)
Now, if we substitute (16), (17) and (19) in (14), we get that
"
#
X
H
b
b
Ccs ≤ max
min
λs log det I + P Hs,Ω Hs,Ω + |Ω| log e +2 log(N +2)+N log(Card(S1 ))
P{Si } (·) Ω⊆[1:N ]∪{0}
≤ max
min
P{Si } (·) Ω⊆[1:N ]∪{0}
|
s
X
s
b s,Ω H
b H + (N + 1) log e + 2 log(N +2)+N log(Card(S1 )) .
λs log det I + P H
s,Ω
|
{z
}
GAP
{z
}
Ccs,iid
(20)
The main observation in (20) is that an i.i.d Gaussian distribution on the inputs and a fixed schedule are within
a constant additive gap from the information-theoretic cut-set upper bound on the capacity of the 1-2-1 network.
With this, we can argue that Ccs,iid is within a constant gap of the capacity. This is due to the fact that Ccs,iid can
be achieved using QMF as in [9] or Noisy Network Coding as in [11].
January 9, 2018
DRAFT
14
Due to the special structure of the Gaussian 1-2-1 network we can further simplify Ccs,iid by making use of the
b s,Ω in (20). In particular, recall that, since every row (and column) in H
b s,Ω has at most one non-zero
structure of H
b s is a diagonal matrix (not necessarily square).
element, then there exists a permutation matrix Πs,Ω such that Πs,Ω H
Thus we have
(a)
H
H
b s,Ω H
b s,Ω
b s,Ω H
b s,Ω
log det I + P H
= log det I + P Πs,Ω H
ΠTs,Ω
min{|Ω|,|Ωc |}
(b)
X
=
2
b s,Ω ]i,i
log 1 + P [Πs,Ω H
,
(21)
i=1
where: (i) the equality in (a) follows since permutation matrices are orthogonal matrices and thus multiplying by
them only permutes the singular values of a matrix; (ii) the equality in (b) follows since the permuted channel
b s,Ω can be represented as a parallel MIMO channel with min{|Ω|, |Ωc |} active links. We can rewrite
matrix Πs,Ω H
the expression in (21) as
T
b s,Ω H
b s,Ω
log det I + P H
=
2
b
log 1 + P [H]j,i
X
(i,j):
i∈s+ ∩Ω, j∈s− ∩Ωc ,
j∈si,t , i∈sj,r
2
log 1 + P |hji | .
X
=
(22)
(i,j):
i∈s+ ∩Ω, j∈s− ∩Ωc ,
j∈si,t , i∈sj,r
2
Thus, by letting `j,i = log 1 + P |hji | , we arrive at the following expression for Ccs,iid
Ccs,iid =
=
max
X
min
λ:kλk1 =1 Ω⊆[1:N ]∪{0}
λ≥0
max
λ:kλk1 =1 Ω⊆[1:N ]∪{0}
λ≥0
=
max
min
s
X
λs
s
1{j∈si,t ,
X
1{i∈Ω,
min
X
λ:kλk1 =1 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
λ≥0
j∈Ωc
=
max
min
j∈Ωc }
λ:kλk1 =1 Ω⊆[1:N ]∪{0}
(i,j):i∈Ω,
λ≥0
j∈Ωc
λs 1{j∈si,t ,
i∈sj,r } `j,i
X
λs `j,i
s:
j∈si,t ,
i∈sj,r
(s)
X
X
s
max
i∈sj,r } 1{i∈Ω, j∈Ωc } `j,i
(i,j)∈[0:N +1]2
λ:kλk1 =1 Ω⊆[1:N ]∪{0}
(i,j)∈[0:N +1]2
λ≥0
=
`j,i
(i,j):
i∈s+ ∩Ω, j∈s− ∩Ωc ,
j∈si,t , i∈sj,r
X
min
X
λs
`j,i ,
(23)
(s)
where `j,i is defined as
X
(s)
`j,i =
λs `j,i .
s:
j∈si,t ,
i∈sj,r
This concludes the proof that the capacity C of the Gaussian 1-2-1 network described in (4) can be characterized
to within a constant gap as expressed in (5).
January 9, 2018
DRAFT
15
A PPENDIX B
G AUSSIAN FD D IAMOND 1-2-1 N ETWORK : P ROOF OF L EMMA 4 AND L EMMA 5(L1)
In this section, we prove Lemma 4 for FD and Lemma 5(L1) by analyzing the Gaussian FD 1-2-1 FD network
with a diamond topology. In this network the source communicates with the destination by hopping through one
layer of N non-interfering relays. For this network the LP P1 in (8) can be further simplified by leveraging the
two following implications of the sparse diamond topology:
1) In a Gaussian 1-2-1 diamond network, we have N disjoint paths from the source to the destination, each
passing through a different relay. We enumerate these paths with the index i ∈ [1 : N ] depending on which
relay is in the path. Moreover, each path i ∈ [1 : N ] has a FD capacity equal to Ci = min {`i,0 , `N +1,i };
2) In the Gaussian FD 1-2-1 diamond network, each relay i ∈ [1 : N ] appears in only one path from the source
to the destination. Thus, when considering constraints (P1b) and (P1c) in (8) for i ∈ [1 : N ] gives us that
xi
Ci
≤1
`i,0
&
xi
Ci
≤ 1.
`N +1,i
(24)
3) Note that Ci = min{`i,0 , `N +1,i }. Therefore, one of the coefficients Ci /`i,0 or Ci /`N +1,i in (24) is equal to 1.
This implies that a feasible solution of P1d , has x1 ≤ 1 and x2 ≤ 1. Therefore, the constraints xi ≤ 1, ∀i ∈
[1 : N ], albeit redundant, can be added to the LP without reducing the feasibility region.
4) In the Gaussian FD 1-2-1 diamond network, the constraints due to the source and destination nodes, namely
(P1b) for i = 0 and (P1c) for i = N + 1 in (8) gives us that
X
X
Ci
Ci
xi
≤ 1,
xi
≤ 1.
`i,0
`N +1,i
i∈[1:N ]
(25)
i∈[1:N ]
Note that the constraints in (25) make the constraints (24) redundant.
By considering the two implications above, we can readily simplify P1 in (8) for Gaussian FD 1-2-1 networks with
a diamond topology as follows
X
P1d : Ccs,iid = max
xi Ci
i∈[1:N ]
(P1a)d 0 ≤ xi ≤ 1
∀i ∈ [1 : N ],
X
C
i
(P1b)d
xi
≤ 1,
`i,0
i∈[1:N ]
X
Ci
(P1c)d
xi
≤ 1,
`N +1,i
(26)
i∈[1:N ]
which is the LP we have in Lemma 4.
To prove Lemma 5(L1), we observe that for a bounded LP, there always exists an optimal corner point.
Furthermore, at any corner point in the LP P1d , we have at least N constraints satisfied with equality among
(1a)d , (1b)d and (1c)d . Therefore, we have at least N − 2 constraints in (1a)d satisfied with equality that make
linearly independent equations (since (1b)d and (1c)d combined represent only two constraints). Furthermore, recall
that as mentioned earlier all constraints xi ≤ 1 are redundant. Thus, at least N −2 relays are turned off (i.e., xi = 0),
i.e., at most two relays are sufficient to characterize the approximate capacity of any N -relay Gaussian FD 1-2-1
network with a diamond topology.
January 9, 2018
DRAFT
16
A PPENDIX C
G AUSSIAN HD D IAMOND 1-2-1 N ETWORK : P ROOF OF L EMMA 4 AND L EMMA 5(L2)
We prove Lemma 4 for HD diamond networks in the first subsection and later prove Lemma 5(L2) in the
following subsection.
A. Proof of Lemma 4 for an HD diamond network
Throughout this section, we slightly abuse notation by defining `i = `i,0 and ri = `N +1,i . Based on this definition,
we can write the approximate capacity expression (5b) as
Ccs,iid
X
X
X
λs `i +
= max
min
λs :λs ≥0 Ω⊆[1:N ]∪{0}
P
c
s:
s
i∈Ω
λs =1
i∈Ω
i∈s0,t ,
0∈si,r
X
s:
(N +1)∈si,t ,
i∈sN +1,r
λs
ri
N
X
X
1{i∈Ωc }
λ
`
+
1
s
i
{i∈Ω}
λs :λs ≥0 Ω⊆[1:N ]∪{0}
s:
P
i=1
λ =1
= max
s
min
N
X
X
λs `i ,
= max
min
min
λs :λs ≥0 Ω⊆[1:N ]∪{0}
P
s:
i=1
s λs =1
0,t ,
i∈s
0∈s
i,r
X
s:
(N +1)∈si,t ,
i∈sN +1,r
i∈s0,t ,
0∈si,r
s
λs
ri
X
.
r
λ
s i
s:
(N +1)∈si,t ,
(27)
i∈sN +1,r
Our first directive is to show that the approximate capacity Ccs,iid in (27) is equivalent to solving the LP P4
P4 :
maximize
N
X
λ`i `i
i=1
subject to
∀i ∈ [1 : N ],
(P4a) λ`i `i = λri ri
(P4b)
N
X
λ`i ≤ 1,
i=1
(P4c) λ`i + λri ≤ 1
N
X
λri ≤ 1,
(28)
i=1
∀i ∈ [1 : N ],
where: (i) fi = λ`i `i = λri ri represents the data flow through the i-th relay; (ii) λ`i (respectively, λri ) represents
the fraction of time in which the link from the source to relay i (respectively, from relay i to the destination) is
active. Note that since the network is operating in HD, then in (27), we have that |si,t | + |si,r | ≤ 1, ∀i ∈ [1 : N ]
which is captured by the constraint (P4c) above.
January 9, 2018
DRAFT
17
To show the first direction (i.e., a feasible schedule in (27) gives a feasible point in the LP P1), we define the
following transformation
X
∀i ∈ [1 : N ] : fi = min
λs `i ,
s:
0,t ,
i∈s
0∈s
i,r
λ `i =
fi
,
`i
λri =
X
r
λ
s i
s:
(N +1)∈si,t ,
i∈sN +1,r
fi
.
ri
(29)
Using this transformation, we have that
∀i ∈ [1 : N ]
X
N
N
N
X
X
X
X
1
fi
r
λ
λs `i ,
=
min
λ`i =
s i
`
`
s:
i=1
i=1 i
i=1 i
i∈ss:0,t ,
(N +1)∈si,t ,
0∈s
i∈sN +1,r
i,r
!
N
X
1 X
min{`i , ri }
≤
λs min{`i , ri } ≤
≤1
`
`i
s
i=1 i
N
N
N
X
X
X 1
X
X fi
=
min
λri =
λs `i ,
λ
r
s i
r
r
s:
s:
i=1 i
i=1
i=1 i
j∈s
,
(N
+1)∈s
,
i,t
i,t
i∈sj,r
i∈sN +1,r
!
N
X
min{`i , ri }
1 X
≤
λs min{`i , ri } ≤
≤1
r
ri
s
i=1 i
X
X
1
1
r
λ
λs `i ,
λ`i +λri =
+
min
s i
`i
ri
s:
s:
i∈s
,
(N
+1)∈s
,
0,t
i,t
0∈s
i∈sN +1,r
i,r
X
X
X
X
r
`
i
i
=min
λs ,
λs
λ
+min
,
λ
s
s
`i
ri
s:
s:
s:
s:
i∈s
,
i∈s
,
(N
+1)∈s
,
(N
+1)∈s
,
i,t
i,t
0∈s0,t
0∈s0,t
i∈sN +1,r
i∈sN +1,r
i,r
i,r
λ`i `i = fi = λri ri ,
X
≤
λs +
s:
i∈s0,t ,
0∈si,r
X
s:
(N +1)∈si,t ,
i∈sN +1,r
X
λs
λs = 1
≤
s
=⇒ (P4a)
=⇒ (P4b)
=⇒ (P4b)
=⇒ (P4c).
Thus, a feasible schedule in (27) gives a feasible point in the LP P4 in (28). Furthermore, by substituting (29)
in (27), we get that the rate achieved by the schedule is
X
N
N
N
X
X
X
X
min
λs `i ,
λ
r
=
f
=
λ`i `i
s i
i
s:
s:
i=1
i=1
i=1
i∈s0,t ,
(N +1)∈si,t ,
0∈s
i,r
January 9, 2018
i∈sN +1,r
DRAFT
18
which is equal to the objective function value of P4 in (28).
To prove the opposite direction (i.e., P4 → (27)), we show that we can map an optimal solution in P4 to a feasible
point (schedule) in (27) with a rate equal to the optimal value of P4. First, note that the LP P4 has 2N variables.
As a result, a corner point in P4, should have at least N constraints from (P4b), (P4c) and (P4d) satisfied with
equality (we already have N other equality constraints due to (P4a). We now prove an interesting property about
optimal corner points in P4 which facilitates our proof.
Property 1. For any optimal corner point {λ?`i , λ?ri } in P4, there exists an i0 ∈ [1 : N ] such that λ?`i0 + λ?ri0 = 1.
Proof. To prove this property, we are going to consider three cases depending on which constraints are satisfied
with equality at an optimal corner point.
1) Both conditions in (P4b) are not satisfied with equality: In this case, a corner point has at least N constraints
among (P4c) and (P4d) satisfied with equality. It is not difficult to see that, in order for the corner point to be
optimal, at least for one i0 we have λ?`i0 + λ?ri0 = 1, otherwise we have a non-optimal zero value for the objective
function.
2) Only one condition in (P4b) is not satisfied with equality: In this case, a corner point has at least N − 1
constraints among (P4c) and (P4d) satisfied with equality. Thus, there exists at most one i such that 0 < λ`i +λri <
1. Additionally, by adding the conditions in (P4b), we get that
1<
N
X
(λ`i + λri ) < 2.
(30)
i=1
Thus, there exists one i0 such that λ`i0 + λri0 = 1, otherwise, we cannot satisfy the lower bound in (30).
3) Both conditions in (P4b) are satisfied with equality: In this case, a corner point has at least N − 2 constraints
among (P4c) and (P4d) satisfied with equality. Thus, there exist at most two i such that 0 < λ`i + λri < 1.
Furthermore, adding the constraints in (P4b) implies that
N
X
(λ`i + λri ) = 2.
(31)
i=1
The two aforementioned observations imply that all N − 2 equalities cannot be from (P4d), otherwise we have
PN
that i=1 (λ`i + λri ) < 2, which contradicts (31). Thus, there exists a constraint in (P4c) that is satisfied with
equality, i.e., λ?`i0 + λ?ri0 = 1 for some i0 ∈ [1 : N ].
We now use Property 1 to show that, for any optimal point in P4, we can find a feasible schedule in (27) that
gives a rate equal to the objective function in P4. For an optimal point (λ?`1 , λ?r1 , . . . , λ?`N , λ?rN ), let i0 be the index
such that λ?`i0 + λ?ri0 = 1 (such an index exists thanks to Property 1). Thus, we have the following condition for
January 9, 2018
DRAFT
19
our optimal point
λ?`i0 + λ?ri0 = 1,
X
λ?`i ≤ 1 − λ?`i0 = λ?ri0 ,
(32)
i∈[1:N ]\{i0 }
X
λ?ri ≤ 1 − λ?ri0 = λ?`i0 .
i∈[1:N ]\{i0 }
Note that, any state s in the 1-2-1 Gaussian HD diamond network activates at most two links in the network:
a link between the source and the m-th relay and/or well as the link between the n-th relay and the destination.
For brevity, in our construction we will denote with sm,n the state that activates the link from the source to the
i-th relay in the diamond network as well as the links from the n-th relay to the destination (If either link is not
activated, the corresponding index is ∅). We also use λsm,n to denote the fraction of time during which this network
state is active. Using this notation, we can construct the following schedule from the given optimal point in P4
∀i ∈ [1 : N ]\{i0 },
λsi,i0 = λ?`i ,
λs∅,i0 = λ?ri0 −
X
λ?`i ,
i∈[1:N ]\{i0 }
∀i ∈ [1 : N ]\{i0 },
λsi0 ,i = λ?ri ,
λsi0 ,∅ = λ?`i0 −
X
(33)
λ?ri .
i∈[1:N ]\{i0 }
The activation time of all other states, except those described above, is set to zero.
From (32), we know that all values defined in (33) are positive. We can verify that the generated schedule is
feasible, i.e., the sum of all λ has to add up to one as follows
X
X
X
(a)
(b)
λs = λs0,i0 +
λsi,i0 +
λsi0 ,i + λsi0 ,0 = λ?`i0 + λ?ri0 = 1,
s
i∈[1:N ]\{i0 }
i∈[1:N ]\{i0 }
where: (i) the equality in (a) follows from the definitions in (33) and (ii) the equality in (b) follows from Property 1.
In order to conclude the mapping from P4 to (27), we need to verify that the rate achieved with the constructed
schedule in (33) is equal to the optimal value of the LP P4. From (27), we get that
N
X
X
X
min
λs `i ,
λs ri
s:
s:
i=1
i∈s0,t ,0∈si,r
(N +1)∈si,t ,i∈sN +1,r
X
X
X
X
X
=
min
λs `i ,
λs ri + min
λs `i ,
λs ri
s:
s:
s:
s:
i∈[1:N ]\{i0 }
i∈s0,t ,
i∈s0,t ,
(N +1)∈si,t ,
(N +1)∈si,t ,
0∈si,r
0∈si,r
i∈sN +1,r
i∈sN +1,r
n
o
X
X
X
=
min λsi,i0 `i , λsi0 ,i ri + min
λsi0 ,j `i0 ,
λsk,i0 ri0
i∈[1:N ]\{i0 }
j∈[0:N ]\{i0 }
k∈[0:N ]\{i0 }
X
X
?
min λ`i `i , λ?ri ri + min λ`i0 `i0 , λri0 ri0 =
min λ?`i `i , λ?ri ri .
(34)
=
i∈[1:N ]\{i0 }
January 9, 2018
i∈[1:N ]
DRAFT
20
Now note that, since the optimal corner point in P4 is feasible in P4, then λ?`i `i = λ?ri ri , ∀i ∈ [1 : N ]. Thus the
PN
expression in (34) can be rewritten as i=1 λ?`i `i , which is the optimal objective function value in P4. Thus, we
can now conclude that (27) is equivalent to P4.
We are now going to relate the LP P4 discussed above to the LP in Lemma 4. Recall that for a two hop
Half-Duplex path with link capacities `i and ri , the capacity is given by
Ci =
`i ri
.
`i + ri
Thus, we ca write the LP P1d as
P1d :
N
X
maximize
xi
i=1
`i ri
`i + ri
(P1a)d 0 ≤ xi ≤ 1
subject to
(P1b)d
N
X
xi
`i
≤ 1,
`i + ri
xi
ri
≤ 1.
`i + ri
i=1
d
(P1c)
N
X
∀i ∈ [1 : N ],
i=1
(35)
We are now going to show that P1 is equivalent to the LP P4 and, as a consequence, it is to the formulation of
Ccs,iid in (27). To do this, we are going to show how a feasible point in P4 can be transformed into a feasible point
in P1d and vice versa.
1) P4 → P1d . Define xi to be
xi = λ`i
`i + ri
,
ri
∀i ∈ [1 : N ].
(36)
Using this transformation, we get that the constraints in P4 imply the following
(P4b) :
1≥
N
X
λ `i =
i=1
i=1
(P4b) :
1≥
N
X
N
X
(P4c) :
∀i ∈ [1 : N ],
(P4d) :
∀i ∈ [1 : N ],
ri
`i + ri
N
X
=⇒ (P1c)d
N
X
`i
`i
xi
=
ri
`i + ri
i=1
i=1
`i
ri
`i
(P4a)
1 ≥ λ`i + λri = λ`i 1 +
= xi
1+
= xi
ri
`i + ri
ri
`i + ri
0 ≤ λ`i
= xi
ri
(P4a)
λ ri =
i=1
xi
λ`i
(P4 objective function) :
N
X
λ`i `i =
i=1
N
X
xi
i=1
=⇒ (P1b)d
=⇒ (P1a)d
=⇒ (P1a)d
ri
`i = (P1d objective function).
`i + ri
Thus for any feasible point in P4, we get a feasible point in P1d using the transformation in (36) that has the
same objective function with the same value as the original point in P4.
2) P1d → P4. Define λ`i and λri to be
λ `i = x i
January 9, 2018
ri
,
`i + ri
λri = xi
`i
,
`i + ri
∀i ∈ [1 : N ].
(37)
DRAFT
21
Note that the transformation above directly implies condition (P4a) in P4. Now, we are going to show that the
constraints in P1d when applied to (37) imply the rest of the constraints in P4 as follows
ri
`i
(P1a)d : 1 ≥ xi = xi
+
= λ`i + λri
=⇒ (P4c)
`i + ri
`i + ri
ri
(P1a)d : 0 ≤ xi
= λ`i
=⇒ (P4d)
ri + `i
N
N
X
X
`i
(P1b)d : 1 ≥
=
xi
λri
=⇒ (P4b)
`i + ri
i=1
i=1
(P1c)d :
1≥
N
X
i=1
N
xi
X
ri
=
λ`i
`i + ri
i=1
(P1d objective function) :
N
X
i=1
=⇒ (P4b)
N
xi
X
ri
`i =
λ`i `i = (P4 objective function).
`i + ri
i=1
d
Thus the two problems P1 and P4 are equivalent. This concludes the proof of Lemma 4 for the HD case.
B. Proof of Lemma 5(L2) for an HD diamond network
We first prove the following property of the optimal corner points in the LP P1d in (35).
Property 2. If we have a 1-2-1 Gaussian HD diamond network, then for any optimal corner point solution of P1d ,
at least one of the constraints in (P1b)d and (P1c)d is satisfied with equality.
Proof. We are going to prove Property 2 by contradiction. Note that, since the LP P1d has N variables, then any
corner point in P1d has at least N constraints satisfied with equality. Now, assume that we have an optimal point
(x?1 , x?2 , . . . , x?N ) such that neither (P1b)d nor (P1c)d is satisfied with equality. This implies that the constraints
satisfied with equality are only of the type (P1a)d . Thus, from the constraints in (P1a)d , we have that x?i ∈
PN
{0, 1}, ∀i ∈ [1 : N ]. Additionally, (P1b)d and (P1c)d being strict inequalities implies that i=1 x?i < 2. Thus,
there exists at most one i0 , such that x?i0 = 1, while x?j = 0, ∀j ∈ [1 : N ]\{i0 }.
Now, if we pick some k 6= i0 and set x?k = ε > 0 such that both (P1b)d and (P1c)d are still satisfied, then we
rk
, which contradicts the fact that (x?1 , x?2 , . . . , x?N ) is an optimal solution.
increase the objective function by ε ``kk+r
k
Now using Property 2, we are going to prove Lemma 5(L2) by considering the following two cases: (i) There
exists an optimal corner point for which only one of the constraints in (P1b)d and (P1c)d is satisfied with equality,
and (ii) all optimal corner points have both (P1b)d and (P1c)d satisfied with equality.
1) An optimal corner point exists with only one among (P1b)d and (P1c)d satisfied with equality. We
denote this optimal corner point as (x?1 , x?2 , . . . , x?N ). Since only one among (P1b)d and (P1c)d is satisfied
with equality, then at least N − 1 constraints of the type (P1a)d are satisfied with equality. Also note that,
PN
since only one among (P1b)d and (P1c)d is satisfied with equality, then this implies that i=1 xi < 2. This
implies that, although we have at least N − 1 constraints in (P1a)d satisfied with equality, we have at most
one i0 such that x?i0 = 1. As a result, at least N − 2 of the constraints satisfied with equality from (P1a)d are
January 9, 2018
DRAFT
22
of the form xi = 0. This proves that at least N − 2 relays are not utilized at this optimal corner point, which
proves Lemma 5(L2) in this case.
2) All optimal corner points have (P1b)d and (P1c)d satisfied with equality. Pick an optimal corner point
and denote it as (x?1 , x?2 , . . . , x?N ). Define Fx? = {i|0 < x?i < 1} and Ix? = {i|x?i = 1}, i.e., the sets of indices
of the variables with non-integer and unitary values, respectively. The fact that both (P1b)d and (P1c)d are
PN
satisfied with equality implies that i=1 x?i = 2, which implies that |Ix? | ≤ 2. Additionally, since we are
considering a corner point, then we have that at least N − 2 constraints of the type (P1a)d are satisfied with
equality. This implies that |Fx? | ≤ 2. Note that, if |Fx? | + |Ix? | ≤ 3 for all optimal corner points, then we
have proved Lemma 5(L2) for this case. Thus, we now show that the events {|Fx? | = 2} and {|Ix? | = 2} are
mutually exclusive (i.e., disprove the possibility that |Fx? | + |Ix? | = 4). This follows by observing the following
relation
2=
N
X
x?i =
i=1
X
x?i +
X
?
i∈Ix
?
i∈[1:N ]\Ix
X
x?i =
x?i + |Ix? |.
?
i∈[1:N ]\Ix
Thus
X
|Ix? | = 2 =⇒
x?i = 0 =⇒ |Fx? | = 0,
?
i∈[1:N ]\Ix
which proves that the two events are mutually exclusive. This concludes the proof of Lemma 5(L2).
A PPENDIX D
P ROOF OF L EMMA 5(L3)
The proof of Lemma 5(L3) for the FD case follows directly from 5(L1) by taking only the two paths (relays)
needed to achieve the Ccs,iid . Without loss generality, we assume that relays 1 and 2 are the relays in question.
Then we have using the optimal fractions x?1 and x?2 that
Ccs,iid = x?1 C1 + x?2 C2
(P1a)d
≤
C1 + C2 ,
which proves that either C1 or C2 are greater than or equal half Ccs,iid .
To prove Lemma 5(L3) for the HD case, note that for a HD network Ci in P1d is given by
Ci =
`i,0 `N +1,i
.
`i,0 + `N +1,i
Thus, by adding the constraints (P1b)d and (P1c)d , we have the following implication for any feasible point in
P1d
2≥
X
i∈[1:N ]
=
X
i∈[1:N ]
January 9, 2018
xi
X
Ci
Ci
+
xi
`i,0
`N +1,i
i∈[1:N ]
X
X
`N +1,i
`i,0
xi
+
xi
=
xi .
`i,0 + `N +1,i
`i,0 + `N +1,i
i∈[1:N ]
(38)
i∈[1:N ]
DRAFT
23
Now, assume without loss of generality that the path through relay 1 has the largest HD approximate capacity.
Then, for any optimal point x?i that solves P1d in the HD case, we have
X
X
Ccs,iid =
x?i Ci ≤
x?i C1 ≤ 2C1 .
i∈[1:N ]
i∈[1:N ]
This proves that the approximate capacity of the best path in the network is at least half the of Ccs,iid .
A PPENDIX E
E QUIVALENCE BETWEEN P3 AND P1
In this section, we prove the equivalence between the LPs P1 and P3 (which, as proved in Section IV is equivalent
to P2), hence concluding the proof of Theorem 2. In particular, our proof consists of two steps.
We first show that the LP in P3 is equivalent to the LP P5 below
P
P5 : Ccs,iid = max p∈P Fp
(P5a) Fp ≥ 0
(P5b)
(P5c)
(P5d)
(P5e)
∀p ∈ P,
Fp = λp`p.nx(i),i `p.nx(i),i
Fp = λp`i,p.pr(i) `i,p.pr(i)
P
p
p∈Pi λ`p.nx(i),i ≤ 1
P
p
p∈Pi λ`i,p.pr(i) ≤ 1
∀i ∈ p\{N + 1}, ∀p ∈ P,
(39)
∀i ∈ p\{0}, ∀p ∈ P,
∀i ∈ [0 : N ],
∀i ∈ [1 : N + 1],
and then show that P5 is equivalent to P1.
P3 → P5. For (i, j) ∈ p such that j = p.nx(i), define the variable λp`j,i to be
Fp
.
`j,i
λp`j,i =
(40)
Note that, the definition above automatically satisfies the constraints (P5a), (P5b) and (P5c) in P5. Then, by always
using the definition in (40), we can equivalently rewrite the constraint (P3b) as
X
(P3b) :
λp`j,i ≤ λ`j,i ,
∀(j, i) ∈ [1 : N + 1] × [0 : N ].
p∈P,
(i,j)∈p,
j=p.nx(i)
Now, if we fix î ∈ [0 : N ] and add the left-hand side and right-hand side of (P3b) for (j, i) ∈ [1 : N + 1] × {î},
then we get
∀î ∈ [0 : N ],
X
X
j∈[1:N +1]
p∈P,
(î,j)∈p,
j=p.nx(î)
λp`
j,î
=⇒
X
λp`
p.nx(î),î
p∈Pî
≤
X
λ`j,î
j∈[1:N +1]
≤
X
(P3c)
λ`j,î ≤ 1 =⇒ (P5d).
j∈[1:N +1]
Similarly, by adding the constraints in (P3b) for a fixed ĵ ∈ [1 : N +1], one can show that, under the transformation
in (40), the constraint in (P5e) is satisfied. Thus, for any feasible point in P3, we can get a feasible point in P5
using the transformation in (40). Regarding the objective function, note that we did not perform any transformation
January 9, 2018
DRAFT
24
on the variables Fp from P3 to P5. It therefore follows that the objective function value achieved in P3 is the same
as the one achieved in P5.
P5 → P3. Given a feasible point in P5, we define the following variables for each link in the network
X
λ`j,i =
λp`j,i .
p∈P,
(i,j)∈p,
j=p.nx(i)
Based on this transformation, we automatically have that (P3e) is satisfied. Moreover, we have that
(P5a) :
(P5d) :
∀p ∈ P,
0 ≤ Fp
X
X p
λ`j,i =
λ`p.nx(i),i =
1≥
∀i,
p∈P,
(i,j)∈p
p∈Pi
(P5e) :
∀i,
1≥
X
X
p∈P,
(i,j)∈p,
j=p.nx(i)
Fp
=
`j,i
X
λp`i,p.pr(i) =
X
λ`j,i
=⇒ (P3c)
j∈[1:N +1]\{i}
X
λ`i,j =
p∈P,
(j,i)∈p
p∈Pi
(P5b)&(P5c) :
=⇒ (P3a)
X
λ`i,j
=⇒ (P3d)
j∈[0:N ]\{i}
λp`j,i = λ`j,i
=⇒ (P3b).
p∈P,
(i,j)∈p,
j=p.nx(i)
Furthermore, note that the objective function in P5 and P3 is the same. Thus, a feasible point in P5 can be mapped
to a feasible point in P3 with the same objective function value. In conclusion, the problems P2, P3 and P5 are
equivalent. We now show that P5 is equivalent to P1 in Theorem 2.
P5 → P1. Define xp to be
xp =
Fp
,
Cp
∀p ∈ P.
(41)
Using this transformation, we get that the constraints in P4 imply the following
(P5a) :
(P5d) :
∀p ∈ P,
0 ≤ Fp = xp Cp
X p
(P5b) X
∀i ∈ [0 : N ], 1 ≥
λ`p.nx(i),i =
p∈Pi
(P5e) :
∀i ∈ [1 : N +1],
1≥
X
p∈Pi
(P5c)
λp`i,p.pr(i) =
p∈Pi
=⇒ (P1a)
Fp
`p.nx(i),i
Fp
X
p∈Pi
=
`i,p.pr(i)
X
p∈Pi
=
xp Cp
`p.nx(i),i
(9)
=
X
p
xp fp.nx(i),i
X xp Cp (9) X
p
=
xp fi,p.pr(i)
`i,p.pr(i)
p∈Pi
=⇒ (P1b)
p∈Pi
=⇒ (P1c).
p∈Pi
Moreover, we have that
(P5 objective function) :
X
p∈P
Fp =
X
xp Cp = (P1 objective function).
p∈P
Thus, for any feasible point in P5, we get a feasible point in P1 using the transformation in (41) that has the
objective function with the same value as the original point in P5.
P1 → P5. Define Fp , λp`p.nx(i),i and λp`i,p.pr(i) as
Fp = xp Cp ,
January 9, 2018
λp`p.nx(i),i =
xp Cp
∀i ∈ p\{N + 1},
`p.nx(i),i
λp`i,p.pr(i) =
xp Cp
∀i ∈ p\{0}
`i,p.pr(i)
(42)
DRAFT
25
that hold ∀p ∈ P. Note that the transformation above directly implies conditions (P5b) and (P5c) in P5. Now, we
are going to show that the constraints in P1 when applied to (42) imply the rest of the constraints in P5 as follows
(P1a) : ∀p ∈ P,
Fp
Cp
X
(9) X
p
=
1≥
xp fp.nx(i),i
0 ≤ xp =
(P1b) : ∀i ∈ [0 : N ]
1≥
xp Cp
`p.nx(i),i
p∈Pi
p∈Pi
(P1c) : ∀i ∈ [1 : N + 1]
=⇒ (P5a)
X
(9)
p
xp fi,p.pr(i)
=
p∈Pi
=
X
λp`p.nx(i),i
=⇒ (P5d)
p∈Pi
X xp Cp
X p
=
λ`i,p.pr(i)
`i,p.pr(i)
p∈Pi
=⇒ (P5e)
p∈Pi
Moreover, we have that
(P1 objective function) :
X
p∈P
xp Cp =
X
Fp = (P5 objective function).
p∈P
Thus, for any feasible point in P1, we get a feasible point in P5 using the transformation in (42) that has the objective
function with the same value as the original point in P1. Thus, the two problems P1 and P5 are equivalent. In
conclusion, the problems P1, P2, P3 and P5 are equivalent. This concludes the proof of Theorem 2.
R EFERENCES
[1] N. Alliance, “5g white paper,” Next generation mobile networks, white paper, 2015.
[2] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models
for future wireless communication system design,” IEEE Transactions on Communications, vol. 63, no. 9, pp. 3029–3056, 2015.
[3] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter wave channel modeling and cellular
capacity evaluation,” IEEE journal on selected areas in communications, vol. 32, no. 6, pp. 1164–1179, 2014.
[4] H. Shokri-Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, and M. Zorzi, “Millimeter wave cellular networks: A mac layer perspective,”
IEEE Transactions on Communications, vol. 63, no. 10, pp. 3437–3458, 2015.
[5] Y. Niu, Y. Li, D. Jin, L. Su, and A. V. Vasilakos, “A survey of millimeter wave communications (mmwave) for 5g: opportunities and
challenges,” Wireless Networks, vol. 21, no. 8, pp. 2657–2676, 2015.
[6] P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE Transactions on Information Theory, vol. 46, no. 2, pp. 388–404,
2000.
[7] S. Yi, Y. Pei, and S. Kalyanaraman, “On the capacity improvement of ad hoc wireless networks using directional antennas,” in Proceedings
of the 4th ACM international symposium on Mobile ad hoc networking & computing.
ACM, 2003, pp. 108–116.
[8] G. Kramer, “Models and theory for relay channels with receive constraints,” in 42nd Annual Allerton Conference on Communication,
Control, and Computing, Sept. 2004, pp. 1312–1321.
[9] A. Özgür and S. N. Diggavi, “Approximately achieving Gaussian relay network capacity with lattice-based QMF codes,” IEEE Transactions
on Information Theory, vol. 59, no. 12, pp. 8275–8294, December 2013.
[10] S. Lim, Y.-H. Kim, A. El Gamal, and S.-Y. Chung, “Noisy network coding,” IEEE Transactions on Information Theory, vol. 57, no. 5,
pp. 3132 –3152, May 2011.
[11] M. Cardone, D. Tuninetti, R. Knopp, and U. Salim, “Gaussian half-duplex relay networks: improved constant gap and connections with
the assignment problem,” IEEE Transactions on Information Theory, vol. 60, no. 6, pp. 3559 – 3575, June 2014.
January 9, 2018
DRAFT
| 7cs.IT
|
Ensuring Spreadsheet Integrity with Model Master
Jocelyn Paine,
Visiting Fellow, Institute of Learning and Research Technology, University of Bristol
http://www.ifs.org.uk/~popx/
ABSTRACT
We have developed the Model Master (MM) language for describing spreadsheets, and tools for converting MM programs
to and from spreadsheets. The MM decompiler translates a spreadsheet into an MM program which gives a concise
summary of its calculations, layout, and styling. This is valuable when trying to understand spreadsheets one has not seen
before, and when checking for errors. The MM compiler goes the other way, translating an MM program into a
spreadsheet. This makes possible a new style of development, in which spreadsheets are generated from textual
specifications. This can reduce error rates compared to working directly with the raw spreadsheet, and gives important
facilities for code reuse. MM programs also offer advantages over Excel files for the interchange of spreadsheets.
1 INTRODUCTION
Spreadsheets are alarmingly error-prone to write. To quote [Panko 2000], “given data from recent field audits,
most large spreadsheets probably do contain significant errors”. The most recent audits he cites found errors in
at least 86% of spreadsheets audited. In a 1997 feature entitled “Fatal Addition” [Ward 1997], New Scientist
reported that 90% of the spreadsheets audited in a study carried out by Coopers and Lybrand were found to
have errors. Given the billions of spreadsheets in use, this leaves the worlds of business and finance horribly
vulnerable to programming mistakes. Studies show the chances of any given spreadsheet cell containing an
error are somewhere between 0.3% and 3%, so that a spreadsheet of only 100 cells has about a 30% chance of
having one error or more.
Creating the spreadsheet is not the only problem. Spreadsheets can be difficult to read, and hence to debug and
maintain. Look at the example below - a small professionally-written educational model for calculating income
elasticities, shown with formulae visible – and make your best attempt to read it:
Firstly, the use of cell locations – C9, C10 - in formulae makes them hard to understand, because the location’s
name conveys no information about its purpose. This presumably increases the reader’s cognitive load, because
of the time needed to match locations with their targets, and the need to hold information in short-term memory
while doing so. Excel does provide facilities for naming cells, but many developers don’t use these, so trying to
understand cell locations is a common readability problem.
Secondly, even in such a small spreadsheet, an annoying amount of horizontal scrolling is needed before one
can see all the formulae. This is partly because the spreadsheet contains not only the formulae but also a lot of
text and empty cells used for layout. User-interface design experts agree that horizontal scrolling should be
avoided wherever possible. In a larger spreadsheet, the reader would have to search a considerable amount of
screen area, both to find all the formulae, and to match the cell locations with their targets. It seems reasonable
to assume that if the reader cannot see the entire spreadsheet in one go, he or she will have to remember parts of
it, further adding to cognitive load. As one developer said: “I end up with vast numbers of notes on little bits of
paper”. If the spreadsheet has more than one worksheet, flicking between them is an extra inconvenience.
A third difficulty, illustrated by the “Lazy Days” example below [Rajalingham et al 2000], is repeated
formulae, due to a calculation being replicated over several time periods, departments, or other entities. This
example contains identical calculations of total wage and average wage for each type of personnel:
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
Lazy Days Staff Budget Costs 1995-1996
Staff
Basic
Overtime Total
Average
Numbers Wages £ Wages £ Wages £ Wages £
Managers
Grade 1
Grade 2
Grade 3
Grand Totals
1
3
9
12
25
17700
45540
122340
102350
287930
0
1400
2000
0
3400
17700
46940
124340
102350
291330
17700.00
15646.67
13815.56
8529.17
11653.20
There is much repeated structure here, but the fact that it is repeated is not explicit. The columnar layout and the
nature of the application make this likely, but the reader needs to examine and compare all the formulae before
being sure.
To overcome such problems, we are developing a tool, named Model Master or MM, that “decompiles”
spreadsheets, generating concise specifications of their calculations. The example below shows a specification
generated by the decompiler for the elasticity model:
attributes <
new_quantity
old_quantity
new_real_income
old_real_income
demand_change
real_income_change
income_elasticity
good_type
>
where
demand_change = new_quantity / old_quantity – 1
real_income_change =
new_real_income / old_real_income - 1 and
income_elasticity =
demand_change / real_income_change and
good_type =
if( income_elasticity > 0,
"So, this product is a normal good.",
"So, this product is an inferior good."
)
and
This lists the variables or “attributes” represented by the cells, and the equations relating them. It gives an
alternative view of the spreadsheet, which we believe to be valuable when checking for errors and when trying
to understand spreadsheets written by other people.
The significant point about these listings is that they can be regarded as programs. Not only can we decompile
existing spreadsheets into this form; we could go in the other direction and compile such programs into
spreadsheets. We have in fact devised a complete programming language for describing spreadsheets, which
we also call MM, and a compiler that translates MM programs into spreadsheets.
What are the benefits of generating spreadsheets from MM programs, rather than writing directly in Excel? One
is that typing errors and other mistakes in an MM program are more likely to be detected than errors in the
spreadsheet itself. This is because spreadsheets have little redundancy: if the programmer mistypes a cell
location, or through a slip of the cursor accidentally types a formula into the wrong cell, the resulting code still
means something to the spreadsheet. On the other hand, when programming MM, one works with names such
as balance and start_time, rather than with cell locations. The programmer must give a list of all such
names – to “declare” them – when writing the program. So if a typing error is made, blance for example,
MM can check the name against this list, realising that since it doesn’t occur there, there was an error.
As the example above demonstrates, MM programs consist of almost nothing other than lists of variables
together with the equations or formulae relating them. So if the variable names are well chosen –
total_wage rather than tot_w, hours_per_week and not hrpwk – the programs ought to be
comprehensible to non-programmers, so long as they understand the business processes being modelled.
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
Although in some organisations, spreadsheets are regarded as private to their author and not to be checked by
anyone else, others may find this useful in managing development.
Another benefit of MM is code reuse. MM programs can be divided up into separate modules. These can be put
into program libraries and included in a program just by mentioning the name of the file where they are stored.
Authors can therefore build up a central repository of modules, reducing duplicated work and the risk of cutand-paste errors. We regard this as extremely important, and emphasise that it is not a trivial claim. This feature
of MM stems from mathematical techniques used in designing the OBJ family of algebraic specification
languages [Goguen and Tracz 2000], and provides the simplest and most comprehensive module system that
we know of.
MM makes possible a style of spreadsheet development similar to that practised with conventional languages
such as Fortran and Java. In this, the MM program is primary, and the spreadsheet is merely a means of running
it and displaying the results: something to be generated whenever one needs to do a calculation, but otherwise
to be ignored. This is extremely different from the way in which most Excel programmers work, and we
suspect few would be willing to give up the convenience of interactive model construction and testing possible
when you work directly with the spreadsheet. However, because of MM’s benefits, particularly with respect to
code reuse, we believe it important to find out how it can best be used to complement and increase the integrity
of this kind of development. We are therefore continuing work on the compiler and decompiler, and on an
integrated user interface which allows them to be invoked directly from Excel.
We also believe that MM programs, whether written from scratch or decompiled out of an existing spreadsheet,
are a better medium for interchanging spreadsheets than are Excel XLS files. We are investigating MM for this
purpose.
In the rest of this paper, we describe the MM language, compiler, and decompiler from the user’s point of view,
and discuss the benefits of MM as a medium for spreadsheet interchange. We then give some details of the
implementation, and finally summarise our progress and the work yet to be done. We start with an account of
the language: without this, it is difficult to understand the compiler and decompiler. The main point of the paper
is to present MM in enough detail that we may receive worthwhile criticisms and suggestions for improvement.
2 THE MM LANGUAGE AND COMPILER
2.1 Attributes, Objects, and Equations
MM views the world in terms of attributes which have values, usually numbers or strings, and which are interrelated by equations. Attributes may be single-valued, or may range over regions of time or space, with a value
for each point in the region. It is sometimes convenient to think of them as grouped together to make objects.
The simple example below is an MM description of a company:
attributes <
incomings [ 1995:2004 ]
outgoings [ 1995:2004 ]
profit
[ 1995:2004 ]
>
where
profit[ all t ] = incomings[ t ] – outgoings[ t ]
(We do realise that as a business application, this program is trivial. This is not a case of the “toy language”
syndrome, whose sufferers design languages that won’t scale up to realistic applications; we are using MM to
build real models. But we do need to simplify so that we don’t obscure key concepts.)
To generate a spreadsheet, we pass the program text to the MM compiler, which checks for errors, and if none
are found, produces an Excel file. Were we writing the spreadsheet directly, we would almost certainly use
successive rows to represent successive years, allocating attributes to adjacent columns. This is what the
compiler does, heading each column with the identifier used for its attribute. Appendix 1 shows a generated
spreadsheet in the queuing simulation example.
2.2 Layout
MM separates the appearance of a spreadsheet from its calculations. All MM programs are written in two parts:
the variables and equations, and a separate section stipulating how these are to be arranged and styled. It is this
section that deals with matters such as cell formats, background colours, the allocation of variables to cells, and
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
whether a variable is to run down a column or along a row. This means that the results of calculations can be
displayed in a variety of ways without changing the program. An interesting example of this is discussed in the
section below on Multidimensional Tables.
We based the layout language on HTML, the language in which Web pages are written, largely because this is
now very well-developed for expressing layout. And of course, many people have written Web pages and will
have some knowledge of it. For those who do not, we shall explain the basics. For the purposes of this paper,
the important thing about it is that an HTML document consists of plain text interspersed with tags that specify
the document’s appearance. For example, <H1>The MM user manual</H1> creates a main heading;
<P>…</P> enclose a paragraph; and the tags <TABLE>…</TABLE>, <TR>…</TR> and
<TD>…</TD> are used to display tables, where <TABLE>…</TABLE> enclose the entire table,
<TR>…</TR> enclose a row, and <TD>…</TD> enclose cells within the row. Our layout language is
similar, augmented with special tags that insert attributes into the tables. Two examples are shown in the section
on the decompiler.
2.3 Code Reuse; Named Constants
Had we stored the program above in a file called company, we could reuse it in another program:
include “company”
company2 =
company
where
incomings[ 1995 ] = 1000
and
incomings[ all t > 1995 ] = incomings[ t-1 ] * 1.2
Here, we have extended company by adding an initial value for the 1995 value of incomings, and a growth
law for the other values. This example illustrates code reuse. It also shows how one equation can be applied to a
number of points within an attribute’s domain, as with the formula for profit earlier.
Instead, we could add some new attributes:
include “company”
constant average_wage = 500
company3 =
company plus attributes <
workforce [1995:2004]
>
where
outgoings[ all t ] = workforce[ t ] * average_wage
In this program, we have given company the new attribute workforce representing the number of
employees. We calculate outgoings from it (again, in a terribly naïve way). This also shows that we can
define constants to make code more readable.
In models that simulate some process that spans many regions of time or space, with the same thing happening
in each region, constants often enable us to change the number of regions by altering just one number. The
queuing simulation in Appendix A has two examples. To achieve the same effect on the raw spreadsheet would
require a lot of error-prone copying and editing of cell references.
2.4 Parameterisation: Defining Templates
In the above, the company’s life span is hard-wired into the code. To change it, we would need to do a global
find-and-replace of the years 1995 and 2004. We can avoid this by using parameters to our object definitions.
Below, our definition of company has become company_template and acquired two parameters, T1 and
T2. These specify the time over which the attributes range. We then reconstruct company from this template.
The great thing about this is that we can construct general templates and then use them in a whole variety of
applications just by providing suitable parameters.
company_template( T1:integer, T2:integer ) = attributes <
incomings [ T1:T2 ]
outgoings [ T1:T2 ]
profit
[ T1:T2 ]
>
where
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
profit[ all t ] = incomings[ t ] – outgoings[ t ]
company = company_template( 1995, 2004 )
2.5 Function Definitions
The equations in MM programs can call built-in Excel functions, as shown by the queuing simulation, which
calls min and rand amongst others. We considered that it might be useful for users to be able to define their
own functions in terms of these built-ins. We implemented this in a prototype, but have omitted it from the
current version because of difficulties over its interaction with templates.
2.6 Ranges as Arguments; Slices
Many of Excel’s built-in functions accept cell-range arguments. For example, min(C4:D10) finds the
smallest number anywhere within the rectangle whose top-left corner is C4 and whose bottom right is D10. We
clearly need an equivalent in MM, which we provide by using the range operator. So min( range
incomings ) would find the smallest number anywhere in the incomings column. Using the word
range before the attribute name just adds a bit of redundancy to confirm that the user did indeed intend a
range and not just one value.
Sometimes, we need to pass only some of an attribute’s values. This is done by “slicing” – using subscripts to
select a rectangular subregion. For example, min( range incomings[1995:2000] ). The queuing
simulation does this. It sets up a series of servers (representing shop assistants or similar), with one column for
each. Each row represents one customer; the entry for the customer in row R and the server in column S
specifies when S could start processing R. For each customer, the simulation needs to search for the server with
the earliest start time. It does this by slicing out the row, and then passing all the server-columns in it to min.
This explains what is happening with potential_start_time in the example.
2.7 Ranging over Kinds of Employee: a More Realistic Example
From the examples above, you might gain the impression that MM is useful only for applications where all the
attributes range over time. This is not so, as demonstrated by the “Lazy Days” spreadsheet shown in the
Introduction. Here is the MM program that would produce it (without the data, which we assume to be typed
directly into the spreadsheet):
base employee_kind =
{ “Managers”, “Grade 1”, “Grade 2”, “Grade 3”,
“Grand Totals” }
attributes <
staff_numbers [ employee_kind
basic_wages
[ employee_kind
overtime_wages [ employee_kind
total_wages
[ employee_kind
average_wage
[ employee_kind ]
]
]
]
]
>
where
total_wages[ all e ] =
basic_wages[ e ] + overtime_wages[ e ]
average_wage[ all e ] =
and
total_wages[ e ] / staff_numbers[ e ]
What we are doing here is to define a domain called employee_kind, and to say that our attributes all have
one value for each element of this domain. (We call this domain a “base”, a word taken from the mathematics
underlying MM.) This is analogous to what we did when making attributes range over the years 1995 to 2004,
except that we have replaced the year numbers by symbolic values like “Grade 1”. To save space, we have not
shown the layout definitions that would produce the above spreadsheet, but essentially, they would consist of
<table>, <tr> and <td> tags used so as to position the attributes and the headers Lazy Days Staff
Budget Costs 1995-1996, Average, and so on as they appear in the spreadsheet grid.
2.8 Multidimensional Tables
There is no reason for attributes to range over only two dimensions. Had we holographic 3-d computer displays
rather than flat monitors, we could imagine a cubical spreadsheet which tabulates the yearly costs due to a
company’s departments: years running down columns, cost categories (overheads, lighting, transport) running
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
from left to right across the front of the display, and department running from front to back, with a sheet for I.T.
at the front, one for Personnel behind it, and a third for Sales behind that. Such attributes are easy to declare:
base lifespan = [ 1995:2004 ]
base cost_type = { “Overheads”, “Lighting”, “Transport” }
base department = { “I.T.”, “Personnel”, “Sales” }
attribute <
costs : lifespan * cost_type * department
>
Even without a cubical display, we can display such a table in a variety of ways. We might have a number of
tabbed panels, with a worksheet on each panel and one panel per year, department running down and cost type
running across. We could swap cost type and department. Or we could put all the information on one
worksheet, stacking successive years one under the other. And so on.
In such situations, to separate presentation from layout is particularly valuable. Excel enables this to be done to
a limited extent by using pivot tables. However, these are an ad-hoc trick. We consider that using a separate
layout section is the correct way to go, and we are working on notations to make it easy to express how
multidimensional attributes are to be displayed. (Technically, we base these on the fact that the mapping from
multidimensional attribute base to two-dimensional worksheet surface is a linear transformation. This makes
storage allocation uniform, because the storage allocator can represent all such mappings as matrices.)
There are some spreadsheets, for example Storeys [ProFunda page], that are better than Excel for displaying
multidimensional tables. MM would be particularly valuable as a front-end to these.
2.9 Units and Dimensional Analysis
There is an Oxford story, recounted in [Morris 1978], that when John Keble was looking after the books of
Oriel College, his accounts showed an inexplicable deficiency of between £1800 and £1900. It was eventually
discovered that he had added the date to the college liabilities. We have spreadsheets now … and they do
nothing to prevent such errors. However, a paper on the programming language Algol68 [Cleaveland 1975]
indicates a cure. Cleaveland suggests that it should be possible to declare variables to have units such as meters,
seconds or pounds sterling. Knowing the properties of the arithmetic operators, a compiler could then check
that expressions were dimensionally correct. If we declared A to be seconds and B to be meters, adding A to B
would be invalid (adding length to time makes no sense), but multiplying them would be permitted. Hopefully,
this would enable a large number of errors to be detected. The mechanism would easily extend to financial
programs, those used by Oriel College included.
We have experimented with this in MM, by augmenting the programming language in two ways. Firstly, the
user was allowed to declare names for units, such as cm, lb, £K and $. Secondly, he or she could attach these
to attributes, telling the compiler for example that costs has units £K. Appendix 2 shows an example of the
compiler checking unit errors.
Although the feature is useful, we have left it out of the current version of MM. One reason is that there are
technical problems about the way it interacts with templates. Another is that it complicates function definition,
because a function that operates on quantities with units must specify what units each argument can have, and
the result’s units in terms of these. For example, were we to define square(N) as N*N, we would need to
specify that if N has units cm, then the result has units cm*cm, and so on. Yet another problem is that consistent
use of units is perhaps too verbose. For example, if we declare a constant length: cm = 1 cm we
have told MM that it has units cm. Strictly speaking, if we do not give the quantity assigned to it – 1 – a unit
too, we have committed an error, since a dimensionless number is not the same as one with a dimension. There
is a tension here between conciseness and correctness which we have not yet resolved.
3 THE MM DECOMPILER
3.1 Program Transformations
Going from spreadsheets to MM is harder than going in the other direction, because there are many different
MM programs that could generate any given spreadsheet. For example, consider a trivial spreadsheet with only
two occupied cells:
Value
1
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
This can always be represented as an MM program with two single-valued attributes. But it could also have
originated from a program with one attribute ranging over two points, or from a program with no attributes and
two headings, or from a program with one attribute and one heading. We compare two of these possibilities
below:
attributes < a >
attributes < s a >
where a=1
where s=”Value”
layout
and
a=1
<table>
layout
<tr>
<table>
<td>Value</td>
<tr>
<td><attr name=”a”/></td>
<td><attr name=”s”/></td>
</tr>
<td><attr name=”a”/></td>
</table>
</tr>
</table>
Although the decompiler can make some attempt to decide what the user intended – a column of numbers with
text above is probably one multi-valued attribute plus a static heading stating its name – it will in general need
to ask the user for more information.
The crucial insight here was that any spreadsheet can be trivially rewritten as an MM program, provided that
MM allows any attribute to be arranged anywhere in a spreadsheet, at any position relative to another attribute.
(If it didn’t, there would be some arrangements not expressible in MM). Once we have this program, we can
transform it in various ways, using hints provided by the user or gleaned by examining the spreadsheet.
As an example, consider the Staff Numbers column from the “Lazy Days” spreadsheet shown in the
Introduction. The trivial equivalent of this in MM (naming attributes after their cells) is this:
attributes < b2 b3 b5 b6 b7 b8 b9 >
where
b2=“Staff” and b3=”Numbers” and
b5=1 and b6=3 and b7=9 and b8=12 and b9=25
With a hint from the user, or by noticing that the two name cells contain text and are unrelated to any other
cells, we can assume them to be headings, and remove them from the list of attributes, moving them into the
layout section:
attributes < b5 b6 b7 b8 b9 >
where
b5=1 and b6=3 and b7=9 and b8=12 and b9=25
The remaining attributes are all numbers. With another hint from the user, or by assuming that a column with
text above it corresponds to one attribute, we could replace b5 to b9 by one multi-valued attribute:
attributes < b5b9[1:5] >
where
b5b9[1]=1 and b5b9[2]=3 and b5b9[3]=9 and b5b9[4]=12 and
b5b9[5]=25
We call this “rebasing”.
With yet another user hint, or by using the heuristic that text above a columnar attribute is most likely its name,
we could replace b5b9 by a meaningful name:
attributes < StaffNumbers[1:5] >
where
StaffNumbers[1]=1 and StaffNumbers[2]=3 and
StaffNumbers[3]=9 and StaffNumbers[4]=12 and
StaffNumbers[5]=25
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
We can do the same with the other columns. Of course, when an attribute is renamed or rebased, we have to
transform any equations that refer to it. This may entail replacing names in the equation, but may also require
changing or adding subscripts. For example, any reference to b9 would have to be replaced by b5b9[5]
and then by StaffNumbers[5].
3.2 Pattern-wizards and rolled equations
In the section on compiler implementation, we talk about “unrolling” equations. This is what the compiler does
when it takes an equation like total_wages[ all e ] = basic_wages[ e ] +
overtime_wages[ e ] that applies to more than one point in an attribute and generates one instance of
the equation for each of these points, i.e. one for each cell. To produce concise readable programs, the
decompiler has to do the reverse: “rolling”.
It would be ideal if the decompiler could do this automatically. However, recognising when to is not easy, and
for the moment, the user will need to specify rolling and the range over which to roll as another program
transformation. Chris Browne, an expert on spreadsheets and their history, has suggested [Chris Browne’s
Linux spreadsheet page] that we could start collecting common patterns of repetition and incorporating them
into “wizards” that search for instances and offer to apply the most appropriate transformation. He has also
suggested that the decompiler be used in the same way as people use algebra manipulation programs: we can
provide a large repertoire of transformations, but the user may need to apply a lot of trial and error before
finding a “best” MM program for a given spreadsheet.
Isakowitz, Schocken and Lucas [Isakowitz et al 1995] describe a decompiler which has a fair amount of
success at rolling equations: we have the impression that the spreadsheets they tried had a very repetitive form
generated by drag-and-drop copies. We have been told that these occur very frequently, due to the need to
replicate calculations across time periods, so when rolling equations, it probably is worth looking for such
patterns.
3.3 Error-checking
Once the decompiler has generated its equations, we can inspect them for errors, such as cells that should
contain formulae but which have actually been hard-coded (section 5 of [Rajalingham et al 2000] under “Hard
coding”). This will, we believe, be the decompiler’s most important use as an analysis tool.
Because of the tasks the decompiler has to perform, however, it can also display other information. It has to
calculate dependencies between cells, and to discover which cells are used but not initialised, or initialised but
not used. Once we have this information, displaying it – either inside Excel by changing cell colours, or in a
separate analysis window - is not difficult.
The equations and dependency diagrams are also useful to people who will run but never modify a spreadsheet.
In this connection, we are using some routines from the decompiler to extract programs from Excel and
implement them for the Web, in a project for the Institute of Learning and Research Technology (ILRT) at the
University of Bristol. This involves putting a series of economic models onto the Web as an educational
resource – part of the Biz/Ed Virtual Learning Arcade - for students of business and economics [Biz/Ed page].
4 MM AS A STANDARD FOR SPREADSHEET INTERCHANGE
We are investigating MM as a medium for the interchange of spreadsheets. At present, XLS files are the most
common means of transferring but we believe that using MM programs offers significant benefits. To transfer a
spreadsheet developed in MM, one would just send the MM program. To transfer one that was not, one would
decompile. As we have already mentioned in the Decompiler section, if one is prepared to help the decompiler
by suggesting appropriate program transformations, the resulting program should be easier to understand than
its parent spreadsheet.
These are the advantages we believe MM files have for interchange:
•
Unlike XLS files, MM programs are text files, which are easy to process with general-purpose tools such
as editors. This facilitates debugging when writing programs that process them.
•
XLS is a complicated format, and descriptions of it are fairly hard to find, whereas the MM language has a
simple syntax which we would be happy to publicise. When writing the MM compiler, we took care to
write the parser as a separate module, with a well-defined API and clear documentation describing the data
structures produced. We could also make this generally available.
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
•
One may sometimes want to process the spreadsheet formulae in other ways than evaluating them: for
example, to produce an annotated cross-reference listing for reference documentation. This is easier to do
with MM programs than with XLS files, partly because of the complexity of the latter, and partly because
MM programs make more of the structure apparent, for example the replication of calculations over
different time periods.
•
MM completely separates the program from its appearance. The latter can be changed simply by
modifying the layout section.
5 IMPLEMENTATION
5.1 The Compiler
The compiler uses standard methods from compiler technology, but differs from other compilers in that it
allocates storage not for a one-dimensional machine memory, but for the two-dimensional, and visible, medium
of a spreadsheet. It has 7 passes:
1. Input: reads and parses the input and builds the syntax tree. We use the freeware JavaCC [JavaCC page]
parser-generator, which translates a grammar into a Java parser.
2. Semantic checking: analyses the syntax tree for errors such as undeclared identifiers and objects defined in
terms of themselves. One theme that occurs here and in the decompiler and Web-based spreadsheet
evaluator is the need to discover dependencies between attributes. There are standard directed-acyclicgraph algorithms for this.
3. Identifier substitution: replaces all identifiers by their definitions.
4. Colimit: performs an algebraic operation analogous to multiplying together all the object definitions in the
tree, resulting in one big object definition for the entire program.
5. Storage allocation: allocates locations for all attributes, taking them from the layout section if there is one.
This generates a map associating each attribute with a cell address.
6. Code generation: “unrolls” equations, for example replacing profit[ all t ] = incomings[ t
] – outgoings[ t ] by one equation for each value of t. Then it replaces attributes by their cell
addresses. This gives us a “cellmap”: a map associating each used cell with the formula it contains.
7. Output: writes out the formulae together with instructions for placing them into their cells.
The compiler is written in Java, for portability. Why worry about this, when Excel only runs under Windows?
Firstly, for our Web work, we want to generate code for our own Web-based spreadsheet engine, which may
not be running on a Windows server. That will not concern many users, but another point is that Excel is not the
only spreadsheet, and we do not want to restrict the range of machines on which MM can be used. Unix, for
example, has the public-domain Oleo, Gnumeric and Dismal spreadsheets amongst others (a list can be found
at [Spreadsheet FAQ page]), and MM would be as useful with these as with Excel. We have also mentioned its
potential for use with ProFunda.
Because of portability, we are committed to Java, but the language has undoubted disadvantages for writing
compilers. It is verbose; it is object-oriented to an extent which just does not suit compiler-writing; and there are
certain idioms very useful here which Java has trouble with, such as higher order functions. For these reasons,
we have written large parts of MM in Kawa [Kawa page], a Java implementation of the functional
programming language Scheme. Kawa programs can be linked with Java, and so do not compromise
portability.
Mathematical Background
MM was developed from a branch of mathematics known as category theory. Here is not the place to discuss
details. Briefly stated however, category theory, like logic, is a tool for studying mathematical and
computational concepts, one concerned much more with their form than with their content. It comes equipped
with very general equivalents of common notions such as sum and product which can be applied to a vast and
varied range of different situations. The computer scientist Joseph Goguen [Goguen 1975, Goguen 1992] has
put it to use, together with another branch of mathematics known as sheaf semantics, in answering the questions
“what is an object?”, and “what does it mean to say a system is composed out of objects?”. It turns out that his
formulation can be implemented as a new style of programming, which we call System Limit Programming –
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
“limit” being the mathematical operation used to assemble components into a system. This gave us the
attribute-and-base notation for MM, and – together with the ideas described in [Goguen and Tracz 2000] – the
module structure.
5.2 The Decompiler
The decompiler has four passes:
1. Input: reads formulae from a spreadsheet file, again parsing it with JavaCC. Converts these into the same
internal representation as that generated by pass 6 of the compiler, and then builds a cell-map.
2. Dependency analysis: builds a graph of cell dependencies, enabling input-only and output-only cells to be
identified. An optional phase here checks for text cells with no dependants or dependees and assumes them
to be headers and other static text, removing them from the cellmap.
3. Program transformation: reads and obeys transformation commands, rewriting attribute lists and equations
as directed.
4. Output: pretty-prints the result as an MM program.
It is interesting to compare the decompiler with the factoring algorithm described by Isakowitz, Schocken and
Lucas [Isakowitz et al 1995]. We developed our decompiler independently, but there are similarities. Like us,
the authors recommend separating layout, which they call “editorial” information, from program, and their
algorithm does so. It works in two stages, first allowing the user to split the spreadsheet into regions
corresponding to separate attribute groups (we would call these objects: Isakowitz et al refer to them as
relations). We do not do this yet, regarding the spreadsheet as “flat”, with all attributes in the same object, but it
probably is useful. We are considering identifying blocks automatically, by using a depth-first regionnumbering algorithm to find all disconnected regions of the spreadsheet.
Their second stage removes the static text, and then converts each relation into an MM-like language which the
authors call FRL. Unlike us, the authors see this just as a convenient internal representation, and not as
something to be understood by the user. FRL does not have explicit bases: in our terms, all attributes are based
from 1 upwards.
6 PROGRESS AND FURTHER WORK
We have built several prototypes of the MM compiler, testing language features such as dimensional analysis,
and are now completing a version which uses the same internal representations as, and will be integrated with,
the decompiler. For the latter, some work still has to be done on allowing the user to describe program
transformations conveniently, and on heuristics for identifying attribute names and automatically acquiring
other useful information from the spreadsheet.
At present, both the compiler and decompiler are command-line-driven programs which communicate with
Excel via SYLK files [SYLK page]. This is convenient for development, because SYLK is a textual
representation, much easier to read and to edit than XLS. However, Microsoft have not bothered to maintain it
– one source states that it has not been revised since Excel 2.0 – and it can not represent all the style information
used in later versions of Excel. Nor can it handle multiple worksheets. To overcome this, we intend to connect
MM directly to Excel so that it can be run as an Excel add-on, taking information from, and passing it to, Excel
cells and cell ranges. This will also make it possible to provide the user with an integrated development
environment for MM, which we regard as very important.
The largest model on which we have tested the compiler is the queuing simulation below; the decompiler has
so far been run only on very small spreadsheets, of similar size to the examples in the introduction. We shall be
trying it out on more spreadsheets of this kind when putting them onto the Web, as we describe under the
decompiler section on Error Checking. This will be the only usability evaluation under the current funding,
though we hope to set up a joint project with Clemson University to do more formal evaluation.
For the longer term, we note as we did in the introduction that MM enables two kinds of spreadsheet
development. One is the conventional kind where one works directly with the spreadsheet. Here, MM is useful
just as an analysis and error-detection tool, through its decompiler. The other is similar to that practised with
conventional languages such as Fortran and Java. In this, the MM program is primary, and the spreadsheet is
merely a means of running it and displaying the results: something to be generated whenever one needs to do a
calculation, but otherwise to be ignored. Using MM in this way, one can take full advantage of the facilities for
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
code reuse, and can easily restructure spreadsheets as described under the section on Multidimensional Tables.
We need to find out how best to combine the two kinds of development so as to provide the convenience of
interactive inputting and testing with the readability, integrity and maintainability of MM programs.
7 ACKNOWLEDGEMENTS
I have done most of MM’s design and implementation in my own time. However, funding for putting the
models onto the Web, and for the latest work on the compiler and decompiler was provided via the ILRT by the
Joint Information Systems Committee 5/99 call for teaching and learning resources. I wish to thank Jack
Ponton of the Department of Chemical Engineering at the University of Edinburgh for telling me about SYLK
files, Margaret van Biene-Hershey for a valuable discussion on spreadsheet integrity as well as a delightful visit
to Harmelen, Rob Kemmeren, and the anonymous referees for their comments on the draft version of this
paper.
APPENDIX 1: QUEUING SIMULATION IN MM
The program below is an MM version of a queuing simulation. We designed it by reverse-engineering a
spreadsheet written by Thomas Grossman, a professor of management science at Calgary University. He has
written a variety of simulations in order to help students understand queuing behaviour [Grossman 1999]. Ours
is derived from his basic four-server queue, downloadable as 4vanilla.xls from the Web page cited with
this paper. Because the original spreadsheet, while interesting, is relatively complex and difficult to understand,
we thought it was a good test for MM.
The program demonstrates several features of MM.
•
Comments for program documentation.
•
Attributes ranging over more than one dimension. The attribute potential_start_time has one
dimension representing time: this runs from row to row, as with the other attributes. Its other dimension
represents servers, and runs along the columns: there is one column for each server.
•
Constant declarations make changing the spreadsheet structure easy. potential_start_time is
declared as ranging over event*[1:N], where event is itself declared as [1:10]. In other words,
potential_start_time ranges over the two-dimensional region [1:10,1:4]. The N referred to
here is declared as a constant equal to 4. Simply changing this to another value would automatically change
the number of columns allocated to the servers. Consider how little effort this is compared to modifying the
spreadsheet itself!
•
The use of Excel’s built-in functions if, match, min and rand.
•
Cell ranges in function arguments. In the function call min( range potential_start_time[e]
), e is time: it selects one of the rows. potential_start_time has one column for each server,
so the expression range potential_start_time[e] causes a cell range containing all these
columns within the row to be passed to min.
•
Commands for setting cell formats, and for placing headings in columns. We have now replaced these by
the HTML-based layout language. Here is the program:
/* Queue.mm */
This is a process-driven simulation of queuing. It has two main components: a
set of customers, and a set of servers. The number of customers and servers is
fixed before starting. Customers enter a shop, go to a server, queue, get
served, and leave. Each customer interacts once only and then leaves.
We could slice up time by allocating equally-spaced time points to successive
rows. Instead, following Grossman, each row represents one customer's complete
sequence of transactions: arriving, starting to queue, being served, and
leaving. We handle this in MM by declaring the relevant attributes to run over a
base (domain of observations) 1:10 similar, where the e'th point represents the
e'th customer. These attributes are, for example, customer_number, arrival_time,
service_start_time and service_time, all quantified over 'event'.
Again following Grossman, we allocate one column to represent each server. The
only attribute of a server that we use directly is the potential start time that
the customers would be served by it. So the columns hold these times. In MM, we
do this by declaring the attribute potential_start_time's base to have a second
dimension, running from 1 to N, the number of servers.
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
By default, the MM compiler puts the name of each attribute above its column.
You can change this with the 'name' qualifier. The 'format' qualifier changes
the format of the cells for the attribute values.
*/
base event = [1:10];
// Each point represents one customer.
constant N = 4;
// The number of servers.
< customer_number : event
name "Customer" br "#"
// The e'th element is the number of the e'th customer.
interarrival_time : event
name "Interarrival" br "duration"
// The e'th element is the time between the arrivals of
// the e-1'th and e'th customers. For customer 1, it's
// the time between start and the customer's arrival.
interarrival_time_mins : event
name "Interarrival" br
"duration" br "(mins)"
format 0.00
// The same thing but in minutes, for display.
arrival_time : event
name "Arrival"
format hh:mm
// The e'th element is the time at which the e'th
// customer arrives.
potential_start_time : event * [1:N]
name "Potential" br "start"
format hh:mm
// The e,N'th element is the time at which the e'th
// customer could start being served by server N, given
// that it may already be busy serving someone else.
next_server : event
name "Server" br "#"
// The e'th element is the server that will be used for
// the e'th customer.
service_start_time : event
name "Service" br "start"
format hh:mm
// The e'th element is the time at which the e'th
// customer starts being served.
service_end_time : event
name "Service" br "end"
format hh:mm
// The e'th element is the time at which the e'th
// customer finishes being served.
service_time : event
name "Service" br "duration"
// The e'th element is the time taken to serve the e'th
// customer.
service_time_mins : event
name "Service" br "duration" br
"(mins)"
format 0.00
// The same thing but in minutes, for display.
start name "Start"
format hh:mm
// When the simulation begins (time at which the servers
// open).
>
where
customer_number[all e] = e and
// Every customer has a unique number. We display this,
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
// but don't use it in the calculation.
interarrival_time[all e] = 10 * rand() / ( 24 * 60 ) and
// The interarrival times are taken from a uniform
// distribution over the interval 0..10 minutes. Excel
// represents times as 1 day = 1, and its rand() function
// returns numbers from the interval 0.0..1.0, so we
// scale down by dividing by 24*60 to bring the range to
// 1 minute, then multiplying by 10.
interarrival_time_mins[all e] =
interarrival_time[e] * 24*60 and
// Scale to 1 minute = 1, for nice display.
service_time[all e] = 20 * rand() / ( 24 * 60 ) and
// Service times are generated in the same way.
service_time_mins[all e] = service_time[e] * 24*60 and
// Scale to 1 minute = 1, for nice display.
arrival_time[all e>1] =
arrival_time[e-1] + interarrival_time[e] and
arrival_time[1] = start + interarrival_time[1] and
// The arrival time of each customer is the arrival time
// of the previous one (or the start time, for customer
// 1) plus the generated interarrival time.
potential_start_time[all e>1,all N] =
if( next_server[e-1] = N,
service_end_time[e-1], arrival_time[e] ) and
potential_start_time[1,all N] = start and
// potential_start_time[e,N] is the earliest time at
// which server N can serve customer e, given that it may
// already be busy. For all servers,
// the potential start time for the first customer is the
// start of the simulation. The potential start time for
// other customers is the time at which the
// server finishes with the previous customer, if it has
// one; otherwise the customer's arrival time.
service_start_time[all e] =
min( range potential_start_time[e] ) and
// The time when we can actually start is the minimum of
// the potential start times. We calculate this by using
// Excel's min function. The MM 'range' construct
// delivers the range of potential_start_time
// across all servers.
service_end_time[all e] =
service_start_time[e] + service_time[e] and
// The time when the customer finishes being served is
// their start time plus the generated service time.
next_server[all e] =
match( service_start_time[e],
range potential_start_time[e], 0 ) and
// To find the first free server for customer e, we use
// Excel's match function. This scans the potential start
// times until it finds service_start_time, and returns
// the index.
start = 9 / 24
// The start time is 9:00 am.
This is the spreadsheet MM generates:
Customer InterarrivalInterarrivalArrival
Potential
#
duration duration
start
(mins)
1 0.000845
1.22
09:01
09:00
2 0.005899
8.50
09:09
09:02
3 0.005184
7.46
09:17
09:10
4 0.001579
2.27
09:19
09:21
5 0.004946
7.12
09:26
09:26
6 0.001618
2.33
09:28
09:29
7 0.002856
4.11
09:33
09:33
8 0.001633
2.35
09:35
09:33
9 0.001845
2.66
09:38
09:52
10 0.005995
8.63
09:46
09:46
Server
#
09:00
09:09
09:17
09:19
09:27
09:28
09:34
09:35
09:38
09:45
09:00
09:09
09:17
09:19
09:26
09:28
09:33
09:35
09:38
09:46
09:00
09:09
09:17
09:19
09:26
09:28
09:33
09:35
09:38
09:46
Service
start
1
1
1
2
1
2
1
1
2
2
09:00
09:02
09:10
09:19
09:26
09:28
09:33
09:33
09:38
09:45
Service
end
09:02
09:10
09:21
09:27
09:29
09:34
09:33
09:52
09:45
09:57
Service Service Start
duration duration
(mins)
0.001776
2.56
09:00
0.005462
7.87
0.007887
11.36
0.005732
8.25
0.002219
3.20
0.004124
5.94
0.000268
0.39
0.013551
19.51
0.004895
7.05
0.008421
12.13
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
APPENDIX 2: ERROR DETECTION
A2.1 USE OF IDENTIFIERS
The extract below demonstrates error-checking. We edited the queuing simulation so that it contained a
duplicated identifier and two undeclared ones. The compiler has detected these and reported them to the listing
file.
1: /* Queue.mm */
2:
...
97:
98:
service_time : event
99:
name "Service" br "duration"
100:
// The e'th element is the time taken to serve the e'th
101:
// customer.
102:
103:
service_time : event
^
Error: Duplicate attribute service_time in object
...
132:
133:
arrival_time[all e>1] =
134:
arrival_time[e-1] + interarrival_tim[e] and
^
Error: Undeclared identifier interarrival_tim
...
139:
140:
potential_start_time[all e>1,all N] =
141:
if( next_server[e-1] = M,
^
Error: Undeclared identifier M
A2.2 UNITS
The extract below shows the compiler detecting unit mismatch errors in constant definitions. Two of these are
errors across the + operator: the compiler assumes that adding quantities with different units is not allowed. The
other is an error in the ^ (power) operator, where it is assumed that one can only raise a quantity to a
dimensionless power. So raising a length to the power 2 to get an area is permitted, but raising it to the power of
2 times another length does not make sense.
1: /* UnitsIncompatibleErrors.mm */
2:
3:
4:
5:
6:
7:
unit cm
unit sec
unit £
constant a = 1 cm + 2 sec
^
Error: The left-hand argument has units cm, but the
right-hand argument has units sec.
...
15: constant e = (1 cm) ^ (2 cm)
^
Error: Operator ^ expects something with no units here,
not units cm.
...
22:
23: constant i = 2(cm/sec) + 3cm*1£
^
Error: The left-hand argument has units cm * sec^-1, but
the right-hand argument has units cm * £.
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
REFERENCES
Biz/Ed page. http://www.bized.ac.uk/ 10:30am 21/4/2001.
Chen, Y. and Chen, H.C. (July 2000), “Visual Checking of Spreadsheets”, Proceedings of EuSpRIG 2000, pages 75-85..
Cleaveland, J.C. (1975), “Meaning and Syntactic Redundancy”, New Directions in Algorithmic Languages 1975 (ed. S.A.Schumann),
IRIA.
Chris Browne’s Linux spreadsheet page. http://vip.hex.net/~cbbrowne/spreadsheets.html 10:40am 21/4/2001.
Goguen, J.A. (1975), “Objects”, Int. J. General Systems, vol 1, pages 237-243.
Goguen, J.A. (1992), “Sheaf Semantics for Concurrent Interacting Objects”, Math. Struct. In Comp. Science, vol 2, pages 159-91.
Goguen, J.A. and Tracz, W. (April 2000), “An Implementation-Oriented Semantics for Module Composition”, Foundations of
Component-based Systems (ed. Gary Leavens and Murali Sitaraman), Cambridge. http://www-cse.ucsd.edu/users/goguen/ps/will.ps.gz is
a more complete version of the paper (7 March 1997, revised 15 October 1998), but the authors state that it has a few errors.
Grossman, T.A. (May-June 1999), “Teachers’ Forum: Spreadsheet Modelling and Simulation Improves Understanding of Queues”,
Interfaces, vol 29(3), pages 88-103. There are links to the paper and to the Excel queuing simulations at
http://www.acs.ucalgary.ca/~grossman/simulation/index.html 9:45am 16/4/2001.
Isakowitz, T., Schocken, S. and Lucas, H.C. (January 1995), “Towards a Logical/Physical Theory of Spreadsheet Modelling”, ACM
Transactions on Information Systems, vol 13(1), pages 1-37.
JavaCC page. This was http://www.metamata.com/, but Metamata have just (24/4/2001) been taken over by Webgain. We are told the
new page will shortly appear on http://www.webgain.com/ .
Kawa page. This contains links to downloads and a user manual, at http://www.gnu.org/software/kawa/ 9:30am 5/4/2001.
Morris, J. (1978), Oxford, Oxford University Press.
Panko, R.P. (July 2000), “Spreadsheet Errors: What We Know, What We Think We Can Do”, Proceedings of EuSpRIG 2000, pages 717.
ProFunda page. This gives information about the Storeys spreadsheet, at http://www.profunda.com/ 2:50pm 16/4/2001.
Rajalingham, K., Chadwick, D.R. and Knight, B. (July 2000), “Classification of Spreadsheet Errors”, Proceedings of EuSpRIG 2000,
pages 23-34..
Spreadsheet FAQ page. This is the Frequently Asked Questions list for the comp.apps.spreadsheets newsgroup. It contains sections on
free and commercial Unix spreadsheets, amongst other useful information, at http://www.faqs.org/faqs/spreadsheets/faq/ 3:30pm
16/4/2001.
SYLK page. This has links to information about the SYLK file format and to code for parsing SYLK files, at
http://www.worldgate.ca/~rschulz/misc/ 3:20pm 16/4/2001.
Ward, M. (August 16 1997),. “Fatal Addition”, New Scientist.
Extracted from Controlling the Subversive Spreadsheet – Risks, Audit and Development Methods
Proceedings of EuSpRIG 2001 Conference ISBN: 1 86166 179 7
Copyright © 2001 European Spreadsheet Risks Interest Group www.eusprig.org
| 6cs.PL
|
arXiv:1604.04969v1 [math.GR] 18 Apr 2016
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL
GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
FUCAI LIN AND CHUAN LIU
Abstract. Given a Tychonoff space X, let A(X) be the free Abelian topological
group over X in the sense of Markov. For every n ∈ N, let An (X) denote the
subspace of A(X) that consists of words of reduced length at most n with respect to
the free basis X. In this paper, we show that A4 (X) is a k-space if and only if A(X)
is a k-space for the non-metrizable Lašnev space X, which gives a complementary
for one result of K. Yamada’s. In addition, we also show that, under the assumption
of ♭ = ω1 , the subspace A3 (X) is a k-space if and only if A(X) is a k-space for
the non-metrizable Lašnev space X. However, under the assumption of ♭ > ω1 , we
provide a non-metrizable Lašnev space X such that A3 (X) is a k-space but A(X) is
not a k-space.
1. Introduction
In 1941, Markov [11] introduced the concept of the free Abelian topological group
A(X) over a Tychonoff space X. Since then, the free Abelian topological groups have
been a source of various examples and also an interesting topic of study in the theory of
topological groups, see [2]. For every n ∈ N, let An (X) denote the subspace of A(X) that
consists of words of reduced length at most n with respect to the free basis X. Recently,
K. Yamada has showed the following two theorems:
Theorem 1.1. [16] Let X be a metrizable space. Then the following statements are
equivalent.
(1) An (X) is a k-space for each n ∈ N.
(2) A4 (X) is a k-space.
(3) either X is locally compact and the set X ′ of all non-isolated points of X is
separable, or X ′ is compact.
Theorem 1.2. [17] For the hedgehog space J(κ) of spininess κ ≥ ℵ0 , the space An (J(κ))
is a k-space for each n ∈ N, but A(J(κ)) is not a k-space.
For a metrizable space X, it follow from Theorems 1.1 and 1.2 that the subspace
A4 (X) with the k-space property implies each An (X) is a k-space, and that each An (X)
is a k-space does not imply A(X) is a k-space. Therefore, it is natural to ask what
the relations of subspaces A4 (X), An (X) and A(X) with the k-space property for a
non-metrizable space X. In this paper, we obtain an unexpected result and show that
for any non-metrizable Lašnev space X (i.e. a non-metrizable, closed image of a metric
space), the subspace A4 (X) is a k-space if and only if A(X) is a k-space. Moreover,
we also discuss the relations of subspaces A3 (X) and A4 (X) with the k-space property
2000 Mathematics Subject Classification. primary 22A30; secondary 54D10; 54E99; 54H99.
Key words and phrases. Free Abelian topological groups; k-spaces; Lašnev spaces; non-metrizable
spaces; sequential spaces.
The first author is supported by the NSFC (Nos. 11571158, 11201414, 11471153), the Natural
Science Foundation of Fujian Province (No. 2012J05013) of China and Training Programme Foundation
for Excellent Youth Researching Talents of Fujian’s Universities (JA13190).
1
2
FUCAI LIN AND CHUAN LIU
for a non-metrizable space X. Our results give a complementary for some results of K.
Yamada’s in literature.
2. Preliminaries
In this section, we introduce the necessary notation and terminologies. First of all,
let N denote the set of all positive integers. For a space X, we always denote the sets
of all non-isolated points and isolated points of X by NI(X) and I(X) respectively. All
spaces are Tychonoff unless stated otherwise. For undefined notation and terminologies,
the reader may refer to [2], [4] and [7].
Let X be a topological space and A ⊆ X be a subset of X. The closure of A in
X is denoted by A and the diagonal of X is denoted by ∆(X). Moreover, A is called
bounded if every continuous real-valued function f defined on A is bounded. Recall that
X is said to have a Gδ -diagonal if ∆(X) is a Gδ -set in X × X. The space X is called a
k-space provided that a subset C ⊆ X is closed in X if C ∩ K is closed in K for each
compact subset K of X. If there exists a family of countably many compact subsets
{Kn : n ∈ N} of X such that each subset F of X is closed in X provided that F ∩ Kn is
closed in Kn for each n ∈ N, then X is called a kω -space. Note that every kω -space is a
k-space. A subset P of X is called a sequential neighborhood of x ∈ X, if each sequence
converging to x is eventually in P . A subset U of X is called sequentially open if U is a
sequential neighborhood of each of its points. The space X is called a sequential space if
each sequentially open subset of X is open. The space X is said to be Fréchet-Urysohn
if, for each x ∈ A ⊂ X, there exists a sequence {xn } such that {xn } converges to x
and {xn : n ∈ N} ⊂ A. Let κ be an infinite
L cardinal. For each α ∈ κ, let Tα be a
sequence converging to xα 6∈ Tα . Let T :=
α∈κ (Tα ∪ {xα }) be the topological sum of
S
{Tα ∪ {xα } : α ∈ κ}. Then Sκ := {x} ∪ α∈κ Tα is the quotient space obtained from T
by identifying all the points xα ∈ T to the point x.
Let P be a family of subsets of X. Then, P is called a k-network [12] if for every
compact subset K of X and an arbitrary
S open set U containing K in X there is a finite
subfamily P ′ ⊆ P such that K ⊆ P ′ ⊆ U . Recall that a space X is an ℵ-space
(resp. ℵ0 -space) if X has a σ-locally finite (resp. countable) k-network. Recall that a
topological space X is said to be Lašnev if it is the closed image of some metric space.
The following two well known facts about the Lašnev spaces shall be used in this paper.
Fact 1: A Lašnev space is metrizable if it contains no closed copy of Sω .
Fact 2: A Lašnev space is an ℵ-space if it contains no closed copy of Sω1 .
Let X be a non-empty Tychonoff space. Throughout this paper, −X := {−x : x ∈ X},
which is just a copy of X. For every n ∈ N, An (X) denotes the subspace of A(X) that
consists of all words of reduced length at most n with respect to the free basis X. Let 0
be the neutral element of A(X) (i.e., the empty word). For every n
L∈ N and
L an element
x1 + x2 + · · · + xn is also called a form for (x1 , x2 , · · · , xn ) ∈ (X
−X {0})n . This
word g is called reduced if it does not contains 0 or any pair of consecutive symbol of the
form x − x. It follows that if the word g is reduced and non-empty, then it is different
from the neutral element 0 of A(X). In particular, each element g ∈ A(X) distinct from
the neutral element can be uniquely written in the form g = r1 x1 + r2 x2 + · · · + rn xn ,
where n ≥ 1, ri ∈ Z \ {0}, xi ∈ X, and xi 6= xj for i 6= j, and the support of g =
r1 x1 + r2 x2 + · · · + rn xn isS defined as supp(g) := {x1 , · · · , xn }. Given a subset K of
A(X), we put supp(K) := g∈K supp(g). For every n ∈ N, let
M
M
in : (X
−X
{0})n → An (X)
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
3
be the natural mapping defined by
in (x1 , x2 , ...xn ) = x1 + x2 + ... + xn
L
L
for each (x1 , x2 , ...xn ) ∈ (X
−X {0})n .
3. main results
First, we give a characterization of a non-metrizable Lašnev space X such that A4 (X)
is a k-space, see Theorem 3.9. In order to obtain this result, we first prove some propositions and lemmas.
Proposition 3.1. If A(X) is a sequential space, then either X is a discrete space or
A(X) contains a closed copy of Sω .
Proof. Assume on the contrary that X is not a discrete space. Since A(X) is sequential
and X is closed in A(X), the space X is sequential, hence there are a point x ∈ X and a
non-trivial sequence {xn : n ∈ N} ⊂ X \ {x} such that {xn : n ∈ N} converges to x. For
each k ∈ N, let Lk S
:= {kxn − kx : n ∈ N}; then Lk ⊂ A2k (X) \ A2k−1 (X) and Lk → 0 as
k → ∞. Let L := {Lk : k ∈ N} ∪ {0}. Next it suffices to show that L is a closed copy
of Sω .
Suppose L is not closed in A(X) or contains no copy of Sω . Then, since A(X) is
sequential, there is a sequence {yn : n ∈ N} ⊂ L such that yn → y ∈
/ L and {yn } meets
infinitely many Lk ’s. This fact implies that there is a subsequence K ⊂ {yn : n ∈ N}
such that K ∩ An (X) is finite, hence K is discrete by [2, Corollary 7.4.3], which is a
contradiction. Therefore, L is a closed copy of Sω .
The following lemma was proved in [9].
Lemma 3.2. [9] Suppose X is a sequential topological group, then either X contains
no closed copy of Sω or every closed first-countable subspace of X is locally countably
compact.
By Lemma 3.2, we can show the following proposition.
Proposition 3.3. Let X be a first-countable space. If A(X) is sequential, then X is
locally countably compact.
Proof. If X is discrete, then it is obvious that X is locally countably compact. Assume
that X is not discrete, then it follows from Proposition 3.1 that A(X) contains a closed
copy of Sω . Therefore, by Lemma 3.2, every first-countable closed subspace of A(X) is
locally countably compact. Since X is first-countable and closed in A(X), the space X
is locally countably compact.
By Proposition 3.3 and the following lemma, we have Corollary 3.5, which was proved
in [1]. The concept of stratifiable space can be seen in [3].
Lemma 3.4. [13] Let X be a stratifiable space. Then A(X) is also a stratifiable space.
Corollary 3.5. [1] If X is a metrizable space and A(X) is a k-space, then X is locally
compact.
Proof. Since X is metrizable, it follows from Lemma 3.4 that A(X) is stratifiable, hence
it is a sequential space. By Proposition 3.3, the space X is locally countably compact,
then it is locally compact since every countably compact subset of a metrizable space is
compact.
4
FUCAI LIN AND CHUAN LIU
Next we recall twoSspaces M1 and M3 , which were introduced in [16].
Let M1 := {x} ∪ ( {Xi : i ∈ N}) such that each Xi is S
an infinite, countable, discrete
and open subspace of M1 and the family {Vk = {x}L
∪ {Xi : i ≥ k} : k ∈ N} is a
neighborhood base of the point x in M1 . Let M3 := {Cα : α < ω1 }, where, for each
α < ω1 , the set Cα := {x(n, α) : n ∈ N} ∪ {xα } with x(n, α) → xα as n → ∞.
Lemma 3.6. Let m0 and n0 be two natural numbers. If Am0 (X) contains a closed copy
of Sω and An0 (X) contains a closed copy of the space M1 , then Am0 +n0 (X) is not a
sequential space.
Proof. Assume that Am0 +n0 (X) is a sequential space. Let {x0 } ∪ {x(n, m) : n, m ∈ N}
be a closed copy of Sω in Am0 (X), where x(n, m) → x0 as m → ∞ for each n ∈ N;
let {y0 } ∪ {y(n, m) : n, m ∈ N} be a closed copy of space M1 in An0 (X), where the set
{y(n, m) : m ∈ N} is discrete and open in An0 (X) for each n ∈ N. Put
H := {x(n, m) − y(n, m) : n, m ∈ N}.
We claim that H contains no non-trivial convergent sequence. Assume on the contrary
that there exists a non-trivial sequence {x(n(k), m(k)) − y(n(k), m(k)) : k ∈ N} converging to z ∈ Am0 +n0 (X). Without loss of generality, we may assume that x(n(k), m(k)) −
y(n(k), m(k)) 6= x(n(l), m(l)) − y(n(l), m(l)) if k 6= l. If there exists some p ∈ N such that
{k ∈ N : n(k) = p} is an infinite set. By the assumption of the sequence, there exists a
subsequence {x(n(lk ), m(lk ))−y(n(lk ), m(lk )) : k ∈ N} of {x(n(k), m(k))−y(n(k), m(k)) :
k ∈ N} such that n(lk ) = p for each k ∈ N and {m(lk ) : k ∈ N} is an infinite set. Then
we have −y(n(lk ), m(lk )) = x(n(lk ), m(lk )) − y(n(lk ), m(lk )) − x(n(lk ), m(lk )) → z − x0
as k → ∞, which is a contradiction. Therefore, there does not exist p ∈ N such that
{k ∈ N : n(k) = p} is an infinite set. Similar to the above proof, ones can see that there
does not exist p ∈ N such that {k ∈ N : m(k) = p} is an infinite set. Therefore, we
have x(n(k), m(k)) = x(n(k), m(k)) − y(n(k), m(k)) + y(n(k), m(k)) → z + y0 as k → ∞,
which is a contradiction. Thus H contains no non-trivial convergent sequence.
By the above claim, it follows that H is sequentially closed in Am0 +n0 (X), then H is
closed in Am0 +n0 (X) since Am0 +n0 (X) is a sequential space. However, it is easy to see
that x0 − y0 ∈ H and x0 − y0 6∈ H, which is a contradiction. Thus Am0 +n0 (X) is not a
sequential space.
The following lemma was proved in [16], which plays an important role in the proof
of our main theorem.
Lemma 3.7. [16, Theorem 3.6] The subspace A4 (M3 ) is not a k-space.
Lemma 3.8. Let X be a Lašnev space. If A2 (X) is a k-space, then X is an ℵ-space.
Proof. By Fact 2, it suffices to show that X contains no any closed copy of Sω1 . Assume
that X contains a closed copy of Sω1 . Then it follows from [18, Corollary 2.2] that A2 (X)
is not a sequential space. However, by Lemma 3.4, the subspace A2 (X) is a stratifiable
space, hence it has a Gδ -diagonal, and then A2 (X) is a sequential space since a k-space
with a Gδ -diagonal is sequential [7].
Recall that a space is called ω1 -compac if every uncountable subset of X has a cluster
point. Now we can show one of our main theorems.
Theorem 3.9. Let X be a non-metrizable Lašnev space. Then the following statements
are equivalent.
(1) A(X) is a sequential space.
(2) A(X) is a k-space.
(3) An (X) is a k-space for each n ∈ N.
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
5
(4) A4 (X) is a k-space.
(5) X is a topological sum of a space with a countable k-network consisting of compact
subsets and a discrete space.
Proof. The implications of (1) ⇒ (2), (2) ⇒ (3) and (3) ⇒ (4) are trivial. The implication
of (2) ⇒ (1) follows from Lemma 3.4. It suffices to show (4) ⇒ (5) and (5) ⇒ (2).
(4) ⇒ (5). First, we show the following claim.
Claim 1: The subspace NI(X) is ω1 -compact in X.
Assume on the contrary that there exists a closed, discrete and uncountable subset
{xα : α < ω1 } in NI(X). Since X is paracompact and NI(X) is closed in X, there is an
uncountable and discrete collection of open subsets {Uα : α < ω1 } in X such that xα ∈ Uα
for each α < ω1 . Since X is Fréchet-Urysohn, for each α < ω1 , let {x(n, α) : n ∈ N} be
a non-trivial sequence converging to xα in X. For each α < ω1 , let
Cα := {x(n, α) : n ∈ N} ∪ {xα }
and put
M3 :=
[
{Cα : α < ω1 }.
Since M3 is a closed subset of X and X is a Lašnev space, it follows from [15] that
the subspace A4 (M3 ) is homeomorphic to a closed subset of A4 (X), thus A4 (M3 ) is
sequential. However, by Lemma 3.7,the subspace A4 (M3 ) is not sequential, which is a
contradiction. Therefore, Claim 1 holds.
Since X a non-metrizable Lašnev space, the space X must contain a closed copy of
Sω by Fact 1. Moreover, since A4 (X) is a sequential space, the subspace A2 (X) is also a
sequential space. Then, by Lemma 3.6, the space X contains no closed copy of the space
M1 . In addition, X is an ℵ-space by Lemma 3.8. Hence there exists a σ-locally finite
k-network P in X such that P is compact for each P ∈ P. Since every compact subset
of X is metrizable in a Lašnev space [7], the family P is a σ-locally-finite k-network
consisting of separable metric subsets of X. Let
P1 := {P : P ∈ P, P ∩ NI(X) 6= ∅}.
By the ω1 -compactness of NI(X), the family {P ∈ P : P ∩ NI(X) 6= ∅} is countable
since at most countably many elements of an arbitrary locally-finite family intersect an
ω1 -compact subset. Therefore, we have |P1 | < ω1 . Note that |P ∩ I(X)| ≤ ω, hence
P1 ∪ {{y} : y ∈ P ∩ I(X), P ∈ P1 }
S
is a countable k-network of X1 := P1 . It is easy to prove that X1 is sequentially open,
thus it is open in X.LLet X2 := X \ X1 . Then X2 is an open and closed discrete subset
of X, thus X = X1 X2 . L
(5) ⇒ (2). Let X := X1 X2 , where X1 is a k-space with a countable k-network
consisting of compact subsets in X1 and X2 is a discrete space. It follows from [14] that
A(X) ∼
= A(X1 ) × A(X2 ). Since X1 is a k-space with a countable k-network consisting
of compact subsets, the space X1 is a kω -space, hence it follows from [2, Theorem 7.4.1]
that A(X1 ) is a kω -space, then A(X) is a k-space since A(X2 ) is discrete.
It is natural to consider that if it can be replaced “A4 (X)” with “A3 (X) or A2 (X)” in
Theorem 3.9. Next we shall give a complete answer to this question. First, the following
proposition shows that it can not be replaced “A4 (X)” with “A2 (X)” in Theorem 3.9.
L
M3 . Then A2 (X) is a sequential space,
Proposition 3.10. Let the space X := Sω
but A(X) is not a sequential space.
6
FUCAI LIN AND CHUAN LIU
L
L
Proof. First of all, it is not difficult to verify that (X L−X L{0})2 is a sequential space.
Then, by [1, Proposition 4.8], the mapping i2 : (X
−X {0})2 → A2 (X) is closed,
which shows that A2 (X) is sequential. Assume that A(X) is sequential, then A(M3 ) is
sequential. Hence NI(M3 ) is separable [1, Theorem 2.11]. This is a contradiction, thus
A(X) is not sequential.
In order to prove that it can not be replaced “A4 (X)” with “A3 (X)” in Theorem 3.9,
we must recall some concepts.
Consider ω ω, the collection of all functions from ω to ω. We define a quasi-order ≤∗
on ω ω by specifying that if f, g ∈ω ω, then f ≤∗ g if f (n) ≤ g(n) for all but finitely many
n ∈ ω. A subset F of ω ω is bounded if there is a g ∈ω ω such that f ≤∗ g for all f ∈ F ,
and is unbounded otherwise. We denote by ♭ the smallest cardinality of an unbounded
family in ω ω. It is well known that ω < ♭ ≤ c, where c denotes the cardinality of the
continuum.
Let UX be the universal uniformity on a space X and put P := {P ⊂ UX : |P | ≤ ω}.
For each P = {Ui }i∈N ∈ P ω , let
W (P ) := {x1 − y1 + x2 − y2 + ... + xk − yk : (xi , yi ) ∈ Ui , i ≤ k, k ∈ N}
and
W := {W (P ) : P ∈ P ω }.
In [16], K. Yamada showed the following important result, which gives a neighborhood
base of 0 in A(X).
Theorem 3.11. [16, Theorem 2.3] The family W is a neighborhood base of 0 in A(X).
The following theorem shows that we can replace “A4 (X)” with “A3 (X)” in Theorem 3.9 under the assumption of ♭ = ω1 . Note that the following proofs contain some
ideas in [6].
Theorem 3.12. Assume ♭ = ω1 . For a non-metrizable Lašnev space X, the subspace
A3 (X) is a sequential space if and only if A(X) is a sequential space.
Proof. Clearly, it suffices to show the necessity. In order to show the sequentiality of
A(X), it suffices to show that the subspace NI(X) is ω1 -compact in X by the proof of
(4) ⇒ (5) of Theorem 3.9. Next we shall show that the subspace NI(X) is ω1 -compact
in X.
Suppose that ♭ = ω1 and A3 (X) is a sequential space. Then there exists a collection
{fα ∈ω ω : α < ω1 } such that if f ∈ω ω, then there exists α < ω1 with fα (n) > f (n) for
infinitely may n ∈ ω. Since X is a non-metrizable Lašnev space, the space X contains a
closed copy of Sω by Fact 1. We rewrite the copy of Sω as Y := {y}∪{y(n, m) : m, n, ∈ ω},
where y(n, m) → y as m → ∞ for each n ∈ N.
Assume on the contrary that the subspace NI(X) is not ω1 -compact in X. Then, by
viewing the
L proof of (3) ⇒ (4) in Theorem 3.9, we can see that X contains a closed copy
of M3 := {Cα : α < ω1 }, where, for each α ∈ ω1 , the set Cα := {x(n, α) : n ∈ ω}∪{xα }
and x(n, α) → xα as n → L
∞. Moreover, without loss of generality, we may assume that
Y ∩ M3 = ∅. Let Z := Y
M3 . Then we can define a uniform base U of the universal
uniformity on Z as follows. For each α < ω1 and k ∈ ω, let
Vk,α := {x(m, α) : m ≥ k} ∪ {xα }
and
Uk,α := (Vk,α × Vk,α ) ∪ ∆α ,
where ∆α is the diagonal of Cα × Cα . For each f ∈ω ω and g ∈ω1 ω, let
[
U (g, f ) := {Ug(α),α : α < ω1 } ∪ (Vf × Vf ) ∪ ∆Z ,
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
7
where Vf := {y} ∪ {y(n, m) : m ≥ f (n), n ∈ ω} and ∆Z is the diagonal of Z × Z. Put
U := {U (g, f ) : g ∈ω1 ω, f ∈ω ω}.
Then, the family U is a uniform base of the universal uniformity on the space Z. Put
W := {W (P ) : P ∈ U ω }. Then it follows from Theorem 3.11 that W is a neighborhood
base of 0 in A(Z).
For each α < ω1 , let
Hα := {y(n, m) + x(n, α) − xα : m ≤ fα (n), n ∈ ω}.
Let H := ∪α<ω1 Hα . Then H ⊂ A3 (Z). Then we have the following claim.
Claim 1: The point y ∈ H \ H in A3 (X).
Proof of Claim 1. Since X is a Lašnev space and Z is closed in X, the group A(Z)
is topologically homeomorphic to a closed subgroup of A(X). Therefore, it suffices to
show that y ∈ H \ H in A3 (Z). Obviously, the family {(y + U ) ∩ A3 (Z) : U ∈ W} is a
neighborhood base of y in A3 (Z). Next we shall prove (y + U ) ∩ A3 (Z) ∩ H 6= ∅ for each
U ∈ W, which implies y ∈ H \ H in A3 (Z). Fix an U ∈ W. Then there exist a sequence
{hi }i∈ω in ω1 ω and a sequence {gi }i∈ω in ω ω such that
U = {x1 − y1 + x2 − y2 + ... + xn − yn : (xi , yi ) ∈ U (hi , gi ), i ≤ n, n ∈ ω}.
Let
B := {x′ − y + x′′ − y ′′ : (x′ , y) ∈ Vg1 × Vg1 , (x′′ , y ′′ ) ∈ Uh1 (α),α , α < ω1 }.
Then B ⊂ U . By the assumption, there exists α < ω1 such that fα (k) ≥ g1 (k) for
infinitely many k. Pick a k ′ > h1 (α) such that (y(k ′ , fα (k ′ )), y) ∈ Vg1 × Vg1 . Then
y(k ′ , fα (k ′ )) − y + x(k ′ , α) − xα ∈ U,
hence
y + y(k ′ , fα (k ′ )) − y + x(k ′ , α) − xα
= y(k ′ , fα (k ′ )) + x(k ′ , α) − xα
∈ ((y + U ) ∩ A3 (Z)) ∩ Hα .
The proof of Claim 1 is completed.
Since A3 (X) is sequential and H is not closed in A3 (X) by Claim 1, there is a sequence
L ⊂ H such that L → z for some z ∈ A3 (X) \ H. Then the set supp(L ∪ {z}) is bounded
by [1], then the closure of supp(L ∪ {z}) is compact since X is a Lašnev space. If L
meets infinitely many Hα ’s, then the set supp(L ∪ {z}) contains infinitely many xα ’s.
This is a contradiction since {xα : α < ω1 } is discrete in X. If L ∪ {z} is contained in
the union of finitely many Hα ’s, then the set supp(L ∪ {z}) contains an infinite subset
{y(ni , mi ) : i ∈ N} of {y(n, m) : n, m ∈ N} such that ni 6= nj if i 6= j, which is a
contradiction. Therefore, the subspace NI(X) is ω1 -compact in X.
Finally, we shall give an example to show that it can not be replaced “A4 (X)” with
“A3 (X)” in Theorem 3.9 under the assumption of ♭ > ω1 . First, we give a technical
lemma.
Let C ⊂ M3 be a convergent sequence with the limitL
point. Since the subspaces
L
M3 respectively, it follows
C
Sω , M3 and Sω are all P ∗ -embedded in the space Sω
from [2, Problem 7.7.B.] that the following three subgroups
M
M
M
M
A(C
Sω , Sω
M3 ), A(M3 , Sω
M3 ) and A(Sω , Sω
M3 )
L
L
M3 ) are topologically isomorphic to the groups A(C
Sω ), A(M
in A(Sω
L3 ) and A(Sω )
respectively. Then, by [2, Corollary 7.4.9] and [16, Theorem 4.9], A(C
Sω ), A(M3 )
and A(Sω ) are all sequential, hence we have the following lemma.
8
FUCAI LIN AND CHUAN LIU
Lemma 3.13. The following three subspaces
M
M
M
M
A3 (C
Sω , Sω
M3 ), A3 (M3 , Sω
M3 ) and A3 (Sω , Sω
M3 )
L
M3 ).
are all sequential and closed in A(Sω
Let S(A) denote the set of all limit points of a subset A in a space X. In order to
show the last theorem, we need the following technical lemma.
L
M3 ).
Lemma 3.14. Let B be a sequentially closed subset of A3 (Sω
(a) If
B ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vf × Vf } 6= ∅
for any f ∈ω ω, then y ∈ B.
(b) If
B ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vh(α),α × Vh(α),α , α < ω1 } =
6 ∅
ω1
for any h ∈ ω, then y ∈ B.
(c) If
B ∩ {y + x′ − y ′ + x′′ − y ′′ : (x′ , y ′ ) ∈ Vf × Vf , (x′′ , y ′′ ) ∈ Vg × Vg } =
6 ∅
for any f, g ∈ω ω, then y ∈ B.
L
M3 ). Then B1 is sequentially closed in the
Proof. (a) Let B1 :=
L B ∩ A3 (Sω , Sω
M3 ), hence it is closed by Lemma 3.13. Since
subspace A3 (Sω , Sω
B1 ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vf × Vf } = B ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vf × Vf } =
6 ∅
for any f ∈ω ω, we have y = y + y − y ∈ B1 = B1 ⊂ B
(b) Since B ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vh(α),α × Vh(α),α , α < ω1 } 6= ∅ for any h ∈ω1 ω, it
is easy to see that there exists β such that
B ∩ {y + x′ − y ′ : (x′ , y ′ ) ∈ Vh(β),β × Vh(β),β } 6= ∅
for any h ∈ω1 ω. Then it easily obtain that y = y + xβ − xβ ∈ S(B) = B.
L
M3 ). By [2, Theorem 7.4.5] and Lemma 3.13, the
(c) Let B1 := B L
∩ A3 (Sω , Sω
M3 ) is closed and sequential, which implies that B1 is closed in
subspace AL
3 (Sω , Sω
A3 (Sω , Sω
M3 ). Let
A := B1 ∩ {y + x′ − y ′ + x′′ − y ′′ : (x′ , y ′ ) ∈ Vf × Vf , (x′′ , y ′′ ) ∈ Vg × Vg , f, g ∈ω ω}.
Then y ∈ A, and A ⊂ B1 . It implies that y ∈ B1 ⊂ B.
Theorem 3.15. Assume ♭ > ω1 . There exists a non-metrizable Lašnev space X such
that A3 (X) is a sequential space but A(X) is not a sequential space.
L
M3 ; then X is a non-metrizable Lašnev space. However, A(X) is
Proof. Let X = Sω
not a sequential space. Indeed, assume on the contrary that A(X) is a sequential space.
Since A(M3 ) is topologically homeomorphic to a closed subset of A(X), the group A(M3 )
is sequential, then it follows from [1, Theorem 2.11] that NI(M3 ) is separable, which is a
contradiction since NI(M3 ) is a uncountable closed discrete subset in M3 .
Next, we shall prove that A3 (X) is a sequential space. Assume on the contrary that
A3 (X) is not a sequential space. Then there exists a subset H in A3 (X) such that H
is a sequentially closed subset in A3 (X), but not closed in A3 (X). Then there exists a
point x0 belonging to A3 (X) such that x0 ∈ H \ H in A3 (X). Let H1 := A2 (X) ∩ H and
H2 := (A3 (X) \ A2 (X)) ∩ H. Since H1 is sequentially closed and A2 (X) is sequential,
the set H1 is closed. If x0 ∈ H1 , then x0 ∈ H1 = H1 ⊂ H, this is a contradiction.
Hence x0 ∈ H2 . Then since A3 (X) is paracompact (in fact, A3 (X) is stratifiable), there
exist open subsets W1 , W2 in A3 (X) such that x0 ∈ W1 , H1 ⊂ W2 and W1 ∩ W2 = ∅.
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
9
Let H ′ := W2 ∩ H ⊂ A3 (X) \ A2 (X). Then the point x0 belongs to H ′ , the set H ′ is
sequentially closed and H ′ does not have any limit point in X ∪ (−X). Without loss of
generality, we may assume H = H ′ .
Since the length of each element of H is 3, the length of x0 is 1 or 3. However, the
length of x0 can not be 3. Assume on the contrary that x0 ∈ A3 (X) \ A2 (X); then let V
be an open neighborhood of x0 in A3 (X) \ A2 (X) such that V ⊂ A3 (X) \ A2 (X). Then
V ∩ H is sequentially closed in the sequential subspace A3 (X) \ A2 (X), hence V ∩ H is
closed and x0 ∈ H, this is a contradiction. Therefore, the length of x0 is 1. Without loss
of generality, we may assume that x0 ∈ X. In order to obtain a contradiction, we shall
show that x0 = y, but y 6∈ H. We divide the proof in two claims.
Claim 2: We have x0 = y.
Proof of Claim 2. It suffices to show that x0 6= x for each x ∈ X \ {y}. Assume on
the contrary that x0 ∈ X \ {y}. In order to obtain a contradiction, we divide the proof
in three cases.
Subcase 1: For some n, m ∈ ω, we have x0 = y(n, m).
Clearly, ones can choose f ∈ω ω such that y(n, m) ∈
/ Vf . Fix an arbitrary g ∈ω1 ω,
ω
and then let P := {U (g, f ), U (g, f ), ...} ∈ U . Then (x0 + W (P )) ∩ H 6= ∅. Since
x0 = y(n, m) ∈
/ Vf , each element of (x0 + W (P )) ∩ H has the form x0 + x′ − y ′ , where
′ ′
(x , y ) ∈ U (g, f ). Put
B := {x′ − y ′ : x0 + x′ − y ′ ∈ (x0 + W (P )) ∩ H, (x′ , y ′ ) ∈ U (g, f )}.
Then B ⊂ −x0 + H, and it follows from x0 ∈ H that 0 ∈ B. Moreover, it is obvious
that −x0 + H is sequentially closed in A4 (X). Hence the set (−x0 + H) ∩ A2 (X) is
sequentially closed in A2 (X), then it follows that (H − x0 ) ∩ A2 (X) is closed since A2 (X)
is a sequential space. However, since B ⊂ (−x0 + H) ∩ A2 (X), we have
0 ∈ (−x0 + H) ∩ A2 (X) = (−x0 + H) ∩ A2 (X)
in A2 (X), which shows that x0 ∈ H, this is a contradiction.
Subcase 2: For some n ∈ ω and β ∈ ω1 , we have x0 = x(n, β).
Fix arbitrary f ∈ ω ω and g ∈ ω1 ω with g(β) > n. Then (x0 + W (P )) ∩ H 6= ∅,
where P = (U (g, f )) ∈ U ω . Obviously, each element of (x0 + W (P )) ∩ H has the form
x0 + x′k − yk′ ∈ H, where (x′k , yk′ ) ∈ U (g, f ). By a proof analogous to Subcase 1, we can
see that x0 ∈ H, which is a contradiction.
Subcase 3: For some β ∈ ω1 , we have x0 = xβ .
Obviously, we have x0 ∈ (x0 + W (P )) ∩ H for any P ∈ U ω . Fix f ∈ ω ω and g ∈ ω1 ω.
Let
A := (x0 + W (P )) ∩ H,
C := {x0 + x′ − y ′ : (x′ , y ′ ) ∈ U (g, f ), x′ 6= xβ , y ′ 6= xβ , x′ 6= y ′ } ∩ H
and
D := {x′ + x′′ − y ′′ : (x′′ , y ′′ ) ∈ U (g, f ), x′ ∈ Vg(β),β \ {xβ }, x′ 6= y ′′ } ∩ H.
Clearly, we have A = C ∪ D. If x0 ∈ C, then, by a proof analogous to Subcase 1, we can
see that x0 ∈ H, which is a contradiction. Hence we assume x0 ∈ D.
Let
M
M
B1 := H ∩ A3 (Cβ
Sω , Sω
M3 )
and
B2 := H ∩ A3 (M3 , Sω
M
M3 ),
where Cβ := {xβ } ∪ {x(n, β) : n ∈ N}. By L
Lemma 3.13,
B2 are
L the sets B1 and L
M3 )
M3 ) and A3 (M3 , Sω
Sω , Sω
sequentially closed subsets of the spaces A3 (Cβ
10
FUCAI LIN AND CHUAN LIU
respectively, therefore, closed in A(X). Then B1 ∪ B2 is closed in A(X), hence we have
x0 ∈ D ⊂ B1 ∪ B2 ⊂ H, which is a contradiction.
Therefore, we have x0 = y. The proof of Claim 2 is completed.
Claim 3: The point y ∈
/ H in A(X).
proof of Claim 3. Obviously, it suffices to show that there exist two mappings h ∈ω ω
and g ∈ω1 ω such that (y + W (P )) ∩ H = ∅, where P = (U (g, h), U (g, h), ...). Since
y 6∈ H, it follows from Lemma 3.14 that we have the following statements:
(a) there exists h1 ∈ω ω such that {y + x′ − y ′ : (x′ , y ′ ) ∈ Vh1 × Vh1 } ∩ H = ∅;
(b) there exists g1 ∈ω1 ω such that
{y + x′ − y ′ : (x′ , y ′ ) ∈ Vg1 (α),α × Vg1 (α),α , α < ω1 } ∩ H = ∅;
(c) there exist h2 , h3 ∈ω ω such that
{y + x′ − y ′ + x′′ − y ′′ : (x′ , y ′ ) ∈ Vh2 × Vh2 , (x′′ , y ′′ ) ∈ Vh3 × Vh3 } ∩ H = ∅.
Therefore, without loss of generality, we may assume that
H ⊂ {y(n, m) + x1 − y1 : n, m ∈ N, x1 , y1 ∈ Cα , α < ω1 }.
Fix an arbitrary α < ω1 . Since y 6∈ H, there exist a fα ∈ ω ω and n(α) ∈ N such that
{y(n, m) + x1 − y1 : n, m ∈ N, m ≥ fα (n), x1 , y1 ∈ Vn(α),α } ∩ H = ∅.
ω
(⋆)
By the assumption of ♭ > ω1 , there is a f ∈ ω such that fα ≤ f for each α < ω1 . For
each α < ω1 , let k(α) be the smallest natural number such that fα (k) ≤ f (k) whenever
k ≥ k(α). We claim that, for each α < ω1 , there exists n′ (α) ∈ N with n′ (α) ≥ n(α)
such that
∗
{y(j, m) + x1 − y1 : j < k(α), m ∈ N, x1 , y1 ∈ Vn′ (α),α } ∩ H = ∅.
(⋆⋆)
Assume the converse. Then it is easy to see that there is a convergent sequence of H with
the limit point in X, which is a contradiction with the above assumption. Let g ∈ω1 ω
with g(α) = n′ (α) and P = (U (g, f ), U (g, f ), ...). It follows from (⋆) and (⋆⋆) ones can
see
{y(n, m) + x1 − y1 : y(n, m) ∈ Vf , (x1 , y1 ) ∈ Vg(α),α × Vg(α),α , α < ω1 , n, m ∈ N} ∩ H = ∅.
Then we have (y + W (P )) ∩ H = ∅. The proof of Claim 3 is completed.
By Claims 2 and 3, we obtain a contradiction. Hence A3 (X) is not a sequential
space.
References
[1] A.V. Arhangel’skiı̌, O.G. Okunev and V.G. Pestov, Free topological groups over metrizable spaces,
Topology Appl., 33(1989), 63–76.
[2] A. Arhangel’skiı̌ and M. Tkachenko, Topological groups and related structures, Atlantis Press, Paris;
World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2008.
[3] C.J. Borges, On stratifiable spaces, Pacific J. Math., 17(1)(1966), 1–16.
[4] R. Engelking, General Topology (revised and completed edition), Heldermann Verlag, Berlin, 1989.
[5] M.I. Graev, Free topological groups, Izvestiya Akad. Nauk SSSR Ser. Mat. 12(1948), 279–323.
[6] G. Gruenhage, k-spaces and products of closed images of metric spaces, Proc. Amer. Math. Soc.,
80(3)(1980), 478–482.
[7] G. Gruenhage, Generalized metric spaces, K. Kunen, J.E. Vaughan eds., Handbook of Set-Theoretic
Topology, North-Holland, (1984), 423–501.
[8] J.A. Guthrie, A characterization of ℵ0 -spaces, General Topology Appl., 1 (1971), 105–110.
[9] C. Liu and S. Lin, Generalized metric spaces with algebraic structures, Topology Appl., 157(2010),
1966–1974.
[10] S. Lin and Y. Tanaka, Point-countable k-networks, closed maps, and related results, Topology Appl.,
59 (1994), 79–86.
[11] A.A. Markov, On free topological groups, Dokl. Akad. Nauk. SSSR, 31(1941), 299–301.
THE k-SPACES PROPERTY OF FREE ABELIAN TOPOLOGICAL GROUPS OVER NON-METRIZABLE LAŠNEV SPACES
11
[12] P. O′ Meara, On paracompactness in function spaces with the compact-open topology, Proc. Amer.
Math. Soc., 29 (1971), 183–189.
[13] O.V. Sipacheva, On the stratifiability of free Abelian topological groups. Topology Proc. 18(1993),
271–311.
[14] B.V.S. Thomas, Free topological groups, Gen. Topol. Appl., 4(1974), 51–72.
[15] V.V. Uspenskij, Free topological groups of metrizable spaces, Math. USSR. Izv., 37(3)(1991), 657–
680.
[16] K. Yamada, Characterization of a metrizable space X such that every An (X) is a k-space, Topology
Appl., 49(1993), 75–94.
[17] K. Yamada, Free Abelian topological groups and k-spaces, Glas. Math., 31(1996), 8-91.
[18] K. Yamada, Tightness of free Abelian topological groups and of finite products of sequential fans,
Topology Proc., 22(1997), 363–381.
(Fucai Lin): School of mathematics and statistics, Minnan Normal University, Zhangzhou
363000, P. R. China
E-mail address: [email protected]; [email protected]
(Chuan Liu): Department of Mathematics, Ohio University Zanesville Campus, Zanesville,
OH 43701, USA
E-mail address: [email protected]
| 4math.GR
|
XX
TRIÈST: Counting Local and Global Triangles in Fully-dynamic Streams
with Fixed Memory Size
LORENZO DE STEFANI and ELI UPFAL, Brown University
ALESSANDRO EPASTO, Google
MATTEO RIONDATO, Two Sigma Investments, LP
arXiv:1602.07424v2 [cs.DS] 28 Jun 2016
“Ogni lassada xe persa” 1 – Proverb from Trieste, Italy.
We present trièst, a suite of one-pass streaming algorithms to compute unbiased, low-variance, highquality approximations of the global and local (i.e., incident to each vertex) number of triangles in a
fully-dynamic graph represented as an adversarial stream of edge insertions and deletions.
Our algorithms use reservoir sampling and its variants to exploit the user-specified memory space at all
times. This is in contrast with previous approaches, which require hard-to-choose parameters (e.g., a fixed
sampling probability) and offer no guarantees on the amount of memory they use. We analyze the variance
of the estimations and show novel concentration bounds for these quantities.
Our experimental results on very large graphs demonstrate that trièst outperforms state-of-the-art
approaches in accuracy and exhibits a small update time.
CCS Concepts: •Mathematics of computing → Graph enumeration; Probabilistic algorithms;
•Information systems → Data stream mining; •Human-centered computing → Social networks;
•Theory of computation → Dynamic graph algorithms; Sketching and sampling;
Additional Key Words and Phrases: Cycle counting; Reservoir sampling; Subgraph counting;
ACM Reference Format:
Lorenzo De Stefani, Alessandro Epasto, Matteo Riondato, and Eli Upfal. 2016. TRIÈST: Counting Local
and Global Triangles in Fully-dynamic Streams with Fixed Memory Size. ACM XX, X, Article XX ( 2016),
49 pages.
DOI: 0000001.0000001
1. INTRODUCTION
Exact computation of characteristic quantities of Web-scale networks is often impractical
or even infeasible due to the humongous size of these graphs. It is natural in these cases to
resort to efficient-to-compute approximations of these quantities that, when of sufficiently
high quality, can be used as proxies for the exact values.
In addition to being huge, many interesting networks are fully-dynamic and can be represented as a stream whose elements are edges/nodes insertions and deletions which occur in
an arbitrary (even adversarial) order. Characteristic quantities in these graphs are intrinsically volatile, hence there is limited added value in maintaining them exactly. The goal is
rather to keep track, at all times, of a high-quality approximation of these quantities. For
efficiency, the algorithms should aim at exploiting the available memory space as much as
possible and they should require only one pass over the stream.
We introduce trièst, a suite of sampling-based, one-pass algorithms for adversarial fullydynamic streams to approximate the global number of triangles and the local number of triA preliminary report of this work appeared in the proceedings of ACM KDD’16 as [De Stefani et al. 2016].
This work was supported in part by NSF grant IIS-1247581 and NIH grant R01-CA180776.
Authors’ addresses: Lorenzo De Stefani and Eli Upfal, Department of Computer Science, Brown University,
email: {lorenzo,eli}@cs.brown.edu; Alessandro Epasto, Google Inc., email: [email protected]; Matteo
Riondato, Two Sigma Investments LP, email: [email protected].
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. Copyrights for third-party components of this
work must be honored. For all other uses, contact the owner/author(s).
c 2016 Copyright held by the owner/author(s). 0000-0000/2016/-ARTXX $15.00
DOI: 0000001.0000001
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:2
Lorenzo De Stefani et al.
angles incident to each vertex. Mining local and global triangles is a fundamental primitive
with many applications (e.g., community detection [Berry et al. 2011], topic mining [Eckmann and Moses 2002], spam/anomaly detection [Becchetti et al. 2010; Lim and Kang 2015],
ego-networks mining [Epasto et al. 2015a] and protein interaction networks analysis [Milo
et al. 2002].)
Many previous works on triangle estimation in streams also employ sampling (see Sect. 3),
but they usually require the user to specify in advance an edge sampling probability p that
is fixed for the entire stream. This approach presents several significant drawbacks. First,
choosing a p that allows to obtain the desired approximation quality requires to know or
guess a number of properties of the input (e.g., the size of the stream). Second, a fixed p
implies that the sample size grows with the size of the stream, which is problematic when
the stream size is not known in advance: if the user specifies a large p, the algorithm may
run out of memory, while for a smaller p it will provide a suboptimal estimation. Third, even
assuming to be able to compute a p that ensures (in expectation) full use of the available
space, the memory would be fully utilized only at the end of the stream, and the estimations
computed throughout the execution would be suboptimal.
Contributions. We address all the above issues by taking a significant departure from
the fixed-probability, independent edge sampling approach taken even by state-of-the-art
methods [Lim and Kang 2015]. Specifically:
— We introduce trièst (TRI angle Estimation from ST reams), a suite of one-pass
streaming algorithms to approximate, at each time instant, the global and local number
of triangles in a fully-dynamic graph stream (i.e., a sequence of edges additions and deletions in arbitrary order) using a fixed amount of memory. This is the first contribution that
enjoys all these properties. trièst only requires the user to specify the amount of available
memory, an interpretable parameter that is definitively known to the user.
— Our algorithms maintain a sample of edges: they use the reservoir sampling [Vitter
1985] and random pairing [Gemulla et al. 2008] sampling schemes to exploit the available
memory as much as possible. To the best of our knowledge, ours is the first application
of these techniques to subgraph counting in fully-dynamic, arbitrarily long, adversarially
ordered streams. We present an analysis of the unbiasedness and of the variance of our estimators, and establish strong concentration results for them. The use of reservoir sampling
and random pairing requires additional sophistication in the analysis, as the presence of
an edge in the sample is not independent from the concurrent presence of another edge.
Hence, in our proofs we must consider the complex dependencies in events involving sets of
edges. The gain is worth the effort: we prove that the variance of our algorithms is smaller
than that of state-of-the-art methods [Lim and Kang 2015], and this is confirmed by our
experiments.
— We conduct an extensive experimental evaluation of trièst on very large graphs, some
with billions of edges, comparing the performances of our algorithms to those of existing
state-of-the-art contributions [Jha et al. 2015; Lim and Kang 2015; Pavan et al. 2013]. Our
algorithms significantly and consistently reduce the average estimation error by up to 90%
w.r.t. the state of the art, both in the global and local estimation problems, while using the
same amount of memory. Our algorithms are also extremely scalable, showing update times
in the order of hundreds of microseconds for graphs with billions of edges.
In this article, we extend the conference version [De Stefani et al. 2016] in multiple ways.
First of all, we include all proofs of our theoretical results and give many additional details
that were omitted from the conference version due to lack of space. Secondly, we strengthen
the analysis of trièst, presenting tighter bounds to the variance of its variants. Thirdly, we
show how to extend trièst to approximate the count of triangles in multigraphs. Additionally, we include a whole subsection of discussion of our results, highlighting their advantages,
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:3
disadvantages, and limitations. Finally, we expand our experimental evaluation, reporting
the results of additional experiments and giving additional details on the comparison with
existing state-of-the-art methods.
Paper organization. We formally introduce the settings and the problem in Sect. 2. In
Sect. 3 we discuss related works. We present and analyze trièst and discuss our design
choices in Sect. 4. The results of our experimental evaluation are presented in Sect. 5. We
draw our conclusions in Sect. 6. Some of the proofs of our theoretical results are deferred
to Appendix A.
2. PRELIMINARIES
We study the problem of counting global and local triangles in a fully-dynamic undirected
graph as an arbitrary (adversarial) stream of edge insertions and deletions.
Formally, for any (discrete) time instant t ≥ 0, let G(t) = (V (t) , E (t) ) be the graph
observed up to and including time t. At time t = 0 we have V (t) = E (t) = ∅. For any t > 0,
at time t + 1 we receive an element et+1 = (•, (u, v)) from a stream, where • ∈ {+, −} and
u, v are two distinct vertices. The graph G(t+1) = (V (t+1) , E (t+1) ) is obtained by inserting
a new edge or deleting an existing edge as follows:
(t)
E ∪ {(u, v)} if • = “ + ”
(t+1)
E
=
.
E (t) \ {(u, v)} if • = “ − ”
If u or v do not belong to V (t) , they are added to V (t+1) . Nodes are deleted from V (t) when
they have degree zero.
Edges can be added and deleted in the graph in an arbitrary adversarial order, i.e., as
to cause the worst outcome for the algorithm, but we assume that the adversary has no
access to the random bits used by the algorithm. We assume that all operations have effect:
if e ∈ E (t) (resp. e 6∈ E (t) ), (+, e) (resp. (−, e)) can not be on the stream at time t + 1.
Given a graph G(t) = (V (t) , E (t) ), a triangle in G(t) is a set of three edges
{(u, v), (v, w), (w, u)} ⊆ E (t) , with u, v, and w being three distinct vertices. We refer to
{u, v, w} ⊆ V (t) as the corners of the triangle. We denote with ∆(t) the set of all triangles
(t)
in G(t) , and, for any vertex u ∈ V (t) , with ∆u the subset of ∆(t) containing all and only
the triangles that have u as a corner.
Problem definition. We study the Global (resp. Local) Triangle Counting Problem in
Fully-dynamic Streams, which requires to compute, at each time t ≥ 0 an estimation of
(t)
|∆(t) | (resp. for each u ∈ V an estimation of |∆u |).
Multigraphs. Our approach can be further extended to count the number of global and
local triangles on a multigraph represented as a stream of edges. Using a formalization
analogous to that discussed for graphs, for any (discrete) time instant t ≥ 0, let G(t) =
(V (t) , E (t) ) be the multigraph observed up to and including time t, where E (t) is now a bag
of edges between vertices of V (t) . The multigraph evolves through a series of edges additions
and deletions according to the same process described for graphs. The definition of triangle
in a multigraph is also the same. As before we denote with ∆(t) the set of all triangles in
G(t) , but now this set may contain multiple triangles with the same set of vertices, although
each of these triangles will be a different set of three edges among those vertices. For any
(t)
vertex u ∈ V (t) , we still denote with ∆u the subset of ∆(t) containing all and only the
triangles that have u as a corner, with a similar caveat as ∆(t) . The problems of global and
local triangle counting in multigraph edge streams are defined exactly in the same way as
for graph edge streams.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:4
Lorenzo De Stefani et al.
3. RELATED WORK
The literature on exact and approximate triangle counting is extremely rich, including exact
algorithms, graph sparsifiers [Tsourakakis et al. 2009, 2011], complex-valued sketches [Kane
et al. 2012; Manjunath et al. 2011], and MapReduce algorithms [Pagh and Tsourakakis
2012; Park et al. 2014; Park and Chung 2013; Park et al. 2016; Suri and Vassilvitskii 2011].
Here we restrict the discussion to the works most related to ours, i.e., to those presenting
algorithms for counting or approximating the number of triangles from data streams. We
refer to the survey by Latapy [2008] for an in-depth discussion of other works. Table I
presents a summary of the comparison, in terms of desirable properties, between this work
and relevant previous contributions.
Many authors presented algorithms for more restricted (i.e., less generic) settings than
ours, or for which the constraints on the computation are more lax [Bar-Yossef et al. 2002;
Buriol et al. 2006; Jowhari and Ghodsi 2005; Kutzkov and Pagh 2013]. For example, Becchetti et al. [2010] and Kolountzakis et al. [2012] present algorithms for approximate triangle
counting from static graphs by performing multiple passes over the data. Pavan et al. [2013]
and Jha et al. [2015] propose algorithms for approximating only the global number of triangles from edge-insertion-only streams. Bulteau et al. [2015] present a one-pass algorithm
for fully-dynamic graphs, but the triangle count estimation is (expensively) computed only
at the end of the stream and the algorithm requires, in the worst case, more memory than
what is needed to store the entire graph. Ahmed et al. [2014] apply the sampling-and-hold
approach to insertion-only graph stream mining to obtain, only at the end of the stream
and using non-constant space, an estimation of many network measures including triangles.
None of these works has all the features offered by trièst: performs a single pass over the
data, handles fully-dynamic streams, uses a fixed amount of memory space, requires a single
interpretable parameter, and returns an estimation at each time instant. Furthermore, our
experimental results show that we outperform the algorithms from [Jha et al. 2015; Pavan
et al. 2013] on insertion-only streams.
Lim and Kang [2015] present an algorithm for insertion-only streams that is based on
independent edge sampling with a fixed probability: for each edge on the stream, a coin
with a user-specified fixed tails probability p is flipped, and, if the outcome is tails, the edge
is added to the stored sample and the estimation of local triangles is updated. Since the
memory is not fully utilized during most of the stream, the variance of the estimate is large.
Our approach handles fully-dynamic streams and makes better use of the available memory
space at each time instant, resulting in a better estimation, as shown by our analytical and
experimental results.
Vitter [1985] presents a detailed analysis of the reservoir sampling scheme and discusses
methods to speed up the algorithm by reducing the number of calls to the random number
generator. Random Pairing [Gemulla et al. 2008] is an extension of reservoir sampling to
handle fully-dynamic streams with insertions and deletions. Cohen et al. [2012] generalize
and extend the Random Pairing approach to the case where the elements on the stream are
key-value pairs, where the value may be negative (and less than −1). In our settings, where
the value is not less than −1 (for an edge deletion), these generalizations do not apply and
the algorithm presented by Cohen et al. [2012] reduces essentially to Random Pairing.
4. ALGORITHMS
We present trièst, a suite of three novel algorithms for approximate global and local
triangle counting from edge streams. The first two work on insertion-only streams, while
the third can handle fully-dynamic streams where edge deletions are allowed. We defer the
discussion of the multigraph case to Sect. 4.4.
Parameters. Our algorithms keep an edge sample S containing up to M edges from the
stream, where M is a positive integer parameter. For ease of presentation, we realistically
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:5
Table I. Comparison with previous contributions
Work
Single
pass
Fixed
space
Local
counts
Global
counts
Fully-dynamic
streams
[Becchetti et al. 2010]
[Kolountzakis et al. 2012]
[Pavan et al. 2013]
[Jha et al. 2015]
[Ahmed et al. 2014]
[Lim and Kang 2015]
This work
7
7
3
3
3
3
3
3/7†
7
3
3
7
7
3
3
7
7
7
7
7
3
7
3
3
3
3
7
3
7
7
7
7
7
7
3
† The
required space is O(|V (t) |), which, although not dependent on the number of
triangles or on the number of edges, is not fixed in the sense that it can be fixed
a-priori.
assume M ≥ 6. In Sect. 1 we motivated the design choice of only requiring M as a parameter
and remarked on its advantages over using a fixed sampling probability p. Our algorithms
are designed to use the available space as much as possible.
Counters. trièst algorithms keep counters to compute the estimations of the global and
local number of triangles. They always keep one global counter τ for the estimation of the
global number of triangles. Only the global counter is needed to estimate the total triangle
count. To estimate the local triangle counts, the algorithms keep a set of local counters τu
for a subset of the nodes u ∈ V (t) . The local counters are created on the fly as needed, and
always destroyed as soon as they have a value of 0. Hence our algorithms use O(M ) space
(with one exception, see Sect. 4.2).
Notation. For any t ≥ 0, let GS = (V S , E S ) be the subgraph of G(t) containing all and
only the edges in the current sample S. We denote with NuS the neighborhood of u in GS :
S
= NuS ∩ NvS the shared neighborhood of u and
NuS = {v ∈ V (t) : (u, v) ∈ S} and with Nu,v
S
v in G .
Presentation. We only present the analysis of our algorithms for the problem of global
triangle counting. For each presented result involving the estimation of the global triangle
count (e.g., unbiasedness, bound on variance, concentration bound) and potentially using
other global quantities (e.g., the number of pairs of triangles in ∆(t) sharing an edge), it is
straightforward to derive the correspondent variant for the estimation of the local triangle
(t)
count, using similarly defined local quantities (e.g., the number of pairs of triangles in ∆u
sharing an edge.)
4.1. A first algorithm – trièst-base
We first present trièst-base, which works on insertion-only streams and uses standard
reservoir sampling [Vitter 1985] to maintain the edge sample S:
— If t ≤ M , then the edge et = (u, v) on the stream at time t is deterministically inserted in
S.
— If t > M , trièst-base flips a biased coin with heads probability M/t. If the outcome is
heads, it chooses an edge (w, z) ∈ S uniformly at random, removes (w, z) from S, and
inserts (u, v) in S. Otherwise, S is not modified.
After each insertion (resp. removal) of an edge (u, v) from S, trièst-base calls the proceS
dure UpdateCounters that increments (resp. decrements) τ , τu and τv by |Nu,v
|, and τc
S
by one, for each c ∈ Nu,v
.
The pseudocode for trièst-base is presented in Alg. 1.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:6
Lorenzo De Stefani et al.
ALGORITHM 1 trièst-base
Input: Insertion-only edge stream Σ, integer M ≥ 6
1: S ← ∅, t ← 0, τ ← 0
2: for each element (+, (u, v)) from Σ do
3:
t←t+1
4:
if SampleEdge((u, v), t) then
5:
S ← S ∪ {(u, v)}
6:
UpdateCounters(+, (u, v))
7: function SampleEdge((u, v), t)
8:
if t ≤ M then
9:
return True
10:
else if FlipBiasedCoin( M
) = heads then
t
11:
(u0 , v 0 ) ← random edge from S
12:
S ← S \ {(u0 , v 0 )}
13:
UpdateCounters(−, (u0 , v 0 ))
14:
return True
15:
return False
16: function UpdateCounters((•, (u, v)))
S
17:
Nu,v
← NuS ∩ NvS
S
18:
for all c ∈ Nu,v
do
19:
τ ←τ •1
20:
τc ← τc • 1
21:
τu ← τu • 1
22:
τv ← τv • 1
4.1.1. Estimation. For any pair of positive integers a and b such that a ≤ min{M, b} let
ξa,b
1
if b ≤ M
.
a−1
Y
b
b
−
a
b
−
i
=
otherwise
=
M
M −a
M −i
i=0
.
−1
As shown in the following lemma, ξk,t
is the probability that k edges of G(t) are all in S at
time t, i.e., the k-th order inclusion probability of the reservoir sampling scheme. The proof
can be found in App. A.1.
Lemma 4.1. For any time step t and any positive integer k ≤ t, let B be any subset of
E (t) of size |B| = k ≤ t. Then, at the end of time step t,
0 if k > M
Pr(B ⊆ S) =
.
−1
ξk,t
otherwise
We make use of this lemma in the analysis of trièst-base.
(t)
Let, for any t ≥ 0, ξ (t) = ξ3,t and let τ (t) (resp. τu ) be the value of the counter τ at
the end of time step t (i.e., after the edge on the stream at time t has been processed by
trièst-base) (resp. the value of the counter τu at the end of time step t if there is such
a counter, 0 otherwise). When queried at the end of time t, trièst-base returns ξ (t) τ (t)
(t)
(resp. ξ (t) τu ) as the estimation for the global (resp. local for u ∈ V (t) ) triangle count.
4.1.2. Analysis. We now present the analysis of the estimations computed by trièst-base.
Specifically, we prove their unbiasedness (and their exactness for t ≤ M ) and then show an
exact derivation of their variance and a concentration result. We show the results for the
global counts, but results analogous to those in Thms. 4.2, 4.4, and 4.5 hold for the local
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:7
triangle count for any u ∈ V (t) , replacing the global quantities with the corresponding local
ones. We also compare, theoretically, the variance of trièst-base with that of a fixedprobability edge sampling approach [Lim and Kang 2015], showing that trièst-base has
smaller variance for the vast majority of the stream.
4.1.3. Expectation. We have the following result about the estimations computed by
trièst-base.
Theorem 4.2. We have
ξ (t) τ (t) = τ (t) = |∆(t) | if t ≤ M
h
i
E ξ (t) τ (t) = |∆(t) | if t > M .
The trièst-base estimations are not only unbiased in all cases, but actually exact for
t ≤ M , i.e., for t ≤ M , they are the true global/local number of triangles in G(t) .
To prove Thm. 4.2, we need to introduce a technical lemma. Its proof can be found in
Appendix A.1. We denote with ∆S the set of triangles in GS .
Lemma 4.3. After each call to UpdateCounters, we have τ = |∆S | and τv = |∆Sv |
for any v ∈ VS s.t. |∆Sv | ≥ 1.
From here, the proof of Thm. 4.2 is a straightforward application of Lemma 4.3 for the
case t ≤ M and of that lemma, the definition of expectation, and Lemma 4.1 otherwise.
The complete proof can be found in App. A.1.
4.1.4. Variance. We now analyze the variance of the estimation returned by trièst-base
for t > M (the variance is 0 for t ≤ M .)
Let r(t) be the total number of unordered pairs of distinct triangles from ∆(t) sharing an
(t)
edge,2 and w(t) = |∆2 | − r(t) be the number of unordered pairs of distinct triangles that
do not share any edge.
Theorem 4.4. For any t > M , let f (t) = ξ (t) − 1,
g(t) = ξ (t)
(M − 3)(M − 4)
−1
(t − 3)(t − 4)
and
h(t) = ξ (t)
(M − 3)(M − 4)(M − 5)
− 1 (≤ 0).
(t − 3)(t − 4)(t − 5)
We have:
h
i
Var ξ(t)τ (t) = |∆(t) |f (t) + r(t) g(t) + w(t) h(t).
(1)
In our proofs, we carefully account for the fact that, as we use reservoir sampling [Vitter
1985], the presence of an edge a in S is not independent from the concurrent presence of another edge b in S. This is not the case for samples built using fixed-probability independent
edge sampling, such as mascot [Lim and Kang 2015]. When computing the variance, we
must consider not only pairs of triangles that share an edge, as in the case for independent
edge sampling approaches, but also pairs of triangles sharing no edge, since their respective
presences in the sample are not independent events. The gain is worth the additional sophistication needed in the analysis, as the contribution to the variance by triangles no sharing
edges is non-positive (h(t) ≤ 0), i.e., it reduces the variance. A comparison of the variance
2 Two
distinct triangles can share at most one edge.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:8
Lorenzo De Stefani et al.
of our estimator with that obtained with a fixed-probability independent edge sampling
approach, is discussed in Sect. 4.1.6.
Proof of Thm. 4.4. Assume |∆(t) | > 0, otherwise the estimation is deterministically
(t)
correct and has variance 0 and the thesis holds. Let λ ∈ ∆(t) and δλ be as in the proof
(t)
(t)
of Thm. 4.2. We have Var[δλ ] = ξ − 1 and from this and the definition of variance and
covariance we obtain
i
i
h
h
X (t)
X X
(t)
Var ξ (t) τ (t) = Var
δλ =
Cov δλ , δγ(t)
λ∈∆(t)
=
X
h
λ∈∆(t) γ∈∆(t)
(t)
Var δλ
i
+
λ∈∆(t)
i
h
(t)
Cov δλ , δγ(t)
X
λ,γ∈∆(t)
λ6=γ
= |∆(t) |(ξ (t) − 1) +
h
i
(t)
Cov δλ , δγ(t)
X
λ,γ∈∆(t)
λ6=γ
= |∆(t) |(ξ (t) − 1) +
h
i
h i h i
(t)
(t)
E δλ δγ(t) − E δλ E δγ(t)
X
λ,γ∈∆(t)
λ6=γ
= |∆(t) |(ξ (t) − 1) +
h
i
(t)
E δλ δγ(t) − 1 .
X
(2)
λ,γ∈∆(t)
λ6=γ
Assume now |∆(t) | ≥ 2, otherwise we have r(t) = w(t) = 0 and the thesis holds as the
second term on the r.h.s. of (2) is 0. Let λ and γ be two distinct triangles in ∆(t) . If λ and
(t) (t)
2
γ do not share an edge, we have δλ δγ = ξ (t) ξ (t) = ξ3,t
if all six edges composing λ and
(t) (t)
γ are in S at the end of time step t, and δλ δγ = 0 otherwise. From Lemma 4.1 we then
have that
5
h
i
Y
M −j
1
(t)
(t)
2
2
2
= ξ3,t
E δλ δγ(t) = ξ3,t
Pr δλ δγ(t) = ξ3,t
= ξ3,t
ξ6,t
t−j
j=3
= ξ (t)
(M − 3)(M − 4)(M − 5)
.
(t − 3)(t − 4)(t − 5)
(3)
(t) (t)
2
if all five edges composing
If instead λ and γ share exactly an edge we have δλ δγ = ξ3,t
λ and γ are in S at the end of time step t, and
then have that
(t) (t)
δλ δγ
= 0 otherwise. From Lemma 4.1 we
4
h
i
Y
1
M −j
(t)
(t)
2
2
2
E δλ δγ(t) = ξ3,t
Pr δλ δγ(t) = ξ3,t
= ξ3,t
= ξ3,t
ξ5,t
t−j
j=3
= ξ (t)
(M − 3)(M − 4)
.
(t − 3)(t − 4)
(4)
The thesis follows by combining (2), (3), (4), recalling the definitions of r(t) and w(t) , and
slightly reorganizing the terms.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:9
4.1.5. Concentration. We have the following concentration result on the estimation returned
by trièst-base. Let h(t) denote the maximum number of triangles sharing a single edge in
G(t) .
Theorem 4.5. Let t ≥ 0 and assume |∆(t) | > 0.3 For any ε, δ ∈ (0, 1), let
s
3h(t) + 1
(3h(t) + 1)e
3
−2
Φ = 8ε
ln
.
δ
|∆(t) |
If
1
M ≥ max tΦ 1 + ln2/3 (tΦ) , 12ε−1 + e2 , 25 ,
2
then |ξ (t) τ (t) − |∆(t) || < ε|∆(t) | with probability > 1 − δ.
The roadmap to proving Thm. 4.5 is the following:
(1) we first define two simpler algorithms, named indep and mix. The algorithms use,
respectively, fixed-probability independent sampling of edges and reservoir sampling (but
with a different estimator than the one used by trièst-base);
(2) we then prove concentration results on the estimators of indep and mix. Specifically,
the concentration result for indep uses a result by Hajnal and Szemerédi [1970] on graph
coloring, while the one for mix will depend on the concentration result for indep and on a
Poisson-approximation-like technical result stating that probabilities of events when using
reservoir sampling are close to the probabilities of those events when using fixed-probability
independent sampling;
(3) we then show that the estimates returned by trièst-base are close to the estimates
returned by mix;
(4) finally, we combine the above results and show that, if M is large enough, then the
estimation provided by mix is likely to be close to |∆(t) | and since the estimation computed
by trièst-base is close to that of mix, then it must also be close to |∆(t) |.
Note: for ease of presentation, in the following we use φ(t) to denote the estimation returned
by trièst-base, i.e., φ(t) = ξ (t) τ (t) .
The indep algorithm. The indep algorithm works as follows: it creates a sample Sin by
sampling edges in E (t) independently with a fixed probability p. It estimates the global
number of triangles in G(t) as
(t)
(t)
φin =
τin
,
p3
(t)
where τin is the number of triangles in Sin . This is for example the approach taken by
mascot-c [Lim and Kang 2015].
The mix algorithm. The mix algorithm works as follows: it uses reservoir sampling (like
trièst-base) to create a sample Smix of M edges from E (t) , but uses a different estimator
than the one used by trièst-base. Specifically, mix uses
3
t
(t)
τ (t)
φmix =
M
3 If
|∆(t) | = 0, our algorithms correctly estimate 0 triangles.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:10
Lorenzo De Stefani et al.
as an estimator for |∆(t) |, where τ (t) is, as in trièst-base, the number of triangles in GS
(t)
(trièst-base uses φ(t) = M t(t−1)(t−2)
as an estimator.)
(M −1)(M −2) τ
We call this algorithm mix because it uses reservoir sampling to create the sample, but
computes the estimate as if it used fixed-probability independent sampling, hence in some
sense it “mixes” the two approaches.
Concentration results for indep and mix. We now show a concentration result for indep.
Then we show a technical lemma (Lemma 4.7) relating the probabilities of events when
using reservoir sampling to the probabilities of those events when using fixed-probability
independent sampling. Finally, we use these results to show that the estimator used by mix
is also concentrated (Lemma 4.9).
Lemma 4.6. Let t ≥ 0 and assume |∆(t) | > 0.4 For any ε, δ ∈ (0, 1), if
s
p≥
3
2ε−2 ln
3h(t) + 1
δ
3h(t) + 1
|∆(t) |
(5)
then
(t)
Pr |φin − ∆(t) || < ε|∆(t) | > 1 − δ .
Proof. Let H be a graph built as follows: H has one node for each triangle in G(t) and
there is an edge between two nodes in H if the corresponding triangles in G(t) share an edge.
By this construction, the maximum degree in H is 3h(t) . Hence by the Hajanal-Szeméredi’s
theorem [Hajnal and Szemerédi 1970] there is a proper coloring of H with at most 3h(t) + 1
(t)
|
colors such that for each color there are at least L = 3h|∆(t) +1
nodes with that color.
(t)
Assign an arbitrary numbering to the triangles of G (and, therefore, to the nodes of
H) and let Xi be a Bernoulli random variable, indicating whether the triangle i in G(t)
is in the sample at time t. From the properties of independent sampling of edges we have
Pr(Xi = 1) = p3 for any triangle i. For any color c of the coloring of H, let Xc be the set
of r.v.’s Xi such that the node i in H has color c. Since the coloring of H which we are
considering is proper, the r.v.’s in Xc are independent, as they correspond to triangles which
do not share any edge and edges are sampled independent of each other. Let Yc be the sum
of the r.v.’s in Xc . The r.v. Yc has a binomial distribution with parameters |Xc | and p3t . By
the Chernoff bound for binomial r.v.’s, we have that
Pr |p−3 Yc − |Xc || > ε|Xc | < 2 exp −ε2 p3 |Xc |/2
< 2 exp −ε2 p3 L/2
δ
≤ (t)
,
3h + 1
where the last step comes from the requirement in (5).Then by applying the union bound
over all the (at most) 3h(t) + 1 colors we get
Pr(∃ color c s.t. |p−3 Yc − |Xc || > ε|Xc |) < δ .
4 For
|∆(t) | = 0, indep correctly and deterministically returns 0 as the estimation.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
Since φin (t) = p−3
X
color
XX:11
Yc , from the above equation we have that, with probability at least
c
1 − δ,
(t)
|φin − |∆(t) || ≤
X
color
≤
c
X
color
p−3 Yc −
c
X
color
|Xc |
c
|p−3 Yc − |Xc || ≤
X
color
ε|Xc | ≤ ε|∆(t) | .
c
The above result is of independent interest and can be used, for example, to give concentration bounds to the estimation computed by mascot-c [Lim and Kang 2015].
We remark that we can not use the same approach from Lemma 4.6 to show a concentration result for trièst-base because it uses reservoir sampling, hence the event of having a
triangle a in S and the event of having another triangle b in S are not independent.
We can however show the following general result, similar in spirit to the well-know
Poisson approximation of balls-and-bins processes [Mitzenmacher and Upfal 2005]. Its proof
can be found in App. A.1.
Fix the parameter M and a time t > M . Let Smix be a sample of M edges from E (t)
obtained through reservoir sampling (as mix would do), and let Sin be a sample of the edges
in E (t) obtained by sampling edges independently with probability M/t (as indep would
do). We remark that the size of Sin is in [0, t] but not necessarily M .
Lemma 4.7. Let f : 2E
of E (t) to {0, 1} . We have
(t)
→ {0, 1} be an arbitrary binary function from the powerset
√
Pr (f (Smix ) = 1) ≤ e M Pr (f (Sin ) = 1) .
(t)
We now use the above two lemmas to show that the estimator φmix computed by mix is
concentrated. We will first need the following technical fact.
Fact 4.8. For any x ≥ 5, we have
ln x(1 + ln2/3 x) ≤ ln2 x .
Lemma 4.9. Let t ≥ 0 and assume |∆(t) | < 0. For any ε, δ ∈ (0, 1), let
(t)
+1
3h(t) + 1
−2 3h
Ψ = 2ε
ln e
.
δ
|∆(t) |
If
√
√
1
3
3
M ≥ max t Ψ 1 + ln2/3 t Ψ
, 25
2
then
(t)
Pr |φmix − |∆(t) || < ε|∆(t) | ≥ 1 − δ .
Proof. For any S ⊆ E (t) let τ (S) be the number of triangles in S, i.e., the number of
(t)
triplets of edges in S that compose a triangle in G(t) . Define the function g : 2E → R as
3
t
g(S) =
τ (S) .
M
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:12
Lorenzo De Stefani et al.
Assume that we run indep with p = M/t, and let Sin ⊆ E (t) be the sample built by indep
(through independent sampling with fixed probability p). Assume also that we run mix with
parameter M , and let Smix be the sample built by mix (through reservoir sampling with a
(t)
(t)
reservoir of size M ). We have that φin = g(Sin ) and φmix = g(Smix ). Define now the binary
(t)
function f : 2E → {0, 1} as
1 if |g(S) − |∆(t) || > ε|∆(t) |
.
f (S) =
0 otherwise
We now show that, for M as in the hypothesis, we have
s
√ 3h(t) + 1
3h(t) + 1
p ≥ 3 2ε−2
ln
e
M
.
δ
|∆(t) |
(6)
Assume for now that the above is true. From this, using Lemma 4.6 and the above fact
about g we get that
δ
(t)
Pr |φin − |∆(t) || > ε|∆(t) | = Pr (f (Sin ) = 1) < √
.
e M
From this and Lemma 4.7, we get that
Pr (f (Smix ) = 1) ≤ δ
which, from the definition of f and the properties of g, is equivalent to
(t)
Pr |φmix − |∆(t) || > ε|∆(t) | ≤ δ
and the proof is complete. All that is left is to show that (6) holds for M as in the hypothesis.
Since p = M/t, we have that (6) holds for
(t)
√
+1
3h(t) + 1
3
3 −2 3h
M ≥ t 2ε
ln
Me
δ
|∆(t) |
(t)
(t)
3h + 1
1
+1
3 −2 3h
= t 2ε
ln e
+ ln M
.
(7)
δ
2
|∆(t) |
We now show that (7) holds.
√
√
√
Let A = t 3 Ψ and let B = t 3 Ψ ln2/3 t 3 Ψ . We now show that A3 + B 3 is greater or
equal to the r.h.s. of (7), hence M 3 = (A + B)3 > A3 + B 3 must also be greater or equal to
the r.h.s. of (7), i.e., (7) holds. This really reduces to show that
B 3 ≥ t3 2ε−2
3h(t) + 1 1
ln M
|∆(t) | 2
(8)
as the r.h.s.of (7) can be written as
A3 + t3 2ε−2
3h(t) + 1 1
ln M .
|∆(t) | 2
We actually show that
1
B 3 ≥ t3 Ψ ln M
2
(9)
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:13
which implies (8) which, as discussed, in turn implies (7). Consider the ratio
√
√
√
3
2
1 3
B3
ln2 (t 3 Ψ)
ln2 (t 3 Ψ)
2 t Ψ ln (t Ψ)
√ .
≥
=
=
√
ln M
t3 Ψ 12 ln M
t3 Ψ 12 ln M
ln t 3 Ψ 1 + ln2/3 t 3 Ψ
(10)
√
We now show that t 3 Ψ ≥ 5. By the assumptions t > M ≥ 25 and by
√
√
t
3
t Ψ≥ p
≥ t
3
|∆(t) |
which holds because |∆(t)√| ≤ t3/2 (in a graph with t edges there can not be more than t3/2
triangles) we have that t 3 Ψ ≥ 5. Hence Fact 4.8 holds and we can write, from (10):
√
√
ln2 (t 3 Ψ)
ln2 (t 3 Ψ)
√
√ ≥
√ ≥ 1,
ln t 3 Ψ 1 + ln2/3 t 3 Ψ
ln2 t 3 Ψ
which proves (9), and in cascade (8), (7), (6), and the thesis.
Relationship between trièst-base and mix. When both trièst-base and mix use a
(t
sample of size M , their respective estimators φ(t) and φmix are related as discussed in the
following result, whose straightforward proof is deferred to App. A.1.
Lemma 4.10. For any t > M we have
(t)
(t)
φ(t) − φmix ≤ φmix
4
.
M −2
Tying everything together. Finally we can use the previous lemmas to prove our concentration result for trièst-base.
Proof of Thm. 4.5. For M as in the hypothesis we have, from Lemma 4.9, that
(t)
Pr φmix ≤ (1 + ε/2)|∆(t) | ≥ 1 − δ .
(t)
(t)
Suppose the event φmix ≤ (1 + ε/2)|∆(t) | (i.e., |φmix − |∆(t) || ≤ ε|∆(t) |/2) is indeed verified.
From this and Lemma 4.10 we have
ε (t)
4
6
(t)
|φ(t) − φmix | ≤ 1 +
|∆ |
≤ |∆(t) |
,
2
M −2
M −2
where the last inequality follows from the fact that ε < 1. Hence, given that M ≥ 12ε−1 +
e2 ≥ 12ε−1 + 2, we have
ε
(t)
|φ(t) − φmix | ≤ |∆(t) | .
2
Using the above, we can then write:
(t)
(t)
|φ(t) − |∆(t) || = |φ(t) − φmix + φmix − |∆(t) ||
(t)
(t)
≤ |φ(t) − φmix | + |φmix − |∆(t) ||
ε
ε
≤ |∆(t) | + |∆(t) | = ε|∆(t) |
2
2
which completes the proof.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:14
Lorenzo De Stefani et al.
4.1.6. Comparison with fixed-probability approaches. We now compare the variance of trièstbase to the variance of the fixed probability sampling approach mascot-c [Lim and Kang
2015], which samples edges independently with a fixed probability p and uses p−3 |∆S | as
the estimate for the global number of triangles at time t. As shown by Lim and Kang [2015,
Lemma 2], the variance of this estimator is
Var[p−3 |∆S |] = |∆(t) |f¯(p) + r(t) ḡ(p),
where f¯(p) = p−3 − 1 and ḡ(p) = p−1 − 1.
Assume that we give mascot-c the additional information that the stream has finite
length T , and assume we run mascot-c with p = M/T so that the expected sample size at
(t)
the end of the stream is M .5 Let Vfix be the resulting variance of the mascot-c estimator
at time t, and let V(t) be the variance of our estimator at time t (see (1)). For t ≤ M ,
(t)
V(t) = 0, hence V(t) ≤ Vfix .
For M < t < T , we can show the following result. Its proof is more tedious than interesting
so we defer it to App. A.1.
8α
Lemma 4.11. Let 0 < α < 1 be a constant. For any constant M > max( 1−α
, 42) and
(t)
any t ≤ αT we have V(t) < Vfix .
For example, if we set α = 0.99 and run trièst-base with M ≥ 400 and mascot-c with
p = M/T , we have that trièst-base has strictly smaller variance than mascot-c for 99%
of the stream.
(t)
What about t = T ? The difference between the definitions of Vfix and V(t) is in the
¯
presence of f (M/T ) instead of f (t) (resp. ḡ(M/T ) instead of g(t)) as well as the additional
term w(t) h(M, t) ≤ 0 in our V(t) . Let M (T ) be an arbitrary slowly increasing function of
¯
(T )/T )
)/T )
= limT →∞ ḡ(Mg(T
= 1, hence,
T . For T → ∞ we can show that limT →∞ f (Mf(T
(T )
)
(T )
informally, V(T ) → Vfix , for T → ∞.
A similar discussion also holds for the method we present in Sect. 4.2, and explains the
results of our experimental evaluations, which shows that our algorithms have strictly lower
(empirical) variance than fixed probability approaches for most of the stream.
4.1.7. Update time. The time to process an element of the stream is dominated by the computation of the shared neighborhood Nu,v in UpdateCounters. A Mergesort-based
algorithm for the intersection requires O (deg(u) + deg(v)) time, where the degrees are
w.r.t. the graph GS . By storing the neighborhood of each vertex in a Hash Table (resp. an
AVL tree), the update time can be reduced to O(min{deg(v), deg(u)}) (resp. amortized
time O(min{deg(v), deg(u)} + log max{deg(v), deg(u)})).
4.2. Improved insertion algorithm – trièst-impr
trièst-impr is a variant of trièst-base with small modifications that result in higherquality (i.e., lower variance) estimations. The changes are:
(1) UpdateCounters is called unconditionally for each element on the stream, before
the algorithm decides whether or not to insert the edge into S. W.r.t. the pseudocode
in Alg. 1, this change corresponds to moving the call to UpdateCounters on line 6 to
before the if block. mascot [Lim and Kang 2015] uses a similar idea, but trièst-impr is
significantly different as trièst-impr allows edges to be removed from the sample, since it
uses reservoir sampling.
5 We are giving mascot-c a significant advantage: if only space M were available, we should run mascot-c
with a sufficiently smaller p0 < p, otherwise there would be a constant probability that mascot-c would
run out of memory.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:15
(2) trièst-impr never decrements the counters when an edge is removed from S.
W.r.t. the pseudocode in Alg. 1, we remove the call to UpdateCounters on line 13.
(3) UpdateCounters performs a weighted increase of the counters using η (t) =
max{1, (t − 1)(t − 2)/(M (M − 1))} as weight. W.r.t. the pseudocode in Alg. 1, we replace “1” with η (t) on lines 19–22 (given change 2 above, all the calls to UpdateCounters
have • = +).
The resulting pseudocode for trièst-impr is presented in Alg. 2.
ALGORITHM 2 trièst-impr
Input: Insertion-only edge stream Σ, integer M ≥ 6
1: S ← ∅, t ← 0, τ ← 0
2: for each element (+, (u, v)) from Σ do
3:
t←t+1
4:
UpdateCounters(u, v)
5:
if SampleEdge((u, v), t) then
6:
S ← S ∪ {(u, v)}
7: function SampleEdge((u, v), t)
8:
if t ≤ M then
9:
return True
10:
else if FlipBiasedCoin( M
) = heads then
t
11:
(u0 , v 0 ) ← random edge from S
12:
S ← S \ {(u0 , v 0 )}
13:
return True
14:
return False
15: function UpdateCounters(u, v)
S
16:
Nu,v
← NuS ∩ NvS
17:
η = max{1, (t − 1)(t − 2)/(M (M − 1))}
S
18:
for all c ∈ Nu,v
do
19:
τ ←τ +η
20:
τc ← τc + η
21:
τu ← τu + η
22:
τv ← τv + η
Counters. If we are interested only in estimating the global number of triangles in G(t) ,
trièst-impr needs to maintain only the counter τ and the edge sample S of size M , so
it still requires space O(M ). If instead we are interested in estimating the local triangle
counts, at any time t trièst-impr maintains (non-zero) local counters only for the nodes u
such that at least one triangle with a corner u has been detected by the algorithm up until
time t. The number of such nodes may be greater than O(M ), but this is the price to pay
to obtain estimations with lower variance (Thm. 4.13).
4.2.1. Estimation. When queried for an estimation, trièst-impr returns the value of the
corresponding counter, unmodified.
4.2.2. Analysis. We now present the analysis of the estimations computed by trièst-impr,
showing results involving their unbiasedness, their variance, and their concentration around
their expectation. Results analogous to those in Thms. 4.12, 4.13, and 4.15 hold for the
local triangle count for any u ∈ V (t) , replacing the global quantities with the corresponding
local ones.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:16
Lorenzo De Stefani et al.
4.2.3. Expectation. As in trièst-base, the estimations by trièst-impr are exact at time
t ≤ M and unbiased for t > M . The proof of the following theorem follows the same steps
as the one for Thm 4.2, and can be found in App. A.2.
Theorem 4.12. We have τ (t) = |∆(t) | if t ≤ M and E τ (t) = |∆(t) | if t > M .
4.2.4. Variance. We now show an upper bound to the variance of the trièst-impr estimations for t > M . The proof relies on a very careful analysis of the covariance of two
triangles which depends on the order of arrival of the edges in the stream (which we assume
to be adversarial). For any λ ∈ ∆(t) we denote as tλ the time at which the last edge of
λ is observed on the stream. Let z (t) be the number of unordered pairs (λ, γ) of distinct
triangles in G(t) that share an edge g and are such that:
(1) g is neither the last edge of λ nor γ on the stream; and
(2) min{tλ , tγ } > M + 1.
Theorem 4.13. Then, for any time t > M , we have
h
i
t−1−M
Var τ (t) ≤ |∆(t) |(η (t) − 1) + z (t)
.
(11)
M
The bound to the variance presented in (11) is extremely pessimistic and loose. Specifically,
(t)
it does not contain the negative contribution to the variance given by the |∆2 | − z (t)
triangles that do not satisfy the requirements in the definition of z (t) . Among these pairs
there are, for example, all pairs of triangles not sharing any edge, but also many pairs of
triangles that share an edge, as the definition of z (t) consider only a subsets of these. All
these pairs would give a negative contribution to the variance, i.e., decrease the r.h.s. of (11),
whose more correct form would be
(t)
|∆ |
t−1−M
− z (t) ωM,t
+
|∆(t) |(η (t) − 1) + z (t)
2
M
where ωM,t is (an upper bound to) the minimum negative contribution of a pair of triangles
that do not satisfy the requirements in the definition of z (t) . Sadly, computing informative
upper bounds to ωM,t is not straightforward, even in the restricted setting where only pairs
of triangles not sharing any edge are considered.
To prove Thm. 4.13 we first need Lemma 4.14, whose proof is deferred to App. A.2.
For any time step t and any edge e ∈ E (t) , we denote with te the time step at which e is
on the stream. For any λ ∈ ∆(t) , let λ = (`1 , `2 , `3 ), where the edges are numbered in order
of appearance on the stream. We define the event Dλ as the event that `1 and `2 are both
in the edge sample S at the end of time step tλ − 1.
Lemma 4.14. Let λ = (`1 , `2 , `3 ) and γ = (g1 , g2 , g3 ) be two disjoint triangles, where
the edges are numbered in order of appearance on the stream, and assume, w.l.o.g., that the
last edge of λ is on the stream before the last edge of γ. Then
Pr(Dγ | Dλ ) ≤ Pr(Dγ ) .
We can now prove Thm. 4.13.
Proof of Thm. 4.13. Assume |∆(t) | > 0, otherwise trièst-impr estimation is deterministically correct and has variance 0 and the thesis holds. Let λ ∈ ∆(t) and let δλ be a
random variable that takes value ξ2,tλ −1 if both `1 and `2 are in S at the end of time step
tλ − 1, and 0 otherwise. Since
Var [δλ ] = ξ2,tλ −1 − 1 ≤ ξ2,t−1 ,
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:17
we have:
h
i
Var τ (t) = Var
X
λ∈∆(t)
=
X
λ∈∆(t)
X
δλ =
Var [δλ ] +
X
Cov [δλ , δγ ]
λ∈∆(t) γ∈∆(t)
X
λ,γ∈∆
λ6=γ
≤ |∆(t) |(ξ2,t−1 − 1) +
Cov [δλ , δγ ]
(t)
X
(E[δλ δγ ] − E[δλ ]E[δγ ])
λ,γ∈∆(t)
λ6=γ
≤ |∆(t) |(ξ2,t−1 − 1) +
X
λ,γ∈∆
λ6=γ
(E[δλ δγ ] − 1) .
(12)
(t)
For any λ ∈ ∆(t) define qλ = ξ2,tλ −1 . Assume now |∆(t) | ≥ 2, otherwise we have r(t) =
w = 0 and the thesis holds as the second term on the r.h.s. of (12) is 0. Let now λ and γ
be two distinct triangles in ∆(t) (hence t ≥ 5). We have
(t)
E [δλ δγ ] = qλ qγ Pr (δλ δγ = qλ qγ )
The event “δλ δγ = qλ qγ ” is the intersection of events Dλ ∩ Dγ , where Dλ is the event that
the first two edges of λ are in S at the end of time step tλ − 1, and similarly for Dγ . We
now look at Pr(Dλ ∩ Dγ ) in the various possible cases.
Assume that λ and γ do not share any edge, and, w.l.o.g., that the third (and last)
edge of λ appears on the stream before the third (and last) edge of γ, i.e., tλ < tγ . From
Lemma 4.14 and Lemma 4.1 we then have
1
Pr(Dλ ∩ Dγ ) = Pr(Dγ |Dλ ) Pr(Dλ ) ≤ Pr(Dγ ) Pr(Dλ ) ≤
.
qλ qγ
Consider now the case where λ and γ share an edge g. W.l.o.g., let us assume that tλ ≤ tγ
(since the shared edge may be the last on the stream both for λ and for γ, we may have
tλ = tγ ). There are the following possible sub-cases :
g is the last on the stream among all the edges of λ and γ. In this case we have tλ =
tγ . The event “Dλ ∩ Dγ ” happens if and only if the four edges that, together with g,
compose λ and γ are all in S at the end of time step tλ − 1. Then, from Lemma 4.1 we
have
Pr(Dλ ∩ Dγ ) =
1 (M − 2)(M − 3)
1
M (M − 1)
1
1
≤
≤
≤
.
ξ4,tλ −1
qλ (tλ − 3)(tλ − 4)
qλ (tλ − 1)(tλ − 2)
qλ qγ
g is the last on the stream among all the edges of λ and the first among all the edges of γ.
In this case, we have that Dλ and Dγ are independent. Indeed the fact that the first
two edges of λ (neither of which is g) are in S when g arrives on the stream has no
influence on the probability that g and the second edge of γ are inserted in S and are
not evicted until the third edge of γ is on the stream. Hence we have
Pr(Dλ ∩ Dγ ) = Pr(Dγ ) Pr(Dλ ) =
1
.
qλ qγ
g is the last on the stream among all the edges of λ and the second among all the edges of γ.
In this case we can follow an approach similar to the one in the proof for Lemma 4.14
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:18
Lorenzo De Stefani et al.
and have that
1
.
qλ qγ
The intuition behind this is that if the first two edges of λ are in S when g is on the
stream, their presence lowers the probability that the first edge of γ is in S at the same
time, and hence that the first edge of γ and g are in S when the last edge of γ is on the
stream.
g is not the last on the stream among all the edges of λ. There are two situations to
consider, or better, one situation and all other possibilities. The situation we consider
is that
(1) g is the first edge of γ on the stream; and
(2) the second edge of γ to be on the stream is on the stream at time t2 > tλ .
Suppose this is the case. Recall that if Dλ is verified, than we know that g is in S at
the beginning of time step tλ . Define the following events:
— E1 : “the set of edges evicted from S between the beginning of time step tλ and the
beginning of time step t2 does not contain g.”
— E2 : “the second edge of γ, which is on the stream at time t2 , is inserted in S and
the edge that is evicted is not g.”
— E3 : “the set of edges evicted from S between the beginning of time step t2 + 1 and
the beginning of time step tγ does not contain either g or the second edge of γ.”
We can then write
Pr(Dλ ∩ Dγ ) ≤ Pr(Dγ ) Pr(Dλ ) ≤
Pr(Dγ | Dλ ) = Pr(E1 | Dλ ) Pr(E2 | E1 ∩ Dλ ) Pr(E3 | E2 ∩ E1 ∩ Dλ ) .
We now compute the probabilities on the r.h.s., where we denote with 1t2 >M (1) the
function that has value 1 if t2 > M , and value 0 otherwise:
tY
2 −1
M
M M −1
1−
Pr(E1 | Dλ ) =
+
j
j
M
j=max{tλ ,M +1}
=
tY
2 −1
j=max{tλ ,M +1}
j−1
max{tλ − 1, M }
=
;
j
max{M, t2 − 1}
M − 1t2 >M (1)
M − 1t2 >M (1)
M
=
;
max{t2 , M }
M
max{t2 , M }
tγ −1
Y
M
M M −2
1−
Pr(E3 | E2 ∩ E1 ∩ Dλ ) =
+
j
j
M
Pr(E2 | E1 ∩ Dλ ) =
j=max{t2 +1,M +1}
tγ −1
=
Y
j=max{t2 +1,M +1}
j−2
max{t2 , M } max{t2 − 1, M − 1}
=
.
j
max{tγ − 2, M − 1} max{tγ − 1, M }
Hence, we have
Pr(Dγ | Dλ ) =
max{tλ − 1, M }(M − 1t2 >M (1)) max{t2 − 1, M − 1}
.
max{M, t2 − 1} max{(tγ − 2)(tγ − 1), M (M − 1)}
With a (somewhat tedious) case analysis we can verify that
Pr(Dγ | Dλ ) ≤
1 max{M, tλ − 1}
.
qγ
M
Consider now the complement of the situation we just analyzed. In this case, two edges
of γ, that is, g and another edge h are on the stream before time tλ , in some non-relevant
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:19
order (i.e., g could be the first or the second edge of γ on the stream). Define now the
following events:
— E1 : “h and g are both in S at the beginning of time step tλ .”
— E2 : “the set of edges evicted from S between the beginning of time step tλ and the
beginning of time step tγ does not contain either g or h.”
We can then write
Pr(Dγ | Dλ ) = Pr(E1 | Dλ ) Pr(E2 | E1 ∩ Dλ ) .
If tλ ≤ M + 1, we have that Pr(E1 | Dλ ) = 1. Consider instead the case tλ > M + 1. If
Dλ is verified, then both g and the other edge of λ are in S at the beginning of time step
tλ . At this time, all subsets of E (tλ −1) of size M and containing both g and theother
tλ −3
edge of λ have an equal probability of being S, from Lemma A.1. There are M
−2 such
tλ −4
sets. Among these, there are M
sets
that
also
contain
h.
Therefore,
if
t
>
M + 1,
λ
−3
we have
tλ −4
M −2
M −3
Pr(E1 | Dλ ) = tλ −3 =
.
tλ − 3
M −2
Considering what we said before for the case tλ ≤ M + 1, we then have
M −2
Pr(E1 | Dλ ) = min 1,
.
tλ − 3
We also have
tγ −1
Pr(E2 | E1 ∩ Dλ ) =
Y
j=max{tλ ,M +1}
j−2
max{(tλ − 2)(tλ − 1), M (M − 1)}
=
.
j
max{(tγ − 2)(tγ − 1), M (M − 1)}
Therefore,
M −2
Pr(Dγ | Dλ ) = min 1,
tλ − 3
max{(tλ − 2)(tλ − 1), M (M − 1)}
.
max{(tγ − 2)(tγ − 1), M (M − 1)}
With a case analysis, one can show that
Pr(Dγ | Dλ ) ≤
1 max{M, tλ − 1}
.
qγ
M
To recap we have the following two scenarios when considering two distinct triangles γ and
λ:
(1) if λ and γ share an edge and, assuming w.l.o.g. that the third edge of λ is on the
stream no later than the third edge of γ, and the shared edge is neither the last among
all edges of λ to appear on the stream nor the last among all edges of γ to appear on the
stream, then we have
Cov[δλ , δγ ] ≤ E[δλ δγ ] − 1 = qλ qγ Pr(δλ δγ = qλ qγ ) − 1
1 max{M, tλ − 1}
max{M, tλ − 1}
t−1−M
≤ qλ qγ
−1≤
−1≤
;
qλ qγ
M
M
M
where the last inequality follows from the fact that tλ ≤ t and t − 1 ≥ M .
For the pairs (λ, γ) such that tλ ≤ M + 1, we have max{M, tλ − 1}/M = 1 and therefore
Cov[δλ , δγ ] ≤ 0. We should therefore only consider the pairs for which tλ > M + 1. Their
number is given by z (t) .
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:20
Lorenzo De Stefani et al.
(2) in all other cases, including when γ and λ do not share an edge, we have E[δλ δγ ] ≤ 1,
and since E[δλ ]E[δγ ] = 1, we have
Cov[δλ , δγ ] ≤ 0 .
Hence, we can bound
X
Cov[δλ , δγ ] ≤ z (t)
λ,γ∈∆(t)
λ6=γ
t−1−M
M
and the thesis follows by combining this into (12).
4.2.5. Concentration. We now show a concentration result on the estimation of trièstimpr, which relies on Chebyshev’s inequality [Mitzenmacher and Upfal 2005, Thm. 3.6]
and Thm. 4.13.
Theorem 4.15. Let t ≥ 0 and assume |∆(t) | > 0. For any ε, δ ∈ (0, 1), if
(s
)
2(t − 1)(t − 2) 1 1
2z (t) (t − 1)
M > max
+ + , 2 (t) 2
4 2 δε |∆ | + 2z (t)
δε2 |∆(t) | + 2
then |τ (t) − |∆(t) || < ε|∆(t) | with probability > 1 − δ.
Proof. By Chebyshev’s inequality it is sufficient to prove that
Var[τ (t) ]
<δ .
ε2 |∆(t) |2
We can write
Var[τ (t) ]
1
≤ 2 (t)
2
(t)
2
ε |∆ |
ε |∆ |
(t) t − 1 − M
(η(t) − 1) + z
.
M |∆(t) |
Hence it is sufficient to impose the following two conditions:
Condition 1.
δ
η(t) − 1
> 2 (t)
2
ε |∆ |
(t − 1)(t − 2) − M (M − 1)
1
,
> 2 (t)
M (M − 1)
ε |∆ |
(13)
which is verified for:
M (M − 1) >
As t > M , we have
2(t − 1)(t − 2)
.
δε2 |∆(t) | + 2
2(t−1)(t−2)
δε2 |∆(t) |+2
> 0. The condition in (13) is thus verified for:
s
!
1
2(t − 1)(t − 2)
M>
4 2 (t)
+1+1
2
δε |∆ | + 2
Condition 2.
δ
t−1−M
> z (t)
,
2
M ε2 |∆(t) |2
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:21
which is verified for:
2z (t) (t − 1)
.
+ 2z (t)
M>
δε2 |∆(t) |2
The theorem follows.
In Thms. 4.13 and 4.15, it is possible to replace the value z (t) with the more interpretable
r(t) , which is agnostic to the order of the edges on the stream but gives a looser upper
bound to the variance.
4.3. Fully-dynamic algorithm – trièst-fd
trièst-fd computes unbiased estimates of the global and local triangle counts in a fullydynamic stream where edges are inserted/deleted in any arbitrary, adversarial order. It
is based on random pairing (RP) [Gemulla et al. 2008], a sampling scheme that extends
reservoir sampling and can handle deletions. The idea behind the RP scheme is that edge
deletions seen on the stream will be “compensated” by future edge insertions. Following
RP, trièst-fd keeps a counter di (resp. do ) to keep track of the number of uncompensated
edge deletions involving an edge e that was (resp. was not) in S at the time the deletion
for e was on the stream.
When an edge deletion for an edge e ∈ E (t−1) is on the stream at the beginning of time
step t, then, if e ∈ S at this time, trièst-fd removes e from S (effectively decreasing the
number of edges stored in the sample by one) and increases di by one. Otherwise, it just
increases do by one. When an edge insertion for an edge e 6∈ E (t−1) is on the stream at
the beginning of time step t, if di + do = 0, then trièst-fd follows the standard reservoir
sampling scheme. If |S| < M , then e is deterministically inserted in S without removing
any edge from S already in S, otherwise it is inserted in S with probability M/t, replacing
an uniformly-chosen edge already in S. If instead di + do > 0, then e is inserted in S
with probability di /(di + do ); since it must be di > 0, then it must be |S| < M and no
edge already in S needs to be removed. In any case, after having handled the eventual
insertion of e into S, the algorithm decreases di by 1 if e was inserted in S, otherwise it
decreases do by 1. trièst-fd also keeps track of s(t) = |E (t) | by appropriately incrementing
or decrementing a counter by 1 depending on whether the element on the stream is an
edge insertion or deletion. The pseudocode for trièst-fd is presented in Alg. 3 where the
UpdateCounters procedure is the one from Alg. 1.
4.3.1. Estimation. We denote as M (t) the size of S at the end of time t (we always have
(t)
(t)
M (t) ≤ M ). For any time t, let di and do be the value of the counters di and do at the
(t)
(t)
end of time t respectively, and let ω (t) = min{M, s(t) + di + do }. Define
κ(t) = 1 −
2 (t) (t)
(t) (t)
(t)
(t)
X
s + d + do
s
d + do
i
j=0
j
ω (t) − j
i
ω (t)
For any three positive integers a, b, c s.t. a ≤ b ≤ c, define6
ψa,b,c =
6 We
follow the convention that
0
0
.
a−1
Y c−i
c
c−a
.
=
b
b−a
b−i
i=0
= 1.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
.
(14)
XX:22
Lorenzo De Stefani et al.
ALGORITHM 3 trièst-fd
Input: Fully-dynamic edge stream Σ, integer M ≥ 6
1: S ← ∅, di ← 0, do ← 0, t ← 0, s ← 0
2: for each element (•, (u, v)) from Σ do
3:
t←t+1
4:
s←s•1
5:
if • = + then
6:
if SampleEdge (u, v) then
7:
UpdateCounters(+, (u, v))
. UpdateCounters is defined as in Alg. 1.
8:
else if (u, v) ∈ S then
9:
UpdateCounters(−, (u, v))
10:
S ← S \ {(u, v)}
11:
di ← di + 1
12:
else do ← do + 1
13: function SampleEdge(u, v)
14:
if do + di = 0 then
15:
if |S| < M then
16:
S ← S ∪ {(u, v)}
17:
return True
18:
else if FlipBiasedCoin( M
) = heads then
t
19:
Select (z, w) uniformly at random from S
20:
UpdateCounters(−, (z, w))
21:
S ← (S \ {(z, w)}) ∪ {(u, v)}
22:
return True
di
23:
else if FlipBiasedCoin di +d
= heads then
o
24:
S ← S ∪ {(u, v)}
25:
di ← di − 1
26:
return True
27:
else
28:
do ← do − 1
29:
return False
When queried at the end of time t, for an estimation of the global number of triangles,
trièst-fd returns
(
0 if M (t) < 3
(t)
(t) (t)
(t)
ρ = τ (t)
(s −1)(s(t) −2)
ψ (t) (t) = κτ (t) Ms(t) (M
(t) −1)(M (t) −2) othw.
κ(t) 3,M ,s
trièst-fd can keep track of κ(t) during the execution, each update of κ(t) taking time O(1).
Hence the time to return the estimations is still O(1).
4.3.2. Analysis. We now present the analysis of the estimations computed by trièst-impr,
showing results involving their unbiasedness, their variance, and their concentration around
their expectation. Results analogous to those in Thms. 4.16, 4.17, and 4.18 hold for the
local triangle count for any u ∈ V (t) , replacing the global quantities with the corresponding
local ones.
4.3.3. Expectation. Let t∗ be the first t ≥ M + 1 such that |E (t) | = M + 1, if such a time
step exists (otherwise t∗ = +∞).
Theorem 4.16. We have ρ(t) = |∆(t) | for all t < t∗ , and E ρ(t) = |∆(t) | for t ≥ t∗ .
The proof, deferred to App. A.1, relies on properties of RP and on the definitions of κ(t)
and ρ(t) . Specifically, it uses Lemma A.6, which is the correspondent of Lemma 4.1 but for
RP, and some additional technical lemmas (including an equivalent of Lemma 4.3 but for
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:23
RP) and combine them using the law of total expectation by conditioning on the value of
M ( t).
4.3.4. Variance
Theorem 4.17. Let t > t∗ s.t. |∆(t) | > 0 and s(t) ≥ M . Suppose we have d(t) =
(t)
+ di ≤ αs(t) total unpaired deletions at time t, with 0 ≤ α < 1. If M ≥ 2√α10 −α 7 ln s(t)
for some α < α0 < 1, we have:
h i
Var ρ(t) ≤ (κ(t) )−2 |∆(t) | ψ3,M (1−α0 ),s(t) − 1 + (κ(t) )−2 2
−1
2
+ (κ(t) )−2 r(t) ψ3,M
(1−α0 ),s(t) ψ5,M (1−α0 ),s(t) − 1
(t)
do
The proof of Thm. 4.17 is deferred to App. A.3. It uses two results on the variance of
ρ(t) conditioned on a specific value of M (t) (Lemmas A.9 and A.10), and an analysis of the
probability distribution of M (t) (Lemma A.11 and Corollary A.12). These results are then
combined using the law of total variance.
4.3.5. Concentration. The following result relies on Chebyshev’s inequality and on
Thm. 4.17, and the proof (reported in App. A.3) follows the steps similar to those in
the proof for Thm. 4.13.
(t)
(t)
Theorem 4.18. Let t ≥ t∗ s.t. |∆(t) | > 0 and s(t) ≥ M . Let d(t) = do + di
for some 0 ≤ α < 1. For any ε, δ ∈ (0, 1), if for some α < α0 < 1
(
1
M > max √ 0
7 ln s(t) ,
a −α
v
u
(t) (s(t) − 1)(s(t) − 2)
u
2s
3
+ 2 ,
(1 − α0 )−1 t
(t)
δε2 |∆(t) |(κ(t) )2 + 2 |∆|∆(t)|−2
|
)
(1 − α0 )−1
r(t) s(t)
3
δε2 |∆(t) |2 (κ(t) )−2 + 2r(t)
≤ αs(t)
then |ρ(t) − |∆(t) || < ε|∆(t) | with probability > 1 − δ.
4.4. Counting global and local triangles in multigraphs
We now discuss how to extend trièst to approximate the local and global triangle counts
in multigraphs.
4.4.1. trièst-base on multigraphs. trièst-base can be adapted to work on multigraphs as
follows. First of all, the sample S should be considered a bag, i.e., it may contain multiple
copies of the same edge. Secondly, the function UpdateCounters must be changed as
presented in Alg. 4, to take into account the fact that inserting or removing an edge (u, v)
from S respectively increases or decreases the global number of triangles in GS by a quantity
that depends on the product of the number of edges (c, u) ∈ S and (c, v) ∈ S, for c in the
shared neighborhood (in GS ) of u and v (and similarly for the local number of triangles
incidents to c).
For this modified version of trièst-base, that we call trièst-base-m, an equivalent
version of Lemma 4.3 holds. Therefore, we can prove a result on the unbiasedness of trièstbase-m equivalent (i.e., with the same statement) as Thm. 4.2. The proof of such result is
also the same as the one for Thm. 4.2.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:24
Lorenzo De Stefani et al.
ALGORITHM 4 UpdateCounters function for trièst-base on multigraphs
1: function UpdateCounters((•, (u, v)))
S
2:
Nu,v
← NuS ∩ NvS
S
3:
for all c ∈ Nu,v
do
4:
yc,u ← number of edges between c and u in S
5:
yc,v ← number of edges between c and v in S
6:
yc ← yc,u · yc,v
7:
τ ← τ • yc
8:
τc ← τc • yc
9:
τ u ← τ u • yc
10:
τ v ← τ v • yc
To analyze the variance of trièst-base-m, we need to take into consideration the fact
that, in a multigraph, a pair of triangles may share two edges, and the variance depends
(t)
(also) on the number of such pairs. Let r1 be the number of unordered pairs of distinct
(t)
triangles from ∆(t) sharing an edge and let r2 be the number of unordered pairs of distinct
triangles from ∆(t) sharing two edges (such pairs may exist in a multigraph, but not in a
(t)
(t)
(t)
simple graph). Let q (t) = |∆2 | − r1 − r2 be the number of unordered pairs of distinct
triangles that do not share any edge.
Theorem 4.19. For any t > M , let f (t) = ξ (t) − 1,
g(t) = ξ (t)
(M − 3)(M − 4)
−1
(t − 3)(t − 4)
and
h(t) = ξ (t)
(M − 3)(M − 4)(M − 5)
− 1 (≤ 0),
(t − 3)(t − 4)(t − 5)
and
j(t) = ξ (t)
M −3
−1 .
t−3
We have:
h
i
(t)
(t)
Var ξ(t)τ (t) = |∆(t) |f (t) + r1 g(t) + r2 j(t) + q (t) h(t).
The proof follows the same lines as the one for Thm. 4.4, with the additional steps needed
(t)
to take into account the contribution of the r2 pairs of triangles in G(t) sharing two edges.
4.4.2. trièst-impr on multigraphs. A variant trièst-impr-m of trièst-impr for multigraphs can be obtained by using the function UpdateCounters defined in Alg. 4, modi(t)
(t)
fied to increment7 the counters by η (t) yc , rather than yc , where η (t) = max{1, (t − 1)(t −
2)/(M (M − 1))}. The result stated in Thm. 4.12 holds also for the estimations computed
by trièst-impr-m. An upper bound to the variance of the estimations, similar to the one
presented in Thm. 4.13 for trièst-impr, could potentially be obtained, but its derivation
would involve a high number of special cases, as we have to take into consideration the
order of the edges in the stream.
4.4.3. trièst-fd on multigraphs. trièst-fd can be modified in order to provide an approximation of the number of global and local triangles on multigraphs observed as a stream of
7 As
in trièst-impr, all calls to UpdateCounters in trièst-impr-m have • = +. See also Alg. 2.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:25
edge deletions and deletions. It is however necessary to clearly state the data model. We
assume that for all pairs of vertices u, v ∈ V (t) , each edge connecting u and v is assigned
a label that is unique among the edges connecting u and v. An edge is therefore uniquely
identified by its endpoints and its label as (u, v), label). Elements of the stream are now in
the form (•, (u, v), label) (where • is either + or −). This assumption, somewhat strong, is
necessary in order to apply the random pairing sampling scheme [Gemulla et al. 2008] to
fully-dynamic multigraph edge streams.
Within this model, we can obtain an algorithm trièst-fd-m for multigraphs by adapting
trièst-fd as follows. The sample S is a set of elements ((u, v), label). When a deletion
(−, (u, v), label) is on the stream, the sample S is modified if and only if ((u, v), label)
belongs to S. This change can be implemented in the pseudocode from Alg. 3 by modifying
line 8 to be
“else if ((u, v), label) ∈ S then” .
Additionally, the function UpdateCounters to be used is the one presented in Alg. 4.
We can prove a result on the unbiasedness of trièst-fd-m equivalent (i.e., with the
same statement) as Thm. 4.16. The proof of such result is also the same as the one for
Thm. 4.16. An upper bound to the variance of the estimations, similar to the one presented
in Thm. 4.17 for trièst-fd, could be obtained by considering the fact that in a multigraph
two triangles can share two edges, in a fashion similar to what we discussed in Thm. 4.19.
4.5. Discussion
We now briefly discuss over the algorithms we just presented, the techniques they use, and
the theoretical results we obtained for trièst, in order to highlight advantages, disadvantages, and limitations of our approach.
On reservoir sampling. Our approach of using reservoir sampling to keep a random sample
of edges can be extended to many other graph mining problems, including approximate
counting of other subgraphs more or less complex than triangles (e.g., squares, trees with a
specific structure, wedges, cliques, and so on). The estimations of such counts would still be
unbiased, but as the number of edges composing the subgraph(s) of interest increases, the
variance of the estimators also increases, because the probability that all edges composing
a subgraph are in the sample (or all but the last one when the last one arrives, as in the
case of trièst-impr), decreases as their number increases. Other works in the triangle
counting literature [Jha et al. 2015; Pavan et al. 2013] use samples of wedges, rather than
edges. They perform worse than trièst in both accuracy and runtime (see Sect. 5), but
the idea of sampling and storing more complex structures rather than simple edges could
be a potential direction for approximate counting of larger subgraphs.
On the analysis of the variance. We showed an exact analysis of the variance of trièstbase but for the other algorithms we presented upper bounds to the variance of the estimates. These bounds can still be improved as they are not currently tight. For example,
we already commented on the fact that the bound in (11) does not include a number of
negative terms that would tighten it (i.e., decrease the bound), and that could potentially
be no smaller than the term depending on z (t) . The absence of such terms is due to the
fact that it seems very challenging to obtain non-trivial upper bounds to them that are valid
for every t > M . Our proof for this bound uses a careful case-by-case analysis, considering
the different situations for pair of triangles (e.g., sharing or not sharing an edge, and considering the order of edges on the stream). It may be possible to obtain tighter bounds to
the variance by following a more holistic approach that takes into account the fact that the
sizes of the different classes of triangle pairs are highly dependent on each other.
Another issue with the bound to the variance from (11) is that the quantity z (t) depends on the order of edges on the stream. As already discussed, the bound can be made
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:26
Lorenzo De Stefani et al.
independent of the order by loosening it even more. Very recent developments in the sampling theory literature [Deville and Tillé 2004] presented sampling schemes and estimators
whose second-order sampling probabilities do not depend on the order of the stream, so
it should be possible to obtain such bounds also for the triangle counting problem, but a
sampling scheme different than reservoir sampling would have to be used, and a careful
analysis is needed to establish its net advantages in terms of performances and scalability
to billion-edges graphs.
On the trade-off between speed and accuracy. We concluded both previous paragraphs
in this subsection by mentioning techniques different than reservoir sampling of edges as
potential directions to improve and extend our results. In both cases these techniques are
more complex not only in their analysis but also computationally. Given that the main goal
of algorithms like trièst is to make it possible to analyze graphs with billions (and possibly
more) nodes, the gain in accuracy need to be weighted against expected slowdowns in
execution. As we show in our experimental evaluation in the next section, trièst, especially
in the trièst-impr variant, actually seems to strike the right balance between accuracy
and tradeoff, when compared with existing contributions.
5. EXPERIMENTAL EVALUATION
We evaluated trièst on several real-world graphs with up to a billion edges. The algorithms
were implemented in C++,and ran on the Brown University CS department cluster.8 Each
run employed a single core and used at most 4 GB of RAM. The code is available from
http://bigdata.cs.brown.edu/triangles.html.
Datasets. We created the streams from the following publicly available graphs (properties
in Table II).
Patent (Co-Aut.) and Patent (Cit.). The Patent (Co-Aut.) and Patent (Cit.) graphs
are obtained from a dataset of ≈ 2 million U.S. patents granted between ’75 and ’99 [Hall
et al. 2001]. In Patent (Co-Aut.), the nodes represent inventors and there is an edge
with timestamp t between two co-inventors of a patent if the patent was granted in
year t. In Patent (Cit.), nodes are patents and there is an edge (a, b) with timestamp t
if patent a cites b and a was granted in year t.
LastFm. The LastFm graph is based on a dataset [Celma Herrada 2009; The Koblenz
Network Collection (KONECT)] of ≈ 20 million last.fm song listenings, ≈ 1 million
songs and ≈ 1000 users. There is a node for each song and an edge between two songs
if ≥ 3 users listened to both on day t.
Yahoo!-Answers. The Yahoo! Answers graph is obtained from a sample of ≈ 160 million
answers to ≈ 25 millions questions posted on Yahoo! Answers [Datasets]. An edge
connects two users at time max(t1 , t2 ) if they both answered the same question at
times t1 , t2 respectively. We removed 6 outliers questions with more than 5000 answers.
Twitter. This is a snapshot [Boldi et al. 2011; Kwak et al. 2010] of the Twitter followers/following network with ≈ 41 million nodes and ≈ 1.5 billions edges. We do not have
time information for the edges, hence we assign a random timestamp to the edges (of
which we ignore the direction).
Ground truth. To evaluate the accuracy of our algorithms, we computed the ground truth
for our smaller graphs (i.e., the exact number of global and local triangles for each time
step), using an exact algorithm. The entire current graph is stored in memory and when an
edge u, v is inserted (or deleted) we update the current count of local and global triangles
by checking how many triangles are completed (or broken). As exact algorithms are not
scalable, computing the exact triangle count is feasible only for small graphs such as Patent
8 https://cs.brown.edu/about/system/services/hpc/grid/
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:27
(Co-Aut.), Patent (Cit.) and LastFm. Table II reports the exact total number of triangles
at the end of the stream for those graphs (and an estimate for the larger ones using trièstimpr with M = 1000000).
Table II. Properties of the dynamic graph streams analyzed. |V |, |E|, |Eu |, |∆|
refer respectively to the number of nodes in the graph, the number of edge addition
events, the number of distinct edges additions, and the maximum number of
triangles in the graph (for Yahoo! Answers and Twitter estimated with trièstimpr with M = 1000000, otherwise computed exactly with the naïve algorithm).
Graph
|V |
|E|
|Eu |
|∆|
Patent (Co-Aut.)
1,162,227
3,660,945
2,724,036
3.53 × 106
Patent (Cit.)
2,745,762
13,965,410
13,965,132
6.91 × 106
LastFm
681,387
43,518,693
30,311,117
1.13 × 109
109
109
7.86 × 1010
1.20 × 109
3.46 × 1010
Yahoo! Answers
2,432,573
1.21 ×
Twitter
41,652,230
1.47 × 109
1.08 ×
5.1. Insertion-only case
We now evaluate trièst on insertion-only streams and compare its performances with
those of state-of-the-art approaches [Jha et al. 2015; Lim and Kang 2015; Pavan et al.
2013], showing that trièst has an average estimation error significantly smaller than these
methods both for the global and local estimation problems, while using the same amount
of memory.
Estimation of the global number of triangles. Starting from an empty graph we add one
edge at a time, in timestamp order. Figure 1 illustrates the evolution, over time, of the
estimation computed by trièst-impr with M = 1,000,000. For smaller graphs for which
the ground truth can be computed exactly, the curve of the exact count is practically
indistinguishable from trièst-impr estimation, showing the precision of the method. The
estimations have very small variance even on the very large Yahoo! Answers and Twitter
graphs (point-wise max/min estimation over ten runs is almost coincident with the average
estimation). These results show that trièst-impr is very accurate even when storing less
than a 0.001 fraction of the total edges of the graph.
Comparison with the state of the art. We compare quantitatively with three state-of-theart methods: mascot [Lim and Kang 2015], Jha et al. [Jha et al. 2015] and Pavan et
al. [Pavan et al. 2013]. mascot is a suite of local triangle counting methods (but provides
also a global estimation). The other two are global triangle counting approaches. None of
these can handle fully-dynamic streams, in contrast with trièst-fd. We first compare the
three methods to trièst for the global triangle counting estimation. mascot comes in two
memory efficient variants: the basic mascot-c variant and an improved mascot-i variant.9
Both variants sample edges with fixed probability p, so there is no guarantee on the amount
of memory used during the execution. To ensure fairness of comparison, we devised the
following experiment. First, we run both mascot-c and mascot-i for ` = 10 times with a
fixed p using the same random bits for the two algorithms run-by-run (i.e. the same coin
tosses used to select the edges) measuring each time the number of edges Mi0 stored in
9 In
the original work [Lim and Kang 2015], this variant had no suffix and was simply called mascot. We
add the -i suffix to avoid confusion. Another variant mascot-A can be forced to store the entire graph with
probability 1 by appropriately selecting the edge order (which we assume to be adversarial) so we do not
consider it here.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:28
Lorenzo De Stefani et al.
Triangle Estimation vs Time
7x106
Triangle Estimation vs Time
1.2x109
ground truth
max est.
min est.
avg est.
Triangles
5x106
ground truth
max est.
min est.
avg est.
1x109
Triangles
6x106
4x106
3x106
2x106
8x108
6x108
4x108
1x106
2x108
0
0
7
10
7
10
7
10
5x
3.
3x
7
10
7
10
5x
2.
2x
7
Triangles
Triangles
3x10
4x1010
3x1010
10
2.5x1010
2x1010
1.5x1010
2x1010
1x1010
10
5x109
0
0
4x
1.
9
9
10
10
9
2x
8
10
8
10
8
10
8
10
10
Time
(d) Twitter
1.
1x
8x
6x
4x
2x
0
9
10
9
10
8
10
8
10
8
10
8
10
2x
1.
1x
8x
6x
4x
2x
0
Time
(c) Yahoo! Answers
5x
max est.
min est.
avg est.
3.5x1010
5x1010
1x10
1.
10
10
6x10
Triangle Estimation vs Time
4x1010
max est.
min est.
avg est.
7x1010
1x
6
10
Time
(b) LastFm
Triangle Estimation vs Time
8x1010
5x
0
7
10
7
10
4x
1.
2x
1.
7
10
1x
6
10
8x
6
10
6x
6
10
4x
6
10
2x
0
Time
(a) Patent (Cit.)
Fig. 1. Estimation by trièst-impr of the global number of triangles over time (intended as number of
elements seen on the stream). The max, min, and avg are taken over 10 runs. The curves are indistinguishable
on purpose, to highlight the fact that trièst-impr estimations have very small error and variance. For
example, the ground truth (for graphs for which it is available) is indistinguishable even from the max/min
point-wise estimations over ten runs. For graphs for which the ground truth is not available, the small
deviations from the avg suggest that the estimations are also close to the true value, given that our algorithms
gives unbiased estimations.
the sample at the end of the stream (by construction this the is same for the two variants
run-by-run). Then, we run our algorithms using M = Mi0 (for i ∈ [`]). We do the same to
fix the size of the edge memory for Jha et al. [Jha et al. 2015] and Pavan et al. [Pavan
et al. 2013].10 This way, all algorithms use the same amount of memory for storing edges
(run-by-run).
We use the MAPE (Mean Average Percentage Error) to assess the accuracy of the global
triangle estimators over time. The MAPE measures the average percentage of the prediction
error with respect to the ground truth, and is widely used in the prediction literature [Hyndman and Koehler 2006]. For t = 1, . . . , T , let ∆
PT
triangles at time t, the MAPE is defined as T1 t=1
(t)
be the estimator of the number of
|∆(t) |−∆
|∆(t) |
(t)
.11
10 More precisely, we use M 0 /2 estimators in Pavan et al. as each estimator stores two edges. For Jha et
i
al. we set the two reservoirs in the algorithm to have each size Mi0 /2. This way, all algorithms use Mi0 cells
for storing (w)edges.
11 The MAPE is not defined for t s.t. ∆(t) = 0 so we compute it only for t s.t. |∆(t) | > 0. All algorithms we
consider are guaranteed to output the correct answer for t s.t. |∆(t) | = 0.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:29
In Fig. 2(a), we compare the average MAPE of trièst-base and trièst-impr as well as
the two mascot variants and the other two streaming algorithms for the Patent (Co-Aut.)
graph, fixing p = 0.01. trièst-impr has the smallest error of all the algorithms compared.
We now turn our attention to the efficiency of the methods. Whenever we refer to one
operation, we mean handling one element on the stream, either one edge addition or one
edge deletion. The average update time per operation is obtained by dividing the total time
required to process the entire stream by the number of operations (i.e., elements on the
streams).
Figure 2(b) shows the average update time per operation in Patent (Co-Aut.) graph,
fixing p = 0.01. Both Jha et al. [Jha et al. 2015] and Pavan et al. [Pavan et al. 2013]
are up to ≈ 3 orders of magnitude slower than the mascot variants and trièst. This is
expected as both algorithms have an update complexity of Ω(M ) (they have to go through
the entire reservoir graph at each step), while both mascot algorithms and trièst need
only to access the neighborhood of the nodes involved in the edge addition.12 This allows
both algorithms to efficiently exploit larger memory sizes. We can use efficiently M up to 1
million edges in our experiments, which only requires few megabytes of RAM.13 mascot is
one order of magnitude faster than trièst (which runs in ≈ 28 micros/op), because it does
not have to handle edge removal from the sample, as it offers no guarantees on the used
memory. As we will show, trièst has much higher precision and scales well on billion-edges
graphs.
Table III. Global triangle estimation MAPE for trièst and mascot. The rightmost
column shows the reduction in terms of the avg. MAPE obtained by using trièst.
Rows with Y in column “Impr.” refer to improved algorithms (trièst-impr and
mascot-i) while those with N to basic algorithms (trièst-base and mascot-c).
Max. MAPE
p
Avg. MAPE
Graph
Impr.
mascot
trièst
mascot
trièst
Change
Patent
(Cit.)
N
Y
N
Y
0.01
0.01
0.1
0.1
0.9231
0.1907
0.0839
0.0317
0.2583
0.0363
0.0124
0.0037
0.6517
0.1149
0.0605
0.0245
0.1811
0.0213
0.0070
0.0022
-72.2%
-81.4%
-88.5%
-91.1%
Patent
(CoA.)
N
Y
N
Y
0.01
0.01
0.1
0.1
2.3017
0.1741
0.0648
0.0225
0.3029
0.0261
0.0175
0.0034
0.8055
0.1063
0.0390
0.0174
0.1820
0.0177
0.0079
0.0022
-77.4%
-83.4%
-79.8%
-87.2%
LastFm
N
Y
N
Y
0.01
0.01
0.1
0.1
0.1525
0.0273
0.0075
0.0048
0.0185
0.0046
0.0028
0.0013
0.0627
0.0141
0.0047
0.0031
0.0118
0.0034
0.0015
0.0009
-81.2%
-76.2%
-68.1%
-72.1%
Given the slow execution of the other algorithms on the larger datasets we compare
in details trièst only with mascot.14 Table III shows the average MAPE of the two
approaches. The results confirm the pattern observed in Figure 2(a): trièst-base and
12 We
observe that Pavan et al. [Pavan et al. 2013] would be more efficient with batch updates. However,
we want to estimate the triangles continuously at each update. In their experiments they use batch sizes of
million of updates for efficiency.
13 The experiments by Jha et al. [2015] use M in the order of 103 , and in those by Pavan et al. [2013], large
M values require large batches for efficiency.
14 We attempted to run the other two algorithms but they did not complete after 12 hours for the larger
datasets in Table III with the prescribed p parameter setting.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:30
Lorenzo De Stefani et al.
trièst-impr both have an average error significantly smaller than that of the basic mascotc and improved mascot variant respectively. We achieve up to a 91% (i.e., 9-fold) reduction
in the MAPE while using the same amount of memory. This experiment confirms the theory:
reservoir sampling has overall lower or equal variance in all steps for the same expected total
number of sampled edges.
Avg. MAPE vs Algorithm
Avg. Update Time vs Algorithm
Avg. Micros/Operation
Avg. Mape
1
0.1
0.01
10000
1000
100
10
1
O
AL
ET
N
VA
.
PA
AL
ET
A
I
JH
TO
C
AS
C
M
T-
C
AS
M
PR
IM
.
SE
BA
.
AL
.
ET
I
AL
ST
IE
ST
IE
TR
TR
N
ET
T-
O
C
T-
E
PR
AS
-IM
O
C
VA
A
PA
JH
AS
M
C
ST
IE
AS
M
TR
-B
ST
IE
TR
(a) MAPE
(b) Update Time
Fig. 2. Average MAPE and average update time of the various methods on the Patent (Co-Aut.) graph
with p = 0.01 (for mascot, see the main text for how we computed the space used by the other algorithms)
– insertion only. trièst-impr has the lowest error. Both Pavan et al. and Jha et al. have very high update
times compared to our method and the two mascot variants.
To further validate this observation we run trièst-impr and the improved mascot-i
variant using the same (expected memory) M = 10000. Figure 3 shows the max-min estimation over 10 runs and the standard deviation of the estimation over those runs. trièstimpr shows significantly lower standard deviation (hence variance) over the evolution of the
stream, and the max and min lines are also closer to the ground truth. This confirms our
theoretical observations in the previous sections. Even with very low M (about 2/10000 of
the size of the graph) trièst gives high-quality estimations.
Triangle Estimation vs Time
Std. Dev. of Triangle Estimation vs Time
1.2x108
ground truth
1.2x109 max est. TRIEST-IMPR
min
est.
TRIEST-IMPR
9
1x10
max est. MASCOT-I
min est. MASCOT-I
8x108
Std. Dev.
Triangles
1.4x109
6x108
4x108
2x108
0
1x10
8
8x10
7
std dev TRIEST-IMPR
std dev MASCOT-I
6x107
4x10
7
2x10
7
0
3.
7
10
7
5x
7
Time
(b) Standard deviation
10
7
10
7
5x
3x
2.
10
10
7
5x
10
2x
1.
6
10
1x
7
7
10
7
5x
5x
0
3.
10
7
10
7
5x
3x
2.
10
10
7
5x
2x
1.
6
10
1x
10
5x
0
Time
(a) Ground truth, max, and min
Fig. 3. Accuracy and stability of the estimation of trièst-impr with M = 10000 and of mascot-i with same
expected memory, on LastFM, over 10 runs. trièst-impr has a smaller standard deviation and moreover
the max/min estimation lines are closer to the ground truth. Average estimations not shown as they are
qualitatively similar.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:31
Local triangle counting. We compare the precision in local triangle count estimation of
trièst with that of mascot [Lim and Kang 2015] using the same approach of the previous experiment. We can not compare with Jha et al. and Pavan et al. algorithms as
they provide only global estimation. As in [Lim and Kang 2015], we measure the Pearson
coefficient and the average ε error (see [Lim and Kang 2015] for definitions). In Table IV
we report the Pearson coefficient and average ε error over all timestamps for the smaller
graphs.15 trièst (significantly) improves (i.e., has higher correlation and lower error) over
the state-of-the-art mascot, using the same amount of memory.
Table IV. Comparison of the quality of the local triangle estimations between our algorithms and the stateof-the-art approach in [Lim and Kang 2015]. Rows with Y in column “Impr.” refer to improved algorithms
(trièst-impr and mascot-i) while those with N to basic algorithms (trièst-base and mascot-c). In
virtually all cases we significantly outperform mascot using the same amount of memory.
Avg. ε Err.
Avg. Pearson
Graph
Impr.
p
mascot
trièst
Change
mascot
trièst
Change
Y
0.1
0.05
0.01
0.99
0.97
0.85
1.00
1.00
0.98
+1.18%
+2.48%
+14.28%
0.79
0.99
1.35
0.30
0.47
0.89
-62.02%
-52.79%
-34.24%
N
0.1
0.05
0.01
0.97
0.92
0.32
0.99
0.98
0.70
+2.04%
+6.61%
+117.74%
1.08
1.32
1.48
0.70
0.97
1.34
-35.65%
-26.53%
-9.16%
Y
0.1
0.05
0.01
0.41
0.24
0.05
0.82
0.61
0.18
+99.09%
+156.30%
+233.05%
0.62
0.65
0.65
0.37
0.51
0.64
-39.15%
-20.78%
-1.68%
N
0.1
0.05
0.01
0.16
0.06
0.00
0.48
0.24
0.003
+191.85%
+300.46%
+922.02%
0.66
0.67
0.86
0.60
0.65
0.68
-8.22%
-3.21%
-21.02%
Y
0.1
0.05
0.01
0.55
0.34
0.08
0.87
0.71
0.26
+58.40%
+108.80%
+222.84%
0.86
0.91
0.96
0.45
0.63
0.88
-47.91%
-31.12%
-8.31%
N
0.1
0.05
0.01
0.25
0.09
0.01
0.52
0.28
0.03
+112.40%
+204.98%
+191.46%
0.92
0.92
0.70
0.83
0.92
0.84
-10.18%
0.10%
20.06%
LastFm
Patent (Cit.)
Patent (Co-aut.)
Memory vs accuracy trade-offs. We study the trade-off between the sample size M vs the
running time and accuracy of the estimators. Figure 4(a) shows the tradeoffs between the
accuracy of the estimation (as MAPE) and the size M for the smaller graphs for which
the ground truth number of triangles can be computed exactly using the naïve algorithm.
Even with small M , trièst-impr achieves very low MAPE value. As expected, larger M
corresponds to higher accuracy and for the same M trièst-impr outperforms trièst-base.
Figure 4(b) shows the average time per update in microseconds (µs) for trièst-impr
as function of M . Some considerations on the running time are in order. First, a larger
edge sample (larger M ) generally requires longer average update times per operation. This
is expected as a larger sample corresponds to a larger sample graph on which to count
triangles. Second, on average a few hundreds microseconds are sufficient for handling any
update even in very large graphs with billions of edges. Our algorithms can handle hundreds
of thousands of edge updates (stream elements) per second, with very small error (Fig. 4(a)),
and therefore trièst can be used efficiently and effectively in high-velocity contexts. The
larger average time per update for Patent (Co-Auth.) can be explained by the fact that the
15 For
efficiency, in this test we evaluate the local number of triangles of all nodes every 1000 edge updates.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:32
Lorenzo De Stefani et al.
graph is relatively dense and has a small size (compared to the larger Yahoo! and Twitter
graphs). More precisely, the average time per update (for a fixed M ) depends on two main
factors: the average degree and the length of the stream. The denser the graph is, the higher
the update time as more operations are needed to update the triangle count every time the
sample is modified. On the other hand, the longer the stream, for a fixed M , the lower is the
frequency of updates to the reservoir (it can be show that the expected number of updates
to the reservoir is O(M (1 + log( Mt ))) which grows sub-linearly in the size of the stream
t). This explains why the average update time for the large and dense Yahoo! and Twitter
graphs is so small, allowing the algorithm to scale to billions of updates.
Avg Micros/Operation vs M
MAPE
1
M=100000 Base
M=500000 Base
M=1000000 Base
M=100000 Impr
M=500000 Impr
M=1000000 Impr
0.1
0.01
10000
M=100000 Impr
M=500000 Impr
M=1000000 Impr
1000
100
10
1
er
itt
o
ho
ya
tw
fm
st
la
ut
oa
t-c
en
t
pa
fm
st
la
it
t-c
en
t
pa
(a) M vs MAPE
it
t-c
en
t
pa
0.001
Avg Micros/Operation
MAPE vs M
(b) M vs Update Time (trièst-impr)
Fig. 4. Trade-offs between M and MAPE and avg. update time (µs) – edge insertion only. Larger M implies
lower errors but generally higher update times.
Alternative edge orders. In all previous experiments the edges are added in their natural
order (i.e., in order of their appearance).16 While the natural order is the most important
use case, we have assessed the impact of other ordering on the accuracy of the algorithms.
We experiment with both the uniform-at-random (u.a.r.) order of the edges and the random
BFS order: until all the graph is explored a BFS is started from a u.a.r. unvisited node and
edges are added in order of their visit (neighbors are explored in u.a.r. order). The results
for the random BFS order and u.a.r. order (Fig. 5) confirm that trièst has the lowest error
and is very scalable in every tested ordering.
5.2. Fully-dynamic case
We evaluate trièst-fd on fully-dynamic streams. We cannot compare trièst-fd with the
algorithms previously used [Jha et al. 2015; Lim and Kang 2015; Pavan et al. 2013] as they
only handle insertion-only streams.
In the first set of experiments we model deletions using the widely used sliding window
model, where a sliding window of the most recent edges defines the current graph. The
sliding window model is of practical interest as it allows to observe recent trends in the
stream. For Patent (Co-Aut.) & (Cit.) we keep in the sliding window the edges generated
in the last 5 years, while for LastFm we keep the edges generated in the last 30 days. For
Yahoo! Answers we keep the last 100 millions edges in the window17 .
16 Excluding
Twitter for which we used the random order, given the lack of timestamps.
sliding window model is not interesting for the Twitter dataset as edges have random timestamps.
We omit the results for Twitter but trièst-fd is fast and has low variance.
17 The
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
Avg. MAPE vs Algorithm
Avg. MAPE vs Algorithm
1
1
Avg. MAPE
Avg. MAPE
XX:33
0.1
0.01
0.1
A
I
T-
O
ET
ET
N
VA
.
PA
AL
JH
.
AL
C
T-
O
C
AS
M
C
AS
M
E
PR
-IM
AS
-B
ST
IE
ST
IE
.
AL
.
ET
AL
I
(a) BFS order
TR
TR
N
ET
T-
E
PR
C
T-
O
C
VA
A
PA
JH
AS
M
O
AS
-IM
ST
-B
ST
C
IE
AS
M
TR
IE
TR
0.01
(b) U.a.r. order
Fig. 5. Average MAPE on Patent (Co-Aut.), with p = 0.01 (for mascot, see the main text for how we
computed the space used by the other algorithms) – insertion only in Random BFS order and in u.a.r.
order. trièst-impr has the lowest error.
Figure 6 shows the evolution of the global number of triangles in the sliding window
model using trièst-fd using M = 200,000 (M = 1,000,000 for Yahoo! Answers). The
sliding window scenario is significantly more challenging than the addition-only case (very
often the entire sample of edges is flushed away) but trièst-fd maintains good variance and
scalability even when, as for LastFm and Yahoo! Answers, the global number of triangles
varies quickly.
Continuous monitoring of triangle counts with trièst-fd allows to detect patterns that
would otherwise be difficult to notice. For LastFm (Fig. 6(c)) we observe a sudden spike
of several order of magnitudes. The dataset is anonymized so we cannot establish which
songs are responsible for this spike. In Yahoo! Answers (Fig. 6(d)) a popular topic can
create a sudden (and shortly lived) increase in the number of triangles, while the evolution
of the Patent co-authorship and co-citation networks is slower, as the creation of an edge
requires filing a patent (Fig. 6(a) and (b)). The almost constant increase over time18 of the
number of triangles in Patent graphs is consistent with previous observations of densification
in collaboration networks as in the case of nodes’ degrees [Leskovec et al. 2007] and the
observations on the density of the densest subgraph [Epasto et al. 2015b].
Table V shows the results for both the local and global triangle counting estimation
provided by trièst-fd. In this case we can not compare with previous works, as they only
handle insertions. It is evident that precision improves with M values, and even relatively
small M values result in a low MAPE (global estimation), high Pearson correlation and
low ε error (local estimation). Figure 7 shows the tradeoffs between memory (i.e., accuracy)
and time. In all cases our algorithm is very fast and it presents update times in the order
of hundreds of microseconds for datasets with billions of updates (Yahoo! Answers).
Alternative models for deletion. We evaluate trièst-fd using other models for deletions
than the sliding window model. To assess the resilience of the algorithm to massive deletions
we run the following experiments. We added edges in their natural order but each edge
addition is followed with probability q by a mass deletion event where each edge currently
in the the graph is deleted with probability d independently. We run experiments with
q = 3,000,000−1 (i.e., a mass deletion expected every 3 millions edges) and d = 0.80 (in
expectation 80% of edges are deleted). The results are shown in Table VI.
18 The
decline at the end is due to the removal of the last edges from the sliding window after there are no
more edge additions.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:34
Lorenzo De Stefani et al.
Triangle Estimation vs Time
1.4x106
Triangle Estimation vs Time
1.6x106
ground truth
avg est.-std dev
avg est.
Triangles
Triangles
1x10
6
ground truth
1.4x106 avg est.+std dev
1.2x106 avg est.-std dev
1.2x106 avg est.+std dev
800000
600000
400000
1x10
avg est.
6
800000
600000
400000
200000
200000
0
0
7
10
7
0
7
0
ground truth
avg est.+std dev
avg est.-std dev
avg est.
avg est.+std dev
avg est.-std dev
avg est.
2x1010
Triangles
Triangles
3x
7
10
1
5x
2.
2x
1
5x
Triangle Estimation vs Time
2.5x1010
6x107
4x10
1.
8x107
7
Triangle Estimation vs Time
8
10
6
Time
(b) Patent (Cit.)
1.2x108
1x10
1x
10
6
10
6x
6
10
5x
6
6
10
4x
10
3x
6
10
2x
6
10
1x
0
Time
(a) Patent (Co-Aut.)
5x
0
-200000
7
1.5x1010
1x1010
5x109
0
2x107
-5x109
5x
9
10
9
10
9
Time
(d) Yahoo! Answers
2.
2x
10
9
10
5x
1.
1x
7
7
10
8
10
8x
7
10
7x
7
10
6x
7
10
5x
7
10
4x
7
10
3x
7
10
10
2x
1x
0
Time
(c) LastFm
5x
0
0
Fig. 6. Evolution of the global number of triangles in the fully-dynamic case (sliding window model for
edge deletion). The curves are indistinguishable on purpose, to remark the fact that trièst-fd estimations
are extremely accurate and consistent. We comment on the observed patterns in the text.
Avg Micros/Operation vs M
Avg Micros/Operation
10000
M=200000
M=500000
M=1000000
1000
100
10
1
o
ho
ya
nt
fm
te
st
la
pa
nt
te
pa
ut
oa
-c
it
-c
Fig. 7. Trade-offs between the avg. update time (µs) and M for trièst-fd.
We observe that trièst-fd maintains a good accuracy and scalability even in face of a
massive (and unlikely) deletions of the vast majority of the edges: e.g., for LastFM with
M = 200000 (resp. M = 1,000,000) we observe 0.04 (resp. 0.006) Avg. MAPE.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:35
Table V. Estimation errors for trièst-fd.
Avg. Global
Avg. Local
Graph
M
MAPE
Pearson
ε Err.
LastFM
200000
1000000
0.005
0.002
0.980
0.999
0.020
0.001
Pat. (Co-Aut.)
200000
1000000
0.010
0.001
0.660
0.990
0.300
0.006
Pat. (Cit.)
200000
1000000
0.170
0.040
0.090
0.600
0.160
0.130
Table VI. Estimation errors for trièst-fd– mass deletion experiment,
q = 3,000,000−1 and d = 0.80.
Avg. Global
Avg. Local
Graph
M
MAPE
Pearson
ε Err.
LastFM
200000
1000000
0.040
0.006
0.620
0.950
0.53
0.33
Pat. (Co-Aut.)
200000
1000000
0.060
0.006
0.278
0.790
0.50
0.21
Pat. (Cit.)
200000
1000000
0.280
0.026
0.068
0.510
0.06
0.04
6. CONCLUSIONS
We presented trièst, the first suite of algorithms that use reservoir sampling and its variants to continuously maintain unbiased, low-variance estimates of the local and global
number of triangles in fully-dynamic graphs streams of arbitrary edge/vertex insertions
and deletions using a fixed, user-specified amount of space. Our experimental evaluation
shows that trièst outperforms state-of-the-art approaches and achieves high accuracy on
real-world datasets with more than one billion of edges, with update times of hundreds of
microseconds.
APPENDIX
A. ADDITIONAL THEORETICAL RESULTS
In this section we present the theoretical results (statements and proofs) not included in
the main body.
A.1. Theoretical results for trièst-base
Before proving Lemma 4.1, we need to introduce the following lemma, which states a well
known property of the reservoir sampling scheme.
Lemma A.1 ([Vitter 1985, Sect. 2]). For any t > M , let A be any subset of E (t) of
size |A| = M . Then, at the end of time step t,
Pr(S = A) =
1
=
|E (t) |
M
1
t
M
,
i.e., the set of edges in S at the end of time t is a subset of E (t) of size M chosen uniformly
at random from all subsets of E (t) of the same size.
Proof of Lemma 4.1. If k > min{M, t}, we have Pr(B ⊆ S) = 0 because it is impossible for B to be equal to S in these cases. From now on we then assume k ≤ min{M, t}.
−1
If t ≤ M , then E (t) ⊆ S and Pr(B ⊆ S) = 1 = ξk,t
.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:36
Lorenzo De Stefani et al.
Assume instead that t > M , and let B be the family of subsets of E (t) that 1. have size
M , and 2. contain B:
B = {C ⊂ E (t) : |C| = M, B ⊆ C} .
We have
|B| =
|E (t) | − k
M −k
=
t−k
M −k
.
(15)
From this and and Lemma A.1 we then have
X
Pr(B ⊆ S) = Pr(S ∈ B) =
Pr(S = C)
C∈B
=
t−k
M −k
t
M
t−k
M −k
t−k Qk−1 t−i
i=0 M −i
M −k
=
=
k−1
Y
i=0
M −i
−1
= ξk,t
.
t−i
A.1.1. Expectation
Proof of Lemma 4.3. We only show the proof for τ , as the proof for the local counters
follows the same steps.
The proof proceeds by induction. The thesis is true after the first call to UpdateCounS
= ∅,
ters at time t = 1. Since only one edge is in S at this point, we have ∆S = 0, and Nu,v
so UpdateCounters does not modify τ , which was initialized to 0. Hence τ = 0 = ∆S .
Assume now that the thesis is true for any subsequent call to UpdateCounters up to
some point in the execution of the algorithm where an edge (u, v) is inserted or removed
from S. We now show that the thesis is still true after the call to UpdateCounters that
follows this change in S. Assume that (u, v) was inserted in S (the proof for the case of an
edge being removed from S follows the same steps). Let S b = S \ {(u, v)} and τ b be the
value of τ before the call to UpdateCounters and, for any w ∈ VS b , let τwb be the value
of τw before the call to UpdateCounters. Let ∆Su,v be the set of triangles in GS that
have u and v as corners. We need to show that, after the call, τ = |∆S |. Clearly we have
b
b
∆S = ∆S ∪ ∆Su,v and ∆S ∩ ∆Su,v = ∅, so
b
|∆S | = |∆S | + |∆Su,v |
b
S
We have |∆Su,v | = |Nu,v,
| and, by the inductive hypothesis, we have that τ b = |∆S |. Since
S
UpdateCounters increments τ by |Nu,v,
|, the value of τ after UpdateCounters has
S
completed is exactly |∆ |.
We can now prove Thm. 4.2 on the unbiasedness of the estimation computed by trièstbase (and on their exactness for t ≤ M ).
Proof of Thm. 4.2. We prove the statement for the estimation of global triangle
count. The proof for the local triangle counts follows the same steps.
If t ≤ M , we have GS = G(t) and from Lemma 4.3 we have τ (t) = |∆S | = |∆(t) |, hence
the thesis holds.
Assume now that t > M , and assume that |∆(t) | > 0, otherwise, from Lemma 4.3,
we have τ (t) = |∆S | = 0 and trièst-base estimation is deterministically correct. Let
(t)
λ = (a, b, c) ∈ ∆(t) , (where a, b, c are edges in E (t) ) and let δλ be a random variable that
takes value ξ (t) if λ ∈ ∆S (i.e., {a, b, c} ⊆ S) at the end of the step instant t, and 0 otherwise.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
From Lemma 4.1, we have that
h i
1
1
(t)
= ξ (t) (t) = 1 .
E δλ = ξ (t) Pr({a, b, c} ⊆ S) = ξ (t)
ξ3,t
ξ
XX:37
(16)
We can write
X
ξ (t) τ (t) =
(t)
δλ
λ∈∆(t)
and from this, (16), and linearity of expectation, we have
i
h i
h
X
(t)
E δλ = |∆(t) | .
E ξ (t) τ (t) =
λ∈∆(t)
A.1.2. Concentration
Proof of Lemma 4.7. Using the law of total probability, we have
Pr (f (Sin ) = 1) =
t
X
Pr (f (Sin ) = 1 | |Sin | = k ) Pr(|Sin | = k)
k=0
≥ Pr (f (Sin ) = 1 | |Sin | = M ) Pr(|Sin | = M )
≥ Pr (f (Smix ) = 1) Pr (|Sin | = M ) ,
(17)
where the last inequality comes from Lemma A.1: the set of edges included in Smix is
a uniformly-at-random subset of M edges from E (t) , and the same holds for Sin when
conditioning its size being M .
√
√
Using the Stirling approximation 2πn( ne )n ≤ n! ≤ e n( ne )n for any positive integer n,
we have
t−M
M
t−M
t
M
Pr (|Sin | = M ) =
t
t
M
√√
t
t t 2πe−t
M M (t − M )t−M
≥ √ √
tt
e2 M t − M e−t M M (t − M )t−M
1
≥ √
.
e M
Plugging this into (17) concludes the proof.
Fact A.2. For any x > 2, we have
x2
4
≤1+
.
(x − 1)(x − 2)
x−2
Proof of Lemma 4.10. We start by looking at the ratio between M t(t−1)(t−2)
(M −1)(M −2) and
3
(t/M ) . We have:
3
t(t − 1)(t − 2)
M
M2
(t − 1)(t − 2)
1≤
=
M (M − 1)(M − 2)
t
(M − 1)(M − 2)
t2
2
M
≤
(M − 1)(M − 2)
4
≤1+
M −2
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:38
Lorenzo De Stefani et al.
where the last step follows from Fact A.2. Using this, we obtain
φ
(t)
−
(t)
φmix
3
t(t − 1)(t − 2)
t
(t)
= τ
−τ
M (M − 1)(M − 2)
M
!
3
3
t
M
t(t − 1)(t − 2)
(t)
= τ
−1
M
M (M − 1)(M − 2)
t
3
t
4
≤ τ (t)
M
M −2
4
(t)
= φmix
.
M −2
(t)
A.1.3. Variance comparison. We now prove Lemma 4.11, about the fact that the variance
of the estimations computed by trièst-base is smaller, for most of the stream, than the
variance of the estimations computed by mascot-c [Lim and Kang 2015]. We first need
the following technical fact.
Fact A.3. For any x > 42, we have
8
x2
≤1+
.
(x − 3)(x − 4)
x
Proof of Lemma 4.11. We focus on t > M > 42 otherwise the theorem is immediate.
We show that for such conditions f (M, t) < f¯(M/T ) and g(M, t) < ḡ(M/T ). Using the fact
that t ≤ αT and Fact A.2, we have
t(t − 1)(t − 2)
T3
− 3
M (M − 1)(M − 2) M
α3 T 3
M2
T3
<
−
M 3 (M − 1)(M − 2) M 3
4
T3
α3 T 3
1
+
−
≤
M3
M −2
M3
3
3
T
4α
≤ 3 α3 +
−1 .
M
M −2
f (M, t) − f¯(M/T ) =
(18)
Given that T and M are ≥ 42, the r.h.s. of (18) is non-positive iff
α3 +
4α3
−1≤0 .
M −2
Solving for M we have that the above is verified when M ≥
given our assumption that M >
42 ≤ M and for any 0.6 ≤ α < 1
≤ 0 and f (M, t) < f¯(M/T ).
8α
max( 1−α
, 42): for any 0
4α3
8α
we have 1−α
3 + 2 < 1−α
4α3
1−α3
+ 2. This is always true
< α < 0.6, we have
4α3
1−α3
+2 <
≤ M . Hence the r.h.s. of (18) is
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:39
We also have:
T
t(t − 1)(t − 2)(M − 3)(M − 4)
−
(t − 3)(t − 4)M (M − 1)(M − 2) M
t
t2
T
<
−
M (t − 3)(t − 4) M
t
8
T
≤
1+
−
,
M
t
M
g(M, t) − ḡ(M/T ) =
(19)
where the last inequality follow from Fact A.3, since t > M > 42. Now, from (19) since
t ≤ αT and t > M , we can write:
T
8α
g(M, t) − ḡ(M/T ) <
α+
−1 .
M
M
The r.h.s. of this equation is non-positive given the assumption M >
ḡ(M/T ).
8α
1−α ,
hence g(M, t) <
A.2. Theoretical results for trièst-impr
A.2.1. Expectation
Proof of Thm. 4.12. If t ≤ M trièst-impr behaves exactly like trièst-base, and
the statement follows from Lemma 4.2.
Assume now t > M and assume that |∆(t) | > 0, otherwise, the algorithm deterministically
returns 0 as an estimation and the thesis follows. Let λ ∈ ∆(t) and denote with a, b, and c the
edges of λ and assume, w.l.o.g., that they appear in this order (not necessarily consecutively)
on the stream. Let tλ be the time step at which c is on the stream. Let δλ be a random
variable that takes value ξ2,tλ −1 if a and b are in S at the end of time step tλ − 1, and 0
otherwise. Since it must be tλ − 1 ≥ 2, from Lemma 4.1 we have that
Pr (δλ = ξ2,tλ −1 ) =
1
.
ξ2,tλ −1
(20)
When c = (u, v) is on the stream, i.e., at time tλ , trièst-impr calls UpdateCounters
S
S
| is the number of triangles with
|ξ2,tλ −1 , where |Nu,v
and increments the counter τ by |Nu,v
S∪{c}
(u, v) as an edge in ∆
. All these triangles have the corresponding random variables
taking the same value ξ2,tλ −1 . This means that the random variable τ (t) can be expressed
as
X
τ (t) =
δλ .
λ∈∆(t)
From this, linearity of expectation, and (20), we get
h
i
X
X
X
E τ (t) =
E[δλ ] =
ξ2,tλ −1 Pr (δλ = ξ2,tλ −1 ) =
ξ2,tλ −1
λ∈∆(t)
λ∈∆(t)
λ∈∆(t)
1
ξ2,tλ −1
= |∆(t) | .
A.2.2. Variance
Proof of Lemma 4.14. Consider first the case where all edges of λ appear on the
stream before any edge of γ, i.e.,
t`1 < t`2 < t`3 < tg1 < tg2 < tg3 .
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:40
Lorenzo De Stefani et al.
The presence or absence of either or both `1 and `2 in S at the beginning of time step t`3
(i.e., whether Dλ happens or not) has no effect whatsoever on the probability that g1 and
g2 are in the sample S at the beginning of time step tg3 . Hence in this case,
Pr(Dγ | Dλ ) = Pr(Dγ ) .
Consider now the case where, for any i ∈ {1, 2}, the edges g1 , . . . , gi appear on the stream
before `3 does. Define now the events
— Ai : “the edges g1 , . . . , gi are in the sample S at the beginning of time step t`3 .”
— Bi : if i = 1, this is the event “the edge g2 is inserted in the sample S during time step tg2 .”
If i = 2, this event is the whole event space, i.e., the event that happens with probability
1.
— C: “neither g1 nor g2 were among the edges removed from S between the beginning of
time step t`3 and the beginning of time step tg3 .”
We can rewrite Dγ as
Dγ = Ai ∩ Bi ∩ C .
Hence
Pr(Dγ | Dλ ) = Pr (Ai ∩ Bi ∩ C | Dλ )
= Pr (Ai | Dλ ) Pr (Bi ∩ C | Ai ∩ Dλ ) .
(21)
We now show that
Pr (Ai | Dλ ) ≤ Pr (Ai ) .
If we assume that t`3 ≤ M + 1, then all the edges that appeared on the stream up until the
beginning of t`3 are in S. Therefore,
Pr (Ai | Dλ ) = Pr(Ai ) = 1 .
Assume instead that t`3 > M +1. Among the t`3M−1 subsets of E (t`3 −1) of size M , there are
t`3 −3
that contain `1 and `2 . From Lemma A.1, we have that at the beginning of time t`3 ,
M −2
S is a subset of size M of E (t`3 −1) chosen uniformly at random. Hence, if we condition on
`3 −3
the fact that {`1 , `2 } ⊂ S, we have that S is chosen uniformly at random from the tM
−2
`3 −3−i
subsets of E (t`3 −1) of size M that contain `1 and `2 . Among these, there are tM
that
−2−i
also contain g1 , . . . , gi . Therefore,
t`3 −3−i
i−1
Y M −2−j
Pr(Ai | Dλ ) = Mt`−2−i
=
.
3 −3
t −3−j
j=0 `3
M −2
From Lemma 4.1 we have
Pr(Ai ) =
1
ξi,t`3 −1
=
i−1
Y
M −j
,
t −1−j
j=0 `3
where the last equality comes from the assumption t`3 > M + 1. From the same assumption
x
and from the fact that for any j ≥ 0 and any y ≥ x > j it holds x−j
y−j ≤ y , then we have
Pr(Ai | Dλ ) ≤ Pr(Ai ) .
This implies, from (21), that
Pr(Dγ | Dλ ) ≤ Pr(Ai ) Pr(Bi ∩ C | Ai ∩ Dλ ) .
(22)
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:41
Consider now the events Bi and C. When conditioned on Ai , these event are both independent from Dλ : if the edges g1 , . . . , gi are in S at the beginning of time t`3 , the fact that
the edges `1 and `2 were also in S at the beginning of time t`3 has no influence whatsoever
on the actions of the algorithm (i.e., whether an edge is inserted in or removed from S).
Thus,
Pr(Ai ) Pr(Bi ∩ C | Ai ∩ Dλ ) = Pr(Ai ) Pr(Bi ∩ C | Ai ) .
Putting this together with (22), we obtain
Pr(Dγ | Dλ ) ≤ Pr(Ai ) Pr(Bi ∩ C | Ai ) ≤ Pr(Ai ∩ Bi ∩ C) ≤ Pr(Dγ ) ,
where the last inequality follows from the fact that Dγ = Ai ∩ Bi ∩ C by definition.
A.3. Theoretical results for trièst-fd
A.3.1. Expectation. Before proving Thm. 4.16 we need the following technical lemmas.
The following is a corollary of [Gemulla et al. 2008, Thm. 1].
Lemma A.4. For any t > 0, and any j, 0 ≤ j ≤ s(t) , let B (t) be the collection of subsets
of E (t) of size j. For any B ∈ B (t) it holds
1
Pr S = B | M (t) = j = |E (t) | .
j
That is, conditioned on its size at the end of time step t, S is equally likely to be, at the end
of time step t, any of the subsets of E (t) of that size.
The next lemma is an immediate corollary of [Gemulla et al. 2008, Thm. 2]
Lemma A.5. Recall the definition of κ(t) from (14). We have
κ(t) = Pr(M (t) ≥ 3) .
The next lemma follows from Lemma A.4 in the same way as Lemma 4.1 follows from
Lemma A.1.
Lemma A.6. For any time step t and any j, 0 ≤ j ≤ s(t) , let B be any subset of E (t)
of size |B| = k ≤ s(t) . Then, at the end of time step t,
0
if k > j
1
Pr B ⊆ S | M (t) = j =
otherwise .
ψ
k,j,s(t)
The next two lemmas discuss properties of trièst-fd for t < t∗ , where t∗ is the first
time that |E (t) | had size M + 1 (t∗ ≥ M + 1).
Lemma A.7. For all t < t∗ , we have:
(t)
(1 ) do = 0; and
(2 ) S = E (t) ; and
(3 ) M (t) = s(t) .
Proof. Since the third point in the thesis follows immediately from the second, we focus
on the first two points.
The proof is by induction on t. In the base base for t = 1: the element on the stream must
be an insertion, and the algorithm deterministically inserts the edge in S. Assume now that
it is true for all time steps up to (but excluding) some t ≤ t∗ − 1. We now show that it is
also true for t.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:42
Lorenzo De Stefani et al.
Assume the element on the stream at time t is a deletion. The corresponding edge must
be in S, from the inductive hypothesis. Hence trièst-fd removes it from S and increments
(t)
the counter di by 1. Thus it is still true that S = E (t) and do = 0, and the thesis holds.
Assume now that the element on the stream at time t is an insertion. From the inductive
hypothesis we have that the current value of the counter do is 0.
If the counter di has currently value 0 as well, then, because of the hypothesis that t < t∗ ,
it must be that |S| = M (t−1) = s(t−1) < M . Therefore trièst-fd always inserts the edge
(t)
in S. Thus it is still true that S = E (t) and do = 0, and the thesis holds.
If otherwise di > 0, then trièst-fd flips a biased coin with probability of heads equal to
di
di
=
= 1,
di + do
di
therefore trièst-fd always inserts the edge in S and decrements di by one. Thus it is still
(t)
true that S = E (t) and do = 0, and the thesis holds.
The following result is an immediate consequence of Lemma A.5 and Lemma A.7.
Lemma A.8. For all t < t∗ such that s(t) ≥ 3, we have κ(t) = 1.
We can now prove Thm. 4.16.
Proof of Thm. 4.16. Assume for now that t < t∗ . From Lemma A.7, we have that
s(t) = M (t) . If M (t) < 3, then it must be s(t) < 3, hence |∆(t) | = 0 and indeed the algorithm
returns ρ(t) = 0 in this case. If instead M (t) = s(t) ≥ 3, then we have
ρ(t) =
τ (t)
.
κ(t)
From Lemma A.8 we have that κ(t) = 1 for all t < t∗ , hence ρ(t) = τ (t) in these cases. Since
(an identical version of) Lemma 4.3 also holds for trièst-fd, we have τ (t) = |∆S | = |∆(t) |,
where the last equality comes from the fact that S = E (t) (Lemma A.7). Hence ρ(t) = |∆(t) |
for any t ≤ t∗ , as in the thesis.
Assume now that t ≥ t∗ . Using the law of total expectation, we can write
(t)
h i min{s
i
X,M } h
(t)
E ρ
E ρ(t) | M (t) = j Pr M (t) = j .
=
(23)
j=0
Assume that |∆(t) | > 0, otherwise, the algorithm deterministically returns 0 as an esti(t)
mation and the thesis follows. Let λ be a triangle in ∆(t) , and let δλ be a random variable
that takes value
ψ3,M (t) ,s(t)
1
s(t) (s(t) − 2)(s(t) − 2)
=
(t)
(t)
(t)
(t)
κ
M (M − 1)(M − 2) κ(t)
if all edges of λ are in S at the end of the time instant t, and 0 otherwise. Since (an identical
version of) Lemma 4.3 also holds for trièst-fd, we can write
ρ(t) =
X
(t)
δλ
.
λ∈∆(t)
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:43
Then, using Lemma A.5 and Lemma A.6, we have, for 3 ≤ j ≤ min{M, s(t) },
h
i
X ψ3,j,s(t)
ψ3,j,s(t)
(t)
(t)
(t)
(t)
Pr δλ =
|M =j
E ρ |M =j =
κ(t)
κ(t)
(t)
λ∈∆
= |∆(t) |
ψ3,j,s(t)
1
1
= (t) |∆(t) |,
(t)
ψ
κ
κ
3,j,s(t)
(24)
and
h
i
E ρ(t) | M (t) = j = 0, if 0 ≤ j ≤ 2.
(25)
Plugging this into (23), and using Lemma A.5, we have
h i
1
E ρ(t) = |∆(t) | (t)
κ
min{s(t) ,M }
X
Pr(M (t) = j) = |∆(t) | .
j=3
A.3.2. Variance. We now move to prove Thm. 4.17 about the variance of the trièst-fd
estimator. We first need some technical lemmas.
Lemma A.9. For any time t ≥ t∗ , and any j, 3 ≤ j ≤ min{s(t) , M }, we have:
h
i
−1
2
Var ρ(t) |M (t) = j = (κ(t) )−2 |∆(t) | ψ3,j,s(t) − 1 + r(t) ψ3,j,s
−
1
(t) ψ
(t)
5,j,s
−1
(t)
2
+w
ψ3,j,s(t) ψ6,j,s(t) − 1
(26)
An analogous result holds for any u ∈ V (t) , replacing the global quantities with the corresponding local ones.
Proof. The proof is analogous to that of Theorem 4.4, using j in place of M , s(t)
in place of t, ψa,j,s(t) in place of ξa,t , and using Lemma A.6 instead of Lemma 4.1. The
additional (k (t) )−2 multiplicative term comes from the (k (t) )−1 term used in the definition
of ρ(t) .
−1
2
The term w(t) ψ3,j,s
is non-positive.
(t) ψ6,j,s(t) − 1
Lemma A.10. For any time t ≥ t∗ , and any j, 6 < j ≤ min{s(t) , M }, if s(t) ≥ M we
have:
h
i
−1
2
−
1
, for i ≥ j
Var ρ(t) |M (t) = i ≤ (κ(t) )−2 |∆(t) | ψ3,j,s(t) − 1 + r(t) ψ3,j,s
(t) ψ
5,j,s(t)
h
i
−1
2
Var ρ(t) |M (t) = i ≤ (κ(t) )−2 |∆(t) | ψ3,3,s(t) − 1 + r(t) ψ3,5,s
− 1 , for i < j
(t) ψ
5,5,s(t)
An analogous result holds for any u ∈ V (t) , replacing the global quantities with the corresponding local ones.
−1
2
is nonProof. The proof follows by observing that the term w(t) ψ3,j,s
(t) ψ6,j,s(t) − 1
positive, and that (26) is a non-increasing function of the sample size.
The following lemma deals with properties of the r.v. M (t) .
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:44
Lorenzo De Stefani et al.
(t)
(t)
Lemma A.11. Let t > t∗ , with s(t) ≥ M . Let d(t) = do + di denote the total number
of unpaired deletions at time t.19 The sample size M (t) follows the hypergeometric distribution:20
( (t) (t) (t) (t)
s
d
s +d
for max{M − d(t) , 0} ≤ j ≤ M .
(t)
j
M −j
M
Pr M = j =
(27)
0
otherwise
We have
h
i
E M (t) = M
s(t)
,
s(t) + d(t)
(28)
and for any 0 < c < 1
i
h
Pr M (t) > E M (t) − cM ≥ 1 −
1
.
e2c2 M
(29)
Proof. Since t > t∗, from the definition of t∗ we have that the M (t) has reached size
M at least once (at t∗ ). From this and the definition of d(t) (number of uncompensated
deletion), we have that M (t) can not be less than M − d(t) . The rest of the proof for (27)
and for (28) follows from [Gemulla et al. 2008, Thm. 2].
The concentration bound in (29) follows from the properties of the hypergeometric distribution discussed by Skala [2013].
The following is an immediate corollary from Lemma A.11.
Corollary A.12. Consider the execution of trièst-fd at time t > t∗ . Suppose we
have d(t) ≤ αs(t) , with 0 ≤ α < 1 and s(t) ≥ M . If M ≥ 2√α10 −α c0 ln s(t) for α < α0 < 1, we
have:
1
Pr M (t) ≥ M (1 − α0 ) > 1 −
c0 .
s(t)
We can now prove Thm. 4.17.
Proof of Thm. 4.17. From the law of total variance we have:
h
Var ρ
(t)
i
M
X
=
h
i
Var ρ(t) |M (t) = j Pr M (t) = j
j=0
+
M
h
i2
X
E ρ(t) |M (t) = j (1 − Pr M (t) = j ) Pr M (t) = j
j=0
−2
j−1 h
M X
i
h
i
X
E ρ(t) |M (t) = j Pr M (t) = j E ρ(t) |M (t) = i Pr M (t) = i .
j=1 i=0
19 While
20 We
(t)
(t)
both do and di are r.v.s, their
sum is not.
use here the convention that 00 = 1, and k0 = 1.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:45
As shown in (24) and (25), for any j = 0, 1, . . . , M we have E ρ(t) |M (t) = j ≥ 0. This in
turn implies:
M
h i X
h
i
Var ρ(t) ≤
Var ρ(t) |M (t) = j Pr M (t) = j
j=0
+
M
X
h
i2
E ρ(t) |M (t) = j (1 − Pr M (t) = j ) Pr M (t) = j .
(30)
j=0
Let us consider separately the two main components of (30). From Lemma A.10 we have:
M
X
h
i
Var ρ(t) |M (t) = j Pr M (t) = j =
(31)
j=0
M
X
M (1−α0 )
h
(t)
Var ρ |M
(t)
i
= j Pr M
j≥M (1−α0 )
≤ (κ(t) )−2
≤ (κ(t) )−2
(t)
= j) +
X
h
i
Var ρ(t) |M (t) = j Pr M (t) = j
j=0
−1
2
(t)
0
|∆(t) | ψ3,j,s(t) − 1 + r(t) ψ3,j,s
−
1
×
Pr
M
>
M
(1
−
α
)
(t) ψ
(t)
5,j,s
!
(t) 3
(t)
s
s
+ r(t)
Pr M (t) ≤ M (1 − α0 )
(32)
|∆(t) |
6
6
According to our hypothesis M ≥
√1
7 ln s(t) ,
2 α0 −α
thus we have, from Corollary A.12:
Pr M (t) ≤ M (1 − α0 )) ≤
1
.
(s(t) )7
As r(t) < |∆(t) |2 and |∆(t) | ≤ (s(t) )3 we have:
!
(t) 3
(t)
(t) −2
(t) s
(t) s
(κ )
|∆ |
+r
Pr M (t) ≤ M (1 − α0 ) ≤ (κ(t) )−2
6
6
We can therefore rewrite (32) as:
M
X
h
i
Var ρ(t) |M (t) = j Pr M (t) = j ≤ (κ(t) )−2 |∆(t) | ψ3,M (1−α0 ),s(t) − 1
j=0
−1
2
+ (κ(t) )−2 r(t) ψ3,M
(1−α0 ),s(t) ψ5,M (1−α0 ),s(t) − 1 + 1 .
(33)
2
PM
Let us now consider the term j=0 E ρ(t) |M (t) = j (1 − Pr M (t) = j ) Pr M (t) = j .
Recall that, from (24) and (25), we have E ρ(t) |M (t) = j = |∆(t) |(κ(t) )−1 for j = 3, . . . , M ,
(t) (t)
and E ρ |M = j = 0 for j = 0, 1, 2. From Corollary A.12 we have that for j ≤ (1−α0 )M
and M ≥ 2√α10 −α 7 ln s(t)
Pr M (t) = j ≤ Pr M (t) ≤ (1 − α0 )M ≤
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
1
7 ,
s(t)
XX:46
Lorenzo De Stefani et al.
and thus:
(1−α0 )M
X
j=0
h
i2
(1 − α0 )M |∆(t) |2 (κ(t) )−2
E ρ(t) |M (t) = j (1 − Pr M (t) = j ) Pr M (t) = j ≤
7
s(t)
≤ (1 − α0 )(κ(t) )−2 ,
(34)
where the last passage follows since, by hypothesis, M ≤ s(t) .
Let us now consider the values j > (1 − α0 )M , we have:
M
X
h
i2
E ρ(t) |M (t) = j (1 − Pr M (t) = j ) Pr M (t) = j
j>(1−α0 )M
≤ α0 M |∆(t) |2 (κ(t) )−2 1 −
M
X
Pr M (t) = j
j>(1−α0 )M
≤ α0 M |∆(t) |2 (κ(t) )−2 1 − Pr M (t) > (1 − α0 )M
≤
α0 M |∆(t) |2 (κ(t) )−2
7
s(t)
≤ α0 (κ(t) )−2 ,
(35)
where the last passage follows since, by hypothesis, M ≤ s(t) .
The theorem follows from composing the upper bounds obtained in (33), (34) and (35)
according to (30).
A.3.3. Concentration. We now prove Thm. 4.18 about trièst-fd.
Proof of Thm. 4.18. By Chebyshev’s inequality it is sufficient to prove that
Var[ρ(t) ]
<δ .
ε2 |∆(t) |2
From Lemma 4.17, for M ≥ √a10 −α 7 ln s(t) we have:
h i
Var ρ(t) ≤ (κ(t) )−2 |∆(t) | ψ3,M (1−α0 ),s(t) − 1
−1
2
+ (κ(t) )−2 r(t) ψ3,M
ψ
−
1
(t)
0
(1−α ),s
5,M (1−α0 ),s(t)
+ (κ(t) )−2 2
Let M 0 = (1 − α0 )M . In order to verify that the lemma holds, it is sufficient to impose the
following two conditions:
Condition (1).
(κ(t) )−2 |∆(t) | ψ3,M (1−α0 ),s(t) − 1 + 2
δ
>
.
2
ε2 |∆(t) |2
As by hypothesis |∆(t) | > 0, we can rewrite this condition as:
|∆(t) |−2
(t) −2
(κ
)
ψ
0 ),s(t) − (
(t)
3,M
(1−α
δ
|∆ |
>
2
ε2 |∆(t) |
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:47
which is verified for:
M 0 (M 0 − 1)(M 0 − 2) >
2s(t) (s(t) − 1)(s(t) − 2)
,
(t)
δε2 |∆(t) |(κ(t) )2 + 2 |∆|∆(t)|−2
|
v
u
(t)
(t)
(t)
u 2s (s − 1)(s − 2)
3
M0 > t
+ 2,
(t)
δε2 |∆(t) |(κ(t) )2 + 2 |∆|∆(t)|−2
|
v
u
(t) (s(t) − 1)(s(t) − 2)
u
2s
3
M > (1 − α0 )−1 t
+ 2 .
(t)
δε2 |∆(t) |(κ(t) )2 + 2 |∆|∆(t)|−2
|
Condition (2).
δ
(κ(t) )−2
−1
2
ψ
−
1
.
> 2 (t) 2 r(t) ψ3,M
(t)
0
(1−α ),s
5,M (1−α0 ),s(t)
2
ε |∆ |
(36)
As we have:
(t) −2 (t)
(κ )
r
−1
2
ψ3,M
(1−α0 ),s(t) ψ5,M (1−α0 ),s(t)
(t) −2 (t)
− 1 ≤ (κ ) r
s(t)
−1
6M (1 − α0 )
The condition (36) is verified for:
(1 − α0 )−1
M>
3
r(t) s(t)
δε2 |∆(t) |2 (κ(t) )−2 + 2r(t)
.
The theorem follows.
ACKNOWLEDGMENTS
This work was supported in part by NSF grant IIS-1247581 and NIH grant R01-CA180776.
REFERENCES
N. K. Ahmed, N. Duffield, J. Neville, and R. Kompella. Graph Sample and Hold: A framework for big-graph analytics. In Proceedings of the 20th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD ’14, pages 1446–1455. ACM,
2014.
Z. Bar-Yossef, R. Kumar, and D. Sivakumar. Reductions in streaming algorithms, with
an application to counting triangles in graphs. In Proceedings of the Thirteenth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA ’02, pages 623–632. SIAM, 2002.
L. Becchetti, P. Boldi, C. Castillo, and A. Gionis. Efficient algorithms for large-scale local
triangle counting. ACM Transactions on Knowledge Discovery from Data, 4(3):13:1–
13:28, 2010.
J. W. Berry, B. Hendrickson, R. A. LaViolette, and C. A. Phillips. Tolerating the community
detection resolution limit with edge weighting. Physical Review E, 83(5):056119, 2011.
P. Boldi, M. Rosa, M. Santini, and S. Vigna. Layered label propagation: A multiresolution coordinate-free ordering for compressing social networks. In Proceedings of the 20th
International Conference on World Wide Web, WWW ’11, pages 587–596. ACM, 2011.
L. Bulteau, V. Froese, K. Kutzkov, and R. Pagh. Triangle counting in dynamic graph
streams. Algorithmica, pages 1–20, 2015. ISSN 1432-0541. .
L. S. Buriol, G. Frahling, S. Leonardi, A. Marchetti-Spaccamela, and C. Sohler. Counting
triangles in data streams. In Proceedings of the 25th ACM SIGMOD-SIGACT-SIGART
Symposium on Principles of Database Systems, PODS ’06, pages 253–262. ACM, 2006.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
XX:48
Lorenzo De Stefani et al.
Ò. Celma Herrada. Music recommendation and discovery in the long tail. Technical report,
Universitat Pompeu Fabra, 2009.
E. Cohen, G. Cormode, and N. Duffield. Don’t let the negatives bring you down: sampling
from streams of signed updates. ACM SIGMETRICS Performance Evaluation Review,
40(1):343–354, 2012.
Y. R. W. Datasets. Yahoo! Answers browsing behavior version 1.0. http://webscope.
sandbox.yahoo.com.
L. De Stefani, A. Epasto, M. Riondato, and E. Upfal. TRIÈST: Counting local and global triangles in fully-dynamic streams with fixed memory size. In Proceedings of the 22nd ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16.
ACM, 2016.
J.-C. Deville and Y. Tillé. Efficient balanced sampling: The cube method. Biometrika, 91
(4):893–912, 2004. .
J.-P. Eckmann and E. Moses. Curvature of co-links uncovers hidden thematic layers in the
World Wide Web. Proceedings of the National Academy of Sciences, 99(9):5825–5829,
2002.
A. Epasto, S. Lattanzi, V. Mirrokni, I. O. Sebe, A. Taei, and S. Verma. Ego-net community
mining applied to friend suggestion. Proceedings of the VLDB Endowment, 9(4):324–335,
2015a.
A. Epasto, S. Lattanzi, and M. Sozio. Efficient densest subgraph computation in evolving
graph. In Proceedings of the 24th International Conference on World Wide Web, WWW
’15, pages 300–310. ACM, 2015b.
R. Gemulla, W. Lehner, and P. J. Haas. Maintaining bounded-size sample synopses of
evolving datasets. The VLDB Journal, 17(2):173–201, 2008.
A. Hajnal and E. Szemerédi. Proof of a conjecture of P. Erdős. In Combinatorial theory
and its applications, II (Proc. Colloq., Balatonfüred, 1969), pages 601–623, 1970.
B. H. Hall, A. B. Jaffe, and M. Trajtenberg. The NBER patent citation data file: Lessons,
insights and methodological tools. Technical report, National Bureau of Economic Research, 2001.
R. J. Hyndman and A. B. Koehler. Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4):679–688, 2006.
M. Jha, C. Seshadhri, and A. Pinar. A space-efficient streaming algorithm for estimating transitivity and triangle counts using the birthday paradox. ACM Transactions on
Knowledge Discovery from Data, 9(3):15:1–15:21, 2015.
H. Jowhari and M. Ghodsi. New streaming algorithms for counting triangles in graphs.
In Computing and Combinatorics: 11th Annual International Conference, COCOON ’05,
pages 710–716. Springer, 2005.
D. M. Kane, K. Mehlhorn, T. Sauerwald, and H. Sun. Counting arbitrary subgraphs in data
streams. In A. Czumaj, K. Mehlhorn, A. Pitts, and R. Wattenhofer, editors, Automata,
Languages, and Programming, volume 7392 of Lecture Notes in Computer Science, pages
598–609. Springer, 2012.
M. N. Kolountzakis, G. L. Miller, R. Peng, and C. E. Tsourakakis. Efficient triangle counting
in large graphs via degree-based vertex partitioning. Internet Mathematics, 8(1–2):161–
185, 2012.
K. Kutzkov and R. Pagh. On the streaming complexity of computing local clustering
coefficients. In Proceedings of the 6th ACM International Conference on Web Search and
Data Mining, WSDM ’13, pages 677–686. ACM, 2013.
H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a social network or a news
media? In Proceedings of the 19th International Conference on World Wide Web, WWW
’10, pages 591–600. ACM, 2010.
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
TRIÈST: Counting Triangles in Fully-dynamic Streams with Fixed Memory Size
XX:49
M. Latapy. Main-memory triangle computations for very large (sparse (power-law)) graphs.
Theoretical Computer Science, 407(1):458–473, 2008.
J. Leskovec, J. Kleinberg, and C. Faloutsos. Graph evolution: Densification and shrinking
diameters. ACM Transactions on Knowledge Discovery from Data, 1(1):2, 2007.
Y. Lim and U. Kang. MASCOT: Memory-efficient and accurate sampling for counting
local triangles in graph streams. In Proceedings of the 21st ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD ’15, pages 685–694. ACM,
2015.
M. Manjunath, K. Mehlhorn, K. Panagiotou, and H. Sun. Approximate counting of cycles
in streams. In European Symposium on Algorithms, ESA ’11, pages 677–688. Springer,
2011.
R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon. Network motifs:
simple building blocks of complex networks. Science, 298(5594):824–827, 2002.
M. Mitzenmacher and E. Upfal. Probability and computing: Randomized algorithms and
probabilistic analysis. Cambridge University Press, 2nd edition, 2005.
R. Pagh and C. E. Tsourakakis. Colorful triangle counting and a MapReduce implementation. Information Processing Letters, 112(7):277–281, Mar. 2012.
H. Park, F. Silvestri, U. Kang, and R. Pagh. MapReduce triangle enumeration with guarantees. In Proceedings of the 23rd ACM International Conference on Conference on
Information and Knowledge Management, CIKM ’14, pages 1739–1748. ACM, 2014. .
H.-M. Park and C.-W. Chung. An efficient MapReduce algorithm for counting triangles
in a very large graph. In Proceedings of the 22nd ACM International Conference on
Conference on Information & Knowledge Management, CIKM ’13, pages 539–548. ACM,
2013.
H.-M. Park, S.-H. Myaeng, and U. Kang. PTE: Enumerating trillion triangles on distributed systems. In Proceedings of the 22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, KDD ’16. ACM, 2016.
A. Pavan, K. Tangwongsan, S. Tirthapura, and K.-L. Wu. Counting and sampling triangles
from a graph stream. Proceedings of the VLDB Endowment, 6(14):1870–1881, 2013.
M. Skala. Hypergeometric tail inequalities: ending the insanity. arXiv preprint, 1311.5939,
2013.
S. Suri and S. Vassilvitskii. Counting triangles and the curse of the last reducer. In Proceedings of the 20th International Conference on World Wide Web, WWW ’11, pages
607–614. ACM, 2011.
The Koblenz Network Collection (KONECT). Last.fm song network dataset. http://konect.
uni-koblenz.de/networks/lastfm_song.
C. E. Tsourakakis, U. Kang, G. L. Miller, and C. Faloutsos. Doulion: counting triangles
in massive graphs with a coin. In Proceedings of the 15th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD ’09, pages 837–846. ACM,
2009.
C. E. Tsourakakis, M. N. Kolountzakis, and G. L. Miller. Triangle sparsifiers. Journal of
Graph Algorithms and Applications, 15(6):703–726, 2011.
J. S. Vitter. Random sampling with a reservoir. ACM Transactions on Mathematical
Software, 11(1):37–57, 1985.
Received Month Year; revised Month Year; accepted Month Year
ACM Journal Name, Vol. XX, No. X, Article XX, Publication date: 2016.
| 8cs.DS
|
1
An Analytical Framework for Modeling a
Spatially Repulsive Cellular Network
arXiv:1701.02261v3 [cs.IT] 30 Sep 2017
Chang-Sik Choi, Jae Oh Woo, and Jeffrey G. Andrews
Abstract
We propose a new cellular network model that captures both deterministic and random aspects of
base station deployments. Namely, the base station locations are modeled as the superposition of two
independent stationary point processes: a random shifted grid with intensity λg and a Poisson point
process (PPP) with intensity λp . Grid and PPP deployments are special cases with λp → 0 and λg → 0,
with actual deployments in between these two extremes, as we demonstrate with deployment data.
Assuming that each user is associated with the base station that provides the strongest average received
signal power, we obtain the probability that a typical user is associated with either a grid or PPP base
station. Assuming Rayleigh fading channels, we derive the expression for the coverage probability of the
typical user, resulting in the following observations. First, the association and the coverage probability
of the typical user are fully characterized as functions of intensity ratio ρλ = λp /λg . Second, the user
association is biased towards the base stations located on a grid. Finally, the proposed model predicts
the coverage probability of the actual deployment with great accuracy.
I. I NTRODUCTION
A. Motivation and Related Work
The topological distribution of base stations (BSs) is a first order effect in terms of the signalto-interference-and-noise ratio (SINR) of a cellular network. BSs are usually intentionally spread
out from one another, which is referred to as repulsion in studies [2]–[5]. However, those works
did not analytically derive SINR expressions, primarily because it is quite difficult to do so
for repulsive network model. Two examples of actual BS deployments are given in Figures 1
Chang-sik Choi, Jae Oh Woo, and Jeffrey G. Andrews are with the Wireless and Networking and Communication Group,
Department of Electrical and Computer Engineering, The University of Texas at Austin, Texas, 78701, USA. (e-mail: [email protected], [email protected], [email protected])
Part of this work has been presented in IEEE ISIT 2017 [1]
2
and 2, in which repulsion can be observed both subjectively and objectively, by using statistical
metrics. The goal of this paper is to propose a novel model and analytical framework for cellular
networks in view of repulsive BS deployments.
B. Background: From Poisson to More General Models
The modeling and analysis of BS locations using the Poisson point process (PPP) has become
popular in the last five years because the independence of the PPP helps allow computable—and
in some cases quite simple—SINR distribution expressions. The PPP-based models describe
randomly distributed BSs and often yield a simple expression for the interference [6], [7].
In [8] the coverage probability of the typical user—meaning the probability that the SINR is
above a certain value—was derived in a closed form, and was extended in [9]–[11] to multi-tier
heterogeneous cellular networks (HCNs). The benefits of HCNs were investigated further in [12]
and [13]. The PPP model was also used for finding the average rate [14], obtaining an energyefficient algorithm [15] and [16], and deriving the signal-to-interference (SIR) of millimeter
cellular networks [17]. A PPP approximation was studied in [18]–[21]. However, the PPP model
ignores the repulsive nature of spatial topology observed in cellular networks.
In order to mathematically represent repulsion between BSs, [22]–[24] considered repulsive
point processes given in [25, Chapter 5] and attempted to derive the SINR expressions analytically. Several specific point processes were used. For example, [22] considered the Ginibre point
process, [23] provided tractable but complicated analysis for the more general determinantal point
process, and [24] used the Matérn hard-core point process, which introduces an exclusion region
around each BS.
C. Contributions
This paper proposes to model BS locations using a stationary point process that is constructed
by the superposition of two independent stationary point processes. The first is a standard PPP
with density λp and the second is a random shifted grid with intensity λg . The realizations of
the proposed point process and its association regions are given in Figure 3 for multiple values
of ρλ . Possible real-world examples where such a model could be applicable including
•
A vehicular network where fixed infrastructure modeled by grid and randomly located
vehicles modeled by Poisson points can transmit information to other vehicles.
3
•
A device-to-device (D2D) network with fixed BSs and randomly scattered D2D devices; a
given mobile receiver is able to receive transmissions from either one.
•
A HCN with repulsively (nearly in a grid) macro BSs overlaid with small cells randomly
scattered over space.
As can be seen heuristically in Figures 1 and 2, actual data also conforms to such a model where
BSs are placed both randomly and regularly.
The theoretical contributions of this paper are as follows.
Analytical framework capturing repulsion between BSs. We propose to model cellular
networks as a combination of two extreme sub-structures. The repulsive BSs of transmit power
pg are modeled by a random shifted grid with intensity λg and the random BSs with transmit
power pp are modeled by a PPP with intensity λp . We capture repulsion by the intensity ratio
ρλ =
λp
.
λg
In fact, any network model between the grid and the PPP can be reproduced by simply
varying ρλ . When ρλ → 0 the proposed model describes the lattice BS model: meanwhile,
when ρλ → ∞ the proposed model corresponds to the Poisson BS model. The interpolation is
demonstrated in Figure 3. Consequently, the proposed model is readily applicable to practical
deployments compared to models with only lattice [26] or PPP BSs [8].
Derivation of the nearest distance distribution, user association probability, interference,
and the SIR coverage probability. First, the nearest distance distribution of the proposed
stationary point process is characterized. Then, by assuming a typical user is associated with the
BS that provides the strongest average received power, the association probability of the typical
user is derived. Assuming Rayleigh fading channels, the Laplace transform of interference seen
at the typical user is derived and the expression for the coverage probability of the typical user is
obtained. The association probability, the interference, and the coverage probability are described
by functions of system parameters including the intensity ratio ρλ and the power ratio η =
pp
.
pg
Insights and observations from the proposed model. First, in user association, a bias towards
the repulsive BSs is observed. For instance, when ρλ = η = 1, the typical user at the origin is
more likely to be associated with a grid BS than a randomly located BS. Second, the proposed
model is robust and consistent for different values of α. We numerically show that the proposed
model predict the coverage probability of the actual deployment very accurately.
The rest of this paper is organized as follows. Section II discusses the proposed system model
and defines performance metrics. Section III investigates mathematical properties of the proposed
point process and its tessellation. Section IV discusses the user association, interference, the
4
TABLE I
K EY N ETWORK PARAMETERS
Parameter
Notation
Random shifted grid BS (intensity)
Φg (λg )
Homogeneous Poisson BS (intensity)
Φp (λp )
The proposed point process
Φ = Φg ∪ Φp
Intensity ratio
ρλ = λp /λg
Fading random variable
H
Transmit power ratio
η = pp /pg
Nearest distance from the origin to the Φg
Rg
Nearest distance from the origin to the Φp
Rp
The smallest element of grid
S0 ∈ R2
The area of a set A
ν2 (A)
An uniform random variable on a set A
The area of the typical Voronoi cell w.r.t. grid (Poisson) BSs
Uniform[A]
V¯g (V¯p )
The event that typical user is associated with Φg (Φp )
Ag (Ap )
coverage probability of the typical user, and its bounds. Section V concludes the paper.
II. S YSTEM M ODEL
A. Spatial Model
To model the locations of the grid BSs, a random shifted grid is used. The random shifted
grid is given by shifting the vertexes of the standard square 2-d grid via a single uniform random
variable while maintaining the entire grid structure. The locations of the BSs are given by
X
Φg =
δs·k+U ,
(1)
k∈Z2
where δx denotes the dirac-delta function at x, s denotes the distance between the nearest vertexes,
2
and U denotes the uniform random variable on − 2s , 2s . The above expression shows that all
points of the square grid are shifted by a random variable U as a group. This is different from the
perturbed lattice produced by shifting each point using i.i.d. uniform Ui . The perturbed lattice is
inappropriate for the modeling purpose since the deterministic structure between BSs no longer
exists and repulsion between BSs is diluted. On the other hand, the random shifted grid is a
stationary point process because the joint distribution,
{Φg (B1 + x), Φg (B2 + x), . . . , Φg (Bk + x)},
(2)
5
33.55
BS location
33.54
33.53
33.52
latitude
33.51
33.5
33.49
33.48
33.47
33.46
33.45
-112.2
-112.15
-112.1
-112.05
longitude
Fig. 1. Actual BS deployment for the area of latitude and longitude [33.45, 33.55] × [−112.2, −112.05]. The pair correlation
value is 0.90 and the corresponding estimated intensity ratio is 2.3.
does not depend on the location of x ∈ R2 for any finite Borel set Bk ∈ B(R2 ) [27, Definition
3.2.1.]. Since it is a stationary point process, it admits the intensity parameter denoted by λg . It
is given by the average number of points in a unit area λg = 1/s2 . Throughout this paper, the
random shifted grid is referred to as stationary grid or just grid. On the same space that Φg is
considered, an independent PPP with intensity parameter λp is considered
X
Φp =
δXi .
(3)
i∈N
Since the two point processes are stationary, the superposition of the point processes also
yields a stationary point process. The proposed point process is given by
Φ = Φg ∪ Φp .
(4)
its intensity parameter is λ = λg + λp . We introduce a prefatory parameter ρλ defined by
ρλ =
λp
,
λg
(5)
6
33.44
BS location
33.43
33.42
33.41
latitude
33.4
33.39
33.38
33.37
33.36
33.35
33.34
-112.2
-112.15
-112.1
-112.05
longitude
Fig. 2. Actual BS deployment in the area of latitude and longitude [33.34, 33.44] × [−112.2, −112.05]. The pair correlation
value is 0.776 and the corresponding estimated intensity ratio is 1.11
which will be seen repeatedly in this paper. Note that repulsion between BSs can be modeled in
many different ways as presented in [5], [22]–[24]; this paper proposes one way of characterizing
repulsion using (5). Under the proposed framework, if ρλ tends to zero, the proposed network
describes the grid almost everywhere. If ρλ tends to infinity, the proposed network describes the
PPP almost everywhere. In this sense, the proposed model is designed to interpolate the grid
and the PPP. It is clarified in Figure 3.
In addition to the proposed point process Φ, another stationary independent PPP Φr is considered to describe the locations of downlink cellular users. Under the stationary framework, a
typical user located at the origin is considered.
Remark 1: The proposed point process, in particular the random shifted grid, is able to address
repulsion between pairs of BSs by changing the intensity λg w.r.t. λp . This compares favorably
to the models using perturbed lattice [28] where the repulsion is softly induced by having only
7
Fig. 3. The grid BSs and PPP BSs are demonstrated by squares and triangles, respectively. The interpolation of the proposed
model is described for multiple values of ρλ . We maintain the total number of BSs to minimize the edge effect. The dashed
lines indicate the Voronoi boundaries.
one point per grid. In other words, the perturbed lattice model potentially places two BSs very
close to each other, which is observed rarely in actual cellular deployments. In addition, the
proposed model captures the repulsion by a single scalar parameter λg . This approach contrasts
to the Ginibre point process [22] or determinantal point process [23] where the repulsions are
modeled for all pairs of points and require a multi-dimensional parameter.
B. Transmission Model
The received signal power at distance R is given by P HR−α where P denotes the transmit
power, H denotes Rayleigh fading that follows exponential distribution with mean one, and R−α
denotes path loss with α > 2. The transmit power P of the BS at X is given by
pg . for X ∈ Φg
PX =
pp . for X ∈ Φp
8
This paper assumes that cellular downlink users are associated with BSs that provide the strongest
average power. The BS associated with the typical user is given by
PXi HXi
PXi
?
X = arg max E
= arg max
,
α
kXi − 0k
kXi kα
Xi ∈Φ
Xi ∈Φ
(6)
where the subscripts on H and P are used to specify the transmitter and k · k denotes the
Euclidean distance. Using X ? , the received signal power of the typical user at the origin is given
by
S = PX ? HX ? kX ? k−α ,
(7)
and the interference is given by
X
I=
Xi
PXi HXi kXi k−α .
(8)
∈Φ\{X ? }
C. System Performance Metric
In this paper, we use the coverage probability of the typical user to assess the coverage
probability of all users averaged across the space. Provided that noise power at the user is
minuscule compared to the received signal power or the interference powers, which is commonly
assumed in practical dense cellular networks, the coverage probability of the typical user is
captured by the SIR at the typical user
SIR =
PX ? HkX ? k−α
S
,
=P
−α
I
Xi ∈Φ\{X ? } PXi HkXi k
(9)
where subscripts on H are omitted. The typical transmission at the typical user is successful if
the above typical SIR is greater than a threshold T . The coverage probability of the typical user
is written as pcov = P0 (SIR ≥ T ) where we use the Palm notation.
III. P ROPERTIES OF THE P ROPOSED M ODEL : N EAREST D ISTANCE D ISTRIBUTION AND THE
VORONOI T ESSELLATION
This section discusses mathematical properties of the proposed point process for BSs using
stochastic geometry.
9
A. Nearest Distance Distribution
The distances from the typical point to the nearest points of Φg or Φp are defined by
Rg = inf kXi k,
(10)
Rp = inf kXi k,
(11)
Xi ∈Φg
Xi ∈Φp
respectively. The cumulative distribution function (CDF) and probability distribution function
(PDF) for Rp are given by
P(Rp ≤ r) = 1 − exp(−πλp r2 ),
(12)
fRp (r) = 2πλp r exp(−πλp r2 ).
(13)
since Φp is a homogeneous Poisson point process.
Lemma 1: The CDF of Rg is given by
0
2
πr
s2
q
P(Rg < r) =
s
π−4 arccos( 2r
2
2
)
r
2 s2
− 2 + 4rs2 − 1
2
1
if r < 0
if 0 ≤ r ≤ 2s ,
if
s
2
<r≤
if r >
√s ,
2
(14)
√s ,
2
and its PDF is given by
2πrs−2
fRg (r) = 2πrs−2 −
0
if 0 ≤ r ≤ 2s ,
√
8r
arccos( 2rs ) if 2s < r ≤ 2s 2,
s2
(15)
otherwise,
where the constant s denotes the width of square lattice, which is equal to √1 .
λg
Proof: Appendix A
The distance from the origin to the nearest point of Φ is described by
P(R ≤ r) = 1 − (1 − P(Rp ≤ r))(1 − P(Rg ≤ r)),
where P(Rg < r) and P(Rp < r) are given by (12) and (14), respectively.
(16)
10
B. Voronoi Tessellation and Average Fraction of Typical Voronoi Cells
Corollary 1: Let V¯p and V¯g denote the average area for the typical cells w.r.t. Φp and Φg ,
respectively. The average area fractions are given by
Z √2
Z √2
√
π 2
V¯p
2
2
(1 − r )f (r) dr +
r arccos(1/r) − r − 1 f (r) dr,
=
4
V̄p + V̄g
0
1
V¯g
V¯p
=1−
,
V̄p + V̄g
V̄p + V̄g
(17)
(18)
where the function f (r) is given by
f (r) = 2π
ρ
ρλ
λ
r exp −π r2 .
4
4
(19)
Proof: The Voronoi cell centered around X ? is described by
VX ? = {x ∈ R2 |kxk ≤ kx − Xj k∀Xj ∈ Φ}.
The area fraction of the typical cell is then
E0 (1{Rg > Rp }
V¯p
= 0
V̄p + V̄g
E (1{Rg > Rp } + E0 (1{Rp > Rg })
= E0 (1 {Rg > Rp })
Z ∞
=
P (Rg ≥ r) fRp (r) dr.
0
By replacing P(Rg > r) with expression (14), we have
Z √2
Z √2
√
π 2
V¯p
(1 − r )f (r) dr +
r2 arccos(1/r) − r2 − 1 f (r) dr,
=
4
V̄p + V̄g
0
1
ρλ 2
ρλ
where f (r) = 2π 4 r exp −π 4 r .
C. Application to Any Pairwise Repulsive Network
The applicability of the network model is supported by discussing the pair correlation function
of proposed point process. It is known that the pair correlation function is a fundamental
mathematical metric for explaining the pairwise repulsion between two points [25, Section 4.5.].
The concept was utilized in cellular architecture by the authors of [4].
Definition 1: Consider a simple point process Ξ. The pair correlation function between x, y ∈ Ξ
is defined by
κ(x, y) =
λ(x, y)
,
λ(x)λ(y)
(20)
11
where λ(x, y) is the factorial second-order moment measure of Ξ, and λ(x) is the first-order
moment measure of Ξ at X.
To describe the pair correlation function of our proposed model Φ, we define two sets D =
{(x, y) : x 6= y ∈ R2 } and G = {(x, y) : x ∼g y} where x ∼g y means x − y ∈ s · k for some
k ∈ Z2 . In other words, D is a collection of all diagonal points in R2 × R2 , and G is a collection
of all possible pair grid points from each realization of Φg .
Proposition 1: The pair correlation function of the proposed Φ is given by
1
(1 + ρλ )2
∀(x, y) ∈ R2 × R2 \ (D ∪ G) ,
(21)
1
ρλ = p
− 1 ∀(x, y) ∈ R2 × R2 \ (D ∪ G) .
1 − κ(x, y)
(22)
κ(x, y)=1 −
which is independent of x, y. Equivalently,
Proof: To find the factorial second moment density function λ(x, y) for (x, y) ∈ R2 × R2 \
(D ∪ G), we can consider three sub-cases where each point can be from either the grid or the
PPP excluding both from the grid. Since Φp and Φg are independent,
λ(x, y) = λ2p + λp λg + λg λp .
(23)
Then by using λ(x) = λ(y) = λp + λg for all x, y ∈ R2 , (21) follows immediately.
Since D ∪ G has zero volume in R2 × R2 , we can regard the pair-correlation function κ(x, y)
of Φ as a constant function κ(x, y) = κ for all x, y ∈ R2 . In general, for any simple point process
Ξ, the pair correlation function explains the pairwise repulsion between two points. Since we
have the one-to-one relation between ρλ and κ, our proposed point process Φ is able to capture
pairwise repulsion by a single scalar constant. In the following example, we illustrate one method
that the proposed model and its results can be applied to a spatial model Ξ with arbitrary BSs
utilizing the average pair correlation κ̂ in [25, Section 4.7.4].
Example 1: The actual BS deployment map and the applicability of our model: Figures
1 and 2 demonstrate two actual BS deployment maps of the United States at latitude and
longitude specified by [33.45, 33.55]×[−112.2, −112.05] and [33.34, 33.44]×[−112.2, −112.05],
respectively. We use a open source software R to find the averaged pair correlation values given
by 0.90 and 0.766, respectively. Using (22), the intensity ratios of those maps are given by 2.3
and 1.11, respectively.
12
IV. U SER A SSOCIATION , I NTERFERENCE ,
AND
C OVERAGE P ROBABILITY
Throughout this paper, we write Ag and Ap to denote the event that the typical user being
associated with either Φg or Φp , respectively.
A. Typical User Association
Theorem 1: The probability that the typical user is associated with the random BS is
Z √2
Z √2
√
π 2
0
2
2
(1 − r )f (r) dr +
r arccos(1/r) − r − 1 f (r) dr.
P (Ap ) =
4
0
1
(24)
Accordingly, the probability that the typical user is associated with the grid BS is
P0 (Ag ) = 1 − P0 (Ap ),
(25)
ρλ η 2/α
ρλ η 2/α 2
f (r) = 2π
r exp −π
r .
4
4
(26)
where f (r) is given by
Proof: Appendix B
Note that the expressions given in (24) or (25) are functions of the intensity ratio ρλ and power
ratio η. Note that if η = 1 the association probability reduces to the average area fraction provided
in Corollary 1. In the following, we derive the upper and lower bounds for the association
probability.
Corollary 2: The association probability P0 (Ap ) is lower and upper bounded by
P0 (Ap ) ≥ 1 +
π/2 − 1
e−πρ/2 − 1
+√ √
β − γ,
ρ
ρ( 2 − 1)
(27)
P0 (Ap ) ≤ 1 +
π/2 − 1
e−πρ/2 − 1
+√ √
β,
ρ
ρ( 2 − 1)
(28)
2 Rz 2
where ρ = ρλ η 2/α , erf(z) = √ 0 e−t dt,
π
p
p
β = erf( πρ/2) − erf( πρ/4),
γ = 0.0956(e−πρ/4 − e−πρ/2 ).
(29)
(30)
Proof: Appendix C.
Figure 4 demonstrates the association probability, its simulated value, and its lower and upper
bounds. The lower and upper bounds are very tight. The bounds are described very simple
function will show its best use to study the user association in repulsive networks.
Probability of being associated with Poisson point process
13
0.7
0.6
0.5
Bias toward Φ g @ ρ λ =1
0.4
0.36
∆(P(Ap ))=0.14
0.3
λ g =λ p ; E[Φ g ]=E[Φ p ]
0.2
Analytic expression
An analytic upper bound
An analytic lower bound
Simulation
Poisson point processes
0.1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Intensity ratio
Fig. 4. P(Ap ) (square mark) is juxtaposed with P(Ap0 ) (x mark) for ρλ ∈ (0.2, 2). For ρλ = 1, the proposed model with
repulsion gives P(Ap ) = 0.36 while the model with only PPPs produces P(Ap ) = 0.5.
Remark 2: To illustrate the influence of repulsion to user association, let us consider the same
transmit power for grid BSs and random BSs for the moment. In this case, for λp = λg ,
P(Ap ) = 0.36 and P(Ag ) = 0.64.
In other words, our model with repulsion yields P0 (Ap ) < P0 (Ag ). This contrasts to the user
association in cellular networks only modeled by independent PPPs. Specifically, the association
probability of the typical user in the independent PPP BSs Φp and Φp0 with intensity λp and λp0
was described by
P0 (Ap ) =
λp
λp0
and P0 (Ap0 ) =
.
λp0 + λp
λg + λp
(31)
In this case, for λp = λp0 ,
P0 (Ap ) = 0.5 and P0 (Ap0 ) = 0.5.
(32)
In other words, by comparing the proposed model to the model with only PPPs, even for the
same intensity, the typical user is more likely to be associated with the grid BSs. Figure 4 shows
14
P0 (Ap ) in the proposed model and in the model with only PPPs. The difference of association
probabilities is evident and it is alluded by the deterministic structure of the model. Specifically,
in the network model with the grid structure, a random user finds its associated BS at a distance
always less than infinity; √1
2
λg
in this paper. In other words, the void probability of the proposed
point process has a finite support, (0, √1 ). However, the void probability of the models with
2
λg
only PPPs has an infinite support, (0, ∞).
B. Coverage Probability under Exponential Fading
To elaborate the derivation of the coverage probability, consider two PPPs denoted by Φp and
Φp0 . The SIR coverage probability is given
P0 (SIR > T ) = P0 (SIR > T, Ap0 ) + P0 (SIR > T, Ap )
= P0 (SIR > T | A0p ) P0 (A0p ) + P0 (SIR > T | Ap ) P0 (Ap ),
(33)
where sometimes produces a simple closed-form expression [6].
On the other hand, in the proposed BS model the complete independence property [27, Chapter
2.] does not hold and (33) produces complex expression. In particular, given the association of
users, the formula for the interference needs to be derived from conditional point processes such
that there are no points of Φp and Φg inside a certain exclusion area.
to denote the conditional point processes of Φg and Φp such that no
and Φ!∼r
We use Φ!∼r
p
g
points of Φg and Φp exist inside the ball of radius r centered at zero, respectively. Note that users
are assumed to be associated with the BSs with the maximum average receive power. Therefore,
provided that the typical user is associated with a grid BS at distance kU k, the radius of the
exclusion ball for Φp is given by kU kη 1/α . In the same vein, provided that the typical user is
associated with a Poisson BS at the distance R, the radius of the exclusion ball for the grid
is given by Rη −1/α . In the following, we omit the Palm notation for simpler expression. The
15
coverage probability is given by
P (SIR > T, Ag )
P ? H ? /kX ? kα
PX X
> T, Rp > kX ? kη 1/α
α
PXi HXi /kXi k
= P
Xi ∈Φ\{X ? }
? α
P ? H ? /kX k
d
PX X
> T, Rp > kX ? kη 1/α X ? = U
α
PXi HXi /kXi k
(a)
= EU P
Xi ∈Φ\{X ? }
α
(b)
= EU P·|U
P
PU HU /kU k
1/α
1/α
P
R
>
kU
kη
>
T
R
>
kU
kη
p
p
·|U
PXi HXi /kXi kα
Xi ∈Φ\{U }
(c)
= EU
P·|U
P
pg HU /kU kα
> T
P·|U (Rp > kU kη 1/α )
α
PXi HXi /kXi k
(34)
!∼kU kη 1/α
Xi ∈Φg \{U }∪Φp
where (a) is given by conditioning on the random variable kU k which is given by (15), (b)
is obtained by the Bayes rule with the notation P·|U = P(·|U ), and (c) is acquired from
incorporating the exclusion ball. In the expectation of (34), we have
P·|U
P
pg HU /kU kα
>
T
PXi HXi /kXi kα
!∼kU kη 1/α
Xi ∈Φg \{U }∪Φp
(d)
= E·|U 1
α
pg HU /kU k
P
>T
PXi HXi /kXi kα
!∼kU kη 1/α
Xi ∈Φg \{U }∪Φp
α −1
= E·|U P H > T kU k pg
IΦg \{U } + I !∼kU kη1/α
Φp
(e)
α −1
= E·|U exp −T kU k pg
IΦg \{U } + I !∼kU kη1/α
Φp
= LIΦg \{U } T kU kα p−1
LI
g
!∼kU kη 1/α
Φp
T kU kα p−1
,
g
(35)
where (d) is given by P(A) = E[1{A}] and (e) is obtained by using CDF of exponential
distribution. In (35), LIΦg \{U } (·) denotes the Laplace transform of the interference created by
16
Φg \ {U }. Incorporating (35) to (34), we have
LI
P(SIR > T, Ag ) = E LIΦg \{U } T kU kα p−1
g
!∼kU kη 1/α
Φp
T kU kα p−1
g
P(Rp > kU kη
1/α
) ,
(36)
where
P·|U (Rp > kU kη 1/α ) = P(Rp > kU kη 1/α |R) = exp(−πλp kU k2 η 2/α )
and fkU k (u) = fRg (u) which is given in (15).
Consequently, the derivation of the coverage expression is reduced into obtaining the condi!∼kU kη 1/α
tional expectations of functionals on the conditional point processes: Φg \ {U } and Φp
.
Similarly, we have
P (SIR > T, Ap )
P ? H ? /kX ? kα
PX X
> T, Rg > kX ? kη −1/α
PXi HXi /kXi kα
= P
(37)
Xi ∈Φ\{X ? }
= ER P
P ? H ? /kX ? kα
d
PX X
> T, Rg > kX ? kη −1/α kX ? k = R
α
PXi HXi /kXi k
(38)
Xi ∈Φ\{X ? }
P
= ER
·|R
P
pp HR /kRkα
P·|R (Rg > Rη −1/α ) .
>
T
PXi HXi /kXi kα
Xi ∈Φ!∼R
∪Φ!∼Rη
p
p
(39)
−1/α
Similarly to the above, we have
h
i
α −1
−1/α
P(SIR > T, Ap ) = ER LIΦ \Rη−1/α T Rα p−1
L
T
R
p
P
(R
>
Rη
)
, (40)
IΦ!∼R
g
·|R
p
p
g
p
where
P·|R (Rg > Rη −1/α ) = P(Rg > Rη −1/α |R)
=
ν2 (S0 \ B(0, Rη −1/α ))
,
ν2 (S0 )
FR (r) = 2πλp r exp(−πλp r2 ).
Consequently, deriving the coverage probability of the typical user is reduced into obtaining the
conditional expectations of the functionals on the conditional point processes, Φ!∼Rη
g
Φ!∼R
p .
−1/α
and
17
C. Interference Seen at the Typical User
∼kU kη 1/α
The Laplace transforms of interference of Φ∼R
p , Φp
, Φ!∼Rη
g
−1/α
and Φg \{U } are given.
∼kU kη 1/α
Φp
Lemma 2: The Laplace transforms of interference from Φ∼R
and
p
Z ∞
pp ξu1−α
LIΦ∼R (ξ) = exp −2πλp
du ,
−α
p
R 1 + pp ξu
Z ∞
pp ξu1−α
LI ∼kU kη1/α (ξ) = exp −2πλp
du .
−α
Φp
kU kη 1/α 1 + pp ξu
are
(41)
(42)
Proof: Since we assumed a homogeneous PPP, applying the radius of the exclusion ball
and the Markov property directly produces the expressions.
−1/α
and Φg \ {U } are
Lemma 3: The Laplace transforms of interference from Φ!∼Rη
g
Z
Y
1
LI !∼Rη−1/α (ξ) =
PU 0 (du0 ),
0
−α
1
+
p
ξku
+
s(z
,
z
)k
−1/α
Φg
g
1
2
S0 \B(0,Rη
)
2
(43)
(z1 ,z2 )∈Z
Y
LIΦg \{U } (ξ) =
(z1 ,z2 )∈Z2 \(0,0)
1
,
1 + pg ξkU + s(z1 , z2 )k−α
(44)
d
where U 0 = Uniform[S0 \ B(0, rη −1/α )].
Proof: Appendix D.
D. Scaling of the Coverage Probability
Theorem 2: The coverage probability of the typical user is given by
Z
pcov =
2 η 2/α
e−πρλ kûk
Sˆ0
1
Y
(z1 ,z2 )∈Z2
1+
T kûkα
kû+(z1 ,z2 )kα
e
−2πρλ
T ηkûkα w1−α
kûkη 1/α 1+T ηsα kûkα w−α
R∞
dw
PÛ (dû)
6=(0,0)
Z
+
0
∞
Z
T r̂ α w1−α
ν2 (Sˆ00 ) −2πρλ Rr̂∞ 1+T
α w−α dw
r̂
e
ν2 (Sˆ0 )
Ŝ0 \B(r̂η −1/α )
2
Z
Y
PÛ 0 (dû0 )
1+
(z1 ,z2 )
η −1 T r̂α
kû+(z1 ,z2 )kα
fRˆp (r̂) dr̂,
(45)
0
where Û 0 = Uniform[Sˆ0 ] , Sˆ00 = Sˆ0 \ B(0, r̂η −1/α ), Û = Uniform[Sˆ0 ] , Sˆ0 := [−1/2, 1/2] ×
2
[−1/2, 1/2], and fRˆp (r̂) = 2πρλ r̂e−πρλ r̂ .
Proof: Appendix E
The coverage expression for the typical user is exactly found in Theorem 2 and it accurately
quantifies the influence of repulsion provided that ρλ of a network can be accurately estimated.
From (22), the one-to-one relationship between a pair-wise correlation function and the ρλ is
18
given. Therefore, in any network topology where the correlations function can be estimated,
equation (45) can be utilized to predict the SIR of the typical user.
Remark 3: From Theorem 2, we confirm that the proposed model is scalable; the coverage
probability of the typical user is invariant w.r.t. the intensity ratio ρλ . This is similar to the
homogeneous Poisson network [8] where the coverage probability is invariant w.r.t. intensity λ.
Note that, however, the path loss function we considered is a distance-based path loss function
with a single exponent and the specific path loss function is known to produce the favorable SIR
scale invariant result of Poisson cellular networks given in [8] because the order of interference
and the order of signal are the same as the density increases [29]. In practical studies, it is
well known that the received signal power is bounded from above. As in [30], under a practical
dual slope truncated path loss function, the scale invariant property of SIR coverage probability
does not hold. In the same vein, the scaling-invariance property in this paper disappears under
different path loss functions and we will numerically examine those functions in Section IV-F.
E. Bounds on Interference and Coverage Expressions
In order to maintain the applicability of the proposed model and to address computational
issues, bounds for interference and SIR coverage will be considered.
Proposition 2: Consider an arbitrary square truncation window W . The Laplace transform of
interference from Φg conditioning on Ag yields
P
− z
ξpg
2
α
1 ,z2 ∈Z \(0,0) kU +s(z1 ,z2 )k
LIΦg \{U } (ξ) ≥ e
,
(46)
1
.
1 + ξpg kU + s(z1 , z2 )k−α
Y
LIΦg \{U } (ξ) ≤
2
(z1 ,z2 )∈Z ∩W
(47)
On the other hand, the Laplace transform of interference from Φg conditioning on Ap yields
2π
LI
−
(ξ) ≥ e
Z
LI !∼Rη−1/α (ξ) ≤
R∞
Rη −1/α
!∼Rη −1/α
Φg
Φg
S00
v log 1+ξpg v −α
(
0)
ν2 (S0
Y
2
(z1 ,z2 )∈Z ∩W
) dv
,
1
PU 0 (dv).
1 + ξpg kv + s(z1 , z2 )k−α
(48)
(49)
Proof: Given in Appendix F
As the size of window kW k increases, upper bounds (47) and (49) become tighter at the expense
of a marginal computational complexity. We will see the tightness of upper bounds in Figure 6.
In order to derive the simplest coverage expression, W = S0 for the moment.
19
Proposition 3: The coverage probability of the typical user is lower bounded by
Z
R
PZ2 \(0,0)
ηT kukα v 1−α
T kukα
dv
−2πλp ∞ 1/α
−πλp kuk2 η 2/α − z1 ,z2
ku+s(z1 ,z2 )kα
1+ηT
kukα v −α
kukη
pcov ≥
e
e
PU (du)
e
S0
Z
∞
+
0
ν2 (S00 ) −
e
ν2 (S0 )
2π
R∞
u log 1+η −1 T r α u−α
rη −1/α
0)
ν2 (S0
(
) du
−2πλp
R∞
r
T r α u1−α
1+T r α u−α
du
fR (r) dr.
(50)
The coverage probability of the typical user is upper bounded by
Z
pcov ≤
−2πλp
e−πλp
S0
Z ∞
+
0
kuk2 η 2/α
e
ηT kukα v 1−α
kukη 1/α 1+ηT kukα v −α
R∞
1+T
Z
1−α
T rα v
ν2 (S00 ) −2πλp Rr∞ 1+T
dv
r α v −α
e
ν2 (S0 )
S0
dv
PU (du)
1
PU 0 (du) fRp (r) dr.
(1 + η −1 T rα kuk−α )
(51)
Proof: The expressions (46) and (47) evaluated at ξ = T kU kp−1
g yield the lower and upper
bounds for P(SIR > T, Ag ). Similarly, the expressions (48) and (49) evaluated at ξ = T Rp p−1
p
produces lower and upper bounds for P(SIR > T, Ap ).
The expression (51) requires much less computation compared to (45) because we essentially
replace the summation of infinite grid interference by the strongest grid interference. Numerical
validation in Section IV-F will justify it. It supports that interference from longer distances
becomes much smaller and it allows the approximation of total interference to the strongest
interference for the grid BSs.
F. Numerical Analysis and Validation
We compare the coverage probability bounds to the numerically obtained coverage probability
of the typical user. We perform the Monte Carlo simulation where the number of iterations is
more than 105 . We depict the coverage probability (45), the lower bound (50), and the upper
bound (51) in Fig. 5 and 6. These figures demonstrate that the lower bound and the upper
bounds are very tight for a wide range of SIR thresholds. The bounds become tighter as the
window size W increases. The acquired bounds show that interference from further BSs is very
small and we can approximate the total interference by the strongest interference. The variable
ρλ play a more noticeable role for higher α.
20
1
,=4 ; 6 =4 (Sim)
,=4 ; 6 =4 (LB)
0.9
,=4,; =0.25 (Sim)
Coverage probability of the typical user
6
,=4,; 6 =0.25 (LB)
0.8
,=3, ; 6 =4 (Sim)
,=4
0.7
,=3, ; 6 =4 (LB)
,=3,; 6 =0.25 (Sim)
,=3,; 6 =0.25 (LB)
0.6
0.5
0.4
0.3
,=3
0.2
0.1
0
-20
-15
-10
-5
0
5
10
15
20
Threshold (dB)
Fig. 5. This figure describes the coverage probabilities of the simulation (Sim) and corresponding lower bounds. It delineates
α = 3, 4 and ρλ = 0.25, 4. We consider W = [−3s/2, 3s/2]2 . To minimize the undesired edge effect, The average number of
BSs is maintained to be 100 in the simulation window, E[Φ(K)] = λg + λp = 100 for the multiple values of ρλ .
In Fig. 7, the coverage probability of the typical user is investigated under different practical
path loss functions such as
l1 (r) = min(C0 , rα ),
min(C0 , r−α1 )
l2 (r) =
min(C1 r−α1 , C2 r−α2 )
(52)
for 0 < r < r1
(53)
for r1 < r
where C0 , C1 , C2 , r1 > 0 are arbitrary constants that ensure the functions continuous. A more
rigorous analysis of coverage probability under different path loss functions is left as future
work.
Fig. 8 illustrates the coverage probability of the typical user for different values of ρλ . It
demonstrates only Poisson (ρλ ≈ ∞) and only grid BS (ρλ ≈ 0) for α = 4. The coverage
probability of the proposed network is located between that of grid BSs and that of PPP BSs.
Remark 4: In Fig. 8 and 9, the coverage probability is shown to be increasing w.r.t ρ−1
λ . The
behavior of the typical user in the proposed model contrasts to the behavior of the typical user in
21
1
,=4,; 6 =2 (Sim)
Coverage probability of the typical user
0.9
,=4,; =2 (UB)
6
,=3,; =2 (Sim)
6
0.8
,=3,; =2 (UB)
,=4
6
,=4,; =1 (Sim)
6
0.7
,=4,; =1 (UB)
6
,=3,; 6 =1 (Sim)
0.6
,=3,; 6 =1 (UB)
,=4,; 6 =0.5 (Sim)
0.5
,=4,; 6 =0.5 (UB)
,=3,; 6 =0.5 (Sim)
0.4
,=3,; 6 =0.5 (UB)
0.3
0.2
,=3
0.1
0
-20
-15
-10
-5
0
5
10
15
20
Threshold (dB)
Fig. 6. It describes the coverage probabilities of simulations (Sim) and the corresponding upper bounds (UB). It delineates
α = 3, 4 and ρλ = 0.5, 2. We assume W = [−3s/2, 3s/2]2 . To minimize the undesired edge effect, we maintain the average
number of BSs in the simulation space at 100, E[Φ(K)] = λg + λp = 100 for the multiple values of ρλ .
Poisson networks [8] where the SIR does not scale w.r.t. the number of BSs. In other words, if
ρλ decreases, i.e., more repulsive BSs are added, the SIR increases. This observation corresponds
to the remark made in [8] that the fixed grid is the performance upper bound of the coverage
probability. A more rigorous analysis of the growth rate of coverage probability is left for future
work.
G. Fitness of the Model: Actual BS Deployment Data and Coverage Probability
In this section, we conduct an empirical study based on real BS deployment data shown in Fig.
1 and 2. For each data set, we estimate the averaged pair correlation value κ and the intensity
ratio ρλ as follows.
•
Step 1. The pair correlation function is estimated using open source software R.
•
Step 2. ρλ is obtained using (22).
•
Step 3. The coverage probability of the typical user is obtained using Theorem 2.
22
coverage probability (Network densitifaction when 6 g =25)
1
Coverage probability of the typical user
; 6 =0.1
truncated
dual slope
0.9
; =0.5
6
; =1.0
6
0.8
; =2.0
truncated single slope
6
; =10
6
0.7
; =0.1
6
; 6 =0.5
0.6
; 6 =1.0
; 6 =2.0
0.5
; 6 =10
0.4
0.3
0.2
0.1
0
-40
-30
-20
-10
0
10
20
30
SINR when SNR is 132dB with Pt =23dBm, 10MHz
Fig. 7. The coverage probability of the typical user where the path loss functions (52) and (53) are considered.
1
Poisson
; 6 : 0.5
Coverage probability of the typical user
0.9
;6 : 1
0.8
; 6: 2
; -1
:1
6
; 6: 4
Grid
0.7
; -1
=4
6
0.6
; -1
=1
6
0.5
0.4
; -1
=0.5
6
0.3
; -1
:0
6
0.2
0.1
0
-30
-20
-10
0
10
20
SIR threshold (dB)
Fig. 8. The coverage probability of the typical user increases as the intensity ratio ρλ decreases.
30
23
1
0.9
0.8
,=2
,=2.5
,=3
,=3.5
,=4
P(SIR>1)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10 -2
10 -1
10 0
Inverse of intensity ratio
Fig. 9. The typical SIR coverage probability with threshold 1 for multiple values of α.
Fig. 10 and 11 show the coverage probability obtained from actual deployment and the coverage probability obtained from the proposed model. The figures demonstrate that the proposed
model predicts the coverage probability of the actual deployment very accurately. In addition, the
parameter ρλ of the proposed model is robust and consistent with the change of α. It contrasts
to the β-Ginibre point process model in [22] where the key parameter β should be modified
for different path loss exponents; this modification compromises the robustness of the Ginibre
model.
V. C ONCLUSION
We developed a scalable modeling technique representing the repulsive BSs and random BSs.
We derived the nearest BS distance, the cellular user association probability, the interference
seen at the typical user, and the SIR coverage probability. Numerical observations confirmed
that the provided bounds for numerical purposes are tight. Our findings include the following
three points. First, the typical user is more likely to be associated with a repulsive BS rather
than a random BS. Second, we have the expressions for the user association and the coverage
24
1
Proposed model ( ,=4, ; 6 =2.3)
actual deployment (Fig1, ,=4)
Proposed model ( ,=3.5, ; 6 =2.3)
Coverge probability of the typical user
0.9
actual deployment (Fig1, ,=3.5)
Proposed model ( ,=3, ; 6 =2.3)
0.8
actual deployment (Fig1, ,=3)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-20
-15
-10
-5
0
5
10
15
20
SIR threshold (dB)
Fig. 10. The coverage probability of the actual deployment given in Figure 1 and proposed model with ρλ = 2.3. The proposed
model predicts the actual coverage probability very accurately. Note that the ρλ is obtained from physical deployment and the
model is robust and consistent for α = 3, 3.5, and 4.
given by ρλ and η. Finally, the proposed model accurately predicts the coverage probability of
the actual cellular deployment.
A PPENDIX A
P ROOF OF L EMMA 1
Denote a ball of radius r centered at x in R2 by B(x, r). The nearest distance Rg is
Rg = inf{R > 0|Φg (B(0, R)) 6= ∅}
(a)
= inf{R > 0| ∪Xi ∈Φg B(Xi , R) ∩ 0 6= ∅}.
(54)
where (a) is obtained by using the stationarity of Φg . Let us define a square of width s centered
at the origin
s s s s
S0 = − ,
× − ,
.
2 2
2 2
(55)
25
1
Proposed model ( ,=4, ; 6 =1.1)
actual deployment (Figure 2, ,=4)
Proposed model ( ,=3.5, ; 6 =1.1)
Coverage probability of the typical user
0.9
actual deployment (Figure 2, ,=3.5)
Proposed model ( ,=3, ; 6 =1.1)
0.8
actual deployment (Figure 2, ,=3)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-20
-15
-10
-5
0
5
10
15
20
SIR threshold (dB)
Fig. 11. The coverage probability of the actual deployment given in Figure 1 and proposed model with ρλ = 1.1. The proposed
model predicts the actual coverage probability very accurately. Note that the ρλ is obtained from physical deployment and the
model is robust and consistent for α = 3, 3.5, and 4..
Then, the CDF of Rg is exactly obtained by
P(Rg < r) = P(inf{R > 0| ∪Xi ∈Φg B(Xi , R) ∩ 0 6= ∅} < r)
ν2 (S0 ∩ B(0, r))
ν2 (S0 )
0
2
πr
s2
(c)
q
=
s
π−4 arccos( 2r
2
2
)
r
2 s2
− 2 + 4rs2 − 1
2
1
(b)
=
(56)
if r < 0
if 0 ≤ r ≤
if
s
2
<r≤
if r >
s
2
√s ,
2
(57)
√s ,
2
where ν2 (A) denotes the area of set A. We have (b) from the fact that the CDF of the first touch
distance to a stationary point process is equivalent to the volume fraction of the stationary point
process [25, Section 6.3.1] and have (c) by merely computing the area. Taking the derivative of
(14) w.r.t. r completes the proof.
26
A PPENDIX B
P ROOF OF T HEOREM 1
If the typical user is associated with the nearest BS of the PPP at distance Rp , it implies that
received power from the nearest PPP is greater than the received power from the nearest grid.
The event Ap is given by
Ap =
pp
pg
>
Rpα
Rgα
,
(58)
= Acp
(59)
whereas the event Ag is given by
Ag =
pp
pg
< α
α
Rp
Rg
since U = Ap ∪ Ag . Denoting the indicator function by 1{·}, we have
pp
pg
0
0
P (Ap ) = E 1
> α
Rpα
Rg
Z ∞
=
P Rg ≥ rη −1/α fRp (r) dr
(60)
0
(a)
∞
Z
=
0
ν2 (S0 \ B(0, rη −1/α ))
2
2πλp re−πλp r dr
ν2 (S0 )
s 1/α
η
2
Z
=
2
(s2 − πr2 η −2/α )2πλp re−πλp r dr
0
√s η 1/α
2
Z
+
s2 − 4rs sin(θ) − 2r2
s 1/α
η
2
√
(b)
Z
=
0
2
√
π
(1 − r2 )f (r) dr +
4
Z
2
π
2
− 2θ
2
2πλp re−πλp r dr
r2 arccos(1/r) −
√
r2 − 1 f (r) dr,
(61)
1
where (a) is given from (56) and (b) is given by change of variables. From equation (61), P0 (Ag )
is obtained.
A PPENDIX C
P ROOF OF C OROLLARY 2
We develop the bounds by obtaining lower and upper bounds for the second integration in
(24). We consider
g(r) = r2 arccos(1/r) −
for 1 ≤ r ≤
√
r2 − 1,
√
2. One can easily check that g(r) is a differentiable function w.r.t r and that
g(r) is a convex function since its second derivative w.r.t r is greater than zero for the interval
27
√
(1, 2). Then, by fully exploiting the differentiability and the convexity of g(r), we propose
two linear bounds
√
g( 2) − g(1)
√
g(r) ≥
(r − 1.19) + 0.1662,
2−1
√
g( 2) − g(1)
√
(r − 1),
g(r) ≤
2−1
(62)
(63)
√
for 1 ≤ r ≤ 2. We obtain the lower bound by applying the mean value theorem to the interval
√
(1, 2) by approximating the tangential line numerically. We obtain the upper bound by finding
the smallest linear function greater than g(r). Replacing g(r) of (61) with (62) and (63), we
acquire the lower and upper bounds, respectively.
A PPENDIX D
P ROOF OF L EMMA 3
For the Laplace transform of the interference from the grid, we have
−s P
p HkXi k−α
!∼Rη −1/α g
Xi ∈Φg
LI !∼Rη−1/α (ξ) = E·|R e
Φg
−s P
p HkXi k−α
(a)
!∼Rη −1/α g
Xi ∈Φg
= EΦg |R EH|R e
|Φg
(b)
1
1 + pg ξkXi k−α
Y
= EΦg |R
Xi ∈Φ!∼Rη
g
−1/α
2
(c)
= EU 0 |R
Z
Y
(z1 ,z2 )
Z
=
1 + pg
ξ||u0
1
+ s(z1 , z2 )||−α
2
Z
Y
S0 \B(0,Rη −1/α ) (z ,z )
1 2
1
PU 0 (dv),
1 + pg ξkv + s(z1 , z2 )k−α
(64)
where (a) is obtained by conditioning on the locations of the grid, (b) is obtained by the Laplace
transform of an exponential distribution with parameter one, and (c) follows from the fact that all
points of grid are measurable w.r.t the uniform random variable U 0 that takes a value uniformly
in the set S00 = S0 \ B(0, Rη −1/α ).
28
On the other hand, we have
h P
i
−s X ∈Φg \{U } pg H/kXi kα
i
LIΦg \kU k (ξ) = E·|U e
ii
h h P
−s X ∈Φg \{u} pg HkXi k−α
i
= E·|U E e
2
=
Z
Y
−α
E·|U [e−sHpg kXi k ]
(z1 ,z2 )6=(0,0)
2
=
Z
Y
(z1 ,z2 )6=(0,0)
1
.
1 + pg ξkU + s(z1 , z2 )k−α
(65)
Note that given that the typical user is associated with kU k, the nearest BS of the grid does not
comprise the interference. Since Φg is measurable w.r.t the associated BS at U , we have (65)
A PPENDIX E
P ROOF OF T HEOREM 2
The expressions (42) and (44) evaluated at ξ = T kukα p−1
g with
P(Rg > rη −1/α ) =
ν2 (S0 \ B(0, rη −1/α ))
ν2 (S 0 )
=
,
ν2 (S0 )
ν2 (S0 )
(66)
gives P(SIR > T, Ag ). Similarly, expressions (41) and (43) evaluated at ξ = T rα p−1
p with
2 η 2/α
P(Rp > kukη 1/α ) = e−πλp kuk
(67)
give P(SIR > T, Ap ). Therefore, the coverage probability is given by
Z
Z2
R∞
Y
T ηkukα v 1−α
1
2 2/α −2πλp
dv
kukη 1/α 1+T ηkukα v −α
PU (du)
pcov =
e−πλp kuk η e
T kukα
1 + ||u+s(z1 ,z2 )||α
S0
(z1 ,z2 )6=(0,0)
Z Y
Z ∞
Z2
R ∞ T rα v1−α
0
ν2 (S0 ) −2πλp r 1+T rα v−α dv
e
PU 0 (dv) fR (r) dr
(68)
+
η −1 T rα
ν
(S
)
0
2
0
1+
S
0
α
(z1 ,z2 )
||v+s(z1 ,z2 )||
Furthermore, by change of variable, the first term is given by
Z
R
Y
T η||u||α v 1−α
1
2 2/α
−2πλp ∞ 1/α
dv
1+T ηkukα v −α
kukη
e−πλp kuk η
e
PU (du)
T kukα
1
+
S0
α
2
||u+s(z1 ,z2 )||
(z1 ,z2 )∈Z \(0,0)
Z
R
Y
sα T ηkûkα v 1−α
1
2 2/α
dv
−2πλp ∞ 1/α
u=sû
1+T ηsα kûkα v −α
skûkη
=
e−πρλ kûk η
PÛ (dû)
e
T kûkα
1
+
Sˆ0
α
2
kû+(z1 ,z2 )k
(z1 ,z2 )∈Z \(0,0)
Z
R
Y
T ηkûkα w1−α
1
2 2/α
−2πρλ ∞ 1/α
dw
v=sw
1+T ηsα kûkα w−α
kûkη
=
e−πρλ kûk η
e
PÛ (dû),
T kûkα
1
+
Sˆ0
kû+(z1 ,z2 )kα
(z ,z )∈Z2 \(0,0)
1
2
29
where Û = Uniform([− 12 , 21 ] × [− 21 , 12 ]) and then PÛ (du) = 1 · dx dy.
In the same vein, the second term is given by
Z ∞
Z
Z2
R ∞ T rα v1−α
0
Y
ν2 (S0 ) −2πλp r 1+T rα v−α dv
e
ν2 (S0 )
0
S0 \B(0,rη −1/α ) (z ,z ) 1 +
1 2
∞
Z
u0 =sû0
=
0
r=sr̂
∞
Z
=
0
T r α v 1−α
ν2 (S00 ) −2πλp Rr∞ 1+T
dv
r α v −α
e
ν2 (S0 )
1
η −1 T rα
||u0 +s(z1 ,z2 )||α
2
Z
Y
Z
−1/α
Sˆ0 \B(0, rη s
)
(z1 ,z2 )
s
1+
Z
=
0
∞
R
α 1−α
ν2 (Sˆ00 )
2 −2πρλ ∞ T r̂ w
r̂ 1+T r̂ α w−α dw
2πρλ r̂e−πρλ r̂ e
ν2 (Sˆ0 )
s−α η −1 T rα
||û0 +(z1 ,z2 )||α
PÛ 0 (dû0 ) fRp (r) dr
2
Z
R ∞ sα T r̂α v1−α
ν2 (Sˆ00 )
−πρλ r̂2 −2πλp sr̂ 1+sα T r̂α v−α dv
2πρλ r̂e
e
ν2 (Sˆ0 )
Z
Y
Ŝ0 \B(0,r̂η −1/α )
v=sw
PU 0 (du0 ) fR (r) dr
1
(z1 ,z2 )
1+
η −1 T r̂α
kû+(z1 ,z2 )kα
PÛ 0 (dû0 )
2
Z
Y
Z
Ŝ0 \B(0,r̂η −1/α )
(z1 ,z2 )
1
1+
η −1 T r̂α
kû+(z1 ,z2 )kα
PÛ 0 (dû0 ) dr̂,
where Û 0 = Uniform([− 12 , 21 ] × [− 21 , 12 ] \ B(0, r̂η −1/α )). Consequently, the coverage probability
of the typical user is given by a function of ρλ and η.
A PPENDIX F
P ROOF OF P ROPOSITION 2
First, we develop upper and lower bounds for the Laplace transform of the interference given
in (36). Note that the interference is comprised of signals from both grid and PPP. The Laplace
transform of interference created by the PPP is well-known and does not require a bound. We
need to find the upper and lower bounds for the interference created by the grid in (36). We
have
P
LIΦg \{U } (ξ) = e
≥ e−
where we use log
1
1+x
(z1 ,z2 )∈Z2 ∩W \(0,0)
P
log
(z1 ,z2 )∈Z2 ∩W \(0,0)
1
1+ξpg kU +s(z1 ,z2 )k−α
ξpg kU +s(z1 ,z2 )k−α
,
(69)
≥ −x for x ∈ (0, ∞).
In order to find an upper bound for (36), we define a measurable function ϕ(·) such that
Y
1
ϕ(U, ξ) =
.
1 + ξpg ||U + s(z1 , z2 )||−α
2
(z1 ,z2 )∈Z
Since the element 1/(1 + ξpg kU + s(z1 , z2 )k−α ) takes a value between zero and one, we have
Y
1
LIΦg \{U } (ξ) = ϕ(U, ξ) ≤
,
(70)
1 + ξpg kU k−α
2
(z1 ,z2 )∈Z ∩W
30
where the inequality follows by truncating the function ϕ(U, ξ) by finite terms. We denote the
truncation window by W . Applying (69) or (70) into (36), we obtain the lower and the upper
bounds, respectively.
Secondly, we develop lower and upper bounds for the Laplace transform of the interference
created by the grid. In order to find the lower bound for (40), we have
P
log
!∼Rη −1/α
Xi ∈Φg
LI
!∼Rη −1/α
Φg
1
1+ξpg kXi k−α
(ξ) = E·|R e
R
−1/α
− R2 log(1+ξpg kxk−α )Φ!∼Rη
(dx)
g
,
= E·|R e
Moreover,
Z
ξpg
!∼Rη −1/α
log 1 +
E·|R exp −
Φg
(dx)
kxkα
R2
Z
ξpg
!∼Rη −1/α
Φg
(dx)
≥ exp − E·|R
log 1 +
kxkα
R2
Z
ξpg
!∼Rη −1/α
E[Φg
(dx)] ,
= exp −
log 1 +
kxkα
R2
(71)
where we use Jensen’s inequality and Campbells mean value formula [7]. We denote the conditional intensity measure of the grid by M
Φ!∼Rη
g
M
Φ!∼Rη
g
−1/α
1{x ∈/
−1/α
(dx) and it is
B(Xi , Rη −1/α )}
dx
ν2 (S0 \ B(Rη −1/α ))
(dx) =
S
Xi ∈s·Z2
1{x ∈ R2 \B(0, Rη −1/α )}
≤
ν2 (S0 \ B(0, Rη −1/α ))
dx.
(72)
In other words, the intensity measure is upper bounded by the same uniform measure in S0 \
B(0, rη −1/α ) that is nonzero for all the grid points except the origin. By inserting (72) into (71),
we have
2π
LI
−
!∼rη −1/α
Φg
(ξ) ≥ e
ξpg
u log 1+ α
du
u
Rη −1/α
−1/α
ν2 (S0 \B(0,Rη
))
R∞
.
In order to find the upper bound for (40), we use the measurable function ϕ(·). Then,
Y
1
ϕ(U 0 , ξ) =
0
1 + ξpg ||U + s(z1 , z2 )||−α
2
(z1 ,z2 )∈Z
≤
Y
(z1 ,z2 )∈Z2 ∩W
1 + ξpg
kU 0
1
,
+ s(z1 , z2 )k−α
(73)
(74)
31
by considering the window W . Using ϕ(U 0 ) ≤ ϕtrun (U 0 ) =⇒ E[ϕ(U )] ≤ E[ϕtrun (U 0 )], we have
Z
Y
1
LI !∼rη−1/α (ξ) ≤
PU 0 (du).
(75)
1 + ξpg ku + s(z1 , z2 )k−α
Φg
S0 \B(0,Rη −1/α )
2
(z1 ,z2 )∈Z ∩W
Applying the expression given in (73) or (75) to (40), we obtain the lower and upper bounds.
R EFERENCES
[1] C. S. Choi, J. O. Woo, and J. G. Andrews, “On the coverage probability of a spatially correlated network,” in Proc. IEEE
ISIT, June 2017, pp. 466–470.
[2] C.-H. Lee, C.-Y. Shih, and Y.-S. Chen, “Stochastic geometry based models for modeling cellular networks in urban areas,”
Wireless Networks, vol. 19, no. 6, pp. 1063–1072, 2013.
[3] Y. J. Chun, M. O. Hasna, and A. Ghrayeb, “Modeling heterogeneous cellular networks interference using Poisson cluster
processes,” IEEE J. Sel. Areas Commun., vol. 33, no. 10, pp. 2182–2195, Oct. 2015.
[4] D. B. Taylor, H. S. Dhillon, T. D. Novlan, and J. G. Andrews, “Pairwise interaction processes for modeling cellular network
topology,” in Proc. IEEE GLOBECOM, 2012, pp. 4524–4529.
[5] I. Nakata and N. Miyoshi, “Spatial stochastic models for analysis of heterogeneous cellular networks with repulsively
deployed base stations,” Performance Evaluation, vol. 78, pp. 7–17, 2014.
[6] M. Haenggi and R. K. Ganti, Interference in large wireless networks.
Now Publishers Inc, 2009.
[7] F. Baccelli and B. Blaszczyszyn, Stochastic geometry and wireless networks: Theory. Now Publishers Inc, 2009, vol. 1.
[8] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans.
Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011.
[9] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and analysis of K-tier downlink heterogeneous
cellular networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 550–560, Apr. 2012.
[10] H.-S. Jo, Y. J. Sang, P. Xia, and J. G. Andrews, “Heterogeneous cellular networks with flexible cell association: A
comprehensive downlink SINR analysis,” IEEE Trans. Wireless Commun., vol. 11, no. 10, pp. 3484–3495, Oct. 2012.
[11] S. Mukherjee, “Distribution of downlink SINR in heterogeneous cellular networks,” IEEE J. Sel. Areas Commun., vol. 30,
no. 3, pp. 575–585, Apr. 2012.
[12] H. S. Dhillon, R. K. Ganti, and J. G. Andrews, “Load-aware modeling and analysis of heterogeneous cellular networks,”
IEEE Trans. Wireless Commun., vol. 12, no. 4, pp. 1666–1677, Apr. 2013.
[13] Q. Ye, B. Rong, Y. Chen, M. Al-Shalash, C. Caramanis, and J. G. Andrews, “User association for load balancing in
heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 2706–2716, Jun. 2013.
[14] M. Di Renzo, A. Guidotti, and G. E. Corazza, “Average rate of downlink heterogeneous cellular networks over generalized
fading channels: A stochastic geometry approach,” IEEE Trans. Commun., vol. 61, no. 7, pp. 3050–3071, Jul. 2013.
[15] Y. S. Soh, T. Q. Quek, M. Kountouris, and H. Shin, “Energy efficient heterogeneous cellular networks,” IEEE J. Sel. Areas
Commun., vol. 31, no. 5, pp. 840–850, May 2013.
[16] K. Huang and V. K. Lau, “Enabling wireless power transfer in cellular networks: Architecture, modeling and deployment,”
IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 902–912, Feb. 2014.
[17] T. Bai and R. W. Heath, “Coverage and rate analysis for millimeter-wave cellular networks,” IEEE Tran. Wireless Commun.,
vol. 14, no. 2, pp. 1100–1114, 2015.
[18] R. W. Heath, M. Kountouris, and T. Bai, “Modeling heterogeneous network interference using Poisson point processes,”
IEEE Trans. Signal Processing, vol. 61, no. 16, pp. 4114–4126, Aug. 2013.
32
[19] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Using Poisson processes to model lattice cellular networks,” in Proc.
IEEE INFOCOM, 2013, pp. 773–781.
[20] A. Guo and M. Haenggi, “Asymptotic deployment gain: A simple approach to characterize the SINR distribution in general
cellular networks,” IEEE Trans. Commun., vol. 63, no. 3, pp. 962–976, Mar. 2015.
[21] H. Wei, N. Deng, W. Zhou, and M. Haenggi, “Approximate SIR analysis in general heterogeneous cellular networks,”
IEEE Trans. Commun., vol. 64, no. 3, pp. 1259–1273, Mar. 2016.
[22] N. Deng, W. Zhou, and M. Haenggi, “The Ginibre point process as a model for wireless networks with repulsion,” IEEE
Trans. Wireless Commun., vol. 14, no. 1, pp. 107–121, Jan 2015.
[23] Y. Li, F. Baccelli, H. S. Dhillon, and J. G. Andrews, “Statistical modeling and probabilistic analysis of cellular networks
with determinantal point processes,” IEEE Trans. Commun., vol. 63, no. 9, pp. 3405–3422, Sept. 2015.
[24] A. M. Ibrahim, T. ElBatt, and A. El-Keyi, “Coverage probability analysis for wireless networks using repulsive point
processes,” arXiv preprint arXiv:1309.3597, 2013.
[25] S. N. Chiu, D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry and its applications. John Wiley & Sons, 2013.
[26] M. Haenggi, “Interference in lattice networks,” arXiv preprint arXiv:1004.0027, 2010.
[27] D. J. Daley and D. Vere-Jones, An introduction to the theory of point processes: volume I: general theory and structure.
Springer Science & Business Media, 2007.
[28] P. Mitran and C. Rosenberg, “On fractional frequency reuse in imperfect cellular grids,” in Proc. IEEE WCNC, April 2012,
pp. 2967–2972.
[29] F. Baccelli and A. Biswas, “On scaling limits of power law shot-noise fields,” Stochastic Models, vol. 31, no. 2, pp.
187–207, 2015.
[30] X. Zhang and J. G. Andrews, “Downlink cellular network analysis with multi-slope path loss models,” IEEE Trans.
Commun., vol. 63, no. 5, pp. 1881–1894, May 2015.
| 7cs.IT
|
arXiv:1603.03220v5 [math.ST] 27 Dec 2017
Convergence Rates for a Class of Estimators Based on
Stein’s Method
Chris J. Oates1,4 , Jon Cockayne2 , François-Xavier Briol2 , Mark Girolami3,4
1
Newcastle University
2
University of Warwick
3
Imperial College London
4
Alan Turing Institute
December 29, 2017
Abstract
Gradient information on the sampling distribution can be used to reduce the variance of Monte Carlo estimators via Stein’s method. An important application is that of
estimating an expectation of a test function along the sample path of a Markov chain,
where gradient information enables convergence rate improvement at the cost of a linear
system which must be solved. The contribution of this paper is to establish theoretical
bounds on convergence rates for a class of estimators based on Stein’s method. Our
analysis accounts for (i) the degree of smoothness of the sampling distribution and test
function, (ii) the dimension of the state space, and (iii) the case of non-independent
samples arising from a Markov chain. These results provide insight into the rapid convergence of gradient-based estimators observed for low-dimensional problems, as well
as clarifying a curse-of-dimension that appears inherent to such methods.
Keywords: asymptotics, control functionals, reproducing kernel, scattered data, variance
reduction
1
Introduction
This paper considers methods to estimate the integral
Z
f dΠ
of a test function f against a distribution Π based on evaluation of f at a finite number n
2
of
R inputs.
R Our2 work is motivated by challenging settings in which (i) the variance σ (f ) =
(f − f dΠ) dΠ is large relative to n, and (ii) the distribution Π is only available up to
1
an unknown normalisation constant. Such problems arise in Bayesian statistics when the
cost of sampling from the posterior is prohibitive, requiring that posterior expectations are
approximated based on a small number n of evaluations of the integrand. Indeed, the intrinsic
accuracy of ergodic averages, such as obtained via Markov chain Monte Carlo (MCMC)
methods (Robert and Casella, 2013), can lead to unacceptably high integration error when n
is small. This paper considers a class of estimators inspired by Stein’s method (Stein, 1972),
based on integration-by-parts in this context:
Z
Z Z
d
log π dΠ,
(1)
f dΠ = −
f dx ·
dx
subject to boundary conditions, where π is a density for Π. These estimators ensure an
1
integration error oP (n− 2 ), provided that gradient information on the sampling distribution
can be obtained. This is often the case; indeed, sophisticated software for automatic differentiation of statistical models has been developed (e.g. Carpenter et al., 2015; Maclaurin et
al., 2015).
Main Contribution: The primary contribution of this paper is to establish convergence
rates for a class of estimators based on Stein’s method. These estimators, first described
in Oates et al. (2017), require as input both function evaluations {f (xi )}ni=1 and gradient
evaluations {∇ log π(xi )}ni=1 , where the states {xi }ni=1 themselves can be either independent
or correlated draws from Π. Our central results are asymptotic rates for integration error;
these enable us to compare and quantify the improvement in estimator precision relative to
standard Monte Carlo methods and in doing so we fill a theoretical void.
The estimators that we consider can be viewed as a control variate (or ‘control functional’)
method, and this concept is discussed next.
Control Functionals: The classical control variate method
R proceeds by seeking a collection
k
of non-trivial statistics {ψi }i=1 , such that each satisfies ψi dΠ = 0. Then a surrogate
function
f 0 = f − a1 ψ1 − · · · − ak ψk
R
R
is constructed such that automatically f 0 dΠ = f dΠ and, for suitably chosen {ai }ki=1 , a
variance reduction σ 2 (f 0 ) < σ 2 (f ) might be obtained; for further details see e.g. Rubinstein
and Marcus (1985). For specific problems it is sometimes possible to identify control variates,
for example based on physical considerations (e.g. Assaraf and Caffarel, 2003). For Monte
Carlo integration based on Markov chains, it is sometimes possible to construct control
variates based on statistics relating to the sample path. In this direction, the problem of
constructing control variates for discrete state spaces was essentially solved by Andradóttir
et al. (1993) and for continuous state spaces, recent contributions include Hammer and
Tjelmeland (2008); Dellaportas and Kontoyiannis (2012); Li et al. (2016); Mijatović and
Vogrinc (2015, 2017). Control variates can alternatively be constructed based on gradient
information on the sampling distribution (Assaraf and Caffarel, 1999; Mira et al., 2013; Oates
et al., 2017).
The estimators considered here stem from a recent development that extends control
variates to control functionals. This idea is motivated by the observation that the methods
2
listed above are (in effect) solving a misspecified regression problem, since in general f does
not belong to the linear span of the statistics {ψi }ki=1 . The recent work by Mijatović and
Vogrinc (2015); Oates et al. (2017) alleviates model misspecification by increasing the number
k of statistics alongside the number n of samples so that the limiting space spanned by the
statistics {ψi }∞
i=1 is dense in a class of functions that contains the test function f of interest.
Both methods provide a non-parametric alternative to classical control variates whose error
1
is oP (n− 2 ). Of these two proposed solutions, Mijatović and Vogrinc (2015) is not considered
here since it is unclear how to proceed when Π is known only up to a normalisation constant.
On the other hand the control functional method of Oates et al. (2017) is straight-forward
to implement when gradients {∇ log π(xi )}ni=1 are provided. Understanding the theoretical
properties of this method is the focus of the present research.
Technical Contribution: This paper establishes that the estimators of Oates et al. (2017)
a∧b
1
incur an integration error OP (n− 2 − d + ), where a is related to the smoothness of the density
π, b is related to the smoothness of the test function f , d is the dimension of the domain
of integration and > 0 can be arbitrarily small (a notational convention used to hide
logarithmic factors). This analysis provides important insight into the strong performance
that has been observed for these estimators in certain low-dimensional applications (Oates et
al., 2017; Liu and Lee, 2017). Indeed, recall that the (naı̈ve) computational cost associated
with these methods, i.e. the cost of solving a linear system, is c = O(n3 ). This cost can also
involve a large constant factor when hyper-parameters are to be jointly estimated. Thus,
1
whilst for standard Monte Carlo methods an estimator error of OP (c− 2 ) can be achieved at
computational cost c, for gradient-based control functionals
1 1 a∧b
1 d−a∧b
error for cost c = OP (c 3 )− 2 − d +
= OP c− 6 + 3d + .
This demonstrates that gradient-based control functionals have asymptotically lower error for
the same fixed computational cost c whenever a ∧ b > d, which occurs when both the density
π and the test function f are sufficiently smooth. In the situation where the computational
bottleneck is evaluation of f , not solution of the linear system, then the computational gain
can be even more substantial. At the same time, the critical dependence on d highlights
the curse-of-dimension that appears inherent to such methods. Going forward, these results
provide a benchmark for future high-dimensional development.
Relation to Other Acceleration Methods: Accelerated rates of convergence can be
achieved by other means, including quasi-Monte Carlo (QMC; Niederreiter, 2010). Consider
the ratio estimator:
Pn
Z
1
i=1 f (xi )π(xi )
n
Pn
(2)
f dΠ ≈
1
i=1 π(xi )
n
For appropriate randomised point sets {xi }ni=1 , the ratio estimator converges at a rate limited
by the least smooth of f · π and f , i.e. limited by a∧b
(at least, in the absence of additional
d
conditions on the mixed partial derivatives, which we have not assumed)1 . See Dick et al.
1
In this section the notation a and b is used as a shorthand for the “smoothness” of, respectively, π and
3
(2016) for a recent study of this approach in the context of Bayesian inference for an unknown
parameter in a partial differential equation model.
The method studied herein can be contrasted with QMC methods in at least two respects:
(1) The states {xi }ni=1 can be independent (or correlated) draws from Π, which avoids the
need to specifically construct a point set. This is an important benefit in cases where the
domain of integration is complicated - indeed, our results hold for any domain of integration
for which an interior cone condition can be established. (2) The estimator studied herein
is unbiased, whereas ratio estimators of the form in Eqn. 2 will be biased in general. The
unbiased nature of the estimator, in common with standard Monte Carlo methods, facilitates
convenient diagnostics to estimate the extent of Monte Carlo error and is therefore useful.
Recent work from Delyon and Portier (2016) and Azaı̈s et al. (2016) considered estimators
of the form
Z
n
1 X f (xi )
(3)
f dΛ ≈
n i=1 π̂(xi )
where π̂ is a kernel density estimate for π = dΠ/dΛ based on a collection of (possibly
1
correlated) draws {xi }ni=1 from Π. Again, theoretical results established an error of oP (n− 2 )
. However, this approach applies to
with an explicit rate gated by a term of the form a∧b
d
integrals with respect to a known, normalised reference measure Λ rather than with respect
to Π.
Outline: Below in Sec. 2 we describe the class of estimators that were considered and
present our main theoretical results, including the case of non-independent samples arising
from a Markov chain sample path. Our theoretical analysis combines error bounds from the
scattered data approximation literature with stability results for Markov chains; proofs are
contained in the electronic supplement. Numerical results in Sec. 3 confirm these error rates
are realised. Finally the importance of our findings is discussed in Sec. 4.
2
Methods
First we fix notation before describing the estimation method.
2.1
Set-up and Notation
Consider an open and bounded set X ⊂ Rd , d ∈ N, with boundary ∂X . Let B = B(X ∪ ∂X )
denote the Borel σ-algebra on X ∪ ∂X and equip (X ∪ ∂X , B) with the reference measure Λ
induced from the restriction of Lebesgue measure on Rd . Further, consider a random variable
X on X ∪ ∂X with distribution Π and suppose Π admits a density π = dΠ/ dΛ.
The following notation
Pwill be used: N0 := N ∪ {0}, a ∧ b := min(a, b), a+ := max(a, 0),
1 = [1, . . . , 1]> , kxk22 := di=1 x2i , ∇x := [∂/∂x1 , . . . , ∂/∂xd ]> , 1A (x) = 1 is the indicator of
f . The precise mathematical definition of a and b differs between manuscripts and the results discussed here
should not be directly compared.
4
the event x ∈ A.R Write
L2 (X , Π) for the vector space of measurable functions f : X → R for
R
which σ 2 (f ) := (f − f dΠ)2 dΠ exists and is finite. Write C k (X ) for the set of measurable
functions for which continuous partial derivatives exist on X up to order k ∈ N0 . A function
g : X × X → R is said to be in C2k (X ) if ∂ 2k g/∂xi1 . . . ∂xik ∂x0j1 . . . ∂x0jk is C 0 (X × X ) for all
i1 , . . . , ik , j1 , . . . , jk ∈ {1, . . . , d}. The notation kf k∞ := supx∈X |f (x)| will be used.
2.2
Control Functionals
This section introduces the control functional method for integration, a non-parametric
extension of the classical control variate method. Recall that the trade-off between random
sampling and deterministic approximation in the context of integration is well-understood
(Bakhvalov, 1959). Our starting point is, in a similar vein, to establish a trade-off between
random sampling and stochastic approximation.
We assume throughout that the test function f belongs to L2 (X , Π) and that the boundary ∂X is piecewise smooth. Consider an independent sample from Π, denoted D = {xi }ni=1 .
n
This is partitioned into disjoint subsets D0 = {xi }m
i=1 and D1 = {xi }i=m+1 , where 1 ≤ m < n.
Although m, n are fixed, we will be interested in the asymptotic regime where m = O(nγ )
for some γ ∈ [0, 1]. Consider constructing an approximation fm ∈ L2 (X , Π) to f , based on
D0 . Stochasticity
in fm is induced via the sampling distribution of elements in D0 . The
R
integral fm dΠ is required to be analytically tractable; we will return to this point.
The estimators that we study take the form
Im,n
Z
n
X
1
f (xi ) − fm (xi ) − fm dΠ .
:=
n − m i=m+1
(4)
R
Such sample-splitting estimators are unbiased, i.e. ED1 [Im,n ] = f dΠ, where the expectation here is with respect to the sampling distribution Π of the n − m random variables
that constitute D1 , and is conditional on fixed D0 . The corresponding estimator variance,
again conditional on D0 , is VD1 [Im,n ] = (n − m)−1 σ 2 (f − fm ). This formulation encompasses
control variates as a special case where fm = a1 ψ1 + · · · + ak ψk , k ∈ N, and D0 are used to
select suitable values for the coefficients {ai }ki=1 (see e.g. Rubinstein and Marcus, 1985).
To go beyond control variates and achieve an error of oP (n−1/2 ), we must construct
increasingly accurate approximations fm to f . Indeed, under the scaling m = O(nγ ), if the
expected functional approximation error satisfies ED0 [σ 2 (f − fm )] = O(m−δ ) for some δ ≥ 0,
then
"
2 #
Z
ED0 ED1
Im,n − f dΠ
= O(n−1−γδ ).
(5)
Here we have written ED0 for the expectation with respect to the sampling distribution Π of
the m random variables that constitute D0 . The rate above is optimised by taking γ = 1,
so that an optimal sample-split satisfies m/n → ρ for some ρ ∈ (0, 1] as n → ∞; this will be
assumed in the sequel.
5
When Π is given via an un-normalised density, this framework
can only be exploited if it
R
is possible to construct approximations fm whose integrals fm dΠ are available in closedform. If and when this is possible, the term in parentheses in Eqn. 4 is known as a control
functional. Oates et al. (2017) showed how to build a flexible class of control functionals
based on Stein’s method; the key points are presented next.
2.3
Stein Operator
To begin, we make the following assumptions on the density π:
(A1) π ∈ C a+1 (X ∪ ∂X ) for some a ∈ N0 .
(A2) π > 0 in X .
The gradient function ∇x log π(·) is well-defined and C a (X ∪ ∂X ) by (A1,2). Crucially, gradients can be evaluated even when π is only available un-normalised. Consider the following
Stein operator:
Sπ : C 1 (X ) × · · · × C 1 (X ) → C 0 (X )
φ(·) 7→ Sπ [φ](·) := ∇x · φ(·) + φ(·) · ∇x log π(·)
(6)
This definition can be motivated in several ways, including via Schrödinger Hamiltonians
(Assaraf and Caffarel, 1999) and via the generator method of Barbour applied to an overdamped Langevin diffusion (Gorham and Mackey, 2015). The choice of Stein operator is not
unique and some alternatives are listed in Gorham et al. (2016).
For functional approximation we follow Oates et al. (2017) and study approximations of
the form
fm (·) := β + Sπ [φ](·)
(7)
where β ∈ R is a constant and Sπ [φ](·) acts as a flexible function, parametrised by the choice
of φ ∈ C 1 (X ) × · · · × C 1 (X ). Under regularity
R assumptions introduced below, integrationby-parts (Eqn.R 1) can be applied to obtain Sπ [φ] dΠ = 0 (Lemma 1). Thus, for this class
of functions, fm dΠ permits a trivial closed-form and Sπ [φ] is a control functional (i.e.
integrates to 0).
The choice of β and φ can be cast as an optimisation problem over a Hilbert space and
this will be the focus next.
2.4
Stein Operators on Hilbert Spaces
This section formulates the construction of fm as approximation in a Hilbert space H+ ⊂
L2 (X , Π). This construction first appeared in Oates et al. (2017) and was subsequently
explored in several papers (e.g. Liu et al., 2016; Chwialkowski et al., 2016; Gorham and
Mackey, 2017).
6
First we restrict each component function φi : X → R to belong to a Hilbert space H
with inner product h·, ·iH . Moreover we insist that H is a (non-trivial) reproducing kernel
Hilbert space (RKHS), i.e. there exists a (non-zero) symmetric positive definite function
k : X × X → R such that (i) for all x ∈ X we have k(·, x) ∈ H and (ii) for all x ∈ X and
h ∈ H we have h(x) = hh, k(·, x)iH (see Berlinet and Thomas-Agnan, 2004, for background).
The vector-valued function φ : X → Rd is defined in the Cartesian product
space Hd :=
P
H × · · · × H, itself a Hilbert space with the inner product hφ, φ0 iHd = di=1 hφi , φ0i iH .
To ensure H ⊆ C 1 (X ) we make an assumption on k that will be enforced by construction
through selection of the kernel:
(A3) k ∈ C2b+1 (X ∪ ∂X ) for some b ∈ N0 .
2.4.1
Boundary Conditions
Two further assumptions are made on π. To this end, denote by Q(k) the set of densities
q = dQ/ dΛ on (X ∪ ∂X , B) such that (a) q ∈ C 1 (X ∪ ∂X ), (b) q > 0 in X , and (c) for all
i = 1, . . . , d we have ∇xi log q ∈ L2 (X ∪ ∂X , Q0 ) for all distributions Q0 on (X ∪ ∂X , B). Let
R(k) denote the set of densities q for which q(x)k(x, ·) = 0 for all x ∈ ∂X .
(A2̄) π ∈ Q(k)
(A4) π ∈ R(k)
The assumption (A2̄) was first discussed in Chwialkowski et al. (2016); note in particular
that (A2̄) implies (A2). A constructive approach to ensure (A4) holds is to start with an
arbitrary RKHS H̃ with reproducing kernel k̃ and let B : H̃ → im(B) be a linear operator
such that Bϕ(x) := δ(x)ϕ(x), where δ(·) is a smooth function such that π(·)δ(·) vanishes on
∂X . Then H = im(B) is a RKHS whose kernel k is defined by k(x, x0 ) = δ(x)δ(x0 )k̃(x, x0 ).
This construction will be used in Sec. 3. The following Lemma shows that Sπ [φ] is a control
functional:
R
Lemma 1. Under (A1-4), if φ ∈ Hd then Sπ [φ] dΠ = 0.
Now, consider the set H0 := Sπ [Hd ], whose elements Sπ [φ] result from application of the
Stein operator Sπ to elements φ of the Hilbert space Hd . Oates et al. (2017, Thm. 1) showed
that H0 can be endowed with the gradient-based reproducing kernel
k0 (x, x0 ) := (∇x · ∇x0 ) k(x, x0 ) + (∇x log π(x)) · (∇x0 k(x, x0 ))
(8)
0
0
0
0
+ (∇x0 log π(x )) · (∇x k(x, x )) + (∇x log π(x)) · (∇x0 log π(x )) k(x, x ).
From (A1,2̄,3)
that H0 ⊆ C a∧b (X ∪ ∂X ). Moreover, under (A1,2̄,3,4), the kernel
R it follows
0
k0 satisfies k0 (x, x )Π(dx) = 0 for all x0 ∈ X . Indeed, the function k0 (·, x0 ) belongs to H0
by definition and Lemma 1 shows that all elements of H0 have zero integral.
7
2.4.2
Approximation in H+
Now we can be specific about how β and φ are selected. Write HR for the RKHS of constant
functions, characterised by the kernel kR (x, x0 ) = c, c > 0, for all x, x0 ∈ X . Denote the
norms associated to HR and H0 respectively by k · kHR and k · kH0 . Write
H+ := HR + H0 = {β + ψ : β ∈ HR , ψ ∈ H0 }.
Equip H+ with the norm kf k2H+ := kβk2HR + kψk2H0 . It can be shown that H+ is a RKHS
with kernel k+ (x, x0 ) := kR (x, x0 ) + k0 (x, x0 ) (Berlinet and Thomas-Agnan, 2004, Thm. 5,
p24). From (A1-3) it follows that H+ ⊆ C a∧b (X ).
The choice of β and φ is cast as a least-squares optimisation problem:
fm := arg min khk2H+ s.t. ∀ i = 1, . . . , m, h(xi ) = f (xi ),
h ∈ H+ .
Pm
By the representer theorem (Schölkopf et al., 2001) we have fm (x) =
i=1 ai k+ (x, xi )
>
where the coefficients a = [a1 , . . . , am ] are the solution of the linear system K+ a = f0 where
K+ ∈ Rm×m , [K+ ]i,j = k+ (xi , xj ), f0 ∈ Rm×1 , [f+ ]i = f (xi ). In situations where K+ is
not full-rank, we define fm ≡ 0. Numerical inversion of this system is associated with a
O(m3 ) cost and may in practice require additional numerical regularisation; this is relatively
standard.
2.5
Theoretical Results
Our novel analysis, next, builds on results from the scattered data approximation literature
(Wendland, 2004) and the study of the stability properties of Markov chains (Meyn and
Tweedie, 2009).
2.5.1
The Case of Independent Samples
First we focus on scattered data approximation and state two assumptions that are central
to our analysis:
(A5) π > 0 on X ∪ ∂X
(A6) f ∈ H+ .
Here (A5) extends (A2) in requiring also that π > 0 on ∂X . (A6) ensures that the problem
is well-posed. Define the fill distance
hD0 := sup
min kx − xi k2 .
x∈X i=1,...,m
The proof strategy that we present here decomposes into two parts; (i) first, error bounds are
obtained on the functional approximation error σ 2 (f − fm ) in terms of the fill distance hD0 ,
(ii) second, the fill distance hD0 is shown to vanish under sampling (with high probability).
For (ii) to occur, we require an additional constraint on the geometry of X :
8
(A7) The domain X ∪ ∂X satisfies an interior cone condition, i.e. there exists an angle
θ ∈ (0, π/2) and a radius r > 0 such that for every x ∈ X ∪ ∂X there exists a unit
vector ξ such that the cone
C(x, ξ, θ, r) := {x + λy : y ∈ Rd , kyk2 = 1, y > ξ ≥ cos θ, λ ∈ [0, r]}
is contained in X ∪ ∂X .
The purpose of (A7) is to rule out the possibility of ‘pinch points’ on ∂X (i.e. ≺-shaped
regions), since intuitively sampling-based approaches can fail to ‘get into the corners’ of
the domain. The limiting behaviour of the fill distance under sampling enters through the
following technical result:
Lemma 2. Let g : [0, ∞) → [0, ∞) be continuous, monotone increasing, and satisfy g(0) = 0
and limx↓0 g(x) exp(x−3d ) = ∞. Then under (A5,7) we have
1
ED0 [g(hD0 )] = O(g(m− d + )),
where > 0 can be arbitrarily small.
Our first main result can now be stated:
Theorem 1. Assume (A1,2̄,3-7). Recall that we partition the set D as D0 ∪ D1 where
|D0 | = m and |D1 | = n − m. There exists
R h > 0, independent of m, n, such that the
estimator Im,n is an unbiased estimator of f dΠ with
"
2 #
Z
a∧b
ED0 ED1 1hD0 <h Im,n − f dΠ
= O (n − m)−1 m−2 d +
where > 0 can be arbitrarily small.
a∧b
Thus for m = O(n), this result establishes an overall error of O(n−1−2 d + ), as claimed.
This establishes that these estimates are more efficient than standard Monte Carlo estimators
when a∧b > 0. Or, when the cost of solving a linear system is taken into account, the method
is more efficient on a per-cost basis when a∧b > d. This provides new insight into the first set
of empirical results reported in Oates et al. (2017) where, for assessment purposes, samples
were generated independently from known, smooth densities. There, control functionals were
constructed based on smooth kernels and integration errors were shown to be substantially
reduced.
On the negative side, this result illustrates a curse of dimension that appears to be
intrinsic to the method. We return to this point in Sec. 4.
The results above hold for independent samples, yet the main area of application for
control functionals is estimation based on the MCMC output. In the next section we prove
that the assumption of independence can be relaxed.
9
2.5.2
The Case of Non-Independent Samples
In practice, samples from posterior distributions are often obtained via MCMC methods.
Our analysis must therefore be extended to the non-independent setting: Consider the case
where {xi }ni=1 are generated by a reversible Markov chain targeting Π. We make the following
stochastic stability assumption:
(A8) The Markov chain is uniformly ergodic.
Then our first step is to extend Lemma 2 to the non-independent setting:
Lemma 3. The conclusion of Lemma 2 holds when {xi }ni=1 are generated via MCMC, subject
to (A8).
Non-independence presents us with the possibility that two of the states xi , xj ∈ D0 are
identical (for instance, when a Metropolis-Hastings sample is used and a rejection occurs).
Under our current definition, such an event would cause the kernel matrix K+ to become
singular and the control functional to become trivial fm = 0. It is thus necessary to modify
the construction. Specifically, we assume that D0 has been pre-filtered such that any repeated
states have been removed. Note that this does not ‘introduce bias’, since we are only prefiltering D0 , not D1 . This reduces the effective number m of points in D0 by at most a
constant factor and has no impact on the asymptotics.
With this technical point safely surmounted, we present our second main result:
Theorem 2. The conclusion of Theorem 1 holds when {xi }ni=1 are generated via MCMC,
subject to (A8).
This result again demonstrates that control functionals are more cost-efficient than standard
Monte Carlo when a ∧ b > d and that efficiency is limited by the rougher of the density π
and the test function f . This helps to explain the second set of empirical results obtained
in Oates et al. (2017), where excellent performance was reported on problems that involved
smooth densities, smooth kernels and MCMC sampling methods. On the other hand, we
again observe a curse of dimension that is inherent to control functionals and, indeed, control
variates in general.
2.6
Commentary
Several points of discussion are covered below, on the appropriateness of the assumptions,
the strength of the results and aspects of implementation.
On the Assumptions: Assumptions (A1,2̄,3,7) are not unduly restrictive. The boundary
condition (A4) has previously been discussed in Oates et al. (2017). Below we discuss the
remaining assumptions, (A5,6,8).
Our entire analysis was predicated on (A5), the assumption that π is bounded away from
0 on the compact set X ∪ ∂X . This ensured that π was equivalent to Lebesgue measure on
X ∪ ∂X and enabled this change of measure in the proofs. This is clearly a restrictive set-up
10
as certain distributions of interest do vanish, however the assumption was intrinsic to our
theoretical approach.
Our analysis also relied on (A6), i.e. that f belongs to the function space H+ . It it is thus
natural to examine this assumption in more detail. To this end, we provide the following
lemma. Recall that a RKHS H is c-universal if it is dense as a set in (C 0 (X ∪ ∂X ), k · k∞ ).
Lemma 4. Assume (A2̄,3,4). If H is c-universal then H+ is dense as a set in (L2 (X ∪
∂X , Π), k · k2 ).
The notion of c-universality was introduced by Steinwart (2001), who showed that many
widely-used kernels are c-universal on compact sets. Indeed, Prop. 1 of Micchelli et al.
(2006) proves that a RKHS with kernel k is c-universal if and only if the map Π0 7→ Π0 [k(·, ·)],
from the space of finite signed Borel measures Π0 to the RKHS H, is injective, which is a
weak requirement. It is not, however, clear whether (A4), (A5) can both hold when k is
also c-universal. Further work will therefore be required to better assess the consequences of
f∈
/ H+ . This might proceed in a similar vein to the related work of Narcowich et al. (2005);
Kanagawa et al. (2016).
The last assumption to discuss is (A8); uniform ergodicity of the Markov chain. Since π is
absolutely continuous with respect to Lebesgue measure on X ∪ ∂X , in practice any Markov
chain that targets Π will typically be uniformly ergodic. Indeed, Roberts and Rosenthal
(1998b) constructed an example where a pinch point in the domain caused a Gibbs sampler
targeting a uniform distribution to fail to be geometrically ergodic; their construction violates
our (A7).
On the Results: The intuition for the results in Thms. 1 and 2 can be described as
‘accurate estimation with high probability’, since the condition hD0 < h is satisfied when
the samples D0 cover the state space X , which occurs with unit probability in the m → ∞
limit. There are two equivalent statements that can be made unconditionally on hD0 < h:
(i) Firstly, one can simply re-define fm = 0 whenever h ≥ h0 , i.e. when the states D0
are poorly spaced we revert to the usual Monte Carlo estimator. (ii) Secondly, one could
augment D0 with additional fixed states, such as a grid, {gi }G
i=1 , to ensure that hD0 < h
is automatically satisfied. However, we find both of these equivalent approaches to be less
aesthetically pleasing, since in practice this requires that h be explicitly computed.
The condition hD0 < h suggests that the asymptotics hold in the same regime where QMC
methods could also be successful. However, as explained in Sec. 1, the method of Oates et
al. (2017) carries some advantages
R over the QMC approach that could be important. First, it
provides unbiased estimation of f dΠ, which enables straight-forward empirical assessment.
Second, the fact that it is based on MCMC output renders it more convenient to implement.
On the sharpness of our results, we refer to Sec. 11.7 of Wendland (2004) where an
overview of the strengths and weaknesses of results in the scattered data approximation
literature is provided.
On the Data-Split: It is required to partition samples into sets D0 and D1 , whose sizes
must be specified. Substituting ρ = m/n into the conclusion of Thm. 1 and minimising this
11
expression over ρ ∈ (0, 1] leads to an optimal value
ρ∗ =
ν
1+ν
where
ν=2
a∧b
.
d
(9)
Thus, when a ∧ b d we have ρ∗ ≈ 1 and the optimal method is essentially a numerical
quadrature method (i.e. all samples assigned to D0 ). Conversely, when a ∧ b d we have
ρ∗ ≈ 0 and the optimal method becomes a Monte Carlo method (i.e. all samples assigned
to D1 ).
On the Bandwidth: For the experiments reported next we considered radial kernels of the
form
kx − x0 k2
0
k̃(x, x ) = ϕ
h
where h > 0 is a bandwidth parameter and ϕ is a radial basis function, to be specified. An
appropriate value for the bandwidth h must therefore be selected. An important consideration is that if h is selected based on D0 but not on D1 then the estimator Im,n remains
unbiased. To this end, we propose to select h via maximisation of the log-marginal likelihood
1
m
1
log |K+ | − log 2π
log p(f0 |D0 , h) = − f0> K−1
+ f0 −
2
2
2
which arises from the duality with Gaussian processes and approximation in RKHS (see e.g.
Berlinet and Thomas-Agnan, 2004).
On an Extension: An extension of the estimation method was also considered. Namely,
for each i one can build an approximation f (−i) ∈ H+ to be used as a control functional for
f (xi ), based on D \ {xi }. This results in a leave-one-out (LOO) estimator
Z
n
1X
(−i)
(−i)
In :=
f (xi ) − f (xi ) − f
dΠ
n i=1
(10)
that again remains unbiased. The performance of In can be expected to compare favourably
with that of Im,n , but the computational cost of In is larger at O(n4 ).
On Computation: It is important to emphasise the ease with which these estimators can
be implemented. In the c → ∞ limit, explicit evaluation of Eqn. 4 is particularly straightforward:
> −1
1 K0 f0
1
>
−1
1
(11)
Im,n =
1
f1 − K10 K0 f0 −
n−m
1> K−1
0 1
where f1 ∈ Rn−m×1 , [f1 ]i = f (xm+i ), K0 ∈ Rm×m , [K0 ]i,j = k0 (xi , xj ), K10 ∈ Rn−m×m
and [K10 ]i,j = k0 (xm+i , xj ). An implementation called control_func.m is available on the
Matlab File Exchange to download.
12
3
Numerical Results
First, in Sec. 3.1, we assessed whether the theoretical results are borne out in simulation
experiments. Then, in Sec. 3.2, we applied the method to a topical parameter estimation
problem in uncertainty quantification for a groundwater flow model.
3.1
Simulation
To construct a test-bed for the theoretical results we considered the simple case where Π
is the uniform distribution on X = [0, 1]d . The test functions that we considered took the
form f (x) = 1 + sin(2πωx1 ) where ω was varied to create a problem that was either ‘easy’
(ω = 1) or ‘hard’ (ω = 3). The importance of the first coordinate x1 aimed to reflect the ‘low
effective dimension’ phenomena that is often encountered. From symmetry of the integrand,
the true integral is 1.
For estimation we took the radial basis function ϕ to have variable smoothness and
compact support, as studied in Wendland (1995). Explicit formulae for the ϕ and their
derivatives
are contained in the electronic supplement. To enforce (A4) we took δ(x) =
Qd
i=1 xi (1 − xi ) which vanishes on ∂X . The data-split fraction ρ and the bandwidth h were
each optimised as described in Sec. 2.6. Optimisation for h was performed through 10
iterations of the Matlab function fminbnd constrained to h ∈ [0, 10].
Three estimators were considered; the standard Monte Carlo estimator, the control functional (CF) estimator Im,n in Eqn. 11 and the LOO estimator In . In the case of the LOO
estimator, the bandwidth h was re-optimised in building each of the n control functionals
f (−i) . (A1,2̄,3-5,7) were satisfied in this experiment. Thus, for f ∈ H+ , Thm. 1 entails
b
a mean squared integration error for Im,n of O(n−1−2 d + ), since π(x) = 1 ∈ C a+1 for all
a ∈ N0 . However, the theoretical analysis does not take into account automatic selection of
the bandwidth h; this will be assessed through experiment.
Independent Samples: To study estimator performance, we repeatedly generated collections of n independent uniform random variables {xi }ni=1 and evaluated all three estimators
on this set. The procedure was repeated several times to obtain estimates (along with
standard errors) for the average mean square errors (MSE) that were incurred. Results are
displayed in Fig. 1. In these experiments the MSE appeared to decrease at least as rapidly as
the rates that were predicted. Also, as predicted, the estimator performance quickly deteriorated as the dimension d was increased. Indeed, for d = 3, 4 an improvement over standard
Monte Carlo was no longer observed. The LOO estimator In in general out-performed the
CF estimator Im,n , as expected, but at an increased computational cost. The integration
error was in general larger for the hard integrand.
Dependent Samples: The effect of correlation among the xi was also explored. For this,
we considered a random walk xi = xi−1 + ei on the d-torus with {ei }ni=1 drawn uniformly
on [−, ]d and x0 = 0. This is a Markov chain with invariant distribution Π. The objective
was to assess estimator performance as a function of the step size parameter ; results for
n = 100 are shown in Fig. 2. Compared to Fig. 1, the MSE was larger in general when
13
Dimension; d = 1
MSE; Easy Integrand
100
Dimension; d = 2
Dimension; d = 3
Dimension; d = 4
10-5
10-10
10-15
MSE; Hard Integrand
100
10-5
Monte Carlo
CF (b = 1)
CF (b = 2)
CF (b = 3)
LOOCF (b = 1)
LOOCF (b = 2)
LOOCF (b = 3)
10-10
10-15
101
102
Num. Samples; n
101
102
101
Num. Samples; n
102
Num. Samples; n
101
102
Num. Samples; n
Figure 1: Simulation results; the case of independent samples. An ‘easy’ and a ‘hard’
integrand were considered. The mean square error (MSE) was estimated for the standard
Monte Carlo estimator, the control functional (CF) estimator Im,n and the leave-one-out
(LOOCF) estimator In , and plotted against the number n of samples used. The CF and
LOOCF estimators were based on kernels of smoothness b ∈ {1, 2, 3}. Standard errors are
also displayed.
< 0.5. This reflects reduction in effective sample size of the set D0 used to build the control
functional.
3.2
Application to Partial Differential Equations
Our theoretical results are illustrated with a novel application to an inverse problem arising
in a partial differential equation (PDE) model. Specifically, we considered the following
14
Dimension; d = 1
MSE; Easy Integrand
100
Dimension; d = 2
Dimension; d = 3
Dimension; d = 4
10-5
10-10
10-15
MSE; Hard Integrand
100
10-5
Monte Carlo
CF (b = 1)
CF (b = 2)
CF (b = 3)
LOOCF (b = 1)
LOOCF (b = 2)
LOOCF (b = 3)
10-10
10-15
0
0.5
Step Size;
1
0
0.5
1
Step Size;
0
0.5
1
Step Size;
0.2 0.4 0.6 0.8
Step Size;
Figure 2: Simulation results; the case of dependent samples. An ‘easy’ and a ‘hard’ integrand
were considered. The mean square error (MSE) was estimated for the standard Monte
Carlo estimator, the control functional (CF) estimator Im,n and the leave-one-out (LOOCF)
estimator In , where samples from a random walk of length n = 100 was used. The MSE was
plotted against the step size of the random walk. The CF and LOOCF estimators were
based on kernels of smoothness b ∈ {1, 2, 3}. Standard errors are also displayed.
elliptic diffusion problem with mixed Dirichlet and Neumann boundary conditions:
∇x · [κ(x; θ)∇x w(x)] = 0
if x1 , x2 ∈ (0, 1)
x1
if x2 = 0
w(x) =
1 − x1 if x2 = 1
∇x1 w(x) = 0
if x1 ∈ {0, 1}.
This PDE serves as a simple model of steady-state flow in aquifers and other subsurface
systems; κ can represent the permeability of a porous medium while w represents the hydraulic head. The aim is to make inferences on the field κ in a setting where the underlying
solution w is observed with noise on a regular grid of M 2 points xi,j , i, j = 1, . . . , M . The
observation model p(y|θ) takes the form y = {yi,j } where yi,j = w(xi,j ) + i,j and i,j are
independent normal random variables with standard deviation σ = 0.1.
15
10-1
Dimension; d = 1
Dimension; d = 3
Dimension; d = 2
Dimension; d = 4
RMSE
10-2
10-3
10-4
Monte Carlo
CF (b = 2)
101
102
m
103
101
102
103 101
m
102
m
103 101
102
103
m
Figure 3: Experimental results; an experiment to approximate the posterior mean of the
parameters θ ∈ [−10, 10]d that govern a permeability field. The figure shows root mean
square error (RMSE) for (i) the standard Monte Carlo estimator based on 2m posterior
samples, and (ii) the control functional (CF) estimator, where m samples are used to train
the control functional and the remaining m samples are used to estimate the expectation.
[Results are shown for the first parameter θ1 ; results for other parameters were similar. The
Matérn kernel of order 7/2 was employed; b = 2 in our notation.]
Following Stuart and Teckentrup
(2017), the field κ was endowed with a prior distribution
Pd
of the form log κ(x; θ) = i=1 θi κi (x), where the κi are Fourier basis functions and θi are
their associated coefficients. For the inference we imposed a uniform prior p(θ) ∝ 1 over
the domain [−10, 10]d . Our aim was to obtain accurate estimates for the posterior mean of
the parameter θ. The posterior density p(θ|y) ∝ p(θ)p(y|θ) is available up to an unknown
normalising constant p(y). Each evaluation of the likelihood necessitates the solution of the
PDE; control functionals offer the possibility to reduce the number of likelihood evaluations,
and hence the computational cost, required to achieve a given estimator precision.
As an aside, we note that the standard approach to inference employs a numerical integrator for the forward-solve, typically based on finite element methods. This would provide
us with gradient information on the posterior, but would also introduce some bias due to discretisation error. To ensure that we obtain exact gradient information, we instead exploited
a probabilistic meshless method due to Cockayne et al. (2016) as our numerical integrator. Automatic differentiation was performed using the Autograd package (Maclaurin et al.,
2015).
The key assumptions of our theory were verified. Smoothness of the prior, together with
ellipticity, imply (A1) holds for all a ∈ N. (A2̄,5) hold since the prior and likelihood are
well-behaved. (A7) holds since the domain of integration was a hyper-cuboid. Samples
from the posterior p(θ|y) were obtained using a Metropolis-adjusted Langevin sampler with
fixed proposal covariance; this ensured that (A8) was satisfied. Remaining assumptions were
satisfied by construction of the kernel k: Following the approach outlined in Section 2.4, we
16
took k̃(θ, θ 0 ) to be the standard Matérn kernel of order 72 , so that b = 2, and then formed
k(θ, θ 0 ) as the product of k̃(θ, θ 0 ) and δ(θ)δ(θ 0 ), where the boundary function δ satisfies
δ(θ) = 1 on θ ∈ [−9, 9]d , δ(θ) = 0 when θi ∈ {−10, 10} for some i, and δ was infinitely
differentiable on [−10, 10]d . With this construction, (A3) holds. (A4) holds since k has a
root at θi ∈ {−10, 10} for each i. The constant c = 1 was fixed. However the conclusion of
Lemma 4 cannot be directly applied here since H is not c-universal (k vanishes at θi = ±10).
Observations were generated from the model with data-generating parameter θ = 1 and
collected over a coarse grid of M 2 = 36 locations. Samples of size n were obtained from the
posterior and divided equally between the training set D0 and test set D1 . The performance
of gradient-based control functionals was benchmarked against that of standard Monte Carlo
with all n samples used. We note that, in all experiments, all values of θ encountered were
contained in [−9, 9]d . Thus it does not matter that we did not specify δ explicitly above,
emphasising the weakness of assumption (A4) in practical application.
Results are shown in Figure 3. For dimensions d = 1 and 2, the estimator that uses
control functionals achieved a dramatic reduction in asymptotic variance compared to the
Monte Carlo benchmark. On the other hand, for d = 3, 4, the curse of dimension is clearly
evident for the control functional method.
4
Conclusion
This paper has established novel asymptotic analysis for a class of estimators based on
Stein’s method. Our analysis makes explicit the contribution of the smoothness a of the
distribution Π, the smoothness b of the test function f and the dimension d of the domain of
integration. As such, these results provide a rigorous theoretical explanation for the excellent
performance in low-dimensions observed in previous work.
Several extensions of this work are suggested: (i) Our results focused on compact domains, since this is the usual setting for results in the scattered data approximation literature.
However, the estimation method does not itself require that the domain of integration be
compact. Extending this analysis to the unbounded-domain setting appears challenging at
present and remains a goal for future research. (ii) Alternative literatures to the scattered
data literature could form the basis of an analysis of control functionals, such as e.g. recent
work by Migliorati et al. (2015). These efforts have the advantage of providing L2 error
bounds, rather than L∞ error bounds and might facilitate the extension to unbounded domains. (iii) Generally, our theoretical results clarify the need to develop estimation strategies
that do not suffer from the curse of dimension. While this curse is intrinsic to functional
approximation in general, due to the need to explore the state space, the observation that
many test functions of interest are of low ‘effective dimension’ suggests that more regularity
on the function space could reasonably be assumed. (iv) Recent work in Liu and Lee (2017)
imposed an additional constraint on the coefficients ai in Sec. 2.4.2. It would be interesting
to extend our analysis to this context.
17
Acknowledgements: The authors wish to thank Aretha Teckentrup, Motonobu Kanagawa, Lester Mackey and anonymous referees for their useful feedback. CJO was supported by the ARC Centre of Excellence for Mathematical and Statistical Frontiers. CJO
and MG were supported by the Lloyds-Turing Programme on Data-Centric Engineering.
FXB was supported by EPSRC [EP/L016710/1]. MG was supported by the EPSRC grants
[EP/J016934/3, EP/K034154/1, EP/P020720/1], an EPSRC Established Career Fellowship,
the EU grant [EU/259348] and a Royal Society Wolfson Research Merit Award.
References
Andradóttir, S., Heyman, D. P. and Ott, T. J. (1993) Variance reduction through smoothing
and control variates for Markov Chain simulations. ACM T. M. Comput. S., 3, 167-189.
Assaraf, R. and Caffarel, M. (1999) Zero-variance principle for Monte Carlo algorithms.
Phys. Rev. Lett., 83(23), 4682-4685.
Assaraf, R. and Caffarel, M. (2003) Zero-variance zero-bias principle for observables in quantum Monte Carlo: Application to forces. J. Chem. Phys., 119, 10536.
Azaı̈s, R., Delyon, B. and Portier, F. (2016) Integral estimation based on Markovian design.
arXiv:1609.01165.
Bakhvalov, N.S. (1959) On the approximate calculation of multiple integrals (in Russian).
Vestnik MGU, Ser. Math. Mech. Astron. Phys. Chem., 4, 3-18.
Berlinet, A. and Thomas-Agnan, C. (2004) Reproducing Kernel Hilbert Spaces in Probability
and Statistics. Kluwer Academic, Boston.
Carpenter, B., Hoffman, M. D., Brubaker, M., Lee, D., Li, P. and Betancourt, M. (2015) The
Stan math library: Reverse-mode automatic differentiation in C++. arXiv:1509.07164.
Chwialkowski, K., Strathmann, H. and Gretton, A. (2016) A kernel test of goodness of fit.
In: Proceedings of the 33rd International Conference on Machine Learning.
Cockayne, J., Oates, C. J., Sullivan, T. and Girolami, M. (2016) Probabilistic numerical
methods for PDE-constrained Bayesian inverse problems. In: Proceedings of the 36th International Workshop on Bayesian Inference and Maximum Entropy Methods in Science
and Engineering.
Dellaportas, P. and Kontoyiannis, I. (2012) Control variates for estimation based on reversible Markov chain Monte Carlo samplers. J. R. Statist. Soc. B. Stat. Methodol., 74,
133-161.
Delyon, B. and Portier, F. (2016). Integral approximation by kernel smoothing. Bernoulli,
22(4), 2177-2208.
18
Dick, J., Gantner, R.N., Gia, Q.T.L. and Schwab, C. (2016) Higher order quasi-Monte Carlo
integration for Bayesian estimation. arXiv:1602.07363.
Gorham, J. and Mackey, L. (2015) Measuring sample quality with Stein’s method. In: Proceedings of the 28th Annual Conference on Neural Information Processing Systems.
Gorham, J., Duncan, A.B., Vollmer, S.J. and Mackey, L. (2016) Measuring sample quality
with diffusions. arXiv:1611.06972.
Gorham, J. and Mackey, L. (2017) Measuring sample quality with kernels. arXiv:1703.01717.
Hammer, H. and Tjelmeland, H. (2008) Control variates for the Metropolis-Hastings algorithm. Scand. J. Stat., 35, 400-414.
Kanagawa, M., Sriperumbudur, B.K. and Fukumizu, K. (2016) Convergence guarantees for
kernel-based quadrature rules in misspecified settings. In: Proceedings of the 29th Annual
Conference on Neural Information Processing Systems.
Li, W., Chen, R. and Tan, Z. (2016) Efficient sequential Monte Carlo with multiple proposals
and control variates. J. Am. Stat. Assoc., 111(513), 298-313.
Liu, Q., Lee, J.D. and Jordan, M.I. (2016) A kernelized Stein discrepancy for goodness-offit tests and model evaluation. In: Proceedings of the 33rd International Conference on
Machine Learning.
Liu, Q. and Lee, J.D. (2017) Black-box importance sampling. In: Proceedings of the 21st
International Conference on Artificial Intelligence and Statistics.
Maclaurin, D., Duvenaud, D., Johnson, M., Adams, R.P. (2015) Autograd: Reverse-mode
differentiation of native Python. http://github.com/HIPS/autograd.
Meyn, S.P. and Tweedie, R.L. (2009) Markov Chains and Stochastic Stability, 2ed. Cambridge University Press.
Micchelli, C.A., Xu, Y. and Zhang, H. (2006) Universal kernels. J. Mach. Learn. Res., 7,
2651-2667.
Migliorati, G., Nobile, F. and Tempone, R. (2015) Convergence estimates in probability
and in expectation for discrete least squares with noisy evaluations at random points. J.
Multivariate Anal., 142, 167-182.
Mira, A., Solgi, R. and Imparato, D. (2013) Zero variance Markov chain Monte Carlo for
Bayesian estimators. Stat. Comput., 23, 653-662.
Mijatović, A. and Vogrinc, J. (2015) On the Poisson equation for Metropolis-Hastings chains.
arXiv:1511.07464.
19
Mijatović, A. and Vogrinc, J. (2017) Asymptotic variance for random walk Metropolis chains
in high dimensions: Logarithmic growth via the Poisson equation. arXiv:1707.08510.
Narcowich, F., Ward, J. and Wendland, H. (2005) Sobolev bounds on functions with scattered zeros, with applications to radial basis function surface fitting. Math. Comput.,
74(250), 743-763.
Niederreiter, H. (2010) Quasi-Monte Carlo Methods. John Wiley & Sons, Ltd.
Oates, C. J., Chopin, N. and Girolami, M. (2017) Control functionals for Monte Carlo
integration. J. R. Statist. Soc. B. Stat. Methodol., 79(3), 695-718.
Robert, C. and Casella, G. (2013) Monte Carlo Statistical Methods. Springer.
Roberts, G. O. and Rosenthal, J. S. (1998b) On convergence rates of Gibbs samplers for
uniform distributions. Ann. Appl. Prob., 8(4), 1291-1302.
Rubinstein, R. Y. and Marcus, R. (1985) Efficiency of multivariate control variates in Monte
Carlo simulation. Oper. Res., 33, 661-677.
Schölkopf, B., Herbrich, R. and Smola, A. J. (2001) A generalized representer theorem. Lect.
Notes Comput. Sc., 2111, 416-426.
Smola, A., Gretton, A., Song, L. and Schölkopf, B. (2007) A Hilbert space embedding
for distributions. In: Proceedings of the 18th International Conference on Algorithmic
Learning Theory, 13-31. Springer-Verlag, Berlin.
Stein, C. (1972). A bound for the error in the normal approximation to the distribution of
a sum of dependent random variables. In: Proceedings of the 6th Berkeley Symposium on
Mathematical Statistics and Probability, 2, 583-602.
Steinwart, I. (2001) On the influence of the kernel on the consistency of support vector
machines. J. Mach. Learn. Res., 2, 67-93.
Stuart, A. M. and Teckentrup, A. L. (2017) Posterior consistency for Gaussian process
approximations of Bayesian posterior distributions. Math. Comput., to appear.
Wendland, H. (1995) Piecewise polynomial, positive definite and compactly supported radial
functions of minimal degree. Adv. Comput. Math., 4(1), 389-396.
Wendland, H. (2004) Scattered Data Approximation. Cambridge University Press.
20
5
Supplement
In this electronic supplement we establish correctness of the theoretical results in the main
text.
5.1
Additional Notation
R
R
Let kf k2 := ( f 2 dΠ)1/2 . Write vol(A) := 1A dΛ and dx := dΛ(x). LetHL∞ (X ) denote the
set of measurable functions for which supx∈X |f (x)| < ∞. The notation ∂X will be used to
denote a surface integral over ∂X , n(x) to denote the unit normal vector to ∂X and S(dx)
to denote a surface element.
5.2
Proofs
Proof of Lemma 1. From the Moore-Aronszajn theorem (Aronszajn, 1950), Hd can be expressed as
(
)
∞
∞
X
X
Hd = φ : X → Rd s.t. φi (x) =
ai,j k(x, xi,j ) where
a2i,j k(xi,j , xi,j ) < ∞ .
j=1
j=1
Note that from (A3) the kernel k is continuous and thus bounded on the compact set X ∪∂X .
It follows that the series representation given in the Moore-Aronszajn theorem is uniformly
convergent. Then, for φ ∈ Hd represented in Moore-Aronszajn form,
Z
Sπ [φ](x)Π(dx) =
Z X
d
(∇xi + ∇xi log π(x))
i=1
=
d X
∞
X
∞
X
ai,j k(x, xi,j ) Π(dx)
j=1
Z
(∇xi + ∇xi log π(x))k(x, xi,j ) Π(dx)
ai,j
i=1 j=1
=
d X
∞
X
Z
=
d X
∞
X
i=1 j=1
∇xi {π(x)k(x, xi,j )} dx
ai,j
X ∪∂X
i=1 j=1
I
ai,j
∂X
π(x)k(x, xi,j ) ni (x)S(dx)
{z
}
|
=
0,
(∗)
where the order of summation and calculus operations can be interchanged due to uniform
convergence of the series representation on the compact set X ∪ ∂X . The term (∗) equals
zero for all xi,j ∈ X ∪ ∂X by (A4).
Proof of Lemma 2. In rigorously establishing this result there are six main steps. Initially
we fix x ∈ X ∪ ∂X and aim to show that with high probability there exists a state xj close
to x. Then we consider implications for the distribution of the fill distance.
21
𝑋
𝑋
𝑥
𝑟
𝜃
𝑥
𝑔𝑖
𝐵(𝑦, 𝜖)
(a) Step #1
𝑦
𝐶(𝑥, 𝜉, 𝜃, 𝑟)
(b) Step #2
Figure 4: Schematic of the steps involved in proving Lemma 2.
Step #1: Construct a reference grid. Since X is bounded, without loss of generality suppose
d
X ∪ ∂X ⊆ [0, 1]d . For M ∈ N, define a uniform grid of reference points {gi }G
i=1 ⊂ [0, 1]
M −2
consisting of all G = M d states of the form g = (g1 , . . . , gd ) where gi ∈ {0, M1−1 , . . . , M
, 1}
−1
(Fig. 4a). We will require that this grid has a sufficiently fine resolution; specifically we
suppose that
√
d(1 + sin θ)
+ 1,
M≥
2r sin θ
where r, θ are defined by the interior cone condition that X ∪ ∂X is assumed to satisfy.
Step #2: Find a grid point gi near to x. Define
√
d
1
+1 .
R=
2
sin θ
We claim that there exists a grid point gi such that kx − gi k2 < MR−1 . Indeed, from the
interior cone condition there exists a cone C(x, ξ, θ, r) ⊆ X (Fig. 4a). Define
√
d
h=
.
2(M − 1) sin θ
The fine grid resolution implies h ≤ r/(1 + sin θ). From Wendland (2004), Lemma 3.7, it
follows that the cone C(x, ξ, θ, r) contains the ball B(y, ) with centre y = x + hξ and radius
d
= h sin θ (Fig. 4b). Now,
the fill distance for the grid {gi }G
i=1 relative to the space [0, 1]
√
can be shown to equal d/(2(M − 1)), which is exactly equal to the radius . It follows that
22
𝑋
𝑋
𝑥
𝐵(𝑔𝑖 ,
1
)
𝑀−1
𝑥
𝑔𝑖
𝑔𝑖
𝐶𝑖
𝐵(𝑔𝑖 + Δ𝜉𝑖 , Δ)
(a) Step #3
(b) Step #4
Figure 5: Schematic of the steps involved in proving Lemma 2.
B(y, ) must contain a grid point gi for some i ∈ {1, . . . , G}. From the triangle inequality;
kx − gi k2 ≤ kx − yk2 + ky − gi k2
| {z } | {z }
≤
=h
√
√
d
d
R
=
+
=
2(M − 1) sin θ 2(M − 1)
M −1
This establishes the claim.
Step #3: A set of points that ‘cover the grid’ must contain at least one point that is near
to x. Consider the event
1
E = ∀i∃j : kgi − xj k2 ≤
.
M −1
Conditional on E, there exists xj such that kgi − xj k2 ≤
on E, we have
1
.
M −1
It follows that, conditional
kx − xj k2 ≤ kx − gi k2 + kgi − xj k2
R
1
R+1
≤
+
=
M −1 M −1
M −1
where we have used the result of Step #1. Thus the event E, or ‘covering the grid’, implies
R+1
the event [hD0 ≤ M
] (Fig. 5a).
−1
Step #4: The grid is covered with high probability. Next, we upper-bound the probability
PD0 [E c ] of the event E c . For this, note that for all i ∈ {1, . . . , G} there exists a unit vector
23
ξi such that the cone Ci = C(gi , ξi , θ, r) is contained in X (Fig. 5b). It follows that
1
1
∩X
≥ vol B gi ,
∩ Ci
vol B gi ,
M −1
M −1
1
= vol C gi , ξi , θ,
M −1
where the final inequality follows since M1−1 ≤ r and the intersection of a cone with a ball is
another cone. Now, from Wendland (2004), Lemma 3.7, the ball with centre gi + ∆ξi and
radius
sin θ
∆=
(M − 1)(1 + sin θ)
is contained in the cone C(gi , ξi , θ, M1−1 ). Continuing,
1
vol C gi , ξi , θ,
≥ vol(B(gi + ∆ξi , ∆))
M −1
sin θ d d/2
) π
( 1+sin
θ
.
= V :=
d
Γ( 2 + 1)(M − 1)d
Now, we have
1
M −1
G
X
1
.
≤
PD0 ∀j, kgi − xj k2 >
M −1
i=1 |
{z
}
PD0 [E ] = PD0 ∃i : ∀j, kgi − xj k2 >
c
(∗)
The probability (∗) can be bounded above using independence of the samples. Indeed, the
probability that a random draw from Π lands in B(gi , M1−1 ) can be lower-bounded by V
times η = inf x∈X ∪∂X π(x) > 0 since, under (A1,5), π can be decomposed as a mixture
density with two components, the first uniform over X ∪ ∂X with mixture weight η and the
second a component with density π̃(x) ∝ π(x) − η and mixture weight 1 − η. Thus we have
1
PD0 ∀j, kgi − xj k2 >
≤ (1 − V η)m .
(12)
M −1
Note that this is the only place in the proof that independence of the {xi } is required.
Step #5: The fill distance is small with high probability. The result of Step #4 can be
used to derive a lower bound on the distribution function of the fill distance:
R+1
PD0 hD0 >
≤ PD0 [E c ] ≤ G(1 − V η)m
M −1
!m
sin θ d d/2
( 1+sin
) π η
θ
= G 1− d
Γ( 2 + 1)(M − 1)d
24
R+1
Letting ζ = M
implies that M = 1 + R+1
and G = (1 + R+1
)d . In this parametrisation,
−1
ζ
ζ
when M > R + 2 we have 0 < ζ < 1 and hence
#m
d "
sin θ d d/2
)
π
η
(
R+1
PD0 [hD0 > ζ] ≤
1+
1 − d1+sin θ
ζd
d
ζ
Γ( 2 + 1)(R + 1)
d
R+2
≤
(1 − Cd ζ d )m ,
(13)
ζ
where we have written
Cd =
sin θ d d/2
( 1+sin
) π η
θ
Γ( d2 + 1)(R + 1)d
.
R+1
While Eqn. 13 holds only for ζ of the form M
, it can be made to hold for all 0 < ζ < 1 by
−1
−d
replacing Cd with C̃d = 2 Cd . This is because for any 0 < ζ < 1 there exists M > 1 such
R+1
that ζ̃ = M
satisfies ζ2 ≤ ζ̃ < ζ, along with the fact that PD0 [hD0 > ζ] ≤ PD0 [hD0 > ζ̃].
−1
Step #6: Putting it all together. From X ⊆ [0, 1]d we have that hD0 ∈ [0, 1] and hence,
since g is continuous and [0, 1] is compact, suph∈[0,1] g(h) < ∞. Without loss of generality,
and with probability one, we have g(hD0 ) ≤ 1. (This is without loss of generality since we can
simply redefine g as g/g(1).) From the reverse Markov inequality, for all ζ < ED0 [g(hD0 )],
PD0 [g(hD0 ) > ζ] ≥
ED0 [g(hD0 )] − ζ
1−ζ
and upon rearranging
ED0 [g(hD0 )] ≤ ζ + (1 − ζ)PD0 [g(hD0 ) > ζ].
(14)
Since g is continuous and monotone, g −1 exists and we have PD0 [g(hD0 ) ≤ ζ] = PD0 [hD0 ≤
g −1 (ζ)]. This allows us to combine Eqns. 13 and 14, obtaining
d
R+2
(1 − C̃d (g −1 (ζ))d )m
ED0 [g(hD0 )]
≤ ζ + (1 − ζ)
g −1 (ζ)
d
R+2
≤ ζ+
(1 − C̃d (g −1 (ζ))d )m .
−1
g (ζ)
Now, letting ζ = g(m−δ ) for some fixed δ, subject to
that
1
3d
≤ δ < d1 , and varying m, we have
ED0 [g(hD0 )] ≤ g(m−δ ) + (R + 2)d mdδ (1 − C̃d m−dδ )m
|
{z
}
(∗∗)
where (∗∗) = O(mdδ exp(−C̃d m1−dδ )). Indeed, since log(1 − x) ≤ −x for all |x| < 1,
log[mdδ (1 − C̃d m−dδ )m ] = dδ log(m) + m log(1 − C̃d m−dδ )
≤ dδ log(m) − C̃d m1−dδ = log[mdδ exp(−C̃d m1−dδ )].
25
(15)
Writing x = m−δ ,
mdδ exp(−C̃d m1−dδ )
x−d exp(−C̃d xd−1/δ )
=
g(m−δ )
g(x)
−d
x exp(−C̃d x−2d )
1
≤
(take δ = ).
g(x)
3d
Under the hypothesis on the limiting behavior of g(x) as x ↓ 0, we have that
x−d exp(−C̃d x−2d )
mdδ exp(−C̃d m1−dδ )
=
lim
m→∞
x↓0
g(m−δ )
g(x)
−d
x exp(−x−3d )
≤ lim
= 0
x↓0
g(x)
lim
Thus the right hand side of Eqn. 15 is asymptotically minimised by taking δ as large as
possible, subject to (∗∗) converging exponentially fast, i.e. δ = d1 − for > 0 arbitrarily
small. We therefore conclude that ED0 [g(hD0 )] = O(g(m−1/d+ )), as required.
Proof of Theorem 1. Unbiasedness follows directly from the structure of splitting estimators
(Eqn. 4). From Eqn. 5 it suffices to consider the rate δ at which σ 2 (f − fm ) vanishes. First,
note that
2
Z
Z
2
0
0
0
σ (f − fm ) =
f (x) − fm (x) − [f (x ) − fm (x )]Π(dx ) Π(dx)
2
Z
Z
0
0
0
2
[f (x ) − fm (x )]Π(dx )
=
[f (x) − fm (x)] Π(dx) −
Z
≤
[f (x) − fm (x)]2 Π(dx).
Second, observe that from (A1-3) the kernel k+ ∈ C2a∧b (X ∪ ∂X ). Thus from Theorem 11.13
of Wendland (2004) there exists h > 0, C > 0 such that, whenever hD0 < h, we have
|f (x) − fm (x)| ≤ Cha∧b
D0 kf kH+
for all x ∈ X . An unconditional bound, after squaring and integrating according to Π, is
Z
2(a∧b)
1hD0 <h [f (x) − fm (x)]2 Π(dx) ≤ C 2 hD0 kf k2H+ .
Third, combining the first two parts and taking an expectation over the m samples in D0
gives
2(a∧b)
ED0 [1hD0 <h σ 2 (f − fm )] ≤ C 2 kf k2H+ ED0 [hD0
2(a∧b)
Finally, from Lemma 2 with g(x) = x2(a∧b) we have that ED0 [hD0
arbitrarily small. The result follows from Eqn. 5.
26
].
] = O(m−2
a∧b
+
d
) for > 0
Proof of Lemma 3. In the proof of Lemma 2, independence of the samples {xi }ni=1 was only
used to establish Eqn. 12. Below we derive an almost equivalent inequality that is valid for
non-independent samples arising from the Markov chain sample path.
From Roberts and Rosenthal (1998a, Prop. 1) a uniformly ergodic, reversible Markov
chain is strongly uniformly ergodic; i.e. there exists N ∈ N and 0 < υ < 1 such that for all
x ∈ X the N -step transition density P N (x, ·) satisfies the minorisation condition P N (x, ·) ≥
υΠ[·]. Manipulating this minorisation condition gives P N (x, Ac ) = 1−P N (x, A) ≤ 1−υΠ(A)
for A ∈ A.
Fix A = B(gi , M1−1 ). Denote the initial distribution of the Markov chain by Π1 and, for
notational simplicity only, assume Π1 has a density π1 = dΠ1 / dΛ. Define the hitting time
τA := min{n : xn ∈ A given x1 ∼ Π1 }. Then for n ≥ N we have the following bound:
Z
Z
···
π1 (x1 )P (x1 , x2 ) . . . P (xn−1 , xn ) dxn . . . dx1
P[τA > n] =
x1 ∈Ac
xn ∈Ac
Z
Z
=
···
π1 (x1 )P (x1 , x2 ) . . . P (xn−N −1 , xn−N )
x1 ∈Ac
···
×
xn−N +1 ∈Ac
|
xn−N ∈Ac
Z
Z
P (xn−N , xn−N +1 ) . . . P (xn−1 , xn )
dxn−N . . . dx1
dxn−N +1 dxn
xn ∈Ac
{z
}
≤P N (xn−N ,Ac )
≤ [1 − υΠ(A)]
Z
Z
×
···
x1 ∈Ac
π1 (x1 )P (x1 , x2 ) . . . P (xn−N −1 , xn−N )
dxn−N . . . dx1
xn−N ∈Ac
. . . ≤ [1 − υΠ(A)]bn/N c
Z
Z
π1 (x1 )P (x1 , x2 ) . . . P (xn−bn/N cN −1 , xn−bn/N cN )
×
···
dxn−bn/N cN . . . dx1
x1 ∈Ac
xn−bn/N cN ∈Ac
≤ [1 − υΠ(A)]bn/N c
Employing this bound, we have
PD0 ∀j, kgi − xj k2 >
1
M −1
As before we have Π(A) ≥ V η and hence
PD0 ∀j, kgi − xj k2 >
= PD0 [τA > m] ≤ [1 − υΠ(A)]bm/N c .
1
M −1
≤ (1 − υV η)bm/N c .
Eqn. 16 is essentially identical to Eqn. 12 up to the inclusion of a factor 0 < υ < 1 and a
factor 1/N . Arguing as in Lemma 2 with C̃d replaced by υN −1 × C̃d completes the proof.
Proof of Theorem 2. We appeal to Theorem 1 of Roberts and Rosenthal (2008). This states
that reversible, geometrically ergodic Markov chains {xi }ni=1 are variance bounding, meaning
27
that there exists K < ∞ such that
" n
#
X
1
lim V
h(xi ) ≤ Kσ 2 (h)
n→∞ n
i=1
(16)
holds for all h ∈ L2 (X , Π). Suppose that the chain starts at stationarity, so that Eqn. 16 is
equivalent to the statement
!2
Z
n
X
1
h(xi ) − h dΠ ≤ Kσ 2 (h).
(17)
lim nE
n→∞
n i=1
(An arbitrary initial condition x1 ∼ Π1 can be handled using standard renewal theory; we do
not present the details here.) Applying Eqn. 17 to the control functional estimator produces
"
2 #
Z
lim (n − m)ED1
Im,n − f dΠ
≤ Kσ 2 (f − fm ).
(18)
n→∞
Arguing as in the proof of Theorem 1, we have
2(a∧b)
ED0 [1hD0 <h σ 2 (f − fm )] ≤ C 2 kf k2H+ ED0 [hD0
].
(19)
2(a∧b)
Finally, it remains only to show that the scaling relation ED0 [hD0 ] = O(n−2(a∧b)/d+ ), where
> 0 can be arbitrarily small, holds in the non-independent setting. This was precisely the
content of Lemma 3.
Denote k : X × X → Rd for the vector k(x, x0 ) = 1k(x, x0 ). The shorthand notation
Sq [k(x, x0 )] will be used to refer to Sq [k(·, x0 )](x); i.e. the action of the Stein operator on
the first argument of a bivariate function.
Lemma 5. Assume (A3). For each q ∈ Q(k) ∩ R(k) it holds that
Z
Sq [k(x, x0 )]Q(dx) = 0,
(20)
where the left hand side exists for all x0 ∈ X .
Proof. This is a generalisation of Lemma 1. From (A3) and q ∈ Q(k), the function Sq [k(·, x0 )]
exists on X for each x0 ∈ X . In particular Sq [k(·, x0 )] ∈ L2 (X ∪ ∂X , Q), so that the left hand
side of Eqn. 20 exists. Then
Z
Z
0
Sq [k(x, x )]Q(dx) =
[∇x · k(x, x0 ) + k(x, x0 ) · ∇x log q(x)] Q(dx)
Z
=
[q(x)∇x · k(x, x0 ) + k(x, x0 ) · ∇x q(x)] dx
Z
=
∇x · {q(x)k(x, x0 )} dx
I
(∗)
=
q(x)k(x, x0 ) ·n(x) S(dx) = 0,
{z
}
∂X |
(∗∗)
where (∗) is integration by parts and (∗∗) equals zero for all x0 ∈ ∂X since q ∈ R(k).
28
Proof of Lemma 4. Recall that, for compact X ∪∂X , the notion of c-universality is equivalent
to cc-universality, where the (weaker) topology of compact convergence is used in place of
the k · k∞ norm topology (Carmeli et al., 2010, Defn. 4.1).
For densities q, q 0 on (X ∪ ∂X , B) with q = dQ/ dΛ, q 0 = dQ0 / dΛ, define
Z
Z
0
0
T (q, q , k) :=
Sq [k(x, ·)]Q(dx) − Sq0 [k(x, ·)]Q0 (dx) ,
H
whenever the right hand side exists. Observe, as in Lemma 5, that T (q, q 0 , k) is well-defined
for all q, q 0 ∈ Q(k).
This proof then proceeds in two steps:
Step #1: First, we follow the proof of Chwialkowski et al. (2016, Thm. 2.1) in order to
establish that
q, q 0 ∈ Q(k) ∩ R(k), T (q, q 0 , k) = 0
=⇒
q = q0.
Indeed:
Z
0
Z
S [k(x, ·)]Q(dx) −
T (q, q , k) =
q0
Sq0 [k(x, ·)]Q0 (dx)
{z
}
|
H
=0 from Lemma 5
=
Z X
d
∇xi k(x, ·) + k(x, ·)∇xi log q (x) Q(dx)
0
i=1
=
H
Z X
d
∇xi k(x, ·) + k(x, ·)∇xi log q(x) Q(dx)
i=1
|
+
{z
}
(∗)
Z X
d
k(x, ·)∇xi log q (x) − k(x, ·)∇xi log q(x) Q(dx)
0
i=1
H
|
{z
}
(∗∗)
where
Z
(∗) =
Sq [k(x, ·)]Q(dx),
which is the zero function from Lemma 5 and (∗∗) is the mean embedding (Smola et al.,
2007) of the function
d
X
q 0 (x)
g(x) :=
∇xi log
.
q(x)
i=1
Note that g ∈ L2 (X ∪ ∂X , Q) follows from q, q 0 ∈ Q(k). By assumption this embedding
satisfies
Z X
d
q 0 (x)
∇xi log
k(x, ·)Q(dx) = 0.
q(x)
i=1
H
29
Since H is cc-universal and g ∈ L2 (X ∪ ∂X , Π), the embedding is zero if and only if g = 0
(Carmeli et al., 2010, Thm. 4.4c). Thus
d
X
∇xi log
i=1
q 0 (x)
= 0.
q(x)
This implies that q 0 and q are proportional on X and, since q, q 0 are both densities in C 1 (X ∪
∂X ), we must conclude q = q 0 .
Step #2: It is sufficient to prove that H+ is dense in (C 1 (X ∪ ∂X ), k · k2 ), since this set
is dense in (L2 (X ∪ ∂X , Π), k · k2 ). Now, suppose H+ is not dense in (C 1 (X ∪ ∂X ), k · k2 ).
Then there exists 0 6= f ∈ C 1 (X ∪ ∂X ) such that
Z
Z
f dΠ = 0,
f ψ dΠ = 0 ∀ ψ ∈ H0 ,
the second requirement representing orthogonality of f with respect to H0 . Compactness of
X ∪ ∂X implies f ∈ L∞ (X ∪ ∂X ). Let
Q :=
c+f
Π,
c
Q0 := Π,
c := 1 + kf k∞ ,
so that Q, Q0 are both distributions with Q 6= Q0 . Moreover, under (A2̄,4), both Q, Q0 admit
densities q, q 0 such that q, q 0 ∈ Q(k) ∩ R(k). Indeed, q ∈ Q(k) since
(a) q ∈ C 1 (X ∪ ∂X )
(b) q > 0 on X ;
∇ f
xi
(c) ∇xi log q = c+f
+ ∇xi log π ∈ L2 (X ∪ ∂X , Q0 ) for all distributions Q0 on (X ∪ ∂X , B),
since ∇xi f ∈ C 0 (X ∪ ∂X ) and c + f ≥ 1 on X ;
and q ∈ R(k) since q(x)k(x, x0 ) = c+fc(x) π(x)k(x, x0 ), where c+fc(x) is bounded on X and
π(x)k(x, x0 ) = 0 for x ∈ ∂X and x0 ∈ X ∪ ∂X .
The purpose of this construction becomes clear from plugging this choice of q, q 0 into the
operator T :
Z
Z
0
T (q, q , k) =
Sq0 [k(x, ·)]Q(dx) − Sq0 [k(x, ·)]Q0 (dx)
H
=
1
c
Z
Sq0 [k(x, ·)] f (x)Π(dx)
| {z }
(∗∗∗)
.
H
The function (∗ ∗ ∗) belongs to H0 ; the definition of f then implies that this integral is zero
(orthogonality with respect to H0 ) and so we conclude that T (q, q 0 , k) = 0. From Step #1
we then conclude that q = q 0 . This is a contradiction and so H+ has been shown to be dense
in (C 1 (X ∪ ∂X ), k · k2 ).
30
Smoothness
b=0
(i.e. k ∈ C21 )
b=1
(i.e. k ∈ C22 )
b=2
(i.e. k ∈ C23 )
Function
ϕ(z)
ϕ(1) (z)
ϕ(2) (z)
ϕ(z)
ϕ(1) (z)
ϕ(2) (z)
ϕ(z)
ϕ(1) (z)
ϕ(2) (z)
Expression
[(` + 1)z + 1](1 − z)`+1
+
−(`2 + 3` + 2)z(1 − z)`+
(`2 + 3` + 2)[(` + 1)z − 1](1 − z)`−1
+
`+2
1
2
2
3 [(` + 4` + 3)z + 3(` + 2)z + 3](1 − z)+
− 13 (`2 + 7` + 12)z[(` + 1)z + 1](1 − z)`+1
+
1 2
2 2
2
2
(`
+
7`
+
12)(`
z
+
4`z
−
`z
+
3z
− 1)(1 − z)`+
3
`+3
1
3
2
3
2
2
15 [(` + 9` + 23` + 15)z + (6` + 36` + 45)z + 15(` + 3)z + 15](1 − z)+
`+2
1
2
2 2
2
2
− 15 (` + 11` + 30)(` z + 4`z + 3`z + 3z + 6z + 3)z(1 − z)+
`+1
1
2
3 3
2 3
3
2
3
2
15 (` +11`+30)(` z +9` z +23`z +6`z −3`z +15z +15z −3z −3)(1−z)+
Table 1: Wendland’s compact support functions and their first and second derivatives. Here
k
` = b d2 +b+2c. Note that z+
should be interpreted as (z+ )k rather than (z k )+ . The associated
d
3
native spaces are equivalent to H 2 +b+ 2 (Rd ).
5.3
Formulae for Wendland Kernels
In this final section, explicit formulae for derivatives of the compact support radial functions
of Wendland (1995) are provided in Table 1. To derive the kernel k+ , let r = kx − x0 k2 .
Then the identities
x − x0 (1) r
ϕ
∇x k̃ = −∇x0 k̃ =
hr
h
1 (2) r
∇x · ∇x0 k̃ = − 2 ϕ
h
h
and
∇x k(x, x0 ) = [∇x δ(x)]δ(x0 )k̃(x, x0 ) + δ(x)δ(x0 )∇x k̃(x, x0 )
∇x0 k(x, x0 ) = δ(x)[∇x0 δ(x0 )]k̃(x, x0 ) + δ(x)δ(x0 )∇x0 k̃(x, x0 )
∇x · ∇x0 k(x, x0 ) = [∇x δ(x)] · [∇x0 δ(x0 )]k̃(x, x0 ) + δ(x0 )[∇x δ(x)] · [∇x0 k̃(x, x0 )]
+δ(x)[∇x0 δ(x0 )] · [∇x k̃(x, x0 )] + δ(x)δ(x0 )∇x · ∇x0 k̃(x, x0 )
can be used to obtain the kernel k+ = c + k0 in closed-form, in combination with Eqn. 8.
References
Aronszajn, N. (1950) Theory of reproducing kernels. T. Am. Math. Soc., 68(3), 337-404.
Carmeli, C., De Vito, E., Toigo, A. and Umanita, V. (2010) Vector valued reproducing kernel
Hilbert spaces and universality. Anal. Appl., 8, 19-61.
Roberts, G. O. and Rosenthal, J. S. (1998a) Two convergence properties of hybrid samplers.
Ann. Appl. Prob., 8(2), 397-407.
Roberts, G. O. and Rosenthal, J. S. (2008) Variance bounding Markov chains. Ann. Appl.
Prob., 18(3), 1201-1214.
31
| 10math.ST
|
Between Homomorphic Signal Processing and Deep
Neural Networks: Constructing Deep Algorithms
for Polyphonic Music Transcription
Li Su∗
arXiv:1706.08231v1 [cs.SD] 26 Jun 2017
∗
Institute of Information Science, Academia Sinica, Taiwan
E-mail: [email protected] Tel: +886-2788-3799
Abstract—This paper presents a new approach in understanding how deep neural networks (DNNs) work by applying
homomorphic signal processing techniques. Focusing on the task
of multi-pitch estimation (MPE), this paper demonstrates the
equivalence relation between a generalized cepstrum and a DNN
in terms of their structures and functionality. Such an equivalence relation, together with pitch perception theories and the
recently established rectified-correlations-on-a-sphere (RECOS)
filter analysis, provide an alternative way in explaining the role
of the nonlinear activation function and the multi-layer structure,
both of which exist in a cepstrum and a DNN. To validate the
efficacy of this new approach, a new feature designed in the same
fashion is proposed for pitch salience function. The new feature
outperforms the one-layer spectrum in the MPE task and, as
predicted, it addresses the issue of the missing fundamental effect
and also achieves better robustness to noise.
I. I NTRODUCTION
Automatic music transcription (AMT) refers to the task
of converting acoustic music signals into symbolic notation,
such as the onset time, offset time, pitch, and others. Since
music is typically composed of multiple overlapped sources
with diverse spectral patterns spreading over a wide frequency
range, AMT is a highly challenging task, and is considered to
be the holy grail in the field in machine listening of music
[1], [2]. The most fundamental technique required for AMT
is pitch detection. Depending on the types of source signals
and the scenario in application, pitch detection algorithms are
designed in different ways, such as single-pitch detection for
monophonic music, or melody tracking and and multi-pitch
estimation (MPE) for polyphonic music. The main focus of
this paper is on the MPE task, the core task in building
an AMT system. Solutions to the MPE task include featurebased approaches such as pitch salience functions [3], [4], [5],
nonnegative matrix factorization (NMF) [6], probability latent
component analysis (PLCA) [7], [8] as well as convolutional
sparse coding (CSC) [9], [10], to name but a few.
Recently, deep learning approaches have gained increasing
attention in polyphonic pitch detection [11], [12], [13], [14],
[15]. In [13], Verma and Schafer used time-domain DNNs
for melody tracking in polyphonic music, and found that the
learned networks resemble traditional pitch detection methods:
the first layer behaves like a spectral analyzer while the second
layer behaves like a comb filter for saliency mapping. Although such observations keep providing indirect evidence on
the physical meanings of deep learning, it is still a highly empirical technique, with its superior performance unexplainable
and its limitation unknown. Moreover, unlike its success in
other fields such as image processing and speech recognition,
deep learning approaches have not been proven state-of-the-art
in the problems of pitch detection. In a recent work on MPE,
Kelz shows that when the training data contains concurrent
notes, neural networks suffer from the entanglement issue,
which somehow implies a limited deduction power of deep
learning on this problem [15]. These facts reveal the urgency
of understanding the theory behind deep learning, and suggest
a direction to ask if conventional signal processing techniques
can help one understand deep learning.
Recent studies focusing on how to understand deep learning are based on diverse approaches, including scattering
transform [16], [17], [18], Taylor expansion [19], generative
modeling [20], renormalization groups [21], probability theory [22], tensor analysis [23], saliency map or layer-wise
relevance propagation [24], [19], [25] and cascaded filtering
[26], [27]. Particularly, in [26], [27], Kuo interprets the convolutional neural networks (CNN) as the so-called REctifiedCOrrelations on a Sphere (RECOS) filters, which answer the
question why nonlinear activation functions and multi-layer
structures are useful: a nonlinear activation function eliminates
the information negatively correlated to the anchor vectors
(i.e., frequently occurring patterns or dictionary atoms) and,
when the filters are cascaded in layers, the system improves
itself in each processing step by avoiding confusion with
negative-correlation samples, noise, and the background. Since
it attempts to bridge the gap between different languages used
in signal processing and deep learning, such an approach is
noteworthy for the research in audio signal processing among
all the above-mentioned approaches.
This paper will show that the structure similar to the RECOS
filter can also be applied in MPE and, maybe surprisingly,
such a structure do exist in many traditional pitch detection
algorithms, including the autocorrelation function (ACF), cepstrum (a.k.a. homomorphic signal processing) [28], as well
as the YIN algorithm [29], where the RECOS filters are
represented as the Fourier transform followed by a piecewise
multiplication. In other words, those traditional pitch detection
functions inherently have a DNN-like structure. To the best
of the author’s knowledge, this fact has never been discussed
before. In addition, by reviewing the issues such as the missing
fundamental effect in pitch perception, this paper will also
provide an perceptual interpretation of the nonlinear activation
function in a DNN. Moreover, to prove that the proposed
statements do inspire one to understand how a deep algorithm
works on MPE, a novel feature computed with a three-layer
network, named the generalized cepstrum of spectrum (GCoS),
is proposed based on the statements. Experiments on piano and
mixed-instrument datasets show that the GCoS outperforms
the baseline method by showing its ability in detecting missing
fundamentals and its robustness to noise.
This paper is organized as follows. Section II reviews two
classic topics on the role of nonlinearity in pitch detection,
the first being the theories explaining the missing fundamental
effects, and the second being the cepstrum. Section III gives
a generalized framework for pitch detection algorithms and
discusses its relation to DNNs and RECOS filters. Remaining
parts go for experiments, results and conclusions.
II. N ONLINEARITY AND PITCH DETECTION
In this section, the nonlinearity of pitch detection will be
investigated in the context of perception and signal processing.
The missing fundamental effect will be taken as an example of
study in the context of psychoacoustics, while cepstrum will
be taken as an example in the context of signal processing.
In later sections, such investigation will be shown useful to
explain the use of nonlinear activation in pitch detection.
A. Missing fundamentals
The missing fundamental effect refers to the phenomenon
that one can perceive a pitch when the source signal contains
few or no spectral components corresponding to that pitch
and its low-order harmonics [30]. More specifically, one can
perceive the pitch of a note at frequency f0 when there
is no spectral component at f0 , and even no 2f0 in the
spectrum – there could be only spectral components at nf0 ,
(n + 1)f0 , · · · , (n + k)f0 for some n, k > 1. As a result, pitch
detection algorithms cannot be designed by using spectrum
only. Though seemingly unusual, the missing fundamental
effect does exist commonly in musical signals; it is commonly
found in the low-pitch notes of various kinds of instruments
such as the 88-key piano, where the physical size of the
instrument is too small to effectively radiate the signal with
long acoustic wavelength. The same also applies to the music
played with a cellphone, where one can still hear bass notes,
although distorted, out of a tiny Lo-Fi speaker.
The missing fundamental effect has been studied and explained with mainly three different approaches.1 First, the
period or time theory holds that a pitch is determined by the
minimal period of the waveform rather than the fundamental
frequency of the signal. In contrast to the place theory which
is based on Fourier transform, time theory explains the missing fundamental effect: the period of a signal with spectral
1 The discussion here mainly focuses on the main ideas of the theories rather
than the relations and debates among them in the history. For more historical
backgrounds of the pitch perception theories, readers are referred to [30].
Fig. 1. Illustration of a pitch detection algorithm based on the cepstrum.
components at nf0 , (n + 1)f0 , · · · , (n + k)f0 obviously has
a period at 1/f0 . The idea that “pitch is period” is the basic
assumption of many pitch detection algorithms, including the
cepstrum which will be introduced later.
Second, the theories incorporating nonlinear effects are
many: for example, Helmholtz claimed that one can hear the
missing fundamental because of the nonlinear effect in the
ear; two sinusoidal components at frequencies f1 and f2 would
generate a component at the frequencies mf1 +nf2 , m, n ∈ Z
[31]. Although this theory was rejected by Schouten through
further experiments [30], it is still inspiring for signal processing: missing fundamentals can be detected by producing
cross terms through nonlinear effects. Another example is the
neural cancellation filters proposed by de Cheveigné [32],
which introduces the concept of neural networks for MPE.
Third, the pattern matching theory proposed by Goldstein
holds that the auditory system resolves the sinusoidal components of a complex sound, and encodes its pitch which best
fits the harmonic series of components no matter whether
the f0 component exists in the source signal or not [33].
The idea of pattern matching is the basis of many MPE
algorithms, such as those using NMF or PLCA [6], [2].
However, for MPE, a pattern matching approach also faces
challenges such as overlapped harmonics and indeterminacy
of the lowest note. Since this paper mainly focuses on the
role of deep structure and nonlinearity in pitch detection, the
pattern matching approach will not be discussed.
B. Cepstrum
The cepstrum, also referred to as homomorphic signal
processing, is a classic signal processing method with three
operations: a Fourier transform, a nonlinear transform (usually a logarithm function), followed by an inverse Foueier
transform. [34]. Since its invention in 1963 [35], the cepstrum
and its derived features have been applied in various signal
processing tasks, such as deconvolution, image enhancement,
speech recognition, and pitch detection, to name but a few. A
thorough review of the cepstrum can be found in [36], [34].
The cepstrum-based pitch detection algorithm works under
a general assumption: the fast-varying part of a spectrum,
usually the regularly-spaced harmonic peaks, are important
for pitch detection, while the slow-varying part, such as the
spectral envelope, should be discarded since it is not related to
pitch, as shown in Fig. 1. To separate the fast-varying part from
the slow-varying one, the spectrum is transformed into the
cepstrum in the quefrency (or lag) domain through an inverse
Fig. 2. Conceptual illustration of Equations (1)-(3). Homomorphic signal
processing (i.e. cepstrum) is equivalent to a 2-layer network.
Fourier transform.2 As a result, the non-pitch information
would lie in the low-quefrency (i.e. short-time) region. This
term can be discarded by means of a high-pass filter. Notice
that since the levels of the harmonic peaks in the spectrum
vary largely, a nonlinear scaling (usually a logarithm scale) on
the spectrum before the inverse Fourier transform is required.
Finally, the detected pitch frequency is then the inverse of lag
corresponding to the maximal value of the filtered cepstrum,
as shown in the right part of Fig. 1. The relation between a
filtered cepstrum and a DNN will be discussed in the next
section.
III. G ENERALIZED F ORMULATION OF P ITCH S ALIENCE
F UNCTIONS
Since the main focus of this paper is on frame-level transcription of polyphonic music, all features are described as
the signal processing of an N -point, time-domain segment
x ∈ RN , a frame of music signal in computing the short-time
Fourier transform (STFT). All vector throughout this paper
are zero-indexed, i.e. x = [x[0], x[1], · · · , x[n], x[N − 1]].
The N -point windowed discrete Fourier transform (DFT) is
denoted by the operator F ∈ CN ×N . Similarly, the N -point
inverse DFT is represented by F−1 . Denote |x| as the absolute
value of each element in x. Consider the following equations:
z(1) = σ (1) |Fx| + b(1) ,
(1)
z(2) = σ (2) W(2) F−1 z(1) + b(2) ,
(2)
z(3) = σ (3) W(3) Fz(2) + b(3) .
(3)
where σ (i) is a element-wise nonlinear transform function such
that for γi > 0, i = 1, 2, 3,
γ
x i,
x > 0;
(i)
σ (x) =
(4)
0,
x ≤ 0,
W(i) ∈ RN ×N is the weighting matrix for filtering the feature
of interest, and b(i) is the bias vector.
Consider the the first bN/2c-th elements (i.e. the positivefrequency or the positive-quefrency part) of z(i) . First, assume
W(i) = I and b(i) = 0 to simplify the discussion. By
definition, z(1) := z(1) [k] in (1) is the magnitude spectrum
2 Since quefrency and lag have the same unit as time, quefrency, lag and
time are used interchangeably in the following discussion.
of x estimated from the DFT. The k-th element of z(1) ,
z(1) [k], is the spectral element at the corresponding frequency
f [k] = kfs /N , k = 0, 1, · · · , bN/2c, where fs is the sampling
frequency. Then, z(2) in (2) is known as the generalized
autocorrelation function (GACF) or the generalized cepstrum
(GC) in the literature, since this formulation is equivalent to
the ACF when γ = 2, and is also equivalent to the cepstrum
up to a linear transformation when γ → 0 [28], [37].3 As the
inverse DFT of a spectral feature in the frequency domain,
z(2) is also said to be a feature in the lag or quefrency domain,
which has the same unit as time. Therefore, for convenience,
the elements of z(2) are indexed by n in this paper. The n-th
element, z(2) [n], represents the salience of the feature at the
corresponding lag q[n] = n/fs , n = 0, 1, · · · , bN/2c. The
GC has been known as a good feature for MPE by setting
0 < γ1 < 1, such as γ1 = 0.67 [38], 0.6 [39], 0.25 [40]
and 0.1 [41]. The generalized ACF is also known as the root
cepstrum, and it has been shown more robust to noise than
the logarithm cepstrum in the literature of speech processing
[42], [43]. Interestingly, z(2) with either γ1 = 2 (i.e. ACF) or
γ1 = 0 (i.e. cepstrum) works only for single-pitch detection
rather than MPE [44], [38]. One may view GC as a more
stable feature than cepstrum since the logarithmic operation is
numerically unstable.
Finally, z(3) in (3) is again a feature indexed by k in
frequency domain, and z(3) [k] represents its k-th element
corresponded to the frequency f [k]. The feature z(3) in Equation (3) is less discussed in the literature of pitch detection.
An exception is [45], where Peeters considered using the
ACF of the magnitude spectrum as a feature for single-pitch
detection, and showed its advantage in identifying the missing
fundamentals unseen in a spectrum. This is because that
although the peak at the fundamental frequency is attenuated,
the ACF still captures the periodicity in between the peaks of
high-order harmonics by mixing those peaks up and producing
a cross-term at the true fundamental frequency. The ACF of
spectrum is a special case of z(3) , where γ1 = 2, γ2 = 1, and
γ3 = 1. It is therefore straightforward to generalize z(3) by
making both γi tunable parameters, where 0 < γi ≤ 2. Since
z(3) is computed by a DFT, a nonlinear activation function
and another subsequently on z(1) , it can be interpreted as
the Generalized Cepstrum of Spectrum (GCoS). This novel
feature is firstly proposed and its superior performance to other
traditional spectral features will be shown in the experiments.
As mentioned in Section II-B, the terms lying in both the
low-k and low-n indexes are uninformative for MPE since
they are unrelated to pitch. The function of W(i) and b(i)
is therefore to discard the terms unrelated to pitch or outside
the pitch detection region. Assuming that b(i) is slow-varying,
the contribution of b(i) would simply be the terms discarded
by W(i+1) . Therefore, for simplicity, b(i) is set to be 0
3 If the term xγ in (4) is replaced by (xγ − 1) /γ, then σ
γ,δ (x) |x>δ
converges to log x as γ → 0 [28], [37]. In this case, z(2) is the cepstrum.
throughout this paper, and W is a diagonal matrix such that
1,
l > kc (i = 1, 3) or nc (i = 2) ;
(i)
W [l, l] =
(5)
0,
otherwise .
where kc and qc are the indices of the cutoff frequency and
cutoff quefrency, respectively. It means that W(2) is a highpass filter with cutoff quefrency at qc = nc /N and W(3) is
a high-pass filter with cutoff frequency at fc = kc fs /N .4 A
straightforward suggestion to the value of fc and qc is the
lowest pitch frequency and the shortest pitch period. In this
paper, fc = 27.5 Hz (frequency of A0) and qc = 0.24 ms
(period of C8).
Besides spectrum, cepstrum, and GCoS, there are also other
types of pitch salience functions which can be summarized by
Equations (1)-(3). For example, the pitch salience function of
the YIN algorithm is [29]:
N
X
2
(x[q] − x[q + n]) ,
(6)
q=1
When the source signal is assumed stationary, derivations
show that (6) cab be represented by z(2) , where γ1 = 2, γ2 =
1, W(2) = −2I, and b(2) = 2z(2) [0]. This bias term makes
the algorithm more stable to amplitude changes.5
A. Relation to DNNs
As illustrated in Fig. 2, Equations (1)-(3) explicitly resemble a DNN with three fully-connected layers. Specifically, a
Fourier spectrum is equivalent to a one-layer network, an ACF,
cepstrum, or GC are equivalent to a two-layer network, and an
ACF of spectrum or GCoS are then equivalent to a three-layer
network.
Equations (1)-(3) also possess some important characteristics which are different from common DNNs:
1) The fully-connected layers are complex numbers (i.e.,
the DFT matrix) while the commonly-used DNNs are
typically real-valued. Notice that since z(1) is symmetric, the Fourier transform in the second and the third
layers can be replaced by a real-valued discrete cosine
transform (DCT) matrix without changing the result.
2) The nonlinear activation function is a power function;
the widely-used rectified linear unit (ReLU) function is
a special case where γ = 1. However, previous studies
consistently suggest a rather sharp nonlinearity where
γ < 1, since taking γ = 1 does not work well in MPE
[38], [39], [40], [41].
3) Since the fully-connected layers are determined clearly
by the Fourier transform, each feature in each layer has a
4 More precisely, W(2) should be called a “long-pass lifter” rather than
“high-pass filter” in order to distinguish the filtering processing in the
quefrency domain from the one in the frequency domain. This paper uses
“high-pass filter” for both cases to simplify the terminology.
5 Equations (1)-(3) and the discussion on ACF are based on the assumption
that the source signal is stationary. If the source signal is non-stationary, the
formulation of (1)-(3) as well as the resulting weighting and bias factors
here should be slightly modified Since the Wiener-Khinchin theorem does
not hold anymore. The resulting network should be formulated in a structure
like a CNN and will not be discussed here.
Fig. 3. The spectra (z(1) , top), GCoS (z(3) with γ1 = 0.24 and γ2 = 0.6,
middle), and ACF of spectrum (z(3) with γ1 = 2 and γ2 = 1, bottom) of an audio sample from 17.91 to 18.09 seconds of ‘MAPS MUSalb se2 ENSTDkCl.wav’ (i.e. Cataluña in Albéniz’s Suite Española, Op. 47)
from the MAPS dataset. The pitch frequencies of the four notes consisting
the sample are labeled.
clear dimensional unit. For example, z(1) and z(3) are in
the frequency domain while z(2) is in the time domain.
Moreover, it should be emphasized that, although the network parameters (i.e., the Fourier coefficients) mentioned here
are predetermined rather than learned, they do share similar
properties (i.e., frequency selection) in performing pitch detection [13]. Therefore, the analogy between a cepstrum and a
DNN is not only physically plausible but also provides a new
way for one to better understand why and how a DNN works
in pitch detection problems. This fact can be seen from the
following example.
B. Examples and interpretation
Fig. 3 shows the spectrum (top) with γ1 = 0.24, GCoS
with γ1 = 0.24 and γ2 = 0.6 (middle), and ACF of spectrum
(bottom) of a piano signal with four pitches D#2 (77.78 Hz),
A#2 (116.54 Hz), F#3 (185.00 Hz), and A#3 (233.08 Hz). In
the last two cases γ3 is set to 1.6 The sampling frequency is
fs = 44100 Hz, and the window function h is the Blackman
window. To investigate the effect of noise interference on the
features, a contaminated signal with pink noise such that the
SNR = 10 dB is also considered. All of the illustrated spectra
and GCoS features are normalized to unit l2 -norm.
The clean and noisy spectra both show that the fundamental
frequencies of D#2 and A#2 are weak comparing to their highorder harmonics. There are numbers of peaks unrelated to the
true fundamental frequencies and their harmonics, especially
in the noisy spectrum. There is also no clear trend in the
6 A different value of γ merely changes the scale of the output saliency
3
function but does not change the result of pitch detection. Therefore, for
simplicity, γ3 is set to be 1 for all cases in this paper.
spectra such that one can set a threshold function to discard
those peaks. As a result, the spectrum is not an effective pitch
salience function, since it is sensitive to noise and is unable
to identify weak fundamental frequencies.
These problems are well resolved in the GCoS. As shown
in the middle of Fig. 3. By setting γ1 = 0.24 and γ2 = 0.6,
the GCoS feature enhances two identifiable peaks at the
frequencies of D#2 and A#2. For the noisy signal, most of
the fluctuation peaks are eliminated except for a low-frequency
peak at 35 Hz, which is the only unwanted cross term produced
by the nonlinear activation function. Moreover, the GCoS of
the noisy signal is nearly identical to the one of the clean
signal; such robustness to noise cannot be seen in spectrum.
The reasons why GCoS works well are explained as follows.
From the discussion in Section II-B, pitch-related information
is the periodic components in the signal. Moreover, the Fourier
transform of a periodic signal typically has a harmonic pattern,
which is also periodic, while the non-periodic components
would lie in the low-frequency or low-quefrency regions.
Therefore, the non-periodic components in the i-th layer are
eliminated by the high-pass filter W(i+1) in the (i+1)-th step.
The periodic components are preserved and enhanced through
the Fourier transform F in each step. Those components
having negative correlation to the DFT basis, mostly noise
and non-periodic ones, are eliminated through rectification in
σ (i) . In other words, similar to the discussion in [26], [27],
W(i) F is a set of RECOS filters for pitch detection, where its
anchor vectors form the Fourier basis.
For the ACF of spectrum, there could also be an alternative
interpretation: since z(3) is an ACF, the anchor vector could
be the time shifting of the input spectrum (i.e. z(1) ) itself. It
computes the correlation of the spectrum to the time shifting
of the itself and discards the components inducing negative
correlation. Such a viewpoint also explains the reason why
the ACF of spectrum is useful in single-pitch detection [45].
It should be noticed that, however, the ACF of spectrum
is not feasible for MPE. The bottom of Fig. 3 shows that,
although there are some cross terms produced in the ACF of
spectrum, the peaks of these cross terms are actually not at
the frequencies of D#2 and A#2, possibly because of other
nonlinear effects produced by other terms. This is the reason
why a sharp nonlinear activation function, i.e. 0 < γi ≤ 1, is
usually suggested for MPE.
C. Combining features in different layers
The idea of combining frequency-domain and time-domain
features has been proposed by Peeters [45] and Emiya et al.
[46] for single-pitch detection, and by Su et al. for MPE.
The idea works well in detecting the pitch frequencies by
utilizing the complementary structure of the two features: the
frequency-domain features with harmonic peaks and the timedomain features with sub-harmonic peaks. The frequencydomain feature is prone to upper octave errors while robust to
lower octave errors [47], [48], and the time-domain feature
be vice versa.7 Therefore, combining them together could
7 Octave
error means the estimated pitch is off by one or multiples of octave.
suppress both the upper and the lower octave errors. In other
words, fusion of two features at different layers, one in the
time domain and the other in the frequency domain, gives rise
to an improved pitch salience function L := L z(i) , z(i+1) .
More specifically, we consider
N
(7)
L z(1 or 3) , z(2) = z(1 or 3) [k]z(2) b e
k
where the time-domain feature z(2) is nonlinearly mapped into
the frequency domain, and d·c is the rounding function. For
example, [45] considered some cases including the combination of the ACF of spectrum (z(3) , γ1 = 1, γ2 = 2, γ3 = 1)
and cepstrum (z(2) , γ1 = 0, γ2 = 1). In this paper, z(2) and
z(3) are constructed with the same set of networks, while the
parameters γ (i) are arbitrary values between 0 and 2. Besides,
to improve the performance, more steps for post-processing
and pitch selection are used.
IV. E XPERIMENT SETTINGS
Since this paper is to investigate the multi-layer construction
of pitch-related features, a learning-based framework for MPE
is not considered now. the MPE method combining z(1) and
z(2) introduced in [4] is taken as the baseline method for
comparison. The purpose of the experiment is to compare
two frequency-domain features, GCoS (z(3) ) and magnitude
spectrum (z(1) ), and show that z(3) outperforms z(1) in the
following senses:
1) elegant design: the feature can be used directly without
pseudo whitening and adaptive thresholding like in [49],
[4] to filter out the spectral envelope in z(1) while
achieve a more succinct feature representation.
2) detect missing fundamentals: the method can be designed for detecting missing fundamentals without handcrafted rules in [4].
3) robustness to noise: z(3) outperforms z(1) because it
cancels the uncorrelated parts with more layers. This
effect could be more obvious with low SNR.
The source code of the proposed as well as the baseline
methods will be announced publicly.
A. Pre-processing
After computing the frequency-domain feature and the timedomain feature, the same procedure in [4] is adopted to compute the pitch profiles of the both features for pitch selection.
The features pairs
z(3) , z(2) are mapped into the pitch pro
(3) (2)
files z̄ , z̄
through a filterbank which center frequencies
are according to the equal-tempered scale indexed by the pitch
number p such that p = P(f ) = 69 + b12 log2 f /440e, where
b·e denotes the round function. For A0 (f0 = 440 Hz), p = 69.
Then, the elements corresponding to the pitch number p in z(3)
and z(2) are merged into the p-th element respectively in z̄(3)
and z̄(2) through max pooling. That is,
k0
(3) 0
= p , (8)
z̄(3) [p] = max
z
[k
]
s.t.
P
f
s
k0
N
fs
z̄(2) [p] = max
z(2) [n0 ] s.t. P
= p.
(9)
0
n
n0
TABLE I
MPE R ESULT ON MAPS AND TRIOS DATASETS . P RECISION (P), RECALL
(R) AND F- SCORE (F) IN % ARE LISTED .
Dataset
MAPS
TRIOS
Pitch
All
Bass
Treble
All
Bass
Treble
P
69.91
64.14
70.57
70.57
85.99
68.25
Proposed
R
68.94
48.71
72.03
62.60
54.00
64.55
F
69.42
55.37
71.29
66.34
66.34
66.35
P
70.51
86.51
71.14
81.67
80.64
81.92
Baseline
R
68.26
39.38
70.95
53.22
54.37
52.96
F
69.37
54.13
71.05
64.45
64.95
64.33
It is worth mentioning that this process resembles a maxpooling layer in a CNN. However, to make the feature fit the
log-frequency scale for pitch detection, the filter size and the
stride vary with the frequency and the time index.
B. Pitch selection process
The criteria of selecting the pitch and the sparsity constraints
for reducing false positives proposed in [4] are applied both
the proposed and baseline methods. Specifically, by setting
the harmonic/subharmonic series are with length four (i.e.
the fundamental frequency/period and the first three harmonics/subharmonics) and the sparse constraint parameter δ = 0.8,
a pitch pi is a true positive if
1) z̄(3) [pi ], z̄(3) [pi + 12], z̄(3) [pi + 19], z̄(3) [pi + 24] > 0
2) z̄(2) [pi ], z̄(2) [pi − 12], z̄(2) [pi − 19], z̄(2) [pi − 24] > 0
3) kz̄(3) [pi : pi +24]k0 < 25δ or kz̄(2) [pi −24 : pi ]k0 < 25δ
where pi : pi + 24 means the indices from pi to pi + 24. All
pitch activations satisfying these conditions form a piano roll.
Finally, a median filter with length of 25 frames is applied on
the resulting piano roll in order to smooth the note activation
and prune the notes with too short duration.
Notice that unlike [4], the rules of selecting missing fundamentals and stacked harmonics are not applied in this works.
Experiments will show that the proposed method can properly catch the information of missing fundamentals without
introducing these rules.
C. Datasets
We consider two MPE datasets in our evaluation. The first
dataset, MAPS, is a widely used piano transcription dataset
created by Emiya et al. [50].8 It contains audio recordings
played on Yamaha Disklavier, an automatic accompaniment
piano, in accordance with MIDI-aligned ground-truth pitch
annotations. Following previous work [6], [7], [8], [47], [51],
[52], we use the first 30 seconds of the 30 music pieces in the
subset ENSTDkCl in our evaluation, totalling 15 minutes.
The second dataset, TRIOS, consists of five pieces of fully
synthesized music in trios form [53].9 Each piece contains
piano, two other pitched instruments (the two instruments
are different for different pieces). One piece has non-pitched
percussions. The pitch annotations were for the three pitched
instruments in each piece, whose length ranges from 17 to 53
seconds, totalling 3 minutes and 11 seconds.
8 http://www.tsi.telecom-paristech.fr/aao/en/2010/07/08/
9 http://c4dm.eecs.qmul.ac.uk/rdr/handle/123456789/27
Fig. 4. F-scores of the MAPS and TRIOS datasets with different SNR levels
for pink noise. (◦): proposed. (): baseline.
D. Parameters and evaluation metrics
The sampling frequency for all source signals is 44.1 kHz.
For computation of the frame-level features, the BlackmanHarris window with size of 0.18 sec is used and the hop size
is 0.01 sec. The pitch range considered in evaluation is from
A1 (55.55Hz) to C7 (2205 Hz); both the ground truth and
the experiment results are restricted in this range. Further, to
identify the performance on the missing fundamentals, the bass
notes, defined as the notes below C3, and the treble notes,
defined as the notes above C3, are also evaluated separately.
Following the standard in MIREX MF0 challenge,10 the performance of MPE is evaluated using the micro-average framelevel Precision (P), Recall (R), and F-score (F). After counting
the number of true positives (Ntp ), false positives (Nf p ) and
false negatives (Nf n ) over all the frames within a dataset,
The evaluation metrics are defined as: P = Ntp /(Ntp + Nf p ),
R = Ntp /(Ntp + Nf n ) and F = 2P R/(P + R).
To evaluate the robustness to noise for both the baseline
and proposed methods, we use the audio degradation toolbox
(ADT) [54] to add pink noise with signal-to-noise ratio (SNR)
ranging from 30 dB (least noisy) to 0 dB (most noisy) to every
piece in the dataset.
V. R ESULTS AND DISCUSSION
Table I lists the resulting precision, recall and F-score
evaluated with 3 sets of pitch ranges: All (A1 to C7), Bass, (A1
to C3), and Treble (C3 to C7). For all pitch sets and datasets,
the proposed method outperforms the baseline method in Fscores. This implies an overall improvement of using GCoS.
For the Bass set, the F-score is improved by 1.24% in MAPS
and by 1.39% in TRIOS. Such an improvement is more than
the ones in the sets of All or Treble in the MAPS dataset. This
fact indicates that the proposed feature can better recognize
missing fundamentals even without the hand-crafted rules [4]
in selecting missing fundamentals. However, the improvement
of the proposed method in either P or R in both datasets
behaves inconsistently. In the Bass set of MAPS, the proposed
method improves R by 9.33% while degrades P by 24.37%.
Conversely, in TRIOS, the proposed method improves P while
degrades R. The possible reason of such a difference is the
10 http://www.music-ir.org/mirex/wiki/2016:Multiple Fundamental
Frequency Estimation %26 Tracking
property of the input data: TRIOS contains less low-frequency
noise since it is constructed with synthetic data. As a result,
there are less unwanted low-frequency terms, making more
of the cross terms be true pitches, and therefore improving
precision more than recall.
Fig. 4 shows that the improvement of F-score in noisy
source data is much more than the improvement in clean data.
The lower the SNR is, the more improvement is found in the
proposed method. Specifically, when the SNR is lower than 10
dB, the improvement in both datasets is over 5%. This verifies
the statement that the GCoS is more robust to noise since it
has one more layer of RECOS filters in refining the features.
A. Discussion
The idea of using multiple Fourier transforms can be
interpreted by one intuition: the Fourier transform of a periodic signal also have periodic (i.e., harmonic) patterns and,
perceptually, the strength of such periodic patterns depends
on an appropriate nonlinear scaling on the input, and the
nonlinear scaling function has some perceptual basis, such as
the dB scale or Stevens’ power law [55]. This intuition directly
suggests promising future directions of applying other kinds
of nonlinear activation function and more than three layers of
Fourier transform to construct the pitch salience function.
Another remaining issue is that the discussion in this paper
still not includes the learning aspect. In fact, the parameter γi
cannot be learned from gradient descent and back propagation,
since the gradient of σ (i) at zero diverges. Therefore, there is
a need to find a differentiable nonlinear activation function
which could replace the role of the power function in MPE
or, to investigate the potential of other optimization methods
such as coordinate descent. Besides, since using the DFT
matrix works well in MPE, it is also important to investigate a
scenario where a learning-based DNN for the MPE task with
the network parameters initialized with a DFT or DCT matrix.
VI. C ONCLUSIONS
This paper has presented a signal-processing perspective for
one to better understand the reason why deep learning works
by demonstrating a new MPE algorithm which generalizes
the concept of homomorphic signal processing and DNN
structures. Containing layers of DFT matrices to extract the
periodic components, high-pass filters to discard non-periodic
components, and nonlinear activation functions to eliminate
the components negatively correlated with the DFT matrix,
the algorithm has shown superior performance in detecting
missing fundamentals and in noisy sources. Although the MPE
task is merely one specific pattern recognition problem, this
approach can still be viewed as an example of RECOS filter
analysis for one to better understand deep learning. With
the positive experiment results, this paper has positioned one
more step to bridge the philosophical and methodological gaps
between signal processing and deep learning, and, provides
new cues in demystifying and understanding deep learning on
other pattern recognition tasks.
R EFERENCES
[1] E. Benetos, S. Dixon, D. Giannoulis, H. Kirchhoff, and A. Klapuri,
“Automatic music transcription: Breaking the glass ceiling,” in ISMIR,
2012.
[2] ——, “Automatic music transcription: challenges and future directions,”
J. Intelligent Information Systems, vol. 41, no. 3, pp. 407–434, 2013.
[3] C. Yeh, A. Röbel, and X. Rodet, “Multiple fundamental frequency
estimation and polyphony inference of polyphonic music signals,” IEEE
Trans. Audio, Speech, Lang. Proc., vol. 18, no. 6, pp. 1116–1126, 2010.
[4] L. Su and Y.-H. Yang, “Combining spectral and temporal representations
for multipitch estimation of polyphonic music,” Audio, Speech, and
Language Processing, IEEE/ACM Transactions on, vol. 23, no. 10, pp.
1600–1612, 2015.
[5] L. Su, T.-Y. Chuang, and Y.-H. Yang, “Exploiting frequency, periodicity and harmonicity using advanced time-frequency concentration
techniques for multipitch estimation of choir and symphony,” in ISMIR,
2016.
[6] E. Vincent, N. Bertin, and R. Badeau, “Adaptive harmonic spectral decomposition for multiple pitch estimation,” IEEE Trans. Audio, Speech,
Lang. Proc., vol. 18, no. 3, pp. 528–537, 2010.
[7] E. Benetos and S. Dixon, “A shift-invariant latent variable model for
automatic music transcription,” Computer Music Journal, vol. 36, no. 4,
pp. 81–94, 2012.
[8] E. Benetos, S. Cherla, and T. Weyde, “An effcient shift-invariant model
for polyphonic music transcription,” in Proc. Int. Workshop on Machine
Learning and Music, 2013.
[9] A. Cogliati, Z. Duan, and B. Wohlberg, “Context-dependent piano music
transcription with convolutional sparse coding,” IEEE/ACM Transactions
on Audio, Speech, and Language Processing, vol. 24, no. 12, pp. 2218–
2230, 2016.
[10] ——, “Piano transcription with convolutional sparse lateral inhibition,”
IEEE Signal Processing Letters, vol. 24, no. 4, pp. 392–396, 2017.
[11] K. Han and D. Wang, “Neural network based pitch tracking in very
noisy speech,” IEEE/ACM Transactions on Audio, Speech and Language
Processing (TASLP), vol. 22, no. 12, pp. 2158–2168, 2014.
[12] S. Sigtia, E. Benetos, and S. Dixon, “An end-to-end neural network
for polyphonic piano music transcription,” IEEE/ACM Transactions on
Audio, Speech and Language Processing (TASLP), vol. 24, no. 5, pp.
927–939, 2016.
[13] V. Prateek and R. W. Schafer, “Frequency estimation from waveforms
using multi-layered neural networks,” in INTERSPEECH, 2016, pp.
2165–2169.
[14] R. Kelz, M. Dorfer, F. Korzeniowski, S. Böck, A. Arzt, and G. Widmer,
“On the potential of simple framewise approaches to piano transcription,” arXiv preprint arXiv:1612.05153, 2016.
[15] R. Kelz and G. Widmer, “An experimental analysis of the entanglement
problem in neural-network-based music transcription systems,” arXiv
preprint arXiv:1702.00025, 2017.
[16] J. Bruna and S. Mallat, “Invariant scattering convolution networks,”
IEEE transactions on pattern analysis and machine intelligence, vol. 35,
no. 8, pp. 1872–1886, 2013.
[17] S. Mallat, “Understanding deep convolutional networks,” Phil. Trans. R.
Soc. A, vol. 374, no. 2065, p. 20150203, 2016.
[18] T. Wiatowski and H. Bölcskei, “A mathematical theory of deep
convolutional neural networks for feature extraction,” arXiv preprint
arXiv:1512.06293, 2015.
[19] G. Montavon, S. Lapuschkin, A. Binder, W. Samek, and K.-R. Müller,
“Explaining nonlinear classification decisions with deep taylor decomposition,” Pattern Recognition, vol. 65, pp. 211–222, 2017.
[20] J. Dai, Y. Lu, and Y.-N. Wu, “Generative modeling of convolutional
neural networks,” arXiv preprint arXiv:1412.6296, 2014.
[21] P. Mehta and D. J. Schwab, “An exact mapping between the variational renormalization group and deep learning,” arXiv preprint
arXiv:1410.3831, 2014.
[22] A. B. Patel, T. Nguyen, and R. G. Baraniuk, “A probabilistic theory of
deep learning,” arXiv preprint arXiv:1504.00641, 2015.
[23] N. Cohen, O. Sharir, and A. Shashua, “On the expressive power of deep
learning: A tensor analysis,” in Conference on Learning Theory, 2016,
pp. 698–728.
[24] S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, and
W. Samek, “On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,” PloS one, vol. 10, no. 7, p.
e0130140, 2015.
[25] W. Samek, A. Binder, G. Montavon, S. Lapuschkin, and K.-R. Müller,
“Evaluating the visualization of what a deep neural network has learned,”
IEEE transactions on neural networks and learning systems, 2017.
[26] C.-C. J. Kuo, “Understanding convolutional neural networks with a
mathematical model,” Journal of Visual Communication and Image
Representation, vol. 41, pp. 406–413, 2016.
[27] ——, “The cnn as a guided multilayer recos transform [lecture notes],”
IEEE Signal Processing Magazine, vol. 34, no. 3, pp. 81–89, 2017.
[28] T. Kobayashi and S. Imai, “Spectral analysis using generalized cepstrum,” IEEE Trans. Acoust., Speech, Signal Proc., vol. 32, no. 5, pp.
1087–1089, 1984.
[29] D. A. Cheveigné and H. Kawahara, “YIN, a fundamental frequency
estimator for speech and music,” J. Acoustical Society of America, vol.
111, no. 4, pp. 1917–1930, 2002.
[30] T. D. Rossing, F. R. Moore, and P. A. Wheeler, The science of sound.
Addison Wesley San Francisco, 2002, vol. 3.
[31] H. Von Helmholtz, “On the sensations of tone (english translation aj
ellis, 1885, 1954),” 1877.
[32] A. De Cheveigné, “Separation of concurrent harmonic sounds: Fundamental frequency estimation and a time-domain cancellation model of
auditory processing,” The Journal of the Acoustical Society of America,
vol. 93, no. 6, pp. 3271–3290, 1993.
[33] J. L. Goldstein, “An optimum processor theory for the central formation
of the pitch of complex tones,” The Journal of the Acoustical Society of
America, vol. 54, no. 6, pp. 1496–1516, 1973.
[34] A. Oppenheim and R. Schafer, Discrete-Time Signal Processing, 3rd ed.
Prentice Hall, 2009.
[35] B. P. Bogert, M. J. Healy, and J. W. Tukey, “The quefrency alanysis
of time series for echoes: Cepstrum, pseudo-autocovariance, crosscepstrum and shape cracking,” in Proceedings of the symposium on time
series analysis, vol. 15, 1963, pp. 209–243.
[36] A. V. Oppenheim and R. W. Schafer, “From frequency to quefrency:
A history of the cepstrum,” IEEE Signal Processing Magazine, vol. 21,
no. 5, pp. 95–106, 2004.
[37] K. Tokuda, T. Kobayashi, T. Masuko, and S. Imai, “Mel-generalized
cepstral analysis: a unified approach to speech spectral estimation.” in
Proc. Int. Conf. Spoken Language Processing, 1994.
[38] T. Tolonen and M. Karjalainen, “A computationally efficient multipitch
analysis model,” IEEE Speech Audio Processing, vol. 8, no. 6, pp. 708–
716, 2000.
[39] S. Kraft and U. Zölzer, “Polyphonic pitch detection by iterative analysis
of the autocorrelation function,” in Proc. Int. Conf. Digital Audio Effects,
2014, pp. 1–8.
[40] H. Indefrey, W. Hess, and G. Seeser, “Design and evaluation of doubletransform pitch determination algorithms with nonlinear distortion in the
frequency domain-preliminary results,” in Proc. IEEE Int. Conf. Acoust.
Speech Signal Process, 1985, pp. 415–418.
[41] A. Klapuri, “Multipitch analysis of polyphonic music and speech signals
using an auditory model,” IEEE Trans. Audio, Speech, Lang. Proc.,
vol. 16, no. 2, pp. 255–266, 2008.
[42] J. S. Lim, “Spectral root homomorphic deconvolution system,” IEEE
Trans. Acoust., Speech, Signal Proc., vol. 27, no. 3, pp. 223–233, 1979.
[43] P. Alexandre and P. Lockwood, “Root cepstral analysis: A unified view.
application to speech processing in car noise environments,” Speech
Communication, vol. 12, no. 3, pp. 277–288, 1993.
[44] L. Rabiner, M. Cheng, A. Rosenberg, and C. McGonegal, “A comparative performance study of several pitch detection algorithms,” IEEE
Trans. Acoust., Speech, Signal Proc., vol. 24, no. 5, pp. 399–418, 1976.
[45] G. Peeters, “Music pitch representation by periodicity measures based
on combined temporal and spectral representations,” in Proc. IEEE Int.
Conf. Acoust. Speech Signal Proc., 2006.
[46] V. Emiya, B. David, and R. Badeau, “A parametric method for pitch
estimation of piano tones,” in Proc. IEEE Int. Conf. Acoust. Speech
Signal Proc., 2007, pp. 249–252.
[47] C.-T. Lee, Y.-H. Yang, and H. H. Chen, “Multipitch estimation of
piano music by exemplar-based sparse representation,” IEEE Trans.
Multimedia, vol. 14, no. 3, pp. 608–618, 2012.
[48] L. Su and Y.-H. Yang, “Resolving octave ambiguities: A cross-dataset
investigation,” in Int. Conf. Computer Music/Sound Music Computing
(ICMC/SMC), 2014, pp. 962–966.
[49] A. P. Klapuri, “Multiple fundamental frequency estimation based on
harmonicity and spectral smoothness,” IEEE Trans. Speech Audio Proc.,
vol. 11, no. 6, pp. 804–816, 2003.
[50] V. Emiya, R. Badeau, and B. David, “Multipitch estimation of piano
sounds using a new probabilistic spectral smoothness principle,” IEEE
Trans. Audio, Speech, Lang. Proc., vol. 18, no. 6, pp. 1643–1654, 2010.
[51] N. Keriven, K. O’Hanlon, and M. D. Plumbley, “Structured sparsity
using backwards elimination for automatic music transcription,” in Proc.
IEEE Int. Works. Machine Learning for Sig. Proc., 2013, pp. 1–6.
[52] K. O’Hanlon, H. Nagano, and M. D. Plumbley, “Structured sparsity for
automatic music transcription,” in Proc. IEEE Int. Conf. Acoust. Speech
Signal Proc., 2012, pp. 441–444.
[53] J. Fritsch, “High quality musical audio source separation,” Master’s
thesis, Queen Mary Centre for Digital Music, 2012.
[54] M. Mauch and S. Ewert, “The audio degradation toolbox and its
application to robustness evaluation.” in Proc. Int. Soc. Music Information Retrieval Conf., 2013, pp. 83–88, https://code.soundsoftware.ac.
uk/projects/audio-degradation-toolbox (v0.2).
[55] S. S. Stevens, “On the psychophysical law.” Psychological review,
vol. 64, no. 3, p. 153, 1957.
| 9cs.NE
|
arXiv:1704.00699v2 [math.DS] 14 Jan 2018
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
CLINTON T. CONLEY, STEVE C. JACKSON, DAVID KERR, ANDREW S. MARKS,
BRANDON SEWARD, AND ROBIN D. TUCKER-DROB
Abstract. We show that every probability-measure-preserving action of a countable
amenable group G can be tiled, modulo a null set, using finitely many finite subsets
of G (“shapes”) with prescribed approximate invariance so that the collection of tiling
centers for each shape is Borel. This is a dynamical version of the Downarowicz–Huczek–
Zhang tiling theorem for countable amenable groups and strengthens the Ornstein–Weiss
Rokhlin lemma. As an application we prove that, for every countably infinite amenable
group G, the crossed product of a generic free minimal action of G on the Cantor set is
Z-stable.
1. Introduction
A discrete group G is said to be amenable if it admits a finitely additive probability
measure which is invariant under the action of G on itself by left translation, or equivalently
if there exists a unital positive linear functional ℓ∞ (G) → C which is invariant under
the action of G on ℓ∞ (G) induced by left translation (such a functional is called a left
invariant mean). This definition was introduced by von Neumann in connection with the
Banach–Tarski paradox and shown by Tarski to be equivalent to the absence of paradoxical
decompositions of the group. Amenability has come to be most usefully leveraged through
its combinatorial expression as the Følner property, which asks that for every finite set
K ⊆ G and δ > 0 there exists a nonempty finite set F ⊆ G which is (K, δ)-invariant in
the sense that |KF ∆F | < δ|F |.
The concept of amenability appears as a common thread throughout much of ergodic
theory as well as the related subject of operator algebras, where it is known via a number
of avatars like injectivity, hyperfiniteness, and nuclearity. It forms the cornerstone of
the theory of orbit equivalence, and also underpins both Kolmogorov–Sinai entropy and
the classical ergodic theorems, whether explicitly in their most general formulations or
implicitly in the original setting of single transformations (see Chapters 4 and 9 of [11]).
A key tool in applying amenability to dynamics is the Rokhlin lemma of Ornstein and
Weiss, which in one of its simpler forms says that for every free probability-measurepreserving action G y (X, µ) of a countably infinite amenable group and every finite set
K ⊆ G and δ > 0 there exist (K, δ)-invariant finite sets T1 , . . . , Tn ⊆ G and measurable
sets A1 , . . . , An ⊆ X such that the sets sAi for i = 1, . . . , n and s ∈ Ti are pairwise disjoint
and have union of measure at least 1 − δ [17].
The proportionality in terms of which approximate invariance is expressed in the Følner
condition makes it clear that amenability is a measure-theoretic property, and it is not
Date: November 9, 2017.
1
2
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
surprising that the most influential and definitive applications of these ideas in dynamics (e.g., the Connes–Feldman–Weiss theorem) occur in the presence of an invariant or
quasi-invariant measure. Nevertheless, amenability also has significant ramifications for
topological dynamics, for instance in guaranteeing the existence of invariant probability
measures when the space is compact and in providing the basis for the theory of topological
entropy. In the realm of operator algebras, similar comments can be made concerning the
relative significance of amenability for von Neumann algebras (measure) and C∗ -algebras
(topology).
While the subjects of von Neumann algebras and C∗ -algebras have long enjoyed a symbiotic relationship sustained in large part through the lens of analogy, and a similar relationship has historically bound together ergodic theory and topological dynamics, the last
few years have witnessed the emergence of a new and structurally more direct kind of rapport between topology and measure in these domains, beginning on the operator algebra
side with the groundbreaking work of Matui and Sato on strict comparison, Z-stability,
and decomposition rank [14, 15]. On the side of groups and dynamics, Downarowicz,
Huczek, and Zhang recently showed that if G is a countable amenable group then for
every finite set K ⊆ G and δ > 0 one can partition (or “tile”) G by left translates of
finitely many (K, δ)-invariant finite sets [3]. The consequences that they derive from this
tileability are topological and include the existence, for every such G, of a free minimal
action with zero entropy. One of the aims of the present paper is to provide some insight
into how these advances in operator algebras and dynamics, while seemingly unrelated
at first glance, actually fit together as part of a common circle of ideas that we expect,
among other things, to lead to further progress in the structure and classification theory
of crossed product C∗ -algebras.
Our main theorem is a version of the Downarowicz–Huczek–Zhang tiling result for
free p.m.p. (probability-measure-preserving) actions of countable amenable groups which
strengthens the Ornstein–Weiss Rokhlin lemma in the form recalled above by shrinking
the leftover piece down to a null set (Theorem 3.6). As in the case of groups, one does
not expect the utility of this dynamical tileability to be found in the measure setting,
where the Ornstein–Weiss machinery generally suffices, but rather in the derivation of
topological consequences. Indeed we will apply our tiling result to show that, for every
countably infinite amenable group G, the crossed product C(X) ⋊ G of a generic free
minimal action G y X on the Cantor set possesses the regularity property of Z-stability
(Theorem 5.4). The strategy is to first prove that such an action admits clopen tower
decompositions with arbitrarily good Følner shapes (Theorem 4.2), and then to demonstrate that the existence of such tower decompositions implies that the crossed product is
Z-stable (Theorem 5.3). The significance of Z-stability within the classification program
for simple separable nuclear C∗ -algebras is explained at the beginning of Section 5.
It is a curious irony in the theory of amenability that the Hall–Rado matching theorem
can be used not only to show that the failure of the Følner property for a discrete group
implies the formally stronger Tarski characterization of nonamenability in terms of the
existence of paradoxical decompositions [2] but also to show, in the opposite direction,
that the Følner property itself implies the formally stronger Downarowicz–Huczek–Zhang
characterization of amenability which guarantees the existence of tilings of the group
by translates of finitely many Følner sets [3]. This Janus-like scenario will be reprised
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
3
here in the dynamical context through the use of a measurable matching argument of
Lyons and Nazarov that was originally developed to prove that for every simple bipartite
nonamenable Cayley graph of a discrete group G there is a factor of a Bernoulli action of
G which is an a.e. perfect matching of the graph [13]. Accordingly the basic scheme for
proving Theorem 3.6 will be the same as that of Downarowicz, Huczek, and Zhang and
divides into two parts:
(i) using an Ornstein–Weiss-type argument to show that a subset of the space of
lower Banach density close to one can be tiled by dynamical translates of Følner
sets, and
(ii) using a Lyons–Nazarov-type measurable matching to distribute almost all remaining points to existing tiles with only a small proportional increase in the size of
the Følner sets, so that the approximate invariance is preserved.
We begin in Section 2 with the measurable matching result (Lemma 2.6), which is a
variation on the Lyons–Nazarov theorem from [13] and is established along similar lines.
In Section 3 we establish the appropriate variant of the Ornstein–Weiss Rokhlin lemma
(Lemma 3.4) and put everything together in Theorem 3.6. Section 4 contains the genericity
result for free minimal actions on the Cantor set, while Section 5 is devoted to the material
on Z-stability.
Acknowledgements. C.C. was partially supported by NSF grant DMS-1500906. D.K. was
partially supported by NSF grant DMS-1500593. Part of this work was carried out while
he was visiting the Erwin Schrödinger Institute (January–February 2016) and the Mittag–
Leffler Institute (February–March 2016). A.M. was partially supported by NSF grant
DMS-1500974. B.S. was partially supported by ERC grant 306494. R.T.D. was partially
supported by NSF grant DMS-1600904. Part of this work was carried out during the AIM
SQuaRE: Measurable Graph Theory.
2. Measurable matchings
Given sets X and Y and a subset R ⊆ X × Y , with each x ∈ X we associate its vertical
section Rx = {y ∈ Y : (x, y) ∈ R} and with each y ∈ Y we associate S
its horizontal section
Ry = {x ∈ X : (x, y) ∈ R}. Analogously, for A ⊆ X we put RA = x∈A Rx = {y ∈ Y :
∃x ∈ A (x, y) ∈ R}. We say that R is locally finite if for all x ∈ X and y ∈ Y the sets Rx
and Ry are finite.
If now X and Y are standard Borel spaces equipped with respective Borel measures
µ and ν, we say that R ⊆ X × Y is (µ, ν)-preserving if whenever f : A → B is a Borel
bijection between subsets A ⊆ X and B ⊆ Y with graph(f ) ⊆ R we have µ(A) = ν(B).
We say that R is expansive if there is some c > 1 such that for all Borel A ⊆ X we have
ν(RA ) ≥ cµ(A).
We use the notation f : X ⇀ Y to denote a partial function from X to Y . We say that
such a partial function f is compatible with R ⊆ X × Y if graph(f ) ⊆ R.
Proposition 2.1 (ess. Lyons–Nazarov [13, Theorem 1.1]). Suppose that X and Y are
standard Borel spaces, that µ is a Borel probability measure on X, and that ν is a Borel
measure on Y . Suppose that R ⊆ X × Y is Borel, locally finite, (µ, ν)-preserving, and
4
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
expansive. Then there is a µ-conull X ′ ⊆ X and a Borel injection f : X ′ → Y compatible
with R.
Proof. Fix a constant of expansivity c > 1 for R.
We construct a sequence (fn )n∈N of Borel partial injections from X to Y which are
compatible with R. Moreover, we will guarantee that the set X ′ = {x ∈ X : ∃m ∈ N ∀n ≥
m x ∈ dom(fn ) and fn (x) = fm (x)} is µ-conull, establishing that the limiting function
satisfies the conclusion of the lemma.
Given a Borel partial injection g : X ⇀ Y we say that a sequence (x0 , y0 , . . . , xn , yn ) ∈
X × Y × · · · × X × Y is a g-augmenting path if
• x0 ∈ X is not in the domain of g,
• for all distinct i, j < n, yi 6= yj ,
• for all i < n, (xi , yi ) ∈ R,
• for all i < n, yi = g(xi+1 ),
• yn ∈ Y is not in the image of g.
We call n the length of such a g-augmenting path and x0 the origin of the path. Note
that the sequence (x0 , y0 , y1 , . . . , yn ) in fact determines the entire g-augmenting path, and
moreover that xi 6= xj for distinct i, j < n.
In order to proceed we require a few lemmas.
Lemma 2.2. Suppose that n ∈ N and g : X ⇀ Y is a Borel partial injection compatible
with R admitting no augmenting paths of length less than n. Then µ(X \ dom(g)) ≤ c−n .
Proof. Put A0 = X \ dom(g). Define recursively for i < n sets Bi = RAi and Ai+1 =
Ai ∪ g −1 (Bi ). Note that the assumption that there are no augmenting paths of length
less than n implies that each Bi is contained in the image of g. Expansivity of R yields
ν(Bi ) ≥ cµ(Ai ) and (µ, ν)-preservation of R then implies that µ(Ai+1 ) ≥ ν(Bi ) ≥ cµ(Ai ).
Consequently, 1 ≥ µ(An ) ≥ cn µ(A0 ), and hence µ(A0 ) ≤ c−n .
We say that a graph G on a standard Borel space X has a Borel N-coloring if there is
a Borel function c : X → N such that if x and y are G-adjacent then c(x) 6= c(y).
Lemma 2.3 (Kechris–Solecki–Todorcevic [12, Proposition 4.5]). Every locally finite Borel
graph on a standard Borel space has a Borel N-coloring.
Proof. Fix a countable algebra {Bn : n ∈ N} of Borel sets which separates points (for
example, the algebra generated by the basic open sets of a compatible Polish topology),
and color each vertex x by the least n ∈ N such that Bn contains x and none of its
neighbors.
Analogously, for k ∈ N, we say that a graph on a standard Borel X has a Borel kcoloring if there is a Borel function c : X → {1, . . . , k} giving adjacent points distinct
colors.
Lemma 2.4 (Kechris–Solecki–Todorcevic [12, Proposition 4.6]). If a Borel graph on a
standard Borel X has degree bounded by d ∈ N, then it has a Borel (d + 1)-coloring.
Proof. By Lemma 2.3, the graph has a Borel N-coloring c : X → N. We recursively
build sets An for n ∈ N by A0 = {x ∈ X : c(x) = 0} and An+1 = An ∪ {x ∈ X :
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
5
S
c(x) = n + 1 and no neighbor of x is in An }. Then A = n An is a Borel set which is
G-independent, and moreover is maximal with this property. So the restriction of G to
X \ A has degree less than d, and the result follows by induction.
Lemma 2.5 (ess. Elek–Lippner [5, Proposition 1.1]). Suppose that g : X ⇀ Y is a Borel
partial injection compatible with R, and let n ≥ 1. Then there is a Borel partial injection
g′ : X ⇀ Y compatible with R such that
• dom(g ′ ) ⊇ dom(g),
• g′ admits no augmenting paths of length less than n,
• µ({x ∈ X : g ′ (x) 6= g(x)} ≤ nµ(X \ dom(g)).
Proof. Consider the set Z of injective sequences (x0 , y0 , x1 , y1 , . . . , xm , ym ), where m < n,
such that for all i ≤ m we have (xi , yi ) ∈ R and for all i < m we have (xi+1 , yi ) ∈ R. Equip
Z with the standard Borel structure it inherits as a Borel subset of (X × Y )≤n . Consider
also the locally finite Borel graph G on Z rendering adjacent two distinct
sequences in
F
Z if they share any entries. By Lemma 2.3 there is a partition Z = k∈N Zk of Z into
Borel sets such that for all k, no two elements of Zk are G-adjacent. In other words, we
partition potential augmenting paths into countably many colors, where no two paths of
the same color intersect. Thus we may flip paths of the same color simultaneously without
risk of causing conflicts. Towards that end, fix a bookkeeping function s : N → N such
that s−1 (k) is infinite for all k ∈ N in order to consider each color class infinitely often.
Given a g-augmenting path z = (x0 , y0 , . . . , xm , ym ), define the flip along z to be the
Borel partial function gz : X ⇀ Y given by
(
yi
if ∃i ≤ m x = xi ,
gz (x) =
g(x) otherwise.
The fact that z is g-augmenting ensures that gz is injective. More generally, for any Borel
G-independent set Z aug ⊆ Z of g-augmenting paths, we may simultaneously flip g along
all paths in Z aug to obtain another Borel partial injection (g)Z aug .
We iterate this construction. Put g0 = g. Recursively assuming that gk : X ⇀ Y has
been defined, let Zkaug be the set of gk -augmenting paths in Zs(k) , and let gk+1 = (gk )Z aug
k
be the result of flipping gk along all paths in Zkaug . As each x ∈ X is contained in only
finitely many elements of Z, and since each path in Z can be flipped at most once (after the
first flip its origin is always in the domain of the subsequent partial injections), it follows
that the sequence (gk (x))k∈N is eventually constant. Defining g′ (x) to be the limiting
value, it is routine to check that there are no g′ -augmenting paths of length less than n.
Finally, to verify the third item of the lemma, put A = {x ∈ X : g′ (x) 6= g(x)}.
With each x ∈ A associate the origin of the first augmenting path along which it was
flipped. This is an at most n-to-1 Borel function from A to X \ dom(g), and since R is
(µ, ν)-preserving the bound follows.
We are now in position to follow the strategy outlined at the beginning of the proof.
Let f0 : X ⇀ Y be the empty function. Recursively assuming the Borel partial injection
fn : X ⇀ Y has been defined to have no augmenting paths of length less than n, let fn+1
be the Borel partial injection (fn )′ granted by applying Lemma 2.5 to fn . Thus fn+1 has
no augmenting paths of length less than n + 1 and the recursive construction continues.
6
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
Lemma 2.2 ensures that µ(X \ dom(fn )) ≤ c−n , and thus the third item of Lemma 2.5
ensures that µ({x ∈ X : fn+1 (x) 6= fn (x)}) ≤ (n + 1)c−n . As the sequence (n + 1)c−n is
summable, the Borel–Cantelli lemma implies that X ′ = {x ∈ X : ∃m ∈ N ∀n ≥ m x ∈
dom(fn ) and fn (x) = fm (x)} is µ-conull. Finally, f = limn→∞ fn ↾ X ′ is as desired.
Lemma 2.6. Suppose X and Y are standard Borel spaces, that µ is a Borel measure on
X, and that ν is a Borel measure on Y . Suppose R ⊆ X × Y is Borel, locally finite,
(µ, ν)-preserving graph. Assume that there exist numbers a, b > 0 such that |Rx | ≥ a for
µ-a.e. x ∈ X and |Ry | ≤ b for ν-a.e. y ∈ Y . Then ν(RA ) ≥ ab µ(A) for all Borel subsets
A ⊆ X.
R
R
Proof. Since R is (µ, ν)-preserving we have A |Rx | dµ = RA |Ry ∩ A| dν. Hence
Z
Z
Z
Z
y
b dν = bν(RA ).
|R ∩ A| dν ≤
|Rx | dµ =
a dµ ≤
aµ(A) =
A
A
RA
RA
3. Følner tilings
Fix a countable group G. For finite sets K, F ⊆ G and δ > 0, we say that F is (K, δ)invariant if |KF △F | < δ|F |. Note this condition implies |KF | < (1 + δ)|F |. Recall that
G is amenable if for every finite K ⊆ G and δ > 0 there exists a (K, δ)-invariant set F . A
Følner sequence is a sequence of finite sets Fn ⊆ G with the property that for every finite
K ⊆ G and δ > 0 the set Fn is (K, δ)-invariant for all but finitely many n. Below, we
always assume that G is amenable.
Fix a free action G y X. For A ⊆ X we define the lower and upper Banach densities
of A to be
|A ∩ F x|
|A ∩ F x|
D (A) = sup inf
and
D̄(A) = inf sup
.
F ⊆G x∈X
¯
|F |
|F |
F ⊆G x∈X
F finite
F finite
Equivalently [3, Lemma 2.9], if (Fn )n∈N is a Følner sequence then
D (A) = lim inf
n→∞ x∈X
¯
|A ∩ Fn x|
|Fn |
and
D̄(A) = lim sup
n→∞ x∈X
|A ∩ Fn x|
.
|Fn |
We now define an analogue of ‘(K, δ)-invariant’ for infinite subsets of X. A set A ⊆ X
(possibly infinite) is (K, δ)∗ -invariant if there is a finite set F ⊆ G such that |(KA△A) ∩
F x| < δ|A ∩ F x| for all x ∈ X. Equivalently, A is (K, δ)∗ -invariant if and only if for every
Følner sequence (Fn )n∈N we have limn supx |(KA△A) ∩ Fn x|/|A ∩ Fn x| < δ.
A collection {Fi : i ∈ I} of finite subsets of X is called ǫ-disjoint if for each i there is
an Fi′ ⊆ Fi such that |Fi′ | > (1 − ǫ)|Fi | and such that the sets {Fi′ : i ∈ I} are pairwise
disjoint.
Lemma 3.1. Let K, W ⊆ G be finite, let ǫ, δ > 0, let C ⊆ X, and for c ∈ CS let Fc ⊆ W
be (K, δ(1 − ǫ))-invariant. If the collection
{Fc c : c ∈ C} is ǫ-disjoint and c∈C Fc c has
S
positive lower Banach density, then c∈C Fc c is (K, δ)∗ -invariant.
S
Proof. Set A = c∈C Fc c and set T = W W −1 ({1G } ∪ K)−1 . Since W is finite and each
Fc ⊆ W , there is 0 < δ0 < δ such that each Fc is (K, δ0 (1 − ǫ))-invariant. Fix a finite set
D (A)
(A)
U ⊆ G which is (T, D¯2|T
| (δ − δ0 ))-invariant and satisfies inf x∈X |A ∩ U x| > ¯ 2 |U |. Now
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
7
fix x ∈ X. Let B be the set of b ∈ U x such that T b 6⊆ U x. Note that B ⊆ T −1 (T U x△U x)
and thus
|U |
|T U △U |
·
· |A ∩ U x| < (δ − δ0 )|A ∩ U x|.
|B| ≤ |T | ·
|U |
|A ∩ U x|
P
Set C ′ = {c ∈ C : Fc c ⊆ U x}. Note that the ǫ-disjoint assumption gives (1−ǫ) c∈C ′ |Fc | ≤
|A∩U x|. Also, our definitions of C ′ , T , and B imply that if c ∈ C \C ′ and
S ({1G }∪K)Fc c∩
U x 6= ∅ then (({1G }∪K)Fc c)∩U x ⊆ B. Therefore (KA△A)∩U x ⊆ B∪ c∈C ′ (KFc c△Fc c).
Combining this with the fact that each set Fc is (K, δ0 (1 − ǫ))-invariant, we obtain
X
|(KA△A) ∩ U x| ≤ |B| +
|KFc c△Fc c|
c∈C ′
< (δ − δ0 )|A ∩ U x| +
X
δ0 (1 − ǫ)|Fc |
c∈C ′
≤ (δ − δ0 )|A ∩ U x| + δ0 |A ∩ U x|
= δ|A ∩ U x|.
Since x was arbitrary, we conclude that A is (K, δ)∗ -invariant.
Lemma 3.2. Let T ⊆ G be finite and let ǫ, δ > 0 with ǫ(1 + δ) < 1. Suppose that A ⊆ X
is (T −1 , δ)∗ -invariant. If B ⊇ A and |B ∩ T x| ≥ ǫ|T | for all x ∈ X, then
D (B) ≥ (1 − ǫ(1 + δ)) · D (A) + ǫ.
¯
¯
Proof. This is implicitly demonstrated in [3, Proof of Lemma 4.1]. As a convenience to
the reader, we include a proof here. Fix θ > 0. Since A is (T −1 , δ)∗ -invariant, we can pick
a finite set U ⊆ G which is (T, θ)-invariant and satisfies
inf
x∈X
|A ∩ U x|
> D(A) − θ
¯
|U |
and
sup
x∈X
|T −1 A ∩ U x|
< 1 + δ.
|A ∩ U x|
|A∩U x|
|U |
> D (A) − θ, and set U ′ = {u ∈ U : A ∩ T ux = ∅}. Notice that
¯
|U | − |T −1 A ∩ U x|
|T −1 A ∩ U x| |A ∩ U x|
|U ′ |
=
=1−
·
> 1 − (1 + δ)α.
|U |
|U |
|A ∩ U x|
|U |
Fix x ∈ X, set α =
Since A ∩ T U ′ x = ∅ and |B ∩ T y| ≥ ǫ|T | for all y ∈ X, it follows that |(B \ A) ∩ T ux| ≥ ǫ|T |
for all u ∈ U ′ . Thus there are ǫ|T ||U ′ | many pairs (t, u) ∈ T × U ′ with tux ∈ B \ A. It
follows there is t∗ ∈ T with |(B \ A) ∩ t∗ U ′ x| ≥ ǫ · |U ′ |. Therefore
|A ∩ U x| |(B \ A) ∩ t∗ U ′ x| |U ′ |
|U |
|B ∩ T U x|
≥
+
·
·
′
|T U |
|U |
|U |
|U |
|T U |
−1
> α + ǫ(1 − (1 + δ)α) · (1 + θ)
= (1 − ǫ(1 + δ))α + ǫ · (1 + θ)−1
> (1 − ǫ(1 + δ))(D (A) − θ) + ǫ · (1 + θ)−1 .
¯
Letting θ tend to 0 completes the proof.
8
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
Lemma 3.3. Let X be a standard Borel space and let G y X be a free Borel action. Let
Y ⊆ X be Borel, let T ⊆ G be finite, and let ǫ ∈ (0, 1/2). Then there is a Borel set C ⊆ X
and a Borel function c ∈ C 7→ Tc ⊆ T suchSthat |Tc | > (1− ǫ)|T |, the sets {Tc c : c ∈ C} are
pairwise disjoint and disjoint with Y , Y ∪ c∈C Tc c = Y ∪ T C, and |(Y ∪ T C) ∩ T x| ≥ ǫ|T |
for all x ∈ X.
Proof. Using Lemma 2.4, fix a Borel partition P = {P1 , . . . , Pm } of X such that T x∩T x′ =
∅ for S
all x 6= x′ ∈ Pi and all 1 ≤ i ≤ m. We will pick Borel sets Ci ⊆ Pi and set
C = 1≤i≤m Ci . Set Y0 = Y . Let 1 ≤ i ≤ m and inductively assume that Yi−1 has
been defined. Define Ci = {c ∈ Pi : |Yi−1 ∩ T c| < ǫ|T |}, defineSYi = Yi−1 ∪ T Ci , and for
c ∈ Ci set Tc = {t ∈ T : tc 6∈ Yi−1 }. It is easily seen that C = 1≤i≤m Ci has the desired
properties.
The following lemma is mainly due to Ornstein–Weiss [17], who proved it with an
invariant probability measure taking the place of Banach density. Ornstein and Weiss also
established a purely group-theoretic counterpart of this result which was later adapted to
the Banach density setting by Downarowicz–Huczek–Zhang in [3] and will be heavily used
in Section 5, where it is recorded as Theorem 5.2. The only difference between this lemma
and prior versions is that we simultaneously work in the Borel setting and use Banach
density.
Lemma 3.4. [17, II.§2. Theorem 5] [3, Lemma 4.1] Let X be a standard Borel space and
let G y X be a free Borel action. Let K ⊆ G be finite, let ǫ ∈ (0, 1/2), and let n satisfy
(1 − ǫ)n < ǫ. Then there exist (K, ǫ)-invariant sets F1 , . . . , Fn , a Borel set C ⊆ X, and a
Borel function c ∈ C 7→ Fc ⊆ G such that:
(i) for every c ∈ C there is 1 ≤ i ≤ n with Fc ⊆ Fi and |Fc | > (1 − ǫ)|Fi |;
(ii) theSsets Fc c, c ∈ C, are pairwise disjoint; and
(iii) D ( c∈C Fc c) > 1 − ǫ.
¯
Proof. Fix δ > 0 satisfying (1 + δ)−1 (1 − (1 + δ)ǫ)n < ǫ − 1 + (1 + δ)−1 . Fix a sequence
of (K, ǫ)-invariant sets F1 , . . . , Fn such that Fi is (Fj−1 , δ(1 − ǫ))-invariant for all 1 ≤ j <
i ≤ n.
The set C will be the disjoint union of sets Ci , 1 ≤ i ≤ n. The construction
S
S will be
such that Fc ⊆ Fi and |Fc | > (1 − ǫ)|Fi | for c ∈ Ci . We will define Ai = i≤k≤n c∈Ck Fc c
S
and arrange that Ai+1 ∪ Fi Ci = Ai+1 ∪ c∈Ci Fc c and
D (Ai ) ≥ (1 + δ)−1 − (1 + δ)−1 (1 − ǫ(1 + δ))n+1−i .
¯
S
In particular, we will have Ai = i≤k≤n Fk Ck .
To begin, apply Lemma 3.3 with Y = ∅ and T = Fn to get a Borel set Cn and a Borel
map c ∈ S
Cn 7→ Fc ⊆ Fn such that |Fc | > (1 − ǫ)|Fn |, the sets {Fc c : c ∈ Cn } are pairwise
disjoint, c∈Cn Fc c = Fn Cn , and |Fn Cn ∩ Fn x| ≥ ǫ|Fn | for all x ∈ X. Applying Lemma
3.2 with A = ∅ and B = Fn Cn we find that the set An = Fn Cn satisfies D (An ) ≥ ǫ.
¯
Inductively assume that Cn through Ci+1 have been defined and An through Ai+1 are
defined as above and satisfy (3.1). Using Y = Ai+1 and T = Fi , apply Lemma 3.3 to get
a Borel set Ci and a Borel map c ∈ Ci 7→ Fc ⊆ Fi such that |FcS| > (1 − ǫ)|Fi |, the sets
{Fc c : c ∈ Ci } are pairwise disjoint and disjoint with Ai+1 , Ai+1 ∪ c∈Ci Fc c = Ai+1 ∪ Fi Ci ,
and |(Ai+1 ∪ Fi Ci ) ∩ Fi x| ≥ ǫ|Fi | for all x ∈ X. The set Ai+1 is the union of an ǫ-disjoint
(3.1)
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
9
collection of (Fi−1 , δ(1 − ǫ))-invariant sets and has positive lower Banach density. So by
Lemma 3.1 Ai+1 is (Fi−1 , δ)∗ -invariant. Applying Lemma 3.2 with A = Ai+1 , we find that
Ai = Ai+1 ∪ Fi Ci satisfies
D (Ai ) ≥ (1 − ǫ(1 + δ)) · D (Ai+1 ) + ǫ
¯
¯
ǫ(1 + δ)
(1 − ǫ(1 + δ))
− (1 + δ)−1 (1 − ǫ(1 + δ))n+1−i +
≥
(1 + δ)
1+δ
= (1 + δ)−1 − (1 + δ)−1 (1 − ǫ(1 + δ))n+1−i .
This completes the inductive step and completes the definition of C. It is immediate from
the construction that (i) and (ii) are satisfied. Clause (iii) also follows by noting that (3.1)
is greater than 1 − ǫ when i = 1.
We recall the following simple fact.
Lemma 3.5. [3, Lemma 2.3] If F ⊆ G is (K, δ)-invariant and F ′ satisfies |F ′ △F | < ǫ|F |
then F ′ is (K, (|K|+1)ǫ+δ
)-invariant.
1−ǫ
Now we present the main theorem.
Theorem 3.6. Let G be a countable amenable group, let (X, µ) be a standard probability
space, and let G y (X, µ) be a free p.m.p. action. For every finite K ⊆ G and every δ > 0
there exist a µ-conull G-invariant Borel set X ′ ⊆ X, a collection {Ci : 0 ≤ i ≤ m} of
Borel subsets of X ′ , and a collection {Fi : 0 ≤ i ≤ m} of (K, δ)-invariant sets such that
{Fi c : 0 ≤ i ≤ m, c ∈ Ci } partitions X ′ .
< δ. Apply Lemma 3.4 to get (K, ǫ)-invariant
Proof. Fix ǫ ∈ (0, 1/2) satisfying (|K|+1)6ǫ+ǫ
1−6ǫ
′
′
sets F1 , . . . , Fn , a Borel set C ⊆ X, and a Borel function c ∈ C 7→ Fc ⊆ G satisfying
(i) for every c ∈ C there is 1 ≤ i ≤ n with Fc ⊆ Fi′ and |Fc | > (1 − ǫ)|Fi′ |;
(ii) theSsets Fc c, c ∈ C, are pairwise disjoint; and
(iii) D ( c∈C Fc c) > 1 − ǫ.
¯
S
Set Y = X \ c∈C Fc c. If µ(Y ) = 0 then we are done. So we assume µ(Y ) > 0 and
we let ν denote the restriction of µ to Y . Fix a Borel map c ∈ C 7→ Zc ⊆ Fc satisfying
4ǫ|Fc | < |Zc | < 5ǫ|Fc | for all c ∈ C (it’s clear from the proof of Lemma
S 3.4 that we may
choose the sets Fi′ so that ǫ|Fc | > mini ǫ(1 − ǫ)|Fi′ | > 1). Set Z = c∈C Zc c and let ζ
denote the restriction
of µ to Z (note that µ(Z) > 0).
S
Set W = ni=1 Fi′ and W ′ = W W −1 . Fix a finite set U ⊆ G which is (W ′ , (1/2−ǫ)/|W ′ |)invariant and satisfies inf x∈X |(X \ Y ) ∩ U x| > (1 − ǫ)|U |. Since every amenable group
admits a Følner sequence consisting of symmetric sets, we may assume that U = U −1 [16,
Corollary 5.3]. Define R ⊆ Y ×Z by declaring (y, z) ∈ R if and only if y ∈ U z (equivalently
z ∈ U y). Then R is Borel, locally finite, and (ν, ζ)-preserving. We now check that R is
expansive. We automatically have |Rz | = |Y ∩ U z| < ǫ|U | for all z ∈ Z. By Lemma 2.6 it
suffices to show that |Ry | = |Z ∩ U y| ≥ 2ǫ|U | for all y ∈ Y . Fix y ∈ Y . Let B be the set
of b ∈ U y such that W ′ b 6⊆ U y. Then B ⊆ W ′ (W ′ U y△U y) and thus
|W ′ U △U |
|B|
≤ |W ′ | ·
< 1/2 − ǫ.
|U |
|U |
10
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
Let A be the union of those sets Fc c, c ∈ C, which are contained in U y. Notice that
(X \ Y ) ∩ U y ⊆ B ∪ A. Therefore
|A|
|(B ∪ A) ∩ U y|
|(X \ Y ) ∩ U y|
1
−ǫ+
>
≥
> 1 − ǫ,
2
|U |
|U |
|U |
hence |A| > |U |/2. By construction |Z ∩ A| > 4ǫ|A|. So |Z ∩ U y| ≥ |Z ∩ A| > 2ǫ|U |. We
conclude that R is expansive.
Apply Proposition 2.1 to obtain a G-invariant µ-conull set X ′ ⊆ X and a Borel injection
ρ : Y ∩X ′ → Z with graph(ρ) ⊆ R. Consider the sets Fc ∪{g ∈ U : gc ∈ Y and ρ(gc) ∈ Fc c}
as c ∈ C varies. These are subsets of W ∪ U and thus there are only finitely many such sets
which we can enumerate as F1 , . . . , Fm . We partition C ∩ X ′ into Borel sets C1 , . . . , Cm
with c ∈ Ci if and only if c ∈ X ′ and Fc ∪ {g ∈ U : gc ∈ Y and ρ(gc) ∈ Fc c} = Fi . Since
ρ is defined on all of Y ∩ X ′ , we see that the sets {Fi c : 1 ≤ i ≤ m, c ∈ Ci } partition X ′ .
Finally, for c ∈ Ci ∩ X ′ , if we let Fj′ be such that |Fc △Fj′ | < ǫ|Fj′ |, then
|Fi △Fj′ | ≤ |Fi △Fc | + |Fc △Fj′ | ≤ |ρ−1 (Fc c)| + ǫ|Fj′ |
≤ |Zc | + ǫ|Fj′ | < 5ǫ|Fc | + ǫ|Fj′ | ≤ 6ǫ|Fj′ |.
Using Lemma 3.5 and our choice of ǫ, this implies that each set Fi is (K, δ)-invariant.
4. Clopen tower decompositions with Følner shapes
Let G y X be an action of a group on a compact space. By a clopen tower we mean
a pair (B, S) where B is a clopen subset of X (the base of the tower) and S is a finite
subset of G (the shape of the tower) such that the sets sB for s ∈ S are pairwise disjoint.
By a clopen tower decomposition of X we mean a finite collection {(Bi , Si )}ni=1 of clopen
towers such that the sets S1 B1 , . . . , Sn Bn form a partition of X. We also similarly speak
of measurable towers and measurable tower decompositions for an action G y (X, µ) on a
measure space, with the bases now being measurable sets instead of clopen sets. In this
terminology, Theorem 3.6 says that if G y (X, µ) is a free p.m.p. action of a countable
amenable group on a standard probability space then for every finite set K ⊆ G and
δ > 0 there exists, modulo a null set, a measurable tower decomposition of X with (K, δ)invariant shapes.
Lemma 4.1. Let G be a countably infinite amenable group and G y X a free minimal
action on the Cantor set. Then this action has a free minimal extension G y Y on
the Cantor set such that for every finite set F ⊆ G and δ > 0 there is a clopen tower
decomposition of Y with (F, δ)-invariant shapes.
Proof. Let F1 ⊆ F2 ⊆ . . . be an increasing sequence of finite subsets of G whose union is
equal to G. Fix a G-invariant Borel probability measure µ on X (such a measure exists
by amenability). The freeness of the action G y X means that for each n ∈ N we can
apply Theorem 3.6 to produce, modulo a null set, a measurable tower decomposition Un
for the p.m.p. action G y (X, µ) such that each shape is (Fn , 1/n)-invariant. Let A be the
unital G-invariant C∗ -algebra of L∞ (X, µ) generated by C(X) and the indicator functions
of the levels of each of the tower decompositions Un . Since there are countably many
such indicator functions and the group G is countable, the C∗ -algebra A is separable.
Therefore by the Gelfand–Naimark theorem we have A = C(Z) for some zero-dimensional
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
11
metrizable space Z and a G-factor map ϕ : Z → X. By a standard fact which can be
established using Zorn’s lemma, there exists a nonempty closed G-invariant set Y ⊆ Z
such that the restriction action G y Y is minimal. Note that Y is necessarily a Cantor
set, since G is infinite. Also, the action G y Y is free, since it is an extension of a free
action. Since the action on X is minimal, the restriction ϕ|Y : Y → X is surjective and
hence a G-factor map. For each n we get from Un a clopen tower decomposition Vn of
Y with (Fn , 1/n)-invariant shapes, and by intersecting the levels of the towers in Vn with
Y we obtain a clopen tower decomposition of Y with (Fn , 1/n)-invariant shapes, showing
that the extension G y Y has the desired property.
Let X be the Cantor set and let G be a countable infinite amenable group. The set
Act(G, X) is a Polish space under the topology which has as a basis the sets
Uα,P,F = {β ∈ Act(G, X) : αs A = βs A for all A ∈ P and s ∈ F }
where α ∈ Act(G, X), P is a clopen partition of X, and F is a finite subset of G.
Write FrMin(G, X) for the set of actions in Act(G, X) which are free and minimal. Then
FrMin(G, X) is a Gδ set. To see this, fix an enumeration s1 , s2 , s3 , . . . of G \ {e} (where
e denotes the identity element of the group) and for every n ∈ N and
S nonempty clopen
set A ⊆ X define the set Wn,A of all α ∈ Act(G, X) such that (i) s∈F αs A = X for
some finite set F ⊆ G, and (ii) there exists a clopen partition {A1 , . . . , Ak } of A such that
αsn Ai ∩ Ai = ∅ for all i = 1, . . . , k. Then each Wn,A is open, which means, with A ranging
T over
T the countable collection of nonempty clopen subsets of X, that the intersection
n∈N A Wn,A , which is equal to FrMin(G, X), is a Gδ set. It follows that FrMin(G, X)
is a Polish space.
Theorem 4.2. Let G be a countably infinite amenable group. Let C be the collection of
actions in FrMin(G, X) with the property that for every finite set F ⊆ G and δ > 0 there
is a clopen tower decomposition of X with (F, δ)-invariant shapes. Then C is a dense Gδ
subset of FrMin(G, X).
α
Proof. That C is a Gδ set is a simple exercise. Let G y X be any action in FrMin(G, X).
β
By Lemma 4.1 this action has a free minimal extension G y Y with the property in the
theorem statement, where Y is the Cantor set. Let P be a clopen partition of X and F
a nonempty finite subset of G. Write A1 , . . . , An for the members of the clopen partition
W
−1
−1
s∈F s P. Then for each i = 1, . . . , n the set Ai and its inverse image ϕ (Ai ) under
the extension map ϕ : Y → X are Cantor sets, and so we can find a homeomorphism
ψi : Ai → ϕ−1 (Ai ). Let ψ : X → Y be the homeomorphism which is equal to ψi on Ai for
γ
each i. Then the action G y X defined by γs = ψ −1 ◦ βs ◦ ψ for s ∈ G belongs to C as
well as to the basic open neighborhood Uα,P,F of α, establishing the density of C.
5. Applications to Z-stability
A C∗ -algebra A is said to be Z-stable if A ⊗ Z ∼
= A where Z is the Jiang–Su algebra [10],
with the C∗ -tensor product being unique in this case because Z is nuclear. Z-stability has
become an important regularity property in the classification program for simple separable
nuclear C∗ -algebras, which has recently witnessed some spectacular advances. Thanks to
recent work of Gong–Lin–Niu [6], Elliott–Gong–Lin–Niu [4], and Tikuisis–White–Winter
12
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
[22], it is now known that simple separable unital C∗ -algebras satisfying the universal
coefficient theorem and having finite nuclear dimension are classified by ordered K-theory
paired with tracial states. Although Z-stability does not appear in the hypotheses of this
classification theorem, it does play an important technical role in the proof. Moreover, it
is a conjecture of Toms and Winter that for simple separable infinite-dimensional unital
nuclear C∗ -algebras the following properties are equivalent:
(i) Z-stability,
(ii) finite nuclear dimension,
(iii) strict comparison.
Implications between (i), (ii), and (iii) are known to hold in various degrees of generality.
In particular, the implication (ii)⇒(i) was established in [23] while the converse is known
to hold when the extreme boundary of the convex set of tracial states is compact [1]. It
remains a problem to determine whether any of the crossed products of the actions in Theorem 5.4 falls within the purview of these positive results on the Toms–Winter conjecture,
and in particular whether any of them has finite nuclear dimension (see Question 5.5).
By now there exist highly effectively methods for establishing finite nuclear dimension
for crossed products of free actions on compact metrizable spaces of finite covering dimension [20, 21, 7], but their utility is structurally restricted to groups with finite asymptotic
dimension and hence excludes many amenable examples like the Grigorchuk group. One
can show using the technology from [7] that, for a countably infinite amenable group with
finite asymptotic dimension, the crossed product of a generic free minimal action on the
Cantor set has finite nuclear dimension. Our intention here has been to remove the restriction of finite asymptotic dimension by means of a different approach that establishes
instead the conjecturally equivalent property of Z-stability but for arbitrary countably
infinite amenable groups.
To verify Z-stability in the proof of Theorem 5.3 we will use the following result of
Hirshberg and Orovitz [8]. Recall that a linear map ϕ : A → B between C∗ -algebras is
said to be complete positive if its tensor product id ⊗ ϕ : Mn ⊗ A → Mn ⊗ B with the
identity map on the n × n matrix algebra Mn is positive for every n ∈ N. It is of order zero
if ϕ(a)ϕ(b) = 0 for all a, b ∈ A satisfying ab = 0. One can show that ϕ is an order-zero
completely positive map if and only if there is an embedding B ⊆ D of B into a larger
C∗ -algebra, a ∗ -homomorphism π : A → D, and a positive element h ∈ D commuting with
the image of π such that ϕ(a) = hπ(a) for all a ∈ A [24]. Below - denotes the relation
of Cuntz subequivalence, so that a - b for positive elements a, b in a C∗ -algebra A means
that there is a sequence (vn ) in A such that limn→∞ ka − vn bvn∗ k = 0.
Theorem 5.1. Let A be a simple separable unital nuclear C∗ -algebra not isomorphic to
C. Suppose that for every n ∈ N, finite set Ω ⊆ A, ε > 0, and nonzero positive element
a ∈ A there exists an order-zero complete positive contractive linear map ϕ : Mn → A
such that
(i) 1 − ϕ(1) - a,
(ii) k[b, ϕ(z)]k < ε for all b ∈ Ω and norm-one z ∈ Mn .
Then A is Z-stable.
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
13
The following is the Ornstein–Weiss quasitiling theorem [17] as formulated in Theorem 3.36 of [11]. For finite sets A, F ⊆ G we write
∂F A = {s ∈ A : F s ∩ A 6= ∅ and F s ∩ (G \ A) 6= ∅}.
For λS≤ 1, a collection C of finite subsets of G is said to λ-cover a finite subset A of G if
|A ∩ C| ≥ λ|A|. For β ≥ 0, a collection C of finite subsets of G is said to be β-disjoint if
for each C ∈ C there is a set C ′ ⊆ C with |C ′ | ≥ (1 − β)|C| so that the sets C ′ for C ∈ C
are pairwise disjoint.
Theorem 5.2. Let 0 < β < 12 and let n be a positive integer such that (1 − β/2)n <
β. Then whenever e ∈ T1 ⊆ T2 ⊆ · · · ⊆ Tn are finite subsets of a group G such that
|∂Ti−1 Ti | ≤ (η/8)|Ti | for i = 2, . . . , n, for every (Tn , β/4)-invariant nonempty finite set
E ⊆ G there exist C1 , . . . , Cn ⊆ G such that
S
(i) ni=1 Ti Ci ⊆ E, and
S
(ii) the collection of right translates ni=1 {Ti c : c ∈ Ci } is β-disjoint and (1−β)-covers
E.
Theorem 5.3. Let G be a countably infinite amenable group and let G y X be a free
minimal action on the Cantor set such that for every finite set F ⊆ G and δ > 0 there is a
clopen tower decomposition of X with (F, δ)-invariant shapes. Then C(X) ⋊ G is Z-stable.
Proof. Let n ∈ N. Let Υ be a finite subset of the unit ball of C(X), F a symmetric finite
subset of G containing the identity element e, and ε > 0. Let a be a nonzero positive
element of C(X) ⋊ G. We will show the existence of an order-zero completely positive
contractive linear map ϕ : Mn → C(X) ⋊ G satisfying (i) and (ii) in Theorem 5.1 where
the finite set Ω there is taken to be Υ ∪ {us : s ∈ F }. Since C(X) ⋊ G is generated as a
C∗ -algebra by the unit ball of C(X) and the unitaries us for s ∈ G, we will thereafter be
able to conclude by Theorem 5.1 that C(X) ⋊ G is Z-stable.
By Lemma 7.9 in [18] we may assume that a is a function in C(X). Taking a clopen set
A ⊆ X on which a is nonzero, we may furthermore assume that a is equal to the indicator
function 1A . Minimality implies that the clopen sets sA for s ∈ G cover X, and so by
compactness there is a finite set D ⊆ G such that D −1 A = X.
Equip X with a compatible metric d. Choose an integer Q > n2 /ε.
Let γ > 0, to be determined. Take a 0 < β < 1/n which is small enough so that if T is
a nonempty finite subset of G which is sufficiently invariant
under left translation by F Q
T
and T ′ is a subset of T with |T ′ | ≥ (1 − nβ)|T | then | s∈F Q s−1 T ′ | ≥ (1 − γ)|T |.
Choose an L ∈ N large enough so that (1 − β/2)L < β. By amenability there exist
finite subsets e ∈ T1 ⊆ T2 ⊆ · · · ⊆ TL of G such that |∂Tl−1 Tl | ≤ (β/8)|Tl | for l = 2, . . . , L.
By the previous paragraph, we may also assume that for each l the set Tl is sufficiently
invariant under left translation by F Q so that for all T ⊆ Tl satisfying |T | ≥ (1 − nβ)|Tl |
one has
\
s−1 T ≥ (1 − γ)|Tl |.
(5.1)
s∈F Q
By uniform continuity there is a η > 0 such that |f (x)−f (y)| < ε/(3n2 ) for all f ∈ Υ∪Υ2
and all x, y ∈ X satisfying d(x, y) < η. Again by uniform continuity there is an η ′ > 0
14
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
S
such that d(tx, ty) < η for all x, y ∈ X satisfying d(x, y) < η ′ and all t ∈ L
l=1 Tl . Fix a
clopen partition {A1 , . . . , AM } of X whose members all have diameter less that η ′ .
Let E be a finite subset of G containing TL and let δ > 0 be such that δ ≤ β/4. We will
further specify E and δ below. By hypothesis there is a collection {(Vk , Sk )}K
k=1 of clopen
towers such that the shapes S1 , . . . , SK are (E, δ)-invariant and the sets S1 V1 , . . . , SK VK
partition X. We may assume that for each k = 1, . . . , K the set Sk is large enough so that
X
L
Mn
(5.2)
|Tl | ≤ β|Sk |.
l=1
By a simple procedure we can construct, for each k, a clopen partition Pk of Vk such that
each level of every one of the towers (V, Sk ) for V ∈ Pk is contained in one of the sets
A1 , . . . , AM as well as in one of the sets A and X \ A. By replacing (Vk , Sk ) with these
thinner towers for each k, we may therefore assume that each level in every one of the
towers (V1 , S1 ), . . . , (VK , SK ) is contained in one of the sets A1 , . . . , AM and in one of the
sets A and X \ A.
Let 1 ≤ k ≤ K. Since Sk is (TL , β/4)-invariant, by Theorem 5.2 and our choice of
the sets T1 , . . . , TL we can find Ck,1 , . . . , Ck,L ⊆ Sk such that the collection {Tl c : l =
1, . . . , L, c ∈ Ck,l } is β-disjoint, has union contained in Sk , and (1 − β)-covers Sk . By
β-disjointness, for every l = 1, . . . , L and c ∈ Ck,l we can find a Tk,l,c ⊆ Tl satisfying
|Tk,l,c | ≥ (1 − β)|Tl | so that the collection of sets Tk,l,cc for l = 1, . . . , L and c ∈ Ck,l is
disjoint and has the same union as the sets Tl c for l = 1, . . . , L and c ∈ Ck,l , so that it
(1 − β)-covers Sk .
For each l = 1, . . . , L and m = 1, . . . , M write Ck,l,m for the set of all c ∈ Ck,l such that
(n)
(1)
cVk ⊆ Am , and choose pairwise disjoint subsets Ck,l,m, . . . , Ck,l,m of Ck,l,m such that each
has cardinality ⌊|Ck,l,m |/n⌋. For each i = 2, . . . , n choose a bijection
G (1)
G (i)
Λk,i :
Ck,l,m →
Ck,l,m
l,m
(1)
l,m
(i)
which sends Ck,l,m to Ck,l,m for all l, m. Also, define Λk,1 to be the identity map from
F
(1)
l,m Ck,l,m to itself.
T
F
(1)
′
= ni=1 Tk,l,Λk,i (c) , which satisfies
Let 1 ≤ l ≤ L and c ∈ m Ck,l,m. Define the set Tk,l,c
(5.3)
′
| ≥ (1 − nβ)|Tl | ≥ (1 − nβ)|Tk,l,c |
|Tk,l,c
since each Tk,l,Λk,i (c) is a subset of Tl of cardinality at least (1 − β)|Tl |. Set
\
′
′
Bk,l,c,Q =
sTk,l,c
, Bk,l,c,0 = Tk,l,c
\ F Q−1 Bk,l,c,Q,
s∈F Q
and, for q = 1, . . . , Q − 1, using the convention F 0 = {e},
Bk,l,c,q = F Q−q Bk,l,c,Q \ F Q−q−1 Bk,l,c,Q.
′
. For s ∈ F we have
Then the sets Bk,l,c,0, . . . , Bk,l,c,Q partition Tk,l,c
(5.4)
sBk,l,c,Q ⊆ Bk,l,c,Q−1 ∪ Bk,l,c,Q,
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
15
while for q = 1, . . . , Q − 1 we have
sBk,l,c,q ⊆ Bk,l,c,q−1 ∪ Bk,l,c,q ∪ Bk,l,c,q+1,
(5.5)
for if we are given a t ∈ Bk,l,c,q then st ∈ F Q−q+1 Bk,l,c,Q, while if st ∈ F Q−q−2 Bk,l,c,Q then
t ∈ F Q−q−1 Bk,l,c,Q since F is symmetric, contradicting the membership of t in Bk,l,c,q .
Also, from (5.1) and (5.3) we get
|Bk,l,c,Q| ≥ (1 − γ)|Tl |.
(5.6)
For i = 2, . . . , n, c ∈
F
(i)
m Ck,l,m ,
and q = 0, . . . , Q we set Bk,l,c,q = Bk,l,λ−1(c),q .
k,i
Write Λk,i,j for the composition Λk,i ◦ Λ−1
k,j . Define a linear map ψ : Mn → C(X) ⋊ G
by declaring it on the standard matrix units {eij }ni,j=1 of Mn to be given by
X
X X
utΛk,i,j (c)c−1 t−1 1tcVk
ψ(eij ) =
k,l,m c∈C (j)
k,l,m
′
t∈Tk,l,c
and extending linearly. Then ψ(eij )∗ = ψ(eji ) for all i, j and the product ψ(eij )ψ(ei′ j ′ ) is
1 or 0 depending on whether i = i′ , so that ψ is a ∗ -homomorphism.
F
(j)
For all k and l, all 1 ≤ i, j ≤ n, and all c ∈ m Ck,l,m we set
hk,l,c,i,j =
Q
X
X
q=1 t∈Bk,l,c,q
q
u
−1 −1 1tcV
k
Q tΛk,i,j (c)c t
and put
h=
n
X
X X
hk,l,c,i,i.
k,l,m i=1 c∈C (i)
k,l,m
Then h is a norm-one function which commutes with the image of ψ, and so we can define
an order-zero completely positive contractive linear map ϕ : Mn → C(X) ⋊ G by setting
ϕ(z) = hψ(z).
Note that ϕ(eij ) =
P
k,l,m
P
(j)
c∈Ck,l,m
hk,l,c,i,j .
We now verify condition (ii) in Theorem 5.1 for the elements of the set {us : s ∈ F }.
F
(j)
Let 1 ≤ i, j ≤ n. For all k and l, all c ∈ m Ck,l,m, and all s ∈ F we have
us hk,l,c,i,j u−1
s − hk,l,c,i,j =
Q
X
X
q=1 t∈Bk,l,c,q
q
u
−1
−1 1stcV
k
Q stΛk,i,j (c)c (st)
−
Q
X
X
q=1 t∈Bk,l,c,q
q
u
−1 −1 1tcV ,
k
Q tΛk,i,j (c)c t
and so in view of (5.4) and (5.5) we obtain
kus hk,l,c,i,j u−1
s − hk,l,c,i,j k ≤
1
ε
< 2.
Q
n
16
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
∗
∗
Since each of the elements b = us hk,l,c,i,j u−1
s −hk,l,c,i,j is such that b b and bb are dominated
′
′
by twice the indicator functions of Tk,l,Λ−1 (c) cVk and Tk,l,Λ−1 (c) Λk,i,j (c)Vk , respectively, and
j
j
F
(i′ )
′
′
the sets Tk,l,Λ−1(c) cVk over all k, l, all i = 1, . . . , n, and all c ∈ m Ck,l,m are pairwise
i′
disjoint, this yields
−1
kus ϕ(eij )u−1
s − ϕ(eij )k = max max kus hk,l,c,i,j us − hk,l,c,i,j k <
k,l,m c∈C (j)
k,l,m
ε
n2
and hence, for every norm-one element z = (zij ) ∈ Mn ,
k[us , ϕ(z)]k = kus ϕ(z)u−1
s − ϕ(z)k ≤
n
X
|zij |kus ϕ(eij )u−1
s − ϕ(eij )k
i,j=1
ε
= ε.
n2
Next we verify condition (ii) in Theorem 5.1 for the functions in Υ. Let 1 ≤ i, j ≤ n.
(j)
Let g ∈ Υ ∪ Υ2 . Let 1 ≤ k ≤ K, 1 ≤ l ≤ L, 1 ≤ m ≤ M , and c ∈ Ck,l,m. Then
< n2 ·
(5.7)
h∗k,l,c,i,j ghk,l,c,i,j =
Q
X
X
q=1 t∈Bk,l,c,q
q2
(tcΛk,i,j (c)−1 t−1 g)1tcVk
Q2
and
(5.8)
gh∗k,l,c,i,j hk,l,c,i,j
=
Q
X
X
q=1 t∈Bk,l,c,q
q2
g1tcVk .
Q2
Now let x ∈ Vk . Since Λk,i,j (c)x and cx both belong to Am , we have d(Λk,i,j (c)x, cx) < η ′ .
It follows that for every t ∈ Tl we have d(tΛk,i,j (c)x, tcx) < η by our choice of η ′ , so that
|g(tΛk,i,j (c)x) − g(tcx)| < ε/(3n2 ) by our choice of η, in which case
k(tcΛk,i,j (c)−1 t−1 g − g)1tcVk k = kc−1 t−1 ((tcΛk,i,j (c)−1 t−1 g − g)1tcVk )k
= k(Λk,i,j (c)−1 t−1 g − c−1 t−1 g)1Vk k
= sup |g(tΛk,i,j (c)x) − g(tcx)|
x∈Vk
<
ε
.
3n2
Using (5.7) and (5.8) this gives us
(5.9)
kh∗k,l,c,i,j ghk,l,c,i,j − gh∗k,l,c,i,j hk,l,c,i,j k
= max
max
q=1,...,Q t∈Bk,l,c,q
q2
ε
k(tcΛk,i,j (c)−1 t−1 g − g)1tcVk k < 2 .
Q2
3n
Set w = ϕ(eij ) for brevity. Let f ∈ Υ. For g ∈ {f, f 2 } the functions h∗k,l,c,i,j ghk,l,c,i,j −
(j)
gh∗k,l,c,i,j hk,l,c,i,j over all k, l, and m and all c ∈ Ck,l,m have pairwise disjoint supports, so
that (5.9) yields
ε
kw∗ gw − gw∗ wk < 2 .
3n
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
17
It follows that
kw∗ f 2 w − f w∗ f wk ≤ kw∗ f 2 w − f 2 w∗ wk + kf (f w∗ w − w∗ f w)k <
2ε
3n2
and so
kf w − wf k2 = k(f w − wf )∗ (f w − wf )k
= kw∗ f 2 w − f w∗ f w + f w∗ wf − w∗ f wf k
≤ kw∗ f 2 w − f w∗ f wk + k(f w∗ w − w∗ f w)f k
2ε
ε
ε
< 2 + 2 = 2.
3n
3n
n
Therefore for every norm-one element z = (zij ) ∈ Mn we have
k[f, ϕ(z)]k ≤
n
X
|zij |k[f, ϕ(eij )]k < n2 ·
i,j=1
ε
= ε.
n2
Finally, we verify that the parameters in the construction of ϕ can be chosen so that
1 − ϕ(1) - 1A . By taking the sets S1 , . . . , SK to be sufficiently left invariant (by enlarging
E and shrinking δ if necessary) we may assume that for every k = 1, . . . , K there is an
Sk′ ⊆ Sk such that the set {s ∈ Sk′ : Ds ⊆ Sk } has cardinality at least |Sk |/2. Let
1 ≤ k ≤ K. Take a maximal set Sk′′ ⊆ Sk′ such that the sets Ds for s ∈ Sk′′ are pairwise
disjoint, and note that |Sk′′ | ≥ |Sk′ |/|D −1 D| ≥ |Sk |/(2|D|2 ). Since D −1 A = X, each of the
sets DsVk for s ∈ Sk′′ intersects A, and so the set Sk♯ of all s ∈ Sk such that sVk ⊆ A has
F F F
cardinality at least |Sk |/(2|D|2 ). Define Sk,1 = l,m ni=1 c∈C (i) Bk,l,c,Qc, which is the
k,l,m
set of all s ∈ Sk such that the function ϕ(1) takes the value 1 on sVk . Set Sk,0 = Sk \ Sk,1 .
P
(i)
Since ni=1 |Ck,l,m | ≥ |Ck,l,m | − n for every l and m, by (5.2) we have
n
XX
(i)
|Tl ||Ck,l,m | ≥
l,m i=1
X
|Tl ||Ck,l,m | − M n
l,m
≥
[
X
|Tl |
l
Tl Ck,l − β|Sk |
l
≥ (1 − 2β)|Sk |.
(i)
Since for all l and i and all c ∈ Ck,l,m we have |Bk,l,c,Q| ≥ (1 − γ)|Tl | by (5.6), it follows,
putting λ = (1 − γ)(1 − 2β), that
|Sk,1 | ≥ (1 − γ)
n
XX
(i)
|Tl ||Ck,l,m | ≥ λ|Sk |.
l,m i=1
By taking γ and β small enough we can guarantee that 1 − λ ≤ 1/(2|D|2 ) and hence
|Sk,0 | = |Sk | − |Sk,1 | ≤ (1 − λ)|Sk | ≤ |Sk♯ |,
18
CONLEY, JACKSON, KERR, MARKS, SEWARD, AND TUCKER-DROB
so that there exists an injection θk : Sk,0 → Sk♯ . Define
z=
K X
X
uθk (s)s−1 1sVk .
k=1 s∈Sk,0
A simple computation shows that z ∗ 1A z is the indicator function of
is the support of 1 − ϕ(1), and so putting v = (1 − ϕ(1))1/2 z ∗ we get
FK
k=1 Sk,0 Vk ,
which
v1A v ∗ = (1 − ϕ(1))1/2 z ∗ 1A z(1 − ϕ(1))1/2 = 1 − ϕ(1).
This demonstrates that 1 − ϕ(1) - 1A , as desired.
Combining Theorems 5.1 and 4.2 yields the following.
Theorem 5.4. Let G be a countably infinite amenable group and X the Cantor set. Then
the set of all actions in FrMin(G, X) whose crossed product is Z-stable is comeager, and
in particular nonempty.
Question 5.5. Do any of the crossed products in Theorem 5.4 have tracial state space
with compact extreme boundary (from which we would be able to conclude finite nuclear
dimension by [1] and hence classifiability)? For G = Z a generic action in FrMin(G, X)
is uniquely ergodic, so that the crossed product has a unique tracial state [9]. However,
already for Z2 nothing of this nature seems to be known. On the other hand, it is known
that the crossed products of free minimal actions of finitely generated nilpotent groups on
compact metrizable spaces of finite covering dimension have finite nuclear dimension, and
in particular are Z-stable [21].
References
[1] J. Bosa, N. Brown, Y. Sato, A. Tikuisis, S. White and W. Winter. Covering dimension of C∗ -algebras
and 2-coloured classification. To appear in Mem. Amer. Math. Soc..
[2] T. Ceccherini-Silberstein, P. de la Harpe, and R. I. Grigorchuk. Amenability and paradoxical decompositions for pseudogroups and discrete metric spaces. (Russian) Tr. Mat. Inst. Steklova 224 (1999),
68–111; translation in Proc. Steklov Inst. Math. 224 (1999), 57–97.
[3] T. Downarowicz, D. Huczek, and G. Zhang. Tilings of amenable groups. To appear in J. Reine
Angew. Math.
[4] G. Elliott, G. Gong, H. Lin, and Z. Niu. On the classification of simple amenable C∗ -algebras with
finite decomposition rank, II. arXiv:1507.03437.
[5] G. Elek and G. Lippner. Borel oracles. An analytic approach to constant time algorithms. Proc.
Amer. Math. Soc. 138 (2010), 2939–2947.
[6] G. Gong, H. Lin, and Z. Niu. Classification of finite simple amenable Z-stable C∗ -algebras.
arXiv:1501.00135.
[7] E. Guentner, R. Willett, and G. Yu. Dynamic asymptotic dimension: relation to dynamics, topology,
coarse geometry, and C∗ -algebras. Math. Ann. 367 (2017), 785–829.
[8] I. Hirshberg and J. Orovitz. Tracially Z–absorbing C∗ -algebras. J. Funct. Anal. 265 (2013), 765–785.
[9] M. Hochman. Genericity in topological dynamics. Ergodic Theory Dynam. Systems 28 (2008), 125–
165.
[10] X. Jiang and H. Su. On a simple unital projectionless C∗ -algebra. Amer. J. Math. 121 (1999),
359–413.
[11] D. Kerr and H. Li. Ergodic Theory: Independence and Dichotomies. Springer, Cham, 2016.
[12] A. Kechris, S. Solecki, and S. Todorcevic. Borel chromatic numbers, Adv. in Math. 141 (1999), 1–44.
FØLNER TILINGS FOR ACTIONS OF AMENABLE GROUPS
19
[13] R. Lyons and F. Nazarov. Perfect matchings as IID factors on non-amenable groups. European J.
Combin. 32 (2011), 1115–1125.
[14] H. Matui and Y. Sato. Strict comparison and Z-absorption of nuclear C∗ -algebras. Acta Math. 209
(2012), 179–196.
[15] H. Matui and Y. Sato. Decomposition rank of UHF-absorbing C∗ -algebras. Duke Math. J. 163 (2014),
2687–2708.
[16] I. Namioka, Følner’s conditions for amenable semi-groups. Math. Scand. 15 (1964), 18–28.
[17] D. S. Ornstein and B. Weiss. Entropy and isomorphism theorems for actions of amenable groups. J.
Analyse Math. 48 (1987), 1–141.
[18] N. C. Phillips. Large subalgebras. arXiv:1408.5546.
[19] M. Rørdam and W. Winter. The Jiang–Su algebra revisited. J. Reine Angew. Math. 642 (2010),
129–155.
[20] G. Szabó. The Rokhlin dimension of topological Zm -actions. Proc. Lond. Math. Soc. (3) 110 (2015),
673–694.
[21] G. Szabo, J. Wu, and J. Zacharias. Rokhlin dimension for actions of residually finite groups.
arXiv:1408.6096.
[22] A. Tikuisis, S. White, and W. Winter. Quasidiagonality of nuclear C∗ -algebras. Ann. of Math. (2)
185 (2017), 229–284.
[23] W. Winter. Nuclear dimension and Z-stability of pure C∗ -algebras. Invent. Math. 187 (2012), 259–
342.
[24] W. Winter and J. Zacharias. Completely positive maps of order zero. Münster J. Math. 2 (2009),
311–324.
Clinton T. Conley, Department of Mathematical Sciences, Carnegie Mellon University,
Pittsburgh, PA 15213, U.S.A.
E-mail address: [email protected]
Steve Jackson, Department of Mathematics, University of North Texas, Denton, TX 762035017, U.S.A.
E-mail address: [email protected]
David Kerr, Department of Mathematics, Texas A&M University, College Station, TX
77843-3368, U.S.A.
E-mail address: [email protected]
Andrew Marks, UCLA Department of Mathematics, Los Angeles, CA 90095-1555, U.S.A.
E-mail address: [email protected]
Brandon Seward, Courant Institute of Mathematical Sciences, New York, NY 10012, U.S.A.
E-mail address: [email protected]
Robin Tucker-Drob, Department of Mathematics, Texas A&M University, College Station,
TX 77843-3368, U.S.A.
E-mail address: [email protected]
| 4math.GR
|
Conditional Expectation of a Markov Kernel Given Another
with Some Applications in Statistical Inference and Disease Diagnosis
arXiv:1306.0084v3 [math.ST] 12 Sep 2017
A.G. Nogales
Dpto. de Matemáticas, Universidad de Extremadura
Avda. de Elvas, s/n, 06006–Badajoz, SPAIN.
e-mail: [email protected]
Fax number: +34 924272911
Abstract.
Markov kernels play a decisive role in probability and mathematical statistics theories, and are an
extension of the concepts of sigma-field and statistic. Concepts such as independence, sufficiency,
completeness, ancillarity or conditional distribution have been extended previously to Markov kernels.
In this paper, the concept of conditional expectation of a Markov kernel given another is introduced,
setting its first properties. An application to clinical diagnosis is provided, obtaining a optimality
property of the predictive values of a diagnosis test. In a statistical framework, this new probabilistic
tool is used to extend to Markov kernels the theorems of Rao-Blackwel and Lehmann-Scheffé. To
obtain and example of application of this generalized version of the Lehmann-Scheffé Theorem, some
results are presented to describe a general procedure to construct complete sufficient Markov kernels
from a sufficient one.
As a final statistical scholium, a generalization of a result about the completeness of the family of
nonrandomized estimators is given.
AMS Subject Class. (2010): Primary 60E05 Secondary 62F10, 62P10, 62B05
Key words and phrases: Markov kernel, conditional expectation, clinical diagnosis, unbiased estimation.
Short running title: Conditional expectations for Markov kernels
1
Introduction
Markov kernels (also referred to as stochastic kernels or transition probabilities) play an important role
in probability theory and mathematical statistics. Indeed, the conditional distribution of one random
variable given another is a Markov kernel (here, we use the term random variable as being synonymous of a measurable function between two arbitrary measurable spaces). In fact, as we shall show
below, every Markov kernel is the conditional distribution of some random variable given another. A
transition matrix in Markov chains theory defines a Markov kernel. Sampling probabilities and posterior distributions in Bayesian inference are Markov kernels. In statistical decision theory, randomized
procedures (also named decision rules or, even, strategies) are Markov kernels, while nonrandomized
procedures are statistics. It is well known that, in some situations, the optimum procedure is a randomized one: for example, the fundamental lemma of Neyman and Pearson shows how randomization
is necessary to obtain a most powerful test; Lehmann (2005) also describes many other statistical
situations where the use of randomization is properly justified. Pfanzagl (1994, Example 4.2.2) shows
a testing problem where there is no nonrandomized test at least as good as a certain randomized test.
A Markov kernel can also be considered as a generalization of the concepts of σ-field and random
variable (or statistic, in a statistical framework).
Well known concepts of the theory of probabilities or mathematical statistics, such as independence,
completeness, ancillarity or conditional distribution have been extended to Markov kernels in Nogales
(2013a) and Nogales (2013b). The reader is referred to Heyer (1982) for the corresponding extension
to the concept of sufficiency in the context of informativity for statistical experiments. Notice that
an extension to Markov kernels of the concepts and results of probability and mathematical statistics
should not be considered useless, as it is not the extension to Markov kernels (or transitions) of the
classical theorems of the product measure and Fubini: it is the version for Markov kernel of this
theorems what we need to describe the joint distribution of two random variables X and Y in terms
of the marginal distribution of X and the conditional distributions of Y given a value of X.
On the other hand, the conditional expectation E(Y |X) of an integrable n-dimensional random
variable Y given a random variable X is the main tool in the study of the relationship between them; in
fact, y = E(Y |X = x) is the so-called regression curve (in a wide sense) of Y on X. Basic properties
and results on conditional expectations can be found in almost every graduate text in probability
theory, after its mathematical introduction in Kolmogorov (1933).
In this paper, we introduce a new probabilistic tool: the conditional expectation for Markov
kernels. Its relationship with the concept of conditional distribution of a Markov kernel given another
is established. Some basic properties, two examples of calculation of such a conditional expectation,
and a representation theorem in terms of conditional expectation for random variables are also given.
One of the given examples is applied to clinical diagnosis, where some expectations and conditional
expectations for Markov kernels get a specific meaning. We obtain in particular a optimality property
of the predictive values of a diagnosis test as the point that minimizes two naturally weighted distances
to the correct decisions on the subpopulations of ill and non-ill individuals. As far as we know, this
interpretation of the predictive values appears here for the first time.
As an statistical application, in this paper we make use of such tools to extend to Markov kernels the
theorems of Rao-Blackwel and Lehmann-Scheffé. These well known theorems are major milestones
of mean unbiased estimation theory, going back to Rao (1945) and Blackwell (1947) regard to the
Theorem of Rao-Blackwell, and to Lehmann and Scheffé (1950) regard to the Theorem of LehmannScheffé. The reader is referred to Pfanzagl (1994, p. 105) for a version for statistics of these theorems;
it is assured even there that a more general version of the Rao-Blackwell theorem can be proved
in the same way for randomized estimators when a sufficient and complete statistics exists. In this
paper both results are generalized for randomized estimators when a sufficient and complete Markov
kernel is known. Two examples of sufficient Markov kernel associated to any statistic are given; two
more similar examples are provided for complete Markov kernels. An example of application of the
generalized Lehmann-Scheffé Theorem not covered by the classical version is also given in the last
section of the paper. In fact, some results are presented to describe a general procedure to construct
complete sufficient Markov kernels from a sufficient one. Namely, the example is obtained from a
result on the completeness of the “largest” family of probabilities keeping sufficient a given Markov
2
kernel.
Finally, notice that, as the conditional expectation of a Markov kernel given another is a statistic,
a generalization of a result about the completeness of the family of nonrandomized estimators is also
obtained.
For ease of reading, the proofs of the results have been collected in the last section of the paper.
2
Basic definitions
The concepts presented in this section can be found in Heyer (1982), although the notation has been
modified to highlight the analogy with similar concepts for random variables. See also Dellacherie and
Meyer (1988). In the next, (Ω, A), (Ω1 , A1 ), and so on, will denote measurable spaces. A random
variable is a map X : (Ω, A) → (Ω1 , A1 ) such that X −1 (A1 ) ∈ A, for all A1 ∈ A1 . Its probability
distribution (or, simply, distribution) P X with respect to a probability measure P on A is the image
measure of P by X, i.e., the probability measure on A1 defined by P X (A1 ) := P (X −1 (A1 )). We will
write × instead of ⊗ for the product of σ-fields or measures. Rk will denote the Borel σ-field on Rk .
Definition 1. (Markov kernel) A Markov kernel M1 : (Ω, A)≻−→(Ω1 , A1 ) is a map M1 : Ω×A1 → [0, 1]
such that
(i) ∀ω ∈ Ω, M1 (ω, ·) is a probability measure on A1 ,
(ii) ∀A1 ∈ A1 , M1 (·, A1 ) is A-measurable.
Remarks. 1) Given two random variables Xi : (Ω, A, P ) → (Ωi , Ai ), i = 1, 2, the conditional distribution of X2 given X1 , when
it exists, is a Markov kernel M : (Ω1 , A1 )≻−→(Ω2 , A2 ) such that
R
P (X1 ∈ A1 , X2 ∈ A2 ) = A1 M (ω1 , A2 )dP X1 (ω1 ), for all A1 ∈ A1 and A2 ∈ A2 . We write
P X2 |X1 =ω1 (A2 ) := M (ω1 , A2 ). Reciprocally, every Markov kernel is a conditional distribution; namely,
given a Markov kernel M1 : (Ω, A, P )≻−→(Ω1 , A1 ), it is easily checked that
M1 (ω, A1 ) = (P ⊗ M1 )π1 |π=ω (A1 ),
where π : Ω × Ω1 → Ω and π1 : Ω × Ω1 → Ω1 are the coordinatewise projections and P ⊗ M1 stands
for
R the only probability measure on the product space (Ω × Ω1 , A × A1 ) such that (P ⊗ M1 )(A × A1 ) =
A M1 (ω, A1 ) dP (ω) for all A ∈ A and A1 ∈ A1 .
2) The concept of Markov kernel extends the concepts of random variable and σ-field. A random
variable T1 : (Ω, A) → (Ω1 , A1 ) will be identified with the Markov kernel MT1 : (Ω, A)≻−→(Ω1 , A1 )
defined by
MT1 (ω, A1 ) = δT1 (ω) (A1 ) = IA1 (T1 (ω)),
where δT1 (ω) denotes the Dirac measure -the degenerate distribution- at the point T1 (ω) and IA1 is the
indicator function of the event A1 . The sub-σ-field B ⊂ A will be identified with the Markov kernel
MB : (Ω, A)≻−→(Ω, B) given by MB (ω, B) = δω (B).
Definition 2. (Image of a Markov kernel) The image (or probability distribution) of a Markov kernel
M1 : (Ω, A, P )≻−→(Ω1 , A1 ) on a probability space is the probability measure P M1 on A1 defined by
Z
M1
M1 (ω, A1 ) dP (ω).
P (A1 ) :=
Ω
Remark. Note that
P M1 = (P ⊗ M1 )π1
where π1 : Ω × Ω1 → Ω1 denotes the coordinatewise projection onto Ω1 . So, if f : (Ω1 , A1 ) → R is a
nonnegative or P M1 -integrable function,
Z Z
Z
M1
f (ω1 )M1 (ω, dω1 )dP (ω)
f (ω1 )dP (ω1 ) =
Ω Ω1
Ω1
Z
f (ω1 )d(P ⊗ M1 )(ω, ω1 ).
=
Ω×Ω1
3
Definition 3. (a) (Composition of Markov kernels) The composition of two Markov kernels M1 :
(Ω1 , A1 )≻−→(Ω2 , A2 ) and M2 : (Ω2 , A2 )≻−→(Ω3 , A3 ) is defined as the Markov kernel
M2 M1 : (Ω1 , A1 )≻−→(Ω3 , A3 )
given by
M2 M1 (ω1 , A3 ) =
Z
M2 (ω2 , A3 )M1 (ω1 , dω2 ).
Ω2
(b) (Composition of a Markov kernel and a random variable) Let X1 : (Ω, A) → (Ω1 , A1 ) be a
random variable and M1 : (Ω1 , A1 )≻−→(Ω′1 , A′1 ) a Markov kernel. A new Markov kernel M1 X1 :
(Ω, A)≻−→(Ω′1 , A′1 ) is defined by means of
M1 X1 (ω, A′1 ) := M1 (X1 (ω), A′1 ).
Remark. When MX1 is the Markov kernel corresponding to the random variable X1 , we have that
M1 X1 = M1 MX1 .
3
Expectation and conditional expectation for Markov kernels
Next we introduce the concepts of expectation and conditional expectations for Markov kernels. Let
(Ω, A, P ) be a probability space.
Definition 4. (Expectation of aRMarkov kernel) A Markov kernel M1 : (Ω, A, P )≻−→Rk is said to
be P -integrable if the map ω 7→ Rk xM1 (ω, dx) is P -integrable, i.e., if there exists and is finite the
integral
Z Z
xM1 (ω, dx)dP (ω)
Ω
Rk
or, which is the same, if the distribution (P ⊗ M1 )π2 has finite mean, where π2 : Ω × Rk → Rk denotes
the second coordinatewise projection. In this case, we define the expectation of the Markov kernel M1
as
Z Z
xM1 (ω, dx)dP (ω)
EP (M1 ) :=
Rk
Ω
Definition 5. Let M1 : (Ω, A, P )≻−→Rk be a P -integrable Markov kernel. We define a set function
M1 · P on A by
Z Z
xM1 (ω, dx)dP (ω).
(M1 · P )(A) :=
A
Rk
Note that M1 ·P ≪ P and (M1 ·P )M2 ≪ P M2 , when M2 : (Ω, A, P )≻−→(Ω2 , A2 ) is another Markov
kernel.
Definition 6. (Conditional expectation of a Markov kernel given another) Let M1 : (Ω, A, P )≻−→Rk
be a P -integrable Markov kernel and M2 : (Ω, A, P )≻−→(Ω2 , A2 ) be a Markov kernel. The conditional
expectation EP (M1 |M2 ) is defined by:
EP (M1 |M2 ) :=
d(M1 · P )M2
dP M2
i.e., EP (M1 |M2 ) is the (equivalence class of) real measurable function(s) on (Ω2 , A2 ) such that, for
every A2 ∈ A2 ,
Z
Z
Z
EP (M1 |M2 )dP M2
xM1 (ω, dx)dP (ω) =
M2 (ω, A2 )
A2
Rk
Ω
Z Z
EP (M1 |M2 )(ω2 )M2 (ω, dω2 )dP (ω).
=
Ω
4
A2
The next result yields an integral representation of such a conditional expectation. First, we
refer the reader to Nogales (2013b) for the definition and existence of the conditional distribution
P M1 |M2 of a Markov kernel M1 : (Ω, A, P )≻−→(Ω1 , A1 ) with respect to another Markov kernel M2 :
(Ω, A, P )≻−→(Ω2 , A2 ). Namely, it is defined as a Markov kernel L : (Ω2 , A2 )≻−→(Ω1 , A1 ) such that,
for every pair of events A1 ∈ A1 and A2 ∈ A2 ,
Z
Z
L(ω2 , A1 )dP M2 (ω2 )
M1 (ω, A1 )M2 (ω, A2 )dP (ω) =
Ω
A2
Z Z
L(ω2 , A1 )M2 (ω, dω2 )dP (ω)
=
Ω
A2
Theorem 1. Let M1 and M2 be two Markov kernels as in the previous definition. Then
Z
EP (M1 |M2 )(ω2 ) =
xP M1 |M2 (ω2 , dx)
Rk
(in the sense that the last integral defines a version of the conditional expectation of M1 given M2 ).
More generally, if f : Rk → Rm has nonnegative components or is P M1 -integrable function, then
Z
EP (f M1 |M2 )(ω2 ) =
f (x)P M1 |M2 (ω2 , dx),
Rk
where f M1 is the Markov kernel defined by f M1 (ω, C) := M1 (ω, f −1 (C)), ω ∈ Ω, C ∈ Rm .
The following are two examples of calculation.
Example 1. Given θ ∈ [0, 1], let Ω = {0, 1}, A = P(Ω) and P the probability measure on (Ω, A)
assigning probability θ to the point 1 and 1 − θ to the point 0. For i = 1, 2, consider the Markov kernel
Mi : (Ω, A)≻−→(Ω, A) defined by the stochastic matrix
pi 1 − pi
,
qi 1 − qi
where 0 ≤ pi , qi ≤ 1. Then
P M1 ({0}) =
Z
M1 (ω, {0})dP (ω) = (1 − θ)p1 + θq1
{0,1}
and P M1 ({1}) = (1 − θ)(1 − p1 ) + θ(1 − q1 ). Hence,
Z
Z
Z
M1 (ω, {1})dP (ω) = (1−θ)(1−p1 )+θ(1−q1 ) = P M1 ({1}).
xM1 (ω, dx)dP (ω) =
EP (M1 ) =
{0,1}
R
{0,1}
Moreover, if L := P M2 |M1 : (Ω, A)≻−→(Ω, A), according to Nogales (2013b, Prop. 2), given
ω1 , ω2 ∈ {0, 1},
R
M1 (i, ω1 )M2 (i, ω2 )dP (i)
(1 − θ)M1 (0, ω1 )M2 (0, ω2 ) + θM1 (1, ω1 )M2 (1, ω2 )
{0,1}
R
=
L(ω1 , {ω2 }) =
(1 − θ)M1 (0, ω1 ) + θM1 (1, ω1 )
{0,1} M1 (i, ω1 )dP (i)
Hence
L(0, {1}) =
(1 − θ)p1 (1 − p2 ) + θq1 (1 − q2 )
(1 − θ)M1 (0, 0)M2 (0, 1) + θM1 (1, 0)M2 (1, 1)
=
(1 − θ)M1 (0, 0) + θM1 (1, 0)
(1 − θ)p1 + θq1
and
L(1, {1}) =
(1 − θ)M1 (0, 1)M2 (0, 1) + θM1 (1, 1)M2 (1, 1)
(1 − θ)(1 − p1 )(1 − p2 ) + θ(1 − q1 )(1 − q2 )
=
,
(1 − θ)M1 (0, 1) + θM1 (1, 1)
(1 − θ)(1 − p1 ) + θ(1 − q1 )
5
while L(ω1 , {0}) = 1 − L(ω1 , {1}), ω1 = 0, 1. Finally, for ω1 ∈ {0, 1},
Z
EP (M2 |M1 )(ω1 ) =
xP M2 |M1 (ω1 , dx) = L(ω1 , {1}).
{0,1}
Subexample 1.1: (Application to clinical diagnosis) Consider a diagnosis tests T for a certain
disease D. We write D = 1 (= 0) for an individual having (not having) the disease as determined
by a “gold standard” diagnostic procedure, and T = 1 (= 0) if the diagnostic is positive (negative).
There are several terms that are commonly used in this context: P (D = 1) is called the prevalence
of the disease (on a given population), while s = P (T = 1|D = 1) is the sensitivity of the test and
e = P (T = 0|D = 0) is its specificity. The stochastic matrix
p1 = e
1−e
M1 =
,
q1 = 1 − s
s
describes the transition probabilities from the state i ∈ {0, 1} (the gold standard test is negative -i = 0or positive -i = 1-) to the state j ∈ {0, 1} (the test T is negative -j = 0- or positive -j = 1-). This
way, M1 becomes a Markov kernel from {0, 1} to {0, 1} and its probability distribution P M1 satisfies
P M1 ({1}) = (1 − θ)(1 − e) + θ(1 − s) = P (T = 1),
the probability that any given individual of the population receive a positive diagnostic. If M2 denotes
the gold standard diagnostic test, M2 is (identified with) the identity matrix of order 2 (i.e., p2 = 1
and q2 = 0). So, analogously, P M2 ({1}) = P (D = 1) = θ. Moreover, if L := P M2 |M1 , according to the
example, we have that
EP (M2 |M1 )(1) = L(1, {1}) =
θs
= P (D = 1|T = 1)
(1 − θ)(1 − e) + θs
is the so-called predictive positive value P P V of the diagnosis test T and, in the same way,
EP (M2 |M1 )({0}) = L(0, {1}) = 1 −
(1 − θ)e
= 1 − P (D = 0|T = 0),
(1 − θ)e + θ(1 − s)
P (D = 0|T = 0) being the predictive negative value P N V of T . Now, if
1−e e
N1 =
,
1−s s
we have that
EP (N1 ) = P N1 ({1}) = (1 − θ)e + θs = P (T = 0, D = 0) + P (T = 1, D = 1)
is the “accuracy” of T , i.e., the proportion of true diagnostics of T . Moreover
EP (M2 |N1 )(1) =
P (T = 1, D = 1)
θs
=
θs + (1 − θ)e
P (T = 0, D = 0) + P (T = 1, D = 1)
is the proportion of positive true diagnostics among all true diagnostics of T .
Example 2. For 1 ≤ i ≤ 3, let (Ωi , Ai , µi ) be a σ-finite measure space such that (Ωi , Ai ) is a
standard Borel space for i = 2, 3, and Xi : (Ω, A, P ) → (Ωi , Ai , µi ) is a random variable. We assume
that the joint distribution of X = (X1 , X2 , X3 ) admits a density f with respect to the product measure
µ1 × µ2 × µ3 . We write fij for the joint µi × µj -density of (Xi , Xj ) when 1 ≤ i < j ≤ 3, and fi for
the µi -density of Xi . It is shown in Nogales (2013b, Example 1) that the conditional distributions
M |M
Mi = P Xi |X1 : (Ω1 , A1 )≻−→(Ωi , Ai ), i = 2, 3, and L := P1 2 3 exist, where P1 = P X1 , and that a
density of L(ω3 , ·) with respect to µ2 is the map
Z
f12 (ω1 , ω2 )f13 (ω1 , ω3 )
dµ1 (ω1 )
ω2 7→
f1 (ω1 )f3 (ω3 )
Ω1
6
L is in fact the conditional distribution of a conditional distribution given another conditional distribution! So, when (Ω2 , A2 ) = (Rk , Rk ) and µ2 is the Lebesgue measure, we have that the conditional
expectation EP1 (M2 |M3 ) is the map
Z
Z
f12 (ω1 , x2 )f13 (ω1 , ω3 )
dµ1 (ω1 )dx2
ω3 7→
x2
f1 (ω1 )f3 (ω3 )
k
R
Ω1
For instance, let X = (X1 , X2 , X3 ) be a trivariate normal random variable with null mean and P1 the
X |X
marginal distribution of X1 . For i = 2, 3, consider the Markov kernel Mi = P1 i 1 , the conditional
distribution of Xi given X1 . It is shown in Nogales (2013b) that the conditional distribution L :=
M |M
P1 2 3 of the Markov kernel M2 given M3 with respect to P1 satisfy that L(x3 , ·) is the univariate
normal distribution of mean σ2 ρσ123 ρ13 x3 and variance σ22 (1 − ρ212 ρ213 ), where σi is the standard deviation
of Xi and ρij stand for the correlation coefficient of Xi and Xj . According to the previous result, the
conditional expectation of M2 given M3 is the random variable x3 7→ σ2 ρσ123 ρ13 x3 .
Moreover, it is easily checked that P1M2 = P X2 and EP1 (M2 ) = EP (X2 ).
Note that
Z
EP (M1 |M2 )dP M2 =
Z
Ω2
Ω2
=
Z
Z
xP M1 |M2 (ω2 , dx)dP M2 (ω2 )
Rk
xdP M2 ×M1 (ω2 , x)
Ω2 ×Rk
=
Z
xd P M2 ×M1
Rk
π
(x),
where π : Ω2 × Rk → Rk is the coordinatewise projection and M2 × M1 : (Ω, A)≻−→(Ω2 × Rk, A2 × Rk )
π
satisfies (M2 × M1 )(ω, A2 × A1 ) = M2 (ω, A2 ) · M1 (ω, A1 ), Ai ∈ Ai , i = 1, 2. But P M2 ×M1 = P M1 .
Hence
Z
Z
M2
=
xdP M1 (x) = EP (M1 ).
EP (M1 |M2 )dP
EP M2 (EP (M1 |M2 )) =
Rk
Ω2
This way we obtain the following corollary, which generalizes a known property of usual conditional
expectations.
Corollary 1. Let M1 and M2 be two Markov kernels as in the previous Definition 6. Then
EP M2 (EP (M1 |M2 )) = EP (M1 ).
We can have a representation of conditional expectations for Markov kernels in terms of conditional
expectations for random variables.
Theorem 2. If M1 is P -integrable,
EP (M1 |M2 ) = EP ⊗M2 (M̄1 |π2 ) where M̄1 : (Ω× Ω2 , A× A2 ) → Rk
R
is defined by M̄1 (ω, ω2 ) := Rk xM1 (ω, dx), and π2 is the second coordinatewise projection on Ω × Ω2 .
As a consequence of this representation theorem and Jensen’s Inequality, we have the next result.
Corollary 2. For every Z ∈ L2 (Ω2 , A2 , (P ⊗ M2 )π2 ), we have that
kM̄1 − EP (M1 |M2 )k22 ≤ kM̄1 − Zk22 ,
i.e.,
Z
(M̄1 (ω, ω2 ) − EP (M1 |M2 )(ω2 ))2 d(P ⊗ M2 )(ω, ω2 ) ≤
Ω×Ω2
Z
(M̄1 (ω, ω2 ) − Z(ω2 ))2 d(P ⊗ M2 )(ω, ω2 ),
Ω×Ω2
7
Subexample 1.1 (cont.): (Application to clinical diagnosis) Applying the preceding Corollary
to Subexample 1.1, writing a = Z(0) (a could represent the probability that the decision 0 is taken,
i.e., the test T discards the disease) and b = Z(1) (b could represent the probability that the decision
1 is taken, i.e., the test T confirms the disease), we obtain the following interpretation of predictive
values of a diagnostic test T :
(1 − P N V, P P V ) = arg min(a,b)∈R2 {[(1 − a)2 e + b2 (1 − e)](1 − θ) + [a2 (1 − s) + (1 − b)2 s]θ}.
Notice that, for a non-ill individual (i.e., when D = 0), the right decision will be (a0 , b0 ) = (1, 0),
and (1 − a)2 e + b2 (1 − e) is a weighted squared distance between (a, b) and the optimal point (1, 0)
on {D = 0}; the weights are e = P (T = 0|D = 0) and 1 − e = P (T = 1|D = 0) for the discrepancy
between a and a0 = 1, and that of b and b0 = 0, respectively, as can be expected. Analogously, for an
ill individual (i.e., when D = 1), the right decision is (a1 , b1 ) = (0, 1), and a2 (1 − s) + (1 − b)2 s is also
a properly weighted squared distance between (a, b) and the optimal point (0, 1) on {D = 1}.
Notice finally that, in the daily clinical practice, it is not known whether D = 0 or D = 1 and we
should choose (a, b) in such a way that its simultaneous distance to (1, 0) on {D = 0} and to (0, 1)
on {D = 1} reach a minimum; obviously, this simultaneous squared distance is weighted according to
the sizes of the subpopulations {D = 0} and {D = 1}.
4
Some statistical applications: extension to Markov kernels
of the Rao-Blackwell and the Lehmann-Scheffé theorems
Now, we position ourselves in a statistical context. Let (Ω, A, P) be a statistical experiment (i.e., P
is a family of probability measures on the measurable space (Ω, A)).
The theorems of Rao-Blackwell and Lehmann-Scheffé are central results of unbiased point estimation theory. We pursue in this section a version in the Markov kernel framework.
The concepts defined in the preceding sections can be extended to a statistical framework in a
standard way. The concept of sufficiency for Markov kernels is introduced in Heyer (1982, p.163).
Recall that, given a Markov kernel M1 : (Ω, A, P)≻−→(Ω1 , A1 ) and P ∈ P, the conditional probability
P (A|M1 ) of an event A ∈ A given M1 is defined as the Radon-Nikodym derivative d(IA · P )M1 /dP M1 ,
where IA · P denotes the measure defined on A by (IA · P )(B) = P (A ∩ B). In other words, P (A|M1 )
is the (equivalence class of) real random variable(s) on (Ω1 , A1 ) such that, for every A1 ∈ A1 ,
(1)
Z
A
Z
P (A|M1 )dP M1
Z Z
P (A|M1 )(ω1 )M1 (ω, dω1 )dP (ω)
=
M1 (ω, A1 )dP (ω) =
A1
Ω
A1
Definition 7. (Sufficiency of a Markov kernel) A Markov kernel M1 : (Ω, A, P)≻−→(Ω1 , A1 ) is said to
be sufficient if, for every A ∈ A, there exists a common version fA : (Ω1 , A1 ) → [0, 1] to the conditional
probabilities P (A|M1 ), P ∈ P.
Remarks. 1) The previous definition generalizes that of a sufficient statistic in the sense that a
statistic T1 is sufficient if, and only if, the corresponding kernel MT1 (ω, A1 ) = δT1 (ω) (A1 ) is sufficient.
Also, a sub-σ-field B ⊂ A is sufficient if, and only if, its corresponding kernel MB : (Ω, A)≻−→(Ω, B),
defined by MB (ω, B) := δω (B), is also.
2) Theorem 22.3 of Heyer (1982) shows that a Markov kernel M1 : (Ω, A, P)≻−→(Ω1 , A1 ) is sufficient if, and only if, the σ-field π1−1 (A1 ) is sufficient in the statistical experiment (Ω × Ω1 , A × A1 , {P ⊗
M1 : P ∈ P}), where π1 denotes the coordinatewise projection over Ω1 .
3) (Sufficiency of Markov kernels when densities are available) Suppose that P is dominated
by a σ-finite measure µ on (Ω, A) –µ is tipically the Lebesgue measure in the absolute continuous case and the counting measure in the discrete case–. Let fP be a µ-density of P ∈ P. Let
M1 : (Ω, A, P)≻−→(Ω1 , A1 ) be a Markov kernel and suppose that m1 : (Ω × Ω1 , A × A1 ) → [0, ∞[ is
8
a measurable function such that, for every ω ∈ Ω, m1 (ω, ·) is a µ1 -density of the probability measure
M1 (ω, ·), where µ1 is a σ-finite measure on (Ω1 , A1 ). It is readily shown that
d(P ⊗ M1 )
(ω, ω1 ) = m1 (ω, ω1 ) · fP (ω).
d(µ × µ1 )
According to the previous remark and the factorization theorem, the Markov kernel M1 is sufficient
if, and only if, there exist a measurable function h : (Ω × Ω1 , A × A1 ) → [0, ∞[ and, for each P ∈ P,
a measurable function gP : (Ω1 , A1 ) → [0, ∞[ such that
m1 (ω, ω1 ) · fP (ω) = gP (ω1 ) · h(ω, ω1 ),
∀ω, ω1 .
Here we introduce two examples, one discrete and one continuous, of sufficient Markov kernels not
associated to statistics.
Example 3. Let Ω = {1, 2, 3}, A the σ-field of all subsets of Ω, and P := {Pθ : θ ∈ [0, 1]}, where Pθ
assigns probability θ/3 to the points 1 and 2 and probability 1 − 2θ/3 to the point 3. The Markov
kernel M : (Ω, A)≻−→(Ω, A) defined by the stochastic matrix
1/3 2/3 0
1/3 2/3 0
0
0 1
is sufficient and is not associated to any statistic.
Example 4. Let (Ω, A) = (R+ , R+ ) and P = {Pθ : θ = 0, 1, 2, . . . }, where dPθ (x) = I[θ,θ+1[ (x) dx.
For x ≥ 0, we denote by M (x, ·) the uniform distribution on the interval [⌊x⌋, ⌊x⌋ + 1[, where ⌊x⌋
stands for the integer part of x. The Markov kernel M : (Ω, A)≻−→(Ω, A) is sufficient and is not
associated to any statistic.
Let us recall from Nogales (2013a) the generalization of the concept of completeness to Markov
kernels.
Definition 8. (Completeness of Markov kernels) A Markov kernel M1 : (Ω, A, P)≻−→(Ω1 , A1 ) is said
to be complete (respectively, boundedly complete) if, for every (respectively, bounded) real statistic
f : (Ω1 , A1 , {P M1 : P ∈ P}) → R,
EP M1 f = 0, ∀P ∈ P
=⇒
f = 0, P M1 -almost surely , ∀P ∈ P.
Remarks. 1) A Markov kernel M1 : (Ω, A, P)≻−→(Ω1 , A1 ) is (respectively, boundedly) complete if,
and only if, the σ-field π1−1 (A1 ) on the statistical experiment (Ω×Ω1 , A×A1 , {P ⊗M1 : P ∈ P}) is also,
where π1 denotes the coordinatewise projection over Ω1 , which in turn is equivalent to the (bounded)
completeness of π1 (see Nogales (2103a)). Moreover, if M1 is the Markov kernel corresponding to a
statistic T1 , then M1 is (boundedly) complete if, and only if, T1 is also.
2) (Completeness of Markov kernels when densities are available) Suppose that P is dominated by
a σ-finite measure µ on (Ω, A). Let fP be a µ-density of P ∈ P. Let M1 : (Ω, A, P)≻−→(Ω1 , A1 ) be a
Markov kernel and suppose that m1 : (Ω × Ω1 , A × A1 ) → [0, ∞[ is a measurable function such that,
for every ω ∈ Ω, m1 (ω, ·) is a µ1 -density of the probability measure M1 (ω, ·), where µ1 is a σ-finite
measure on (Ω1 , A1 ). It is readily shown that
d(P ⊗ M1 )
(ω, ω1 ) = m1 (ω, ω1 ) · fP (ω).
d(µ × µ1 )
According to the previous remark, the Markov kernel M1 is complete if and only if for every statistic
f : (Ω1 , A1 ) → R we have that
Z
f (ω1 )m1 (ω, ω1 )fP (ω)d(µ×µ1 )(ω, ω1 ) = 0, ∀P ∈ P =⇒ f = 0, (P ⊗M1 )π1 −c.s., ∀P ∈ P.
Ω×Ω1
9
Here we present two examples of complete Markov kernels not associated to statistics.
Example 5. Let Ω = {1, 2}, A = P(Ω) and P := {Pθ : θ ∈ [0, 1]}, where Pθ assigns probability θ to
the point 1 and probability 1 − θ to the point 2. The Markov kernel M : (Ω, A)≻−→(Ω, A) defined by
the stochastic matrix
p 1−p
q 1−q
is complete for p, q ∈ [0, 1] when p 6= q, and it is not associated to any statistic unless p, q ∈ {0, 1}.
Example 6. Let Ω = R+ , A = R+ and P := {Pθ : θ > 0}, where Pθ denotes the exponential
distribution of parameter θ. For x > 0, we denote by M (x, ·) the uniform distribution on the interval
[x, x + 1[. The Markov kernel M : (Ω, A)≻−→(Ω, A) is complete and is not associated to any statistic.
Now we are ready to obtain a first extension to Markov kernels of the theorem of Lehmann-Scheffé.
Theorem 6 yields a more general result. First, recall that an statistic T : (Ω, A, P) → Rk is said to be
an unbiased estimator of a function f : P → Rk whenever EP (T ) = f (P ), for all P ∈ P. T is said to
be a minimum variance estimator of f if it is unbiased and has less variance than any other unbiased
estimator of f .
Theorem 3. Let M1 : (Ω, A, P)≻−→(Ω1 , A1 ) be a sufficient and complete Markov kernel and T :
(Ω, A, P) → Rk be an unbiased estimator of a function f : P → Rk . If T is a measurable function of
M1 (i.e., there exists a measurable map S : (Ω1 , A1 ) → Rk such that MT = MS M1 ), then T is the
minimum variance unbiased estimator of f .
Now let us recall the definition of unbiased (randomized) estimator.
Definition 9. (Unbiased estimator) An unbiased estimator of a function f : P → Rk is a P-integrable
Markov kernel M : (Ω, A, P)≻−→(Rk , Rk ) such that
Z Z
xM (ω, dx)dP (ω) = f (P ), ∀P ∈ P
EP (M ) :=
Ω
Rk
Theorem 4. Let M1 : (Ω, A, P)≻−→Rk and M2 : (Ω, A, P)≻−→(Ω2 , A2 ) be Markov kernels. If M2 is
sufficient, then there exists a regular conditional probability P M1 |M2 of M1 given M2 which is independent of P ∈ P. There exists also a common version of the conditional expectations EP (M1 |M2 ),
P ∈ P; it will be denoted E(M1 |M2 ).
The next theorem extend to Markov kernels the Rao-Blackwell theorem.
Theorem 5. (Theorem of Rao-Blackwell generalized) Let M1 : (Ω, A, P)≻−→Rk be an estimator of
f : P → R and M2 : (Ω, A, P)≻−→(Ω2 , A2 ) be a sufficient Markov kernel for P. Then E(M1 |M2 ) is
an estimator of f with less convex risk than M1 . If the loss function is strictly convex then, given
P ∈ P, the risk at P of E(M1 |M2 ) is strictly less than the risk at P of M1 unless E(M1 |M2 )π2 = M̄1 ,
P ⊗ M2 -a.s., where M̄1 is defined as in Theorem 2. Finally, if M1 is unbiased, so is E(M1 |M2 ).
Remark. Since E(M1 |M2 ) is a statistic, this theorem shows that the class of non-randomized unbiased
estimators of f is complete in the sense that, for every randomized unbiased estimator M1 of f , there
exists a non-randomized unbiased estimator E(M1 |M2 ) with less convex risk than M1 . Note that this
assertion remains true if the assumption of unbiasedness is dropped. This result generalizes a similar
result when M2 is a statistic rather than a Markov kernel (for instance, see Pfanzagl (1994, p. 105)).
Theorem 6. (Theorem of Lehmann-Scheffé generalized) Let M1 : (Ω, A, P)≻−→Rk be an unbiased
estimator of f : P → Rk and M2 : (Ω, A, P)≻−→(Ω2 , A2 ) be a sufficient and complete Markov kernel
for P. Then E(M1 |M2 ) is the estimator of f which minimizes the convex risk among all unbiased
estimators of f .
10
5
A example of application of the generalized Lehmann-Scheffé
Theorem
In order to provide such an example, we need a sufficient and complete Markov kernel M that is
not associated to any statistic. We start with a sufficient Markov kernel M and describe a general
procedure to construct a greater family of probabilities for which M is complete and remains still
sufficient. This way, if M has finite mean f , the statistic E(M |M ) if the estimator of f which
minimizes the convex risk among all unbiased estimators of f .
The following results describe this procedure and allow us to construct, from Example 4, the desired
application example of Theorem 6. First we fix the adequate framework.
Let (Ω, A, P) be a statistical experiment dominated by a σ-finite measure µ. Let P ∗ be a privileged
dominatingP
probability (i.e. P ∗ is a probability measure on (Ω, A) such that P ≪ P ∗ and is of the
dP
form P ∗ = n 2−n Pθn for some suitable countable subfamily {Pθn : n ≥ 1} of P). Let rP ∈ dP
∗ be a
∗
P -density of P ∈ P.
Let M : (Ω, A, P)≻−→(Ω′ , A′ ) be a Markov kernel and suppose that µ′ is a σ-finite measure such
that, for every ω ∈ Ω, dM (ω, ω ′ ) = m(ω, ω ′ )dµ′ (ω ′ ), where m : (Ω×Ω′ , A×A′ ) → [0, ∞[ is measurable.
It is readily shown that
R
ω ′ )rP (ω)dP ∗ (ω)
dP M
d(P ⊗ M )
′
′
Ω m(ω,
R
=: rPM (ω ′ ).
(ω
)
=
(ω,
ω
)
=
r
(ω),
and
P
′ )dP ∗ (ω)
d(P ∗ ⊗ M )
m(ω,
ω
dP ∗ M
Ω
We are already prepared to obtain a sufficient and complete Markov kernel by the procedure above
mentioned.
Theorem 7. Let M : (Ω, A, P)≻−→(Ω′ , A′ ) be a sufficient Markov kernel and denote Q := Q(P) the
family of all probability measures Q on A such that Q ≪ P ∗ and M is sufficient for the extended
family P ∪ {Q}.
(a) The M is sufficient and complete for Q.
(b) Let n, k ∈ N and suppose (Ω′ , A′ ) = (Rn , Rn ). If M has finite moments of order k and
Qk := Q(P) denotes the subfamily of Q preserving this property of M , then M is sufficient and
complete for Qk .
(c) Q and its subfamilies Qk are convex.
Example 7. (Example 4, cont.) Let (Ω, A) = (R+ , R+ ) and P = {Pθ : θ = 0, 1, 2, . . . }, where
dPθ (x) = I[θ,θ+1[ (x) dx. For x ≥ 0, we denote by M (x, ·) the uniform distribution on the interval
[⌊x⌋, ⌊x⌋ + 1[, where ⌊x⌋ stands for the integer part of x. The Markov kernel M : (Ω, A)≻−→(Ω, A)
is sufficient and is not associated to any statistic. Notice that M has finite mean Eθ (M ) = θ + 1/2.
M|M
Write Lθ for the Markov kernel Pθ
. As a consequence of Proposition 2 of Nogales (2013b), given
x2 ≥ 0, the map
R∞
I[⌊x⌋,⌊x⌋+1[ (x1 ) · I[⌊x⌋,⌊x⌋+1[ (x2 ) dPθ (x)
R∞
x1 ≥ 0 7→ 0
I[⌊x⌋,⌊x⌋+1[ (x2 ) dPθ (x)
0
M|M=x
2
is the uniform distribution on the interval [⌊x2 ⌋, ⌊x2 ⌋ + 1[,
is a density of Lθ (x2 , ·). So, Pθ
and Eθ (M |M = x2 ) = ⌊x2 ⌋ + 1/2 that, as can be expected, does not depend on θ. Consequently,
we can write this conditional expectation as E(M |M ). By the previous theorem, M is sufficient and
complete for the family Q1 (P). Hence, the Theorem 6 applies and shows that E(M |M ) is the unbiased
estimator of the mean of M (in the family Q1 ) which minimizes the convex risk among all its unbiased
estimators.
It is possible to extend this example to the case of a sample of size n by using the next result.
Let M : (Ω, A, P)≻−→(Ω′ , A′ ) be a Markov kernel, where P is a family dimnated by a σ-finite
measure µ: let fP be a µ-density of P ∈ P. Let µ′ be a σ-finite measure on A′ and suppose that there
exists a map m : (Ω × Ω′ , A × A′ ) → [0, ∞[ such that dM (ω, ω ′ ) = m(ω, ω ′ )dµ′ (ω ′ ), for every ω ∈ Ω.
We know that
d(P ⊗ M )
(ω, ω ′ ) = m(ω, ω ′ ) · fP (ω), ∀ω, ω ′ .
d(µ × µ′ )
11
For n ∈ N, a Markov kernel
n
n
M (n) : (Ωn , An , {P n : P ∈ P})≻−→(Ω′ , A′ ),
is defined, for ~
ω ∈ Ωn , by
n
Y
dM (n) (~ω , ~ω ′ )
m(~ωi , ~ωi′ ).
:=
n
′
dµ
i=1
Hence, we have the following result.
Theorem 8. (a) If M is P-sufficient, then M (n) is {P n : P ∈ P}-sufficient.
(b) If M is P-complete, then M (n) is {P n : P ∈ P}-complete.
6
Proofs
Proof of Theorem 1. First note that there exists a regular conditional probability P M1 |M2 (see
Nogales (2013b)). It will be enough to show that, given A2 ∈ A2 ,
Z
Z
Z Z
M2 (ω, A2 )
xM1 (ω, dx)dP (ω) =
xP M1 |M2 (ω2 , dx)dP M2 (ω2 )
Rk
Ω
A2
But by definition of P M1 |M2 , for all A1 , A2 ,
Z
Z
M1 (ω, A1 )M2 (ω, A2 )dP (ω) =
Ω
i.e.,
Z
M2 (ω, A2 )
Ω
Rk
P M1 |M2 (ω2 , A1 )dP M2 (ω2 )
A2
Z
Rk
IA1 (x)M1 (ω, dx)dP (ω) =
Z
A2
Z
Rk
IA1 (x)P M1 |M2 (ω2 , dx)dP M2 (ω2 )
It follows in a standard way that, for any nonnegative or P M1 -integrable measurable function f :
Rk → Rm ,
Z Z
Z
Z
f (x)M1 (ω, dx)dP (ω) =
M2 (ω, A2 )
f (x)P M1 |M2 (ω2 , dx)dP M2 (ω2 )
Ω
Rk
A2
Rk
which gives the proof.
Proof of Theorem 2. Recall that P M2 = (P ⊗ M2 )π2 . Now we define a Markov kernel
M̂1 : (Ω × Ω2 , A × A2 )≻−→Rk by M̂1 ((ω, ω2 ), B) = M1 (ω, B); M̂1 is the extension to Ω × Ω2 of M1 .
We will prove that (P ⊗ M2 )M̂1 |π2 is a regular conditional P -probability of M1 given M2 . We will
use the following result from Nogales (2013b): “If T2 : (Ω, A) → (Ω2 , A2 ) is a random variable and
K2 (ω, A2 ) = δT2 (ω) (A2 ) is its corresponding Markov kernel then, writing P M1 |T2 := P M1 |K2 , we have
P M1 |T2 (·, A1 ) = EP (M1 (·, A1 )|T2 ).” Applying this result in the probability space (Ω × Ω2 , A × A2 , P ⊗
M2 ), we have that, for ω2 ∈ Ω2 and B ∈ Rk ,
(2)
(P ⊗ M2 )M̂1 |π2 (ω2 , B) = EP ⊗M2 (M̂1 (·, B)|π2 = ω2 )
Hence, given A2 ∈ A2 ,
Z
Z
(P ⊗ M2 )M̂1 |π2 =ω2 (B)d(P ⊗ M2 )π2 (ω2 )
(P ⊗ M2 )M̂1 |π2 =ω2 (B)dP M2 (ω2 ) =
A2
A2
Z
EP ⊗M2 (M̂1 (·, B)|π2 = ω2 )d(P ⊗ M2 )π2 (ω2 )
=
A2
Z
M1 (ω, B)d(P ⊗ M2 )(ω, ω2 )
=
ZΩ×A
Z 2
M1 (ω, B)M2 (ω, dω2 )dP (ω)
=
Ω A2
Z
=
M1 (ω, B)M2 (ω, A2 )dP (ω)
Ω
12
which proves that
(P ⊗ M2 )M̂1 |π2 = P M1 |M2
(3)
Moreover, (2) can be rewritten in the form
Z
Z
M̂1 |π2 =ω2
(dx) = EP ⊗M2
IB (x)(P ⊗ M2 )
Rk
IB (x)M1 (·, dx) π2 = ω2
Rk
It follows that, for a nonnegative or integrable measurable function f : Rk → Rm ,
Z
Z
f (x)(P ⊗ M2 )M̂1 |π2 =ω2 (dx) = EP ⊗M2
f (x)M1 (·, dx) π2 = ω2
Rk
Rk
In particular, for m = k and f (x) = x,
Z
Z
M̂1 |π2 =ω2
(dx) = EP ⊗M2
x(P ⊗ M2 )
Rk
xM1 (·, dx) π2 = ω2
Rk
Using (3), we obtain
EP (M1 |M2 ) = EP ⊗M2 (M̄1 |π2 ).
Proof of Theorem 3. Let T ′ : (Ω, A) → R be an arbitrary unbiased estimator of f and
denote Te′ (ω, ω1 ) := T ′ (ω). Hence Te′ is an unbiased estimator of f in the statistical experiment
(Ω × Ω1 , A × A1 , {P ⊗ M1 : P ∈ P}). Since the coordinatewise projection π1 is sufficient, there exists
a version of the conditional expectation X ′ of Te′ given π1 which is independent of P ∈ P. The
Rao-Blackwell theorem shows that X ′ ◦ π1 has less covariance matrix than Te′ .
Since MT = MS M1 , we have that, for all Borel set B ∈ Rk and all ω ∈ Ω,
Z
IB (S(ω1 ))M1 (ω, dω1 )
IB (T (ω)) =
Ω1
Hence, for all ω ∈ Ω, S = T (ω), M1 (ω, ·)-a.s. It follows that
Te(ω, ω1 ) = (S ◦ π1 )(ω, ω1 ),
{P ⊗ M1 : P ∈ P} − a.s.
where Te(ω, ω1 ) = T (ω), for all ω ∈ Ω. So, S is a conditional expectation of Te given π1 for all P ∈ P.
The completeness of π1 shows that S ◦ π1 = X ′ ◦ π1 , {P ⊗ M1 : P ∈ P}-a.s., and this finish the
proof.
Proof of Theorem 4. According to Heyer (1982, Theorem 22.3), M2 is sufficient if, and only
if, the coordinatewise projection π2 : (Ω × Ω2 , A × A2 , {P ⊗ M2 : P ∈ P}) → (Ω2 , A2 ) is sufficient.
Landers and Rogge (1972, Theorem 7) shows the existence of a common regular conditional probability
on Rk given π2 . The result follows from this fact and the following representation of the conditional
distribution of M1 given M2 obtained in the proof of Theorem 2:
P M1 |M2 (ω2 , B) = (P ⊗ M2 )M̂1 |π2 (ω2 , B) = EP ⊗M2 (M̂1 (·, B)|π2 = ω2 )
The second assertion follows from this and Theorem 1.
Proof of Theorem 5. E(M1 |M2 ) is well defined by the previous theorem and it is an unbiased
estimator of f by Corollary 1. Moreover, if W : P × Rk → [0, ∞[ is a convex loss function (i.e.,
W (P, ·) is a convex function for every P ∈ P) then applying the Jensen inequality (see Pfanzagl (1994,
Theorem 1.10.11)), we obtain from Theorem 1 that
Z
M1 |M2
(·, dx)
W (P, EP (M1 |M2 )) = W P,
xP
Rk
Z
≤
W (P, x)P M1 |M2 (·, dx) = EP (W (P, M1 )|M2 ), P M2 − a.s.
Rk
13
where W (P, M1 ) denotes the kernel W (P, ·)M1 . The result follows by integration with respect to P M2 .
Corollary 1 completes the proof in the unbiased case.
Proof of Theorem 6. If the Markov kernel M1′ : (Ω, A, P)≻−→Rk is an arbitrary unbiased
estimator of f then, according to the previous theorem, X1′ := E(M1′ |M2 ) is a nonrandomized unbiased
estimator of f with less convex risk than M1′ . Moreover X1 := E(M1 |M2 ) is an unbiased estimator of f ;
so EP M2 (X1 −X1′ ) = 0 for all P ∈ P. Since M2 is complete, we have that X1 = X1′ , {P M2 : P ∈ P}-a.s.
So X1 has less convex risk than M1′ .
Proof of Theorem 7. (a) Let π ′ : (Ω × Ω′ , A × A′ ) → (Ω′ , A′ ) the coordinatewise projection.
According to Theorem 22.3 of Heyer (1982), M is suffcient for Q if, and only if, the σ-field π ′−1 (A′ )
is sufficient for the family {Q ⊗ M : Q ∈ Q}. Given Q ∈ Q, M is sufficient for P ∪ {Q}. So, being
P ∗ ⊗ M a privileged dominating probability of P ∪ {Q}, there exists sQ : (Ω′ , A′ ) → [0, ∞[ such that
d(Q ⊗ M )
(ω, ω ′ ) = sQ (ω ′ ).
d(P ∗ ⊗ M )
an the theorem of factorization shows that M is suffcient for Q.
To prove the completeness, we proceed by reduction to the absurd assuming the existence of
g : (Ω′ , A′ ) → R such that EQM (g) = 0, for all Q ∈ Q, and QM
0 (g 6= 0) > 0 for some Q0 ∈ Q. Without
M
∗M
, it follows
loss of generality we can suppose that Q0 (g > 0) > 0. Since QM
0 is dominated by P
M
M
that P ∗ (g > 0) > 0. Define α := P ∗ (g > 0)−1 and s′1 := α · I{g>0} . Hence s′1 is a density with
respect to P ∗ M of a probability measure Q′1 on A′ . Write Q for the probability measure on A × A′
whose density with respect to P ∗ ⊗ M is s1 (ω, ω ′ ) := s′1 (ω ′ ). Finally, consider the probability measure
Q1 on A defined by
Z Z
m(ω, ω ′ ) · s′1 (ω ′ ) dµ′ (ω ′ )dP ∗ (ω).
Q1 (A) = Q(A × Ω′ ) =
A
Ω′
It is obvious that Q = Q1 ⊗ M (in fact, if π denotes the coordinatewise projection over Ω, we have
′
that Q1 = Qπ and M = Qπ |π ). Besides, Q1 ∈ Q, because Q1 ≪ P ∗ and
d(Q1 ⊗ M )
(ω, ω ′ ) = s′1 (ω ′ ).
d(P ∗ ⊗ M )
Finally,
EQM
(g) =
1
Z
Ω′
g(ω ′ ) · s′1 (ω ′ ) dP ∗ M (ω ′ ) = α
Z
g(ω ′ ) dP ∗ M (ω ′ ) > 0,
{g>0}
and this contradiction finish the proof of (a).
(b) It is enough to note in the proof of (a) that Q1 ∈ Qk . But, given 1 ≤ i ≤ n,
Z
Z
k
|xi |k dQM
(x)
=
α
|xi |k dP ∗ M (x) < ∞
1
Rn
{g>0}
because P ∗ ∈ Qk .
(c) Trivial.
Proof of Theorem 8. (a) It is known that M is sufficient if, and only if, it is possible to factorize
m(ω, ω ′ ) · fP (ω) = gP (ω ′ ) · h(ω, ω ′ ),
∀ω, ω ′ , P.
So, M being sufficient, we have that
n
n
Y
Y
d(P n ⊗ M (n) )
′
′
gP (~ωi′ ) · h(~ωi , ~ωi′ ),
m(~
ω
,
ω
~
)
·
f
(~
ω
)
=
(~
ω
,
ω
~
)
=
i
P
i
i
d(µn × µ′ n )
i=1
i=1
and this proves the sufficiency of M (n) .
14
(b) The completenes of M means that f = 0, P M -c.s., for all P ∈ P, when f is a real statistic on (Ω′ , A′ .{P M : P ∈ P}) such that EP M (f ) = 0. Assuming that, we will show that M (n) :
n
n
(Ωn , An , {P n : P})≻−→(Ω′ , A′ ) is also complete.
(n)
Let hn be a real statistic on (Ω′n , A′n .{(P n )M : P ∈ P}) such that E(P n )M (n) (hn ) = 0 for every
(n)
P ∈ P. To prove that hn = 0 (P n )M -c.s. for evey P , it will be enough to show that, for every
A′1 , . . . , A′n ∈ A′ and P ∈ P, we have that
E(P n )M (n) (hn · IQni=1 A′i ) = 0.
For this we will follow a recursive process similar to that used in Landers and Rogge (1976) introducing
the events A′i one by one.
Para ello seguiremos un proceso recursivo similar al utilizado en Landers y Rogge (1976) introduciendo los sucesos A′i uno a uno. In fact, the process will be illustrated only with A′1 , since the rest
is entirely analogous.
Notice that the map
Z Y
n
m(~ωi , ~ωi′ )fP (~ωi )dµn (~ω )
ω ′ ∈ Ω′n 7→
~
Ωn i=1
is a µ′n -density of (P n )M
(n)
. So, it is supposed that
0 = E(P n )M (n) (hn ) =
Z
hn (~ω ′ )
Z
Ωn
Ω′n
n
Y
m(~ωi , ~ωi′ )fP (~ωi )dµn (~ω )dµ′n (~ω ′ ).
i=1
Hence, according to Fubini’s Theorem,
Z
Z
0=
Hn (~
ω1′ )
m(~ω1 , ~ω1′ )fP (~ω1 )dµ(~ω1 )dµ′ (~ω1′ ),
Ω′
Ω
where
Hn (~ω1′ )
:=
Z
′
hn (~
ω)
Ω′n−1
Z
n
Y
Ωn−1 i=2
m(~ωi , ~ωi′ )fP (~ωi )dµn−1 (~ω2 , . . . , ~ωn )dµ′n−1 (~ω2′ , . . . , ~ωn′ ).
That is to say, 0 = EP M (Hn ) for each P ∈ P. M being complete, we obtain Hn = 0 P M -c.s., for all
P ∈ P. It follows that, for every A′1 ∈ A′ and P ∈ P,
Z
Hn (~ω ′ )dP M (~ω ′ ),
0=
A′1
as we wanted to prove.
References:
• Blackwell, D.: Conditional expectation and unbiased sequential estimation, Ann. Math. Statist.
18, 105-110 (1947)
• Dellacherie, C., Meyer, P.A.: Probabilities and Potentiel C, North-Holland, Amsterdam (1988)
• Heyer, H.: Theory of Statistical Experiments, Springer, Berlin (1982)
• Kolmogorov, A.N.: Grundbegriffe der Wahrscheinlichkeitsrechnung, Springer, Berlin (1933).
• Landers, D. and Rogge, L.: A note on completeness, Scandinavian J. Statist. 3, 139 (1976).
• Lehmann, E.L., Scheffé, H.: Completeness, similar regions, and unbiased estimation, Sankhyā
10, 305-340; 15, 219-236; Correction 17, 250 (1950, 1955, 1956)
15
• Nogales, A.G.: On Independence of Markov Kernels and a Generalization of Two Theorems of
Basu, Journal of Statistical Planning and Inference 143, 603-610 (2013a)
• Nogales, A.G.: Existence of Regular Conditional Probabilities for Markov kernels, Statistics and
Probability Letters 83, 891-897 (2013b)
• Pfanzagl, J.: Parametric Statistical Theory, de Gruyter, Berlin (1994)
• Rao, C.R.: Information and the accuracy attainable in the estimation of statistical experiments,
Bull. Calcutta Math. Soc. 37, 81-91 (1945)
16
| 10math.ST
|
1
Relation between financial market structure and the real
economy: comparison between clustering methods
arXiv:1406.0496v2 [q-fin.ST] 21 Jan 2015
Nicoló Musmeci1 , Tomaso Aste2,3,∗ , T. Di Matteo1
1 Department of Mathematics, King’s College London, The Strand, London, WC2R 2LS,
UK
2 Department of Computer Science, UCL, Gower Street, London, WC1E 6BT, UK
3 Systemic Risk Centre, London School of Economics and Political Sciences, London,
WC2A2AE, UK
∗ E-mail: [email protected]
Abstract
We quantify the amount of information filtered by different hierarchical clustering methods on correlations between stock returns comparing the clustering structure with the underlying industrial activity
classification. We apply, for the first time to financial data, a novel hierarchical clustering approach,
the Directed Bubble Hierarchical Tree and we compare it with other methods including the Linkage and
k-medoids. By taking the industrial sector classification of stocks as a benchmark partition, we evaluate
how the different methods retrieve this classification. The results show that the Directed Bubble Hierarchical Tree can outperform other methods, being able to retrieve more information with fewer clusters.
Moreover, we show that the economic information is hidden at different levels of the hierarchical structures depending on the clustering method. The dynamical analysis on a rolling window also reveals that
the different methods show different degrees of sensitivity to events affecting financial markets, like crises.
These results can be of interest for all the applications of clustering methods to portfolio optimization
and risk hedging.
Introduction
Correlation-based networks have been extensively used in Econophysics as tools to filter, visualise and
analyse financial market data [1–8]. Since the seminal work of Mantegna on the Minimum Spanning Tree
(MST) [1] they have provided insights into several aspects of financial markets including financial crises
[9–15]. The MST is strictly related [16] to a hierarchical clustering algorithm, namely the Single Linkage
(SL) [17]. Starting from a set of elements (e.g., stocks) and a related distance matrix (e.g., a convenient
transformation of the stocks correlation matrix [1]), the SL performs an agglomerative algorithm that ends
up with a tree (dendrogram) that arranges the elements into a hierarchical structure [16]. The filtering
procedure linked to MST and SL has been succesfully applied to improve portfolio optimization [10].
Another hierarchical clustering method, the Average Linkage (AL), has been shown to be associated to a
slightly different version of spanning tree [18], called Average Linkage Minimum Spanning Tree. Another
variant of Linkage methods, not associated to a spanning tree representation, is the Complete Linkage
(CL) [17].
The MST is the first but not the only correlation-based filtered network studied in the literature. In
particular the Planar Maximally Filtered Graph (PMFG) is a further step from the MST, that is able
to retain a higher amount of information [3, 4, 19], having less strict topological constraint allowing to
keep a larger number of links. The PMFG has been proven to have interesting practical applications, in
particular in the field of investment strategies to hedge risk [20].
Since the MST has associated a clustering method, and the PMFG is a generalization of the MST, it
could be raised the question whether the PMFG provides a clustering method that exploits this higher
amount of information. In a recent work [21] it has been shown that this is the case: the Directed Bubble
2
Hierarchical Tree (DBHT) is a novel hierarchical clustering method that takes advantage of the topology
of the PMFG yielding a clustering partition and an associated hierarchy. (For the DBHT algorithm refer
to supplementary material of [21] or, for a slightly modified version, to [22].) The approach is completely
different from the agglomerative one adopted in the Linkage methods: the idea of DBHT is to use the
hierarchy hidden in the topology of a PMFG, due to its property of being made of three-cliques [21, 23].
In [21] the DBHT hierarchical clustering has been applied to synthetic and biological data, showing that
it can outperform many other clustering methods. Since DBHT exploits the topology of the correlation
network it can be viewed as an example of community detection algorithm in graphs [24].
In this paper we present the first application of DBHT to financial data. To this purpose we have
analysed the correlations among log-returns of N = 342 US stock prices, across a period of 15 years
(1997-2012). We have studied the structure of the clustering and we have compared the results with
other clustering methods, namely: the single Linkage, the average Linkage, the complete Linkage and the
k-medoids [25] (a partitioning method strictly related to the k-means [26]). The perspective of our study
focuses not only on the clusterings, but on the entire hierarchies associated to those clusterings, covering
all the different levels of the hierarchical structures. We have also studied the dynamical evolution of
these structures, describing how the hierarchies change with time. The dynamical perspective is crucial
for applications, in particular for what concerns hedging risk and portfolio optimization. For this reason
we have given a particular attention to the effects of financial crises on the hierarchical structures,
highlighting differences among the clustering methods.
Another aspect we have focused on is the role of the market mode in shaping the clustering structure.
To this aim we have carried out our analyses also on log-returns detrended of the market mode [27]: this
procedure has been proven to provide a more robust clustering that can carry information not evident in
the original correlation matrix [27]. The comparison between the detrended and non-detrended correlation
structures can shed light on the effect of common factors in the collective market dynamics especially
during crisis.
In order to compare quantitatively the amount of information retrieved by the different hierarchical clustering methods, we have taken the Industrial Classification Benchmark (ICB) as a benchmark
community partition for the stocks and then we have compared it with the output of each clustering
method. The idea is to use the degree of similarity between the ICB and the clustering as a proxy for
the amount of information filtered by the methods. The degree of similarity is measured by using tools
as the Adjusted Rand Index [28] and the Hypergeometric hypothesis test [29]. This is not the first work
comparing correlation-based clusterings and industrial sector classification; however to our knowledge
the comparison has been performed only qualitatively so far [30], with the exception of ref. [27] where
however only one clustering method is analysed. In Ref. [31] the authors have compared quantitatively
different methods in terms of amount of filtered information: yet this comparison was performed without looking at the industrial sector classification and by assuming a multivariate Gaussian distribution
for the stocks returns [16]. Our approach is instead model-free. This is a relevant improvement since
multivariate Gaussian models are known to be inaccurate to describe stocks returns [32, 33] and, more
generally, the correlation-based networks obtained from real data have been found to be incompatible
with some widespread models for asset returns [34].
The paper is organized as follows. In Section “Methods” we describe the main tools we have used
to carry out the analyses. In Section “Dataset and preliminary analyses” we present the dataset and
some preliminary empirical analyses on it. In Section “Static analysis” we perform a set of analyses on
correlations and clusterings calculated by taking the whole 15 years period as time window, hence ending
up with only one hierarchical structure of dependences for each method. In particular we compare the
clustering compositions in terms of ICB supersectors for different clustering methods and we measure the
similarity between clusterings and ICB partition. In Section “Dynamical analysis” we perform instead
analyses using a dynamical approach, with moving time windows. In Section “Discussion” we discuss the
results and the future directions of our work. More details on the clustering methods we have analysed
3
are reported in the Supplementary Information (SI).
Methods
To investigate and compare the DBHT with the other clustering methods we have performed a series of
analyses that aim at describing different aspects of each clustering, from their structure to the economic
information they contain. Here we introduce and describe the tools we have used.
Clustering methods
In this paper we compare four different clustering methods, namely: Single Linkage (SL), Average Linkage
(AL), Complete Linkage (CL), Directed Bubble Hierarchical Tree (DBHT), k-medoids. All these methods
require a suitable distance measure and DBHT also requires an associated similarity measure. In this
paper we use the Pearson’s correlation coefficientspρij between pairs of stocks as similarity measure and
we use the associated Euclidean distance Dij = 2(1 − ρij ); ρ and D are N × N symmetric matrices
with, respectively, all ones and all zeros on the diagonal. Let us here briefly list the main features and
requirements of these clustering methods.
• Single Linkage (SL), is a hierarchical clustering algorithm [17]. Given the distance matrix, it
starts assigning to each object its own cluster, and then at each step merges together the least
distant pair of clusters, until only one cluster remains. At every step the distance between clusters
A and B is updated taking the minimum distance between elements inside the clusters:
dA,B =
min
a∈A,b∈B
D(a, b)
(1)
SL is called an agglomerative clustering, since it begins with a partition of N clusters and then
proceeds merging them. The final output is a dendrogram, that is a tree showing the hierarchical
structure. The distance measure in this dendrogram is an ultrametric distance (see for instance [1]).
A cluster partition can be obtained by choosing the number of clusters (that is therefore a free
parameter) and cutting the dendrogram at the appropriate level. The SL algorithm is strictly
related to the one that provides a Minimum Spanning Tree (MST) [16]. The MST is a tree graph,
it contains exactly N − 1 links. There is therefore a strict relation between the two tools. However
the MST retains some information that the SL dendrogram throws away [16] .
• Average Linkage (AL) [17] is a hierarchical clustering algorithm which is constructed in the same
way as the SL, but with Eq. 17 replaced by:
dA,B = mean D(a, b)
a∈A,b∈B
(2)
• Complete Linkage (CL) [17] is another variant of SL, where Eq. 17 is replaced by:
dA,B =
max D(a, b)
a∈A,b∈B
(3)
• Directed Bubble Hierarchical Tree (DBHT) [21], is instead a novel hierarchical clustering
method that exploits the topological property of the PMFG (Planar Maximally Filtered Graph) in
order to assign the clustering. The PMFG is a generalization of the MST, that is included in the
PMFG as a subgraph. It can be constructed by recursively joining together with an edge nodes with
the smallest distance restricting the procedure to only edges that do not violate graph planarity
condition. The PMFG contains the MST as a sub-graph but it retains a larger number of links than
4
the MST (3(N − 2) instead of N − 1). The basic elements of a PMFG are three-cliques (subgraphs
made of three nodes all reciprocally connected). The DBHT exploits this topological structure,
and in particular the distinction between separating and non-separating three-cliques, to identify a
clustering partition of all nodes in the PMFG [21].
The Linkage algorithms look at the sorted list of distances Dij , build the dendrogram by gathering
subsets of stocks with lowest distances and then find the clustering from the dendrogram after
choosing the number of clusters as free parameter. The DBHT instead reverses this order: first
of all the clusters are identified by means of topological considerations on the planar graph, then
the hierarchy is constructed both inter-clusters and intra-clusters. The difference involves therefore
both the kind of information exploited and the methodological approach.
• k-medoids is a partitioning clustering method closely related to k-means [26]. It takes the number
of clusters Ncl as an input. The algorithm is the so called Partitioning Around Medoids (PAM),
and is constructed in the following steps: 1) select randomly Ncl “medoids” among the N elements;
2) assign each element to the closest medoid; 3) for each medoid, replace the medoid with each
point assigned to it and calculate the cost of each configuration; 4) choose the configuration with
the lowest cost; 5) repeat 2)-4) until no changes occur. This method, differently from previous, is
not a hierarchical method and therefore does not provide a dendrogram but only a partition.
Measuring clustering heterogeneity: disparity
Different clustering methods can differ greatly in the clustering output even under the same input conditions. Some clustering methods might provide a subdivision into a few very large clusters and many
small clusters, whereas other might display a more homogeneous distribution. In order to characterize
the distribution of the cluster cardinality with a single quantity we have calculated for each clustering
partition the coefficient of variation:
σS
,
(4)
y=
hSi
where σS is the standard deviation
s
σS =
X
1
(Sa − hSi)2 ,
Ncl − 1 a
(5)
and the normalization factor hSi is the average
hSi =
1 X
Sa ,
Ncl a
(6)
with Sa being the cardinality (number of stocks in) cluster a and Ncl the number of clusters. In the
limit of homogeneous arrangement of stocks among the clusters (i.e. each cluster has the same number of
stocks), we obtain σS = 0 and then y = 0. The higher is the degree of heterogeneity in the distribution
of sizes, the higher is σS and therefore y. In the following we have used the expression “disparity” to
refer to y, in order to stress the fact that we use it as a measure of heterogeneity in clusters’ sizes.
Measuring clustering similarity: Adjusted Rand Index
To measure the amount of economic information in each correlation based clustering we have used the
Adjusted Rand Index (Radj ) [28] which is a tool conceived to compare different clusterings on the same
set of items [35]. An industrial sector classification is indeed nothing but a partition in communities of
5
the N stocks. Therefore we can take the similarity between clustering and industrial sector classification
as a proxy for the information detected by the clustering method.
In particular, given two clusterings on the same set of items, Radj returns a numerical value equal
to 1 for identical clusterings and to 0 for clusterings completely independent which are undistinguishable
from a random choice.
The idea behind this measure is to calculate the number of pairs of objects that are in the same
cluster in both clusterings, and then to compare this number with the one expected under the hypothesis
of independent clusterings. Specifically, and following the notation of [35], let us call X the set of
the N objects (stocks, in our case). Let us call Y = {Y1 , ..., Yk } a clustering which is a partition
of X into communities which are non-empty disjoint subsets of X such that their union equals X:
X = Y1 ∪ ... ∪ Yk [35]. Let us also consider another different clustering Y 0 , containing l clusters. We call
“contingency table” the matrix M = {mij } with coefficients
mij ≡ |Yi ∩ Y 0 j |,
(7)
i.e. the number of objects in the intersection of clusters Yi and Y 0 j . Let us call a the number of pairs
of objects that are in the same cluster both in Y and in Y 0 , and b the number of pairs that are in two
different clusters in both Y and Y 0 . Then the Rand Index is defined as the sum of a and b, normalized
by the total number of pairs in X:
k X
l
X
mij
2(a + b)
=
.
R(Y, Y ) ≡
N (N − 1)
2
i=1 j=1
0
(8)
We then use, as null hypothesis associated to two independent clusterings, the generalized hypergeometric
distribution and define the Adjusted Rand Index as the difference between the Rand Index and its mean
value under the null hypothesis, normalized by the maximum that this difference can reach:
Pk Pl
mij
− t3
i=1
j=1
2
0
Radj (Y, Y ) ≡
,
(9)
1
2 (t1 + t2 ) − t3
where
t1 =
k
X
|Yi |
i
2
, t2 =
l
X
|Y 0 j |
j
2
, t3 =
2t1 t2
.
N (N − 1)
(10)
We have Radj ∈ [−1, 1] , with 1 correspondent to the case of identical clusterings and 0 to two completely
un-correlated clusterings. Negative values instead show anti-correlation between Y and Y 0 (that is, the
number of pairs classified in the same way by Y and Y 0 is less than what expected assuming a random
overlapping between the two clusterings).
Hypergeometric test for cluster-industry overexpression
The Adjusted Rand Index provides an overall measure of similarity between the clustering partition and
the industrial classification. In order to analize how much each industrial sector is retrieved by each
cluster we must look at the stocks respectively within a given cluster and a given sector and measure
the number of stocks in common. If the percentage of stocks in common between a cluster and an
industrial sector is sensitively higher than what expected from a random overlapping of communities we
say that the cluster overexpresses a specific sector. To quantify such overexpression we use a statistical
one-tail hypothesis test, where the null hypothesis is the Hypergeometric distribution which describes
the probability that by random chance two communities of given sizes have in common k objects over a
total of N [36, 37]. In particular, let us call Yi a cluster in our clustering and Yj0 a sector. We want to
6
verify whether Yi overexpresses Yj0 . If k is the number of stocks in common between Yj0 and Yi , and |Yi | ,
|Yj0 | are the cardinalities of the cluster and the sector respectively, then the Hypergeometric distribution
is [36]:
|Y 0 | N −|Yj0 |
j
P (X = k) =
k
|Yi |−k
N
|Yi |
.
(11)
This is the null hypothesis for the test: to be distinguishable by a random overlap the number k of stocks
in common must be significantly different from a random overlap and therefore P (X = k) must be small.
If P (X = k) is less than the significance level, then it is said that the test is rejected. If the test is not
rejected, then it means that we cannot reject the hypothesis that the k stocks in Yi coming from a sector
Yj0 are picked up just by chance, without any preference for that sector. If instead the test is rejected,
we conclude that the cluster Yi overexpresses the sector Yj0 . We have chosen a significance level of 1%,
together with the Bonferroni correction for multiple tests, which reduces remarkably the significance level
of each test [36] (more details in Section “Industries overexpression”).
Dataset and preliminary analyses
The correlation structure studied in this paper concerns N = 342 stocks from the New York Stock
Exchange (NYSE). A complete description of the dataset is in Supplementary Information (SI). We have
analysed the closing daily prices Pi (t) with i = 1, ..., N , during the time between 1 January 1997 to 31
December 2012 (4026 trading days). From the prices, we have calculated the daily log-returns [32, 33]:
ri (t) ≡ log(Pi (t)) − log(Pi (t − 1)).
(12)
From the set of N log-return time series over a time window T = [tstart , tend ] we have then calculated
the N × N correlation matrix ρ(T ), whose elements are given by the Pearson estimator [38]:
hri (t)rj (t)iT
,
ρij (T ) = q
2
[hri (t)iT − hri (t)i2T ][hrj2 (t)iT − hrj (t)i2T ]
(13)
where h...iT represents the average over the time window
T . The clustering analysis is then carried out
p
on the distance matrix D, with elements Dij (T ) = 2(1 − ρij (T )). For the analysis on moving windows
we used an exponentially smoothed version of the Pearson estimator in Eq.13, where terms in the average
are multiplied by a weight wt = w0 exp( t−tθend ) with t ∈ T according to their temporal distance from
the last trading time tend in the window T . This exponential smoothing scheme [39] allows to mitigate
excessive sensitiveness to outliers in remote observations. The parameter θ has been set to θ = T /3
according to previously established criteria [39].
By using this moving time window approach we have performed a set of preliminary analyses on the
average correlation of our set of stocks, looking in particular at the 2007-2008 financial crisis. Specifically,
we have considered a set of n = 100 overlapped time windows Tk (k = 1, ..., n) of length L = 1000 trading
days (four years) with 30 days shift between subsequent windows. The average correlation hρ(Tk )iij on
these windows is shown in Fig. 1 (blue circles) for L = 1000 and n = 100. To test robustness, we have
verified that the results are similar also for other window sizes, namely L = 750 and L = 1250.
We have also studied the detrended log-returns, i.e. log-returns subtracted of the average return over
all the stocks. Specifically, following [27], we have considered a single factor model for each stock i:
ri (t) = αi + βi I(t) + ci (t),
(14)
P
where the common market factor I(t) is the market average return, I(t) = N1 i ri (t) and the residuals,
ci (t), are the log-returns detrended by the market mode. After estimating the coefficients αi and βi with a
7
Average Correlation
0.35
hρR (Tk )iij
hρ(Tk )iij
0.3
0.25
0.2
0.15
2000
2002
2004
2006
t
2008
2010
2012
Figure 1. Demonstration that the average correlation evolves during time with large
changes during periods of market instability. The figure reports the average correlation for each
time window Tk with k = 1, ..., n (n = 100, each time window has length L = 1000 trading days), for
both non-detrended (blue circles) and detrended log-returns (green squares). The average correlation is
highly reduced by detrending the market mode.
linear regression, the residuals ci (t) can be calculated and used to evaluate the new correlation matrix [27].
We denote this matrix, estimated in the time window Tk with ρR (Tk ). We refer to the analyses based on
this kind of correlation matrix as the “detrended case”. These detrended correlation matrices are worth
analyzing since they have been found to provide a richer and more robust clustering [27] that can carry
information not evident in the original correlation matrix [40]. In this paper we have carried the analyses
using both detrended and non-detrended log-returns, compared the two and looked for differences that
might highlight the effect of common factors on the market correlation structure.
In Fig. 1 it is shown the average correlation for these detrended correlation matrices, i.e. hρR (Tk )iij ,
compared to the average correlation for the non-detrended correlation matrices hρ(Tk )iij . As one can
see, the subtraction of the market mode decreases by about 50% the average level of correlation, pointing
out the important role of the market factor in the correlation structure. However, we can still observe
the increase correspondent to the financial crisis in 2007-2008. Moreover, and interestingly, the level of
correlation reduces after a peak in 2009, unlike the non-detrended case. This fact suggests that, although
the market mode plays an important role in terms of average amount of correlation, yet the peak of the
last financial crisis seems not to be only a global market trend. We therefore suggest that it could involve,
to some extent, the internal dynamics among stocks that remains after the subtraction.
Results
Static analysis
DBHT clusters composition
In this section we present results for the PMFG and DBHT clustering method applied to the set of data
described in the previous section. In particular we have computed the PMFG and the correspondent
8
Figure 2. Visualization of the Planar Maximally Filtered Graph (PMFG) and DBHT
clusters, for both non-detrended and detrended log-returns. a) PMFG calculated on the entire
period 1997-2012, using non-detrended log-returns. Stocks of the same color belong to the same DBHT
cluster. b) PMFG calculated on the same data as in a), but using detrended log-returns. Stocks of the
same color belong to the same DBHT cluster.
DBHT clustering in the time period from 1997 to 2012 and we plot it in Fig. 2 a) where we highlight,
with the same color, stocks belonging to the same DBHT cluster. In the same figure (Fig. 2 b)) we plot
the PMFG calculated by using the detrended log-returns (Eq. 14) as comparison. This PMFG looks
more structured than the first one, with more homogeneous clustering sizes.
We have then analysed the DBHT clustering structure in terms of industrial sectors. It is well known
that the hierarchical structure of the stock return correlations shows a deep similarity with the industrial
sectors categorization [1] [30] [40]. This fact supports the intuitive argument that returns of stocks in the
same industrial sector are affected mainly by the same flows of information and economic enviroment. We
can turn around the reasoning and claim that thus a desirable feature of a clustering method on stocks
data is to retrieve, to some extent, the industrial sector classification. We will refer to the Industrial
Classification Benchmark (ICB) [41]; this categorization divides stocks into 19 different supersectors,
that in turn are gathered in 10 different Industries. For more details on the composition of our dataset
in terms of ICB supersectors, refer to SI. Let us point out that we have run all the analyses in this paper
also by using the Yahoo industrial partitioning, obtaining similar results.
In Fig. 3 we report a graphical summary of the clusters obtained applying the DBHT method to the
whole time window of data (1997-2012), by using detrended log-returns (the clustering shown in Fig. 2
b)). In S1 file we describe the same analyses for undetrended log-returns.
The DBHT retrieves a number of clusters, Ncl , equal to 23: to each cluster is associated a bar, whose
height represents the number of stocks in the cluster. Each bar is made of different colours, showing the
composition of each cluster in terms of ICB supersectors. The legend on the left of the graph reports the
corresponding industrial supersectors. Please note that the colours in Fig. 3 identify the ICB supersectors,
and they have nothing to do with colours in Fig. 2, that identify DBHT clusters.
The largest cluster contains 45 stocks (13% of total), the smallest 4. The average size is 14.8. As
we can see, there are several supersectors that are overexpressed by one or more clusters, either alone
or together with other supersectors: Oil & Gas (clusters 7 and 15), Technology (cluster 18), Utilities
9
Single Linkage
45
Basic Resources
40
Number of stocks
Chemicals
300
Construction & Materials
Financial Services
250
Food & Beverage
Health Care
200
Industrial Goods & Services
Insurance
Media150
Oil & Gas
Personal
100& Household Goods
Real Estate
DBHT
Number of stocks
Automobiles & Parts
Banks350
35
30
25
20
15
10
Retail
50
Technology
Telecommunications
0
Travel & Leisure
12345678910
11
12
13
Utilities
Cluster label
5
0
1 2 3 4 5 6 7 8 9 1011121314151617181920212223
Cluster label
Figure 3. Composition of DBHT clusters in terms of ICB supersectors.The x-axis represents
the cluster labels, the y-axis the number of stocks in each cluster. Each colour corresponds to an ICB
supersector (legend on the left hand side). The clustering is obtained by using log-returns detrended by
removing the market mode. See the Fig. S3 in S1 file for the case non-detrended.
(cluster 17), Retail (cluster 9), Health Care (clusters 22 and 23), Food & Beverage (clusters 20 and 21),
Personal & Household Goods (cluster 21), Industrial Goods & Services (clusters 12 and 16), Insurance
(cluster 11) and Telecommunications (cluster 19). Moreover, there are clusters that, although showing a
mixed composition, are composed by supersectors strictly related: clusters 5 and 6 are made of Banks,
Financial Services and Insurance, all supersectors that the ICB gathers in the same industry (Financial)
at the superior hierarchical step. Similarly, cluster 21 is made entirely of Food & Beverage and Personal
& Household Goods stocks, that are both classified in Consumer Goods industry.
Other clustering compositions
We have applied other clustering methods on the same data and compared results with DBHT clustering.
The clustering methods considered are Single Linkage (SL), Average Linkage (AL), Complete Linkage
(CL) and k-medoids. The number of clusters, that unlike the DBHT is a free-parameter for these methods,
has been chosen equal to 23 in this case, in order to compare the bar graphs with the Fig. 3 for DBHT.
We plot in Fig. 4 a), b), c) and d) the clusters composition obtained by using these four clustering
methods, namely SL, AL, CL and k-medoids. As for DBHT, the same analyses by using non-detrended
log-returns are discussed in SI.
First of all we can observe that for SL there is a strong heterogeneity in the size of clusters, with the
presence of a giant cluster containing 318 stocks, and the other clusters made of one, two or three stocks
only. This giant cluster contains stocks of all ICB sectors.
The AL case shows instead a more structured clustering: the size of the largest cluster shrinks to 58
stocks, and 6 different clusters of medium size (20-40 stocks) appear. Moreover, these clusters show a
much higher overexpression of supersectors than SL, such as Technology (cluster 4), Industrial Goods &
10
Average Linkage
Single Linkage
60
350
a)
b)
50
Number of stocks
Number of stocks
300
250
40
200
30
150
20
100
10
50
0
0
1 2 3 4 5 6 7 8 9 1011121314151617181920212223
Complete Linkage
35
45
c)
30
1 2 3 4 5 6 7 8 9 1011121314151617181920212223
Cluster label
Cluster label
k-medoids
40
d)
Number of stocks
Number of stocks
35
25
30
20
25
20
15
15
10
10
5
0
5
1 2 3 4 5 6 7 8 9 1011121314151617181920212223
Cluster label
0
1 2 3 4 5 6 7 8 9 1011121314151617181920212223
Cluster label
Figure 4. Composition of clustering in terms of ICB supersectors. The x-axis represents the
cluster labels, the y-axis the number of stocks in each cluster. Each colour corresponds to an ICB
supersector (the legend is the same as in Fig. 3). The graphs show the results for a) SL clustering, b)
for AL, c) for CL and d) for k-medoids. The clustering is obtained by using log-returns detrended by
removing the market mode.
Services (cluster 5 and 15), Media (cluster 3) and Financial related supersectors (cluster 23). However
there are still 10 clusters whose size is at most 4 stocks.
For the CL and the k-medoids the supersectors overexpression is further improved, becoming as rich
as the DBHT one. Especially CL shows overexpression of Technology (cluster 2), Industrial Goods &
Services (clusters 4 and 8), Utilities (cluster 17), Oil & Gas (cluster 23), Health Care (cluster 14) and
Financial Services (cluster 9). Similar overexpressions are found for the k-medoids case.
These first comparisons are however made under a specific choice of the number of clusters (23),
given by the DBHT. One could wonder what happens changing this parameter, i.e. moving along the
hierarchical structure provided by each clustering method. Let us stress that the DBHT method gives
automatically the number of clusters that is instead an adjustable parameter for the other methods.
However, DBHT can also be analysed for a varying number of clusters by thresholding over the clustering
hierarchical structure. In the following Sections we discuss a set of quantitative analyses that explore all
the hierarchical levels of the DBHT and the other clustering methods.
11
Disparity in the clusters size
8
Single Linkage
6
k-medoids
Complete Linkage
Single Linkage
6
Average Linkage
DBHT
y
Average Linkage
y
DBHT
4
4
a)
2
0
0
8
k-medoids
Complete Linkage
100
200
Ncl
300
b)
2
400
0
0
100
200
Ncl
300
400
Figure 5. Demonstration that different clustering methods show different degrees of
disparity in the clustering structure. The disparity measure y is shown for clusterings at different
hierachical levels as function of Ncl in the dendrograms, for a) non-detrended log-returns and b)
detrended log-returns.
In the previous Section we have seen that the SL shows a giant cluster that contains more than 90%
of stocks, whereas DBHT, CL and k-medoids methods have a more homogenous distribution of cluster
sizes and the AL seems to be an intermediate case.
Let us here check whether this difference in the structure depends on the choice of the number of
clusters for the linkage methods, which might be penalising the SL with respect to the others. In order
to do that, we vary the number of clusters Ncl for each clustering method by cutting the dendrograms at
different levels. For the k-medoids, for which no dendrogram is present, Ncl is simply an input parameter
of the algorithm. We then calculate the measure of disparity introduced in Methods (Eq. 4).
In Fig. 5 we show, for each clustering method, how the disparity measure varies with Ncl . The graph
a) shows the non-detrended case, the graph b) the detrended case. As we can see the SL provides the
higher disparity in both cases, regardless of Ncl , then the AL, CL and k-medoids follow. The DBHT values
are below all of them, this means that the DBHT clustering provides a more homogeneous community
assignment at any level of the correlation hierarchy. Moreover, in the market mode case the SL and the
AL show the highest values of disparity for Ncl in the interval 50-100. The CL and DBHT have instead a
flatter pattern, with the highest values occurring for lower values of Ncl . Looking at the detrended case
in Fig. 5 b), the removal of the market mode smooths also the pattern of the AL, whereas the SL is even
sharper. Overall, subtracting the market mode makes the clusterings more homogeneous, suggesting that
the largest clusters that emerged in SL and AL in the non-detrended case are associated to the market
mode dynamics.
The algorithms of SL and AL are indeed expected to be more sensitive to the market mode. In the
iterative procedure that generates the SL dendrogram, for instance, the correlation between two new
clusters is defined as the maximum correlation between elements of the first cluster and elements of the
second one: since the most part of correlation in the market is due to the market mode [27] such an
algorithm is likely to force many clusters to join the cluster made of the most influencial stocks in the
market, resulting in a giant cluster and high disparity. The AL is less sensitive to this effect as the
inter-clusters correlation is defined as the average of correlations; for the CL the minimum correlation
12
is chosen, resulting -unsurprisingly- in the lowest value of disparity. For what concerns the DBHT it is
probably the topology of PMFG, which is more structured and clusterised than the MST, to provide a
lower sensitivity to the market mode dynamics.
We can conclude that, from the point of view of the disparity measure, the analysed clustering
methods provide quite different structures at any level of the dendrograms. The DBHT yields the most
homogeneous clustering, whereas the SL displays the highest levels of disparity.
Retrieving the industrial sectors
In this Section we quantify the similarity between clustering and ICB, by varying the number of clusters
Ncl . We have used the Adjusted Rand Index introduced in Methods (Eq. 9). In Fig. 6 we show the
results of the first set of analyses. We have applied the five clustering methods to the entire time-period
1997-2012 and then we have varied the number of clusters Ncl . For each clustering obtained in this
way we have calculated the Adjusted Rand Index Radj between that clustering and the ICB supersector
partition. In this way we have explored the entire hierarchical structure.
k-medoids
Complete Linkage
Single Linkage
Average Linkage
DBHT
0.4
a)
0.2
0
0
Radj
Radj
0.4
k-medoids
Complete Linkage
Single Linkage
Average Linkage
DBHT
b)
0.2
0
100
200
Ncl
300
0
100
200
Ncl
300
Figure 6. Demonstration that different clustering methods retrieve different amount of
industrial sector information. The Adjusted Rand Index Radj between clustering and ICB
supersectors is shown for different number of clusters Ncl . In a) correlations are calculated on
non-detrended log-returns, in b) are calculated on detrended log-returns. The vertical dashed line shows
the value (Ncl = 19) correspondent to the actual number of ICB supersectors.
Fig. 6 a) refers to the non-detrended case, Fig. 6 b) to the detrended case. The vertical dashed line in
the graphs identifies the value Ncl = 19, that is the number of ICB supersectors. For all the methods we
observe an increasing trend for low values of Ncl , a maximum and then a decreasing trend toward zero as
Ncl goes to 342. However the five methods show differences for what concerns the value of the maximum
and its position. In the non-detrended case, we find that the highest values of Adjusted Rand Index,
R∗adj , are reached by DBHT (0.419), k-medoids (0.387) and Complete Linkage (0.387). Interestingly these
three values are quite close to each other, maybe indicating this level as the actual maximum similiarity
between correlation clustering and ICB supersectors. However the number of clusters correspondent to
∗
each maximum (Ncl
) depends on the method and is respectively 13, 17 and 39. It is worth noticing that
the maximum for the DBHT and k-medoids occurs very close to the “real” ICB supersectors Ncl = 19
indicated by the dashed line in Fig.6. However, the k-medoids values are sensitively more fluctuating
13
than the two hierachical methods. The Average Linkage and Single Linkage reach instead much lower
∗
R∗adj (respectively 0.352 and 0.184) and much higher Ncl
(respectively 111 and 229).
For what concerns the detrended case, we notice first of all that the maximum values of Radj increase
for all the methods. The natural explanation for this is that the market mode, driving all the stocks
regardless of their industrial supersector, hides to some extent the ICB structure [27]. The CL shows now
the highest degree of similarity (0.510), followed by the AL (0.48, showing the most remarkable increase
with respect to the non-detrended case), the k-medoids (0.467) and DBHT (0.444). The SL is again
∗
the last one in the ranking (0.315). The ranking in Ncl
is instead the same of the market mode case:
∗
lowest Ncl for DBHT (20), followed by k-medoids (25), Complete Linkage (50), Average Linkage (60)
and Single Linkage (101). Let us stress that, although the DBHT has not the highest Radj in this case,
its maximum is the closest to the real Ncl of 19. In general the effect of the market mode subtraction
∗
on Ncl
changes according to the clustering method: for the DBHT, CL and k-medoids the subtraction
∗
raises Ncl
, whereas for AL and SL the same quantity reduces remarkably. The effect is more pronunced
for the Linkage methods, less for DBHT and k-medoids.
Overall we can conclude that varying the number of clusters (Ncl ) the DBHT, k-medoids and CL
outperform the other two clustering methods at retrieving the ICB information. DBHT, k-medoids and
∗
CL have however peaks of R∗adj at different Ncl
values. In particular, the DBHT and k-medoids are
∗
able to retrieve the ICB information at a Ncl that is both lowest and closest to the actual number of
ICB supersectors (19). After subtracting the market mode also the AL reaches the same level of DBHT,
∗
k-medoids and CL, but at too high Ncl
.
Interestingly, the Average and Single Linkage methods have the clusterings with both the lowest values
of Radj and the highest disparity values y: i.e. it appears that the higher the disparity y is, the less
the clustering method is able to retrieve the industrial classification. This has to be due to the presence
of a large cluster when y is very high: this again indicates a strong sensitivity to the market mode (see
previous Section), that hides the intrasector correlations merging many stocks in a single cluster.
Industries overexpression
As we have stated above, the Adjusted Rand Index provides an overall measure of similarity between the
clustering and the ICB partition. Let us now focus on a more refined level of analysis, with which we
quantify to what extent each industrial sector is retrieved by the clustering.
To this aim we have varied the number of clusters Ncl and, for each clustering, we have performed
a one-tail hypothesis test (see Methods, Eq. 11) for each pair cluster/ICB supersector. In Fig. 7 we
show with bar graphs the results of these analyses for each one of the five clustering methods, by using
detrended log-returns. The x-axis shows Ncl , whereas N on the y-axis represents the total number
of hypergeometric tests that have been rejected for that value of Ncl (i.e. how many times an ICB
supersector has been found to be overexpressed by a cluster). The colors on each bar show the number
of overexpressions for each different supersector. We have chosen a significance level for the test equal
to 0.01, together with the conservative Bonferroni correction [36] that reduces the significance level to
0.01/(0.5 ∗ Ncl ∗ NICB ), with NICB being the number of ICB supersectors (19). (In Fig.S6 in S1 file, we
report also the relative numbers N /(0.5 ∗ Ncl ∗ NICB ).)
As we can see, the compositions and trends with Ncl are quite different among different methods. The
DBHT shows the highest number of clusters overexpressing ICB industries, followed by k-medoids, CL, AL
and SL. Let us note that AL and SL show the worst performances also in the Adjusted Rand Index analysis
in Fig. 6 a). However, in that analysis the DBHT, k-medoids and CL were showing very close Adjusted
Rand Index values; the hypothesis test is therefore able to highlight better the differences between these
methods. For what concerns the industry composition, we can see that the DBHT, k-medoids, CL and AL
show a quite homogeneous composition, with almost each ICB supersector overexpressed. The SL instead
shows a much less rich composition, with no more than 6 overexpressed supersectors simultaneously even
at the maximum level of total overexpressions.
14
DBHT
30
Average Linkage
30
20
N
N
20
10
10
a)
0
0
100
200
Ncl
300
b)
0
0
400
Complete Linkage
100
200
Ncl
300
400
Single Linkage
20
20
N
30
N
30
10
10
c)
0
0
100
200
Ncl
300
d)
0
0
400
100
200
Ncl
300
400
Single Linkage
Automobiles & Parts
Banks350
Basic Resources
Number of stocks
Chemicals
300
Construction & Materials
Financial Services
250
Food & Beverage
Insurance
Media150
Personal
100& Household Goods
Real Estate
20
N
Oil & Gas
k-medoids
30
Health Care
200
Industrial Goods & Services
10
Retail
e)
50
Technology
Telecommunications
0
Travel & Leisure
12345678910
11
12
13
Utilities
Cluster label
0
0
100
200
Ncl
300
400
Figure 7. Amount of ICB information retrieved by the clustering methods, in terms of
ICB supersectors overexpressed by each cluster. Each bar graph shows, varying the number of
clusters Ncl , how many times (N ) an ICB supersector is overexpressed by a cluster according to the
Hypergeometric hypothesis test (i.e., number of null-hypothesis tests being rejected). Each colour shows
the number of overexpressions for each ICB supersector. In graphs a)-e) the results for DBHT, AL, CL,
SL and k-medoids clustering are shown respectively. The correlations are calculated on detrended
log-returns. See Fig. S5 in S1 file for the non-detrended case.
15
In terms of N shape the DBHT is quite peaked, quickly dropping to low values for high Ncl values.
The three Linkage methods are instead flatter and more spread along the Ncl axis, a feature that was
evident also in the trend of Radj in Fig. 6. The k-medoids seems to be a mix between these two shapes,
showing however a much higher level of noise and instability in the ICB composition when Ncl changes.
Finally, it is worth noticing that there is a change in the composition at different values of Ncl , and
that similar patterns can be found among the four hierarchical clustering methods (for the k-medoids
no clear patterns can be found, because of the higher level of instability of the method). There are
supersectors that tend to become overexpressed for low values of Ncl and then disappear at intermediate
values: this is the case of Automobiles & Parts, Telecommunications, Insurance and Financial Services.
Others are instead more persistent, appearing along all the x-axis: Utilities, Technology, Oil & Gas. The
most persistent is the latter, that is still overexpressed when all the other supersectors are not expressed
anymore. We can then conclude that not only the ICB partition is hidden at different levels in the
dendrograms (see “Retrieving the industrial sectors” Section) depending on the clustering method, but
also different ICB supersectors are retrieved at different levels. This is probably due to the different
degrees of correlation within different ICB supersectors.
By using non-detrended log-returns we obtain quite similar results, apart from an overall decrease in
N for all the methods, consistently with what found in “Retrieving the industrial sectors”. The details
are shown in Fig. S5 in S1 file.
Dynamical analysis
Here we present a dynamical analysis of the DBHT clustering in the 15 years period ranging from 1
January 1997 to 31 December 2012. We have selected the set of overlapping time windows described in
“Dataset and preliminary analyses” (n = 100 time windows of length L = 1000 trading days) and used a
weighted version of the Pearson estimator (Eq. 13) in order to mitigate excessive sensitiveness to outliers
in remote observations.
In Fig. 8 a) is shown the number of DBHT clusters obtained for each time window, both for nondetrended log-returns (red circles) and detrended log-returns (blue squares). For the first case the number
of clusters ranges between 6 and 19, for the second case the range is 14-26. Let us stress that this is the
“natural” number of clusters automatically provided by the DBHT method, and it is not the result of
any thresholding. Dashed lines are the values correspondent to the clustering obtained using the entire
time period 1997-2012 as time window. A study of the statistical robustness of these clusterings have
been carried out by means of a bootstrapping approach: it turns out that the number of clusters shown
in Fig. 8 a) is robust against resampling of the log-returns time series, as reported in the next section
“Robustness of DBHT: a bootstrapping analysis”.
As observed previously, the number of clusters in the non-detrended case is sistematically lower than
in the detrended case. Moreover, an overall decreasing trend characterizes the non-detrended values
and makes them go below the corresponding dashed line; this decreasing pattern is not present in the
detrended case, that however stays below the correspondent dashed line most of the times either. The
evolution of the disparity y, introduced in Eq. 4, is shown in Fig. 8 b) both for non-detrended and
detrended case. Again the dashed lines are the values for the all period. In the non-detrended case we
see an overall increasing trend, especially after the 2006; an analysis of the sizes distribution shows that
largest cluster contains up to 240 stocks (70% of total number of stocks); moreover, from 2006 onwards
we observe also a much higher fluctuation in the values. This behaviour is of interest since it concerns
the overall influence of the market mode on the correlation structure, with higher y indicating a stronger
influence of the market mode that tends to gather all stocks in one cluster. Indeed, in the detrended case
we find that subtracting the market mode makes the increasing trend disappear with decreasing disparity
values that stay closer to the dashed line, without significant patterns apart from some fluctuations.
In order to better understand the relation between the DBHT clusterings obtained with detrended
and non-detrended log-returns, we have also performed a dynamical Adjusted Rand Index analysis. Now
16
30
2
non-detrended
detrended
a)
25
1.5
y
Ncl
20
15
1
10
0.5
5
0
2000
non-detrended
detrended
2002
2004
b)
2006
t
2008
2010
2012
2014
0
2000
2002
2004
2006
t
2008
2010
2012
2014
0.8
0.7
Radj (Tk )
0.6
0.5
0.4
0.3
0.2
c)
0.1
0
2000 2002 2004 2006 2008 2010 2012
t
2014
Figure 8. Dynamical evolution of the DBHT clustering. Each plot refers to 100 moving time
windows of length 1000 trading days. Specifically, in graph a) we plot the number of DBHT clusters,
Ncl , for both log-returns non-detrended (red circles) and detrended by the market mode (blue squares),
whereas the two dashed horizontal lines are the Ncl values obtained by taking the largest time window
of 4026 trading days. Overall the non-detrended case shows a decreasing trend. In graph b) it is shown
the disparity measures, y, again for the two sets of DBHT clustering (red dots non-detrended, blue dots
detrended), the dashed horizontal lines being the y values from the 4026 length time window. In the
non-detrended case the 2007 marks a transition to higher and more volatile values of y. Finally in graph
c) it is shown the Adjusted Rand Index, Radj , measured at each time window between the detrended
and non-detrended clusterings. A steady decreasing trend is evident.
we compare no longer the clustering and the ICB partition, but the two clusterings (non-detrended
and detrended) at each time window. In Fig. 8 c) the Adjusted Rand Index between the two sets of
DBHT clusters is shown. Interestingly, it shows a steady decreasing trend that drives the similarity
from relatively high values (about 0.7) to values close to zero, indicating complete uncorrelation between
the two clusterings. We can therefore conclude that the influence of the market mode has increased
remarkably over the last 15 years, making the detrended clustering structure more and more different
from the non-detrended one. Let us note that this observation would have not been possible without the
clustering analysis, since from the preliminary dataset measures (see Fig. 1 and Fig. S2 in S1 file for
details) it is not evident any constant pattern either in the average return or in the average correlation.
17
Dynamically retrieving the industrial sectors
Let us here investigate the relation between industrial classification and clustering under a dynamic
perspective. To this end we here perform the previous dynamical analysis by considering the set of 100
overlapping time windows Tk and calculating for each of them the Adjusted Rand Index Radj (Tk ) between
clustering and ICB supersectors classification. Since Radj (Tk ) varies with the chosen threshold and Ncl ,
we select at every time the Ncl that maximizes Radj (Tk ); the numbers that we report are these maximum
values and account therefore for the maximum ability of the clustering methods to retrieve the ICB.
In Figs. 9 a)-e) we show the results for each of the five clustering methods, using returns with market
mode. Interestingly, all of them show a decreasing trend across the time. On average, the DBHT and CL
display the highest similarity with industrial classfication, whereas the Single Linkage the lowest. This
is consistent with what found in the static analyses. We have also highlighted in the graphs the major
events that affected the stock market in the last 15 years. It can be observed that different clustering
methods are affected in different ways by these events. For instance we observe that, if the 2007-2008
crisis and the following recession is evident in all the methods as a significant drop in the similarity, other
events such as 11/09/2001 or the 2002 stock market downturn appear only in the Single and Average
Linkage plots. In particular the 2002 downturn drives a steep decrease in the similarity of SL and AL,
staying at low values until the end of 2005. Instead DBHT, Complete Linkage and k-medoids do not
seem to be significantly affected by these events. This observation points out that the DBHT, CL and kmedoids are more robust than SL and AL against exogenous events in their ability to retrieve an economic
information as the industrial classification. Nonetheless, there are differences also among DBHT, CL and
k-medoids: in particular in the period following the 2008 crisis, DBHT and k-medoids show a peak that
does not appear with CL. Moreover, for the k-medoids the drop in similarity seems to start more than
one year before the 2007. All these features have non-trivial implications for both portfolio optimization
and systemic risk evaluation. We plan to investigate these implications in future works.
Fig. 9 f) shows the number of clusters Ncl that, in each time window Tk , maximizes the Adjusted
Rand Index shown in the previous plots. As we can see, Ncl for SL is always the highest, followed by
AL, CL, k-medoids and DBHT. This is consistent with what we found in the “Static analysis” Section:
different clustering methods “hide” the industrial information at different levels of the hierarchy. SL and
AL, that yield higher Ncl (i.e., lower levels in the hierarchy), are also the methods that show the lowest
level of similarity with industrial classification and the highest degree of disparity.
In Figs. 10 a)-f) we show the same set of plots for the detrended case. The main differences with the
non-detrended case are the following:
• the average similarity with the industrial classification rises for all methods; this confirms in the
dynamical case what we found for the static case;
• the average Ncl is lower for all methods: the absence of market mode “moves” the industrial
classification to higher levels of the hierarchy;
• the strong influence of the 11/09/2001 and 2002 downturn on the SL and AL pattern seems to
disappear, whereas the 2007-2008 crisis is still evident in all the five methods. This could be
explained claiming that the formers are global events in the market, whereas the latter exhibits also
a “local” dynamics;
• the AL shows the most evident change in the dynamical behaviour, displaying a trend much more
similar to the DBHT and CL one. Also in terms of Ncl , it shows values closer to DBHT, CL
and k-medoids than SL. A similar observation was made in the static analysis in “Retrieving the
industrial sectors”, where the AL turned out to perform like DBHT, CL and k-medoids once the
market mode was removed.
18
0.5
0.5
DBHT
Average Linkage
b)
0.4
Radj (Tk )
Radj (Tk )
0.4
0.3
Lehman Brothers default
0.2
a)
USA recession
offically announced
08/07: worldwide
credit crunch spreads
0.1
2000
2002
2004
2006
0.3
Lehman Brothers default
0.2
0.1
t
2008
2010
2012
2014
USA downgrade;
EU debt crisis
07/02:
market downturn
11/09/01
2000
08/07: worldwide
credit crunch spreads
2002
2004
2006
USA recession
offically announced
t
2008
2010
0.5
0.5
2012
2014
Single Linkage
Complete Linkage
Lehman Brothers default
0.4
Radj (Tk )
Radj (Tk )
0.4
0.3
c)
Lehman Brothers default
0.2
0.2
d)
08/07: worldwide
credit crunch spreads
2002
2004
2006
0.1
t
2008
2010
2012
350
300
250
2002
2004
2006
t
2008
2010
DBHT
Single Linkage
Average Linkage
Complete Linkage
k-medoids
2012
2014
f)
200
0.3
Ncl
Radj (Tk )
0.4
e)
150
08/07: worldwide
credit crunch spreads
0.2
2000
USA recession
offically announced
2000
2014
0.5
0.1
USA downgrade;
EU debt crisis
08/07: worldwide
credit crunch spreads
0.3
USA recession
offically announced
0.1
2000
07/02:
market downturn
USA recession
officially announced
USA downgrade;
EU debt crisis
k-medoids
2002
2004
100
2006
t
2008
2010
2012
2014
50
0
2000
2002
2004
2006
t
2008
2010
2012
2014
Figure 9. Dynamical evolution of the similarity between clustering and ICB. It is shown the
Adjusted Rand Index, Radj , calculated at each time window Tk (k = 1, ..., n) between clustering and
ICB partition, for a) DBHT, b) AL, c) CL, d) SL and e) k-medoids method. A drop in the similarity
occurs for all the methods during the 2007-2008 crisis. The AL and SL show decreases also during other
financial events. At each time window the number of clusters, Ncl , has been chosen in order to
maximize the Radj itself: in f) we plot these Ncl values for each clustering method. It is evident as the
maximum similarity clustering-ICB is reached at different hierarchical levels depending on the
clustering method. The correlations are calculated on non-detrended log-returns.
19
DBHT
Average Linkage
Lehman Brothers default
Lehman Brothers default
0.5
Radj (Tk )
Radj (Tk )
0.5
0.4
0.3
a)
0.4
b)
0.3
08/07: worldwide
credit crunch spreads
08/07: worldwide
credit crunch spreads
0.2
2000
2002
2004
2006
t
2008
2010
2012
USA recession
offically announced
0.2
USA recession
offically announced
2000
2014
2002
2004
2006
t
2008
2010
2012
2014
Single Linkage
Complete Linkage
Lehman Brothers default
0.5
Radj (Tk )
Radj (Tk )
0.5
0.4
0.3
c)
d)
0.4
0.3
USA recession
officially announced
08/07: worldwide
credit crunch spreads
2000
2002
2004
2006
t
2008
2010
2012
2000
2014
300
k-medoids
250
200
0.4
2004
2006
t
2008
2010
DBHT
Single Linkage
Average Linkage
Complete Linkage
k-medoids
2012
2014
f)
150
e)
100
Lehman Brothers default
50
0.2
2000
2002
Ncl
Radj (Tk )
0.5
0.3
USA recession
offically announced
0.2
0.2
USA downgrade;
EU debt crisis
Lehman Brothers default
2002
2004
2006
t
2008
2010
2012
2014
0
2000
2002
2004
2006
t
2008
2010
2012
2014
Figure 10. Dynamical evolution of the similarity between clustering and ICB, with
detrended log-returns. a)-f): Same graphs as in Fig. 9, but by using correlations on detrended
log-returns.
Robustness of DBHT: a bootstrapping analysis
In order to test the sensitiveness of the DBHT clustering to the statistical noise, inevitably present in
every correlation estimate, we have performed the Bootstrapping test to our dataset [42]. This method
generates a set of nboot replicas of each original correlation matrix by randomly resampling the log-returns
matrix. On each replica of correlation matrix we can calculate a new DBHT clustering, ending up with
a sample of nboot clusterings on which we can perform statistical analyses. A detailed description of the
20
26
16
a)
14
24
Ncl
Ncl
b)
22
12
10
8
6
Empirical returns series
Bootstrapped returns series
20
18
16
Empirical returns series
Bootstrapped returns series
4
2000 2002 2004 2006 2008 2010 2012 2014
t
14
12
2000 2002 2004 2006 2008 2010 2012 2014
t
Figure 11. Test of robustness for the dynamical DBHT clustering. a) Number of clusters Ncl
as a function of the time t: the black squares correspond to the DBHT clusterings obtained by using the
empirical (non-detrended) log-returns, the blue dots are the average over the 100 Ncl given by the 100
bootstrapping replica correlation matrices (see text for further details). The bar errors in the blue dot
plot is the standard deviation calculated among the same set of 100 Ncl . As one can see the empirical
Ncl is quite robust against the bootstrapping test. b) Same plot as in a), but by using detrended
log-returns.
Bootstrapping method can be found in SI.
In Fig. 11 we show the result of the dynamical Bootstrapping performed over all the 100 time windows
that cover the entire period. We chose nboot = 100: each time window has therefore 100 replica DBHT
clustering associated. The blue points are the average number of clusters over the nboot replicas of DBHT
clusterings, whereas the error bars are the standard deviations calculated over the same sample. The
black squares are the empirical numbers of clusters yielded by the DBHT. The left-hand side plot (a))
is obtained by using non-detrended log-returns, the right-hand side (b)) is obtained by using detrended
log-returns.
The plot of empirical number of clusters is slightly different from what we have shown in Fig. 8
a) because for this bootstrapping analysis we did not use exponential smoothing for the correlations,
but only bare correlations. The exponential smoothing, indeed, creates an asymmetry among the points
in each time series that makes the bootstrapping test inapplicable. Overall, removing the exponential
smoothing affects each Ncl value by 20% on average.
From the plot we can observe that the method is statistically robust, with the most of empirical points
within one standard deviation from the mean of replicas. More importantly, the mean of replicas follows
the general trend of the empirical points; namely, the decreasing trend in the market mode case, and the
drop after the 2007-2008 crisis in the detrended case.
Discussion
In this paper we have presented a set of static and dynamical analyses to quantify empirically the amount
of information filtered from correlation matrices by different hierarchical clustering methods. By taking
the Industrial Classification Benchmark (ICB) as a benchmark community partition we have performed
quantitative analyses on real data without any assumption on the returns distribution.
In particular we have considered three variants of Linkage methods (Single, Average and Complete)
and the k-medoids, and we have compared them with the Directed Bubble Hierarchical Tree (DBHT), a
novel clustering method applied here for the first time to financial data.
The methods show remarkably different performances in retrieving the economic information encoded
21
in the ICB, with big dissimilarities even among the Linkage methods. We have suggested that these
differences should be connected to different degrees of sensitivity to the market mode dynamics, that
in turns are to be ascribed to differences in the methods underlying working principles. Moreover, the
economic information appears to be hidden at different levels of the hierarchical structures depending on
the clustering method. The DBHT and k-medoids methods show the best performances, but the latter
seems to be affected by the noise much more than the DBHT and the Linkage methods. The DBHT
turns out then to be a good mix between the advantages of the k-medoids and those of the Linkages.
The dynamical analysis has also proved that the methods show different degrees of sensitivity to financial
crises. This is again a new result that could give insights into the dynamics of such events, as well as an
indication on which clustering method is more robust for financial applications.
We have also performed each analysis on log-returns detrended by the market mode, by following a
standard procedure in literature [27] [40]. Interestingly the effect of this detrending is very dissimilar
for different methods, with the weakest methods (Average and Single Linkage) improving remarkably
their ability to retrieve industrial sectors. In general the detrending increases the degree of economic
information that the clustering methods retrieve. It also makes the clusterings more homogeneous in
sizes, suggesting that the high heterogeneity in SL and AL must be due to the market mode dynamics.
In future works we plan to extend the present study to other datasets, covering different periods and
different stock exchanges and considering other measures of dependences including non-linear dependences
such as the Kendall’s rank correlation [43] and the Mutual information [44] [45]. Finally, since correlationbased networks and clustering methods have shown to be useful tools for portfolio optimization [20],
[46, 47], we also plan to use these new insights into the hierarchical structures to improve further the
current performances of portfolio optimization tools.
Acknowledgments
The authors wish to thank Bloomberg for providing the data. TDM wishes to thank the COST Action
TD1210 for partially supporting this work. TA acknowledges support of the UK Economic and Social
Research Council (ESRC) in funding the Systemic Risk Centre (ES/K002309/1).
References
1. Mantegna RN (1999) Hierarchical structure in financial markets. Eur Phys J B 11: 193.
2. Onnela JP, Chakraborti A, Kaski K, Kertész J, Kanto A (2003) Asset trees and asset graphs in
financial markets. Phys Scr T106: 48.
3. Aste T, Di Matteo T, Hyde ST (2005) Complex networks on hyperbolic surfaces. Physica A 346:
20.
4. Tumminello M, Aste T, Di Matteo T, Mantegna RN (2005) A tool for filtering information in
complex systems. Proc Natl Acad Sci 102: 10421-10426.
5. Di Matteo T, Aste T (2002) How does the eurodollars interest rate behave?
Finance 5: 122-127.
J Theoret Appl
6. Di Matteo T, Aste T, Mantegna RN (2004) An interest rate cluster analysis. Physica A 339:
181-188.
7. Di Matteo T, Aste T, Hyde ST, Ramsden S (2005) Interest rates hierarchical structure. Physica
A 335: 21-33.
22
8. Bartolozzi M, Mellen C, Di Matteo T, Aste T (2007) Multi-scale correlations in different futures
markets. Eur Phys J B 58: 207-220.
9. Onnela JP, Chakraborti A, Kaski K, Kertész J (2003) Dynamic asset trees and black monday.
Physica A 324: 247-252.
10. Tola V, Lillo F, Gallegati M, Mantegna RN (2008) Cluster analysis for portfolio optimization. J
Econ Dyn Control 32: 235-258.
11. Fenn DJ, Porter MA, Mucha P, McDonald M, Williams S, et al. (2012) Dynamical clustering of
exchange rates. Quantitative Finance 12: 1493.
12. Di Matteo T, Pozzi F, Aste T (2010) The use of dynamical networks to detect the hierarchical
organization of financial market sectors. Eur Phys J B 73: 3-11.
13. Tumminello M, Di Matteo T, Aste T, Mantegna RN (2007) Correlation based networks of equity
returns sampled at different time horizons. Eur Phys J B 55: 209-217.
14. Morales R, Di Matteo T, Aste T (2014) Dependency structure and scaling properties of financial
time series are related. Sci Rep 4: 4589.
15. Aste T, Shaw W, Di Matteo T (2010) Correlation structure and dynamics in volatile markets. New
J Phys 12: 085009.
16. Tumminello M, Lillo F, Mantegna RN (2010) Correlation, hierarchies, and networks in financial
markets. J Econ Behav Organ 75: 40-58.
17. Anderberg MR (1973) Cluster analysis for applications. Academic Press.
18. Tumminello M, Coronnello C, Lillo F, Miccichè S, Mantegna RN (2007) Spanning trees and bootstrap reliability estimation in correlation-based networks. Int J Bifurcat Chaos 17: 2319-2329.
19. Aste T (2012). An algorithm to compute Planar Maximally Filtered Graphs (PMFG). URL
http://www.mathworks.com/matlabcentral/fileexchange/38689-pmfg.
20. Pozzi F, Di Matteo T, Aste T (2013) Spread of risk across financial markets: better to invest in
the peripheries. Sci Rep 3: 1665.
21. Song WM, Di Matteo T, Aste T (2012) Hierarchical information clustering by means of topologically embedded graphs. PLoS ONE 7: e31929.
22. Aste T (2014). An algorithm to compute Directed Bubble Hiererachical Tree (DBHT) clustering.
URL http://www.mathworks.com/matlabcentral/fileexchange/46750-dbht.
23. Song WM, Di Matteo T, Aste T (2011) Nested hierarchies in planar graphs. Discrete Appl Math
159: 2135.
24. Fortunato S (2010) Community detection in graphs. Physics Reports 486: 75-174.
25. Kaufman L, Rousseeuw PJ (1987) Clustering by means of medoids. Statistical Data Analysis Based
on the L1 -Norm and Related Methods : 405-416.
26. MacQueen J (1967) Some methods for classification and analysis of multivariate observations.
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability 1: 281297.
23
27. Borghesi C, Marsili M, Miccichè S (2007) Emergence of time-horizon invariant correlation structure
in financial returns by subtraction of the market mode. Phys Rev E 76: 026104.
28. Hubert L, Arabie P (1985) Comparing partitions. Journal of Classification 2: 193-218.
29. Tumminello M, Miccichè S, Lillo F, Varho J, Piilo J, et al. (2011) Community characterization of
heterogeneous complex systems. J Stat Mech P01019.
30. Coronnello C, Tumminello M, Lillo F, Miccichè S, Mantegna RN (2011) Sector identification in a
set of stock return time series traded at the London Stock Exchange. Acta Physica Polonica 36:
2653-2680.
31. Tumminello M, Lillo F, Mantegna RN (2007) Kullback-Leibler distance as a measure of the information filtered from multivariate data. Phys Rev E Stat Nonlin Soft Matter Phys 76: 031123.
32. Cont R (2001) Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance 1: 223-236.
33. Mantegna RN, Stanley HE (2000) Introduction to Econophysics: Correlations and Complexity in
Finance. Cambridge University Press.
34. Bonanno G, Caldarelli G, Lillo F, Mantegna RN (2003) Topology of correlation-based minimal
spanning trees in real and model markets. Phys Rev E Stat Nonlin Soft Matter Phys 68: 046130.
35. Wagner S, Wagner D (2007) Comparing clusterings - an overview. Technical Report, ITI Wagner,
Faculty of Informatics, Universität Karlsruhe (TH) .
36. Feller W (2008) An introduction to probability theory and its applications, volume 2. John Wiley
& Sons.
37. Tumminello M, Miccichè S, Lillo F, Piilo J, Mantegna RN (2011) Statistically validated networks
in bipartite complex systems. PLoS ONE 6: e17994.doi:10.1371/journal.pone.0017994.
38. Pearson K (1895) Notes on regression and inheritance in the case of two parents. Proceedings of
the Royal Society of London 58: 240-242.
39. Pozzi F, Di Matteo T, Aste T (2012) Exponential smoothing weighted correlations. Eur Phys J B
85: 6.
40. Ross G (2014) Dynamic multifactor clustering of financial networks. Phys Rev E Stat Nonlin Soft
Matter Phys 89: 022809.
41. A guide to industry classification benchmark. URL http://www.icbenchmark.com/.
42. Efron B (1979) Bootstrap methods: Another look at the jackknife. Ann Stat 7: 1-26.
43. Kendall MG (1938) A new measure of rank correlation. Biometrika 30: 81-93.
44. Shannon CE (1948) A mathematical theory of communication. Bell Syst Tech J 27: 379.
45. Fiedor P (2014) Networks in financial markets based on the mutual information rate. Phys Rev E
Stat Nonlin Soft Matter Phys 89: 052801.
46. Nanda S, Mahanty B, Tiwari MK (2010) Clustering indian stock market data for porfolio management. Expert System with Applications 37: 8793-8798.
47. Pai GAV, Michel T (2009) Clustering indian stock market data for porfolio management. Evolutionary Optimization of Constrained K-means Clustered Assets for Diversification in Small Portfolios
13: 1030-1053.
24
Supporting Information
S1. Dataset analysis
Automobiles & Parts
Banks
Basic Resources
Chemicals
Construction & Materials
2%
4%
2%
4%
4% 2% 7%
2%
1%
10%
Financial Services
Food & Beverage
Health Care
5%
Industrial Goods & Services
Insurance
Media
10%
7%
Oil & Gas
Personal & Household Goods
2%
Real Estate
Retail
Technology
7%
16%
Telecommunications
Travel & Leisure
5%
4%
6%
Utilities
Figure 12. Pie chart showing the composition of the entire set of stocks in terms of ICB
supersectors.
The set of stocks has been chosen in order to provide a significant sample of the different industrial
sectors in the market. We have chosen the ICB industrial classification, that yields 19 different Supersectors, that in turns gather in 10 Industries: the percentage of stocks belonging to each ICB supersectors
is reported in Fig. 12 .
In Fig. 13 two plots are
P shown that summarize the main features of this dataset.PThe graphs show the
average price P̄ (t) ≡ N1 i Pi (t) and the average log return of the prices, r̄(t) ≡ N1 i ri (t), as a function
of time. From these plots we can see that both the internet bubble bursting (2002) and the credit crunch
(2007-08) are displayed by the market dynamics. In particular it is evident a steep increase in volatility
for both periods, strongly autocorrelated in time: a well known feature of log-returns dynamics [32]. Such
clusters of volatility can be observed also after the credit crunch, in 2010 and 2012.
S2. Clustering methods: a brief review
Our focus is on the hierarchical structures (dendrograms) that can best describe the interdependencies in
the market, together with a cluster characterization of this structure. To this purpose, the first step is to
define a suitable distance
p between pairs of stocks, knowing the correlation between them. An appropriate
function [1] is Dij = 2(1 − Cij ) : it can be shown that with this choice Dij satisfies the three properties
of a distance measure.
25
Average price
Average log returns
55
0.1
a)
45
0.05
40
0
r̄(t)
P̄ (t)
50
0.15
35
−0.05
30
−0.1
25
−0.15
20
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014
t
b)
−0.2
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014
t
Figure 13. Average price and log-returns of the dataset, from January 1997 to December
2012. a) Average price P̄ (t) of the 342 US stocks in the dataset; b) Average log-return r̄(t) of the same
prices.
We then end up with a set of N × N distance matrices D(tk ) and DR (tk ), to which we apply two
different, well known tools in order to reveal the hidden (unknown) structure of dependencies:
• Single Linkage (SL), that is an hierarchical clustering algorithm. Given the distance matrix, it
starts assigning to each objects its own cluster, and then at each step merges the closest (i.e. least
distant) pairs of clusters into one new cluster, until only one cluster remains. The distances among
two generic clusters A and B is everytime defined and updated according to the formula
dA,B =
min
a∈A,b∈B
D(a, b)
(15)
SL is called an agglomerative clustering, since it begins with a partition of N clusters and then
proceed merging them. The final output of the method is a dendrogram, that is a tree showing
the hierarchical structure found by the SL. The distance measure defined in this dendrogram is
an ultrametric distance [1]. A proper cluster partition of the stocks can be obtained by choosing
the number of clusters (that is therefore a free parameter) and cutting the dendrogram at the
appropriate level.
This algorithm is strictly related to the one that provides a Minimum Spanning Tree (MST) given
the distance matrix D. The MST is a tree graph having the stocks as nodes, and it has been used
as topological tool in Econophysics since the work of Mantegna [1]. It can be generated starting
with an empty graph: after sorting all the correlations in C in descending order, add a weighted
link between the two stocks/nodes with the highest correlation, and then go ahead with the next
highest pair correlation; whenever the new link to add generates a loop, do not add that link and
skip to the next one, until all the list is checked.
This tree contains exactly N − 1 links. It can be shown [16] that the MST algorithm is basically
the SL procedure carried out until the graph is completely connected. There is therefore a strict
relation between the two tools. However the MST retains some information that the SL dendrogram
throws away [16] .
26
• Average Linkage (AL) is a hierarchical clustering algorithm similar to SL. The algorithm is the
same as the one underlying the SL, but with Eq. 17 replaced by:
dA,B = meana∈A,b∈B D(a, b)
(16)
• Complete Linkage (CL) is a third variant of SL, where Eq. 17 is replaced by:
dA,B =
max D(a, b)
a∈A,b∈B
(17)
• Directed Bubble Hierarchical Tree (DBHT) [21], a novel hierarchical clustering method that
exploits the topological property of the PMFG (Planar Maximally Filtered Graph) in order to find
the clustering.
The PMFG is a generalization of the MST, that is included in the PMFG as a subgraph. It
is constructed following the same procedure of the MST, except that the non loop condition is
replaced with the weaker condition of planarity (i.e. each added link must not cut a pre-existent
link). Thanks to this more relaxed topological constraint the PMFG is able to retain a larger
amount of link, and then information, than the MST. In particular it can be shown that each
PMFG contains exactly 3(N − 2) links.
The basic elements of a PMFG are three-cliques, subgraphs made of three nodes all reciprocally
connected (i.e., triangles). The DBHT exploits this topological structure, and in particular the
distinction between separating and non-separating three-cliques, to identify a clustering partition
of all the nodes in the PMFG [21] . A complete hierarchical structure (dendrogram) is then obtained
both inter-clusters and intra-clusters by following a traditional agglomerative clustering procedure.
The Linkage algorithms look at the sorted list of distances dij and build the dendrogram by gathering subsets of stocks with lowest distances; the clustering is then obtained, as we said, from
the dendrogram after choosing the parameter “number of clusters”. The DBHT instead reverses
this order: first of all the clusters are identified by means of topological considerations on the planar graph, then the hierarchy is constructed both inter-clusters and intra-clusters. The difference
involves therefore both the kind of information exploited and the methodological approach.
• k-medoids is a partitioning clustering method closely related to k-means [26]. It takes the number
of clusters Ncl as an input. The algorithm is the so called Partitioning Around Medoids (PAM),
and is as follows:
1. select randomly Ncl “medoids” among the N elements;
2. assign each element to the closest medoid;
3. for each medoid, replace the medoid with each point assigned to it and calculate the cost of
each configuration;
4. choose the configuration with the lowest cost;
5. repeat 2)-4) until no change occurs.
This method, alike the others taken into account here, is not a hierarchical method and does not
provide therefore a dendrogram but only a partition.
27
Single Linkage
Automobiles & Parts
Banks350
DBHT
70
Basic Resources
Financial Services
250
Food & Beverage
Health Care
200
Industrial Goods & Services
Insurance
Media150
Oil & Gas
Personal
100& Household Goods
Real Estate
Retail
50
Technology
Telecommunications
0
Travel & Leisure
12345678910
11
12
13
Utilities
Cluster label
60
Number of stocks
Number of stocks
Chemicals
300
Construction & Materials
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 1011121314151617
Cluster label
Figure 14. Number of stocks and composition of DBHT clusters in terms of ICB
supersectors, on non-detrended log-returns. The composition is shown by using different colours.
S3. Clustering compositions: non-detrended case
DBHT clusters composition
In Fig. 14 we report a graphical summary of the clusters obtained applying the DBHT method to the
whole time window of data (1997-2012), by using non-detrended log-returns.
The DBHT returns a number of clusters, Ncl , equal to 17. Cluster 4, the largest, is made of 62 stocks,
accounting for about the 18% of the total number of stocks; cluster 9, the smallest, contains 4 stocks. The
average size of clusters is 20.1 stocks. As we can see, four clusters show a composition of stocks belonging
to only one ICB supersector : cluster 9 and 13 (Oil & Gas), 11 (Technology) and 14 (Utilities). Similar
cases are cluster 8, made of Technology stocks for more than 86%, cluster 15, within which 91% of stocks
are from Retail, cluster 16 (75% of stocks from Health Care) and cluster 17 (87.5% of stocks from Food
& Beverage). Moreover there are clusters that, although showing a mixed composition, are composed
by supersectors strictly related: the number 6 is made of Banks, Financial Services and Insurance, all
supersectors that the ICB gathers in the same industry (Financial) at the superior hierarchical step.
There are clusters that do not show an overexpression for a particular supersector or industry: this
fact points out that the clustering is after all providing an information that cannot be reduced only
to the industrial classification. In particular clusters 1, 3 and 12 have a heterogeneous composition,
covering almost all the 19 supersectors and with no sector dominating the others. The cluster 4 is an
intermediate case, since even though it overexpresses the Industrial Goods & Services (75%), it contains
stocks belonging to 9 different supersectors. Interestingly the largest clusters (4, 12, 1 and 10) are all
among these types of “mixed” clusters.
Other clustering compositions
We here apply other clustering methods on the same data and compare results with DBHT clustering. The
clustering methods considered are Single Linkage (SL), Average Linkage (AL), Complete Linkage (CL)
28
Average Linkage
Single Linkage
350
350
a)
250
250
200
200
150
150
100
100
50
0
b)
300
Number of stocks
Number of stocks
300
50
0
1 2 3 4 5 6 7 8 9 1011121314151617
1 2 3 4 5 6 7 8 9 1011121314151617
Cluster label
Cluster label
Complete Linkage
k-medoids
140
140
c)
120
Number of stocks
Number of stocks
120
100
80
60
40
100
80
60
40
20
20
0
d)
1 2 3 4 5 6 7 8 9 1011121314151617
Cluster label
0
1 2 3 4 5 6 7 8 9 1011121314151617
Cluster label
Figure 15. Composition of clustering in terms of ICB supersectors, for different
clustering methods, on non-detrended log-returns. The x-axis represents the single cluster
labels, the y-axis the number of stocks in each cluster. Each colour corresponds to an ICB supersector
(the legend is the same as in Fig. 14. The graphs show the results for a) SL clustering, b) for AL, c) for
CL and d) for k-medoids.
and k-medoids. The latter is not a hierarchical clustering method, so it does not provide a dendrogram:
however we analysed it to compare our results with a well established clustering method. The number
of clusters, that unlike the DBHT, is a free-parameter for these methods, has been chosen equal to 17 in
these cases, in order to compare the bar graphs with the Fig. 14 for DBHT. We plot in Fig. 15 a), b),
c) and d) the clusters compositions obtained by using these four clustering methods, namely SL, AL, CL
and k-medoids.
First of all we can observe that for each of them there is a strong heterogeneity in the size of clusters:
SL and AL display two huge clusters of 323 and 322 stocks respectively (almost identical, having 318
stocks in common), with the other clusters made of one, two or three stocks. For both the algorithms
this giant cluster contains stocks of all ICB sectors.
29
For the CL and the k-medoids the situation is quite different. For the CL, the giant cluster (cluster
number 10) is much reduced in size (136 stocks), with also other three clusters (the number 12, 9 and
5) containing a relevant number of stocks (50, 33 and 25 respectively): the main supersectors that are
overexpressed are Technology (cluster 12), Utilities (cluster 5), Retail (cluster 9), Oil & Gas (cluster
16) and Health Care (cluster 2). A very similar structure occurs with the k-medoids, but with the giant
cluster splitting further in two large clusters (7 and 10). However the DBHT clustering is the one showing
the largest degree of homogeneity in size and overexpression of ICB supersectors, at least for this number
of clusters (see Fig. 14 for comparison).
Comparing these results with the same analyses on detrended log-returns (Fig. 3 and 4 in the paper)
we can conclude that the subtraction of the market mode makes all the clusterings methods (with the
exception of SL) more homogenous in size and more able to retrieve the ICB partition. The SL clustering
instead does not seem to be sensitive to this subtraction, and keeps not overexpressing any ICB supersector
even in the detrended case (Fig. 4 a)).
S4. Bootstrapping test of robustness
The basic idea of the Bootstrapping technique is the following [42]: suppose, for a given time window of
length L, we have N time series (one for each stock), each one having length L . We can fit this data
in a N × L matrix, say X, and calculate the correlation matrix for it, say ρ , and a clustering using
the DBHT, say Y . Now let us create a replica X 0 of the matrix X, such that each row of X 0 is drawn
randomly among the rows of X, allowing multiple drawings of the same rows. From X 0 we can again
calculate a correlation matrix ρ0 and a clustering Y 0 .
By repeating this procedure nboot times, we end up with nboot replica of clusterings, each one slightly
different from the original one due to the differences between X and its replicas. This sample of replicas
can be used to test the robustness of any quantity measured in the original clustering Y , e.g. the number
of clusters. This can be done by checking whether the original measure is compatible with the distribution
of replicas, performing e.g. a statistical hypothesis test.
30
30
25
25
20
20
N
N
Average Linkage
DBHT
30
15
10
b)
10
a)
5
0
0
15
5
100
200
Ncl
300
0
0
400
100
30
25
25
20
20
15
c)
10
400
15
d)
10
5
0
0
300
Single Linkage
30
N
N
Complete Linkage
200
Ncl
5
100
200
Ncl
300
400
0
0
100
Single Linkage
Automobiles & Parts
Banks350
Number of stocks
25
Financial Services
250
Food & Beverage
20
Retail
50
Technology
Telecommunications
0
Travel & Leisure
12345678910
11
12
13
Utilities
Cluster label
N
Health Care
200
Industrial Goods & Services
Personal
100& Household Goods
Real Estate
400
k-medoids
Chemicals
300
Construction & Materials
Oil & Gas
300
30
Basic Resources
Insurance
Media150
200
Ncl
15
e)
10
5
0
0
100
200
Ncl
300
400
Figure 16. ICB supersectors overexpression at different levels of the hierarchies. Each bar
graph shows, varying the number of clusters Ncl , how many times (N ) an ICB supersector is
overexpressed by a cluster, according to the Hypergeometric hypothesis test (i.e., number of tests being
rejected). Each colour shows the number of overexpressions for each ICB supersector. In graphs a)-e)
the results for DBHT, AL, CL, SL and k-medoids clustering respectively are shown. The correlations
are calculated on non-detrended log-returns.
31
DBHT
0.5
a)
′
0.3
0.2
0.1
b)
0.4
N
N
′
0.4
0
Complete Linkage
0.5
0.3
0.2
0.1
50
100
150
200
Ncl
250
300
0
50
100
150
200
250
300
Ncl
Figure 17. ICB supersectors overexpression as percentage of pairs cluster/supersector
rejecting the hypergeometric test. Each bar graph shows, varying the number of clusters Ncl , how
many times (N ) an ICB supersector is overexpressed by a cluster according to the Hypergeometric
hypothesis test (i.e., number of tests being rejected), divided by the total number of Hypergeometric
2N
tests performed (0.5 × Ncl × NICB , with NICB the number of ICB supersectors): N 0 = Ncl ×N
. Each
ICB
colour shows the number of overexpressions for each ICB supersector. In graphs a)-b) the results for
DBHT and CL clusterings respectively are shown. The correlations are calculated on detrended
log-returns.
| 5cs.CE
|
www.seipub.org/gpai
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
Cellular Automata and Its Applications in
Bioinformatics: A Review
Pokkuluri Kiran Sree 1, Inampudi Ramesh Babu 2, SSSN Usha Devi .N3
1. Research Scholar, Dept of CSE, JNTU Hyderabad, [email protected]
2. Professor, Dept of CSE, Acharya Nagarjuna Univesity, Guntur. [email protected]
3. Assistant Professor, Dept of CSE, JNTU Kakinada
Abstract
Implicit Modeling (Numerical Techniques)
This paper aims at providing a survey on the problems that
can be easily addressed by cellular automata in
bioinformatics. Some of the authors have proposed
algorithms for addressing some problems in bioinformatics
but the application of cellular automata in bioinformatics is a
virgin field in research. None of the researchers has tried to
relate the major problems in bioinformatics and find a
common solution. Extensive literature surveys were
conducted. We have considered some papers in various
journals and conferences for conduct of our research. This
paper provides intuition towards relating various problems
in bioinformatics logically and tries to attain a common
frame work for addressing the same.
The second methodology addresses this issue through
advancement of the model dependent upon
perceptions and experience. In accordance with this
approach, CA based model takes in companionships
from a set of samples and after that apply this
information base to handle beforehand unseen in line
with this methodology, Cellular Automata (CA) based
classifier can learn associations from examples and
apply this acquired knowledge to handle unknown
cases.
Keywords
Cellular Automata; Multiple Attractors CA; Artificial Immune
System
Introduction
The survey is mainly concerted survey of problems in
bioinformatics like protein coding and promoter
prediction and protein structure prediction. In the
section 3 of this paper, Cellular Automata is
introduced with AIS-MACA (Artificial Immune
System Based Multiple Attractor Cellular Automata).
The next section 4, 5 of the paper introduce the survey
on protein coding region, promoter region prediction.
Then the survey on protein structure prediction is
presented in section 6 and section 7 that a common
frame work tool output based on CA.
Cellular Automata
Artificial Intelligence (AI) is the exploration of
impersonating human mental ability in a computer.
Current AI methods predominantly succumb to two
general categories.
Explicit Modeling (Words, Images)
This is based on rules, frames and case based learning
has succeeded in some domains. This category cannot
handle the unseen cases.
16
1)
Cellular Automata
Cellular automata consist of a grid of cells with a
finite number of states. Cellular Automata (CA) is a
computing model which provides a good platform
for performing complex computations with the
available local information.
CA is defined a four tuple <G, Z, N, F>
Where G -> Grid (Set of cells)
Z -> Set of possible cell states
N -> Set which describe cells
neighborhoods
F -> Transition Function (Rules of
automata)
A CA displays three basic characteristics - locality,
infinite parallelism, and simplicity
Locality: CA is characterized by local connectivity
of its cells. The interactions among the cells are
defined on the basis of locality. Each cell can
communicate with adjacent cells. The transitions
defined between cells carry small amount of
information only. None of the cells connected will
have a global view of the complex system.
Parallelism: Parallel computing environment is
demanded for addressing most of the complex
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
computing problems. Most parallel computers
contain more than a few dozens of processors. CA
can achieve parallelism on a scale larger than
massively parallel computers. CA performs
computations in a distributed fashion on a spatially
extended grid. It differs from the conventional
approach to parallel computation in which a
problem is split into independent sub-problems,
each solved by a different processor; the solution of
sub-problems are subsequently combined to yield
the final solution
Simplicity: The basic unit of Cellular Automata
(CA) is a cell that has a simple structure evolving in
discrete time and space. Wolfram developed and
proposed a uniform cellular automata construction
which is a significant contribution in the field of
computing. This leads to the emergence of single
dimensional array approach with each cell has only
two states and three neighborhoods. It is very easy
to map the frame work of CA to the frame work of
many bioinformatics problems.
As Cellular Automata consists of a number of cells
structured in the form of a grid. The transitions
between the cells may depend on its own state and
the states of its neighboring cells. The equation one
sates that if ith cell have to make a transition, it has
to depend on own state, left neighbor and right
neighbor either.
qi(t + 1) = f (qi−1(t), qi(t), qi+1(t)) ---- Equation1.1
2) Wolfram’s two state three neighborhood
Cellular Automata
Figure 1, 2 demonstrate Wolfram’s two state three
neighborhoods, which consist of eight different
present state combinations. The rule 254 is used for
defining the transitions between the neighbors.
www.seipub.org/gpai
Each CA cell in the grid must be a memory element
(D Flip flop) with some combinational logic - an
XOR and or XNOR Gate (additive). Each cell will
be updated at every clock cycle. The transitions
between the cells will depend on the immediate
neighbors
TABLE1. RULE REPRESENTATION
Possible
Combinations
Binary
Equivalent of
Rule-254
(Next State)
111
110
101
100
011
010
001
000
1
1
1
1
1
1
1
0
The decimal equivalent of the next state function is
defined as the rule number of the CA cell
introduced by Wolfram. In a 2-state 3neighborhood CA, there are 256 distinct next state
functions. Out of 256 rules, rule 254 is
demonstrated. In Cellular Automata rules play a
vital role in developing a good classifier.
3) AIS-MACA (Artificial Immune System Based
Multiple Attractor Cellular Automata)
Multiple Attractor Cellular Automata which is a
special class of fuzzy cellular automata which was
introduced thirty years ago. It uses fuzzy logic to
handle real value attributes. The development
process / implementation of Multiple Attractor
Cellular Automata are administered by AIS
technique, a Clonal Algorithm with the underlying
theory of survival of the fittest gene. Genetic
algorithm framework is adapted to represent the
corresponding CA rule structure. So we have
named this special classifier as AIS-MACA. AISMACA can address many problems in
bioinformatics
FIGURE 1. NEIGHBORHOOD REPRESENTATION
17
www.seipub.org/gpai
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
FIGURE 2. WOLFRAM REPRESENTATION OF FLIP-FLOP CIRCUIT
Protein Coding Regions
Jesus P. Mena-Chalco at al has used Modified GaborWavelet Transform for addressing this issue. In this
connection, numerous coding DNA model-free
systems dependent upon the event of particular
examples of nucleotides at coding areas have been
proposed. Besides, these techniques have not been
totally suitable because of their reliance on an
observationally predefined window length needed for
a nearby dissection of a DNA locale. Authors present a
strategy dependent upon a changed Gabor-wavelet
transform for the ID of protein coding areas. This
novel convert is tuned to examine intermittent sign
parts and presents the focal point of being free of the
window length. We contrasted the execution of the
MGWT and different strategies by utilizing eukaryote
information sets. The effects indicate that MGWT beats
all evaluated model-autonomous strategies regarding
ID exactness. These effects demonstrate that the
wellspring of in any event some pieces of the ID lapses
handled by the past systems is the altered working
scale. The new system stays away from this wellspring
of blunders as well as makes an instrument accessible
for point by point investigation of the nucleotide event.
Changchuan Yin at al has proposed a strategy to
foresee protein coding areas is produced which is
dependent upon the way that the vast majority of exon
arrangements have a 3-base periodicity, while intron
groupings don't have this interesting characteristic.
The technique registers the 3-base periodicity and the
foundation clamor of the stepwise DNA sections of the
target DNA groupings utilizing nucleotide circulations
as a part of the three codon positions of the DNA
successions. Exon and intron successions might be
recognized from patterns of the degree of the 3-base
18
periodicity to the foundation commotion in the DNA
groupings.
Suprakash Datta at al used a DFT based gene
prediction for addressing this problem. Authors
provided theoretical concept of three periodicity
properties observed in protein coding regions in
genomic DNA. They proposed new criteria for
classification based on traditional frequency
approaches of coding regions.
Pradipta Maji has proposed a Cellular Automata (CA)
based example classifier to recognize the coding locale
of a DNA succession. CA is exceptionally
straightforward, proficient, and handles more faultless
classifiers than that has long ago been acquired for a
reach of diverse succession lengths. Far reaching trial
outcomes make that the proposed classifier is a
financially savvy elective in protein coding area
identification issue.
We also proposed a classifier based on MACA for
predicting the protein coding region with an accuracy
of 83%. It is later strengthened with an AIS technique
whose predicted accuracy is more compared to the
earlier.
Promoter Region Prediction
Anne Vanet at al proposed some algorithms for
predicting the promoter regions. This paper displays a
study of right now accessible scientific models and
algorithmically routines for attempting to recognize
promoter successions. The routines concern both
looking in a genome for an awhile ago characterized
agreement and concentrating an accord from a set of
arrangements. Such routines were frequently
customized for either eukaryotes or prokaryotes in
spite of the fact that this does not block utilization of
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
the same system for both sorts of organic entities. The
overview consequently blankets all techniques; in any
case, attention is set on prokaryotic promoter
arrangement identification. Illustrative requisitions of
the fundamental concentrating calculations are given
for three microbes
Vetriselvi Rangnanan at al made a dissection of
different anticipated structural lands of promoter
districts in prokaryotic and in addition eukaryotic
genomes had prior shown that they have a few normal
characteristics, for example, lower steadiness, higher
curve and less bendability, when contrasted and their
neighboring areas. In light of the distinction in
solidness between neighboring upstream and
downstream areas in the region of tentatively dead set
translation begin destinations, a promoter forecast
calculation has been produced to distinguish
prokaryotic promoter groupings in entire genomes
Jih-Wei Hung has developed an effective forecast
calculation that can expand the recognition (power =1 false negative) of promoter. Authors introduce two
strategies that utilize the machine force to ascertain all
conceivable examples which are the conceivable
characteristics of promoters. The primary strategy we
exhibit FTSS (Fixed Transcriptional Start Site) utilizes
the known TSS positions of promoter arrangements to
prepare the score record that helps us in promoter
forecast. The other strategy is NTSS (Nonfixed TSS).
The TSS positions of promoter arrangements utilized
as a part of NTSS are thought to be obscure, and NTSS
won't take irrefutably the positions of Tsss into
attention. By the exploratory effects, our expectation
has higher right rate than different past systems
Marshall S.Z. Horwitz at al have chosen an assembly
of Escherichia coli promoters from irregular DNA
groupings by swapping 19 base sets at the -35
promoter area of the etracycline safety gene te" of the
plasmid pbr322. Substitution of 19 base sets with
artificially blended irregular groupings brings about a
greatest of 419 (something like 3 x 1011) conceivable
swap groupings. From a populace of in the ballpark of
1000 microscopic organisms harboring plasmids with
these irregular substitutions, tetracycline choice has
uncovered numerous practical -35 promoter
successions. These promoters have held just halfway.
Homology to the -35 promoter accord grouping. In
three of these promoters, the agreement operator
moves 10 nucleotides downstream, permitting the
RNA polymerase to distinguish an alternate Pribnow
box from inside the definitive pbr322 succession. Two
www.seipub.org/gpai
of the successions advertise translation
determinedly than the local promoter
more
We also proposed a classifier based on MACA for
predicting the promoter region with an accuracy of
89%. It is later strengthened with an AIS technique
whose predicted accuracy is more compared to the
earlier
Protein Structure Prediction
Jooyoung Lee, Anticipating protein 3d structures from
the amino harsh corrosive arrangement still stays as an
unsolved issue after five decades’ endeavors. In the
event that the target protein has a homologue recently
understood, the undertaking is moderately simple and
high-determination models could be assembled by
replicating the skeleton of the settled structure. Be that
as it may, such a modeling methodology does not help
address the inquiry of how and why a protein
embraces its particular structure. In the event that
structure homologues (sporadically analogues) don't
exist, or exist can't be recognized yet, models must be
developed starting with no outside help. This strategy,
called stomach muscle initio modeling, is crucial for a
complete answer for the protein structure forecast
issue; it can likewise help us comprehend the
physicochemical rule of how proteins crease in nature.
Right now, the precision of abdominal muscle initio
modeling is low and the victory is restricted to little
proteins (<100 buildups). In this section, we give an
audit on the field of stomach muscle initio modeling.
Center will be put on three key components of the
modeling calculations: vigor capacity, conformational
inquiry, and model determination. Advances and
developments of an assortment of calculations will be
talked about.
Scott C. Schmidler has provided a significant part for
statisticians in the period of the Human Genome
Project has improved in the rising territory of
structural bioinformatics". Succession dissection and
structure expectation for biopolymers is a vital venture
on the way to transforming recently sequenced
genomic
information
into
naturally
and
pharmaceutically pertinent data in backing of subatomic solution. We depict our work on Bayesian
models for expectation of protein structure from
succession, in view of dissection of a database of
tentatively resolved protein structures (Datta,
Suprakash et al., 2005)
We have advanced fragment based models of protein
optional structure awhile ago which catch central parts
19
www.seipub.org/gpai
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
of the protein collapsing procedure. These models give
prescient executions at the level of the best accessible
strategies in the. Here we indicate that this Bayesian
schema is regularly summed up to fuse data
dependent upon non-neighborhood arrangement
cooperation’s. We show this thought by displaying a
basic model for strand blending and a Markov chain
Monte Carlo (MCMC) calculation for induction. We
apply the methodology to forecast 3-dimensional
contacts for two illustration proteins.
David E. Kim al worked on PSP and the arrangements
submitted to the server are parsed into putative
dominions and structural models are produced
utilizing either near displaying or a new structure
expectation routines. Assuming that a certain match to
a protein of known structure is found utilizing BLAST,
PSI-BLAST, Ffas03 or 3d-Jury, it is utilized as a pattern
for relative displaying. Assuming that no match is
found, structure expectations are made utilizing the
Rosetta piece insertion strategy again. Test atomic
attractive thunder (NMR) obligations information can
likewise be submitted with an inquiry succession for
Rosetta NMR all over again structure determination.
Other current capacities might be the expectation of
the impacts of transformations on protein–protein
communications utilizing computational interface
alanine filtering. The Rosetta protein configuration and
protein–protein docking systems will soon be
accessible through the server.
We also proposed a classifier based on MACA for
predicting the secondary structure of protein with an
accuracy of 85%. It is later strengthened with an AIS
technique whose predicted accuracy is more compared
to the earlier.
CA Based Integrated Tool For Predicting
Protein Coding and Promoter Region
We have proposed an integrated tool showed in Fig 3.4
for predicting the protein coding and promoter region
using the same classifier. This has recognized an
international accload. We then extended this
framework to predict the secondary structure of the
protein. The classifier is trained with Fickett and
Toung data sets for protein coding regions and E Coli
data sets for promoter prediction. This proposed
algorithm can handle different sequence lengths input
and can give an accuracy of 82%.
FIGURE 3. COMMON PROTEIN CODING & PROMOTER PREDICTION INTERFACE
FIGURE 4. BOUNDARY REPORTING FOR PROMOTER
20
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
www.seipub.org/gpai
Systems,” Proceedings of the von Karman Lecture series
Conclusion
This paper is aimed at bringing an intuition towards
application of CA in bioinformatics. A common frame
work is required for addressing major problems in
bioinformatics (Crick Francis et al., 1970). CA frame
work can be used to address many problems in like
protein structure prediction, RNA structure prediction,
predicting the splicing pattern of any primary
transcript and analysis of information content in DNA,
RNA, protein sequences and structure and many more.
on Intelligent Systems for Aeronautics, Belgium, May
2002.
Mena-Chalco, Jesús P., Helaine Carrer, Yossi Zana, and
Roberto M. Cesar. "Identification of protein coding
regions using the modified Gabor-wavelet transform."
Computational Biology and Bioinformatics, IEEE/ACM
Transactions on 5, no. 2 (2008): 198-207.
Mishra, Rakesh K., Vijaya Gopal, and Dipankar Chatterji.
"Correlation between the DNA supercoiling and the
REFERENCES
initiation of transcription by RNA polymerase in vitro:
Crick, Francis. "Central dogma of molecular biology." Nature
role of the sequences upstream of the promoter region."
FEBS letters 260, no. 2 (1990): 273-276.
227, no. 5258 (1970): pp. 561-563.
Datta, Suprakash, and Amir Asif. "A fast DFT based gene
Pokkuluri Kiran Sree , I.Ramesh Babu , “Investigating an
prediction algorithm for identification of protein coding
Artificial Immune System to Strengthen the Protein
regions." In Acoustics, Speech, and Signal Processing,
Structure
2005.
Identification
Proceedings.
(ICASSP'05).
IEEE
International
Prediction
using
and
Protein
Cellular
Coding
Automata
Region
Classifier”
International Journal of Bioinformatics Research and
Conference on, vol. 5, pp. v-653. IEEE, 2005.
Dieterich, Christoph, Steffen Grossmann, Andrea Tanzer,
Stefan Röpcke, Peter F. Arndt, Peter F. Stadler, and
Martin Vingron. "Comparative promoter region analysis
Applications ,Vol 5,Number 6,pp 647-662, ISSN : 17445493.
Pokkuluri Kiran Sree, I.Ramesh Babu, PSMACA: An
powered by CORG." BMC genomics 6, no. 1 (2005): 24.
Automated Protein Structure Prediction using MACA
E E Snyder and G D Stormo,”Identification of coding regions
(Multiple Attractor Cellular Automata). Journal of
in genomic DNA sequences: an application of dynamic
Bioinformatics and Intelligent Control (JBIC) in Volume 2
programming and neural networks “Nucleic Acids Res.
Number 3, pp 211-215 ,2013
1993 February 11; 21(3): 607–613.
Pokkuluri Kiran Sree at al, Identification of Promoter Region
Fickett, James W., and Chang-Shung Tung. "Assessment of
in Genomic DNA Using Cellular Automata Based Text
protein coding measures." Nucleic acids research 20, no.
Clustering. The International Arab Journal of Information
24 1992: pp. 6441-6450.
Technology (IAJIT), Volume 7, No 1, 2010, pp 75-78.
Gish, Warren, and David J. States. "Identification of protein
Pokkuluri Kiran Sree, I.Ramesh Babu, An extensive report on
coding regions by database similarity search." Nature
Cellular Automata based Artificial Immune System for
genetics 3, no. 3 (1993): 266-272.
strengthening Automated Protein Prediction. Advances
Huang, Jih-Wei. "Promoter Prediction in DNA Sequences."
PhD diss., National Sun Yat-sen University, 2003.
Horwitz, M. S., and Lawrence A. Loeb. "Promoters selected
from random DNA sequences." Proceedings of the
National Academy of Sciences 83, no. 19 (1986): 74057409.
J.
Von.
Neumann,
‘The
Theory of
Self-Reproducing
Automata,” A. W. Burks, Ed.University of Illinois Press,
Urbana and London, 1966.
KrishnaKumar,
“Challenging
K.,
Kaneshige,
Aerospace
J.,
and
Problems
Satyadas,
for
A.,
Intelligent
in Biomedical Engineering Research (ABER) Volume 1
Issue 3, September 2013, pp 45-51.
Pokkuluri Kiran Sree, I.Ramesh Babu, “HMACA: Towards
proposing Cellular Automata based tool for protein
coding, promoter region identification and protein
structure prediction”. International Journal of Research
in Computer Applications & Information Technology,
Volume 1 Number 1, pp 26-31, 2013.
Pokkuluri Kiran Sree, I.Ramesh Babu,” Multiple Attractor
Cellular Automata (MACA) for Addressing Major
Problems” in Bioinformatics in Review of Bioinformatics
21
www.seipub.org/gpai
Global Perspectives on Artificial Intelligence (GPAI) Volume 2 Issue 2, April 2014
and Biometrics (RBB) Volume 2 Issue 3, September 2013,
Cellular Automata Based Pattern Classifier, Proceedings
pp70-76.
of 9th International Conference on Database Systems,
P. Flocchini, F. Geurts, A. Mingarelli, and N. Santoro (2000),
“Convergence and Aperiodicity in Fuzzy Cellular
Automata: Revisiting Rule 90,” Physica D.
P. Maji and P. P. Chaudhuri (2004), “FMACA: A Fuzzy
Salzberg, Steven. "Locating protein coding regions in human
DNA using a decision tree algorithm." Journal of
Computational Biology 2, no. 3 (1995): 473-485.
Cellular Automata Based Pattern Classifier,” Proceedings
Snyder, Eric E., and Gary D. Stormo. "Identification of
of 9th International Conference on Database Systems,
protein coding regions in genomic DNA." Journal of
Korea, pp. 494–505, 2004.
molecular biology 248, no. 1 (1995): 1-18.
P. Maji and P. P. Chaudhuri, “Fuzzy Cellular Automata for
Modeling Pattern Classifier,” IEICE, (2004).
P.Kiran Sree, I .Ramesh Babu, Identification of Protein
Coding Regions in Genomic DNA Using Unsupervised
Yin, Changchuan, and Stephen S-T. Yau. "Prediction of
protein coding regions by the 3-base periodicity analysis
of a DNA sequence." Journal of theoretical biology 247,
no. 4 (2007): 687-694.
FMACA Based Pattern Classifier in International Journal
Wolfram, Stephen. "Universality and complexity in cellular
of Computer Science & Network Security with ISSN:
automata." Physica D: Nonlinear Phenomena 10, no.1,
1738-7906 Volume Number: Vol. 8, No. 1, 2008.
1984: pp. 1-35.
P. Maji and P. P. Chaudhuri (2004), FMACA: A Fuzzy
22
Korea, pp. 494–505, 2004.
| 5cs.CE
|
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
arXiv:1708.09064v1 [math.AG] 30 Aug 2017
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
A BSTRACT. We bring examples of toric varieties blown up at a point in the torus that do
not have finitely generated Cox rings. These examples are generalizations of [6] where
toric surfaces of Picard number 1 were studied. In this article we consider toric varieties of
higher Picard number and higher dimension. In particular, we bring examples of weighted
projective 3-spaces blown up at a point that do not have finitely generated Cox rings.
1. I NTRODUCTION
We work over an algebraically closed field k of characteristic 0.
Our aim in this article is to bring examples of varieties X that do not have finitely generated Cox rings. Our varieties X are toric varieties X∆ blown up at a point t0 in the torus.
In [6] we constructed examples of such toric surfaces X∆ of Picard number 1. In this article we generalize this construction to toric varieties of higher Picard number and higher
dimension.
Let us recall the definition by Hu and Keel [9] of the Cox ring of a normal projective
variety X:
M
Cox(X) =
H0 (X, OX (D)).
[D]∈Cl(X)
Giving a ring structure to this space involves some choices, but finite generation of the resulting k-algebra does not depend on the choices. A normal projective Q-factorial variety
X is called a Mori Dream Space (MDS) if Cox(X) is a finitely generated k-algebra.
The construction in [6] was based on the examples of blowups at a point of weighted
projective planes by Goto, Nishida and Watanabe [7] and the geometric description of
these examples by Castravet and Tevelev [3]. A basic fact about Cox rings is that on a MDS
X every nef divisor is semiample (i.e. there exists a positive multiple of the divisor that has
no base locus and defines a morphism X → Pn ). To prove that X is not a MDS, it suffices
to find a nef divisor D that is not semiample. The examples in [6] have Picard number 2
and there is essentially a unique choice for D. The class of D necessarily has to lie on the
boundary of the (2-dimensional) nef cone. One of the boundary rays is generated by the
class H of the pullback of an ample divisor on X∆ , which is clearly semiample. It follows
that D must lie on the other boundary ray. In the case where X is a surface, this other
The second author was supported by a NSERC Discovery grant.
1
2
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
boundary ray is determined if we can find a curve C of negative self-intersection on X,
different from the exceptional curve.
In general, the existence of a nef divisor D on X that is not semiample is only a sufficient
condition for X being a non-MDS. When X∆ is a weighted projective plane P(a, b, c), then
Cutkosky [4] has shown that X is a MDS if and only if the divisor D as above is semiample.
There are two essential differences in the proof of non-finite generation when going to
higher Picard number or higher dimension. In the case of surfaces X with Picard number
p > 2 we still look for a curve C ⊂ X of negative self-intersection. This curve now defines
a (p − 1)-dimensional face of the nef cone and there is no obvious choice for the nonsemiample divisor D. We show that a general divisor on this face is not semiample.
In dimension greater than 2 we will encounter normal projective varieties X that are
not Q-factorial. For such varieties the Cox ring and MDS are defined in the same way
as above. (This generalizes slightly the definition of Hu and Keel [9] who required a
MDS to be Q-factorial.) In this greater generality, if X has a free class group and a finitely
generated Cox ring, then its cones of effective, moving, semiample and nef divisors are
polyhedral [1, Theorem 4.2, Theorem 7.3, Remark 7.6]. Moreover, the cones of nef Cartier
divisors and semiample Cartier divisors coincide [1, Corollary 7.4]. In our examples we
find nef Cartier divisors D that are not semiample and hence X is not a MDS.
Acknowledgment. We thank Jürgen Hausen for explaining us various details in the
definition of Cox rings.
2. S TATEMENT
OF THE MAIN RESULTS .
We use the terminology of toric varieties from [5]. Let X∆ be the toric variety defined
by a rational convex polytope ∆ and let X be the blowup of X∆ at a general point, which
we can assume to be the identity point t0 = (1, 1, . . . , 1) in the torus. We are interested in
the Cox ring of X.
2.1. The case of surfaces. Let ∆ be a convex plane 4-gon with rational vertices (0, 0),
(0, 1), PL = (xL , yL ), PR = (xR , yR ), where xL < 0 and xR > 0 (see Figure 1). The polygon
can equivalently be defined by the slopes of its sides, s1 , s2 , s3 , s4 . We will assume that
the slope s2 of the side connecting (0, 0) and PR satisfies 0 ≤ s2 < 1. When xR ≤ 1, this
can always be achieved without changing the isomorphism class of X∆ by applying an
integral shear transformation (x, y) 7→ (x, y + ax) for some a ∈ Z to the polytope.
Choose m > 0 such that m∆ is integral. We study lattice points in m∆. Let us denote
by column c in m∆ the set of lattice points with first coordinate x = c.
Theorem 2.1. Let ∆ be a rational plane 4-gon as above. Assume that 0 ≤ s2 < 1 and let m > 0
be sufficiently large and divisible so that m∆ is integral. The variety X = Blt0 X∆ is not a MDS if
the following two conditions are satisfied:
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
3
y
1
s4
(x L,y L)
s3
∆
(xR,yR)
s1
s2
x
0
F IGURE 1. Polygon ∆.
(1) Let w = xR − xL be the width of ∆. Then w < 1.
(2) Let the column mxL + 1 in m∆ consist of n points (mxL + 1, b + i), i = 0, . . . , n − 1.
Then
(a) columns mxR , mxR − 1, . . . , mxR − n + 1 in m∆ have 1, 2, . . . , n lattice points,
respectively;
(b) myL is not equal to b + i, i = 1, . . . , n − 1.
If the width w = 1 or ∆ degenerates to a triangle with slopes s1 = s2 , then X is not a MDS if in
addition to (1’) w ≤ 1 and (2) the following holds:
(3) Let s =
yR −yL
w
be the slope of the line joining the left and right vertices. Then myL 6= b−ns.
Example 2.2. Consider ∆ with (xL , yL ) = (−3/4, 1/2) and (xR , yR ) = (1/4, 3/4).
(−3,2)
4∆
(1,3)
(−2,3)
(1,1)
(−2,−3)
(3,−1)
F IGURE 2. Polygon 4∆ and the corresponding (outer) normal fan.
In this case w = 1 and n = 1. When n = 1 condition (2) of the theorem is vacuously true
and condition (3) states that the single lattice point in column mxL + 1 does not lie on the line
joining the left and right vertices. (These conditions still hold after applying an integral shear
transformation as above, hence the assumption 0 ≤ s2 < 1 is not necessary in the n = 1 case.)
This gives an example of a surface X of Picard number 3 that is not a MDS. Notice that if we move
4
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
the vertex (xR , yR ) to (1/4, 1) or (1/4, 7/6), but not (1/4, 1/2), the theorem applies and we again
get an example of a non-MDS.
When ∆ degenerates to a triangle then Theorem 2.1 reduces to the case considered in
[6]. In the case of a triangle, He [8] has generalized condition (2a) to a weaker one. We
expect that such a generalization also exists in the case of 4-gons.
By a result of Okawa [10], if Y → X is a surjective morphism of (not necessarily Qfactorial) normal projective varieties, and X is not a MDS, then Y is also not a MDS. Thus,
if X = Blt0 X∆ is not a MDS, we can replace X∆ with any toric blowup X∆^ to produce
non-MDS of higher Picard number. Our methods do not give examples of surfaces other
than the ones obtained from a plane 4-gon. The proof below shows that finite generation
of the Cox ring of X only depends on the singularities at the two torus fixed points corresponding to PL , PR and the curve of negative self-intersection C ⊂ X passing through
these points. If X∆ has toric divisors that do not pass through the two torus fixed points,
then these can be contracted.
2.2. Higher dimensional varieties. We first generalize Theorem 2.1 to dimension 3 and
then discuss generalizations to dimension 4 and higher.
Let now ∆ be a rational convex 3-dimensional polytope with vertices (0, 0, 0), (0, 1, 0),
(0, 0, 1), PL = (xL , yL , zL ), PR = (xR , yR , zR ), where xL < 0 and xR > 0. We allow ∆ to
degenerate to a tetrahedron, where the points (0, 0, 0), PL, PR are collinear.
y
1
(xL , yL , zL )
∆
(xR , yR , zR )
0
x
1
z
F IGURE 3. Polytope ∆.
We assume that 0 ≤ yxRR , xzRR < 1. When xR ≤ 1, this can be achieved by applying an
integral shear transformation to the polytope.
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
5
Let m∆ be integral. A slice c of m∆ consists of all lattice points in m∆ with first coordinate x = c. Such a slice forms a right triangle with n lattice points on each side. We say
that the slice has size n.
Theorem 2.3. Let ∆ be a 3-dimensional polytope as above. Assume that 0 ≤ yxRR , xzRR < 1 and let
m > 0 be sufficiently large and divisible so that m∆ is integral. The variety X = Blt0 X∆ is not a
MDS if the following three conditions are satisfied:
(1) Let w = xR − xL be the width of ∆. Then w ≤ 1.
(2) Let the slice mxL + 1 in m∆ have size n with points (mxL + 1, b + i, c + j), i, j ≥ 0,
i + j < n. Then
(a) the slices mxR , mxR − 1, . . . , mxR − n + 1 in m∆ have size 1, 2, . . . , n, respectively;
(b) (myL , mzL ) is not equal to (b + i, c + j) for any i, j ≥ 1, i + j < n.
−yL
−zL
(3) Let sy = yRw
, sz = zRw
be the two slopes of the line joining left and right vertices. Then
(a) (myL , mzL ) 6= (b − nsy , c − nsz );
(b)
(i) if myL = b − nsy and c < mzL < c + n, then sy 6= 0;
(ii) if mzL = c − nsz and b < myL < b + n, then sz 6= 0;
(iii) if myL + mzL = b − nsy + c − nsz and b < myL , c < mzL , then sy + sz 6= −1.
The case n = 1 of the theorem simplifies considerably as follows.
Corollary 2.4. Let ∆ be a 3-dimensional polytope as above and let m > 0 be sufficiently large
and divisible so that m∆ is integral. The variety X = Blt0 X∆ is not a MDS if the following three
conditions are satisfied:
(1) w = xR − xL ≤ 1.
(2) The slice mxL + 1 in m∆ consists of a single lattice point P.
(3) The point P does not lie on the line joining the left and right vertices of m∆.
Theorem 2.3 in particular applies to the case where ∆ is a tetrahedron. The statement
also simplifies in this case.
Corollary 2.5. Let ∆ be a 3-dimensional tetrahedron as above, where the points (0, 0, 0), PL, PR
are collinear. Let m > 0 be sufficiently large and divisible so that m∆ is integral. The variety
X = Blt0 X∆ is not a MDS if the following three conditions are satisfied:
(1) w = xR − xL ≤ 1.
(2) Let the slice mxL + 1 in m∆ have size n. Then the slice mxR − n + 1 in m∆ has size n.
−yL
−zL
, sz = zRw
be the two slopes of the line joining left and right vertices. Then
(3) Let sy = yRw
2
n(sy , sz ) ∈
/Z.
6
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
We will study the tetrahedron case further to find examples where X∆ is a weighted
projective space P(a, b, c, d). Let (xL , xR , y0 , z0) be such that
(xL , yL , zL) = xL (1, y0 , z0),
(xR , yR , zR) = xR (1, y0, z0 ).
Then the 4-tuple of rational numbers (xL , xR , y0 , z0) determines the tetrahedron ∆. The
normal fan to ∆ has rays generated by
1
1
, −1, −1), (y0 + z0 − , −1, −1), (−y0, 1, 0), (−z0, 0, 1).
xL
xR
The slice mxL + 1 in m∆ can be identified with lattice points in the triangle with vertices
(y0 , z0 ), (y0 − x1L , z0 ), (y0, z0 − x1L ). It has size
(2.1)
(y0 + z0 −
1
⌋ − ⌈y0 ⌉ − ⌈z0 ⌉.
xL
Similarly, the slice mxR − n + 1 in m∆ can be identified with lattice points in the triangle
with vertices (n − 1)(y0, z0 ), (n − 1)(y0 − x1R , z0 ), (n − 1)(y0 , z0 − x1R ). It has size
n = 1 + ⌊y0 + z0 −
1 − ⌈(n − 1)(y0 + z0 −
1
)⌉ + ⌊(n − 1)y0 ⌋ + ⌊(n − 1)z0 ⌋.
xR
We can now state Corollary 2.5 in terms of (xL , xR , y0 , z0 ).
Corollary 2.6. Let ∆ be a tetrahedron given by the 4-tuple of rational numbers (xL , xR , y0, z0 ),
with xL < 0 and xR > 0. The variety X = Blt0 X∆ is not a MDS if the following three conditions
are satisfied:
(1) w = xR − xL ≤ 1.
(2) Let
n = 1 + ⌊y0 + z0 −
1
⌋ − ⌈y0 ⌉ − ⌈z0 ⌉.
xL
Then also
n = 1 − ⌈(n − 1)(y0 + z0 −
1
)⌉ + ⌊(n − 1)y0 ⌋ + ⌊(n − 1)z0 ⌋.
xR
(3) n(y0 , z0) ∈
/ Z2 .
Note that the statements of Corollaries 2.4, 2.5 and 2.6 do not depend on the assumption 0 ≤ yxRR , xzRR < 1. The three conditions are the same after applying an integral shear
transformation as above.
Example 2.7. Let xL = −3/5, xR = 6/17, y0 = 1/3, z0 = 1/2. The three conditions of Corollary 2.6 are satisfied with w = 81/85 and n = 1. The normal fan has rays generated by
(5, −2, −2), (−2, −1, −1), (−1, 3, 0), (−1, 0, 2).
These vectors generate the lattice Z3 , and X∆ is the weighted projective space P(17, 20, 18, 27).
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
7
Example 2.8. Let xL = −2/3, xR = 1/3, y0 = 1/2, z0 = 1/2. The three conditions are again
satisfied with w = 1 and n = 1. The normal fan has rays generated by
(5, −2, −2), (2, −3, −3), (−1, 2, 0), (−1, 0, 2).
These vectors generate a sublattice of index 2 in Z3 , and X∆ is the quotient of P(2, 6, 11, 11) by a
2-element subgroup of the torus.
Example 2.9. Let xL = −5/18, xR = 5/7, y0 = 2/5, z0 = 1. Here w = 125/126 < 1 and n = 4.
However,
1
1 − ⌈(n − 1)(y0 + z0 − )⌉ + ⌊(n − 1)y0 ⌋ + ⌊(n − 1)z0 ⌋ = 5,
xR
and hence Corollary 2.6 does not apply to the blowup of X∆ = P(7, 18, 5, 25).
Remark 2.10. Given a polytope ∆, one can project it to the xy-plane or the xz-plane to get a
plane 4-gon. The slice c in m∆ has size no bigger than the corresponding column c in the
projection. This implies that if the projection of ∆ satisfies the conditions of Theorem 2.1
with n = 1, then ∆ satisfies the conditions in Corollary 2.4. Thus, one can construct 3dimensional polytopes by lifting 2-dimensional polygons. However, Examples 2.7 and
2.8 are genuinely new: they can not be reduced to 2-dimensional cases by projection. This
can be seen as follows. The projection of the tetrahedron to the xy-plane is a triangle
determined by (xL , xR , y0). The three conditions of Theorem 2.1 in the case n = 1 are:
(1) w = xR − xL ≤ 1.
(2) 1 = 1 + ⌊y0 − x1L ⌋ − ⌈y0 ⌉.
(3) y0 ∈
/ Z.
In Examples 2.7 and 2.8 the second condition is not satisfied. Similarly, projecting to the
xz-plane, the condition 1 = 1 + ⌊z0 − x1L ⌋ − ⌈z0 ⌉ is not satisfied.
In [6] we gave an algorithm for checking if the blowup of a weighted projective plane
satisfies the assumptions of Theorem 2.1. We will state a similar result in dimension 3.
Consider the weighted projective space P(a, b, c1, c2). We say that (e, f, g1, g2) ∈ Z4>0 is
a relation in degree d if
ea + fb = g1 c1 = g2 c2 = d.
We require for a relation (e, f, g1, g2) that
gcd(e, f, g1) = gcd(e, f, g2) = gcd(g1 , g2) = 1.
(If x, y, z1, z2 are variables of degree a, b, c1, c2 respectively, then xe yf , zg1 1 , zg2 2 are three
monomials of degree d. They correspond to the three lattice points in ∆.)
Theorem 2.11. Let P(a, b, c1, c2 ) be a weighted projective space with a relation (e, f, g1, g2) in
degree d. Then Blt0 P(a, b, c1, c2) is not a MDS if the following three conditions are satisfied:
8
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
(1) Let
w=
d3
.
abc1 c2
Then w ≤ 1.
(2) Consider integers δ1 , δ2 ≤ 0 such that the vector
1
δ2
δ1
(e, −f)
(b, a) +
+
g1 g2
g1 g2
has non-negative integer entries. The set of such (δ1 , δ2 ) forms a slice of size n. Then the
integers γ1 , γ2 ≥ 0 such that
γ1 γ2
n−1
(e, −f)
(b, a) +
+
g1 g2
g1 g2
has non-negative integer entries must also form a slice of size n.
(3) With n as above,
n
(b, a) ∈
/ Z2 .
g1 g2
To check if some P(a, b, c1, c2) satisfies the assumptions of the theorem, we first determine g1 , g2 . The conditions g1 c1 = g2 c2 and gcd(g1 , g2) = 1 imply that g1 = c2 / gcd(c1 , c2),
g2 = c1 / gcd(c1, c2). After that we check that w ≤ 1, find e, f, and compute the two slices.
Table 1 lists examples with a, b, c1, c2 < 50 that were found using a computer. We
have omitted some isomorphic weighted projective spaces from this table. For exam∼ P(da, db, dc1, dc2) for any d > 0. Similarly, if a prime p divides all
ple, P(a, b, c1, c2) =
numbers a, b, c1, c2 except one, we can divide the three numbers by p to get isomorphic
weighted projective spaces. The table lists only spaces P(a, b, c1, c2) where every triple in
{a, b, c1, c2} has no common divisor greater than 1.
Corollaries 2.4, 2.5 and 2.6 have obvious generalizations to higher dimension. Similarly,
Theorem 2.11 can be generalized to dimension r. We need to consider weighted projective
spaces P(a, b, c1, c2, . . . , cr−1) with a relation (e, f, g1, g2, . . . , gr−1). Wherever there is a
term with c1 and c2 (or g1 , g2 ) in Theorem 2.11, we need to add terms with c3 , . . . , cr−1
(or g3 , . . . , gr−1). Table 2 lists weighted projective 4-spaces with a, b, ci < 65. Again, only
normalized numbers are listed.
3. P ROOF
OF
T HEOREM 2.1
We use standard notation from birational geometry. Let N1 (X) (resp. N1 (X)) be the
group of numerical equivalence classes of Cartier divisors (resp. 1-cycles). Let NE(X) ⊂
N1 (X)R be the closed Kleiman-Mori cone of curves, and Nef(X) ⊂ N1 (X)R the dual cone
of nef divisors.
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
P(a, b, c1, c2) (e, f, g1, g2)
P(47, 13, 12, 30) (1, 1, 5, 2)
P(19, 41, 15, 20) (1, 1, 4, 3)
P(43, 17, 15, 20) (1, 1, 4, 3)
P(26, 49, 15, 25) (1, 1, 5, 3)
P(11, 32, 18, 27) (2, 1, 3, 2)
P(13, 28, 18, 27) (2, 1, 3, 2)
P(17, 20, 18, 27) (2, 1, 3, 2)
P(47, 7, 18, 27) (1, 1, 3, 2)
P(23, 44, 18, 45) (2, 1, 5, 2)
P(29, 32, 18, 45) (2, 1, 5, 2)
P(23, 20, 22, 33) (2, 1, 3, 2)
P(25, 16, 22, 33) (2, 1, 3, 2)
P(29, 20, 26, 39) (2, 1, 3, 2)
n
1
3
1
3
2
2
1
1
2
1
1
1
1
P(a, b, c1, c2) (e, f, g1, g2 )
P(31, 16, 26, 39) (2, 1, 3, 2)
P(29, 50, 27, 36) (2, 1, 4, 3)
P(31, 46, 27, 36) (2, 1, 4, 3)
P(35, 38, 27, 36) (2, 1, 4, 3)
P(43, 49, 27, 45) (2, 1, 5, 3)
P(44, 47, 27, 45) (2, 1, 5, 3)
P(17, 33, 28, 42) (3, 1, 3, 2)
P(19, 27, 28, 42) (3, 1, 3, 2)
P(37, 16, 30, 45) (2, 1, 3, 2)
P(23, 27, 32, 48) (3, 1, 3, 2)
P(43, 46, 33, 44) (2, 1, 4, 3)
P(47, 38, 33, 44) (2, 1, 4, 3)
P(49, 34, 33, 44) (2, 1, 4, 3)
9
n
1
2
1
1
1
1
1
1
1
1
1
1
1
TABLE 1. Weighted projective spaces P(a, b, c1, c2), a, b, c1, c2 < 50, with
relation (e, f, g1, g2), that satisfy the conditions of Theorem 2.11.
P(a, b, c1, c2, c3 ) (e, f, g1, g2, g3)
P(47, 13, 12, 30, 60) (1, 1, 5, 2, 1)
P(19, 11, 13, 52, 52) (1, 3, 4, 1, 1)
P(21, 10, 13, 52, 52) (2, 1, 4, 1, 1)
P(19, 41, 15, 20, 60) (1, 1, 4, 3, 1)
P(43, 17, 15, 20, 60) (1, 1, 4, 3, 1)
P(22, 7, 17, 51, 51)
(2, 1, 3, 1, 1)
P(11, 32, 18, 27, 54) (2, 1, 3, 2, 1)
P(13, 28, 18, 27, 54) (2, 1, 3, 2, 1)
n
1
3
1
3
1
1
2
2
P(a, b, c1, c2, c3) (e, f, g1, g2, g3)
P(17, 20, 18, 27, 54) (2, 1, 3, 2, 1)
P(47, 7, 18, 27, 54)
(1, 1, 3, 2, 1)
P(25, 7, 19, 57, 57)
(2, 1, 3, 1, 1)
P(53, 7, 20, 30, 60)
(1, 1, 3, 2, 1)
P(15, 7, 26, 52, 52)
(3, 1, 2, 1, 1)
P(9, 13, 29, 58, 58)
(5, 1, 2, 1, 1)
P(17, 7, 29, 58, 58)
(3, 1, 2, 1, 1)
P(19, 7, 32, 64, 64)
(3, 1, 2, 1, 1)
n
1
1
1
1
1
1
1
1
TABLE 2. Weighted projective spaces P(a, b, c1, c2, c3), a, b, c1, c2, c3 < 65,
with relation (e, f, g1, g2 , g3) that satisfy the conditions of Theorem 2.11 in
dimension 4.
We prove Theorem 2.1 by contradiction. We assume that X is a MDS and produce a
nef divisor D that is not semiample. Note that X being a MDS implies that its nef cone is
polyhedral, generated by a finite number of semiample divisor classes.
Let ∆ be a plane 4-gon as in the theorem. The toric variety X∆ is Q-factorial and has
Picard number 2. The blowup X has Picard number 3. (We will deal with the case where
∆ is a triangle or w = 1 later.) The 4-gon contains two lattice points, (0, 0) and (0, 1).
10
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
Consider the irreducible curve in the torus T defined by the vanishing of the binomial
χ(0,0) − χ(0,1) = 1 − y,
and let C ⊂ X∆ be its closure. Considering C as a Q-Cartier divisor in X∆ , it has class
corresponding to the polygon ∆. This implies that its self-intersection number is
2
C = 2Area(∆) = w.
If now C is the strict transform of C in X, then C has divisor class π∗ C−E, where π : X → X∆
is the blowup map and E is the exceptional divisor. Hence C2 = w − 1 < 0. This implies
that C defines an extremal ray in the cone NE(X) and C⊥ defines a 2-dimensional face in
the 3-dimensional nef cone of X. We will show that a general divisor D ∈ C⊥ ∩ Nef(X) is
not semiample.
Let us start by describing the face of the nef cone defined by C⊥ . A nef divisor in X
has the form H − aE, where a ≥ 0 and H is the pullback of a nef divisor in X∆ . We
may assume that a 6= 0, and even more specifically that a = 1. Indeed, if a = 0 and
(H − aE) · C = 0, then also H = 0 because C is ample on X∆ . The divisor H corresponds to
a convex polygon with sides parallel to the sides of ∆. (The polygon may be degenerate
if some side has length 0). Let us define the width of H as the width of the corresponding
polygon.
Lemma 3.1. A nef divisor H − E lies in C⊥ if and only if the width of H is equal to 1.
Proof. Let ∆ ′ be the polygon corresponding to H and let m > 0 be such that m∆ ′ is integral.
Denote by QL and QR the left and right vertices of m∆ ′ (which are necessarily distinct).
Consider the divisor in T defined by the vanishing of
χQL − χQR .
Let D be its closure in X∆ and let D = π∗ D − mE in X. Then D has class m(H − E).
Let us compute the intersection number D · C. The two curves intersect only in the
torus T . We may multiply the equation χQL − χQR with χ−QL to put it in the form 1 − xi yj .
Here i/m is the width of the polygon ∆ ′ . Now the intersection
V(1 − xi yj ) ∩ V(1 − y)
has i points with multiplicity 1. This implies that
D · C = D · C + mE · E = i − m,
which is zero if and only if i = m.
Let now D be a general nef Q-divisor on X in the class H − E, where H is defined by a
polygon ∆ ′ of width 1. Since D is a general divisor on the 2-dimensional face of Nef(X),
we may assume that ∆ ′ is a 4-gon. We wish to show that D is not semiample. More
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
11
precisely, we show that for any m sufficiently large and divisible, all global sections of
OX (mD) vanish at the T -fixed point corresponding to the left vertex PL .
Let m > 0 be an integer such that m∆ ′ is integral. Let QL , QR be the left and right
vertices of ∆ ′ . Global sections of OX (mD) have the form
X
f=
aq χq
aq ∈ k, f vanishes to order at least m at t0.
q∈m∆ ′
Such a global section f vanishes at the T -fixed point corresponding to PL if and only if
amQL = 0. The condition that f vanishes to order at least m at t0 can be expressed by
saying that all partial derivatives of f up to order m−1 vanish at the point t0 = (1, 1). Now
the vanishing of the coefficient amQL is equivalent to the existence of a partial derivative
D of order at most m − 1 such that for q ∈ m∆ ′
0
if q 6= mQL ,
D(χq )|t0 =
c 6= 0 if q = mQL .
As in [6], it is enough to find such a derivative D after an integral translation of m∆ ′
(which corresponds to multiplication of f with a monomial). We translate m∆ ′ so that
its right vertex mQR has coordinates (m − 2, 0). Then its left vertex mQL has coordinates
(−2, β) for some β ∈ Z. We choose D of the form
D = ∂m−n−1
D̃,
x
vanishes when applied to monomials
where D̃ has order at most n. Note that ∂m−n−1
x
q
i j
m−n−1
χ = x y , 0 ≤ i < m − n − 1. After applying ∂x
to the monomials χq , q ∈ m∆ ′ , the
results with nonzero coefficients can be divided into three sets:
S1 = {x−A−1 yβ },
S2 = {x−A yB+j }j=0,...,n−1 ,
S3 = {xi yj }i,j≥0,i+j<n .
Here β is as above, A = m − n and B ∈ Z. We used here conditions 0 ≤ s2 < 1 and
(2a) of Theorem 2.1 to describe the set S3. It is shown in Lemma 4.1 below that up to a
nonzero constant factor there is a unique partial derivative D̃ of degree n that vanishes
on monomials in S2 and S3 when evaluated at t0 . When applied to the monomial in S1 , its
value at t0 is
nB
(β − B − 1)(β − B − 2) · · · (β − B − n + 1)(β − B −
).
A
We need to check when this expression is nonzero. The condition β 6= B+j, j = 1, . . . , n−1
is precisely condition (2b) in Theorem 2.1. (Notice that condition (2) of Theorem 2.1 only
depends on the configuration of lattice points near the vertices mPl , mPR . The condition
does not change if we replace m∆ with m∆ ′ or its translation.) We claim that the condition
6= 0 can always be satisfied by choosing the divisor D general. Indeed, first
β − B − nB
A
12
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
notice that replacing m by any of its positive multiples preserves the hypothesis of the
theorem. We can vary D in the 2-dimensional face of the nef cone by moving the left
vertex of ∆ ′ up or down. For m fixed, this deformation changes both β and B by the same
amount and leaves A fixed. We can then choose m sufficiently divisible and a new D in
6= 0. This
the 2-dimensional face of the nef cone such that m∆ ′ is integral and β − B − nB
A
finishes the proof of the first half of Theorem 2.1.
Consider now the second half of Theorem 2.1, where w = 1 or ∆ is a triangle. If w = 1,
then the curve C as above has C2 = 0. This implies that C lies on the boundary of the
cone NE(X), but may not define an extremal ray. If C spans an extremal ray of NE(X) we
obtain the desired conclusion proceeding as before. Hence we assume that C⊥ ∩ Nef(X)
is a 1-dimensional face of the nef cone. Since C itself is nef, this 1-dimensional face must
be generated by C, hence D = C. This means that in the proof above we need to use
6= 0. This
∆ ′ = ∆ and we can not deform it. That gives us the extra condition β − B − nB
A
condition with A = m − n, β = myL − myR and B = b − myR is precisely condition (3) in
Theorem 2.1.
In the case of a triangle, X has Picard number two. For any w ≤ 1, C spans an extremal
ray of NE(X) and D = w1 π∗ C − E spans an extremal ray of Nef(X). Thus, we use ∆ ′ = w1 ∆
.
and Condition (3) of the theorem again gives non-vanishing of β − B − nB
A
4. N ON - VANISHING
DERIVATIVES .
In this section we prove the claim about the existence of the derivative D̃ made in the
last section and then generalize this result to dimension 3.
Lemma 4.1. Let A, B, β, n ∈ Z, A > 0, n > 0. Consider three sets of monomials
S1 = {x−A−1 yβ },
S2 = {x−A yB+j }j=0,...,n−1 ,
S3 = {xi yj }i,j≥0,i+j<n .
There exists a nonzero partial derivative D̃ of degree n such that D̃ applied to monomials in S2
and S3 vanishes at t0 = (1, 1). This derivative is unique up to a constant factor. The derivative D̃
applied to the monomial in S1 and evaluated at t0 is
(β − B − 1)(β − B − 2) · · · (β − B − n + 1)(β − B −
nB
).
A
Proof. It was noted by Castravet [2] that the existence of such a partial derivative D̃ is
equivalent to the existence of a plane curve of degree n that passes through the lattice
points (a, b) for xa yb ∈ S2 ∪ S3 . Indeed, we may replace partial derivatives ∂x , ∂y with
logarithmic partial derivatives x∂x , y∂y . Now if p(X, Y) is a polynomial, then
p(x∂x , y∂y )(xa yb )|t0 = p(a, b).
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
13
Instead of the derivative D̃ we will construct such a polynomial p(X, Y).
We use the notation
[X]i = X(X − 1) · · · (X − i + 1).
The general degree n polynomial that vanishes at (a, b) for all xa yb ∈ S3 has the form
p(X, Y) =
n
X
ci [X]n−i [Y]i
i=0
for ci ∈ k. We need that p(a, b) also vanishes when xa yb ∈ S2 . This means that, up to a
constant factor
p(−A, Y) = [Y − B]n .
Note that [−A]n−i [Y]i for i = 0, . . . , n form a basis for the space of all polynomials in Y of
degree at most n. It follows that we can solve for ci uniquely from this equation. However,
we can find p(−A − 1, Y) without solving for ci .
Let us evaluate p(X, Y) at X = −A − 1.
X
A+n−i
A+n
1 X
ci [−A]n−i
p(−A − 1, Y) =
ici [−A]n−i [Y]i .
[Y]i =
p(−A, Y) −
A
A
A
i
i
Similarly we find
p(−A, Y − 1) =
X
i
ci [−A]n−i [Y]i
Y−i
1X
ici [−A]n−i [Y]i .
= p(−A, Y) −
Y
Y i
We can eliminate the sums in the two expressions to get
Ap(−A − 1, Y) = (A + n − Y)p(−A, Y) + Yp(−A, Y − 1)
= (A + n − Y)[Y − B]n + Y[Y − B − 1]n
= [Y − B − 1]n−1 (YA − AB − nB).
Dividing both sides by A and substituting Y = β gives the result.
Let us now generalize the previous lemma to dimension 3. Consider three sets of lattice
points
T1 = {(−A − 1, β, γ)},
T2 = {(−A, B + i, C + j)}i,j≥0,i+j<n ,
T3 = {(l, i, j)}l,i,j≥0,l+i+j<n ,
for some A, B, C, β, γ, n ∈ Z, A > 0, n > 0. We want to find a degree n polynomial
p(X, Y, Z) that vanishes on T2 and T3 , but not on T1 .
The general polynomial that vanishes on T3 has the form
X
cij [X]n−i−j [Y]i [Z]j .
(4.1)
p(X, Y, Z) =
i,j≥0;i+j≤n
14
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
As before we find
p(−A−1, Y, Z) =
A+n
1 X
1 X
icij [−A]n−i−j [Y]i [Z]j −
jcij [−A]n−i−j [Y]i [Z]j ,
p(−A, Y, Z)−
A
A i,j
A i,j
p(−A, Y − 1, Z) = p(−A, Y, Z) −
1X
icij[−A]n−i−j [Y]i [Z]j,
Y i,j
p(−A, Y, Z − 1) = p(−A, Y, Z) −
1X
jcij [−A]n−i−j [Y]i [Z]j .
Z i,j
Eliminating the sums from the three equations we get
Ap(−A − 1, Y, Z) = (A + n − Y − Z)p(−A, Y, Z) + Yp(−A, Y − 1, Z) + Zp(−A, Y, Z − 1).
The polynomial p(X, Y, Z) must vanish at points (−A, Y, Z) ∈ T2 . There is an (n + 1)dimensional space of degree n polynomials in Y, Z that vanish at these points. A basis for
this space is given by [Y − B]d [Z − C]n−d , d = 0, . . . , n. Let p = pd be a polynomial as in
(4.1) with the coefficients cij chosen such that
pd (−A, Y, Z) = [Y − B]d [Z − C]n−d .
When d = n, we get the polynomial from the 2-dimensional case pn (−A, Y, Z) = [Y − B]n ,
which at X = −A − 1 is
pn (−A − 1, Y, Z) = [Y − B − 1]n−1 (Y − B −
nB
).
A
Similarly, the polynomial p0 satisfies
p0 (−A − 1, Y, Z) = [Z − C − 1]n−1 (Z − C −
nC
).
A
For 0 < d < n we can express
Apd (−A − 1, Y, Z) = (A + n − Y − Z)pd (−A, Y, Z) + Ypd (−A, Y − 1, Z) + Zpd (−A, Y, Z − 1)
= (A + n − Y − Z)[Y − B]d [Z − C]n−d + Y[Y − B − 1]d [Z − C]n−d + Z[Y − B]d [Z − C − 1]n−d
= [Y − B − 1]d−1 [Z − C − 1]n−d−1 (A + n − Y − Z)(Y − B)(Z − C)
+Y(Y − B − d)(Z − C) + Z(Y − B)(Z − C − (n − d)) .
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
15
Let us change variables to Ȳ = Y − B, Z̄ = Z − C. The polynomials pd (−A − 1, Y, Z) can
then be simplified to
nC
),
A
nB
pn (−A − 1, Y, Z) = [Ȳ − 1]n−1 (Ȳ −
),
A
(n − d)C
dB
Z̄ −
Ȳ
pd (−A − 1, Y, Z) = [Ȳ − 1]d−1 [Z̄ − 1]n−d−1 Ȳ Z̄ −
A
A
d
nB
n−d
nC
= [Ȳ − 1]d−1 [Z̄ − 1]n−d−1
Z̄(Ȳ −
)+
Ȳ(Z̄ −
) , 0 < d < n.
n
A
n
A
p0 (−A − 1, Y, Z) = [Z̄ − 1]n−1 (Z̄ −
Let β̄ = β − B, γ̄ = γ − C, where (−A − 1, β, γ) is the point in T1 . We need to determine
when pd (−A − 1, β, γ) does not vanish for some d.
Lemma 4.2. There exists 0 ≤ d ≤ n such that pd (−A − 1, β, γ) does not vanish if and only if
the following conditions hold:
(1) (β̄, γ̄) 6= (i, j) for any i, j ≥ 1, i + j < n.
(2) (β̄, γ̄) 6= ( nB
, nC
).
A
A
(3) (a) If β̄ = 0 and 0 < γ̄ < n, then B 6= 0.
(b) If γ̄ = 0 and 0 < β̄ < n, then C 6= 0.
(c) If β̄ + γ̄ = n and 0 < β̄, γ̄ < n, then B + C 6= A.
Proof. Let us call [Ȳ − 1]d−1 [Z̄ − 1]n−d−1 the first part of pd and the remainder the last part.
Similarly for p0 and pn .
Consider cases:
• 0 < β̄, γ̄, β̄ + γ̄ < n. Then the first part of every pd vanishes at (β̄, γ̄).
, nC
). Then the last part of every pd vanishes at (β̄, γ̄).
• (β̄, γ̄) = ( nB
A
A
• β̄ = 0 and 0 < γ̄ < n. If every pd vanishes at (β̄, γ̄), then in particular pn vanishes, which implies B = 0. Conversely, if B = 0, then every pd vanishes. Similar
argument applies to the case γ̄ = 0 and 0 < β̄ < n.
• β̄ + γ̄ = n and 0 < β̄, γ̄ < n. Let β̄ = d > 0, γ̄ = n − d > 0. Then pd is the only
polynomial whose first part does not vanish at (β̄, γ̄). The last part of pd vanishes
if and only if B + C = A.
• All other (β̄, γ̄). There exist two different d such that the first part of pd does not
, nC
).
vanish at (β̄, γ̄). If both last parts vanish at (β̄, γ̄) then (β̄, γ̄) = ( nB
A
A
5. P ROOFS
We start with the proof of Theorem 2.3.
IN DIMENSION
3.
16
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
Let ∆ be the polytope in Theorem 2.3. The variety X∆ is not Q-factorial and has Picard
number 1. (To see the Picard number, consider deformations of the polytope by moving
facets in the normal direction. We may keep one vertex, say the origin, fixed and move
the remaining two facets. There is a one parameter family of such deformations, given
by moving the vertex (0, 1, 0) along the y-axis.) Let H be the class of the Q-Cartier divisor
corresponding to the polytope ∆. Then H generates Pic(X∆ )R . The space Pic(X)R = N1 (X)
is generated by (the pullback of) H and the class E of the exceptional divisor.
We construct a curve C ⊂ X that is analogous to a curve of negative self-intersection on
a surface. The polytope ∆ contains 3 lattice points (0, 0, 0), (0, 1, 0) and (0, 0, 1). Consider
two surfaces in the torus T defined by the vanishing of
χ(0,0,0) − χ(0,1,0) = 1 − y,
χ(0,0,0) − χ(0,0,1) = 1 − z,
and let S̄1 , S̄2 be their closures in X∆ . Then S̄1 and S̄2 are both Q-Cartier divisors in the
class H. Let C̄ be their intersection.
Lemma 5.1. C̄ is an irreducible curve.
Proof. We consider the intersection of C̄ with T -orbits of X∆ . For any T -orbit of dimension
1 or 2, the restriction of at least one Si to the orbit is defined by a monomial equation,
hence that Si does not intersect the T -orbit. This implies that C̄ does not contain any
component in X∆ rT and hence is irreducible.
Let S1, S2 , C be the strict transforms of S̄1 , S̄2 , C̄ in X. Then S1 and S2 both have class
H − E and C = S1 ∩ S2 .
Lemma 5.2. The class of C generates an extremal ray in NE(X). The dual face of Nef(X) is
generated by the class w1 H − E.
Proof. We can compute the intersection number
S̄3i = H3 = 6Vol(∆) = w.
Hence S3i = w − 1 ≤ 0. Now
Si · C = S3i ≤ 0.
Any other irreducible curve C ′ in X does not lie on either S1 or S2 , hence Si · C ′ ≥ 0.
It follows that the class of C lies on the boundary of NE(X), and since this cone is 2dimensional, C generates an extremal ray.
The class
1
H
w
− E is orthogonal to C:
1
1
1
H − E) · C = ( H − E)(H − E)(H − E) = H3 − 1 = 0,
w
w
w
hence it generates a boundary ray of Nef(X).
(
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
17
It now remains to show that a divisor in the class w1 H − E is not semiample. Let m be
as in the theorem, with m∆ integral, and let M = mw ∈ Z. Notice that any positive
integer multiple of m also satisfies the hypotheses of the theorem. Consider the divisor
class M( w1 H − E) = mH − ME. We show that any
X
f(x, y, z) =
cq χq
q∈m∆∩Z3
that vanishes to order at least M at t0 = (1, 1, 1) must have cmPL = 0. This implies that the
T -fixed point corresponding to PL is a base point for M( w1 H − E). This argument run with
m replaced by any of its positive integer multiples, allows us to deduce that w1 H − E is not
semiample.
As in the 2-dimensional case, we need to produce a partial derivative D of order M − 1
such that, when applied to any monomial χq for q ∈ m∆ ∩ Z3 , it vanishes at t0 if and
only if q 6= mPL . To find such D, we first translate m∆ so that mPR becomes equal to
(M − 2, 0, 0). Then mPL moves to (−2, β, γ), where β = myL − myR , γ = mzL − mzR . We
look for D of the form
D = ∂xM−n−1 D̃,
where D̃ has order n. When applying ∂xM−n−1 to monomials χq for q ∈ m∆ ∩ Z3 , the
resulting nonzero terms ap χp correspond to lattice points p that can be divided into three
sets:
T1 = {(−A − 1, β, γ)},
T2 = {(−A, B + i, C + j)}i,j≥0,i+j<n ,
T3 = {(l, i, j)}l,i,j≥0,l+i+j<n ,
where A = M−n, B = b−myR , C = c−mzR . Here we used the assumptions 0 ≤
and (2a) of Theorem 2.3 to describe the set T3 .
yR zR
,
xR xR
<1
Finding a derivative D̃ as above is equivalent to finding a degree n polynomial p(X, Y, Z)
that vanishes on T2 and T3 , but not on T1 . The necessary and sufficient conditions for the
existence of such polynomial are given in Lemma 4.2. We need to check that the assumptions of Theorem 2.3 imply the assumptions of the lemma.
In the notation of Lemma 4.2,
β̄ = β − B = myL − b,
γ̄ = γ − C = mzL − c.
Now condition (1) in the lemma is the same as (2b) in the theorem. For the remaining
is equivalent to myL = b − nsy and
conditions one can compute that the equality β̄ = nB
A
nC
γ̄ = A is equivalent to mzL = c − nsz . This implies that (3a), (3b) in the theorem are the
same conditions as (2), (3) in the lemma, finishing the proof of Theorem 2.3.
18
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
Corollary 2.4 follows directly from Theorem 2.3. Corollary 2.5 is also obtained from this
theorem as follows. Given a tetrahedron as in Corollary 2.5, we apply a shear transformation to arrange 0 ≤ yxRR , xzRR < 1. For a tetrahedron these inequalities imply that myL ≤ b
and mzL ≤ c. Hence conditions (2b) and (3b) of Theorem 2.3 hold trivially. The other conditions of the theorem follow from the three conditions of Corollary 2.5. In condition (2)
of Corollary 2.5 we only required that the slice mxR − n + 1 has size n instead of requiring
that slices mxR , mxR − 1, . . . , mxR − n + 1 have size 1, 2, . . . , n. The stronger condition can
fail if the slice mxR − 1 has size 1 instead of the required 2. However, then by reflecting
the tetrahedron across the yz-plane, we are in the case n = 1, which automatically gives
a non-MDS. Corollary 2.6 is a direct translation of Corollary 2.5 in terms of (xL , xR , y0, z0 ).
Let us now prove Theorem 2.11. The proof is similar to the proof in dimension 2 [6].
Let P(a, b, c1, c2 ) = Proj k[x, y, z1, z2 ], where the variables x, y, z1, z2 have degree a, b, c1,
c2 , respectively. The relation (e, f, g1, g2 ) gives three monomials xe yf , zg1 1 and zg2 2 of degree
d = ae + bf = ci gi .
Consider the degree map deg : R4 → R that maps (u, v, w1, w2) 7→ au + bv + c1 w1 +
∼ R3 with lattice
c2 w2 . The tetrahedron ∆ is then deg−1 (d) ∩ R4≥0 in the space deg−1 (d) =
∼ Z3 . We identify points in ∆ with points in deg−1 (d) as follows:
deg−1 (d) ∩ Z4 =
(0, 0, 0)7→ (e, f, 0, 0),
(0, 1, 0)7→(0, 0, g1, 0),
(0, 0, 1)7→(0, 0, 0, g2),
d
PR 7→ ( , 0, 0, 0),
a
d
PL 7→ (0, , 0, 0).
b
The gcd conditions on the relation (e, f, g1, g2) imply that this identification is compatible
with the isomorphism of lattices.
A homogeneous polynomial of degree d defines a divisor D on P(a, b, c1, c2) with selfintersection number
d3
D3 =
.
abc1c2
This identifies condition (1) in Theorem 2.11 and Corollary 2.5.
To count lattice points in slices of m∆, consider the linear function h(u, v, w1, w2 ) defined by dot product with
d
(f, −e, 0, 0).
c1 c2
We claim that the function h takes value c on slice c. This can be proved by checking that
h vanishes on slice 0 and when evaluated at the vertices PL and PR , it gives the correct
width w.
Consider now lattice points Q in slice mxL + 1 in m∆. We replace these lattice points Q
with Q − mPL . The new points are of the form (u, v, w1, w2 ) ∈ Z4 , u, w1, w2 ≥ 0, v ≤ 0,
EXAMPLES OF NON-FINITELY GENERATED COX RINGS.
19
satisfying the equations
h(u, v, w1, w2 ) = 1
deg(u, v, w1, w2) = 0
⇔
There is a rational point
d
(fu − ev) = 1
c1 c2
au + bv + c1 w1 + c2 w2 = 0.
⇔
1
(b, −a, 0, 0)
g1 g2
satisfying these equations. To get integral points we subtract from this a rational linear
combination of (e, f, −g1, 0) and (e, f, 0, −g2):
(u, v, w1, w2 ) =
1
δ1
δ2
(b, −a, 0, 0) + (e, f, −g1, 0) + (e, f, 0, −g2),
g1 g2
g1
g2
δ1 , δ2 ≤ 0.
Replacing v with −v, the slice mxL + 1 in m∆ can be identified with pairs of integers
δ1 , δ2 ≤ 0 such that
δ1
δ2
1
(b, a) + ( + )(e, −f)
(u, −v) =
g1 g2
g1 g2
has non-negative integer components.
By a similar argument, the slice mxR − n + 1 in m∆ can be identified with pairs of
integers γ1 , γ2 ≥ 0 such that
n−1
γ1 γ2
(b, a) + ( + )(e, −f)
g1 g2
g1 g2
has non-negative integer components.
Finally, the condition n(sy , sz ) ∈ Z2 in Corollary 2.5 is equivalent to the slice mxR − n
in m∆ having a lattice point on the edge joining mPL and mPR . Similarly to the slice
mxR − n + 1 this happens if and only if
n
(b, a)
g1 g2
has integer components.
R EFERENCES
[1] Florian Berchtold and Jürgen Hausen. Cox rings and combinatorics. Transactions of the American Mathematical Society, 359(3):1205–1252, 2007.
[2] Ana-Maria Castravet. Mori dream spaces and blow-ups. arXiv: 1701.04738.
[3] Ana-Maria Castravet and Jenia Tevelev. M0,n is not a Mori dream space. Duke Mathematical Journal,
164(8):1641–1667, 2015.
[4] Steven Dale Cutkosky. Symbolic algebras of monomial primes. J. Reine Angew. Math., 416:71–89, 1991.
[5] William Fulton. Introduction to toric varieties, volume 131 of Annals of Mathematics Studies. Princeton
University Press, Princeton, NJ, 1993. The William H. Roever Lectures in Geometry.
[6] José Luis González and Kalle Karu. Some non-finitely generated Cox rings. Compos. Math., 152(5):984–
996, 2016.
20
JOSÉ LUIS GONZÁLEZ AND KALLE KARU
[7] Shiro Goto, Koji Nishida, and Keiichi Watanabe. Non-Cohen-Macaulay symbolic blow-ups for space
monomial curves and counterexamples to Cowsik’s question. Proc. Amer. Math. Soc., 120(2):383–392,
1994.
[8] Zhuang He. New examples and non-examples of Mori dream spaces when blowing up toric surfaces.
arXiv: 1703.00819.
[9] Yi Hu and Sean Keel. Mori dream spaces and GIT. Michigan Math. J., 48:331–348, 2000. Dedicated to
William Fulton on the occasion of his 60th birthday.
[10] Shinnosuke Okawa. On images of Mori dream spaces. Math. Ann., 364(3-4):1315–1342, 2016.
J.L. G ONZ ÁLEZ , D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF C ALIFORNIA , R IVERSIDE , R IVER SIDE , CA 92521, U NITED S TATES .
K. K ARU , D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF B RITISH C OLUMBIA , VANCOUVER , BC
V6T1Z2, C ANADA .
E-mail address: [email protected], [email protected]
| 0math.AC
|
Deep Learning for Automatic Stereotypical Motor Movement Detection using
Wearable Sensors in Autism Spectrum Disorders
Nastaran Mohammadian Rad 1,2,3,∗, Seyed Mostafa Kia4,5 , Calogero Zarbo1 , Twan van Laarhoven2 , Giuseppe
Jurman1 , Paola Venuti6 , Elena Marchiori2 , Cesare Furlanello1
Fondazione Bruno Kessler, Via Sommarive 18, 38123, Povo, Trento, Italy
arXiv:1709.05956v1 [cs.CV] 14 Sep 2017
Abstract
Autism Spectrum Disorders are associated with atypical movements, of which stereotypical motor movements (SMMs)
interfere with learning and social interaction. The automatic SMM detection using inertial measurement units (IMU)
remains complex due to the strong intra and inter-subject variability, especially when handcrafted features are extracted
from the signal. We propose a new application of the deep learning to facilitate automatic SMM detection using multiaxis IMUs. We use a convolutional neural network (CNN) to learn a discriminative feature space from raw data. We
show how the CNN can be used for parameter transfer learning to enhance the detection rate on longitudinal data. We
also combine the long short-term memory (LSTM) with CNN to model the temporal patterns in a sequence of multi-axis
signals. Further, we employ ensemble learning to combine multiple LSTM learners into a more robust SMM detector.
Our results show that: 1) feature learning outperforms handcrafted features; 2) parameter transfer learning is beneficial
in longitudinal settings; 3) using LSTM to learn the temporal dynamic of signals enhances the detection rate especially
for skewed training data; 4) an ensemble of LSTMs provides more accurate and stable detectors. These findings provide
a significant step toward accurate SMM detection in real-time scenarios.
Keywords: Convolutional Neural Networks, Long Short-Term Memory, Transfer Learning, Ensemble Learning,
Wearable Sensors, Autism Spectrum Disorders
1. Introduction
Autism spectrum Disorder (ASD) is a complex and heterogeneous neuro-developmental disorder. Specific symptoms in ASD are difficulties in social interactions, repetitive or restricted behaviors, and verbal/nonverbal communication difficulties. Prevalence of ASD is reported to
be 1 in 88 individuals [1]. While the majority of studies
have mainly focused on social and communication problems of ASD children, the repetitive and restricted behaviors associated with ASD individuals are also objects
of interest because of their effect on the learning performance and socialization; and also as an indicator of dis∗ Corresponding
author
Email address: [email protected] (Nastaran Mohammadian
Rad )
1 Fondazione Bruno Kessler, Trento, Italy
2 Institute for Computing and Information Sciences, Radboud
University, Nijmegen, The Netherlands
3 Department of Information Engineering and Computer Science,
University of Trento, Trento, Italy
4 Donders Centre for Cognitive Neuroimaging, Donders Institute
for Brain, Cognition and Behaviour, Radboud University, Nijmegen,
The Netherlands
5 Department of Cognitive Neuroscience, Radboud University
Medical Centre, Nijmegen, The Netherlands
6 Department of Psychology and Cognitive Science, University of
Trento, Trento, Italy
Preprint submitted to arxiv
tress [2, 3, 4]. Stereotypical Motor Movements (SMMs) are
the major group of atypical repetitive behaviors in children
with ASD. SMMs occur without evoking stimuli and include hand flapping, body rocking, and mouthing [5, 6, 7].
SMMs have a specific negative effect on the quality of life
of ASD children: they can affect negatively on the performance of children during learning new skills and while
using the learned skills [8]. Furthermore, since these type
of movements are socially abnormal, they cause difficulties in interaction with pairs in the school or other social
settings [9]. In some cases, the severity of SMMs can lead
to a meltdown event and even can cause self-damaging
behaviors [10].
There are three traditional approaches for measuring the
SMMs [11]: 1) paper-and-pencil rating, 2) direct behavioral observation, and 3) video-based methods. Paper-andpencil rating is an interview-based approach which suffers
from the subjectivity in rating. Furthermore, it cannot
accurately detect the intensity, amount, and duration of
SMMs [12]. In the direct behavioral observation approach,
therapists can directly observe and record the sequences of
SMMs. This method is not also a reliable approach due
to the several reasons [11, 13]: first, in high speed movements, it is hard for therapists to accurately observe and
document all SMM sequences instantaneously. Second, determining the start and end time of the SMM sequences
September 19, 2017
is difficult. Third, it is impossible for therapists to concurrently record all environmental conditions and SMMs.
Video-based approaches are based on video capturing, offline coding, and analysis of SMMs. Since multiple coding
sessions of the captured videos are feasible, this observational method is much more accurate than two previous
approaches, but it is time-consuming and unpractical as a
clinical tool[14].
Considering the high prevalence rate of autism in children [15] and the limitations of existing methods for measuring SMMs, it is essential to develop time efficient and
accurate methods for automatic SMM detection. Developing a real-time SMM detection and quantification system
would be advantageous for ASD researchers, caregivers,
families, and therapists. Such a system would provide a
powerful tool to evaluate the adaptation of subjects with
ASD to diverse life contexts within an ecologic approach.
In particular, it helps to mitigate the meltdown behaviors that are anticipated by the increase in atypical behaviors. Any automatic quantification of atypical movements
would indeed help caregivers and teachers to defuse the
mechanism leading to stereotyped behaviors by involving
children in specific activities or social interactions. Such
involvement decreases the frequency of SMMs and gradually alleviates their duration and severity [16, 17]. A
real-time implementation of SMM detection system (see
Figure 1) would help therapists to evaluate the efficacy of
behavioral interventions.
Inertial Measurement Units (IMUs) provide effective
tools for measuring the frequency, intensity, and duration of physical activities over a time period via embedded accelerometer, gyroscope, and magnetometer sensors.
Due to the small size and possibility of embedding in the
mobile phones, IMUs have been adopted as common and
useful sensors for wearable devices to measure the physical activities in either constrained and free-living environments [18, 19, 20]. In recent years, human activity
recognition using IMU sensors has been widely studied.
Most of these studies tried to extract time and frequency
domain features such as mean, standard deviation, skewness, and FFT peaks from raw signals to feed them to
a classifier for activity identification [21, 22]. According
to the achieved results in human activity recognition systems, applying pattern recognition on the collected data
by IMU sensors can reliably and accurately detect physical
activities which are an evidence for the possibility of applying such techniques to automatically detect SMMs in ASD
children [23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36].
Despite meaningful amount of research in this direction,
few challenges for automatic SMM detection using wearable sensors still remain unsolved especially in real-time
applications.
One of the important challenges for accurate SMM detection is to extract a set of effective and robust features
from the IMU signal. As in many other signal processing applications, SMM detection is commonly based on
extracting handcrafted features from the IMU signals. So
far, a wide variety of feature extraction methods have been
used in the literature. Generally two main types of features
are extracted from the accelerometer signal [37]: i) time
domain features, ii) frequency domain features. For time
domain features, some statistical features such as mean,
standard deviation, zero-crossing, energy, and correlation
are extracted from the overlapping windows of the signal.
In the case of frequency features the discrete Fourier transform is used to estimate the power of different frequency
bands. In addition, the Stockwell transform [38] is proposed by [33] for feature extraction from 3-axis accelerometer signals in order to provide better time-frequency resolution for non-stationary signals. In spite of their popularity, manual feature extraction and selection suffer from
subjectivity and time inefficiency [39] that restrict the performance and also the application of SMM detection systems in real-time scenarios.
Another challenge toward developing a real-time SMM
detection system is personalization due to the intra and
inter-subject variability [33, 35]. This challenge, despite
its crucial importance, has been undervalued [33]. Intrasubject variability is mainly due to the high variability
in the intensity, duration, frequency, and topography of
SMMs in each individual with ASD. Inter-subject differences are defined by the same variability across different
individuals. Existence of these two types of variability
within and across ASD persons motivates the necessity
of developing an adaptive SMM detection algorithm that
is capable to adjust to new patterns of behaviors. Feature learning and transfer learning [40] can be considered
as candidate solutions to attack these challenges. To this
end, here we present an extended version of our previous
efforts in [41, 42] with four main contributions: 1) robust
feature learning from multi-sensor IMU signals; 2) enhancing the adaptability of SMM detection system to new data
via parameter transfer learning; 3) improving the detection
rate by incorporating the temporal dynamics of signals in
the feature learning process; and 4) using principles of the
ensemble learning to enhance the detection rate.
To achieve our first goal, we propose a new application of
the deep learning paradigm in order to directly learn discriminating features for detecting SMM patterns. In particular, we use a convolutional neural network (CNN) [43]
to bypass the commonly used feature extraction procedure.
The idea of the CNN is inspired by the visual sensory system of living creatures [44]. Following this idea, LeCun
et al. [45] developed a deep CNN architecture to address
a pattern recognition problem in computer vision. Having fewer connections and parameters due to the weight
sharing property, CNNs are easier to train compared to
other deep neural networks. Currently, CNN solutions are
among the best-performing systems on pattern recognition
systems specifically for the handwritten character [45] and
object recognition [46]. Beyond audio and image recognition systems, CNNs are successfully applied on various
types of signals. Mirowski et al. [47] applied CNN on EEG
signals for seizure detection. In the domain of psychophys2
Data collection using wearable sensors
Data Analysis
Monitoring Children
Figure 1: A real-time automatic SMM detection system. Inertial Measurement Units (IMUs) can be used for data collection. The collected
data can be analyzed locally or remotely to detect SMMs. In case of detecting abnormal movements, an alert is sent to a therapist, caregiver,
or parents.
order to learn the dynamic feature space on the sequence
of IMU data. We further show that combining multiple
LSTM models using an ensemble learning technique can
improve the stability of results. To the best of our knowledge, it is the first time that a recurrent architecture is
used for SMM detection using wearable technologies.
The rest of this paper is organized as follows: in Section 2 we first briefly review the formal definitions and
concepts about CNN, LSTM, parameter transfer learning, and the ensemble of the best base learners approach.
Then using the presented definitions we introduce the proposed CNN and LSTM architectures 1 for SMM detection
on IMU signals. The experimental materials and procedures are explained in this section. Section 3 compares our
experimental results versus the state of the art solutions
in SMM detection. Our results on a simulated dataset
and two real datasets show that feature learning via the
CNN outperforms handcrafted features in SMM classification. Furthermore, it is shown that the parameter transfer learning is beneficial in enhancing the SMM detection
rate when moving to a new dataset. Finally our results
illustrate that including the dynamics of recorded data in
feature learning process improves the classification performance in SMM detection especially when an unbalanced
training set is used in the training phase. In Section 4
we discuss how the proposed deep architecture facilitates
developing real-time SMM detection systems. We further
discuss the main limitation of our method and state the
possible future directions. Section 5 concludes this paper
by summarizing our achievements.
iology, for the first time Martinez et al. [39] proposed a
model based on CNN to predict affective states of fun, excitement, anxiety, and relaxation. Their proposed model
was tested on skin conductance and blood volume pulse
signals. Recent studies show the advantageous of applying
CNN on accelerometer signals for human activity recognition [48, 49, 50].
To fulfill our second goal, we employ the parameter
transfer learning by pre-initializing the parameters of the
CNN [51]. We hypothesize that this capability can be used
to transfer the prior knowledge regarding the distribution
of parameters from one dataset to another dataset that are
collected in a longitudinal study. If successful, our method
can be employed in order to enhance the adaptability of
SMM detection system to new unseen data, thus facilitates
its applications in wild real-world scenarios.
By the definition, SMMs are repetitive behaviors, thus
temporal patterns stored in the sequence of samples are expected to contain valuable information. Our third contribution relies on the fact that the proposed CNN architecture does not fully exploit the temporally structured information in the IMU signal sequences. This is a general issue
in SMM detection, in which the segments of IMU signals
are treated as statistically independent samples. Therefore
the possible long-term dependencies stored in the longer
temporal intervals of the signal are ignored in the detection
process. Long short-term memory (LSTM) [52] as a type
of recurrent neural networks (RNN) has been effectively
used for learning long-term temporal dependencies in sequential data such as speech recognition [53], handwriting
recognition [54], and natural language modeling [55]. Recently, the LSTM has also been successfully used for human activity recognition using wearable technologies as a
classic sequence analysis problem [56, 57, 50]. Considering
these studies in human activity recognition, it is expected
that learning the temporal patterns stored in the consecutive samples of IMU data to provide higher accurate SMM
detectors. Thus, here we propose a deep architecture,
stacking an LSTM layer on top of the CNN architecture, in
2. Methods
2.1. Notation
Let S1 , S2 , . . . , Sc ∈ RL be c time-series of length L
that are recorded by a set of inertial measurement units
1 Here the architecture implies the customization of the structural
parameters of the CNN such as the number of layers, the number
and size of filters, etc.
3
(IMUs) (e.g., accelerometer, gyroscope, and magnetometer sensors) at the sampling frequency of ν Hz. Thus,
T = L/ν represents the length of the signal in seconds.
We refer to each Si as a data channel. Now consider
Xt ∈ Rc×ν for t ∈ {1, 2, . . . , T } as a sample in the raw
feature space that is constructed by concatenating timeseries of c data channels in a given time t. Let yt ∈ {0, 1}
be the label associated to Xt where yt = 1 corresponds
to an SMM sample. In this text, we use boldface capital
letters to represent matrices, boldface lowercase letters to
represent vectors, and italic lowercase letters to represent
scalars. We represent the matrix and element-wise multiplication between A and B matrices by A · B and A B,
respectively. Further, [a, b] represents vector concatenation operation between a and b vectors.
over a pooling window with size of p elements and calν
culates the reduced feature map M0t ∈ Rf × u . In fact, a
pooling layer reduces the resolution of a feature map by
factor of u1 where u is the stride (or step) size. Max-pooling
and average-pooling are two commonly used pooling functions which compute the maximum or average value among
the values in a pooling window, respectively. In averagepooling for m0i,j ∈ M0t , i ∈ {1, . . . , f }, and j ∈ {1, . . . , uν },
we have:
p
m0i,j =
1X
mi,(j−1)×u+k .
p
(3)
k=1
Alternatively, in the max-pooling each element of the
reduced feature map is the maximum value in a corresponding pooling window:
2.2. Feature Learning via Convolutional Neural Network
m0i,j = max(mi,(j−1)×u+1 , mi,(j−1)×u+2 , . . . , mi,(j−1)×u+p ).
The goal of an SMM detector is to predict the probability of being an SMM for a given sample Xt , i.e.,
P (yt = 1 | Xt ). While the raw feature space (Xt ) is sensitive to intra and inter-subject differences, feature learning
can provide the possibility to learn a new feature space
that is robust over time and across different subjects. The
aim of feature learning is to learn a linear or non-linear
mapping function F : Xt 7→ x0t , where x0t ∈ Rd is called
the learned feature space. Then a classifier can be used in
the learned feature space to estimate P (yt = 1 | x0t ).
Convolutional neural networks (CNNs) offer an effective
infrastructure for feature learning. CNN benefits from invariant local receptive fields, shared weights, and spatiotemporal sub-sampling features to provide robustness over
shifts and distortions in the input space [43]. A classic
CNN has a hierarchical architecture that alternates convolutional and pooling layers in order to summarize large
input spaces with spatio-temporal relations into a lower dimensional feature space. A 1D-convolutional layer receives
the input signal Xt ∈ Rc×ν , convolves it with a set of f
filters with the length of m, W ∈ Rf ×c×m , and produces
a feature map Mt ∈ Rf ×ν :
The reduced feature map M0t can be used as the input
to the next convolutional layer, i.e., Xt of the next layer.
In general, the reduced feature map computed by stacking
several convolution, ReLU, and pooling layers is flattened
as a vector before the classification step. The flattening
step is performed by collapsing the rows of M0t in the form
of a vector. The resulting vector is called the learned feature space x0t that represents a new representation of the
original feature space. This new representation is typically fed to a fully connected neural network followed by
a softmax layer for the classification purposes.
In this paper, and for the purpose of SMM detection on
the multi-sensor IMU data, we propose to use a three-layer
CNN to transform the time-series of multiple sensors to a
new feature space. The proposed architecture is shown in
Figure 2. Three convolutional layers are designed to have
4, 4, and 8 filters with the length of 9 samples (i.e., 0.1
seconds), respectively. The length of the pooling window
and pooling stride are fixed to 3 (p = 3) and 2 (u = 2),
respectively. The pooling stride of 2 reduces the length
of feature maps by the factor of 0.5 after each pooling
layer. The output of the third convolutional layer after
flattening provides the learned feature vector. Then, the
learned feature vector is fed to a two-layer fully-connected
network with 8 and 2 neurons that are connected to a
softmax layer. A dropout [59] rate of 0.5 is used in the fully
connected layers to avoid the overfitting problem. Since
only the information in Xt is used to compute x0t and then
predict yt , we refer to the learned feature space via this
CNN architecture as the static feature space.
Mt = Xt ∗ W =
Pc Pm
w
× xj,1+i
Pj=1 Pi=1 1,j,i
cj=1 m
w
×
xj,1+i
2,j,i
i=1
..
.
Pc Pm
j=1
i=1 wf,j,i × xj,1+i
...
...
..
.
...
Pc Pm
i=1 w1,j,i × xj,ν+i
Pcj=1 Pm
j=1
i=1 w2,j,i × xj,ν+i
..
Pc Pm .
i=1 wf,j,i × xj,ν+i
j=1
(4)
(1)
where ∗ represents the convolution operator. The feature
map is then fed to an activation function, generally the
rectified linear unit (ReLU), to add non-linearity to the
network and also to avoid the gradient vanishing problem [58], where:
(2)
2.3. Parameter Transfer Learning via Network Preinitialization
Here max(., .) represents the element-wise max operation.
Finally, in order to reduce the sensitivity of the output
to shifts and distortions, M+
t is passed through a pooling
layer which performs a local averaging or sub-sampling
The quality and characteristics of recorded IMU signals
varies not only from subject to subject but also from time
to time in a single subject. Therefore it is important that
the SMM detector system be able to adapt to new streams
f ×ν
M+
, Mt ).
t = max(0
4
(c)
X
Convolution+ReLU
c=4
W
AVG-Pooling
M
x'
M'
(e)
f=4
p=3
m=9
Softmax
f=4
c=4
f=4
1
1
1
X
CNN Layer 1
Convolution+ReLU
c
S1
W
M
CNN Layer 2
AVG-Pooling
CNN Layer 3
M'
X
c=4
f=4
f=4
W
M
AVG-Pooling
f=8
1
(b)
x'
1
1
Sc
Flattening
f=8
m=9
1
M'
Labels
f=8
c=4
p=3
m=9
c
f=4
Convolution+ReLU
Classifier
p=3
Input
Data
Learned Features
(a)
1
1
1
1
1
(d)
1
Figure 2: (a) The proposed architecture for SMM detection in the static feature space using a three-layer CNN. (b) The first CNN layer.
This layer receives the one-second long time-series of several IMU sensors at time t, i.e., Xt , and transfer it to the first level reduced feature
map M0t . (c) The second CNN layer that uses the first level reduced feature map M0t as its input, and transfer it to the second-level reduced
feature map. (d) The third CNN layer. The reduced feature map of this layer is reshaped to the learned feature vector x0t using the flattening
operation. (e) The learned feature vector is fed to a fully-connected followed by a softmax layer to classify the samples to SMM and no-SMM
classes.
of signals in longitudinal scenarios. In this paper, we explore the possibility of parameter transfer learning via network pre-initialization in order to transfer the learned patterns to the newly seen data in a different time span. In
this direction we first formalize the background theoretical
concepts.
SMM and no-SMM samples. Assume DS , DT , TS , and
TT to represent the source domain, target domain, source
task, and target task, respectively. Transfer learning aims
to benefit from the knowledge in the source domain and
task in order to improve the predictive power in the target
domain when DS 6= DT or TS 6= TT [40].
In the statistical learning theory, the goal is to learn a
task T in a certain domain D. A domain D = {X , ρX } is
defined as a possible conjunction between an input space
X and a marginal probability distribution ρX . For example in the SMM detection context, the recorded IMU
signal for different subjects can be considered as different domains as the marginal probability distribution ρX is
different from one subject to another. Similarly different
domains can be defined by time in longitudinal data collection scenarios. Given a domain D, a task T = {Y, Φ}
is defined as a predictive function Φ from D to the output space Y. For example in this study Φ is the SMM
detector, and Y represents the categorical output space of
In this study, we are interested in the application of
parameter transfer learning via pre-initializing the parameters of a deep neural network, as a well-established technique in the deep learning community, to improve the
SMM prediction performance across different subjects and
time intervals. To this end, we define the source domain
DS as the IMU signal which is collected on several subjects at the time span T1 . Similarly the target domain
DT is defined as the IMU signal which is collected on several subjects at the time span T2 . Assume ΦS be the
learned predictive function, i.e., the CNN classifier, in the
source domain. We use the learned parameters in ΦS to
pre-initialize the parameters of the predictive function in
5
the target domain ΦT . In simpler words, instead of random pre-initialization, we initialize the parameters of CNN
classifier in the target domain with the learned CNN parameters in the source domain. We hypothesize that such
a knowledge transfer via learned parameters improves the
prediction performance in the longitudinal studies where
the data are collected at different time intervals.
Classification
Long Short-Term Memory
2.4. SMM Detection in Dynamic Feature Space using
LSTM
In SMM detection using static feature space (see Section 2.2) only the data in Xt is used to predict yt . Thus
it is implicitly assumed that the sequence of samples over
time are independent and identically distributed (i.i.d).
But in reality, this assumption is not valid as the samples
in consecutive time steps are highly dependent. Therefore,
it is expected that accounting for this dependency would
improve the performance of the SMM detector. Following this hypothesis, we propose to use a long short-term
memory (LSTM) layer to model the temporal dependency
between the consecutive time steps of the recorded signal.
Let x0t be a set of static features that are extracted or
learned from samples in the raw feature space (i.e., from
Xt ). Here we assume the CNN architecture explained in
Section 2.2 is used to compute x0t . Then, let ct ∈ Rq and
ht ∈ Rq to represent the cell state and output of an LSTM
unit at time step t, respectively, where q is the number of
neurons in the LSTM unit. We will refer to ht as the
dynamic feature space. The LSTM unit receives x0t , ht−1 ,
and ct−1 as its inputs, and computes ct and ht as follows:
ct = ft
ht = ot
ct−1 + it
c̃t ,
(1 − e−2×ct )
Feature
Extraction
Time
Figure 3: The proposed architecture for SMM detection in the dynamic feature space using long short-term memory. Each feature extraction block contains a trained three-layer CNN architecture (see
Figure 2).
In this paper, a fully-connected layer with dropout of
0.2 is used to transfer the output of the LSTM layer at
time t, i.e., ht , to the input of the softmax layer zt =
(0) (1)
[zt , zt ]T ∈ R2 :
(1)
P (yt = 1 | xt−τ , xt−τ +1 , . . . , xt ) =
(5)
(1 + e−2×ct )−1 .
Here ft ∈ Rq is called the forget gate vector and its
elements are real numbers between 0 and 1 that decide how
much information to be passed from ct−1 to ct . During
the learning phase, the forget gate learns the forget weight
matrix Wf and the forget bias vector bf . ft is computed
by
0
0
0
(8)
where it ∈ Rq is the input gate vector with elements between 0 and 1. These values determine the level of new
information in c̃t to be transferred to the cell state ct . it
is computed based on Wi and bi as follows:
0
it = (1 + e−(Wi ·[yt−1 ,xt ]+bi ) )−1 .
(9)
Finally, ot ∈ Rq is the output gate vector that filters the
cell state ct to generate the output of the LSTM unit ht :
0
ot = (1 + e−(Wo ·[yt−1 ,xt ]+bo ) )−1 .
(1)
, (11)
Due to the random initialization and using stochastic
optimization algorithms on random mini-batches in training deep learning models, retraining the same model on the
same training set results in heterogeneous approximations
of the target classifier. This heterogeneity is the direct result of reaching different local optimums in optimizing a
complex non-convex error surface. One possible approach
to overcome this problem is ensemble learning (EL) [60].
The main idea behind EL is to combine the knowledge
learned by individual classifiers in order to achieve a more
superior and stable performance. It is shown that in general an ensemble of classifiers works better than every
single classifier due to the statistical, computational, and
representational reasons [61]. Considering the success of
deep learning ensembles in pattern recognition and signal processing applications [62, 63, 64, 65], in this study
we are interested in applying classifier selection voting approach [66] to combine an ensemble of the best base learners.
(7)
(1 + e−2×(Wc ·[yt−1 ,xt ]+bc ) )−1 ,
(0)
ezt + ezt
2.5. Ensemble of the Best Base Learners
Using a tangent hyperbolic function, c̃t ∈ Rq provides
new candidate values between −1 and 1 for ct by learning
Wc and bc :
c̃t = (1 − e−2×(Wc ·[yt−1 ,xt ]+bc ) )
ezt
where τ represents the number of previous time steps that
are used as the input to the LSTM layer. Figure 3 presents
a schematic overview of the proposed architecture.
(6)
ft = (1 + e−(Wf ·[ht−1 ,xt ]+bf ) )−1 .
Feature
Extraction
Feature
Extraction
(10)
6
Algorithm 1 The training and test procedures in the
majority voting on a set of b best models.
coding is undertaken during sessions to annotate the starting and ending time of movements. The captured data
were band-pass filtered with a cut-off frequency of 0.1Hz
to remove the DC components. Then the signal was segmented to 1 second long (i.e., 100 time-points) using a
sliding window. The sliding window was moved along the
time dimension with 10 time-steps resulting in 0.9 overlaps
between consecutive windows 2 .
1: procedure training(C,Xtr ,ytr )
2:
for all ci ∈ C do
3:
Predict ŷ using ci on Xtr .
4:
Evaluate ŷ and store the performance in αi .
5:
6:
7:
8:
9:
end
for i ← 1, l do
Store the best i classifiers in Ci∗ .
Predict ŷ1 , . . . , ŷi using classifiers in Ci∗ on Xtr .
Compute majority voting ỹ on predictions in ŷ1 , . . . , ŷi .
Evaluate ỹ and store the performance in αi .
end
Find the best value for b by b = argmaxi (αi ).
return Cb∗ .
2.6.2. Real Data
We use the data presented in [33] wherein the accelerometer data were collected for 6 subjects with autism in a
longitudinal study3 . The data were collected in the laboratory and classroom environments while the subjects wore
three 3-axis wireless accelerometers and engaged in body
rocking, hand flapping, or simultaneous body rocking and
hand flapping. The accelerometer sensors were attached
to the left and right wrists, and on the torso. Offline annotation based on a recorded video is used to annotate the
data by an expert. Two separate collections are available:
the first collection, here we call it Real Data1, was recorded
by MITes sensors at 60Hz sampling frequency [31]. The
second collection Real Data2, was recorded three years after the first recording on the same subjects using Wockets
sensors with the sampling frequency of 90Hz. The sampling rate of two recordings is equalized by re-sampling the
signal in Real Data1 to 90Hz using linear interpolation.
The cut-off high pass filter at 0.1Hz is applied in order to
remove the DC components of the signal. Similar to [33],
the signal is segmented to 1-second long overlapped intervals using a sliding window. The amount of overlap is set
to 10 time-points resulting in 0.87 overlap between consecutive windows. Table 1 summarizes the number of samples
in no-SMM and SMM classes for each subject. The difference in the number of samples in SMM and no-SMM
classes shows unbalanced nature of the real data, where
in Real Data1 and Real Data2 datasets 31% and 23% of
samples are in the SMM class, respectively.
10:
11:
12:
13: procedure test(Cb∗ ,Xts , yts )
14:
Predict ŷ1 , . . . , ŷb using classifiers in Cb∗ on Xts .
15:
Compute majority voting ỹ on predictions in ŷ1 , . . . , ŷb .
16:
Evaluate ỹ and store the performance in α.
17:
return α.
Let (Xtr , ytr ) and (Xts , yts ) to be the corresponding
sample/target pairs in the training and test sets, respectively. Then assume C = {c1 , c2 , . . . cl } be a set of l base
learners trained on the training set. Our goal is to first
find a set of b best classifiers C ∗ ⊆ C based on a performance measure α on the training set, and then to combine
their prediction on the test set using majority voting in the
prediction phase. Algorithm 1 summarizes this approach.
2.6. Experimental Materials
We assess the performance of the proposed methods on
both simulated and real data. In the following, we describe
the datasets and the procedures that are used for data
preparation.
2.6.1. Simulated Data
In a simulation setting, 5 healthy subjects (3 females and
2 males) are asked to emulate stereotypical movements in a
controlled environment. Each participant wore an EXLs3
sensor1 , a miniaturized electronic device with the function
of real-time IMU, fixed on the right wrist using a wristband (see Figure 4(a)). EXLs3 sensor records three-axis
accelerometer, gyroscope, and magnetometer data (it has
9 data channels in total). The sensor was set to transmit three-axis ±16g acceleration and ±2000dps angular
velocity at the 100Hz sampling rate. The participants
were instructed to perform their normal working activities
such as sitting, writing, and typing; while intermittently
performing hand flapping upon receiving a start/stop cue
from the instructor (see Figure 4(b)-(e)). The total period
of SMMs is organized somehow to keep the distribution
of two classes comparable with real datasets where 27%
of samples are in the SMM class (see Table 1 and Section 2.6.2). The total duration of each experiment was 30
minutes organized in three 10 minutes sessions. Real-time
2.7. Experimental Setups and Evaluation
To investigate the effect of static and dynamic feature
learning and parameter transfer learning on the performance of SMM detection, we conducted four experiments.
Keras library [67] is used in our implementations4 .
2.7.1. Experiment 1: Static Feature Learning
The main aim of this experiment is to compare the effectiveness of feature learning using a deep neural network
versus raw feature space and handcrafted features in an
2 The collected simulated data is made publicly available at https:
//gitlab.fbk.eu/MPBA/smm-detection.
3 The dataset and full description of data are publicly
available
at
https://bitbucket.org/mhealthresearchgroup/
stereotypypublicdataset-sourcecodes/downloads.
4 See https://gitlab.fbk.eu/MPBA/smm-detection to access the
implemented scripts and codes.
1 For the technical description see: http://www.exelmicroel.
com/elettronica_medicale-tecnologia-indossabile-exl-s3_
module.html.
7
(c)
(b)
(a)
(d)
(e)
Figure 4: (a) The configuration of the EXLs3 sensor on the right hand. (b),(c) The simulated data are collected during daily work activities
(e.g., writing, typing, etc.). (d),(e) The subjects are asked to intermittently perform hand flapping upon receiving a start/stop cue from the
instructor.
Table 1: Number of samples in SMM and no-SMM classes in three
datasets.
Data
Simulated Data
Real Data1
Real Data2
Subjects
Sub1
Sub2
Sub3
Sub4
Sub5
Total
Sub1
Sub2
Sub3
Sub4
Sub5
Sub6
Total
Sub1
Sub2
Sub3
Sub4
Sub5
Sub6
Total
No-SMM
13875
11686
13694
12428
13583
65266
21292
12763
31780
10571
17782
12207
106395
18729
22611
40557
38796
22896
2375
145964
SMM
4075
6224
4246
5532
4367
24444
5663
4372
2855
10243
6173
17725
47031
11656
4804
268
8176
6728
11178
42810
All
17950
17910
17940
17960
17950
89710
26955
17135
34635
20814
23955
29932
153426
30385
27415
40825
46972
29624
13553
188774
distribution. The stochastic gradient descent with momentum (the momentum is fixed to 0.9) is used for training the
network. All these steps are performed only on the training data to ensure unbiased error estimation. Due to the
random initialization of weights and employing stochastic
gradient descent algorithm for optimization, results can be
different from one training run to another. Therefore, we
repeated the whole procedure of learning and classification
10 times and the mean and standard variation over runs
are reported. It is important to emphasize that, similar
to [33], in all three parts of this experiment the number
of samples in minority class is used to randomly draw a
balanced training set.
SMM/All
0.23
0.35
0.24
0.31
0.24
0.27
0.21
0.26
0.08
0.49
0.26
0.59
0.31
0.38
0.18
0.01
0.17
0.23
0.82
0.23
2.7.2. Experiment 2: Parameter Transfer Learning
As discussed before, deep neural networks provide the
capability of parameter transfer learning via network preinitialization. We applied this experiment only on two real
datasets in order to investigate the possibility of transferring learned knowledge from one dataset to another in a
longitudinal data collection setting. This experiment is
similar to Experiment 1, except for the network initialization step. Instead of random initialization, here we firstly
train the CNN on one balanced real dataset, e.g., Real
Data1, and then we use the learned parameters for preinitializing the parameters of CNN before training on another balanced real dataset, e.g., Real Data2. Similar to
previous experiment, we repeated the whole experiment
10 times to evaluate the standard deviation of the classification performance.
across-subject SMM detection setting. To evaluate the effect of both feature extraction and feature learning on the
SMM classification performance, first, without any feature
extraction the signals in raw feature space are used as the
input to a support vector machine (SVM) classifier. In this
case, all data channels of each sample Xt are collapsed into
a feature vector (with length of 900 = 9 × 100 in simulated
data and 810 = 9 × 90 in real data case). Second, to
evaluate the detection performance using handcrafted features we extracted all features mentioned in [33] including
time, frequency, and Stockwell transform features, then,
we replicated the across-subject SMM detection experiment in [33]. In this setting we used exactly the same
implementation provided by the authors 1 in the feature
extraction and classification steps. Third, a CNN architecture (see Section 2.2) is used to learn a middle representation of the multi-sensor signal. In this experiment, all effective parameters of CNN (weights and biases) are initialized by drawing small random numbers from the normal
2.7.3. Experiment 3: Training on the Unbalanced Training
Set
As explained, in Experiment 1 and 2 we balanced the
training set based on the number of samples in minority
class. Even though balancing the training set improves
the quality of the trained model but in fact it suffers from
some deficits: 1) by balancing the training set we impose
a wrong prior assumption on the original distribution of
data. As shown in Table 1 in real datasets around 0.3
of samples belong to SMM class, when by balancing the
1 The
code is available at:
https://bitbucket.org/
mhealthresearchgroup/stereotypypublicdataset-sourcecodes/
downloads.
8
Figure 5: The distribution of SMM and no-SMM samples in the 2-dimensional t-SNE space for (a) raw feature space, (b) handcrafted features,
(c) static feature space learned by CNN, (d) static feature space learned by pre-initialized CNN, and (e) dynamic feature space learned by
LSTM. Feature learning increases the separability of samples in two classes compared to raw and handcrafted features.
For all datasets, we used the same configurations for the
LSTM models by fixing τ = 25 and q = 40. We set l =
10 and used the F1-score for the performance metric α
in Algorithm 1. The experiment is repeated 10 times to
evaluate the standard deviation over the mean prediction
performance.
dataset we assume it is 0.5; 2) by balancing the training set
we cannot employ the full richness of the data as we need
to remove significant amount of samples from the training set; 3) in some practical scenarios, such as real-time
adaptation or classification on the sequence of streamed
data, balancing the training set is impractical. Considering these limitations, in this experiment in order to evaluate the effect of balancing on the performance of CNN
model we evaluate the performance of the proposed CNN
architecture in predicting SMMs when unbalanced training sets are used in the training phase.
2.7.6. Evaluation
In all experiments the leave-one-subject-out scheme is
used for model evaluation in order to measure the robustness of the trained model against inter-subject variability.
Due to the unbalanced class distributions in the test set,
we computed the F1-score to evaluate the classification
performance:
2.7.4. Experiment 4: Dynamic Feature Learning
In this experiment, we are interested in answering three
main questions: 1) what are the advantages of learning
temporal representation of IMU signals for reliable SMM
detection? 2) how long is the most informative time interval in IMU signals for detecting abnormal movements? 3)
what is the optimal configuration for the LSTM unit? To
answer these questions, we applied the proposed LSTM architecture in Section 2.4 on the three benchmark datasets
with different values for τ and q, i.e., time steps and neuron
number, respectively. We set τ = {1, 3, 5, 10, 15, 25, 50}
and q = {5, 10, 20, 30, 40, 50}. The LSTM unit is trained
on the extracted features by the CNN using the RMSProp [68] optimizer. The learned dynamic features via
LSTM (ht ) are classified to target classes using a softmax
classifier. It is worthwhile to emphasize that, in this setting, since the order of samples in the training set matters,
balancing the training set is impossible, thus we use the
original unbalanced data.
F1 = 2 ×
P recision × Recall
P recision + Recall
,
where true positive (TP), false positive (FP), and false
negative (FN) rates are used as follows to compute the
precision and recall:
TP
TP + FP
TP
Recall =
.
TP + FN
P recision =
3. Results
3.1. Feature Learning Outperforms Handcrafted Features
The classification performances summarized in Table 2
compare the quality of feature learning via CNN with raw
and handcrafted feature spaces on three datasets. In all
three datasets, the classification performance of SMM detection on the handcrafted and learned features is higher
than the classification performance on the raw feature
space. This observation demonstrates the importance of
2.7.5. Experiment 5: Ensemble of LSTMs
To explore the possible advantage of combining multiple
classifiers, we used the procedure explained in Section 2.5
in order to combine a set of b best base learners. In this
experiment, we used the LSTM models in the Experiment
4 as base learners for the classification of unbalanced data.
9
Table 2: Results for SMM detection using raw, handcrafted, static, dynamic feature spaces, and ensemble learning in three benchmarked
datasets. The results show that feature learning generally outperforms raw and handcrafted feature spaces. In addition, the parameter
transfer learning has a positive effect on the performance of the CNN classifier. Furthermore, training the CNN classifier on unbalanced
training sets causes the performance drop in feature learning and transfer learning scenarios. Using the LSTM network to extract dynamic
features from the signal alleviates this problem to some degrees. Ensemble of LSTMs shows more stable performance compared to single
LSTM classifiers.
Balanced Training Sets
Real Data2
Real Data1
Simulated
Data
Sub
Raw
Features
Handcrafted
Features
Feature
Learning
Transfer
Learning
1
2
3
4
5
Mean
1
2
3
4
5
6
Mean
1
2
3
4
5
6
Mean
0.29
0.84
0.55
0.76
0.38
0.56
0.44
0.32
0.22
0.44
0.56
0.56
0.42
0.47
0.23
0.01
0.32
0.38
0.50
0.32
0.71
0.86
0.76
0.48
0.77
0.72
0.74
0.37
0.50
0.73
0.44
0.46
0.54
0.43
0.26
0.03
0.86
0.73
0.07
0.40
0.78 ± 0.05
0.86 ± 0.03
0.80 ± 0.01
0.85 ± 0.03
0.79 ± 0.01
0.82 ± 0.03
0.74 ± 0.02
0.75 ± 0.02
0.68 ± 0.04
0.92 ± 0.01
0.51 ± 0.04
0.90 ± 0.01
0.74 ± 0.03
0.61 ± 0.11
0.20 ± 0.04
0.02 ± 0.01
0.72 ± 0.03
0.21 ± 0.09
0.36 ± 0.13
0.35 ± 0.08
0.71 ± 0.02
0.73 ± 0.01
0.70 ± 0.03
0.92 ± 0.00
0.68 ± 0.05
0.94 ± 0.01
0.78 ± 0.03
0.68 ± 0.05
0.22 ± 0.04
0.02 ± 0.01
0.77 ± 0.02
0.75 ± 0.09
0.91 ± 0.05
0.56 ± 0.05
feature extraction/learning for detecting SMMs. Furthermore, the comparison between the results achieved by
handcrafted and learned features illustrates the efficacy
of feature learning over the manual feature extraction in
SMM prediction. The learned feature space reaches on average 0.10 and 0.20 higher F1-score than the handcrafted
features in case of simulated data and real data1, respectively, while in case of real data2 its performance declines
by 0.05. These results support the overall efficacy of feature learning versus handcrafted features in extracting robust features for across-subject SMM detection. These
conclusions are even further confirmed in Figures 5(a)-(c),
where the t-distributed Stochastic Neighbor Embedding
(t-SNE) [69] technique is employed to visualize the different feature spaces in a 2-dimensional space. We used
the average of Fisher’s separability score [70] across two tSNE dimensions to quantify the separability of samples in
two classes for different feature spaces. Figure 5(a) shows
2D t-SNE distribution of SMM and no-SMM samples in
the raw feature space, where there is a high overlap between the samples of two classes. This high overlap is
also well-reflected in the low Fisher’s separability score in
raw feature space (0.02). Figure 5(b) depicts the distribution of samples of two classes in handcrafted feature space.
The samples in two classes are barely separable and the
Fisher’s separability score is 0.03. Figure 5(c) displays the
2D t-SNE space for the learned features via the CNN architecture. In this case the separability score is improved
significantly to 0.10.
Feature
Learning
(1 sec)
0.73 ± 0.13
0.78 ± 0.09
0.75 ± 0.13
0.73 ± 0.12
0.80 ± 0.04
0.76 ± 0.11
0.70 ± 0.02
0.63 ± 0.03
0.57 ± 0.08
0.88 ± 0.01
0.51 ± 0.08
0.79 ± 0.07
0.68 ± 0.06
0.33 ± 0.14
0.11 ± 0.03
0.02 ± 0.01
0.71 ± 0.14
0.14 ± 0.09
0.62 ± 0.08
0.32 ± 0.1
Unbalanced Training Sets
Feature
Dynamic
Transfer
Learning
Feature
Learning
(2.5 sec)
Learning
0.95 ± 0.01 0.95 ± 0.01
0.86 ± 0.13 0.95 ± 0.01
0.95 ± 0.03 0.97 ± 0.01
0.97 ± 0.01 0.96 ± 0.02
0.91 ± 0.01 0.90 ± 0.02
0.93 ± 0.06 0.95 ± 0.01
0.71 ± 0.03 0.73 ± 0.04 0.77 ± 0.03
0.63 ± 0.04 0.68 ± 0.04 0.71 ± 0.03
0.59 ± 0.06 0.56 ± 0.13 0.68 ± 0.05
0.88 ± 0.01 0.93 ± 0.00 0.91 ± 0.01
0.58 ± 0.07 0.51 ± 0.04 0.52 ± 0.04
0.81 ± 0.09 0.86 ± 0.12 0.90 ± 0.02
0.70 ± 0.06 0.71 ± 0.08 0.75 ± 0.03
0.36 ± 0.08 0.47 ± 0.15 0.53 ± 0.09
0.16 ± 0.04 0.13 ± 0.05 0.26 ± 0.06
0.02 ± 0.01 0.02 ± 0.01 0.02 ± 0.02
0.83 ± 0.03 0.90 ± 0.02 0.76 ± 0.09
0.09 ± 0.02 0.23 ± 0.17 0.35 ± 0.15
0.70 ± 0.16 0.68 ± 0.21 0.96 ± 0.01
0.36 ± 0.08 0.40 ± 0.13 0.48 ± 0.08
Ensemble
Learning
0.95 ± 0.00
0.96 ± 0.01
0.97 ± 0.00
0.97 ± 0.01
0.91 ± 0.00
0.95 ± 0.01
0.80 ± 0.00
0.74 ± 0.00
0.72 ± 0.01
0.93 ± 0.00
0.51 ± 0.01
0.91 ± 0.00
0.77 ± 0.01
0.59 ± 0.03
0.29 ± 0.02
0.02 ± 0.02
0.87 ± 0.02
0.43 ± 0.08
0.98 ± 0.00
0.53 ± 0.04
3.2. Parameter Transfer Learning is Beneficial in Longitudinal Studies
As mentioned in Section 2.7.2, the aim of our second
experiment was to investigate the possibility of transferring learned knowledge from one dataset to another using
parameter transfer learning. Our results in Table 2 shows
that transferring knowledge from one dataset to another
in a longitudinal study, by pre-initializing the parameters
of CNN model improves the average classification performance of the SMM detectors by 0.04 and 0.21 in Real
Data1 and Real Data2 datasets, respectively.
3.3. Training on Unbalanced Data Decreases the Performance
The results in Table 2 illustrate the negative effect of
using unbalanced training set in training CNN architecture in randomly initialized (feature learning) and preinitialized (transfer learning) scenarios. The performance
of SMM detection in feature learning scenario drops by
0.06 and 0.03 in Real Data1 and Real Data2 datasets,
respectively. This performance drop is even more pronounced in the transfer learning scenario where we observe 0.08 and 0.20 performance drop in the corresponding
datasets.
3.4. Dynamic Feature Learning Outperforms Static Feature Learning
Figure 6 compares the averaged SMM classification performance over subjects in the static feature space via the
10
Figure 6: Comparison between the classification performances of CNN and LSTM for different time-steps (τ ) and number of neurons q, on
three datasets and when an unbalanced training set is used for training the networks. The results show the superiority of the dynamic feature
space over the static feature space. While the number of neurons in the LSTM unit has little effect on the performance, using around 2.5
seconds long interval is the best choice for extracting effective dynamic features from the IMU signals.
improves the detection rates when unbalanced training
sets are used (see Table 2). This observation is consistent
across three datasets.
Figure 7 further explores the superiority of the dynamic
feature representation when the training set is unbalanced.
In the static feature space case, balancing the training
set and enforcing the wrong prior class distribution into
the classification task, despite higher recall rate, affects
negatively the precision of the classifier. In other words,
the classifiers have higher false alarm rate, which could
be problematic in real-world applications. This deficit is
recovered in the case of dynamic feature representation
where the classifier presents higher precision rate and comparable recall with respect to static features. In fact, the
LSTM-based architecture by enforcing the true prior distribution of data into the training process and, at the same
time using all the recorded samples, provides an SMM detection system with higher sensitivity and specificity.
CNN (the green dashed line for plain feature learning and
the blue dotted-dashed line for transfer learning) with the
dynamic feature space via the LSTM, the latter with different values for τ (x-axis) and q (line colors). Here in all
settings, an unbalanced training set is used in the training
phase. The results on three datasets illustrate that learning the temporal representation of signals with an LSTM
unit, consistently across datasets, improves the classification performance compared to the static feature learning
via the CNN. The classification performance improves by
increasing τ , and it reaches its highest performance around
τ = 25. Considering the consistency of the best τ value
for different subjects and different datasets, it can be concluded that using around 25 time-steps, i.e., around 2.5
seconds long interval, for extracting dynamic features is
the best choice for SMM detection purposes. On the other
hand, the results show the negligible effect of the number
of LSTM neurons (q) on the detection performance, thus,
a value around 10 can be considered a reasonable choice.
To further benchmark the advantage of dynamic feature
learning via LSTM, we used the CNN architecture for the
SMM detection on the best length for the time intervals,
i.e., on 2.5 seconds time intervals. The results on the three
datasets are summarized in Table 2 and Figure 6 (the dotted red line). The results confirm the superiority of dynamic feature learning compared to static feature learning
despite using longer time intervals for learning static features.
The effect of learning the temporal representation on
the separability of SMM and no-SMM samples is shown in
Figure 5(e). The higher Fisher’s separability score (0.17)
in the dynamic feature space compared to static feature
spaces can be considered as the basis for the higher classification performance of the proposed architecture, demonstrating the importance of learning dynamic features using
an LSTM based architecture. Furthermore, employing the
dynamic-feature representation computed by the LSTM,
3.5. Ensemble of LSTMs Stabilizes the Performance
The last column of Table 2 summarizes the results of
the ensemble approach. The results show slight boost in
the mean performance compared to single LSTM classifiers, especially on real data2 (see dashed black line in
Figure 6). Figure 7 shows that both precision and recall
contribute equally to this improvement in F1-scores. In
addition to the higher performance, the main advantage
of EL is demonstrated by the low variability of results.
This reduction in the variability is well-reflected in the reduced standard deviation around the mean performance
in real datasets (0.02 and 0.04 reduction in real data1 and
real data2 datasets, respectively). In other words an ensemble of LSTMs provides more reliable SMM detector in
comparison to every single LSTM classifier.
11
Figure 7: A comparison between precision and recall rates of classifiers trained on balanced/unbalanced training sets for static/dynamic
feature representations. Using dynamic feature space provides an SMM detection system with higher sensitivity and specificity.
4. Discussion
the SMM detector in the static feature space, exploiting
the temporal patterns of multi-sensor IMU signals recovers
its performance. This advantage facilitates training highperforming models by exploiting whole data sequences in
real-time SMM detection scenarios. Our effort, for the
first time in the SMM detection context, demonstrated the
superiority of recurrent structures in extracting discriminative temporal patterns from IMU signals. As the final
contribution, considering the important role of ensemble
learning for classifying the stream data in non-stationary
environments [66, 65], we employed ensemble of the best
base learners technique to improve the reliability of the
SMM detector. In summary, our results show that feature learning, transfer learning, ensemble learning, and
learning temporal structures from multi-modal IMU signals improve the performance of SMM detection systems
especially in more realistic scenarios.
4.1. Toward Real-Time Automatic SMM Detection
In this study, we proposed an original application of
deep learning for SMM detection in ASD children using
wearable technologies. To this end, we used a three-layer
CNN architecture for learning a more discriminative and
robust feature space in an across-subject SMM detection
scenario. Our experimental results, on the simulated and
real data, support the superiority of learning middle representation of IMU signal over traditional feature extraction methods in automatic SMM detection. Further, we
showed that the parameter transfer learning via network
pre-initialization provides the infrastructure for effective
knowledge transfer from one dataset to another in a longitudinal setting. We also presented an application of LSTM
in SMM detection context for extracting dynamic features
on the sequence of IMU data. In a comparison with the
static feature space learned via CNN architecture, we illustrated the advantage of employing the temporal information in improving the separability of SMM and no-SMM
samples. We experimentally showed that using around
2.5 seconds long interval for extracting dynamic features
is the best choice for SMM detection purposes. Further,
we showed using around 10 neurons in the LSTM unit
is a reasonable choice in order to extract the dynamics
of samples over time. As a side-advantage of learning a
dynamic feature space, we experimentally demonstrated
the higher performance of our method when, in a real
world setting, the distribution of samples in SMM and
no-SMM classes is highly skewed. We showed, while the
skewness of samples negatively affects the performance of
Developing real-time mobile applications for detecting
the abnormal movements such as SMMs can be considered as a final goal in the context of automatic SMM detection using wearable sensors. At the moment there are
numerous challenges in real-time human activity recognition using wearable sensors, namely [19, 71, 72]: 1) designing effective feature extraction and inference methods; 2)
recognition on realistic data; 3) the adaptability of system
to new users. Addressing these issues demands a huge investment in research toward finding robust and effective
features that can be extracted in a reasonable time from
the stream of IMU signal. Our proposal to learn a middle
representation of the signal, that is robust to the signal
variability of a single subject data over time and also to
the across-subject variability, can be considered as an effec12
tive solution in this direction. In addition, the parameter
transfer learning capability besides the possibility of incremental training of the proposed deep architecture facilitates the online adaptation of an automatic SMM detector
in real-time scenarios. This finding overlooks the subject
specific [73], and monolithic [74, 35] activity recognition
systems opening new frontiers toward adaptable activity
recognition systems which are more appropriate for realtime usages. At the end, the high detection performance
on unbalanced training sets achieved in the dynamic feature space facilitates the application of our method on the
realistic data when the incoming data samples are highly
skewed.
Acknowledgments
The authors would like to thank the members of MPBA
lab for their kind collaboration in collecting the simulated
data.
References
[1] J. Baio, Prevalence of autism spectrum disorders: Autism and
developmental disabilities monitoring network, 14 sites, united
states, 2008. morbidity and mortality weekly report. surveillance summaries. volume 61, number 3., Centers for Disease
Control and Prevention (2012).
[2] J. Hernandez, I. Riobo, A. Rozga, G. D. Abowd, R. W. Picard,
Using electrodermal activity to recognize ease of engagement in
children during social interactions, in: Proceedings of the 2014
ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2014, pp. 307–317.
[3] T. Chaspari, C.-C. Lee, S. Narayanan, Interplay between verbal
response latency and physiology of children with autism during
ECA interactions, Age (years) 8 (2012) 8.
[4] E. Hedman, L. Miller, S. Schoen, D. Nielsen, M. Goodwin,
R. Picard, Measuring autonomic arousal during therapy, in:
Proc. of Design and Emotion, 2012, pp. 11–14.
[5] G. Berkson, R. Davenport, Stereotyped movements of mental
defectives: I. initial survey., American Journal of Mental Deficiency (1962).
[6] M. H. Lewis, A. A. Baumeister, Stereotyped mannerisms in
mentally retarded persons: Animal models and theoretical analyses, International review of research in mental retardation 11
(1982) 123–61.
[7] R. M. Ridley, H. F. Baker, Stereotypy in monkeys and humans,
Psychological medicine 12 (1982) 61–72.
[8] J. W. Varni, O. I. Lovaas, R. L. Koegel, N. L. Everett, An
analysis of observational learning in autistic and normal children, Journal of Abnormal Child Psychology 7 (1979) 31–43.
[9] R. Jones, D. Wint, N. Ellis, The social effects of stereotyped
behaviour, Journal of Intellectual Disability Research 34 (1990)
261–268.
[10] C. H. Kennedy, Evolution of stereotypy into self-injury. (2002).
[11] R. L. Sprague, K. M. Newell, Stereotyped movements: Brain
and behavior relationships., American Psychological Association, 1996.
[12] D. A. Pyles, M. M. Riordan, J. S. Bailey, The stereotypy analysis: An instrument for examining environmental variables associated with differential rates of stereotypic behavior, Research
in Developmental Disabilities 18 (1997) 11–38.
[13] N. C. Gardenier, R. MacDonald, G. Green, Comparison of direct observational methods for measuring stereotypic behavior
in children with autism spectrum disorders, Research in Developmental Disabilities 25 (2004) 99–118.
[14] J. L. Matson, M. Nebel-Schwalm, Assessing challenging behaviors in children with autism spectrum disorders: A review,
Research in Developmental Disabilities 28 (2007) 567–579.
[15] S. Goldman, C. Wang, M. W. Salgado, P. E. Greene, M. Kim,
I. Rapin, Motor stereotypies in children with autism and other
developmental disorders, Developmental Medicine & Child Neurology 51 (2009) 30–38.
[16] S. Lee, S. L. Odom, R. Loftin, Social engagement with peers
and stereotypic behavior of children with autism, Journal of
Positive Behavior Interventions 9 (2007) 67–79.
[17] R. L. Loftin, S. L. Odom, J. F. Lantz, Social interaction and
repetitive motor behaviors, Journal of Autism and Developmental Disorders 38 (2008) 1124–1135.
[18] M. J. Mathie, A. C. Coster, N. H. Lovell, B. G. Celler, Accelerometry: providing an integrated, practical method for longterm, ambulatory monitoring of human movement, Physiological measurement 25 (2004) R1.
4.2. Limitation and Future Work
Even though the deep architecture introduced in this
study provides a significant step toward a more accurate
automatic SMM detection system in real-time scenarios,
but it suffers from a considerable limitation: the proposed
fully supervised scheme for training the SMM detection
model is problematic for its online adaptation. This problem comes from the fact that in real applications the system has no access to the labels of incoming samples during
usage by a new user. Therefore, the adaptation to new unseen data should be performed only based on the input unlabeled data. This limitation motivates future researches
on the online adaptation of the system in an unsupervised
manner. One possible solution in this direction is transductive transfer learning [40] where the basic assumption
is that no labeled data in the target domain are available.
Therefore adopting a transductive transfer learning strategy in the adaptation phase can be considered as a possible
future direction to extend this work.
5. Conclusions
In this study we addressed the problem of automatic
SMM detection in the framework of deep learning. We
used a 3-layer CNN architecture to learn a robust feature space from multi-sensor/modal IMU signals. We illustrated how the proposed architecture can be employed
for parameter transfer learning in order to enhance the
adaptability of SMM detection system to new data. Further, we showed incorporating the temporal dynamics of
the signal in the feature learning process, by combining
the CNN architecture with an LSTM unit, improves the
SMM detection rate in real-world scenarios especially in
case of unbalanced data. We further illustrated the advantage of ensemble learning to provide more stable and
reliable SMM detectors. Our results demonstrate high potentials of deep learning paradigm to address the crucial
challenges toward real-time SMM detection systems using
wearable technologies.
13
[19] O. D. Lara, M. A. Labrador, A survey on human activity recognition using wearable sensors., IEEE Communications Surveys
and Tutorials 15 (2013) 1192–1209.
[20] J. L. Helbostad, B. Vereijken, C. Becker, C. Todd, K. Taraldsen, M. Pijnappels, K. Aminian, S. Mellone, Mobile health
applications to promote active and healthy ageing, Sensors 17
(2017).
[21] J. Baek, G. Lee, W. Park, B.-J. Yun, Accelerometer signal
processing for user activity detection, in: Knowledge-Based Intelligent Information and Engineering Systems, Springer, 2004,
pp. 610–617.
[22] M. Gjoreski, H. Gjoreski, M. Lutrek, M. Gams, How accurately
can your wrist device recognize daily activities and detect falls?,
Sensors 16 (2016).
[23] T. Westeyn, K. Vadas, X. Bian, T. Starner, G. D. Abowd,
Recognizing mimicked autistic self-stimulatory behaviors using
HMMs, in: Ninth IEEE International Symposium on Wearable
Computers (ISWC’05), IEEE, 2005, pp. 164–167.
[24] C.-H. Min, A. H. Tewfik, Y. Kim, R. Menard, Optimal sensor location for body sensor network to detect self-stimulatory
behaviors of children with autism spectrum disorder, in: Engineering in Medicine and Biology Society, 2009. EMBC 2009.
Annual International Conference of the IEEE, IEEE, 2009, pp.
3489–3492.
[25] C.-H. Min, A. H. Tewfik, Automatic characterization and detection of behavioral patterns using linear predictive coding of
accelerometer sensor data, in: Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference
of the IEEE, IEEE, 2010, pp. 220–223.
[26] H. Min Cheol, A. H. Tewfik, Novel pattern detection in children
with autism spectrum disorder using iterative subspace identification, in: 2010 IEEE International Conference on Acoustics,
Speech and Signal Processing, IEEE, 2010, pp. 2266–2269.
[27] N. Gonçalves, J. L. Rodrigues, S. Costa, F. Soares, Automatic
detection of stereotyped hand flapping movements: two different approaches, in: RO-MAN, 2012 IEEE, IEEE, 2012, pp.
392–397.
[28] N. Goncalves, J. L. Rodrigues, S. Costa, F. Soares, Automatic
detection of stereotypical motor movements, Procedia Engineering 47 (2012) 590–593.
[29] J. L. Rodrigues, N. Gonçalves, S. Costa, F. Soares, Stereotyped movement recognition in children with ASD, Sensors and
Actuators A: Physical 202 (2013) 162–169.
[30] F. Albinali, M. S. Goodwin, S. Intille, Detecting stereotypical
motor movements in the classroom using accelerometry and pattern recognition algorithms, Pervasive and Mobile Computing
8 (2012) 103–114.
[31] F. Albinali, M. S. Goodwin, S. S. Intille, Recognizing stereotypical motor movements in the laboratory and classroom: a case
study with children on the autism spectrum, in: Proceedings
of the 11th international conference on Ubiquitous computing,
ACM, 2009, pp. 71–80.
[32] M. S. Goodwin, S. S. Intille, F. Albinali, W. F. Velicer, Automated detection of stereotypical motor movements, Journal of
autism and developmental disorders 41 (2011) 770–782.
[33] M. S. Goodwin, M. Haghighi, Q. Tang, M. Akcakaya, D. Erdogmus, S. Intille, Moving towards a real-time system for automatically recognizing stereotypical motor movements in individuals on the autism spectrum using wireless accelerometry,
in: Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing, ACM, 2014, pp.
861–872.
[34] T. Plötz, N. Y. Hammerla, A. Rozga, A. Reavis, N. Call, G. D.
Abowd, Automatic assessment of problem behavior in individuals with developmental disabilities, in: Proceedings of the 2012
ACM Conference on Ubiquitous Computing, ACM, 2012, pp.
391–400.
[35] U. Großekathöfer, N. V. Manyakov, V. Mihajlovi, G. Pandina,
A. Skalkin, S. Ness, A. Bangerter, M. S. Goodwin, Automated
detection of stereotypical motor movements in autism spectrum
disorder using recurrence quantification analysis, Frontiers in
Neuroinformatics 11 (2017) 9.
[36] A. M. Amiri, N. Peltier, C. Goldberg, Y. Sun, A. Nathan, S. V.
Hiremath, K. Mankodiya, Wearsense: Detecting autism stereotypic behaviors through smartwatches, Healthcare 5 (2017).
[37] N. F. Ince, C.-H. Min, A. Tewfik, D. Vanderpool, Detection
of early morning daily activities with static home and wearable wireless sensors, EURASIP Journal on Advances in Signal
Processing 2008 (2008) 31.
[38] R. G. Stockwell, L. Mansinha, R. Lowe, Localization of the
complex spectrum: the s transform, Signal Processing, IEEE
Transactions on 44 (1996) 998–1001.
[39] H. P. Martinez, Y. Bengio, G. N. Yannakakis, Learning deep
physiological models of affect, Computational Intelligence Magazine, IEEE 8 (2013) 20–33.
[40] S. J. Pan, Q. Yang, A survey on transfer learning, Knowledge
and Data Engineering, IEEE Transactions on 22 (2010) 1345–
1359.
[41] N. Mohammadian Rad, C. Furlanello, Applying deep learning to stereotypical motor movement detection in autism spectrum disorders, in: 2016 IEEE 16th International Conference
on Data Mining Workshops (ICDMW), 2016, pp. 1235–1242.
doi:10.1109/ICDMW.2016.0178.
[42] N. M. Rad, S. M. Kia, C. Zarbo, G. Jurman, P. Venuti,
C. Furlanello, Stereotypical motor movement detection in dynamic feature space, in: 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW), 2016, pp. 487–494.
doi:10.1109/ICDMW.2016.0076.
[43] Y. LeCun, Y. Bengio, Convolutional networks for images,
speech, and time series, The handbook of brain theory and
neural networks 3361 (1995).
[44] K. Fukushima, Neocognitron: A self-organizing neural network
model for a mechanism of pattern recognition unaffected by shift
in position, Biological cybernetics 36 (1980) 193–202.
[45] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based
learning applied to document recognition, Proceedings of the
IEEE 86 (1998) 2278–2324.
[46] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, in: Advances in
neural information processing systems, 2012, pp. 1097–1105.
[47] P. W. Mirowski, Y. LeCun, D. Madhavan, R. Kuzniecky, Comparing SVM and convolutional networks for epileptic seizure
prediction from intracranial EEG, in: Machine Learning for Signal Processing, 2008. MLSP 2008. IEEE Workshop on, IEEE,
2008, pp. 244–249.
[48] J. B. Yang, M. N. Nguyen, P. P. San, X. L. Li, S. Krishnaswamy,
Deep convolutional neural networks on multichannel time series
for human activity recognition, in: Proceedings of the 24th
International Conference on Artificial Intelligence, AAAI Press,
2015, pp. 3995–4001.
[49] M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu,
J. Zhang, Convolutional neural networks for human activity
recognition using mobile sensors, in: Mobile Computing, Applications and Services (MobiCASE), 2014 6th International Conference on, IEEE, 2014, pp. 197–205.
[50] T. Zebin, P. J. Scully, K. B. Ozanyan, Human activity recognition with inertial sensors using a deep learning approach,
in: 2016 IEEE SENSORS, 2016, pp. 1–3. doi:10.1109/ICSENS.
2016.7808590.
[51] F. J. O. Morales, D. Roggen, Deep convolutional feature transfer across mobile activity recognition domains, sensor modalities
and locations, in: Proceedings of the 2016 ACM International
Symposium on Wearable Computers, ACM, 2016, pp. 92–99.
[52] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (1997) 1735–1780.
[53] F. Eyben, M. Wöllmer, B. Schuller, A. Graves, From speech to
letters-using a novel neural network architecture for grapheme
based asr, in: Automatic Speech Recognition & Understanding,
2009. ASRU 2009. IEEE Workshop on, IEEE, 2009, pp. 376–
380.
[54] P. Doetsch, M. Kozielski, H. Ney, Fast and robust training of
recurrent neural networks for offline handwriting recognition,
14
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
in: Frontiers in Handwriting Recognition (ICFHR), 2014 14th
International Conference on, IEEE, 2014, pp. 279–284.
W. Zaremba, I. Sutskever, O. Vinyals, Recurrent neural network
regularization, arXiv preprint arXiv:1409.2329 (2014).
F. J. Ordóñez, D. Roggen, Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition, Sensors 16 (2016) 115.
N. Y. Hammerla, S. Halloran, T. Ploetz, Deep, convolutional,
and recurrent models for human activity recognition using wearables, arXiv preprint arXiv:1604.08880 (2016).
I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT
Press, 2016. http://www.deeplearningbook.org.
N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever,
R. Salakhutdinov, Dropout: a simple way to prevent neural
networks from overfitting., Journal of Machine Learning Research 15 (2014) 1929–1958.
T. G. Dietterich, Ensemble learning, in: M. Arbib (Ed.), The
Handbook of Brain Theory and Neural Networks, volume 2,
MIT Press: Cambridge, MA, 2002, pp. 110–125.
L. I. Kuncheva, Combining Pattern Classifiers:
Methods and Algorithms, 1 ed., Wiley-Interscience, 2004.
URL:
http://www.amazon.com/exec/obidos/redirect?tag=
citeulike07-20&path=ASIN/0471210781.
L. Deng, J. Platt, Ensemble deep learning for speech recognition
(2014).
L.-p. Jin, J. Dong, Ensemble deep learning for biomedical
time series classification, Computational intelligence and neuroscience 2016 (2016).
Y. Guan, T. Ploetz, Ensembles of deep lstm learners for activity
recognition using wearables, arXiv preprint arXiv:1703.09370
(2017).
B. Krawczyk, L. L. Minku, J. Gama, J. Stefanowski, M. Woniak, Ensemble learning for data stream analysis: A survey,
Information Fusion 37 (2017) 132 – 156.
H. M. Gomes, J. P. Barddal, F. Enembreck, A. Bifet, A survey on ensemble learning for data stream classification, ACM
Comput. Surv. 50 (2017) 23:1–23:36.
F. Chollet, Keras, https://github.com/fchollet/keras, 2015.
T. Tieleman, G. Hinton, Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude, COURSERA:
Neural Networks for Machine Learning 4 (2012).
L. v. d. Maaten, G. Hinton, Visualizing data using t-SNE,
Journal of Machine Learning Research 9 (2008) 2579–2605.
R. O. Duda, P. E. Hart, D. G. Stork, Pattern classification,
John Wiley & Sons, 2012.
M. A. Labrador, O. D. L. Yejas, Human Activity Recognition:
Using Wearable Sensors and Smartphones, CRC Press, 2013.
J. W. Lockhart, G. M. Weiss, Limitations with activity recognition methodology & data sets, in: Proceedings of the 2014
ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication, ACM, 2014, pp. 747–
756.
M. Berchtold, M. Budde, H. R. Schmidtke, M. Beigl, An extensible modular recognition concept that makes activity recognition practical, in: Annual Conference on Artificial Intelligence,
Springer, 2010, pp. 400–409.
O. D. Lara, A. J. Pérez, M. A. Labrador, J. D. Posada, Centinela: A human activity recognition system based on acceleration and vital sign data, Pervasive and mobile computing 8
(2012) 717–729.
15
| 1cs.CV
|
arXiv:1608.04120v2 [math.ST] 30 Aug 2016
Yule’s “Nonsense Correlation” Solved!
Philip Ernst ∗, Larry Shepp †, and Abraham Wyner
‡
August 31, 2016
Abstract
In this paper, we resolve a longstanding open statistical problem.
The problem is to mathematically prove Yule’s 1926 empirical finding
of “nonsense correlation” ([15]). We do so by analytically determining
the second moment of the empirical correlation coefficient
R1
R1
R1
W1 (t)W2 (t)dt − 0 W1 (t)dt 0 W2 (t)dt
θ := r
R
R
2 rR
2 ,
R1 2
1
1
1
2
W1 (t)dt − 0 W1 (t)dt
W2 (t)dt − 0 W2 (t)dt
0
0
0
of two independent Wiener processes, W1 , W2 . Using tools from Fredholm integral equation theory, we successfully calculate the second
moment of θ to obtain a value for the standard deviation of θ of nearly
.5. The “nonsense” correlation, which we call “volatile” correlation,
is volatile in the sense that its distribution is heavily dispersed and is
frequently large in absolute value. It is induced because each Wiener
process is “self-correlated” in time. This is because a Wiener process is
an integral of pure noise and thus its values at different time points are
correlated. In addition to providing an explicit formula for the second
moment of θ, we offer implicit formulas for higher moments of θ.
1
Introduction
A fundamental yet crucial question for practitioners of statistics is the following: given a sequence of pairs of random variables {Xk , Yk } (k = 1, 2, . . . , n),
how can we measure the strength of the dependence of X and Y ? The classical Pearson correlation coefficient offers a solution that is standard and often
∗
Rice University, Department of Statistics
Deceased April 23, 2013
‡
The Wharton School of the University of Pennsylvania, Department of Statistics
†
1
powerful. It is also widely used even in situations where little is known about
its empirical properties; for example, when the sequence of random variables
are not identically distributed or independent. The Pearson correlation is often calculated between two time series. Practitioners have developed many
“rules of thumb” to help interpret these values (i.e. a correlation greater
than .9 is generally understood to be large). Such correlations can be difficult to interpret ([15]). It is well known that a “spurious” correlation will be
observed when two time series are themselves dependent on an unobserved
third time series.
However, it is less well known to some practitioners that one may observe
“volatile” correlation in independent time series. The correlation is volatile
in the sense that its distribution is heavily dispersed and is frequently large
in absolute value. Yule observed this empirically and, in his 1926 seminal
paper ([15]), called it “nonsense” correlation, asserting that “we sometimes
obtain between quantities varying with time (time-variables) quite high correlations to which we cannot attach any physical significance whatever, although under the ordinary test the correlation would be held to be certainly
significant.” Yet Yule’s empirical finding remained “isolated” from the literature until 1986 (see [1]), when the authors of [6] and [12] confirmed many
of the empirical claims of “spurious regression” made by the authors of [5].
However, despite these advances, mathematical confirmation of “nonsense”
correlation has remained open. We now turn to closing this matter.
Throughout this work, we will use the word “volatile” in lieu of “nonsense.” We emphasize that volatile correlation is distinctly different from
“spurious” correlation, as the latter refers to a third time series, whereas the
former does not. Despite Yule’s empirical findings, it is often (erroneously)
assumed that a large correlation between such pairs must have a cause (see
[10]) when they definitionally do not. Suppose, for example, that Xi = Si
and Yi = Si0 where Si and Si0 are the partial sums of two independent random
walks. The empirical correlation is defined in the usual way as
Pn
Pn
1 Pn
0
0
i=1 Si Si − n ( i=1 Si )( i=1 Si )
q
θn = qP
.
(1)
Pn
n
1 Pn
1 Pn
2
0
0
2
2
2
i=1 Si − n ( i=1 Si )
i=1 (Si ) − n ( i=1 Si )
Nevertheless, it is sometimes (erroneously) assumed that for large enough n
these correlations should be small. Indeed, large values are quite probable.
The histogram of the empirical correlation of independent random walks is
widely dispersed over nearly the entire range. This was recently mentioned
by [11], which presents (using n = 149) a critique of efforts to reconstruct the
earth’s historical temperatures using correlated time series. We reproduce
2
the histogram here as Figure 1, with n = 10, 000.
Figure 1: Simulated empirical correlation of partial sums of two independent
random walks with n = 10, 000.
The above histogram reports that the middle 95% of the observed correlation coefficients fall in the interval [-.83, .83]. The lesson to be learned
from Figure 1 is that correlation is not always a useful statistic; in the case
of two independent random walks, the observed correlation coefficient has a
very different distribution than that of the nominal t-distribution.
Remark 1. It should be noted that the statistic which
P uses the
P actual random variables, Xk , Xk0 instead of their partial sums, kj=1 Xj , kj=1 Xj0 does
not produce volatile correlation. Indeed, as n → ∞, if the random variables,
Xk , Xk0 , k = 1, . . . are I.I.D. sequences, independent of each other, with positive finite variances, then
θn0
= q P
n
1
n
1
n
Pn
2
k=1 Xk
0
k=1 Xk Xk
−
( n1
Pn
−
1
n
k=1 Xk
Pn
k=1 Xk
q P
1
n
)2
3
1
n
Pn
0
k=1 Xk
n
2
k=1 Xk
−
( n1
Pn
k=1 Xk
)2
is easily seen to tend to zero (by the law of large numbers). This shows that
the volatile correlation is a consequence of using the partial sums in place of
the variables themselves. The reason that the partial sums are self-correlated
(and thereby induce large correlation) seems related to the arcsine law. The
history of Sparre Andersen’s major combinatorial contribution to the proof of
the arcsine law (see [4]) raises the question of whether a formula for discrete
sequences of partial sums can be derived by combinatorial methods employing
cyclic permutations. This would be very elegant, and would greatly simplify
earlier works of Erdös and Kac (see [3]), but seems unlikely.
In lieu of considering the volatile correlation between two independent
random walks, we consider the continuous analog, namely that of two independent Wiener processes. Although it would be ideal to find the full
analytical distribution of the empirical correlation between two independent
Wiener processes, finding the second moment itself suffices as formal confirmation of Yule’s “nonsense” correlation. But such evidence, until now, has
remained elusive.
Formally, let Wi (t), 0 ≤ t ≤ 1, i = 1, 2 denote two independent Wiener
processes on [0, 1]. We analytically find the second moment of the empirical correlation between W1 (t) and W2 (t), where the empirical correlation is
defined as
R1
R1
W1 (t)W2 (t)dt − 0 W1 (t)dt 0 W2 (t)dt
θ=r
R
2 rR
R
2 .
R1 2
1
1
1
2
0 W1 (t)dt
0 W2 (t)dt
0 W1 (t)dt −
0 W2 (t)dt −
R1
0
(2)
Using tools from integral equation theory, we successfully calculate the second moment of θ to obtain a value for the standard deviation of θ of nearly
.5. This volatile correlation is induced because each Wiener process is “selfcorrelated” in time. This is because a Wiener process is an integral of pure
noise and thus its values at different time points are correlated. The correct
intuition for the occurrence of this phenomenon is that a Wiener process
self-correlated, and thus volatile correlation is indeed induced between the
independent Wiener processes.
Of course, it is rather simple to simulate the distribution of θ. We do so
below using a simple Monte Carlo routine. The result of averaging 10,000
simple Monte Carlo iterations of the first ten moments of θ can be found
below in Table 1. The initial validity of the Monte Carlo can be justified
by noting that the odd moments should all be zero, since θ is symmetric.
4
E θ0
1
E θ6
.0609591
E θ1
-.00116634
E θ7
-.00016797
E θ2
.235057
E θ8
.0378654
E θ3
-.000524066
E θ9
-.000105611
E θ4
.109276
E θ10
.0251693
E θ5
-.000283548
Table 1: Moments of θ obtained from 10,000 Monte Carlo iterations
The impact of this problem for practitioners of statistical inference, however, renders it most deserving an analytical solution, and this serves as our
paper’s primary motivation.
The structure of the paper is as follows. In §2, we present results needed
for obtaining the distribution of θ. In §3, we provide implicit formulas for
all moments of θ. Most significantly, we explicitly obtain the following expression for the variance of θ by comparing coefficients of z 2 , which we show
(in Proposition 13) are given by the following double integral:
∞
Z
u1
Z
du1
0
0
r
1
u1 u2
u1 u2 u21 1 −
sinh u1 sinh u2 u1 + u2
u1
cosh u1
sinhu1
−
1
u2
2
1−
u2
cosh u2
sinh u2
u1 − u2
Although it is not possible to calculate the double integral above in elementary terms, it has removable singularities and shockingly converges very
nicely at all points where any of u1 , u2 , or u1 − u2 vanishes.
2
A few results needed for obtaining the distribution of θ
In §2.1, we rewrite θ in an alternate form that will be useful in §3. The
alternate form involves stochastic integrals rather than integrals of a Wiener
process itself. In §2.2, we introduce the function F , which is well suited to
calculating the moments of θ. Furthermore, we explicitly calculate F .
2.1
Rewriting θ
The author of [9] gave the moments of the ratio of a pair of quadratic forms
in normal variables. Our work gives a method for the correlation coefficient
which is a ratio involving three quadratic forms of normal variables as well
as square roots. The three-form problem solved here requires a new method.
We begin this task by rewriting θ.
Recall the definition
R 1 of empirical correlation written in Equation (2).
Noting that mi := 0 Wi (t)dt, i = 1, 2 are the empirical mean values, we
5
du2 .
rewrite θ as:
R1
(W1 (t) − m1 )(W2 (t) − m2 )dt
qR
.
1
2 dt
2 dt
(W
(t)
−
m
)
(W
(t)
−
m
)
1
1
2
2
0
0
0
θ = qR
1
(3)
From Equation (3) it is easy to see, by Cauchy-Schwarz, that −1 ≤ θ ≤ 1.
Proposition 1. We have the equality
X1,2
θ=p
,
X1,1 X2,2
where
(4)
1Z 1
Z
(min(s1 , s2 ) − s1 s2 ) dWi (s1 )dWj (s2 ).
Xi,j =
0
(5)
0
Proof. It is clear that
Y1,2
,
θ=p
Y1,1 Y2,2
where
1
Z
1
Z
W1 (t)W2 (t)dt −
Y1,2 = Y2,1 =
W1 (s)ds
0
1
Z
Wi2 (t)dt
Yi,i =
W2 (t)dt
0
and
1
Z
−
0
1
Z
(6)
0
2
Wi (t)dt , i = 1, 2.
(7)
0
We must show that
Yi,j = Xi,j for i, j ∈ {1, 2}.
By the Fundamental Theorem of Calculus,
1
Z
Xi,j =
Z
1
dWi (s1 )
0
Z
min(s1 ,s2 )
dWj (s2 )
0
1
Z
dt−
0
1
Z
dWi (s1 )
Z
s1
dWj (s2 )
0
0
Z
ds
0
dt.
0
(8)
To simplify the right hand side of Equation (8), we change variables and
integrate over the equivalent regions. The right hand side of Equation (8)
becomes
Z
1
Z
dt
0
1
Z
t
1
Z
dWj (s2 ) −
dWi (s1 )
t
1
Z
ds
0
6
1
Z
dWi (s1 )
s
1
Z
dt
0
1
dWj (s2 ).
t
s2
(9)
Rearranging, we obtain
1
Z
Xi,j
0
1
Z
1
Z
Wi (t)Wj (t)dt −
=
Z
(Wi (1) − Wi (s))ds
0
0
2.2
1
Z
(Wi (1) − Wi (t)) (Wj (1) − Wj (t)) dt −
=
Z
(Wj (1) − Wj (t))dt
0
1
Wi (s)ds
0
1
Wj (t)dt = Yi,j .
0
Defining and calculating F
We define for |a| ≤ 1, βi ≥ 0, i = 1, 2, the integral
β2
β2
aβ1 β2 X1,2 − 21 X1,1 − 22 X2,2
,
F (β1 , β2 , a) = E e
(10)
where the Xi,j are as defined in Equation (5). Under the above conditions,
the
p expectation is finite and thus F is well defined. This is because X1,2 =
θ X1,1 X2,2 and |θ| ≤ 1 so the exponent is at most
2
p
1 p
β1 X1,1 − β2 X2,2 ≤ 0.
(11)
2
Thus the expectand is bounded by unity, and so, for this range, F (β1 , β2 , a) ≤
1.
−
2.2.1
The functions M , K, TM , and TK
Motivated by the definition of Xi,j in Equation (5), we define M by
M (s1 , s2 ) = min(s1 , s2 ) − s1 s2 , s1 , s2 ∈ [0, 1],
(12)
which is the covariance of pinned Wiener process on [0, 1].
For i1 , i2 ∈ {1, 2} and s1 , s2 ∈ [0, 1], we define the kernel function Ki1 ,i2
by
Ki,i (s1 , s2 ) = −βi2 M (s1 , s2 )
and
K1,2 (s1 , s2 ) = K2,1 (s1 , s2 ) = aβ1 β2 M (s1 , s2 ).
Let
K=
K1,1 K1,2
K2,1 K2,2
7
.
(13)
M gives rise to a linear transformation TM : L2 [0, 1] → L2 [0, 1] given by
Z 1
M (s1 , s2 )g(s1 )ds1
(14)
TM (g)(s2 ) =
0
2
where g ∈ L2 [0, 1]. If f~(s) and ~g (s) are elements of L2 [0, 1] , then their
inner product is defined as
Z 1
f~(s) · ~g (s)ds.
0
It is straightforward to check that TM is a self-adjoint, positive-definite, and
continuous linear transformation.
Let M2×2 refer to 2 × 2 matrices with entries in L2 ([0, 1] × [0, 1]).
2
We define L2 [0, 1] = L2 [0, 1] × L2 [0, 1] and note that an element of
2
L2 [0, 1] is an ordered pair of maps from [0, 1] to R. If G is any element of M2×2 (L2 ([0, 1] × [0, 1])), G gives rise to a linear transformation
2
2
TG : L2 [0, 1] → L2 [0, 1] given by
Z 1
TG (~g )(s2 ) =
G(s1 , s2 )~g (s1 )ds1
(15)
0
2
for ~g ∈ L2 [0, 1] . In particular, K gives rise to TK via the above definition.
It is straightforward to check that TK is a self-adjoint, positive-definite, and
continuous linear transformation.
2.2.2
The calculation
The goal of this section is to prove Theorem 2 below.
Theorem 2.
F (β1 , β2 , a) = q
1
sinh c+ sinh c−
c+
c−
,
(16)
where
s
±
±
c = c (β1 , β2 , a) =
(β12 + β22 ) ±
p
(β12 − β22 )2 + 4a2 β12 β22
.
2
(17)
The proof of Theorem 2 has two parts. First, we show (in Proposition
5) that F can be expressed in terms of the Fredholm determinant det(I −
TK ), where TK was defined in Equation (15). Second, we compute this
determinant (in Proposition 7) by calculating the eigenvalues of TK . To
begin, we write F in terms of K.
8
Proposition 3. We have
h 1 P2 P2 R 1 R 1
i
F (β1 , β2 , a) = E e 2 i1 =1 i2 =1 0 0 Ki1 ,i2 (s1 ,s2 )dWi1 (s1 )dWi2 (s2 ) .
(18)
Proof. This follows from plugging in the definition of Ki1 ,i2 from Equations
(13) and (14) and comparing it to the definition of F given in Equation
(10).
Since TK is self-adjoint and positive-definite, there exists a (countable)
~ n (n ∈ N) for TK .
orthonormal basis of (L2 [0, 1])2 consisting of eigenvectors φ
As a matter of notation, we write
φn (1, s1 )
~
φn =
.
φn (2, s1 )
We let αn be the corresponding eigenvalues.
Proposition 4. For i1 , i2 ∈ {1, 2}, s1 , s2 ∈ [0, 1], we have
Ki1 ,i2 (s1 , s2 ) =
∞
X
αn φn (i1 , s1 )φn (i2 , s2 ).
(19)
n=1
~nφ
~ T ∈ M2×2 (L2 ([0, 1] × [0, 1])). The proposition
Proof. Let Gn (s1 , s2 ) = φ
n
P
is equivalent to the equation K = ∞
n=1 αn Gn . Thus, we need only show
that
∞
X
TK =
αn TGn .
(20)
n=1
~ n form a basis for L2 [0, 1] 2 , it suffices to check both
Since the vectors φ
~m.
sides of Equation (20) on an arbitrary basis element φ
~ m yields
By definition, the left hand side of Equation (20) applied to φ
~ m (note that φ~m is chosen among the eigenvectors for TK , i.e., it is one
αm φ
of the φ~n ). So it suffices to show that
~ m ) = δm,n φ
~m.
TGn (φ
We calculate
T
Z 1
φn (1, s2 )
φn (1, s1 )
φm (1, s1 )
~
TGn (φm ) =
ds1
φn (2, s2 )
φn (2, s1 )
φm (2, s1 )
0
Z 1
φn (1, s2 )
~n · φ
~ m ds1 .
=
φ
φn (2, s2 )
0
~ n and φ
~ m are orthonormal,
Since φ
~ m ) = φn (1, s2 ) δm,n = δm,n φ
~m.
TGn (φ
φn (2, s2 )
The proof is now complete.
9
Proposition 5. F (β1 , β2 , a) = (det(I − TK ))−1/2 :=
Q
n (1
− αn )−1/2 .
Proof. By substituting Equation (19) into Equation (18), we obtain
R1
Y 1 P2
2
φ
(i,s
)dW
(s
)
α
)
(
n
n
i
i
i=1
0
E e2
.
F (β1 , β2 , a) =
(21)
n
Letting ξi be real numbers, the right hand side of Equation (21) is equal to
Q
−1/2
.
n (1 − αn )
Now, we proceed to calculate the eigenvalues of TK . We do this by
“guessing” the eigenvectors. Because the entries in K are scalar multiples
of each other, we guess that the eigenvectors are of separable form:
ξ1 φ(s)
~
, s ∈ [0, 1],
(22)
φ(s) =
ξ2 φ(s)
where φ(s) ∈ L2 [0, 1] is an eigenvector of TM and ξi are real numbers. It can
2
be easily shown that these “guessed” eigenvectors span the entire L2 [0, 1] .
√
It is straightforward to verify that the functions ψn (t) := 2sin (πnt)
for n ≥ 1 and t ∈ [0, 1] form an orthonormal basis of L2 [0, 1] consisting
of eigenvectors for TM . The eigenvalue λn corresponding to ψn can be
calculated to be π21n2 , i.e.
TM ψn (t) =
We find that
ξ1 φ(s)
ξ2 φ(s)
1
π 2 n2
ψn (t).
is an eigenvector of TK with eigenvalue α if and
only if
− β12 ξ1 + aβ1 β2 ξ2 =
α
ξ1
λn
(23)
and
aβ1 β2 ξ1 − β22 ξ2 =
α
ξ2 .
λn
(24)
Proposition 6. For each eigenvalue λn of TM , there are two corresponding
eigenvalues γn± of TK , where
p
−(β12 + β22 ) ± (β12 − β22 )2 + 4a2 β12 β22
±
γn = λn
.
2
10
Proof. To find the eigenvalues, we solve Equations (23) and (24) for αn . We
can view the system of Equations (23) and (24) as the system A~v = λ~v ,
where
−β12 aβ1 β2
A=
aβ1 β2 −β22
and
~v =
ξ1
ξ2
and
λ=
α
.
λn
The solutions for α are λn multiplied by the eigenvalues of A. Calculating
the eigenvalues of A finishes the proposition.
Proposition 7. If αn , n ≥ 1 is the list of eigenvalues for TK , then
∞
Y
sinh c+ sinh c−
(1 − αn ) =
,
c+
c−
n=1
where c± is as defined in Equation (17).
Proof. First note that the set {αn , n ≥ 1} is the same as the set {γn± , n ≥ 1}.
Thus,
∞
∞
Y
Y
(1 − αn ) =
(1 − γn+ )(1 − γn− )
n=1
=
n=1
∞
Y
n=1
where
s
±
z =
(z + )2
1− 2 2
π n
Y
∞
(z − )2
sin (z + ) sin (z − )
1− 2 2 =
(25)
,
π n
z+
z−
−(β12 + β22 ) ±
n=1
p
(β12 − β22 )2 + 4a2 β12 β22
.
2
(26)
The final equality in Equation (25) follows from the following product
formula (see [2])
∞
Y
sin (z)
z2
=
1− 2 2 .
z
π n
n=1
11
(27)
Note that for any complex number z,
sin (z)
sinh (−iz)
=
.
z
−iz
Since z ± are purely imaginary, we observe z ± = ic∓ and the c∓ are nonnegative real. In particular, since c∓ = −iz ± , we have that
sinh (−iz + ) sinh (−iz − )
sinh c+ sinh c−
sin (z + ) sin (z − )
=
=
.
z+
z−
−iz +
−iz −
c+
c−
This completes the proof.
Proof of Theorem 2. The proof immediately follows from combining
Proposition 5 and Proposition 7.
Remark 2. Both of c± are nonnegative if |a| ≤ 1, but if
s
2(β12 + β22 )π 2 + π 4
|a| ≥ 1 +
4β12 β22
(28)
then F (β1 , β2 , a) = ∞ since the term involving c− vanishes as c− becomes
imaginary and the term in the denominator becomes sin (−π) = 0. However,
we will only need F (β1 , β2 , a) for |a| < 1 to obtain the distribution of θ.
3
Obtaining the integral equation for the moment
generating function of θ
In this section, we derive a formula for the moments of θ. Our strategy is
to use integral equations to derive the distribution of θ. The methods we
employ draw inspiration from some related ideas and approaches in earlier
works of [7, 8, 13, 14].
3.1
The Form for the Moments of θ
In this chapter, we will generally denote derivatives with subscripts. However, we will sometimes denote derivatives of functions of three arguments
with respect to the third argument by primes rather than subscripts because
it is more natural for Theorem 8 below. For example,
Fi (β1 , β2 , z) =
∂
∂
F (β1 , β2 , z), i = 1, 2; F3 (β1 , β2 , z) = F 0 (β1 , β2 , z) =
F (β1 , β2 , z).
∂βi
∂z
The goal of §3.1 is to prove the following:
12
Theorem 8. The moments of θ, where θ is defined in Equation (2), satisfy
Z
Z ∞
∞
X
z 2n 2n (n!)2 22n
dβ1 ∞ dβ2 0
E θ
=
zF (β1 , β2 , z).
2n
(2n)!
β1 0 β2
0
n=1
We first introduce the function G = G(γ1 , γ2 , a), given by
Z ∞
Z
dβ1 ∞ dβ2
G=
V̄ ,
β1 0 β2
0
(29)
where V̄ is defined as:
F (β1 , β2 , a) − F
√
a
γ1 β1 , β2 , √
γ1
−F
β1 ,
√
a
γ2 β2 , √
γ2
+F
√
√
a
γ1 β1 , γ2 β2 , √
γ1 γ2
.
Remark 3. Note that if γi ≥ 1, βi ≥ 0, i = 1, 2, and |a| < 1, the quantities
a
a
a
√
√
√
√
γ1 β1 , β2 , √
γ1 β1 , γ2 β2 , √
F
, F β1 , γ2 β2 , √
,F
γ1
γ2
γ1 γ2
are finite and well defined. One can (with some algebra) show that the right
hand side of Equation (29) converges and thus G is also well defined.
Proposition 9. Under the conditions of Remark 3, the expression
∂ 2 G(γ1 , γ2 , a)
∂γ1 ∂γ2
is simultaneously equal to
∞
Z
∞
Z
dβ1
0
and
Z
∞
0
0
dβ1
β1
Z
0
∞
β2
β2
β1 β2
aβ1 β2 θ −γ1 21 −γ2 22
dβ2
E e
e
e
4
dβ2 ∂ 2
F
β2 ∂γ1 ∂γ2
√
√
a
γ 1 β 1 , γ 2 β2 , √
γ1 γ2
(30)
.
(31)
Proof. Plugging in the definition of F from Equation (10), Equation (29)
can be rewritten as
Z
G(γ1 , γ2 , a) =
0
∞
dβ1
β1
∞
Z
0
β2 X
β2 X
2X
2X
γ1 β1
γ2 β2
1,1
2,2
1 1,1
2 2,2
dβ2
2
2
E eaβ1 β2 X1,2 e− 2
− e−
e− 2
− e−
.
β2
The second integrand of the above equation is everywhere positive and if
we take the expectation outside the integrals, then for each fixed ω ∈ Ω
(the probability space where Wi (s, ω) are defined), we can replace βi , by
√ βi , i = 1, 2. We then rewrite G(γ1 , γ2 , a) as follows:
Xi,i (ω)
13
∞
Z
dβ1
β1
G(γ1 , γ2 , a) = E
0
∞
Z
0
2
2
2
γ1 β1
β2
γ2 β2
dβ2 aβ1 β2 θ(ω) − β12
e
e 2 − e− 2
e− 2 − e− 2
.
β2
Putting the expectation back inside the integral, we obtain
Z
G(γ1 , γ2 , a) =
0
∞
dβ1
β1
∞
Z
2
2
2
2
γ1 β1
γ2 β2
β2
dβ2 h aβ1 β2 θ i − β21
E e
e
− e− 2
e− 2 − e− 2
. (32)
β2
0
Recall the corollary of Fubini’s theorem that states that integration and
differentiation with respect to a parameter can be interchanged if the integral
of the differentiated integrand converges absolutely. One can show this is
true in Equation (32) for |a| < 1. Thus, for |a| < 1, we obtain
∂ 2 G(γ1 , γ2 , a)
=
∂γ1 ∂γ2
Z
∞
Z
∞
dβ1
0
dβ2
0
β1 β2 aβ1 β2 θ −γ1 β12 −γ2 β22
2 e
2 .
E e
e
4
On the other hand, using the definition of G from Equation (29) directly,
we obtain
∂ 2 G(γ1 , γ2 , a)
∂γ1 ∂γ2
Z
∞
=
0
dβ1
β1
Z
0
∞
dβ2 ∂ 2
F
β2 ∂γ1 ∂γ2
√
√
a
γ 1 β1 , γ 2 β 2 , √
γ1 γ2
.
This completes the proof.
Our next proposition gives our first expression for the moments of θ. We
now fix some notation. Let Fi denote the derivative with respect to the ith
argument, as follows:
∂
Fi (β1 , β2 , z) =
F (β1 , β2 , z), i = 1, 2;
∂βi
∂
F3 (β1 , β2 , z) = F 0 (β1 , β2 , z) =
F (β1 , β2 , z).
∂z
Proposition 10. We have
∞
X
=
(n!)2 22n
z 2n E θ2n
(2n)!
n=0
Z ∞
Z ∞
dβ1
dβ2 2 00
z F (β1 , β2 , z) + zF 0 (β1 , β2 , z) − zβ1 F10 (β1 , β2 , z) − zβ2 F20 (β1 , β2 , z) + β1 β2 F12 (β1 , β2 , z)
β1 0 β2
0
Proof. Our strategy is to simplify expressions (30) and (31) and to set them
equal to each other using Proposition 9. We begin
aβnow
by simplifying the
β
θ
1
2
expression in (30). We do this by expanding E e
as a power series
14
β2
and substitute 2i with ui , i = 1, 2, and do the integrals on ui, term-by
term, to obtain an expression involving the moments, µ2n = E θ2n , of θ
(note that the odd moments, µ2n+1 , n ≥ 0, vanish, by symmetry). Algebraic
manipulation yields a simpler form of the expression in (30), displayed below:
Z
Z ∞
∞
X
a2n
1 ∞
du1
(2u1 )n e−γ1 u1 (2u2 )n e−γ1 u2 du2
µ2n
(2n)!
4
0
0
n=0
2n
∞
2 2n
X
(n!) 2
1
a
µ2n
.
√
γ
γ
(2n)!
4γ
1 2
1 γ2
n=0
=
(33)
Thus we see that the simplified form of expression (30) is 4γ11 γ2 times a
function of z = √γa1 γ2 , and of course the expression (31) must also be of this
√
2
√
γ1 β1 , γ2 β2 , √γa1 γ2
form. Applying the multivariate chain rule to the expression ∂γ∂1 ∂γ2 F
from expression (31) yields
∂
∂γ2
!
a
a
1 β1
−a
√
√
√
√
γ 1 β1 , γ 2 β2 , √
γ 1 β1 , γ 2 β2 , √
+
√ F1
3
1 F3
2 γ1
γ1 γ2
γ1 γ2
2γ12 γ22
which in turn equals
β1 β2
−aβ1
−aβ2
a
a2
F12 + p
F13 + p
F23 + p
F3 + 2 2 F33 . (34)
√
2
3
3
2
3
3
4 γ1 γ2
4γ
4 (γ1 ) (γ2 )
4 (γ1 ) (γ2 )
4 (γ1 ) (γ2 )
1 γ2
Note that in expression (34), for the sake of brevity, we dropped the arguments
a
√
√
γ1 β1 , γ2 β2 , √
(35)
γ1 γ2
of F and its derivatives. Substituting the right hand side of expression (34)
into expression (31) and replacing βi by √βγi i , i = 1, 2, and setting z = √γa1 γ2 ,
we obtain
1
4γ1 γ2
Z
0
∞
dβ1
β1
Z
0
∞
dβ2 2 00
z F + zF 0 − zβ1 F10 − zβ2 F20 + β1 β2 F12 .
β2
(36)
The proposition now follows by equating expression (36) with the expression in (33).
The following proposition completes the proof of Theorem 8.
15
Proposition 11. Equation (33) can be simplified to
Z
Z ∞
∞
X
z 2n 2n (n!)2 22n
dβ1 ∞ dβ2 0
zF (β1 , β2 , z).
E θ
=
2n
(2n)!
β1 0 β2
0
(37)
n=1
Proof. We first consider the third, fourth, and fifth terms in the integrand of
the right hand side of Equation (33). The third and fourth terms give zero
upon integration. For example, the fourth term is again an exact differential
in β2 , so
∞
∞
Z ∞
dβ1 0
dβ1 0
dβ2 −
zF2 (β1 , β2 , z) = z
F (β1 , 0, z) = 0
(38)
β1
β1
0
0
0
β2
− 22 X1,1
because F (β1 , 0, z) = E e
, which is constant in z and so has a
Z
Z
derivative of zero with respect to z. The same argument holds for the third
term. The fifth term gives unity upon integration because the β1 β2 terms in
the numerator and denominator cancel and the integrand becomes an exact
differential, giving
Z
∞
∞
Z
∞
dβ2 F12 (β1 , β2 , z) = −
dβ1
0
Z
0
dβ1 F1 (β1 , 0, z) = F (0, 0, z) = Ee0 = 1.
0
The sum of the first and second terms can be written as:
z 2 F 00 (β1 , β2 , z) + zF 0 (β1 , β2 , z) = z
d
zF 0 (β1 , β2 , z) .
dz
(39)
We may subtract unity from both sides of Equation (33), divide by z, and
integrate to obtain
Z ∞
Z
∞
X
dβ1 ∞ dβ2 0
z 2n 2n (n!)2 22n
E θ
=
zF (β1 , β2 , z).
2n
(2n)!
β1 0 β2
0
(40)
n=1
This finishes the proof.
3.2
Isolating the Moments
The goal of this section is to use Proposition 11 to extract the even moments
of θ. We begin by defining the function S(u) as:
r
S(u) =
u
.
sinh u
16
Recalling the definition of c± in Equation (17), Theorem 2 says
F (β1 , β2 , z) = S(c+ (β1 , β2 , z))S(c− (β1 , β2 , z)).
Differentiating with respect to z and multiplying z yields
z
d
F (β1 , β2 , z) = zS 0 (c+ )(c+ )0 S(c− ) + zS(c+ )S 0 (c− )(c− )0 .
dz
(note that, as before, the prime notation denotes the derivative with respect
to z). For example, the derivative of c+ (with respect to z) is, after some
algebraic manipulation, simplified as:
zb21 b22
1
√ p
.
c+ (b21 − b22 )2 + 4z 2 b21 b22
Define the function T (c) as follows:
T (c) =
1
c
1 S 0 (c)
= 2 1−
cosh c .
c S(c)
2c
sinh c
(41)
Then Equation (40) becomes
Z ∞Z ∞
∞
X
β1 β2 dβ1 dβ2
z 2n 2n (n!)2 22n
.
E θ
= z2
S(c+ )S(c− ) T (c+ ) − T (c− ) p
2 2 2
2 2
2
2n
(2n)!
(β
0
0
1 − β2 ) + 4β1 β2 z
n=1
(42)
We simplify Equation (42) by noting that we can break up the integrand
by finding two regions such that the integral is the same over each of them.
Noting that the integral is symmetric in (β1 , β2 ), we simplify Equation (42)
as
Z ∞
Z β1
∞
X
z 2(n−1) 2n (n!)2 22n
T (c+ ) − T (c− )
E θ
=2
dβ1
β1 β2 S(c+ )S(c− ) + 2
dβ2 .
2n
(2n)!
(c ) − (c− )2
0
0
n=1
(43)
A change of variables will suffice to simplify Equation (43). We proceed by
changing variables from (β1 , β2 ) to (u1 , u2 ), where, for fixed real z ∈ (0, 1),
we set:
u2 (β1 , β2 ) = c− (β1 , β2 , z).
u1 (β1 , β2 ) = c+ (β1 , β2 , z),
The Jacobian of this transformation can be calculated as: J =
(44)
β12 −β22
u21 −u22
√
√
1 − z2,
and for further reference we note that β12 + β22 = u21 +√u22 and β1 β2 1 − z 2 =
u1 u2 , as well as dβ1 dβ2 = J1 du1 du2 . After bringing 1 − z 2 to the left side
and canceling a factor of u21 − u22 , Equation (43) becomes
17
p
1 − z2
ZZ
∞
X
z 2(n−1) 2n (n!)2 22n
u1 u2 S(u1 )S(u2 ) (T (u1 ) − T (u2 ))
q
du1 du2 .
E θ
=2
2n
(2n)!
2
2
U
n=1
(u21 − u22 ) − z 2 (u21 + u22 )
(45)
Here, U is the region of (u1 , u2 ) in which the quantity in the square root of
Equation (45) is positive, i.e.
(
)
r
1−z
U = (u1 , u2 ) : 0 ≤ u2 ≤ u1
.
(46)
1+z
To determine U , first note that the domain of integration in Equation (43)
is bounded by β1 = β2 and β2 = 0 in
√ the first quadrant.√ We make the
following change of variables: u1 = u 1 + v andqu2 = u 1 − v. Noting
that u1 ≥ u2 , the two boundaries become u2 = u1 1−z
1+z and u2 = 0, which
is the description of U in Equation
√(46). The fact that β2 = 0 is equivalent to
u2 = 0 follows from β1 β2 = u1 u2 1 − z 2 along with z ∈ (0, 1), 0 ≤
qu2 ≤ u1 ,
and 0 ≤ β2 ≤ β1 . The fact that β1 = β2 is equivalent to u2 = u1
be derived directly from the formulas for
z in place of a in Equation (17).
For 0 ≤ z < 1, Equation (45) becomes
p
1 − z2
c+
and
c−
1−z
1+z
can
in Equation (17), with
√
√
√
√
Z 1 Z ∞
∞
X
uS(u 1 + v)S(u 1 − v) (T (u 1 + v) − T (u 1 − v)
z 2(n−1) 2n (n!)2 22n
√
du.
E θ
=2
dv
2n
(2n)!
v2 − z2
z
0
n=1
(47)
We now seek to simplify Equation (47). It is easy to check that the right
hand side of Equation (47) converges. Let us define g(v) as:
∞
Z
g(v) =
√
√
√
uS(u 1 + v)S(u 1 − v)T (u 1 + v)du−
0
∞
Z
√
√
√
uS(u 1 + v)S(u 1 − v)T (u 1 − v)du.
0
(48)
By the definition of S and T , and with some algebra, it can be shown that
g(v) is the difference of convergent integrals. We now make the substitution
u
u → √1+v
in the first integral in Equation (48) and the substitution u →
u
√
in the second integral in Equation (48), we can simplify the right hand
1−v
side of Equation (47) as:
Z
2
z
1
dv
√
2
v − z2
Z
0
∞
"
u
S(u)T (u)
S
1+v
r
u
1−v
1+v
!
u
−
S
1−v
r
u
1+v
1−v
!#
du.
(49)
18
We now focus on the term in square brackets in the above expression.
We write its power series representation as:
"
u
S
1+v
r
u
1−v
1+v
!
u
−
S
1−v
r
u
1+v
1−v
!#
=
∞
X
sr (u)v 2r−1 ,
(50)
r=1
for a sequence of functions sr (u), r ≥ 1, which are not easy to obtain. Note
that the right hand side of Equation (50) is analytic in v near v = 0 and is
also odd in v. Placing the right hand side of Equation (50) into expression
(49), and interchanging integrals, expression (49) becomes:
∞ Z
X
r=1
∞
Z
2S(u)T (u)sr (u)du
0
z
1
√
v
v 2(r−1) dv.
v2 − z2
(51)
Substituting v 2 = t2 + z 2 , and noting that vdv = tdt, the second integral in
expression (51) becomes:
Z √1−z 2
Z 1
p
r−1
2
2 r−1
2
z 2 (1 − v 2 ) + v 2
dv,
(z + t ) dt = 1 − z
0
0
√
2
where we
√ have substituted t = v 1 − z . This is quite convenient since the
2
factor 1 − z cancels on both sides of Equation (45). The final result is
the slightly simpler identity below:
Z 1
r−1
∞ X
∞
X
X
z 2(n−1) 2n (n!)2 22n
E θ
=
z 2k
2n
(2n)!
0
r=1
n=1
k=0
!
Z ∞
r−1
(1−v 2 )k v 2(r−1−k) dv
2S(u)T (u)sr (u)du.
k
0
(52)
Comparing coefficients of powers of z gives the moments of the distribution
of θ. Collecting terms with k = n − 1, we can simplify Equation (52) to
obtain the following theorem:
Theorem 12.
Z ∞
∞ Z
2n 2n X 1 r − 1
2 n−1 2(r−n)
(1 − v )
v
dv
2S(u)T (u)sr (u)du,
n 22n r=n 0 n − 1
0
(53)
where the sr are implicitly determined by Equation (50).
E θ2n =
Proposition 13 below gives an even more explicit formula for the second
moment of θ:
19
Proposition 13. The second moment of θ, corresponding to n = 1, can be
calculated explicitly as
∞
Z
u1
Z
r
du1
0
0
1
u1 u2
u1 u2 u21 1 −
sinh u1 sinh u2 u1 + u2
u1
coshu1
sinhu1
−
1
u2
2
1−
u2
coshu2
sinh u2
u1 − u2
du2 .
Proof. For the case n = 1, we simplify Equation (53) as:
E θ2 = 2
Z
∞
Z
1
S(u)T (u)du
0
v
−1
dv
0
∞
X
v 2r−1 sr (u).
(54)
r=1
Placing
the left hand side of Equation (50) into the above equality, we have
E θ2 equals
∞
Z
2
"
1
Z
du
0
v
−1
0
u
S
S(u)T (u)
1+v
r
u
1−v
1+v
!
u
−
S
1−v
r
u
1+v
1−v
!#
dv.
(55)
√
Making the transformation u √
→ u 1 + v in the first term of the bracketed
expression above and u → u 1 − v in the second term of the bracketed
expression above, expression (55) becomes
∞
Z
2
Z
udu
0
1
√
√
√
√
v −1 S(u 1 + v)S(u 1 − v) T (u 1 + v) − T (u 1 − v) dv.
0
√
1 + v and let u2 =
We
now
make
one
final
transformation.
Let
u
=
u
1
√
u
√
u 1 − v. The Jacobian of this transformation is 1−v2 . Note that v =
√
u21 −u22
,
and
1 − v 2 = uu1 u2 2 . After making this transformation, we obtain a
2
2
u1 +u2
simpler form of expression (55), which gives us an explicit formula for the
second moment of θ:
E θ2 =
∞
Z
u1
Z
du1
0
0
2u1 u2
u1 + u2
S(u1 )S(u2 )
T (u1 ) − T (u2 )
du2 .
u1 − u2
(56)
Using the definitions of S and T , we arrive at our final expression for the
second moment of θ as:
∞
Z
u1
Z
du1
0
0
r
u1 u2
u1 u2
sinh u1 sinh u2 u1 + u2
1
u2
1
1−
u1
cosh u1
sinhu1
−
1
u2
2
1−
u2
cosh u2
sinh u2
u1 − u2
This finishes the proof.
Using the above expression for the second moment of θ we numerically
obtain a value of .240522. Recall that the Monte Carlo simulation in Table
1 reported a value of .235057.
Remark 4. One must proceed numerically using Equation (53) to calculate
higher order moments.
20
du2 .
3.3
Open Problems
This work admits many potential extensions. Other Gaussian processes
should be treatable with our methods, for example, a pair of OrnsteinUhlenbeck processes or a pair of correlated Wiener processes. It would be
of particular interest to compare the variances of the correlation coefficient
amongst these cases.
4
Acknowledgments
First and foremost, we are very grateful for the very careful and detailed
work of the anonymous referees, whose detailed and very careful reports
significantly improved the quality of this work. We thank Professor Amir
Dembo for his advice. We are very grateful to Professor Lawrence Brown
for his invaluable input. We are also extremely grateful to Professor Edward George for his careful reading of this manuscript, for his invaluable
suggestions, and, most importantly, for suggesting “volatile correlation” to
describe the phenomenon of “nonsense” correlation.
21
References
[1] J. Aldrich. Correlations genuine and spurious in pearson and yule. Stat
Sci, 10: 364–376, 1995.
[2] R. Boas. Entire Functions. Academic Press, New York, 1954.
[3] P. Erdös and M. Kac. On certain limit theorems of the theory of probability. Bull. Amer. Math. Soc., 52: 292–302, 1946.
[4] W. Feller. An Introduction to Probability Theory and Its Applications.
John Wiley, New York, 1953.
[5] C Granger and D. Newbold. Spurious regression in econometrics. J.
Econometrics, 2: 111–120, 1974.
[6] D.F. Hendry. Economic modelling with cointegrated variables: an
overview. Oxf Bull Econ Stat, 48: 201–212, 1986.
[7] B Logan, C Mallows, S Rice, and L Shepp. Limit distributions of selfnormalized sums. Ann. Probab., 1:788–809, 1973.
[8] B Logan and L Shepp. Real zeros of random polynomials. Proc. London
Math. Soc., 18: 29–35, 1968.
[9] J. Magnus. The exact moments of a ratio of quadratic forms in normal
variables. Ann. Econ. Statist., pages 95–109, 1986.
[10] M. Mann, Z. Zhang, M. K. Hughes, R. S. Bradley, S. K. Miller,
S. Rutherford, and F. Ni. Proxy-based reconstructions of hemispheric
and global surface temperature variations over the past two millenia.
Proc. Natl. Acad. Sci., 105:13252–13257, 2008.
[11] Blakeley B. McShane and Abraham J. Wyner. A statistical analysis of
multiple temperature proxies: Are reconstructions of surface temperatures over the last 1000 years reliable? Ann. Appl. Stat., 5(1):5–44, 03
2011.
[12] P.C.B. Phillips. Understanding spurious regressions in econometrics. J.
Econometrics, 33: 311–340, 1986.
[13] L.C.G. Rogers and L Shepp. The correlation of the maxima of correlated
Brownian motion. J. Appl. Probab., 43: 880–883, 2006.
22
[14] L. Shepp. The joint density of the maximum and its location for a
Wiener process with drift. J. Appl. Probab., 16: 423–427, 1979.
[15] G. U. Yule. Why do we sometimes get nonsense correlations between
time series? a study in sampling and the nature of time series. Roy.
Statist. Soc., 89: 1–63, 1926.
23
| 10math.ST
|
A model of protocell based on the introduction of a
semi-permeable membrane in a stochastic model of catalytic
reaction networks
Roberto Serra
Marco Villani
Department of Physics, Informatics and Mathematics
Modena and Reggio Emilia University
ECLT
European Centre for Living Technology
S. Marco 2940, University of Ca’ Foscari, Venice (IT)
[email protected]
[email protected]
Alessandro Filisetti
ECLT
European Centre for Living Technology
S. Marco 2940, University of Ca’ Foscari, Venice (IT)
[email protected]
Alex Graudenzi
Chiara Damiani
Department of Informatics
Systems and Communication, University of Milan Bicocca
[email protected]
[email protected]
Abstract
The theoretical characterization of the self-organizing molecular structures emerging from ensembles
of distinct interacting chemicals turns to be very important in revealing those dynamics that led to the
transition from the non-living to the living matter as well as in the design of artificial protocells [12, 13,
14].
In this work we aim at studying the role of a semi-permeable membrane, i.e. a very simple protocell
description, in the dynamics of a stochastic model describing randomly generated catalytic reaction sets
(CRSs) of molecules.
Catalytic reaction sets are networks composed of different molecules interacting together leading to the
formation of several reaction pathways. In particular, molecules are representation of molecular species
formed by a concatenation of letters (bricks) from a finite alphabet.
Reaction pathways are formed by two kinds of reaction only, namely condensation, where two molecular
species are glued together in order to create a longer species and cleavage where a molecular species is
divided in two shorter species, Figure 1, Filisetti et al. [3, 4]. We present here only a brief summary of
the principal characteristics and a description of the new features.
An assumption of the model is that each reaction needs a catalyst to occur; hence spontaneous reactions
are neglected at this stage of the model.
To create the reaction scheme, the molecular species involved in the reactions are randomly chosen
at the beginning of the simulation according to a fixed probability p that a random species catalyzes a
random reaction. In accordance with both the species and the reactions created, in different realizations
A. Graudenzi, G. Caravagna, G. Mauri, M. Antoniotti (Eds.):
Wivace 2013 - Italian Workshop on
Artificial Life and Evolutionary Computation
EPTCS 130, 2013, pp. 70–73, doi:10.4204/EPTCS.130.10
c R. Serra et al.
This work is licensed under the
Creative Commons Attribution License.
R. Serra et al.
71
CONDENSATION
K comp
S1 + C → S1C
A
K decomp
K cond
S1C → S1 + C
S1C + S2 → P + C
A
B B A
A A B
B
B
B
A
B
B A A A
B
B
A B B A
B B A
B
B
B
A
A
A B B A
B
A
A A B
B
B
A
B
CLEAVAGE
K cleav
S + C → P1 + P2 + C
A
B B A
B B
B B
A A
A
B
A
B B A
B
B
B A A
B
A
B
Figure 1: A schematic representation of the two allowed reactions. TOP: Condensation reaction. The
white and the grey chains stand for the substrates, the black chain represents the catalyst and the blue
chain stands for the product. The reaction occurs in two steps, the formation of a temporary molecular
complex and the release of the product. BOTTOM: Cleavage reaction. The white chain represents the
substrate, the black chain stands for the catalyst and the blue chains are the two products of the reaction.
the same species can catalyze different reactions leading to the formation of different chemistries, i.e.
different possible artificial worlds.
Several models of catalytic reaction networks have been proposed. Some of them focus on the structural
proprieties of the reaction graph [2, 7, 11] while dynamical investigations in a continuous stirred-tank
reactor (CSTR) have been introduced by Bagley et al. [1], Farmer et al. [2, 10], Hordijk et al. [6, 8, 9]
and Vasas et al. [15].
We introduce in this model a semi-permeable membrane able to entrap all those molecular species longer
than a certain threshold.
In order to reveal the role of the semi-permeable membrane we compare the dynamics of the same chemistry, i.e. the same reaction scheme, within a CSTR and a protocell (kinetic parameters are the same as
well).
An important feature of the model proposed here is that in accordance with the reaction scheme new
molecular species can be created by the dynamics. Each time that a new species is created, according
to p new reactions involving the possible interactions between the new species and the already present
species are created. Thus, the system can grow in time increasing the number of molecular species and
the number of reactions, Filisetti et al. [3, 4].
In order to consider the role of stochastic fluctuations due to the low-number-effects on the overall dynamics, we adopt a stochastic description of the system, thus dynamics are simulated using a hybrid
version the well-known Gillespie algorithm [5].
The first remarkable difference between the dynamics within a CSTR and within a protocell concerns
the overall mass of the system here represented as the total number of bricks present in the system.
In accordance with the nature of the system, a continuous incoming flux of nutrients and a continuous
outgoing flux of all the molecular species, within a CSTR the overall mass reaches always a dynamical equilibrium. On the opposite, a protocell may never reach a dynamical equilibrium. In particular,
72
A model of protocell
while within a CSTR distinct simulations of the same chemistry show the same dynamical equilibrium,
within a protocell the same chemistry reaches different dynamical states starting from the same initial
conditions. A further important outcome is directly connected with the presence of the semi-permeable
membrane. Some molecular species, that are expelled adopting a CSTR framework, remain entrapped
within a protocell if longer than a certain threshold. This process leads to the formation of a pool of catalysts, no longer created by the system dynamics, responsible of the conservation, or of the dismantling,
of certain reaction pathways. It is worthwhile to remark that the direct consequence of this phenomenon
is the formation of different protocell compositions, a necessary requirement for the establishment of a
population of different protocells competing for the same limited resources.
Finally, an important compositional remark concerns the maximum length of the species created by the
dynamics. While within a CSTR, always observing the dynamics of the same chemistry, only species
with a maximum length of 8 are observed, protocells are able to produce species up to length 11 indicating a greater capability to produce more complex compositions.
Further analysis are necessary for a better comprehension of the role played by a semi-permeable membrane on the dynamics of catalytic reaction networks.
References
[1] R J Bagley, J D Farmer, S A Kauffman, N H Packard, A S Perelson & I M Stadnyk (1989):Modeling adaptive
biological systems. Bio Systems 23(2-3), pp. 113–37; discussion 138. Available at http://www.ncbi.
nlm.nih.gov/pubmed/2627562. doi:10.1016/0303-2647(89)90016-6
[2] JD Farmer & SA Kauffman (1986): Autocatalytic replication of polymers. Physica D: Nonlinear Phenomena
220, pp. 50–67. Available at http://linkinghub.elsevier.com/retrieve/pii/0167278986902332.
doi:10.1016/0167-2789(86)90233-2
[3] Alessandro Filisetti, Alex Graudenzi, Roberto Serra, Marco Villani, Davide De Lucrezia, Rudolf M Fuchslin, Stuart A Kauffman, Norman Packard & Irene Poli (2011): A stochastic model of the emergence of
autocatalytic cycles. Journal of Systems Chemistry 2(1), p. 2, doi:10.1186/1759-2208-2-2. Available at
http://www.jsystchem.com/content/2/1/2.
[4] Alessandro Filisetti, Alex Graudenzi, Roberto Serra, Marco Villani, Rudolf M Füchslin, Norman Packard,
Stuart a Kauffman & Irene Poli (2011): A stochastic model of autocatalytic reaction networks. Theory in biosciences = Theorie in den Biowissenschaften, pp. 1–9, doi:10.1007/s12064-011-0136-x. Available at http://www.springerlink.com/content/bll6n41323744577/http://www.ncbi.nlm.nih.
gov/pubmed/21979857.
[5] Daniel T Gillespie (1977): Exact Stochastic Simulation of Coupled Chemical Reactions. The Journal of
Physical Chemistry 81(25), pp. 2340–2361. doi:10.1021/j100540a008
[6] Wim Hordijk, Jotun Hein & Mike Steel (2010): Autocatalytic Sets and the Origin of Life. Entropy 12(7), pp.
1733–1742, doi:10.3390/e12071733. Available at http://www.mdpi.com/1099-4300/12/7/1733/.
[7] Wim Hordijk & Mike Steel (2004): Detecting autocatalytic, self-sustaining sets in chemical reaction systems.
Journal of theoretical biology 227(4), pp. 451–461, doi:10.1016/j.jtbi.2003.11.020. Available at http://
www.ncbi.nlm.nih.gov/pubmed/15038982.
[8] Wim Hordijk & Mike Steel (2012): Predicting template-based catalysis rates in a simple catalytic reaction
model. Journal of theoretical biology 295, pp. 132–8, doi:10.1016/j.jtbi.2011.11.024. Available at http:
//www.ncbi.nlm.nih.gov/pubmed/22142623.
[9] Wim Hordijk, Mike Steel & Stuart Kauffman (2012): The Structure of Autocatalytic Sets: Evolvability,
Enablement, and Emergence, p. 12. arXiv.org Available at http://arxiv.org/abs/1205.0584.
[10] J.D.Farmer, Norman H Packard & Alan S Perelson (1986): The immune system, adaptation, and machine
learning. Physica D 22(2), pp. 187–204. doi:10.1016/0167-2789(86)90240-X
R. Serra et al.
73
[11] S A Kauffman (1986): Autocatalytic sets of proteins. J Theor Biol 119(1), pp. 1–24. doi:10.1016/S00225193(86)80047-9
[12] Steen Rasmussen, Liaohai Chen, Martin Nilsson & Shigeaki Abe (2003): Bridging nonliving and living
matter. Artificial life 9(3), pp. 269–316, doi:10.1162/106454603322392479. Available at http://www.
ncbi.nlm.nih.gov/pubmed/14556688.
[13] Roberto Serra, Timoteo Carletti & Irene Poli (2007): Syncronization phenomena in surface-reaction models
of protocells. Artificial life 123(2), pp. 123–38. doi:10.1038/35053176
[14] J W Szostak, D P Bartel & P L Luisi (2001): Synthesizing life.
doi:10.1038/35053176.
Nature 409(6818), pp. 387–390,
[15] Vera Vasas, Chrisantha Fernando, Mauro Santos, Stuart Kauffman & Eors Szathmary (2012):
Evolution before genes.
Biology Direct 7(1), p. 1, doi:10.1186/1745-6150-7-1.
Available
at
http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=3284417&tool=
pmcentrez&rendertype=abstract.
| 5cs.CE
|
УДК 004.415
Палагин А.В., Петренко Н.Г., Величко В.Ю., Малахов К.С.
К ВОПРОСУ РАЗРАБОТКИ ОНТОЛОГО-УПРАВЛЯЕМОЙ АРХИТЕКТУРЫ
ИНТЕЛЛЕКТУАЛЬНОЙ ПРОГРАММНОЙ СИСТЕМЫ
В работе описана архитектура интеллектуальной программной системы автоматизированного
построения онтологических баз знаний предметных областей и программная модель подсистемы
управляющей графической оболочки.
Ключевые слова: онтология предметной области, инструментальный комплекс онтологического
назначения, управляющая графическая оболочка.
Введение
Одной из основных проблем при создании систем обработки знаний является проблема
автоматизированного построения множеств понятийных структур, отношений между ними и
формально-логического описания, в совокупности составляющих базу знаний (БЗ) предметной области
(ПдО) [1]. Её объём даже для небольших ПдО может составлять несколько тысяч записей. Создание
вручную такой БЗ проблематично. Эффективные и доступные системы автоматизированного
проектирования неизвестны (за небольшим исключением) даже для англоязычных приложений. Для
российского и украинского сегментов указанные системы просто отсутствуют. Поэтому, разработка
интеллектуальной программной системы (далее – инструментального комплекса онтологического
назначения (ИКОН)) автоматизированного построения онтологических баз знаний ПдО является
важной и актуальной. Кроме того, полезность ИКОН существенно возрастёт, если включить в обработку
украинские и российские естественно-языковые источники предметных знаний.
Архитектура инструментального комплекса
ИКОН автоматизированного построения онтологических баз знаний ПдО является системой,
реализующей одно из направлений комплексной технологии Data Mining, а именно – анализ и
обработку больших объёмов неструктурированных данных, в частности лингвистических корпусов
текстов на украинском и/или русском языках, извлечение из последних предметных знаний с
последующим их представлением в виде системно-онтологической структуры или онтологии
предметной области (О ПдО) [1, 2]. Он предназначен для реализации ряда информационных
технологий (ИТ), основными из которых являются:
- автоматическая обработка естественно-языковых текстов (Natural Language Processing) [3];
- извлечение из множества текстовых документов знаний, релевантных заданной ПдО, их системноонтологическая структуризация и формально-логическое представление на одном (или нескольких) из
общепринятых языков описания онтологий (Knowledge Representation).
ИКОН состоит из трёх подсистем и представляет собой интеграцию разного рода
информационных ресурсов, программно-аппаратных средств обработки информации и естественного
интеллекта (ЕИ), которые реализуют совокупность алгоритмов автоматизированного, итерационного
построения понятийных структур предметных знаний, их накопления и/или системной интеграции.
Обобщённая блок-схема ИКОН представлена на рис. 1. На нём приняты следующие сокращения: ЛКТ
– лингвистический корпус текстов, ТД – текстовый документ.
Информационный ресурс среды включает блок формирования ЛКТ, базы данных обработки
языковых структур и библиотеки понятийных структур. Первый компонент представляет собой
различные источники текстовой информации, поступающей на обработку в систему. Второй компонент
представляет собой различные базы данных обработки языковых структур, часть из которых
формируется (наполняется данными) в процессе обработки ТД, а другая часть формируется до
процесса построения О ПдО и, по сути, является электронной коллекцией различных словарей. Третий
компонент представляет собой совокупность библиотек понятийных структур разного уровня
представления (от наборов терминов и понятий до высокоинтегрированной онтологической структуры
междисциплинарных знаний) и является результатом реализации некоторого проекта (проектирования
онтологии ПдО и/или системной интеграции онтологий).
В качестве языка программирования для разработки ИКОН был выбран объектноориентированный язык Java [4] и свободная интегрированная среда разработки приложений (IDE) на
языках программирования Java, JavaFX, Python, PHP, JavaScript, C++ и ряде других – Netbeans IDE [5].
1
Рис.1. Обобщённая блок-схема ИнКом
Программно-аппаратные средства включают блоки обработки языковых и понятийных структур
и управляющую графическую оболочку (УГО). Последняя, во взаимодействии с естественным
интеллектом, осуществляет общее управление процессом реализации связанных информационных
технологий.
Подсистема ЕИ осуществляет подготовку и реализацию процедур предварительного этапа
проектирования, в процессе основного цикла проектирования осуществляет контроль и оценку
(assessment) результатов выполнения отдельных этапов проекта, принимает решение (возможно с
экспертом в заданной ПдО) о степени завершённости последних и (в случае необходимости)
повторении некоторых из них. Ниже описана программная реализация УГО.
Программная реализация управляющей графической оболочки
В состав программно-аппаратных средств ИКОН входит управляющая графическая оболочка, в
которую включены функции управления всеми компонентами ИКОН и системы в целом. Структурная
схема УГО ИКОН представлена на рис. 2. На нём приняты следующие сокращения: GUI – graphical
user interface, OWL – Web Ontology Language, XML – Extensible Markup Language.
УГО взаимодействует с подсистемой ЕИ, осуществляет общее управление процессом
реализации связанных информационных технологий.
УГО выполняет следующие функции:
- во взаимодействии с инженером по знаниям осуществляет предварительное наполнение среды
материалами электронных коллекций энциклопедических, толковых словарей и тезаурусов,
описывающих домен предметных знаний;
- обеспечивает запуск и последовательность исполнения прикладных программ, реализующих
составные информационные технологии проектирования онтологии ПдО и системной интеграции
междисциплинарных знаний (примером составной технологии является автоматизированное
построение тезаурусов ПдО для поисковой системы);
- отображает ход процесса проектирования;
- содержит позиции меню для запуска, как последовательностей, так и отдельных прикладных
программ, используемых в процессе проектирования;
- обеспечивает интерфейс с подсистемой ЕИ;
- индицирует сообщения о текущем состоянии проекта, его наполнении информационными ресурсами;
- обеспечивает обмен информацией между прикладными программами и базами данных через общую
информационную шину.
2
Рис. 2. Структурная схема УГО ИКОН
Одно из главных преимуществ языка программирования Java – кроссплатформенность. Таким
образом, один и тот же код можно запускать под управлением операционных систем Windows, Linux,
FreeBSD, Mac OS и др. Это очень важно, когда программы загружаются посредством глобальной сети
Интернет и используются на различных компьютерных платформах, включая суперкомпьютерные.
Язык Java содержит большое количество уже готовых классов для разработки программного
обеспечения [6], благодаря чему написание кода программы требует меньшее количество временных
затрат.
Для создания графического интерфейса пользователя инструментальным комплексом
использована библиотека графических компонентов Swing [7, 8]. Swing относится к набору библиотек
классов на языке Java – Java Foundation Classes (JFC), предоставляющих программам на Java удобный
API для создания графического интерфейса пользователя. В состав JFC входят, в частности,
библиотека Swing, Java 2D, AWT, Drag & Drop – API и др. Swing поддерживает специфические
динамически подключаемые стили (look-and-feel) [8, 9]. Благодаря look-and-feel возможна адаптация к
графическому интерфейсу платформы, на которой запускается Java-программа (т.е. к компоненту
можно динамически подключить другой, специфический для операционной системы, в том числе и
созданный программистом, стиль). Таким образом, приложения, использующие Swing, могут выглядеть
как “родные” приложения для данной операционной системы.
Одним из основных дополнительных компонент, используемых при создании УГО ИКОН,
является свободно распространяемая Java-реализация Document Object Model (DOM) для eXtensible
Markup Language (XML) – библиотека Java Document Object Model (JDOM) [10], созданная с учётом
особенностей языка и платформы Java. JDOM является уникальным Java-инструментом для работы с
XML.
На рис. 3 представлено окно главного меню УГО, которое содержит элементы управления
следующими функциями и модулями ИКОН:
- модулем конвертирования формата OWL в простой XML формат КВП (рис. 4).
- модулем визуального проектирования (КВП) (рис. 5);
- модулем синтактико-семантического анализа естественно-языковых текстов (рис. 6);
Функционирование УГО ИКОН рассмотрим на примере анализа небольшого фрагмента
естественно-языкового текста из предметной области “информационные технологии”.
1 Склад обчислювальної системи.
2 Характерною їх відмінністю від механічних пристроїв є те, що вони реєструють не переміщення
елементів конструкції, а їх стани.
3 Це надає зручність опрацювання даних, поданих у двійковій системі числення.
4 Сукупність пристроїв, призначених для автоматизації опрацювання даних, називають
обчислювальною технікою.
5 Центральним пристроєм обчислювальної системи, як правило, є комп’ютер.
6 В сучасному розумінні, комп’ютер – це універсальний електронний пристрій, призначений для
автоматизації накопичення, збереження, опрацювання, передачі та відтворення даних.
7 Склад обчислювальної системи називається конфігурацією.
Для каждого предложения текста строятся деревья зависимостей, в которых определяются
синтактико-семантические отношения между лексемами. Далее автоматически выделяются из текста
термины предметной области в соответствии с заданными шаблонами, например: аббревиатура,
существительное, прилагательное – существительное, существительное – прилагательное –
существительное и типами синтактико-семантических отношений: объектное, принадлежность (между
3
двумя существительными), определительное (между прилагательным и существительным),
однородные слова. В качестве терминов, используемых в формальном представлении текста,
автоматически выбираются преимущественно многословные словосочетания, в которых значения
существительных уточняются с помощью связанных прилагательных. Приведем фрагмент списка
терминов, состоящего из отдельных слов и словосочетаний: ПК, обчислювальна техніка, механічний
пристрій, конфігурація, автоматизація, двійкова система числення, електронний пристрій, конструкція
(рис. 6). Вопросы, связанные с методикой и программной поддержкой процесса построения онтологии,
являются предметом отдельного рассмотрения.
Полученные в результате автоматического анализа текста множества понятий и отношений
между ними являются исходными данными для построения онтологии ПдО (рис. 5).
Заключение
В докладе рассмотрена архитектура инструментального комплекса автоматизированного построения
онтологических баз знаний предметных областей, которая, в частности, включает компоненту
управляющей графической оболочки. Описана программная модель УГО, которая включает элементы
управления модулями синтактико-семантического анализа естественно-языковых текстов, визуального
проектирования и конвертирования формата OWL в простой XML формат КВП. В настоящее время
программная реализация ИКОН, в том числе и УГО, находится в стадии завершения.
Рис. 3. Окно главного меню УГО ИКОН
Рис. 4. Модуль конвертирования формата OWL в простой XML формат КВП ИКОН
4
Рис. 5. Модуль визуального проектирования ИКОН
Рис. 6. Модуль синтактико-семантического анализа естественно-языковых текстов
5
Палагін О.В., Петренко М.Г., Величко В.Ю., Малахов К.С. До питання розробки онтологокерованої інтелектуальної програмної системи.
В роботі описана архітектура інтелектуальної програмної системи автоматизованої побудови
онтологічних баз знань предметних областей та програмна модель підсистеми керуючої графічної
оболонки.
Ключові слова: онтологія предметної області, інструментальний комплекс онтологічного
призначення, керуюча графічна оболонка.
Palagin A.V., Petrenko N.G., Velichko V.U., Malakhov K.S. The problem of the development
ontology-driven architecture of intellectual software systems.
The paper describes the architecture of intelligence system for automated construction of ontological
knowledge bases of subject areas and the programming model subsystem management GUI.
Keywords: domain ontology, tool complex ontological destination, GUI management.
Литература
1.
Гаврилова ТА., Хорошевский В.Ф. Базы знаний интеллектуальных систем / Учебник для вузов.
– СПб, Изд-во “Питер”, 2000. – 384 с.
2.
Палагин А.В., Петренко Н.Г. Системно-онтологический анализ предметной области // УСиМ. –
2009. – № 4. – С. 3–14.
3.
Палагин А.В., Крывый С.Л., Величко В.Ю., Петренко Н.Г. К анализу естественно-языковых
объектов // "Intelligent Processing" International Book Series “INFORMATION SCIENCE &
COMPUTING”, Number 9, Supplement to the International Journal “INFORMATION
TECHNOLOGIES & KNOWLEDGE”, Volume 3, 2009 – FOI ITHEA Sofia, Bulgaria. – 2009. – pp. 36–
43.
4.
http://www.oracle.com/technetwork/java/javase/index.html
5.
http://netbeans.org/
6.
Herb Schildt. Herb Schildt’s Java Programming Cookbook. Product line: McGraw-Hill Osborne Media,
2007.
7.
http://download.oracle.com/javase/tutorial/uiswing/
8.
Шилдт Герберт. Swing: Руководство пользователя для начинающих: Пер. с англ. – М.: ООО “И.
Д. Вильямс”, 2007. – 704 с.
9.
Х. М. Дейтел, П. Дж. Дейтел, С. И. Сантри. Технологии программирования на Java 2: Книга 1.
Графика, JavaBeans, интерфейс пользователя. Пер. с англ. – М.: ООО “Бином-Пресс”, 2003. –
560 с.
10.
Hunter Jason. JDOM and XML parsing // ORACLE MAGAZINE SEPTEMBER/OCTOBER 2002 – pp.
68 – 73.
6
| 2cs.AI
|
1
Design and Performance Analysis of Dual and
Multi-hop Diffusive Molecular Communication
Systems
arXiv:1710.00555v1 [cs.IT] 2 Oct 2017
Neeraj Varshney, Student Member, IEEE, Adarsh Patel, Student Member, IEEE, Aditya K.
Jagannatham, Member, IEEE, and Pramod K. Varshney, Fellow, IEEE
Abstract—This work presents a comprehensive performance
analysis of diffusion based direct, dual-hop, and multi-hop molecular communication systems with Brownian motion and drift in
the presence of various distortions such as inter-symbol interference (ISI), multi-source interference (MSI), and counting errors.
Optimal decision rules are derived employing the likelihood ratio
tests (LRTs) for symbol detection at each of the cooperative as
well as the destination nanomachines. Further, closed-form expressions are also derived for the probabilities of detection, false
alarm at the individual cooperative, destination nanomachines,
as well as the overall end-to-end probability of error for sourcedestination communication. The results also characterize the
impact of detection performance of the intermediate cooperative
nanomachine(s) on the end-to-end performance of dual/multi hop
diffusive molecular communication systems. In addition, capacity
expressions are also derived for direct, dual-hop, and multihop molecular communication scenarios. Simulation results are
presented to corroborate the theoretical results derived and also,
to yield insights into system performance.
Index Terms—Cooperative nanomachines, diffusion, Likelihood Ratio Test (LRT), molecular communication, multi-hop
communication, optimal threshold.
I. I NTRODUCTION
Nanoscale molecular communication has garnered significant interest in recent times towards addressing challenging
problems in biomedical, industrial, and surveillance scenarios
[1], [2]. This has lead to the development of novel applications
such as efficient drug delivery and human body monitoring, using communicating nano-robots [3]–[5]. In contrast to
active molecular communication (AMC)-based systems such
as molecular motors and protein filaments, the molecules
in diffusion-based passive molecular communication (PMC)
systems propagate via Brownian motion in a fluidic medium
without requiring additional infrastructure or external energy
[1], [2]. Several research efforts [6]–[17] have been devoted to
exploring various aspects such as developing channel models,
estimating the channel as well as analyzing the performance
of diffusion based molecular systems. It has been shown
in [18] that the molecular concentration decays inversely
as the cube of the distance between the transmitter and
receiver nanomachines, which severely limits the performance
Neeraj Varshney, Adarsh Patel and Aditya K. Jagannatham are with the
Department of Electrical Engineering, Indian Institute of Technology Kanpur,
Kanpur UP 208016, India (e-mail:{neerajv; adarsh; adityaj}@iitk.ac.in).
Pramod K. Varshney is with the Department of Electrical Engineering
& Computer Science, Syracuse University, Syracuse, NY 13244, USA (email:[email protected]).
of such systems. Relay-assisted cooperative communication
has been shown to successfully overcome this impediment
by significantly enhancing the communication range. This
leads to a substantial improvement in the end-to-end reliability of communication, thus making it a very promising
technology for such systems [19]. However, the decoding
accuracy at the destination in relay-assisted molecular communication depends critically on the detection performance
of the intermediate cooperative nanomachine(s) that act as
relays. The end-to-end performance can further deteriorate due
to other degrading effects such as inter-symbol interference
(ISI), multi-source interference (MSI), and counting errors at
the cooperative as well as destination nanomachines [20]. ISI
at the receiving nanomachine arises due to Brownian motion,
which results in the molecules emitted by the transmitting
nanomachine at the beginning of given time slot arriving
stochastically in subsequent time-slots. On the other hand, MSI
arises from the transmissions of other sources using the same
type of molecules. Some works in the existing literature [21]–
[29] propose techniques for receiver design and detection for
direct source-destination diffusive molecular communication.
However, to the best of our knowledge, none of the works in
the existing literature have analyzed the impact of detection
performance of the intermediate cooperative nanomachine(s)
on the end-to-end performance of the relay-assisted dual and
multi-hop molecular communication in the presence of ISI,
MSI, and counting errors and is, one of the central aims of
the work presented in this paper. Next, we present a detailed
overview and comparative survey of related works in the
existing literature on relay-assisted molecular communication.
A. Related Work
Some works in the existing literature have analyzed the
performance of single relay-assisted molecular communication
systems. In [30], a sense-and-forward relaying strategy has
been proposed for diffusion based molecular communication
between two nodes formed from synthetic bacteria. The analysis is extended in [31] to decode-and-forward (DF) relaying
of M -ary information symbols where it was demonstrated that
optimal combining of the direct and relayed outputs improves
communication reliability. Authors in [32] derived an expression to characterize the average error probability for a two-hop
DF molecular network and subsequently proposed mitigation
techniques for the self-intereference that arises at the relay due
to the detection and emission of the same type of molecules.
The analysis is further extended to amplify-and-forward (AF)
relaying with fixed and variable amplification factors in [33].
The optimal amplification factor at the relay node to minimize
the average error probability of the network is also derived
therein based on an approximation of the same. The authors
in [34] proposed an energy efficient scheme for the information molecule synthesis process employing a simultaneous
molecular information and energy transfer (SMIET) relay.
Performance analysis was presented for the resulting bit error
probability and cost of synthesis. However, the above works
[30]–[34] do not consider practical effects such as MSI, ISI
and counting errors at any of the receiver nanomachines. The
work in [35] analyzed the bit error rate (BER) performance
of a dual-hop DF molecular communication system in which
the time-dependent molecular concentrations are influenced by
noise and ISI resulting from channel. However, the work in
[35] does not consider the effect of MSI while analyzing the
performance.
In the context of multi-hop communication, the design and
analysis of repeater cells in Calcium junction channels was
investigated in [36], where signal molecules released by the
transmitter are amplified by intermediate repeaters in order
to aid them reach the receiver. The authors in [32] extended
the error probability analysis to multi-hop links in [37]. In
[38], a diffusion-based multi-hop network between bacterial
colonies was analyzed with several bacteria agents acting as
a single node. Further, the use of bacteria and virus particles
as information carriers in a multi-hop network was proposed
in [39] and [40], respectively.
The work in [41] recently analyzed the error rate performance of a diffusion based DF dual-hop molecular communication system inside a blood vessel of a human body. The
analysis therein formulates an optimization problem to find the
optimal threshold which minimizes the bit error probability
(BEP) at the destination nanomachine. However, analytical
results are not presented therein for the same. None of the
related studies in the existing literature consider the problem
of optimal detection at the destination nanomachine using LRT
for binary hypothesis testing in diffusion based dual-hop and
multi-hop molecular communication systems inside the blood
vessels of the human body. Moreover, the presence of practical
aberrations such as the ISI, MSI, and counting errors and
the impact of the detection performance of the intermediate
cooperative nanomachines on the end-to-end performance of
relay-assisted dual and multi-hop molecular communication
has not been studied in the literature.
B. Contributions
The main contributions of the paper can be logically organized into three parts as described.
1) Initially, the LRT-based optimal decision rule is determined at the destination nanomachine in the presence of
ISI, MSI, and counting errors in a diffusion based direct
molecular communication system with drift. Further,
closed-form expressions are derived for the probabilities
of detection and false alarm are derived to analytically
characterize the detection performance at the destination
nanomachine. These results are subsequently used to determine the probability of error and the channel capacity
for molecular communication between the source and
destination nanomachines.
2) In the second part of the paper, the above results are
extended to a dual-hop scenario with DF relaying in
which the cooperative nanomachine decodes the symbol using the number of molecules received from the
source nanomachine and forwards it to the destination
nanomachine in the subsequent time-slot. In contrast
to [41], closed-form expressions are presented for the
optimal decision rule including optimal threshold, resulting detection performance, and capacity. Further, the
probabilities of detection and false alarm are also derived
at the intermediate cooperative nanomachines.
3) The Final part presents a general framework to analyze
the performance of a multi-hop molecular communication scenario with N ≥ 2 cooperative nanomachines.
The optimal LRT based decision rule, probabilities of
detection and false alarm are derived at each cooperative
and destination nanomachine incorporating the performance of the intermediate cooperating nonomachines.
The results are subsequently used to obtain analytical
results for the end-to-end probability of error and capacity for the aforementioned multi-hop scenario.
C. Organization
The rest of the paper is organized as follows. The system
model for diffusion based direct communication between the
source and destination nanomachines is presented in Section
II-A. Comprehensive analyses of the probabilities of detection,
and false alarm at the individual cooperative and destination
nanomachines, along with the end-to-end probability of error
and capacity are presented in Section II-B, Section II-C,
and Section II-D respectively. This is followed by similar
results for the dual-hop and multi-hop communication cases
in Sections III and IV respectively. Simulation results for all
the above scenarios are presented in Section V, followed by
conclusions in Section VI.
II. D IRECT S OURCE -D ESTINATION C OMMUNICATION
A. System Model
Consider a diffusion based molecular communication system inside a blood vessel in which the source nanomachine
communicates with the destination nanomachine directly in a
fluid medium with positive drift. In this system, the source
nanomachine is assumed to be a point source with the destination nanomachine located at a distance of dsd from the
source nanomachine. Further, similar to [8], [20], [41] and
the references therein, both nanomachines are assumed to be
synchronized in time and stationary. The channel is divided
into time slots of duration τ as shown in Fig. 1, where the
jth slot is defined as the time period [(j − 1)τ, jτ ] with
j ∈ {1, 2, · · · }. At the beginning of each time slot, the source
nanomachine emits either the same type of molecules in the
propagation medium for information symbol 1 generated with
a prior probability β or remains silent for information symbol
0. Let Q0 [j] denote the number of type1 0 molecules released
by the source nanomachine to transmit the information symbol
x[j] = 1 at the beginning of the jth slot. The molecules
released by the source nanomachine diffuse freely in the
propagation medium and are assumed to be in Brownian
motion with a positive drift in the direction of information
transmission from the source to destination nanomachines.
Similar to the works in [20], [41] and the references therein,
this work also assumes that the transmitted molecules do
not interfere or collide with each other. Moreover, once
these molecules reach the destination nanomachine, they are
assumed to be immediately absorbed by it and not propagate
further in the medium. Finally, the destination nanomachine
decodes the information symbol transmitted by the source
nanomachine using the number of molecules received at the
end of each time slot.
Fig. 1. Diffusion based molecular communication over time-slotted
channel [20], where [(j − 1)τ, jτ ] is the current slot and q0sd is
the probability that a molecule reaches the destination nanomachine
within the current slot.
The molecules emitted by the source nanomachine arrive
stochastically at the destination nanomachine and degrade over
time. This is owing to the fact that the life expectancy of a
sd
molecule decays with time. Let qj−i
denote the probability
that a molecule transmitted in slot i ∈ {1, 2, · · · , j} arrives
at the destination nanomachine in time slot j, which can be
obtained as [8, Eq. (1)]
Z (j−i+1)τ
Z ∞
sd
qj−i
=
f (t)
g(u)dudt,
(1)
(j−i)τ
t
where f (t) is the probability density function (PDF) of the first
hitting time, i.e., the time required for a molecule to reach the
destination nanomachine, and g(u) denotes the PDF of the
molecular life expectancy. The first hitting time f (t) follows
the inverse Gaussian distribution with PDF [43], [44]
r
λ(t − µ)2
λ
, t > 0,
(2)
exp −
f (t) =
2πt3
2µ2 t
where the mean µ and the shape parameter λ are given as,
d2sd
µ = dvsd and λ = 2D
respectively, with dsd denoting the
distance between the source and destination nanomachines,
v, D denoting the drift velocity and diffusion coefficient of
1 Subscript 0 in Q [j] is simply used to denote a type of molecules
0
considered for transmission by the source nanomachine. In the multi-hop
scenario, we assume that each cooperative nanomachines sends different types
of molecules. Therefore, we later used subscripts 1, 2, · · · , N to denote
N different types of molecules transmitted by cooperative nanomachines
R1 , R2 , · · · , RN respectively. Further, it is worth mentioning that nanomachines such as eukaryotic cells [42] can be genetically modified to emit
different types of molecules.
the molecules respectively. Further, similar to [8], [20], the life
expectancy is modeled as an exponential distribution g(u) =
α exp(−αu), u > 0, with mean α1 , where α is referred to as
the degradation parameter.
The number of molecules received at the destination
nanomachine during the time slot [(j − 1)τ, jτ ] can be expressed as
Rsd [j] = Ssd [j] + Isd [j] + Nsd [j] + Csd [j].
(3)
The quantity Ssd [j] is the number of molecules received
in the current slot [(j − 1)τ, jτ ] and follows a binomial
distribution [45], [46] with parameters Q0 [j]x[j] and q0sd ,
i.e., Binomial(Q0 [j]x[j], q0sd ), where x[j] ∈ {0, 1} is the
symbol transmitted by the source nanomachine in the jth
time slot. The quantity Nsd [j] denotes MSI, i.e., noise arising
due to molecules received from other sources, which can be
modeled as a Gaussian distributed random variable with mean
µo and variance σo2 under the assumption that the number of
interfering sources is sufficiently large [47]. Also, note that
the noise Nsd [j] and the number of molecules Ssd [j], Isd [j]
received from the intended transmitter nanomachine are independent [20]. The term Csd [j] denotes the error in counting
the molecules at the destination nanomachine, also termed
as the “counting error”, and can be modeled as a Gaussian
distributed random variable with zero mean and variance σc2 [j]
that depends on the average number of molecules received as,
σc2 [j] = E{Rsd [j]} [20], [48]. The quantity Isd [j] is the ISI
arising in slot j due to transmissions in the previous j − 1
slots and is given as
Isd [j] = Isd [1] + Isd [2] + · · · + Isd [j − 1],
(4)
where Isd [i] ∼ Binomial(Q0 [j−i]x[j−i], qisd ), 1 ≤ i ≤ j−1,
denotes the number of stray molecules received from transmission in the (j − i)th slot. If the number of molecules
released by the source nanomachine is sufficiently large, the
binomial distribution for Ssd [j] can be approximated by the
Gaussian distribution2 with mean µsd [j] = Q0 [j]x[j]q0sd and
2
variance σsd
[j] = Q0 [j]x[j]q0sd (1 − q0sd ), i.e., Ssd [j] ∼
N (Q0 [j]x[j]q0sd , Q0 [j]x[j]q0sd (1 − q0sd )) [49]. Similarly, the
binomial distribution of Isd [i], 1 ≤ i ≤ j − 1 can be
approximated as
Isd [1] ∼ N (µI [1]=Q0 [j − 1]x[j − 1]q1sd ,
..
.
σI2 [1]=Q0 [j − 1]x[j − 1]q1sd (1 − q1sd )),
sd
Isd [j − 1] ∼ N (µI [j − 1]=Q0 [1]x[1]qj−1
,
sd
sd
σI2 [j − 1]=Q0 [1]x[1]qj−1
(1 − qj−1
)).
Further note that Ssd [j] and Isd [i], i = 1, 2, · · · , j − 1 are
independent since the molecules transmitted in different time
slots do not interfere with each other [20], [41].
2 This approximation is reasonable when Q [j]q sd > 5 and Q [j](1 −
0
0
0
q0sd ) > 5 [20].
B. Detection Performance Analysis
where the quantities γsd [j] and αsd [j] are defined as
Using the model in (3), the symbol detection problem at
the destination nanomachine can be formulated as the binary
hypothesis testing problem
αsd [j] =
H0 : Rsd [j] =Isd [1] + Isd [2] + · · · + Isd [j − 1]
+ Nsd [j] + Csd [j]
(5)
H1 : Rsd [j] =Ssd [j] + Isd [1] + Isd [2] + · · · + Isd [j−1]
+ Nsd [j] + Csd [j],
where the null and alternative hypotheses H0 and H1 correspond to the transmission of binary symbols 0 and 1 respectively during the jth time slot. The number of molecules
received Rsd [j] at the destination nanomachine corresponding
to the individual hypotheses are distributed as
2
[j])
H0 : Rsd [j] ∼ N (µsd,0 [j], σsd,0
(6)
2
H1 : Rsd [j] ∼ N (µsd,1 [j], σsd,1
[j]),
2
where the mean µsd,0 [j] and variance σsd,0
[j] under hypothesis H0 are derived as
µsd,0 [j] =µI [1] + µI [2] + · · · + µI [j − 1] + µo
=β
j−1
X
i=1
Q0 [j − i]qisd + µo ,
(7)
=
i=1
(8)
2
and the mean µsd,1 [j] and variance σsd,1
[j] under hypothesis
H1 are determined as
µsd,1 [j] =µsd [j] + µI [1] + µI [2] + · · · + µI [j − 1] + µo
2
σsd,1
[j]
Q0 [j − i]qisd + µo ,
(9)
i=1
2
2
=σsd [j]+σI [1]+σI2 [2]+ · · · +σI2 [j−1]+σo2 +σc2 [j]
j−1
X
[βQ0 [j−i]qisd (1−qisd )
=Q0 [j]q0sd (1 − q0sd ) +
i=1
+β(1 − β)(Q0 [j − i]qisd )2 ] + σo2 +µsd,1 [j]. (10)
Using the conditional PDFs p(Rsd [j]|H0 ) and p(Rsd [j]|H1 )
given in (6), the following result presents the LRT-based decision rule at the destination nanomachine for symbol detection.
Theorem 1: The LRT-based optimal decision rule for deciding 0 and 1 at the destination nanomachine corresponding
to source nanomachine transmission during the jth time slot
is obtained as
H1
′
T (Rsd [j]) = Rsd [j] ≷ γsd
[j],
(14)
Proof: The optimal log likelihood ratio test (LLRT) at the
destination nanomachine can be obtained as
p(Rsd [j]|H1 ) H1
1−β
Λ(Rsd [j]) = ln
.
(15)
≷ ln
p(Rsd [j]|H0 ) H0
β
Substituting the Gaussian PDFs p(Rsd [j]|H1 ) and
p(Rsd [j]|H0 ) from (6), the test statistic Λ(Rsd [j]) can
be obtained as
"s
#
2
σsd,0
[j]
1
Λ(Rsd [j]) = ln
+ 2
2
2
σsd,1
[j]
2σsd,0 [j]σsd,1
[j]
2
2
× (Rsd [j]−µsd,0 [j])2 σsd,1
[j]−(Rsd [j]−µsd,1 [j])2 σsd,0
[j] .
{z
}
|
The expression for f (Rsd [j]) given above can be further
simplified as
× (Q0 [j − i]qisd )2 ] + σo2 +µsd,0 [j],
=Q0 [j]q0sd + β
2
2
µ2sd,1 [j]σsd,0
[j] − µ2sd,0 [j]σsd,1
[j]
.
2
2
σsd,1 [j] − σsd,0 [j]
(16)
[βQ0 [j−i]qisd (1−qisd )+β(1 − β)
j−1
X
+
,f (Rsd [j])
2
σsd,0
[j] =σI2 [1] + σI2 [2] + · · · + σI2 [j − 1] + σo2 + σc2 [j]
j−1
X
2
2
[j]
[j] − µsd,0 [j]σsd,1
µsd,1 [j]σsd,0
,
(13)
2 [j] − σ 2 [j]
σsd,1
sd,0
s
#
"
2
2
2
[j]
2σsd,1
[j]σsd,0
[j]
(1−β) σsd,1
+(αsd [j])2
γsd [j] = 2
2 [j] ln
2
σsd,1 [j]−σsd,0
β
σsd,0
[j]
(11)
H0
′
where γsd
[j] is the optimal decision threshold and is given as
p
′
(12)
γsd
[j] = γsd [j] − αsd [j],
2
2
2
f (Rsd [j]) =Rsd
[j](σsd,1
[j] − σsd,0
[j])
2
2
+ 2Rsd [j](µsd,1 [j]σsd,0
[j] − µsd,0 [j]σsd,1
[j])
2
2
+ (µ2sd,0 [j]σsd,1
[j] − µ2sd,1 [j]σsd,0
[j])
2
2
=(σsd,1
[j]−σsd,0
[j])(Rsd [j]+αsd [j])2
2
2
(µsd,1 [j]σsd,0
[j]−µsd,0 [j]σsd,1
[j])2
−
2 [j] − σ 2 [j]
σsd,1
sd,0
2
2
+ (µ2sd,0 [j]σsd,1
[j] − µ2sd,1 [j]σsd,0
[j]),
(17)
where αsd [j] is defined in (13). Substituting the above equation
for f (Rsd [j]) in (16) and subsequently grouping the terms independent of the received molecules Rsd [j] with the threshold,
the test reduces to
H1
(Rsd [j] + αsd [j])2 ≷ γsd [j],
(18)
H0
where γsd [j] is defined in (14). Further, taking the square root
of both sides with γsd [j] ≥ 0, the above expression can be
simplified to yield the optimal test given in (11).
The detection performance at the destination nanomachine for
the test in (11) is obtained next.
Theorem 2: The average probabilities of detection (PD )
and false alarm (PF A ) at the destination nanomachine corresponding to the transmission by the source nanomachine in
slots 1 to k in the diffusion based molecular communication
nano-network are given as
k
1X d
P [j]
k j=1 D
′
k
γsd [j] − µsd,1 [j]
1X
Q
,
=
k j=1
σsd,1 [j]
PD =
k
1X d
P [j]
k j=1 F A
′
k
1X
γsd [j] − µsd,0 [j]
=
Q
,
k j=1
σsd,0 [j]
(19)
PF A =
(20)
d
where PD
[j] and PFd A [j] denote the probabilities of detection
and false alarm, respectively, at the destination nanomachine
′
in the jth time slot. The quantity γsd
[j] is defined in (12) and
the function Q(·) is the tail probability of the standard normal
random variable.
d
Proof: The probabilities of detection (PD
[j]) and false
d
alarm (PF A [j]) at the destination nanomachine in the jth time
slot can be derived using the decision rule (11) as
d
′
PD
[j] =Pr(T (Rsd [j]) > γsd
[j]|H1 )
′
=Pr(Rsd [j] > γsd [j]|H1 ),
PFd A [j]
(21)
′
γsd
[j]|H0 )
=Pr(T (Rsd [j]) >
′
[j]|H0 ),
=Pr(Rsd [j] > γsd
(22)
where the number of received molecules Rsd [j] is Gaussian
distributed (6) under hypotheses H0 and H1 respectively.
Subtracting their respective means followed by division by
R [j]−µsd,1 [j]
R [j]−µsd,0 [j]
the standard deviations, i.e., sd σsd,1 [j]
and sd σsd,0 [j]
yields standard normal random variables under hypotheses H1
d
and H0 respectively. Subsequently, the expressions for PD
[j]
d
and PF A [j] in (19) and (20) respectively follow employing the
R ∞ x2
definition of the Q(·) function Q(α)= √12π α e− 2 dx [50].
C. Probability of Error Analysis
The end-to-end probability of error for direct communication between the source and destination nanomachines follows
as described in the result below.
Theorem 3: The average probability of error (Pe ) for slots
1 to k at the destination nanomachine in the diffusion based
cooperative molecular nano-network in which the molecules
are in Brownian motion with drift, is given as
′
k
1X
γsd [j] − µsd,1 [j]
Pe =
β 1−Q
k j=1
σsd,1 [j]
′
γsd [j] − µsd,0 [j]
.
(23)
+(1 − β)Q
σsd,0 [j]
Proof: The probability of error Ped [j] in the jth time slot
is defined as [50]
Ped [j] =Pr(decide H0 , H1 true) + Pr(decide H1 , H0 true)
d
=(1 − PD
[j])Pr(H1 ) + PFd A [j]Pr(H0 ),
(24)
where the prior probabilities of the hypotheses Pr(H1 ) and
d
[j] and
Pr(H0 ) are β and 1 − β respectively. The quantities PD
d
PF A [j] denote the probabilities of detection and false alarm at
the destination nanomachine in the jth time slot and are given
in (19) and (20) respectively. Substituting these expressions,
the average probability of error for slots 1 to k is obtained as
stated in (23).
D. Capacity Analysis
This section presents the capacity analysis for the above system considering the impact of ISI, counting errors, and MSI.
Let the discrete random variables X[j] and Y [j] represent the
transmitted and received symbol, respectively, in the jth slot.
The mutual information I(X[j], Y [j]) is
I(X[j], Y [j])
X
=
X
Pr(x[j], y[j]) log2
x[j]∈{0,1} y[j]∈{0,1}
Pr(y[j]|x[j])
, (25)
Pr(y[j])
where the joint and marginal probabilities Pr(x[j], y[j]) and
Pr(y[j]) respectively can be written in terms of the conditional
probability Pr(y[j]|x[j]) and marginal probability Pr(x[j])
as P
Pr(x[j], y[j]) = Pr(y[j]|x[j])Pr(x[j]) and Pr(y[j]) =
Pr(y[j]|x[j])Pr(x[j]) respectively. Substituting the
x[j]∈{0,1}
above expressions for Pr(x[j], y[j]) and Pr(y[j]) in terms of
the conditional probability Pr(y[j]|x[j]) and marginal probability Pr(x[j]) in (25), the mutual information I(X[j], Y [j])
can be obtained as (26), where the conditional probabilities
d
Pr(y[j]∈{0, 1}|x[j]∈{0, 1}) can be written in terms of PD
[j]
and PFd A [j] as
Pr(y[j] = 0|x[j] = 0) =1 − PFd A [j]
′
γsd [j] − µsd,0 [j]
,
=1 − Q
σsd,0 [j]
Pr(y[j] = 1|x[j] = 0) =PFd A [j]
′
γsd [j] − µsd,0 [j]
,
=Q
σsd,0 [j]
d
Pr(y[j] = 0|x[j] = 1) =1 − PD
[j]
′
γsd [j] − µsd,1 [j]
=1 − Q
,
σsd,1 [j]
d
Pr(y[j] = 1|x[j] = 1) =PD
[j]
′
γsd [j] − µsd,1 [j]
=Q
.
σsd,1 [j]
The capacity C[k] of the direct source-destination channel, as
k approaches ∞, can be obtained by maximizing the average
mutual information I(X[j], Y [j]) for slots 1 to k as [8]
C[k] = max
β
k
1X
I(X[j], Y [j]) bits/slot.
k j=1
(27)
III. D UAL - HOP C OMMUNICATION BETWEEN THE S OURCE
AND D ESTINATION NANOMACHINES
A. System Model
Consider now a dual-hop diffusive molecular communication system inside a blood vessel wherein the intermediate
I(X[j], Y [j]) =(1−β) Pr(y[j] = 0|x[j] = 0) log2
Pr(y[j] = 0|x[j] = 0)
(1−β)Pr(y[j] = 0|x[j] = 0)+βPr(y[j] = 0|x[j] = 1)
Pr(y[j] = 1|x[j] = 0)
+Pr(y[j] = 1|x[j] = 0) log2
(1 − β)Pr(y[j] = 1|x[j] = 0) + βPr(y[j] = 1|x[j] = 1)
Pr(y[j] = 0|x[j] = 1)
+ β Pr(y[j] = 0|x[j] = 1) log2
(1 − β)Pr(y[j] = 0|x[j] = 0) + βPr(y[j] = 0|x[j] = 1)
Pr(y[j] = 1|x[j] = 1)
,
+Pr(y[j] = 1|x[j] = 1) log2
(1 − β)Pr(y[j] = 1|x[j] = 0) + βPr(y[j] = 1|x[j] = 1)
nanomachine cooperates with the source nanomachine to relay
its information to the destination nanomachine. This system
model is similar to the one considered in the recent work [41]
where the source, cooperative, and destination nanomachines
lie on a straight line and the direction of molecular drift is
from the source to cooperative and cooperative to destination
nanomachines. The cooperative nanomachine is assumed to
be located at a distance of dsr from the source nanomachine
whereas the destination nanomachine is located at a distance of
drd from the cooperative nanomachine. Similar to the previous
direct communication scenario, the channel is divided into
time slots of duration τ as shown in Figs. 2 and 3, where
the jth and (j + 1)th slots are defined as the time intervals
[(j − 1)τ, jτ ] and [jτ, (j + 1)τ ] respectively.
In this system, the end-to-end communication between the
source and the destination nanomachines occurs in two timeslots. At the beginning of the jth time slot, the source nanomachine emits either Q0 [j] number of type 0 molecules in the
propagation medium for information symbol 1 generated with
a prior probability β or remains silent for information symbol
0. The intermediate nanomachine decodes the symbol received
from the source nanomachine followed by retransmission of
Q1 [j + 1] number of type 1 molecules or remaining silent in
the subsequent (j +1)th time slot to indicate decoded symbols
1, 0 respectively. As described in [4], nanomachines such as
eukaryotic cells can be genetically modified to emit different
type of molecules. Further, since the source and cooperative
nanomachines transmit different types of molecules, the cooperative nanomachine can operate in full duplex mode.
(26)
Diffusion based molecular communication between the
cooperative and destination nanomachines over time-slotted channel,
where [jτ, (j + 1)τ ] is the current slot.
Fig. 3.
pressed as
Rsr [j] = Ssr [j] + Isr [j] + Nsr [j] + Csr [j],
(28)
where Ssr [j] ∼ N (Q0 [j]x[j]q0sr , Q0 [j]x[j]q0sr (1 − q0sr )) is
the number of type 0 molecules received at the cooperative
nanomachine from the current slot [(j − 1)τ, jτ ], with x[j] ∈
{0, 1} denoting the symbol transmitted by the source nanomachine in the jth time slot. The quantity Nsr [j] ∼ N (µo , σo2 ) is
the MSI and Csr [j] ∼ N (0, E{Rsr [j]}) denotes the counting
errors at the cooperative nanomachine. The quantity Isr [j] that
represents the ISI at the cooperative nanomachine in slot j is
Isr [j] = Isr [1] + Isr [2] + · · · + Isr [j − 1],
(29)
where Isr [i] ∼ N (Q0 [j −i]x[j −i]qisr , Q0 [j −i]x[j −i]qisr (1−
qisr )), 1 ≤ i ≤ j − 1 denotes the number of stray molecules
received from the (j − i)th slot.
The cooperative nanomachine decodes the symbol using
the number of received molecules Rsr [j] followed by the
retransmission of the decoded symbol x
b[j]. The number of
type 1 molecules received at the destination nanomachine
during time slot [jτ, (j + 1)τ ] can be expressed as
Rrd [j + 1] =Srd [j + 1] + Ird [j + 1] + Nrd [j + 1]
+ Crd [j + 1],
(30)
Fig. 2. Diffusion based molecular communication between the source
and cooperative nanomachines over time-slotted channel, where [(j −
1)τ, jτ ] is the current slot.
xy
Let qj−i
, xy∈{sr, rd} denote the probability that a molecule
transmitted by node x in slot i ∈ {1, 2, · · · , j} arrives at
node y during time slot j. These probabilities can be obtained
using (1) and setting the source-cooperative and cooperativedestination link distances to dsr and drd respectively. The
number of type 0 molecules received at the cooperative
nanomachine during the time slot [(j − 1)τ, jτ ] can be ex-
where Srd [j + 1] ∼ N (Q1 [j + 1]b
x[j]q0rd , Q1 [j + 1]b
x[j]q0rd (1 −
rd
q0 )) is the number of type 1 molecules received at the
destination nanomachine during the current slot [jτ, (j + 1)τ ].
The quantities Nrd [j + 1] ∼ N (µo , σo2 ) and Crd [j + 1] ∼
N (0, E{Rrd [j + 1]}) denote the MSI and counting errors,
respectively, at the destination nanomachine. Similarly, the ISI
Ird [j + 1] at the destination nanomachine is
Ird [j + 1] = Ird [2] + Ird [3] + · · · + Ird [j],
(31)
rd
where Ird [i] ∼ N (Q1 [j − i + 2]b
x[j − i + 1]qi−1
, Q1 [j − i +
rd
rd
2]b
x[j − i + 1]qi−1 (1 − qi−1 )), 2 ≤ i ≤ j denotes the number
of stray molecules received from the previous (j−i+2)th slot.
B. Detection Performance Analysis
On lines similar to the proof of Theorem 1, one can obtain
the optimal decision rule at the cooperative nanomachine for
the jth time slot as
H1
T (Rsr [j]) = Rsr [j] ≷
H0
2
[j + 1] under
where the mean µrd,0 [j + 1] and variance σrd,0
hypothesis H0 are derived as,
µrd,0 [j + 1] =µI [2] + µI [3] + · · · + µI [j] + µo
=β
′
γsr
[j],
(32)
2
σrd,0
[j + 1]
′
where the optimal decision threshold γsr
[j] is given in (33).
2
The mean µsr,0 [j] and variance σsr,0 [j] under hypothesis H0
are derived using (28) as
µsr,0 [j] =β
j−1
X
i=1
2
σsr,0
[j] =
j−1
X
i=1
j
X
rd
Q1 [j − i + 2]qi−1
+ µo ,
i=2
=σI2 [2] +
j
X
=
i=2
(42)
σI2 [3] + · · · + σI2 [j] + σo2 + σc2 [j + 1]
rd
rd
[βQ1 [j−i+2]qi−1
(1 − qi−1
)+β(1−β)
rd 2
× (Q1 [j−i+2]qi−1
) ]+σo2 +µrd,0 [j+1], (43)
Q0 [j − i]qisr + µo ,
(34)
[βQ0 [j−i]qisr (1−qisr )+β(1 − β)
µrd,1 [j + 1] =µrd [j + 1] + µI [2] + µI [3] + · · · + µI [j] + µo
× (Q0 [j − i]qisr )2 ] + σo2 +µsr,0 [j],
j−1
X
i=1
Q0 [j − i]qisr + µo ,
2
σsr,1
[j] =Q0 [j]q0sr (1 − q0sr ) +
+β(1 − β)(Q0 [j −
j−1
X
=Q1 [j+1]q0rd +β
(35)
2
while the mean µsr,1 [j] and variance σsr,1
[j] under hypothesis
H1 are
µsr,1 [j] =Q0 [j]q0sr + β
2
while the mean µrd,1 [j + 1] and variance σrd,1
[j + 1] under
hypothesis H1 are derived as,
2
σrd,1
[j
+
(37)
(1)
Using the test in (32), the probabilities of detection (PD [j])
(1)
and false alarm (PF A [j]) at the cooperative nanomachine for
the jth time slot can be derived as,
′
γsr [j] − µsr,1 [j]
(1)
PD [j] = Q
,
(38)
σsr,1 [j]
′
γsr [j] − µsr,0 [j]
(1)
PF A [j] = Q
,
(39)
σsr,0 [j]
′
where γsr
[j] is defined in (33). Further, similar to the direct
communication scenario detailed in Section II-A, the symbol
detection problem at the destination nanomachine for transmission by the cooperative relay in the (j + 1)th time slot can
be formulated as the binary hypothesis testing problem
H0 : Rrd [j+1]=Ird [2] + Ird [3] + · · · + Ird [j]
+ Nrd [j + 1] + Crd [j + 1]
H1 : Rrd [j+1]=Srd [j+1]+Ird [2]+Ird [3]+ · · · +Ird [j]
+ Nrd [j + 1] + Crd [j + 1].
σI2 [3] + · · · + σI2 [j]
j
X
rd
[βQ1 [j−i+2]qi−1
rd
rd 2
× (1−qi−1
)+β(1−β)(Q1 [j−i+2]qi−1
) ]
+ σo2 + µrd,1 [j + 1].
(45)
The optimal decision rule at the destination nanomachine for
the dual-hop molecular communication network is obtained
next.
Theorem 4: The optimal detector at the destination
nanomachine corresponding to transmission by the cooperative
nanomachine in the (j + 1)th time slot is obtained as,
H1
′
T (Rrd [j + 1]) = Rrd [j + 1] ≷ γrd
[j + 1],
(46)
H0
′
with the optimal decision threshold γrd
[j + 1] given as,
p
′
γrd
[j + 1] = γrd [j + 1] − αrd [j + 1],
(47)
where the quantities αrd [j + 1] and γrd [j + 1] are defined as,
αrd [j + 1]
2
2
µrd,1 [j+1]σrd,0
[j+1]−µrd,0 [j+1]σrd,1
[j+1]
, (48)
=
2
2
σrd,1 [j+1]−σrd,0 [j+1]
(40)
H0 and H1 denote the null and alternative hypotheses corresponding to the transmission of decoded symbols x
b[j] = 0 and
x
b[j] = 1 respectively by the cooperative nanomachine during
the (j +1)th time slot. The quantities Rrd [j +1] corresponding
to the individual hypotheses in (40) are distributed as
2
H1 : Rrd [j + 1] ∼ N (µrd,1 [j + 1], σrd,1
[j + 1]),
i=2
2
2
=σrd [j + 1] + σI [2] +
+ σo2 + σc2 [j + 1]
i=2
σo2 +µsr,1 [j].
2
H0 : Rrd [j + 1] ∼ N (µrd,0 [j + 1], σrd,0
[j + 1])
rd
Q1 [j−i+2]qi−1
+µo , (44)
=Q1 [j+1]q0rd (1−q0rd )+
(36)
[βQ0 [j−i]qisr (1−qisr )
i=1
i]qisr )2 ]
+ 1]
j
X
(41)
γrd [j + 1]
(
)#
"s
(1)
2
σrd,1
[j+1] (1−β)(1−PF(1)
A [j])−β(1−PD [j])
= ln
2
(1)
(1)
σrd,0
[j+1]
βP [j]−(1−β)P [j]
D
FA
2
2
2σrd,1
[j + 1]σrd,0
[j + 1]
× 2
+(αrd [j + 1])2
2
σrd,1 [j + 1]−σrd,0
[j + 1]
+
2
2
µ2rd,1 [j + 1]σrd,0
[j + 1] − µ2rd,0 [j + 1]σrd,1
[j + 1]
.
2
2
σrd,1 [j + 1] − σrd,0 [j + 1]
(49)
v
s
#
"
u
2 [j]
u 2σ 2 [j]σ 2 [j]
(1−β) σsr,1
sr,1
sr,0
t
′
γsr [j] =
2 [j]−σ 2 [j] ln
2 [j] +
σsr,1
β
σsr,0
sr,0
−
2 [j]−µ
2
µsr,1 [j]σsr,0
sr,0 [j]σsr,1 [j]
2 [j]−σ 2 [j]
σsr,1
sr,0
2
2
µsr,1 [j]σsr,0
[j] − µsr,0 [j]σsr,1
[j]
,
2
2
σsr,1 [j] − σsr,0 [j]
(1)
p(Rrd [j + 1]|H1 ) H1 1 − β
.
≷
p(Rrd [j + 1]|H0 ) H0 β
(50)
The likelihood ratio L(Rrd [j + 1]) is evaluated in (51), where
the PDFs p(Rrd [j + 1]|b
x[j] = 1) and p(Rrd [j + 1]|b
x[j] = 0)
are
p(Rrd [j + 1]|b
x[j] = 1)
2
(Rrd [j+1]−µrd,1 [j+1])
exp −
=q
2
2σrd,1
[j+1]
2
2πσrd,1 [j+1]
1
!
, (52)
p(Rrd [j + 1]|b
x[j] = 0)
!
(Rrd [j+1]−µrd,0 [j+1])2
=q
. (53)
exp −
2
2σrd,0
[j+1]
2
2πσrd,0
[j+1]
1
Substituting the above expressions followed by simplification
as shown in Appendix A, the resulting test can be expressed as
given in (54). The expression for the test statistic f (Rrd [j+1])
in (54) can be further simplified as shown in (55). Substituting
expression for f (Rrd [j + 1]) from (55) in (54) and merging
the terms independent of the number of received molecules
Rrd [j + 1] with the threshold in (55), the test reduces to
2
H1
(Rrd [j + 1] + αrd [j + 1]) ≷ γrd,j+1 ,
(56)
H0
where αrd [j + 1] and γrd [j + 1] are defined in (48) and
(49) respectively. Taking the square root of both sides, where
γrd [j + 1] ≥ 0, yields the optimal test at the destination
nanomachine as given in (46). It can be noted that the
decision rule obtained above incorporates also the detection
performance of the intermediate cooperative nanomachine.
The result below determines the resulting detection performance at the destination nanomachine in the dual-hop network.
Theorem 5: The average probabilities of detection PD and
false alarm PF A at the destination nanomachine corresponding
to transmission by the source nanomachine in slots 1 to k for
the dual-hop diffusive molecular communication system are
given as
k
1X d
P [j + 1],
PD =
k j=1 D
PF A =
k
1X d
P [j + 1],
k j=1 F A
+
2 [j]−µ2 [j]σ 2 [j]
µ2sr,1 [j]σsr,0
sr,0
sr,1
2 [j] − σ 2 [j]
σsr,1
sr,0
(33)
The expressions for the probabilities of detection PD [j]
(1)
and false alarm PF A [j] for cooperative nanomachine are as
obtained in (38) and (39) respectively.
Proof: The optimal LRT L(Rrd [j + 1]) at the destination
nanomachine is
L(Rrd [j + 1]) =
!2
d
[j + 1] and false alarm
where the probabilities of detection PD
d
PF A [j + 1] at the destination in the (j + 1)th slot are given
as,
′
γrd [j + 1] − µrd,1 [j + 1]
(1)
d
PD [j]
PD
[j + 1] = Q
σrd,1 [j + 1]
′
γrd [j + 1] − µrd,0 [j + 1]
(1)
+Q
(1 − PD [j]), (59)
σrd,0 [j + 1]
′
γrd [j + 1] − µrd,1 [j + 1]
(1)
d
PF A [j + 1] = Q
PF A [j]
σrd,1 [j + 1]
′
γrd [j + 1] − µrd,0 [j + 1]
(1)
+Q
(1 − PF A [j]). (60)
σrd,0 [j + 1]
d
Proof: The probability of detection PD
[j + 1] at the
destination nanomachine can be derived using the decision
rule in (46) as
d
PD
[j + 1]
′
=Pr(T (Rrd [j + 1]) > γrd
[j + 1]|H1 )
′
=Pr(Rrd [j + 1] > γrd [j + 1]|b
x[j] = 1)Pr(b
x[j] = 1|H1 )
′
[j + 1]|b
x[j] = 0)Pr(b
x[j] = 0|H1 )
+ Pr(Rrd [j + 1] > γrd
(1)
′
=Pr(Rrd [j + 1] > γrd
[j + 1]|b
x[j] = 1)PD [j]
(1)
′
+ Pr(Rrd [j + 1] > γrd
[j + 1]|b
x[j] = 0)(1−PD [j]), (61)
Similarly, the probability of false alarm PFd A [j + 1] can be
derived as,
PFd A [j + 1]
′
[j + 1]|H0 )
=Pr(T (Rrd [j + 1]) > γsd
′
=Pr(Rrd [j + 1] > γrd
[j + 1]|b
x[j] = 1)Pr(b
x[j] = 1|H0 )
′
x[j]=0)Pr(b
x[j]=0|H0 )
+ Pr(Rrd [j + 1] > γrd [j + 1]|b
(1)
′
=Pr(Rrd [j + 1] > γrd
[j + 1]|b
x[j] = 1)PF A [j]
(1)
′
+ Pr(Rrd [j + 1]>γrd
[j+1]|b
x[j]=0)(1−PF A [j]). (62)
′
x[j] = 1) =
Further,
rd [j + 1] > γrd [j + 1]|b
′ substituting Pr(R
γrd [j+1]−µrd,1 [j+1]
′
and
Pr(R
[j
+1]
>
γ
[j
x[j] =
Q
rd
rd +1]|b
[j+1]
σrd,1
′
γrd
[j+1]−µrd,0 [j+1]
in (61) and (62), one can obtain
0) = Q
σrd,0 [j+1]
d
the final expressions for PD
[j + 1] and PFd A [j + 1] given in
(59) and (60) respectively.
C. Probability of Error Analysis
(57)
(58)
The end-to-end probability of error for dual-hop communication between the source and destination nanomachines is
given by the result below.
Theorem 6: The average probability of error (Pe ) at the
destination nanomachine corresponding to the transmission
L(Rrd [j + 1]) =
=
p(Rrd [j + 1]|b
x[j] = 1)Pr(b
x[j] = 1|H1 ) + p(Rrd [j + 1]|b
x[j] = 0)Pr(b
x[j] = 0|H1 )
p(Rrd [j + 1]|b
x[j] = 1)Pr(b
x[j] = 1|H0 ) + p(Rrd [j + 1]|b
x[j] = 0)Pr(b
x[j] = 0|H0 )
(1)
(1)
(1)
(1)
p(Rrd [j + 1]|b
x[j] = 1)PD [j] + p(Rrd [j + 1]|b
x[j] = 0)(1 − PD [j])
x[j] = 0)(1 − PF A [j])
p(Rrd [j + 1]|b
x[j] = 1)PF A [j] + p(Rrd [j + 1]|b
,
(51)
,f (Rrd [j+1])
z
}|
{
1
2 2
2 2
(R
[j+1]−µ
[j+1])
σ
[j+1]−(R
[j
+
1]−µ
[j+1])
σ
[j+1]
rd
rd,0
rd
rd,1
rd,1
rd,0
2
2
2σrd,0
[j+1]σrd,1
[j+1]
(
)#
"s
(1)
2
H1
σrd,1
[j + 1] (1−β)(1−PF(1)
A [j])−β(1−PD [j])
.
≷ ln
2
(1)
(1)
σrd,0
[j + 1]
H0
βP [j]−(1−β)P [j]
D
(54)
FA
2
2
2
2
2
f (Rrd [j + 1]) =Rrd
[j + 1](σrd,1
[j + 1] − σrd,0
[j + 1]) + 2Rrd [j + 1](µrd,1 [j + 1]σrd,0
[j + 1] − µrd,0 [j + 1]σrd,1
[j + 1])
2
2
+ (µ2rd,0 [j + 1]σrd,1
[j + 1] − µ2rd,1 [j + 1]σrd,0
[j + 1])
#2
"
2
2
µrd,1 [j + 1]σrd,0
[j + 1] − µrd,0 [j + 1]σrd,1
[j + 1]
2
2
=(σrd,1 [j + 1] − σrd,0 [j + 1]) Rrd [j + 1] +
2
2
σrd,1
[j + 1] − σrd,0
[j + 1]
−
2
2
(µrd,1 [j+1]σrd,0
[j+1]−µrd,0 [j+1]σrd,1
[j+1])2
2
2
+(µrd,1 [j+1]σrd,0
[j + 1]−µrd,0 [j+1]σrd,1
[j+1]). (55)
2
2
σrd,1
[j + 1] − σrd,0
[j + 1]
by the source nanomachine in slots 1 to k for the dual-hop
diffusive molecular communication system is given as
Pe =
k
i
1 Xh
d
β 1 − PD
[j + 1] + (1−β)PFd A [j + 1] , (63)
k j=1
d
where the expressions for PD
[j + 1] and PFd A [j + 1] are given
in (59) and (60) respectively.
Proof: This follows on lines similar to proof of Theorem
3.
D. Capacity Analysis
The mutual information I(X[j], Y [j + 1]) can be expressed
as
I(X[j],Y [j + 1])
X
=
X
1
The factor k+1
arises in the above expression in contrast to
1
in
(27)
due
to
the fact that k + 1 time-slots are required to
k
transmit k bits from the source to destination in the dual-hop
scheme.
IV. M ULTI - HOP C OMMUNICATION BETWEEN THE S OURCE
AND D ESTINATION NANOMACHINES
A. System Model
Consider now a general multi-hop molecular communication system that consists of a source, destination and N (≥ 2)
intermediate cooperative nanomachines as shown in Fig. 4.
In this system, the communication between the source and
Pr(y[j + 1]|x[j])Pr(x[j])
x[j]∈{0,1} y[j+1]∈{0,1}
× log2
P
Pr(y[j + 1]|x[j])
,
Pr(y[j + 1]|x[j])Pr(x[j])
(64)
x[j]∈{0,1}
where Pr(x[j]=1)=β, Pr(x[j]=0)=1−β, and Pr(y[j + 1] ∈
d
{0, 1}|x[j] ∈ {0, 1}) can be determined in terms of (PD
[j+1])
d
and (PF A [j + 1]) as
Pr(y[j+1]=0|x[j]=0) =1 − PFd A [j + 1],
Pr(y[j+1]=1|x[j]=0) =PFd A [j + 1],
d
Pr(y[j+1]=0|x[j]=1) =1 − PD
[j + 1],
d
Pr(y[j+1]=1|x[j]=1) =PD
[j + 1].
The capacity C[k] of the dual-hop channel, as k approaches
∞, is now obtained by maximizing the mutual information
I(X[j], Y [j + 1])
k
C[k] = max
β
X
1
I(X[j], Y [j + 1]) bits/slot. (65)
(k + 1) j=1
Schematic diagram of multi-hop diffusion based molecular
communication system with N ≥ 2 cooperative nanomachines.
Fig. 4.
destination nanomachines occurs in a hop-by-hop fashion with
N + 1 different types of molecules. Therefore, the end-to-end
communication requires a total of N + 1 time-slots. At the
beginning of the jth time slot, the source nanomachine either
emits Q0 molecules of type 0 in the propagation medium for
information symbol 1 generated with a prior probability β or
remains silent for information symbol 0. In the subsequent
time slots, the intermediate full-duplex cooperative nanomachines decode the symbol using the number of molecules
received from the previous hop nanomachine as shown in Fig.
4 and subsequently retransmit the decoded symbol to the next
nanomachine. The destination nanomachine finally decodes
the symbol using the number of molecules received from N th
cooperative nanomachine (RN ) in the (j + N )th time slot.
Similar to the dual-hop communication scenario, the number
of molecules received at the cooperative nanomachine R1
during the time slot [(j − 1)τ, jτ ] can be expressed as,
Rsr [j] = Ssr [j] + Isr [j] + Nsr [j] + Csr [j],
(66)
where the quantities Ssr [j], Isr [j], Nsr [j], and Csr [j] are
similar to the ones considered in (28). The expressions for the
(1)
(1)
probabilities of detection PD [j] and false alarm PF A [j] at
R1 in the multi-hop system are identical to the ones obtained
for the dual-hop system in (38) and (39) respectively. This
is due to the fact that the system model for the received
molecules at the cooperative nanomachine corresponding to
the transmission by the source nanomachine in the jth slot is
identical for both the dual-hop and multi-hop systems.
In the subsequent time slots [(j + n − 1)τ, (j + n)τ ], n =
1, · · · , N − 1, the number of molecules received at the cooperative nanomachine Rn+1 corresponding to the transmission
of x
b[j + n − 1] ∈ {0, 1} by the previous hop nanomachine Rn
can be expressed as,
Rrr [j+n]=Srr [j+n]+Irr [j+n]+Nrr [j+n]+Crr [j+n], (67)
x[j + n − 1]q0rr , Qn [j +
where Srr [j + n] ∼ N (Qn [j + n]b
rr
rr
n]b
x[j + n − 1]q0 (1 − q0 )) is the number of type n molecules
received in the current slot [(j + n − 1)τ, (j + n)τ ]. The
quantity Nrr [j + n] ∼ N (µo , σo2 ) is the MSI arising due
to molecules received from the other sources. The term
Crr [j +n] ∼ N (0, E{Rrr [j +n]}) denotes the counting errors
at the cooperative nanomachine Rn+1 . The quantity Irr [j +n]
represents the ISI at the cooperative nanomachine and is given
as,
B. Detection Performance Analysis
The binary hypothesis testing problem for symbol detection
at the destination nanomachine corresponding to the transmission of cooperative nanomachine RN in the (j + N )th time
slot can again be formulated as
H0 :Rrd [j+N ]=Ird [N +1]+Ird [N +2]+ · · · +Ird [j+N −1]
+Nrd [j+N ]+Crd [j+N ]
H1 :Rrd [j+N ]=Srd [j+N ] + Ird [N +1]+Ird [N +2]+ · · ·
+Ird [j+N −1]+Nrd [j+N ]+Crd [j+N ],
(71)
where H0 and H1 denote the null and alternative hypotheses
corresponding to the transmission of the decoded symbol x
b[j+
N − 1] = 0 and x
b[j + N − 1] = 1 respectively by the RN in
the (j + N )th time slot. The number of molecules received at
the destination nanomachine corresponding to the individual
hypotheses in (71) are Gaussian distributed as
2
[j + N ])
H0 : Rrd [j + N ] ∼ N (µrd,0 [j + N ], σrd,0
2
H1 : Rrd [j + N ] ∼ N (µrd,1 [j + N ], σrd,1
[j + N ]),
2
where the mean µrd,0 [j + N ] and variance σrd,0
[j + N ] under
hypothesis H0 are
µrd,0 [j + N ] =µI [N +1]+µI [N +2]+ · · · +µI [j+N −1]+µo
=β
2
σrd,0
[j + N ]
rd
QN [j+2N −i]qi−N
+ µo ,
i=N +1
2
=σI [N + 1] +
+ σo2 + σc2 [j
j+N
X−1
i=N +1
(73)
σI2 [N + 2]+ · · · +σI2 [j + N − 1]
+ N]
rd
rd
[βQN [j+2N −i]qi−N
(1 − qi−N
)+β
rd
× (1 − β)(QN [j+2N −i]qi−N
)2 ]
+ σo2 + µrd,0 [j + N ],
(74)
2
and the mean µrd,1 [j + 1] and variance σrd,1
[j + 1] under
hypothesis H1 are
µrd,1 [j + N ] =µrd [j + N ] + µI [N + 1] + µI [N + 2]
+ · · · + µI [j + N − 1] + µo
Rrd [j+N ]=Srd [j+N ]+Ird [j+N ]+Nrd [j+N ]
+Crd [j+N ],
j+N
X−1
=
Irr [j+n] = Irr [n+1]+Irr [n+2]+ · · · +Irr [j+n−1], (68)
rr
where Irr [i] ∼ N (Qn [j + 2n− i]b
x[j + 2n− i − 1]qi−n
, Qn [j +
rr
rr
2n − i]b
x[j + 2n − i − 1]qi−n (1 − qi−n )), n + 1 ≤ i ≤ j +
n − 1 denotes the number of stray molecules received from
the previous (j + 2n − i)th slot.
The number of molecules received at the destination
nanomachine during the time slot [(j + N − 1)τ, (j + N )τ ]
can be expressed as,
(72)
(69)
=QN [j+N ]q0rd
− 1]q0rd , QN [j
where Srd [j + N ] ∼ N (QN [j + N ]b
x[j + N
+
N ]b
x[j+N −1]q0rd(1−q0rd )) is the number of type N molecules
received at the destination nanomachine in the current slot
[(j+N −1)τ, (j+N )τ ]. The quantity Nrd [j+N ] ∼ N (µo , σo2 )
is the MSI. The term Crd [j + N ] ∼ N (0, E{Rrd [j + N ]})
denotes the counting errors at the destination nanomachine.
The quantity Ird [j + N ] represents the ISI at the destination
nanomachine that is given as,
Ird [j+N ]=Ird [N +1]+Ird [N +2]+ · · · +Ird [j+N −1], (70)
where Ird [i] ∼ N (QN [j + 2N − i]b
x[j + 2N − i −
rd
rd
rd
)),
(1 − qi−N
, QN [j + 2N − i]b
x[j + 2N − i − 1]qi−N
1]qi−N
N + 1 ≤ i ≤ j + N − 1 denotes the number of stray molecules
received from transmission during the previous (j + 2N − i)th
slot.
+β
j+N
X−1
i=N +1
+ µo ,
2
σrd,1
[j
+ N]
2
=σrd
[j
+ N] +
=QN [j+N ]q0rd (1 −
j+N
X−1
i=N +1
(75)
σI2 [N
+ · · · + σI2 [j + N
+
rd
QN [j+2N −i]qi−N
+ 1] + σI2 [N + 2]
− 1] + σo2 + σc2 [j +
q0rd )
N]
rd
rd
[βQN [j+2N −i]qi−N
(1 − qi−N
)
rd
+ β(1 − β)(QN [j+2N −i]qi−N
)2 ]
+ σo2 + µrd,1 [j + N ].
(76)
For the above multi-hop scheme, the optimal test at the destination nanomachine in a multi-hop molecular communication
network is given by the result below.
Theorem 7: The optimal detector at the destination
nanomachine corresponding to the received signal in the
(j + N )th time slot is
H1
′
T (Rrd [j + N ]) = Rrd [j + N ] ≷ γrd
[j + N ],
(77)
H0
′
where the optimal decision threshold γrd
[j + N ] is given as,
p
′
(78)
γrd
[j + N ] = γrd [j + N ] − αrd [j + N ].
The quantity αrd [j + N ] is
αrd [j + N ]
2
2
µrd,1 [j+N ]σrd,0
[j+N ]−µrd,0 [j+N ]σrd,1
[j+N ]
,
=
2
2
σrd,1 [j+N ]−σrd,0 [j+N ]
(79)
with γrd [j + N ] as given in (80) and βrd is defined as
(N )
βrd =
(N )
(1−β)(1−PF A [j+N −1])−β(1−Pd
[j+N −1])
(N )
(N )
βPd [j+N −1]−(1−β)PF A [j+N −1]
, (81)
The associated expressions for the probabilities of detection
(N )
(N )
PD [j + N − 1] and false alarm PF A [j + N − 1] for the N th
cooperative nanomachine can be obtained as given in (116)
and (117) respectively.
Proof: Let ξl (n) denote the state of the nth cooperative
nanomachine (Rn ) with ξl (n) = 1 if Rn decodes the source
information symbol as 1 and ξl (n) = 0 otherwise. Let
ξl = [ξl (1), ξl (2), · · · , ξl (N )], 0 ≤ l ≤ 2N − 1 denote the
set of all the possible 2N binary vector states. The binary
vector ξ0 = [0, 0, · · · , 0, 0] represents the state when all the
cooperative nanomachines decode the source symbol as 0
whereas ξ2K −1 = [1, 1, · · · , 1, 1] represents the state when all
the cooperative nanomachines decode the symbol as 1. Further,
let the set Ψl defined as Ψl = {n|ξl (n) = 1, n = 1, 2, · · · , N }
include all the cooperative nanomachines which decode the
symbol as 1. Employing the above notation, the optimal LRT
L(Rrd [j +N ]) at the destination nanomachine can be obtained
as,
p(Rrd [j + N ]|H1 ) H1 1 − β
L(Rrd [j + N ]) =
≷
.
(82)
p(Rrd [j + N ]|H0 ) H0 β
The test statistic L(Rrd [j + N ]) above is further simplified in
(83), where p(Rrd [j + N ]|ξl ) denotes the PDF of the received
molecules Rrd [j + N ] at the destination nanomachine for the
scenario when the network is in state ξl and is given as,
p(Rrd [j+N ]|ξl )
(
2
N (µrd,0 [j+N ], σrd,0
[j+N ]) if l=0, 2, · · ·, 2N −2,
=
(84)
2
N (µrd,1 [j+N ], σrd,1 [j+N ]) if l=1, 3, · · ·, 2N −1,
The quantity Pr(ξl |Hi ), i∈{0, 1} in (83) represents the conditional probability that the network is in state ξl . Since the
source and each of the cooperative nanomachines employ
different types of molecules for transmission, the probability
Pr(ξl |Hi ) of the system being in state ξl under hypothesis Hi
follows as,
Pr(ξl |Hi ) =
N
Y
n=1
Pr(ξl (n)|Hi ),
(85)
where the probabilities Pr(ξl (n)|H0 ) and Pr(ξl (n)|H1 ) of the
cooperative nanomachine being in state ξl (n) under H0 and
H1 can be expressed as,
(
(n)
PF A [j+n−1],
if n ∈ Ψl ,
(86)
Pr(ξl (n)|H0 ) =
(n)
if n ∈ Ψl ,
1 − PF A [j+n−1],
(
(n)
PD [j+n−1],
if n ∈ Ψl ,
Pr(ξl (n)|H1 ) =
(87)
(n)
if n ∈ Ψl .
1 − PD [j+n−1],
The set Ψl comprises of all the cooperative nanomachines that
(n)
(n)
decode the symbol as 0, and PD [j+n−1] and PF A [j+n−1]
are the probabilities of detection and false alarm of the nth
intermediate cooperative nanomachine. Employing the above
expressions, the probabilities of the system being in states ξl
under H0 and H1 can be written as,
Y
Y (n)
(n)
1−PF A [j+n−1] , (88)
PF A [j+n−1]
Pr(ξl |H0 )=
n∈Ψl
n∈Ψ
n∈Ψl
n∈Ψl
l
Y (n)
Y
(n)
Pr(ξl |H1 )=
PD [j+n−1]
1−PD [j+n−1] . (89)
Further, substituting the Gaussian PDFs from (84) in (83), the
test can be expressed as in (90), where βrd is defined as,
P
[(1 − β)Pr(ξl |H0 ) − βPr(ξl |H1 )]
βrd =
l=0,2,··· ,2N −2
P
[βPr(ξl |H1 ) − (1 − β)Pr(ξl |H0 )]
l=1,3,··· ,2N −1
=
(N )
,
(N )
(1−β)(1−PF A [j+N −1])−β(1−PD [j+N −1])
(N )
(N )
βPD [j+N −1] − (1 − β)PF A [j+N −1]
.(91)
The detailed derivations for the test (90) and βrd above are
given in Appendix B. The expression for f (Rrd [j + N ]) in
(90) can be further simplified as shown in (92). Substituting the resulting expression for f (Rrd [j + N ]) in (90) and
subsequently merging the terms independent of the received
molecules Rrd [j + N ] with the threshold yields the test
2
H1
(Rrd [j + N ] + αrd [j + N ]) ≷ γrd [j + N ],
(93)
H0
where αrd [j + N ] and γrd [j + N ] are defined in (79) and (80)
respectively. Finally, taking the square root of both sides with
γrd [j + N ] ≥ 0, the above expression can be simplified to
yield the optimal test in (77).
Result below characterizes the detection performance at the
destination nanomachine in the multi-hop network.
Theorem 8: The average probabilities of detection PD and
false alarm PF A at the destination nanomachine corresponding
to the transmission by the source nanomachine in slots 1 to k
in the multi-hop molecular communication system with N ≥ 2
cooperative nanomachines are given as,
PD =
PF A =
k
1X d
P [j + N ],
k j=1 D
k
1X d
P [j + N ],
k j=1 F A
(94)
(95)
γrd [j + N ] = ln
+
L(Rrd [j + N ])
2N
−1
P
l=0
= N
2P
−1
l=0
=
"s
2
σrd,1
[j+N ]
2
σrd,0
[j+N ]
βrd
#
2
2
2σrd,1
[j + N ]σrd,0
[j + N ]
+(αrd [j + N ])2
2
2
σrd,1 [j + N ]−σrd,0 [j + N ]
2
2
µ2rd,1 [j + N ]σrd,0
[j + N ] − µ2rd,0 [j + N ]σrd,1
[j + N ]
,
2
2
σrd,1 [j + N ] − σrd,0 [j + N ]
(80)
p(Rrd [j + N ]|ξl )Pr(ξl |H1 )
p(Rrd [j + N ]|ξl )Pr(ξl |H0 )
P
p(Rrd [j + N ]|ξl )Pr(ξl |H1 ) +
l=0,2,··· ,2N −2
P
l=0,2,··· ,2N −2
p(Rrd [j + N ]|ξl )Pr(ξl |H0 ) +
P
l=1,3,··· ,2N −1
P
l=1,3,··· ,2N −1
p(Rrd [j + N ]|ξl )Pr(ξl |H1 )
p(Rrd [j + N ]|ξl )Pr(ξl |H0 )
,
(83)
,f (Rrd [j+N ])
z
}|
{
#
"s
2
2
2
σrd,1
[j + N ]
(Rrd [j+N ]−µrd,0 [j+N ])2 σrd,1
[j+N ]−(Rrd [j + N ]−µrd,1 [j+N ])2 σrd,0
[j+N ] H1
≷ ln
βrd .
2
2
2
2σrd,0
[j+N ]σrd,1
[j+N ]
σrd,0
[j + N ]
H0
(90)
f (Rrd [j + N ])
2
2
2
2
2
=Rrd
[j+N ](σrd,1
[j+N ] − σrd,0
[j+N ]) + 2Rrd [j+N ](µrd,1 [j+N ]σrd,0
[j+N ] − µrd,0 [j+N ]σrd,1
[j+N ])
2
2
+ (µ2rd,0 [j+N ]σrd,1
[j+N ] − µ2rd,1 [j+N ]σrd,0
[j+N ])
#2
"
2
2
µrd,1 [j+N ]σrd,0
[j+N ]−µrd,0 [j+N ]σrd,1
[j+N ]
2
2
=(σrd,1 [j+N ]−σrd,0 [j+N ]) Rrd [j+N ]+
2
2
σrd,1
[j+N ]−σrd,0
[j+N ]
−
2
2
(µrd,1 [j+N ]σrd,0
[j+N ]−µrd,0 [j+N ]σrd,1
[j+N ])2
2
2
+(µrd,1 [j+N ]σrd,0
[j + N ]−µrd,0 [j+N ]σrd,1
[j+N ]).
2
2
σrd,1
[j + N ] − σrd,0
[j + N ]
d
where the probabilities of detection PD
[j +N ] and false alarm
d
PF A [j + N ] at the destination in the (j + N )th slot are given
as,
d
PD
[j
+ N]
′
γrd [j+N ]−µrd,0 [j+N ]
(N )
1−PD [j+N −1]
=Q
σ
[j + N ]
′ rd,0
γrd [j+N ]−µrd,1 [j+N ]
(N )
+Q
PD [j+N −1], (96)
σrd,1 [j + N ]
PFd A [j + N ]
′
γrd [j+N ]−µrd,0 [j+N ]
(N )
=Q
1−PF A [j+N −1]
σ
[j + N ]
′ rd,0
γrd [j+N ]−µrd,1 [j+N ]
(N )
+Q
PF A [j+N −1], (97)
σrd,1 [j + N ]
′
with the threshold γrd
[j + N ] as defined in (78).
d
Proof: The probability of detection PD
[j + N ] at the
destination nanomachine corresponding to the transmission by
RN in the (j + N )th slot can be derived using the test statistic
T (Rrd [j + N ]) given in (77) as
d
PD
[j + N ]
=Pr(T (Rrd[j + N ]) >
+ N ]|H1 )
N
=
2X
−1
l=0
′
[j +
where Pr(ξl |H1 ) is given in (89) and Pr(Rrd [j +N ] > γrd
N ]|ξl ) can be obtained using (84) as,
′
Pr(Rrd [j + N ] > γrd
[j + N ]|ξl )
′
γ
[j+N
]−µ
[j+N ]
rd,0
rd
Q
if l = 0, 2, · · · , 2N −2,
′ σrd,0 [j+N ]
=
Q γrd [j+N ]−µrd,1 [j+N ] if l = 1, 3, · · · , 2N −1.
σrd,1 [j+N ]
′
Pr(Rrd [j + N ] > γrd
[j + N ]|ξl )Pr(ξl |H1 ), (98)
(99)
Finally, employing (89) and (99) in (98), the expression for
d
PD
[j + N ] is obtained as
d
PD
[j + N ]
X
=
l=0,2,··· ,2N −2
Q
′
γrd
[j + N ] − µrd,0 [j + N ]
σrd,0 [j + N ]
Y
Y (n)
(n)
1−PD [j+n−1]
×
PD [j+n−1]
n∈Ψl
n∈Ψ
l
′
γrd
[j + N ] − µrd,1 [j + N ]
+
Q
σrd,1 [j + N ]
l=1,3,··· ,2N −1
Y
Y (n)
(n)
1−PD [j+n−1] , (100)
×
PD [j+n−1]
X
n∈Ψl
′
γrd
[j
(92)
n∈Ψl
d
The above expression for PD
[j + N ] considering 2N possible
states for N cooperative nanomachines reduces to (96).
Similarly, the probability of false alarm PFd A [j + N ] at the
destination nanomachine in the (j + N )th slot can be derived
2
and the mean µrr,1 [j + n] and variance σrr,1
[j + n] under
hypothesis H1 are given as,
as
PFd A [j + N ]
′
=Pr(T (Rrd [j + N ]) > γrd
[j + N ]|H0 )
µrr,1 [j + n] =Qn [j+n]q0rr + β
N
=
2X
−1
l=0
′
Pr(Rrd [j+N ] > γrd
[j+N ]|ξl )Pr(ξl |H0 ), (101)
where Pr(ξl |H0 ) is given in (88). Further, substituting (88) and
(99) in (101), the expression for PFd A [j + N ] follows as
PFd A [j
=
+ N]
X
Q
l=0,2,··· ,2N −2
′
γrd
[j
+ N ] − µrd,0 [j + N ]
σrd,0 [j + N ]
Q
n∈Ψl
d
Similar to proof for PD
[j + N ] shown above, the expression
in (102) can be simplified considering 2N possible states for
N cooperative nanomachines as (97).
Similar to the destination nanomachine, the binary hypothesis testing problem for symbol detection at the cooperative
nanomachine Rn+1 , n = 1, 2, · · · , N − 1, corresponding
to the transmission of the cooperative nanomachine Rn in
(j + n − 1)th time slot can also be formulated as
H0 :Rrr [j+n]=Irr [n+1]+Irr [n+2]+ · · · +Irr [j+n−1]
+Nrr [j+n]+Crr [j+n]
H1 :Rrr [j+n]=Srr [j+n] + Irr [n+1]+Irr [n+2]+ · · ·
+Irr [j+n−1]+Nrr [j+n]+Crr [j+n],
2
H0 : Rrr [j + n] ∼ N (µrr,0 [j + n], σrr,0
[j + n])
2
H1 : Rrr [j + n] ∼ N (µrr,1 [j + n], σrr,1
[j + n]),
2
σrr,0
[j + n] =
i=n+1
j+n−1
X
i=n+1
+ σo2 + µrr,0 [j + n],
H1
′
T (Rrr [j + n]) = Rrr [j + n] ≷ γrr
[j + n],
(109)
H0
′
where the optimal decision threshold γrr
[j + n] is
′
γrr
[j + n] =
where
p
γrr [j + n] − αrr [j + n],
(110)
αrr [j + n]
2
2
µrr,1 [j+n]σrr,0
[j+n]−µrr,0 [j+n]σrr,1
[j+n]
, (111)
=
2
2
σrr,1 [j+n]−σrr,0 [j+n]
(104)
(105)
(1−β)(1−PF A [j+n−1])−β(1−Pd [j+n−1])
(n)
(n)
βPd [j+n−1]−(1−β)PF A [j+n−1]
(n)
. (113)
(n)
The quantities PD [j + n − 1] and PF A [j + n − 1] denote
the probabilities of detection and false alarm for the nth cooperative nanomachine. Further, the probabilities of detection
(n+1)
(n+1)
[j +n] and false alarm PF A [j +n] at the cooperative
PD
nanomachine Rn+1 corresponding to the transmission by Rn
in the (j + n)th slot can be derived using the test statistic
T (Rrr [j + n]) in (109) as
(n+1)
PD
[j + n]
′
[j + n]|H1 )
=Pr(T (Rrr [j + n]) > γrr
=
n
2X
−1
′
Pr(Rrr [j+n] > γrr
[j+n]|ξl )Pr(ξl |H1 ), (114)
n
2X
−1
′
Pr(Rrr [j+n] > γrr
[j+n]|ξl )Pr(ξl |H0 ), (115)
l=0
(n+1)
PF A [j +
n]
′
=Pr(T (Rrr [j + n]) > γrr
[j + n]|H0 )
=
(106)
(n)
(n)
βrr =
rr
rr
[βQn [j+2n−i]qi−n
(1 − qi−n
)+β
rr 2
× (1 − β)(Qn [j+2n−i]qi−n
) ]
(108)
and γrr [j + n] is given in (112), where βrr is defined as
2
where the mean µrr,0 [j + n] and variance σrr,0
[j + n] under
hypothesis H0 are given as,
+ µo ,
rr
rr
[βQn [j+2n−i]qi−n
(1 − qi−n
)
Hence, the optimal test for deciding 0 and 1 at the cooperative nanomachine Rn+1 considering all possible 2n states of
previous n cooperative nanomachines i.e, R1 , R2 , · · · , Rn can
be obtained as
(103)
where H0 and H1 denote the null and alternative hypotheses
corresponding to the transmission of the decoded symbol x
b[j+
n − 1] = 0 and x
b[j + n − 1] = 1 respectively by Rn in the
(j + n − 1)th time slot. The number of molecules received at
Rn+1 corresponding to the individual hypotheses in (103) are
Gaussian distributed as
µrr,0 [j + n] =β
+
−
(107)
q0rr )
+ σo2 + µrr,1 [j + n].
n∈Ψ
′
γrd
[j +
rr
Qn [j+2n−i]qi−n
+ n]
=Qn [j+n]q0rr (1
j+n−1
X
rr
Qn [j+2n−i]qi−n
rr 2
+ β(1 − β)(Qn [j+2n−i]qi−n
) ]
n∈Ψl
j+n−1
X
2
σrr,1
[j
i=n+1
N ] − µrd,1 [j + N ]
σrd,1 [j + N ]
l=1,3,··· ,2N −1
Y
Y (n)
(n)
1−PF A [j+n−1] . (102)
×
PF A [j+n−1]
+
X
i=n+1
+ µo ,
Y (n)
Y
(n)
×
PF A [j+n−1]
1−PF A [j+n−1]
n∈Ψl
j+n−1
X
l=0
γrr [j + n] = ln
+
"s
#
2 [j+n]
2
2
σrr,1
2σrr,1
[j + n]σrr,0
[j + n]
β
+(αrr [j + n])2
rr
2
2
2
σrr,0 [j+n]
σrr,1 [j + n]−σrr,0 [j + n]
2
2
µ2rr,1 [j + n]σrr,0
[j + n] − µ2rr,0 [j + n]σrr,1
[j + n]
,
2
2
σrr,1 [j + n] − σrr,0 [j + n]
On lines similar to proof of Theorem 8, the above expressions
(n+1)
(n+1)
[j+n] and PF A [j+n] can be simplified as
PD
(n+1)
PD
[j
=
+ n]
X
Q
l=0,2,··· ,2n −2
′
γrr
[j + n] − µrr,0 [j + n]
σrr,0 [j + n]
+
X
Q
l=1,3,··· ,2n −1
n∈Ψl
′
γrr
[j +
n] − µrr,1 [j + n]
σrr,1 [j + n]
Y
Y (n)
(n)
×
PD [j+n−1]
1−PD [j+n−1]
n∈Ψl
n∈Ψl
′
γrr
[j+n]−µrr,0 [j+n]
(n)
1−PD [j+n−1]
=Q
σ [j+n]
′ rr,0
γrr [j+n]−µrr,1 [j+n]
(n)
+Q
PD [j+n−1]. (116)
σrr,1 [j+n]
X
Pr(y[j + N ]|x[j])Pr(x[j])
x[j]∈{0,1} y[j+N ]∈{0,1}
Y (n)
Y
(n)
×
PD [j+n−1]
1−PD [j+n−1]
n∈Ψl
The mutual information I(X[j], Y [j + N ]) between X[j] and
Y [j + N ] in multi-hop link can be expressed as
I(X[j],Y [j + N ])
X
=
(112)
× log2
Pr(y[j + N ]|x[j])
, (119)
Pr(y[j + N ]|x[j])Pr(x[j])
P
x[j]∈{0,1}
where Pr(x[j]=1)=β, Pr(x[j]=0)=1−β, and Pr(y[j + N ] ∈
{0, 1}|x[j] ∈ {0, 1}) can be determined as
Pr(y[j+N ]=0|x[j]=0) =1 − PFd A [j + N ],
Pr(y[j+N ]=1|x[j]=0) =PFd A [j + N ],
d
Pr(y[j+N ]=0|x[j]=1) =1 − PD
[j + N ],
d
Pr(y[j+N ]=1|x[j]=1) =PD
[j + N ].
The capacity C[k] of the multi-hop channel, when k approaches ∞, can be obtained by maximizing the mutual
information I(X[j], Y [j + N ]) obtained corresponding to the
transmission by the source-nanomachine in slots 1 to k as
k
C[k]= max
(n+1)
PF A [j
=
+ n]
X
β
Q
l=0,2,··· ,2n −2
′
γrr
[j + n] − µrr,0 [j + n]
σrr,0 [j + n]
+
X
l=1,3,··· ,2n −1
Q
n∈Ψl
′
γrr
[j + n] − µrr,1 [j + n]
σrr,1 [j + n]
Y
Y (n)
(n)
×
PF A [j+n−1]
1−PF A [j+n−1]
n∈Ψl
n∈Ψ
l
′
γrr
[j+n]−µrr,0 [j+n]
(n)
1−PF A [j+n−1]
=Q
σ [j+n]
′ rr,0
γrr [j+n]−µrr,1 [j+n]
(n)
+Q
PF A [j+n−1]. (117)
σrr,1 [j+n]
The average probability of error (Pe ) at the destination
nanomachine corresponding to the transmission by the source
nanomachine in slots 1 to k for multi-hop molecular communication follows as
Pe =
k
i
1 Xh
d
β 1−PD
[j+N ] +(1−β)PFd A [j+N ] .
k j=1
(120)
V. S IMULATION R ESULTS
Y (n)
Y
(n)
×
PF A [j+n−1]
1−PF A [j+n−1]
n∈Ψl
X
1
I(X[j], Y [j+N ]) bits/slot.
(k+N ) j=1
(118)
For simulation purposes, the distance between the source
and destination nanomachines is set as dsd = 30 µm, whereas
the cooperative nanomachines in dual and multi-hop systems
are located at distances dsr = drd = 15 µm and dsr1 =
dr1 r2 = dr2 d = 10 µm respectively. The other parameters are
diffusion coefficient D = 2.2 × 10−11 m2 /s, degradation parameter α = 0.2, slot duration τ ∈ {2, 2.5, 3}s, prior probability β = 0.5, and drift velocity v ∈ {1.5×10−5, 7×10−6 } m/s.
The MSI at each receiving node is modeled as a Gaussian
distributed RV with mean µo = 20 and variance σo2 = 20.
Similar to [20], the results are computed for a total of K = 30
slots with Q0 [j] = Q0 , Q1 [j + 1] = Q1 , Q2 [j + 2] = Q2 , ∀j.
Figs. 5-7 demonstrate the detection performance at the
destination nanomachine in the diffusion based direct, dualhop, and multi-hop molecular communication systems under
various scenarios. For simulation purposes, the probabilities
of detection and false alarm at the cooperative nanomachine
R in dual-hop communication as well as at the cooperative
nanomachine R1 in multi-hop communication are fixed as
(1)
(1)
PD [j] = 0.99 and PF A [j] = 0.01 respectively. Fig. 5 shows
the probability of detection versus probability of false alarm
at the destination for a varying number of molecules at each
transmitting nanomachine. One can observe that an increase in
the number of molecules emitted by the nanomachines results
1
Probability of Detection (PD)
Probability of Detection (PD)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Analytical, Number of Molecules = 35
Analytical, Number of Molecules = 60
Direct (Simulated)
Dual−hop (Simulated)
Multi−hop N=2 (Simulated)
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Analytical
Dual−hop (Simulated) d =15 µm
0.2
Dual−hop (Simulated) drd=20 µm
Dual−hop (Simulated) drd=25 µm
0.1
0
1
rd
Dual−hop (Simulated) drd=30 µm
0
0.1
0.2
0.3
Probability of False Alarm (PFA)
Detection performance of direct, dual-hop, and multi-hop
(N = 2) diffusion based molecular systems for a varying number
of molecules with D = 2.2 × 10−11 m2 /s, τ = 2.5 s, and v =
7 × 10−6 µm/s.
Fig. 5.
0.5
0.6
0.7
0.9
1
ular communication system for various relay-destination distances
with Q0 = Q1 = 100, D = 2.2 × 10−11 m2 /s, τ = 2.0 s, and
v = 7 × 10−6 µm/s.
0
10
Analytical
Direct (Simulated)
Dual−hop (Simulated)
Multi−hop N=2 (Simulated)
Probability of Error (Pe)
0.9
0.8
0.7
0.6
0.5
0.4
Analytical, σ2= 20
0.3
Analytical,
0.2
o
2
σo=
50
Direct (Simulated)
Dual−hop (Simulated)
Multi−hop N=2 (Simulated)
0.1
0
0.8
Fig. 7. Detection performance of a diffusion based dual-hop molec-
1
Probability of Detection (PD)
0.4
Probability of False Alarm (PFA)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
−1
10
P(1)[j]=0.999, P(1)[j]=10−3
−2
10
FA
D
−3
10
−4
10
(1)
(1)
−4
PD [j]=0.9999, PFA[j]=10
1
2
Probability of False Alarm (PFA)
4
6
8
10
Drift Velocity (v)
12
14
−6
x 10
Fig. 6.
Detection performance of direct, dual-hop, and multihop (N = 2) diffusion based molecular systems for a varying
noise variance σo2 with µo = 20, Q0 = Q1 = Q2 = 60,
D = 2.2 × 10−11 m2 /s, τ = 2.0 s, and v = 7 × 10−6 µm/s.
Error rate performance of the direct, dual-hop, and multihop (N = 2) diffusion based molecular systems for a varying drift
velocity v with µo = σo2 = 20, Q0 = Q1 = Q2 = 150, D =
2.2 × 10−11 m2 /s, and τ = 2.5 s.
in a higher probability of detection at the destination for a
fixed value of probability of false alarm. It can also be clearly
seen from Figs. 5-6 that the detection performance at the
destination nanomachine significantly improves as the number
of cooperative nanomachines N increases between the source
and destination nanomachines. One can also observe in Fig. 5
that for a fixed value of the probability of false alarm 0.1 and
number of molecules 60, the probability of detection increases
from 0.24 to 0.99 as the number of cooperative nanomachines
N increases from 0 to 2, where N = 0 represents direct
communication. Similar improvement can also be seen in Fig.
6 for different values of noise variance σo2 .
Fig. 7 shows the probability of detection versus probability
of false alarm at the destination nanomachine for various
distances drd between the cooperative and destination nanomachines in a dual-hop based molecular communication system.
One can observe that the distance drd between cooperative
nanomachine R and destination nanomachine has a significant
impact on the detection performance at the destination. The
probability of detection decreases significantly from approximately 0.96 to 0.16 at probability of false alarm 0.1 as the
distance drd increases from 15 µm to 30 µm.
Fig. 8 shows the probability of error Pe versus the drift
velocity v for a fixed number of molecules Q0 = Q1 = Q2 =
150. First, it can be observed from Fig. 8 that the analytical
Pe values obtained using (23), (63), and (118) for direct,
dual and multi-hop systems coincide with those obtained
from simulations, thus validating the derived analytical results.
One can also observe that the end-to-end performance of the
system is significantly enhanced by cooperative nanomachines
in comparison to the direct source-destination only communication scenario with low drift velocity. Moreover, similar
to the detection performance, the end-to-end probability of
error also decreases progressively as the number of cooperative
nanomachines N increases. Further, the dual-hop and multihop molecular communication systems experience a floor at
high drift velocity, whereas the performance of direct communication system improves with increasing drift velocity.
Fig. 8.
0
0
10
0
10
10
′
Optimal γsd[j]=48
2
for σo=µo=30
−2
′
10
Optimal γsd[j]=37.2
2
for σo=µo=20
′
Optimal γsd[j]=25.4
2
for σo=µo=10
0
10
20
10
−2
10
Optimal γ′ [j+1]=66.3
rd
for σ2=µ =40
−3
10
o
′
Optimal γrd[j+1]=30.4
2
for σo=µo=10
−4
10
Analytical
Direct (Simulated)
−3
10
e
e
−1
10
30
40
50
60
Detection Threshold at the Destination
70
(c)
−1
Probability of Error (P )
(b)
Probability of Error (P )
Probability of Error (Pe)
(a)
o
Analytical
(1)
D
(1)
FA
Dual−hop (Simulated), P [j]=0.9, P [j]=0.1
Dual−hop (Simulated), P(1)[j]=0.99, P(1)[j]=0.01
D
(1)
D
FA
(1)
FA
−1
10
′
Optimal γrd[j+2]=31.9
for Q2=30
−2
10
′
Optimal γrd[j+2]=43.7
for Q2=70
Optimal γ′rd[j+2]=38.14
for Q2=50
−3
10
Analytical
Multi−hop N=2 (Simulated)
Dual−hop (Simulated), P [j]=0.999, P [j]=0.001
80
0
10
20
30
40
50
60
70
80
90
−4
10
100
0
10
Detection Threshold at the Destination
20
30
40
50
60
70
80
90
100
Detection Threshold at the Destination
Error rate of diffusion based molecular systems versus detection threshold for various scenarios with (a) direct communication with
(1)
Q0 = 60 (b) dual-hop communication with Q1 = 60 (c) multi-hop (N = 2) communication with µo = σo2 = 20, PD [j] = 0.999, and
(1)
PF A [j] = 0.001.
Fig. 9.
Fig. 10 shows the capacity of diffusion based direct, dualhop, and multi-hop molecular communication systems, where
the maximum mutual information is achieved for equiprobable
information symbols, i.e., β = 0.5. As depicted in Fig. 10(a),
the capacity of molecular communication systems significantly
reduces as the variance (σo2 ) of MSI increases. Further, one
can also observe that the capacity of the dual-hop and multihop systems depends highly on the detection performance of
the intermediate cooperative nanomachines. As the detection
(1)
(1)
performance (PD [j], PF A [j]) at R in dual-hop and R1 in
multi-hop systems increases from (0.9, 0.1) to (0.99, 0.01),
a significant capacity gain can be achieved at high MSI in
comparison to the direct communication. However, at low
MSI, the direct communication system achieves high capacity
values in comparison with dual and multi-hop systems. This
is due to the fixed probabilities of detection and false alarm
at the intermediate cooperative nanomachine as well as the
1
associated with the effective capacity of the
scaling factor k+N
cooperative transmission. One can also notice that the multi-
Direct
Dual−hop
Multi−hop N=2
P(1)[j]=0.99, P(1)[j]=0.01
FA
D
0.9
Capacity (bits/slot)
Fig. 9 shows the error rate performance versus detection
threshold at the destination nanomachine in various scenarios.
For this simulation, the diffusion coefficient, the slot duration
and the drift velocity are set as, D = 2.2×10−11 m2 /s, τ = 3 s,
and v = 1.5×10−5 µm/s respectively. One can clearly observe
that the optimal threshold obtained using (12), (47) and (78)
for direct, dual-hop and multi-hop systems respectively, is able
to achieve the minimum probability of error at the destination
nanomachine. It can also be seen that the optimal values of
the threshold at the destination depend on several parameters.
The optimal value of the threshold for direct communication
and dual-hop scenarios in Figs. 9(a)-(b) increases from 25.4 to
48 and 30.4 to 66.3 as the MSI increases from µo = σo2 = 10
to µo = σo2 = 30 and µo = σo2 = 10 to µo = σo2 = 40
respectively. A similar observation can also be made for multihop communication in Fig. 9(c) where the optimal threshold
at the destination nanomachine increases from 31.9 to 43.7 as
the number of molecules Q2 increases from 30 to 70.
1
0.8
0.7
0.6
0.5
0.4
0.3
(1)
(1)
PD [j]=0.9, PFA[j]=0.1
0
20
40
(a)
60
80
100
120
140
160
180
200
Noise Variance (σ2o)
1
0.9
Capacity (bits/slot)
This is owing to the fact that the end-to-end performance of
dual and multi-hop systems at high drift velocity is dominated
by the weak source-cooperative nanomachine link with fixed
(1)
(1)
(PD [j], PF A [j]).
0.8
P(1)[j]=0.99, P(1)[j]=0.01
0.7
FA
D
(1)
(1)
PD [j]=0.9, PFA[j]=0.1
0.6
0.5
Direct
Dual−hop
Multi−hop N=2
0.4
(b)
0.3
2
2.2
2.4
2.6
2.8
3
3.2
3.4
3.6
3.8
4
Slot Duration (τ)
Capacity of diffusion based direct, dual-hop and multi-hop
molecular communication systems with Q0 = Q1 = Q2 = 60,
D = 2.2 × 10−11 m2 /s, v = 1.5 × 10−5 µm/s, µo = 20, and (a)
increasing noise variance (σo2 ) with τ = 2.5 s (b) slot duration (τ )
with σo2 = 20.
Fig. 10.
hop system achieves lower capacity values in comparison to
the dual-hop system. This is due to the transmission between
R1 and R2 in the multi-hop system that significantly degrades
the detection performance at the destination nanomachine. On
the other hand, the capacity of the molecular system without
cooperating nanomachines increases significantly with an increase in the slot duration (τ ), as is clearly shown in Fig. 10(b).
However, the capacity values obtained in dual and multi-hop
0.5
0.5
0.4
0.4
0.3
0.2
0.3
0.1
0.2
Q0=Q1=Q2=60
0
0
20
0.1
40
60
80
Molecules at R2
0
10
20
30
40
50
60
70
80
90
Capacity (bits/slot)
Capacity (bits/slot)
0.6
0.6
(b)
0.6
0.7
0.7
1
0.5
0.5
0.5
0.45
0.4
0.4
0.3
0.35
0.3
0.2
0.25
0.1
0.2
0.15
Q0=42, Q1=83, Q2=55
0
0
40
60
0
10
Molecules at R2
20
30
40
50
0.6
0.45
0.5
0.4
0.4
0.35
0.3
0.3
0.2
0.25
0.1
60
70
80
90
0.2
0
0
0.1
20
80
Molecules at R
0.55
(c)
0.55
Capacity (bits/slot)
(a)
0.8
0.05
80
20
0.1
60
40
0.05
40
60
80
Molecules at R1
0.15
Q0=55, Q1=83, Q2=42
20
0
Molecules at R
Molecules at R
1
2
Fig. 11. Capacity of diffusion based multi-hop (N = 2) molecular system for different molecules allocation at each transmitting node with
total molecule budget Q0 + Q1 + Q2 = 180, µo = 20, σo2 = 10, D = 2.2 × 10−11 m2 /s, τ = 2 s, v = 7 × 10−6 µm/s, and distances (a)
dsr1 = dr1 r2 = dr2 d = 10 µm (b) dsr1 = 5 µm, dr1 r2 = 15 µm, dr2 d = 10 µm (c) dsr1 = 10 µm, dr1 r2 = 15 µm, dr2 d = 5 µm.
systems are constant with increasing slot duration and are
(1)
d
d
equal to the ones obtained for PD
[j+1] = PD
[j+2] = PD [j]
(1)
and PFd A [j + 1] = PFd A [j + 2] = PF A [j]. This is due to the
fact the detection performance at the destination nanomachine
(1)
(1)
is limited by the performance metrics (PD [j], PF A [j]) of the
cooperative nanomachine R in dual-hop and R1 in multi-hop
systems.
Fig. 11 finally demonstrates the impact of number of
molecules allocated to each transmitting nanomachine considering a fixed molecule budget Q0 + Q1 + Q2 = Q = 180
and various distance configurations in a multi-hop molecular
communication system. Firstly, One can observe that allocating equal number of molecules, Q0 = Q1 = Q2 = Q
3 = 60,
at the source and cooperative nanomachines R1 and R2 is
only optimal for the scenarios when the distances between
all communicating nanomachines are equal. For the scenarios
when dr1 r2 > dr2 d > dsr1 and dr1 r2 > dsr1 > dr2 d as shown
in Fig. 11(b) and Fig. 11(c) respectively, the system capacity
can be significantly enhanced by allocating a large fraction
of the molecule budget Q to cooperative nanomachine R1 in
comparison to the source and cooperative nanomachine R2 .
VI. C ONCLUSION
This work analyzed the performance of diffusion based
direct, dual-hop, and multi-hop molecular communication systems in the presence of ISI, MSI, and counting errors along
with Brownian motion and drift. Analytical expressions were
derived for the optimal test statistics and optimal thresholds at
the cooperative and destination nanomachines, together with
the resulting probabilities of detection, false alarm as well as
the end-to-end probability of error for each scenario. In addition, the capacity expressions were also derived. It was seen
that the detection performance at the destination nanomachine
can be significantly enhanced by incorporating cooperative
nanomachines between the source and destination nanomachines. Further, this work also demonstrated that the end-toend performance of dual and multi-hop molecular systems is
dominated by the performance of the weaker link. Finally,
future studies can focus on the impact of mobile nanomachines
as well as the optimization of transmitted molecules in each
time-slot on the performance of relay-assisted diffusive mobile
molecular communication.
A PPENDIX A
D ERIVATION OF E XPRESSION (54)
Substituting the expressions for p(Rrd [j + 1]|b
x[j] = 1) and
p(Rrd [j + 1]|b
x[j] = 0) in (51) and cross multiplying with the
threshold, the expression can be simplified as shown in (121)(123). Now, taking the logarithm of both sides of (123), the
resulting expression can be simplified to yield the expression
given in (54).
A PPENDIX B
D ERIVATION OF E XPRESSION (90)
Substituting the expressions given below for p(Rrd [j +
N ]|ξl ), l = 1, 3, · · · , 2N − 1 and p(Rrd [j + N ]|ξl ), l =
0, 2, · · · , 2N − 2 in (83)
p(Rrd [j + N ]|ξl )
√ 21
2πσrd,1 [j+N ]
2
× exp − (Rrd [j+N2]−µrd,1 [j+N ]) , l=1, · · · , 2N −1
2σrd,1 [j+N ]
=
√ 21
2πσrd,0
[j+N ]
2
× exp − (Rrd [j+N2]−µrd,0 [j+N ]) , l=2, · · · , 2N −2.
2σ
[j+N ]
rd,0
(124)
and cross multiplying with the threshold 1−β
β , the expression
can be simplified as shown in (125)-(127). Now, taking the
logarithm of both sides of (127), the resulting expression can
be simplified to yield the expression given in (90). where βrd
is defined as,
P
[(1−β)Pr(ξl |H0 )−βPr(ξl |H1 )]
,2N −2
l=0,2,···
P
βrd =
, (128)
[βPr(ξl |H1 )−(1−β)Pr(ξl |H0 )]
l=1,3,··· ,2N −1
where the probability of the system being in state ξl under H0
and H1 can be expressed as,
Y (n)
Y
(n)
1−PF A [j+n−1] ,
Pr(ξl |H0 )=
PF A [j+n−1]
n∈Ψl
n∈Ψ
n∈Ψl
n∈Ψl
l
Y
Y (n)
(n)
Pr(ξl |H1 )=
PD [j+n−1]
1−PD [j+n−1] .
!
!
(1)
2
2
(1−P
[j])
(R
[j+1]−µ
[j+1])
(R
[j+1]−µ
[j+1])
rd
rd,1
rd
rd,0
D
β q
+q
exp −
exp −
2
2
2σrd,1
[j+1]
2σrd,0
[j+1]
2
2
2σrd,1
[j+1]
2σrd,0
[j+1]
!
!
(1)
(1)
2
2
H1
[j])
(1−P
[j]
P
(R
[j+1]−µ
[j+1])
(R
[j+1]−µ
[j+1])
rd
rd,0
rd
rd,1
FA
, (121)
+q
−
≷ (1−β) q F A
exp −
2
2
2σrd,1
[j+1]
2σrd,0
[j+1]
H0
2σ 2 [j+1]
2σ 2 [j+1]
(1)
PD [j]
rd,1
(1)
rd,0
(1)
2
(βPD [j] − (1 − β)PF A [j])
(Rrd [j + 1] − µrd,1 [j + 1])
q
exp −
2
2σrd,1
[j + 1]
2
σrd,1
[j + 1]
(1)
exp
2
(Rrd [j+1]−µrd,0 [j+1]) (Rrd [j+1]−µrd,1 [j+1])
−
2
2
2σrd,0
[j+1]
2σrd,1
[j+1]
X
(1)
2
[(1 − β)(1 − PF A [j]) − β(1 − PD [j])]
(Rrd [j + 1] − µrd,0 [j + 1])
q
≷
exp −
2
2σrd,0
[j + 1]
2
H0
σrd,0
[j + 1]
H1
2
!
!
H1
≷
H0
s
,
(122)
"
#
(1)
2
σrd,1
[j+1] (1−β)(1−PF(1)
[j])−β(1−P
[j])
A
D
. (123)
2
(1)
(1)
σrd,0
[j+1]
βP [j]−(1−β)P [j]
D
2
Pr(ξl |H1 )
!
(Rrd [j + N ] − µrd,1 [j + N ])
q
β
exp −
2
2σrd,1
[j + N ]
2
2σrd,1 [j + N ]
l=0,2,··· ,2N −2
FA
!
!
(Rrd [j + N ] − µrd,0 [j + N ])2
q
exp −
+
2
2σrd,0
[j + N ]
2
2σrd,0
[j + N ]
l=1,3,··· ,2N −1
!
2
X
H1
(R
[j
+
N
]
−
µ
[j
+
N
])
Pr(ξ
|H
)
rd
rd,1
l
0
q
exp −
≷ (1 − β)
2
2σrd,1
[j + N ]
H0
2σ 2 [j + N ]
N
X
Pr(ξl |H1 )
l=0,2,··· ,2 −2
X
rd,1
2
(Rrd [j + N ] − µrd,0 [j + N ])
q
exp −
+
2
2σrd,0
[j + N ]
2
2σrd,0
[j + N ]
l=1,3,··· ,2N −1
Pr(ξl |H0 )
Substituting now the above expressions in (128), the expression for βrd can be further simplified for different number of
cooperative nanomachines as follows. For N = 2 cooperative
nanomachines, the expression for βrd can be solved considering the four possible states, i.e., ξ0 = [0, 0], ξ1 = [0, 1], ξ2 =
N2
[1, 0] and ξ3 = [1, 1] as βrd = D
, where N2 and D2 are given
2
in (129) and (130) respectively. The expressions in (129) and
(130) can be further simplified to yield the final expression
for βrd for N = 2 cooperative nanomachines as,
(2)
(2)
(1−β) 1−PF A [j+1] −β 1−PD [j+1]
βrd =
. (131)
(2)
(2)
βPD [j+1] − (1 − β)PF A [j+1]
The expression for βrd considering the eight possible states
for N = 3 cooperative nanomachines can be obtained as
N3
, where N3 and D3 are given in (132) and (133)
βrd = D
3
respectively. The expressions in (132) and (133) can be further
simplified to yield the final expression for βrd for N = 3
cooperative nanomachines as,
(3)
(3)
(1−β) 1−PF A [j+2] −β 1−PD [j+2]
. (134)
βrd =
(3)
(3)
βPD [j+2] − (1 − β)PF A [j+2]
On the similar lines, the expression for βrd considering 2N
possible states due to the presence of N cooperative nanoma-
!
,
(125)
chines can be similarly obtained as,
(N )
(N )
(1−β) 1−PF A [j+N −1] −β 1−PD [j+N −1]
βrd =
.
(N )
(N )
βPD [j+N −1] − (1 − β)PF A [j+N −1]
(135)
R EFERENCES
[1] T. Nakano, A. W. Eckford, and T. Haraguchi, Molecular Communication. Cambridge University Press, 2013.
[2] B. Atakan, Molecular Communications and Nanonetworks. Springer,
2016.
[3] S. Hiyama, Y. Moritani, T. Suda, R. Egashira, A. Enomoto, M. Moore,
and T. Nakano, “Molecular communication,” J. Institute Electron Inf.
Commun. Engineers, vol. 89, no. 2, pp. 162–165, 2006.
[4] Y. Moritani, S. Hiyama, and T. Suda, “Molecular communication for
health care applications,” in Proc. IEEE PerCom Workshops, 2006, pp.
1–5.
[5] T. Nakano, M. J. Moore, Y. Okaie, A. Enomoto, and T. Suda, “Cooperative drug delivery through molecular communication among biological
nanomachines,” in Proc. IEEE ICC, 2013, pp. 809–812.
[6] M. Pierobon and I. F. Akyildiz, “A physical end-to-end model for molecular communication in nanonetworks,” IEEE J. Sel. Areas Commun.,
vol. 28, no. 4, pp. 602–611, 2010.
[7] ——, “Information capacity of diffusion-based molecular communication in nanonetworks,” in Proc. IEEE INFOCOM, 2011, pp. 506–510.
[8] T. Nakano, Y. Okaie, and J.-Q. Liu, “Channel model and capacity
analysis of molecular communication with Brownian motion,” IEEE
Commun. Lett., vol. 16, no. 6, pp. 797–800, 2012.
2
1
(Rrd [j+N ]−µrd,1 [j+N ])
q
exp −
2
2σrd,1
[j+N ]
2
σrd,1
[j+N ]
!
X
[βPr(ξl |H1 )−(1−β)Pr(ξl |H0 )]
l=1,3,··· ,2N −1
2
1
(Rrd [j+N ]−µrd,0 [j+N ])
exp −
≷ q
2
2σrd,0
[j+N ]
2
H0
σrd,0 [j+N ]
H1
2
exp
!
X
[(1−β)Pr(ξl |H0 )−βPr(ξl |H1 )], (126)
l=0,2,··· ,2N −2
2
(Rrd [j + N ] − µrd,0 [j + N ])
(Rrd [j + N ] − µrd,1 [j + N ])
−
2
2
2σrd,0
[j + N ]
2σrd,1
[j + N ]
!
H1
≷
H0
s
2
σrd,1
[j + N ]
2
σrd,0
[j + N ]
(2)
(1)
(2)
(1)
N2 =(1−β) 1−PF A [j] 1−PF A [j+1] −β 1−PD [j] 1−PD [j+1]
(1)
(2)
(1)
(2)
+(1−β)PF A [j] 1−PF A [j+1] −βPD [j] 1−PD [j+1] ,
βrd ,
(1)
(2)
(1)
(2)
(1)
(2)
(1)
(2)
D2 =β 1−PD [j] PD [j+1]− 1−β)(1−PF A [j] PF A [j+1]+βPD [j]PD [j+1]−(1−β)PF A [j]PF A [j+1].
[9] T. Nakano, T. Suda, Y. Okaie, M. J. Moore, and A. V. Vasilakos,
“Molecular communication among biological nanomachines: A layered
architecture and research issues,” IEEE Trans. Nanobiosci., vol. 13,
no. 3, pp. 169–197, 2014.
[10] I. Llatser, E. Alarcón, and M. Pierobony, “Diffusion-based channel
characterization in molecular nanonetworks,” in Proc. IEEE INFOCOM
Workshops, 2011, pp. 467–472.
[11] I. Llatser, A. Cabellos-Aparicio, and E. Alarcon, “Networking challenges
and principles in diffusion-based molecular communication,” IEEE
Wireless Commun., vol. 19, no. 5, 2012.
[12] A. Einolghozati, M. Sardari, and F. Fekri, “Molecular communication
between two populations of bacteria,” in Proc. IEEE ITW, 2012, pp.
437–441.
[13] G. Aminian, H. Arjmandi, A. Gohari, M. Nasiri-Kenari, and U. Mitra,
“Capacity of diffusion-based molecular communication networks over
LTI-Poisson channels,” IEEE Trans. Mol. Biol. Multi-Scale Commun.,
vol. 1, no. 2, pp. 188–201, 2015.
[14] A. Ahmadzadeh, V. Jamali, A. Noel, and R. Schober, “Diffusive mobile
molecular communications over time-variant channels,” IEEE Commun.
Lett., 2017.
[15] A. Ahmadzadeh, V. Jamali, and R. Schober, “Statistical analysis of timevariant channels in diffusive mobile molecular communications,” arXiv
preprint arXiv:1704.06298, 2017.
[16] V. Jamali, A. Ahmadzadeh, C. Jardin, H. Sticht, and R. Schober,
“Channel estimation for diffusive molecular communications,” IEEE
Trans. Commun., vol. 64, no. 10, pp. 4238–4252, 2016.
[17] W. Haselmayr, S. M. H. Aejaz, A. T. Asyhari, A. Springer, and
W. Guo, “Transposition errors in diffusion-based mobile molecular
communication,” IEEE Commun. Lett., 2017.
[18] I. Llatser, A. Cabellos-Aparicio, M. Pierobon, and E. Alarcón, “Detection techniques for diffusion-based molecular communication,” IEEE J.
Sel. Areas Commun., vol. 31, no. 12, pp. 726–734, 2013.
[19] K. R. Liu, Cooperative Communications and Networking. Cambridge
University Press, 2009.
[20] A. Singhal, R. K. Mallik, and B. Lall, “Performance analysis of
amplitude modulation schemes for diffusion-based molecular communication,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5681–5691,
2015.
[21] A. Noel, K. C. Cheung, and R. Schober, “Improving receiver performance of diffusive molecular communication with enzymes,” IEEE
Trans. NanoBiosci., vol. 13, no. 1, pp. 31–43, 2014.
[22] R. Mosayebi, H. Arjmandi, A. Gohari, M. N. Kenari, and U. Mitra,
“Diffusion based molecular communication: A simple near optimal
receiver,” in Proc. IWCIT, 2014, pp. 1–4.
[23] A. Noel, K. C. Cheung, and R. Schober, “Optimal receiver design for
diffusive molecular communication with flow and additive noise,” IEEE
Trans. Nanobiosci., vol. 13, no. 3, pp. 350–362, 2014.
[24] L.-S. Meng, P.-C. Yeh, K.-C. Chen, and I. F. Akyildiz, “On receiver
design for diffusion-based molecular communication,” IEEE Trans.
Signal Process., vol. 62, no. 22, pp. 6032–6044, 2014.
[25] X. Wang, Y. Lu, M. D. Higgins, and M. S. Leeson, “An optimal decoding
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
(127)
(129)
(130)
algorithm for molecular communications systems with noise, memory,
and pulse width,” Nano Commun. Netw., vol. 9, pp. 7–16, 2016.
Z. Cheng, Y. Zhu, K. Chi, Y. Li, and M. Xia, “Capacity analysis for
diffusive molecular communication with ISI channel,” Nano Commun.
Netw., 2017.
Y. Fang, A. Noel, N. Yang, A. W. Eckford, and R. A. Kennedy, “Maximum likelihood detection for collaborative molecular communication,”
arXiv preprint arXiv:1704.05623, 2017.
V. Jamali, A. Ahmadzadeh, N. Farsad, and R. Schober, “SCW codes for
maximum likelihood detection in diffusive molecular communications
without channel state information,” arXiv preprint arXiv:1706.02171,
2017.
V. Jamali, N. Farsad, R. Schober, and A. Goldsmith, “Non-coherent
detection for diffusive molecular communications,” arXiv preprint
arXiv:1707.08926, 2017.
A. Einolghozati, M. Sardari, and F. Fekri, “Relaying in diffusion-based
molecular communication,” in Proc. IEEE ISIT, 2013, pp. 1844–1848.
——, “Decode and forward relaying in diffusion-based molecular communication between two populations of biological agents,” in Proc. IEEE
ICC, 2014, pp. 3975–3980.
A. Ahmadzadeh, A. Noel, and R. Schober, “Analysis and design of twohop diffusion-based molecular communication networks,” in Proc. IEEE
GLOBECOM, 2014, pp. 2820–2825.
A. Ahmadzadeh, A. Noel, A. Burkovski, and R. Schober, “Amplify-andforward relaying in two-hop diffusion-based molecular communication
networks,” in Proc. IEEE GLOBECOM, 2015, pp. 1–7.
Y. Deng, W. Guo, A. Noel, A. Nallanathan, and M. Elkashlan, “Enabling
energy efficient molecular communication via molecule energy transfer,”
IEEE Commun. Lett., vol. 21, no. 2, pp. 254–257, 2017.
X. Wang, M. D. Higgins, and M. S. Leeson, “Relay analysis in molecular
communications with time-dependent concentration,” IEEE Commun.
Lett., vol. 19, no. 11, pp. 1977–1980, 2015.
T. Nakano and J. Shuai, “Repeater design and modeling for molecular
communication networks,” in Proc. IEEE INFOCOM Workshops, 2011,
pp. 501–506.
A. Ahmadzadeh, A. Noel, and R. Schober, “Analysis and design of
multi-hop diffusion-based molecular communication networks,” IEEE
Trans. Mol. Biol. Multi-Scale Commun., vol. 1, no. 2, pp. 144–157,
2015.
A. Einolghozati, M. Sardari, A. Beirami, and F. Fekri, “Data gathering
in networks of bacteria colonies: Collective sensing and relaying using
molecular communication,” in Proc. IEEE INFOCOM Workshops, 2012,
pp. 256–261.
S. Balasubramaniam et al., “Multi-hop conjugation based bacteria
nanonetworks,” IEEE Trans. Nanobiosci., vol. 12, no. 1, pp. 47–59,
2013.
F. Walsh and S. Balasubramaniam, “Reliability and delay analysis of
multihop virus-based nanonetworks,” IEEE Trans. Nanotechnol., vol. 12,
no. 5, pp. 674–684, 2013.
N. Tavakkoli, P. Azmi, and N. Mokari, “Performance evaluation and
optimal detection of relay-assisted diffusion-based molecular communi-
(1)
(2)
(3)
(1)
(2)
(3)
N3 =(1−β) 1−PF A [j] 1−PF A [j + 1] 1−PF A [j + 2] −β 1−PD [j] 1−PD [j + 1] 1−PD [j + 2]
(3)
(2)
(1)
(3)
(2)
(1)
+ (1−β) 1−PF A [j] PF A [j + 1] 1−PF A [j + 2] −β 1−PD [j] PD [j + 1] 1−PD [j + 2]
(1)
(2)
(3)
(1)
(2)
(3)
+(1−β)PF A [j] 1−PF A [j + 1] 1−PF A [j + 2] −βPD [j] 1−PD [j + 1] 1−PD [j + 2]
(1)
(2)
(3)
(1)
(2)
(3)
+ (1 − β)PF A [j]PF A [j + 1] 1 − PF A [j + 2] − βPD [j]PD [j + 1] 1 − PD [j + 2] ,
(132)
(3)
(2)
(1)
(3)
(2)
(1)
D3 =β 1−PD [j] 1−PD [j + 1] PD [j + 2]−(1−β) 1−PF A [j] 1−PF A [j + 1] PF A [j + 2]
(3)
(2)
(1)
(3)
(2)
(1)
+β 1−PD [j] PD [j + 1]PD [j + 2]−(1−β) 1−PF A [j] PF A [j + 1]PF A [j + 2]
(1)
(2)
(3)
(1)
(2)
(3)
+ βPD [j] 1 − PD [j + 1] PD [j + 2] − (1 − β)PF A [j] 1 − PF A [j + 1] PF A [j + 2]
(1)
(2)
(3)
(1)
(2)
(3)
+βPD [j]PD [j + 1]PD [j + 2]−(1−β)PF A [j]PF A [j + 1]PF A [j + 2].
[42]
[43]
[44]
[45]
[46]
cation with drift,” IEEE Trans. Nanobiosci., vol. 16, no. 1, pp. 34–42,
2017.
Y. Moritani, S. Hiyama, and T. Suda, “Molecular communication among
nanomachines using vesicles,” in Proc. NSTI Nanotechnol. Conf., 2006,
pp. 705–708.
K. Srinivas, A. W. Eckford, and R. S. Adve, “Molecular communication
in fluid media: The additive inverse Gaussian noise channel,” IEEE
Trans. Inf. Theory, vol. 58, no. 7, pp. 4678–4692, 2012.
S. Kadloor, R. S. Adve, and A. W. Eckford, “Molecular communication
using Brownian motion with drift,” IEEE Trans. NanoBiosci., vol. 11,
no. 2, pp. 89–99, 2012.
N.-R. Kim, A. W. Eckford, and C.-B. Chae, “Symbol interval optimization for molecular communication with drift,” IEEE Trans. Nanobiosci.,
vol. 13, no. 3, pp. 223–229, 2014.
H. B. Yilmaz, C.-B. Chae, B. Tepekule, and A. E. Pusane, “Arrival mod-
(133)
eling and error analysis for molecular communication via diffusion with
drift,” in Proceedings of the Second Annual International Conference
on Nanoscale Computing and Communication, 2015, pp. 86–95.
[47] L.-S. Meng, P.-C. Yeh, K.-C. Chen, and I. F. Akyildiz, “MIMO communications based on molecular diffusion,” in Proc. IEEE GLOBECOM,
2012, pp. 5380–5385.
[48] M. Pierobon and I. F. Akyildiz, “Diffusion-based noise analysis for
molecular communication in nanonetworks,” IEEE Trans. Signal Process., vol. 59, no. 6, pp. 2532–2547, 2011.
[49] M. H. Kabir, S. R. Islam, and K. S. Kwak, “D-MoSK modulation in
molecular communications,” IEEE Trans. Nanobiosci., vol. 14, no. 6,
pp. 680–683, 2015.
[50] S. M. Kay, “Fundamentals of Statistical Signal Processing: Detection
Theory, vol. 2,” 1998.
| 7cs.IT
|
Evolving Deep Convolutional Neural Networks
for Image Classification
Yanan Sun, Bing Xue and Mengjie Zhang
arXiv:1710.10741v2 [cs.NE] 31 Oct 2017
School of Engineering and Computer Science, Victoria University of Wellington
PO Box 600, Wellington 6140, New Zealand
e-mail: {Yanan.Sunm, Bing.Xue, Mengjie.Zhang}@ecs.vuw.ac.nz
where X and Y are the input data and corresponding label,
respectively, F (·) denotes the architecture choosing function
with the given data, G(·) refers to the initialization method
of the connection weights Weight based on the chosen architecture, and L(·) measures the differences between the
true label and the label predicted by the CNN using X
and Weight . Typically, the Gradient Descend (GD)-based approaches, e.g., Stochastic GD (SGD), are utilized to minimize
L(X, Weight , Y ) within the given number of epochs, where
the connection weight values are iteratively updated. Although
L(·) is not differentiable in all occasions, GD-based methods
are preferred due to their effectiveness and good scalability
as the number of connection weights increases. A CNN
commonly has a huge number of connection weights. However,
F (·) and G(·) are countable functions that are discrete and
neither convex or concave, and they are not well addressed in
practice. Furthermore, because the gradient-based optimizers
are heavily dependent on the initial values of the weights
(including biases), it is essential to choose a suitable G(·)
that can help the consecutive GD-based approaches to escape
from local minima. Furthermore, the performance of assigned
architectures cannot be evaluated until the minimization of L(·)
is finished, while the minimization is a progress of multiple
iterations, which in turn increases the difficulty of choosing the
I. I NTRODUCTION
potential optimal F (·). Therefore, the architecture design and
Convolutional neural networks (CNNs) have demonstrated
connection weight initialization strategy should be carefully
their exceptional superiority in visual recognition tasks, such
treated in CNNs.
as traffic sign recognition [1], biological image segmentaTypically, most of the architecture design approaches were
tion [2], and image classification [3]. CNNs are originally
initially developed for the deep learning algorithms in the early
motivated by the computational model of the cat visual cortex
date (e.g., the Stacked Auto-Encoders (SAEs) [10], [11] and
specializing in processing vision and signal related tasks [4].
the Deep Belief Networks (DBNs) [12]), such as Grid Search
Since LetNet-5 was proposed in 1989 [5], [6], which is an
(GS), Random Search (RS) [13], Bayesian-based Gaussion
implementation of CNNs and whose connection weights are
Process (BGP) [14], [15], Tree-structured Parzen Estimators
optimized by the Back-Propagation (BP) algorithm [7], various
(TPE) [16], and Sequential Model-Based Global Optimizavariants of CNNs have been developed, such as VGGNet [8]
tion (SMBO) [17]. Theoretically, GS exhaustively tests all
and ResNet [9]. These variants significantly improve the clascombinations of the parameters to expectedly seize the best
sification accuracies of the best rivals in image classification
one. However, GS cannot evaluate all combinations within an
tasks. Diverse variants of CNNs differ from their architectures
acceptable time in reality. Moreover, GS is difficult to optimize
and weight connections.
the parameters of continuous values [13]. RS could reduce the
Mathematically, a CNN can be formulated by (1) in the
exhaustive adverse of GS, but the absolute “random” severely
context of an image classification task with the input (X, Y ),
challenges the sampling behavior in the search space [18]–[20].
In addition, BGP incurs extra parameters (i.e., the kernels) that
Architecture = F (X, Y )
Weight = G(Architecture )
(1) are arduous to tune. TPE treats each parameter independently,
while the most key parameters in CNNs are dependent (e.g.,
minimize L(X, Weight , Y )
Abstract—Evolutionary computation methods have been successfully applied to neural networks since two decades ago, while
those methods cannot scale well to the modern deep neural networks due to the complicated architectures and large quantities
of connection weights. In this paper, we propose a new method
using genetic algorithms for evolving the architectures and
connection weight initialization values of a deep convolutional
neural network to address image classification problems. In the
proposed algorithm, an efficient variable-length gene encoding
strategy is designed to represent the different building blocks and
the unpredictable optimal depth in convolutional neural networks.
In addition, a new representation scheme is developed for effectively initializing connection weights of deep convolutional neural
networks, which is expected to avoid networks getting stuck into
local minima which is typically a major issue in the backward
gradient-based optimization. Furthermore, a novel fitness evaluation method is proposed to speed up the heuristic search
with substantially less computational resource. The proposed
algorithm is examined and compared with 22 existing algorithms
on nine widely used image classification tasks, including the stateof-the-art methods. The experimental results demonstrate the
remarkable superiority of the proposed algorithm over the stateof-the-art algorithms in terms of classification error rate and the
number of parameters (weights).
Index Terms—Genetic algorithms, convolutional neural network, image classification, deep learning.
the convolutional layer size and its stride, more details can be
seen in Subsection II-A). The methods mentioned above have
shown their good performance in most SAEs and DBNs, but
are not suitable to CNNs. Their success in SAEs and DBNs is
due largely to the architecture construction approaches, which
are greedy layer-wised by stacking a group of building blocks
with the same structures (i.e., the three-layer neural networks).
In each building block, these architecture-search methods are
utilized for only optimizing the parameters, such as the number
of neurons in the corresponding layer. However in CNNs, the
layer-wised method cannot be applied due to their architecture
characteristics of non-stacking routine, and we must confirm
the entire architectures at a time. Furthermore, multiple different building blocks exist in CNNs, and different orders
of them would result in significantly different performance.
Therefore, the architecture design in CNNs should be carefully
treated. Recently, Baker et al. [21] proposed an architecture
design approach for CNNs based on reinforcement learning,
named MetaQNN, which employed 10 Graphic Processing
Unit (GPU) cards with 8-10 days for the experiments on the
CIFAR-10 dataset.
Due to the drawbacks of existing methods and limited
computational resources available to interested researchers,
most of these works in CNNs are typically performed by
experts with rich domain knowledge [13]. Genetic Algorithms
(GAs), which are a paradigm of the evolutionary algorithms
that do not require rich domain knowledge [22], [23], adapt
the meta-heuristic pattern motivated by the process of natural
selection [24] for optimization problems. GAs are preferred in
various fields due to their characteristics of gradient-free and
insensitivity to local minima [25]. These promising properties
are collectively achieved by a repeated series of the selection, mutation, and crossover operators. Therefore, it can be
naturally utilized for the optimization of architecture design
and the connection weight initialization for CNNs. Indeed,
GAs for evolving neural networks can be traced back to
1989 [26]. In 1999, Yao [25] presented a survey about these
different approaches, which are largely for the optimization
of connection weights in the fixed architecture topologies.
In 2002, Stanley and Miikkulainen proposed the NeuronEvolution Augmenting Topology (NEAT) [27] algorithm to
evolve the architecture and connection weights of a small
scale neural network. Afterwards, the HyperNEAT [28], i.e.,
NEAT combined with the compositional pattern producing
networks [29], was proposed to evolve a larger scale neural
network with an indirect encoding strategy. Motivated by
the HyperNEAT, multiple variants [30]–[32] were proposed
to evolve even larger scale neural networks. However, the
major deficiencies of the HyperNEAT-based approaches are:
1) they are only suitable for evolving deep neural networks
with global connection and single building blocks, such as
SAEs and DBNs, but not CNNs where local connection exists
and multiple different building blocks need to be evolved
simultaneously, and 2) hybrid weigh connections (e.g., connections between two layers that are non-adjacent) may be
evolved, which are contrast to the architectures of CNNs.
Indeed, the views have been many years that evolutionary
algorithms are incapable of evolving the architecture and
connection weights in CNNs due to the tremendous number
of related parameters [33]–[35]. Until very recently in 2017,
Google showed their Large Evolution for Image Classification
(LEIC) method specializing at the architecture optimization
of CNNs [36]. LETC is materialized by GAs without the
crossover operator, implemented on 250 high-end computers,
and archives competitive performance against the state-ofthe-art on the CIFAR-10 dataset by training for about 20
days. Actually, by directly using GAs for the architecture
design of CNNs, several issues would raise in nature: 1)
the best architecture is unknown until the performance is
received based on it. However, evaluating the performance of
one individual takes a long time, and appears more severely
for the entire population. This would require much more
computational resources for speeding up the evolution; 2)
the optimal depth of CNNs for one particular problem is
unknown, therefore it is hard to constrain the search space
for the architecture optimization. In this regard, a variablelength gene encoding strategy may be the best choice for
both 1) and 2), but how to assign the crossover operation for
different building blocks is a newly resulted problem; and 3)
the weight initialization values heavily affect the performance
of the confirmed architecture, but addressing this problem
involves a good gene encoding strategy and the optimization
of hundreds and thousands decision variables.
A. Goal
The aim of this paper is to design and develop an effective
and efficient GA method to automatically discover good architectures and corresponding connection weight initialization
values of CNNs (i.e., the first two formulae in (1)) without
manual intervention. To achieve this goal, the objectives below
have been specified:
1) Design a flexible gene encoding scheme of the architecture, which does not constrain the maximal length of
the building blocks in CNNs. With this gene encoding
scheme, the evolved architecture is expected to benefit
CNNs to achieve good performance in solving different
tasks at hand.
2) Investigate the connection weight encoding strategy,
which is capable of representing tremendous numbers of
the connection weights in an economy way. With this
encoding approach, the weight connection initialization
problem in CNNs is expected to be effectively optimized
by the proposed GA.
3) Develop associated selection (including the environmental selection), crossover, and mutation operators that can
cope with the designed gene encoding strategies of both
architectures and connection weights.
4) Propose an effective fitness measure of the individuals
representing different CNNs, which does not require
intensive computational resources.
Input
24x24
Feature maps
4@20x20
Feature maps
4@10x10
Feature maps
8@8x8
Feature maps
8@4x4
Full connection layer
128x1
Flatten
Convolution
Pooling
Convolution
Pooling
Fig. 1. An general architecture of the Convolutional Neural Network.
5) Investigate whether the new approach significantly outperform the existing methods in both classification accuracy
and number of weights.
B. Organization
the feature map and the input data through padding zeros,
the convolutional operations are categorized into two types:
the VALID (without padding) and the SAME (with padding).
Specifically, each element in the feature map is the sum of the
products of each element from the filter and the corresponding
elements this filter overlaps. If the input data is with multiple
channels, say 3, one feature map will also require 3 different
filters, and each filter convolves on each channel, then the
results are summed element-wised.
1
0
0
1
1
0
1
0
1
1
1
1
0
1
0
0
1
2
2
0
0
1
1
0
1
0
1
1
1
1
0
1
2
2
The reminder of this paper is organized as follows: the Filter {width:2, height:2}
2
2
1
1
1
1
1
1
1
1
1
Feature map
background of the CNNs, the related works on the architecture Stride {width:1, height:1}
Input 4x4
3x3
Input 4x4
design and weight initialization approaches of CNNs are
reviewed in Section II. The framework and the details of each
Fig. 2. An illustration of convolutional operation.
step in the proposed algorithm are elaborated in Section III.
The experiment design and experimental results of the proAn example of the VALID convolutional operation is illusposed algorithm are shown in Sections IV and V, respectively. trated by Fig. 2, where the filter has the same width and height
Next, further discussions are made in Section VI. Finally, the equal to 2, the input data is with the size of 4 × 4, and the
conclusions and future work are detailed in Section VII.
stride has the same height and width equal to 1. As shown in
Fig. 2, the generated feature map is with the size of 3 × 3, the
II. BACKGROUND AND R ELATED W ORK
shadow areas in the input data with different colors refer to
A. Architecture of Convolutional Neural Network
the overlaps with the filter at different positions of the input
1) Bone of CNNs: Fig. 1 exhibits an extensive architecture data, the shadow areas in the feature map are the respective
of one CNN, where there are two convolutional operations, resulted convolutional outcomes, and numbers in the filter are
two pooling operations, the resulted four groups of feature the connection weight values. Generally, convolutional results
maps, and the full connection layer in the tail. The last layer, regarding each filter are updated by adding a bias term and
which is a full connection layer, receives the input data by then through a nonlinearity, such as the Rectifier Linear Unit
flattening all elements of the fourth group of feature maps. (ReLU) [37], before they are stored into the feature map. ObviGenerally, the convolutional layers and the pooling layers can ously, the involved parameters in one convolutional operation
be mixed to stack together in the head of the architecture, are the filter width, the filter height, the number of feature
while the full connection layers are constantly stacked with maps, the stride width, the stride height, the convolutional
each other in the tail of the architecture. The numbers in Fig. 1 type, and the connection weight in the filter.
refer to the sizes of the corresponding layer. Particularly, the
3) Pooling: Intuitively, the pooling operation resembles the
input data is with 24 × 24, the output is with 128 × 1, and convolution operation except for the element-wised product
the other numbers denote the feature map configurations. For and the resulted values of the corresponding feature map.
example, 4@20 × 20 implies there are 4 feature maps, each Briefly, the pooling operation employs a predefined window
with the size of 4 × 4.
(i.e., the kernel) to collect the average value or the maximal
In the following, the details of the convolutional layer and value of the elements where it slides, and the slide size is
the pooling layer, which are associated with the convolution also called “stride” as in the convolutional operation. For
and the pooling operations, respectively, are documented in better understanding, an example of the pooling operation is
detail, while the full connection layer is not intended to illustrated by Fig. 3, where the kernel is with the size of 2 × 2,
describe here because it is well-known.
and the both stride width and height are 2, the input data
2) Convolution: Given an input image with the size of is with the size of 4 × 4, the shadows with different colors
n × n, in order to receive a feature map generated by the refer to the two slide positions and the resulted pooling values.
convolutional operations, a filter must be defined in advance. In this example, the maximal pooling operation is employed.
Actually, a filter (it can also be simply seen as a matrix) is Evidently in the pooling operation, the involved parameters
randomly initialized with a predefined size (i.e., the filter width are the kernel width, the kernel height, the stride width, the
and the filter height). Then, this filter travels from the leftmost stride height, and the pooling type.
to the rightmost of the input data with the step size equal to
a stride (i.e., the stride width), and then travels again after B. CNN Architecture Design Algorithm
In this subsection, only LEIC [36] is concerned due to its
moving downward with the step size equal to a stride (i.e.,
the stride height), until reach the right bottom of the input specific intention for the architecture design of CNNs. To this
image. Depending on whether to keep the same sizes between end, we will review the algorithm and then point out the
first initialization method. In the second initialization approach,
the shortage of the first one has been solved, but the major diffi0
1
1
1
0
1
1
1
1
1
culty exists in choosing the parameter of the distribution, such
1
1
0
1
1
1
0
1
1
1
as the range of the uniform distribution, and the mean value as
Kernel {width:2, height:2}
1
1
1
1
1
1
1
1
Stride {width:2, height:2}
Feature map
well as the standard derivation of the Gaussian distribution. To
Input 4x4
2x2
Input 4x4
solve this problem, the Xavier initializer presented a range for
uniform sampling based on the neuron saturation prior using
Fig. 3. An illustration of pooling operation.
the sigmoid activation function [39]. Supposed the number of
neurons in two adjacent layers are n1 and n2 , the values of
limitations, which are used to highlight the necessity of the the weights connecting
these two layers are initialized
within
i
h p
p
corresponding work in the proposed algorithm.
the range of − 6/(n1 + n2 ), 6/(n1 + n2 ) by uniformly
LEIC employed a GA to evolve the architectures of CNNs, sampling. Although the Xavier initializer works better than the
where individuals were evolved from scratch, and each individ- initialization methods from the other two categories, a couple
ual was encoded with a variable-length chromosome. During of major issues exist: 1) It highly relies on the architectures of
the phase of evolution, different kinds of layers can be incor- CNNs, particularity the number of neurons in each layer in the
porated into the individuals by the mutation operation, with networks (e.g., n and n in its formulation). If the optimal
1
2
the expectation that individuals with promising performance architectures of the networks are not found, the resulted
will be generated. Note that, the crossover operation, which initialized parameters perform badly as well, then there is no
are designed for local search in GAs, were not investigated way to evaluate the desired performance of the architectures
in the main part of LEIC. Without any crossover operator, and may mislead the adjustment of the architectures. 2) The
GAs typically require a large population size for reducing Xavier initializer is presented on the usage of the sigmoid
the adverse impact. Therefore, LETC adopted a population activation, while the widely used activation function in CNNs
size of 103 , while that of a magnitude with order 102 is is the ReLU [3], [9], [40], [41].
a general setting in the GA community. In addition, 250
To the best of our knowledge, there has not been any existhigh-end computers were employed for LEIC, which was ing evolutionary algorithm for searching for the connection
caused by the fitness assignment approach utilized in LEIC. weight initialization of deep learning algorithms including
In LEIC, each individual was evaluated with the final image CNNs. The main reason is the tremendous numbers of weights,
classification accuracy, which typically took a long time due to which are difficult to be effectively encoded into the chromothe iterative nature of GD-based optimizers. However, LEIC somes and to be efficiently optimized due to its large-scale
reported an external experiment on the crossover operation, global optimization nature [42].
which was used only for exchanging the mutation probabilities
III. T HE PROPOSED ALGORITHM
and the connection weights trained by SGD. Although it is
In this section, the proposed Evolving deep Convolutional
difficult to reach the actual reason why the crossover has not
been used for evolving the architectures in LETC, it is obvious Neural Networks (EvoCNN) for image classification is docuat least that the crossover operations are not easy to achieve mented in detail.
for chromosomes with different lengths. This would be more A. Algorithm Overview
complicated in CNNs that have multiple different building
blocks.
In summary, the main deficiency of LEIC is its high com- Algorithm 1: Framework of EvoCNN
putational complexity mainly caused by the fitness evaluation 1 P0 ←Initialize the population with the proposed gene
encoding strategy;
strategy it adopts as well as without crossover operations. The
2
t
← 0;
huge computational resource that LEIC requires makes it very
3
while
termination criterion is not satisfied do
intractable in academic environment.
4
Evaluate the fitness of individuals in Pt ;
C. Connection Weight Initialization
5
S ← Select parent solutions with the developed slack
binary tournament selection;
Typically, the initialization methods are classified into three
6
Q
t ← Generate offsprings with the designed genetic
different categorises. The first employs constant numbers to
operators
from S;
initialize all connection weights, such as the zero initializer,
7
P
←Environmental
selection from Pt ∪ Qt ;
t+1
one initializer, and other fixed value initializer. The second is
8
t
←
t
+
1;
the distribution initializer, such as using the Gaussion distribution or uniform distribution to initialize the weights. The third 9 end
covers the initialization approaches with some prior knowl- 10 Select the best individual from Pt and decode it to the
corresponding convolutional neural network.
edge, and the famous Xavier initializer [38] belongs to this
category. Because of the numerous connection weights existAlgorithm 1 outlines the framework of the proposed
ing in CNNs, it is not necessary that all the connection weights
start with the same values, which is the major deficiency of the EvoCNN method. Firstly, the population is initialized based on
max(1,0,1,1)
0
1
1
0
max(0,1,1,0)
0
1
1
0
C
P
C
C
C
P
C
P
C
Convolutional
C
C
P
F
...
F
F
F
...
P
...
C
Pooling
P
Full connection
F
Fig. 4. An example of three chromosomes with different lengths in EvoCNN.
TABLE I
T HE ENCODED INFORMATION IN E VO CNN.
Unit Type
convolutional
layer
pooling layer
full connection
layer
Encoded Information
the filter width, the filter height, the number of feature
maps, the stride width, the stride height, the convolutional type, the standard deviation and the mean value
of filter elements
the kernel width, the kernel height, the stride width, the
stride height, and the pooling type (i.e., the average or
the maximal)
the number of neurons, the standard deviation of
connection weights, and the mean value of connection
weights
the proposed flexible gene encoding strategy (line 1). Then, the
evolution begins to take effect until a predefined termination
criterion, such as the maximum number of the generations,
has been satisfied (lines 3-9). Finally, the best individual is
selected and decoded to the corresponding CNN (line 10) for
final training.
During the evolution, all individuals are evaluated first based
on the proposed efficient fitness measurement (line 4). After
that, parent solutions are selected by the developed slack
binary tournament selection (line 5), and new offspring are
generated with the designed genetic operators (line 6). Next,
representatives are selected from the existing individuals and
the generated offsprings to form the population in the next
generation to participate subsequent evolution (line 7). In
the following subsections, the key steps in Algorithm 1 are
narrated in detail.
B. Gene Encoding Strategy
Table I. Commonly, hundreds of thousands connection weights
may exist in one convolutional or full connection layer, which
cannot be all explicitly represented by a chromosome and
effectively optimized by GAs. Therefore, in EvoCNN, we use
only two statistical real numbers, the standard derivation and
mean value of the connection weights, to represent the numerous weight parameters, which can be easily implemented by
GAs. When the optimal mean value and the standard derivation
are received, the connection weights are sampled from the
corresponding Gaussian distribution. The details of population
initialization in EvoCNN are given in the next subsection
based on the gene encoding strategy introduced above.
C. Population Initialization
Algorithm 2: Population Initialization
Input: The population size N , the maximal number of
convolutional and pooling layers Ncp , and the
maximal number of full connection layers Nf .
Output: Initialized population P0 .
1 P0 ← ∅;
2 while |P0 | ≤ N do
3
part1 ← ∅;
4
ncp ← Uniformaly generate an integer between
[1, Ncp ];
5
while |part1 | ≤ ncp do
6
r ← Uniformly generated a number between
[0, 1];
7
if r ≤ 0.5 then
8
l ← Initialize a convolutional layer with
random settings;
9
else
10
l ← Initialize a pooling layer with random
settings;
11
end
12
part1 ← part1 ∪ l;
13
end
14
part2 ← ∅;
15
nf ← Uniformaly generate an integer between
[1, Nf ];
16
while |part2 | ≤ nf do
17
l ← Initialize a full connection layer with random
settings;
18
part2 ← part2 ∪ l;
19
end
20
P0 ← P0 ∪ (part1 ∪ part2 );
21 end
22 Return P0 .
As introduced in Subsection II-A, three different building
blocks, i.e., the convolutional layer, the pooling layer, and the
full connection layer, exist in the architectures of CNNs. Therefore, they should be encoded in parallel into one chromosome
for evolution. Because the optimal depth of a CNN in solving
one particular problem is unknown prior to confirming its
architecture, the variable-length gene encoding strategy, which
is very suitable for this occasion, is employed in the proposed
EvoCNN method. Furthermore, because the performance of
CNNs is highly affected by their depths [8], [43]–[45], this
For convenience of the elaboration, each chromosome is
variable-length gene encoding strategy makes the proposed
EvoCNN method have chances to reach the best result due separated into two parts. The first part includes the convolutional layers and the pooling layers, and the other part is the
to no constrains on the architecture search space.
In particular, an example of three chromosomes with dif- full connection layers. Based on the convention of the CNN
ferent lengths from EvoCNN is illustrated by Fig. 4, and all architectures, the first part starts with one convolutional layer.
represented information in these three layers are detailed in The second part can be added to only at the tail of the first
part. In addition, the length of each part is set by randomly evaluation in EvoCNN. Because EvoCNN concerns on solving
choosing a number within a predefined range.
image classification tasks, the classification error is the best
Algorithm 2 shows the major steps of the population initial- strategy to assign their fitness. The number of connection
ization, where |·| is a cardinality operator, lines 3-13 show the weights is also chosen as an additional indicator to measure
generation of the first part of one chromosome, and lines 14- the individual’s quality based on the principle of Occam’s
19 show that of the second part. During the initialization of razor [46].
the first part, a convolutional layer is added first. Then, a
With the conventions, each represented CNN is trained
convolutional layer or a pooling layer is determined by the on the training set Dtrain , and the fitness is estimated on
once coin tossing probability and then added to the end, which another dataset Df itness 1 . CNNs are frequently with deep
is repeated until the predefined length of this part is met. For architectures, thus thoroughly training them for receiving the
the second part, full connection layers are chosen and then final classification error would take considerable expenditure
added. Note here that, convolutional layers, pooling layers, of computing resource and a very long time due to the
and full connection layers are initialized with the random large number of training epochs required (>100 epochs are
settings, i.e., the information encoded into them are randomly invariably treated in fully training CNNs). This will make it
specified before they are stored into the corresponding part. much more impracticable here due to the population-based
After these two parts finished, they are combined and returned GAs with multiple generations (i.e., each individual will take a
as one chromosome. With the same approach, a population of full training in each generation). We have designed an efficient
individuals are generated.
method to address this concern. In this method, each individual
is trained with only a small number of epochs, say 5 or 10
D. Fitness Evaluation
epochs, based on their architectures and connection weight
initialization values, and then the mean value and the standard
derivation
of classification error are calculated on each batch
Algorithm 3: Fitness Evaluation
of
D
f itness in the last epoch. Both the mean value and the
Input: The population Pt , the training epoch number k
standard
derivation of classification errors, are employed as the
for measuring the accuracy tendency, the training
fitness
of
one individual. Obviously, the smaller mean value,
set Dtrain , the fitness evaluation dataset Df itness ,
the
better
individual.
When the compared individuals are with
and the batch size num of batch.
the
same
mean
values,
the less standard derivation indicates
Output: The population with fitness Pt .
the
better
one.
1 for each individual s in Pt do
In summary, three indicators are used in the fitness evalu2
i ← 1;
ation,
which are the mean value, standard derivation, and the
3
eval steps ← |Df itness |/num of batch;
number
of parameters. There are several motivations behind
4
while i ≤ k do
this
fitness
evaluation strategy: 1) It is sufficient to investigat5
Train the connection weights of the CNN
ing
only
the
tendency of the performance. If individuals are
represented by individual s;
with
better
performance
in the first several training epochs of
6
if i == k then
CNNs,
they
will
probably
still have the better performance in
7
accy list ← ∅;
the
following
training
epochs
with greater confidence. 2) The
8
j ← 1;
mean
value
and
the
standard
derivation are statistical signifi9
while j ≤ eval steps do
cance
indicators,
thus
suitable
for investigating this tendency,
10
accyj ← Evaluate the classification error
and
the
final
classification
error
can be received by optimizing
on the j-th batch data from Df itness ;
only
the
best
individual
evolved
by the proposed EvoCNN
11
accy list ← accy list ∪ accyj ;
method.
3)
CNN
models
with
less
number of connection
12
j ← j + 1;
weights
are
preferred
by
smart
devices
(more details are
13
end
discussed
in
Section
VI).
14
Calculate the number of parameters in s, the
mean value and standard derivation from
E. Slack Binary Tournament Selection
accy list, assign them to individual s, and
update s from Pt ;
We develop one slack version of the standard binary tour15
end
nament selection, which are documented in Algorithm 4, to
16
i ← i + 1;
select parent solutions for the crossover operations in the
17
end
proposed EvoCNN method. Briefly, two sets of comparisons
18 end
are employed. The comparisons between the mean values of
19 Return Pt .
individuals involves a threshold α, and that comparisons between the parameter numbers involves another threshold β. If
Fitness evaluation aims at giving a quantitative measure
determining which individuals qualify for serving as parent
solutions. Algorithm 3 manifests the framework of the fitness
1 The original training set is randomly split into D
train and Df itness ,
where Df itness is unseen to the CNN training phase, which can give a good
indication of the generalization accuracy on the test set.
units, i.e., the convolutional layer, the pooling layer, and the
full connection layer, are firstly collected into three different
lists based on their orders in the corresponding chromosome,
which refers to the Unit Collection (UC) phase. Then, these
three lists are aligned at the top, and units at the same positions
are performed the crossover. This phase is named the UA
and crossover phase. Finally, the Unit Restore (UR) phase is
employed, i.e., when the crossover operation is completed, the
units in these lists are restored to their original positions of
the associated chromosomes. With these three consequential
phases (i.e., the UC, the UA and crossover, and the UR), two
chromosomes with different lengths could easily exchange
their gene information for crossover. Because the crossover
operation is performed on the unit lists where only the units
with the same types are loaded, this proposed UA crossover
operation is natural (because they have the same origins).
For the remaining units, which do not perform crossover
operations due to no paired ones, are kept at the same position.
Mutation operations may perform on each position of the
units from one chromosome. For a selected mutation point, a
unit could be added, deleted, or modified, which is determined
by a probability of 1/3. In the case of unit addition, a convolutional layer, a pooling layer, or a full connection layer is added
by taking a probability of 1/3. If the mutation is to modify an
existing unit, the particular modification is dependent on the
unit type, and all the encoded information in the unit would be
changed (encoded information on each unit type can be seen in
Table I). Note that all the encoded formation is denoted by real
the parent solution cannot be selected with these comparisons,
numbers, therefore the Simulated Crossover (SBX) [47] and
the individual with smaller standard derivation is chosen.
the polynomial mutation [48] are employed in the proposed
In practice, tremendous number of parameters exist in deep
EvoCNN method due to their notable show in real number
CNNs, which would easily cause the overfitting problem.
gene representations.
Therefore, when the difference between the mean values of
two individuals is smaller than the threshold α, we further con- G. Environmental Selection
sider the number of connection weights. The slight change of
the parameter numbers will not highly affect the performance
Algorithm 5: Environmental Selection
of CNNs. Consequently, β is also introduced.
Input: The elistsm fraction γ, and the current population
By iteratively performing this selection, parent solutions
Pt ∪ Qt .
are selected and stored into a mating pool. In the proposed
Output:
The selected population Pt+1 .
EvoCNN method, the size of the mating pool is set to be the
1 a ← Calculate the number of elites based on γ and the
same of the population size.
predefined population size N from Algorithm 2;
F. Offspring Generation
2 Pt+1 ← Select a individuals that have the best mean
values from Pt ∪ Qt ;
The steps for generating offspring are given as follows:
3 Pt ∪ Qt ← Pt ∪ Qt − Pt+1 ;
step 1): randomly select two parent solutions from the mating
4 while |Pt+1 | < N do
pool;
5
s1 , s2 ← Randomly select two individuals from
step 2): use crossover operator on the selected solutions to
Pt ∪ Qt ;
generate offspring;
6
s ← Employe Algorithm 4 to select one individual
step 3): use mutation operator on the generated offspring;
from s1 and s2 ;
step 4): store offspring, remove the parent solutions from the
7
Pt+1 ← Pt+1 ∪ s;
mating pool, and perform steps 1-3 until the mating
8 end
pool is empty.
9 Return Pt+1 .
The proposed crossover operation can be seen in Fig. 5. To
achieve crossover, we design a method called Unit Alignment
(UA) for recombining two individuals with different chromoThe environmental selection is shown in Algorithm 5. Dursome lengths. During the crossover operation, three different ing the environmental selection, the elitism and the diversity
Algorithm 4: Slack Binary Tournament Selection
Input: Two compared individuals, the mean value
threshold α, and the paramemter number
threshold β.
Output: The selected individual.
1 s1 ← The individual with larger mean value;
2 s2 ← The other individual;
3 µ1 , µ2 ← The mean values of s1 , s2 ;
4 std1 , std2 ← The standard derivations of s1 , s2 ;
5 c1 , c2 ← The parameter numbers of s1 , s2 ;
6 if µ1 − µ2 > α then
7
Return s1 .
8 else
9
if c1 − c2 > β then
10
Return s2 .
11
else
12
if std1 < std2 then
13
Return s1 .
14
else if std1 > std2 then
15
Return s2 .
16
else
17
Return randon one from {s1 , s2 }.
18
end
19
end
20 end
Convolutional
Pooling
C
Full connection
P
Chromosome #1
C1
P1
C2
Convolutional
unit list
C3
F
Chromosome #2
P2
Pooling
unit list
F1
F2
F3
C1
Full connection
unit list
P1
C2
P2
Convolutional
unit list
P3
F1
Pooling
unit list
F2
F3
Full connection
unit list
C1
P1
F1
C1
P1
F1
C2
P2
F2
C2
P2
F2
P3
F3
C3
F3
F4
F4
(a) Unit Collection
crossover
crossover
C1
C1
crossover
P1
crossover
P1
F1
P2
F2
P3
F3
F1
crossover
C2
C2
crossover
P2
C3
F2
crossover
F3
F4
(b) Unit Aligh and Crossover
Offspring #2
Offspring #1
C1
P1
C2
C3
P2
F1
F2
F3
C1
P1
C2
P2
P3
F1
F2
F3
F4
(c) Unit Restore
Fig. 5. An example to illustrate the entire crossover process. In this example, the first chromosome is with length 8 including three convolutional layers, two
pooling layers, and three full connection layers; the other one is with length 9 including two convolutional layers, three pooling layers, and four full connection
layers. In the first step of crossover (i.e., the unit collection), the convolutiona layers, the pooling layers, and the full connection layers are collected from
each chromosome and stacked with the same orders to them in each chromosome (see Fig. 5a). In the second step, the unit lists with the same unit types
are aligned at the top, i.e., the two convolutional layer lists are picked and aligned, and the same operations on the other two lists. When these unit lists
finish the alignment, units at the same positions from the each two lists are paired and performed crossover operation, which can be shown in Fig. 5b. At last,
units from these unit lists are restored based on the positions where they are from (see Fig. 5b). Note in Fig. 5b, the units that have experienced crossover
operations are highlighted with italics and underline fonts, while the units that do not perform the crossover operations remain the same.
are explicitly and elaborately addressed. To be specific, a fraction of individuals with promising mean values are chosen first,
and then the remaining individuals are selected by the modified
binary tournament selection demonstrated in Subsection III-F.
By these two strategies, the elitism and the diversity are
considered simultaneously, which are expected to collectively
improve the performance of the proposed EvoCNN method.
Note here that, the selected elites are removed before the
tournament selection gets start to work (line 3 of Algorithm 5),
while the individuals selected for the purpose of diversity are
kept in the current population for the next round of tournament
selection (lines 4-8 of Algorithm 5) based on the convention
in the community.
and then the decoded CNN will be deeply trained with a larger
number of epochs by SGD for future usage.
IV. E XPERIMENT D ESIGN
In order to quantify the performance of the proposed
EvoCNN, a series of experiments are designed and performed
on the chosen image classification benchmark datasets, which
are further compared to state-of-the-art peer competitors. In the
following, these benchmark datasets are briefly introduced at
first. Then, the peer competitors are given. Finally, parameter
settings of the proposed EvoCNN method participating into
these experiments are documented.
A. Benchmark Datasets
H. Best Individual Selection and Decoding
In these experiments, nine widely used image classification
At the end of evolution, multiple individuals are with benchmark datasets are used to examine the performance of
promising mean values but different architectures and connec- the proposed EvoCNN method. They are the Fashion [49], the
tion weight initialization values. In this regard, there will be Rectangle [50], the Rectangle Images (RI) [50], the Convex
multiple choices to select the “Best Individual”. For example, Sets (CS) [50], the MNIST Basic (MB) [50], the MNIST
if we are only concerned with the best performance, we with Background Images (MBI) [50], Random Background
could neglect their architecture configurations and consider (MRB) [50], Rotated Digits (MRD) [50], and with RD plus
only the classification accuracy. Otherwise, if we emphasis Background Images (MRDBI) [50] benchmarks.
Based on the types of the classification objects, these
the smaller number of parameters, corresponding decision
could be made. Once the “Best Individual” is confirmed, benchmarks are classified into three different categories. The
the corresponding CNN is decoded based on the encoded first category includes only Fashion, which is for recognizing
architecture and connection weight initialization information, 10 fashion objects (e.g., trousers, coats, and so on) in the
50, 000 training images and 10, 000 test images. The second final classification accuracy and is invisible to public. As for
one is composed of the MNIST [6] variants including the other architecture design approaches introduced in Section I,
MB, MBI, MRB, MRD, and the MRDBI benchmarks for such as GS, RS, among others, they are not scalable to CNNs
classifying 10 hand-written digits (i.e., 0-9). Because the and not suitable to be directly compared as well.
MNIST has been easily achieved 97%, these MNIST variants
In the experiments, state-of-the-art algorithms that have
are arbitrarily added into different barriers (e.g., random back- reported promising classification errors on the chosen benchgrounds, rotations) from the MNIST to improve the complexity marks are collected as the peer competitors of the proof classification algorithms. Furthermore, these variants have posed EvoCNN method. To be specific, the peer com12, 000 training images and 50, 000 test images, which further petitors on the Fashion benchmark are collected from the
challenges the classification algorithms due to the mush less dataset homepage2. They are 2C1P2F+Dropout, 2C1P, 3C2F,
training data while more test data. The third category is for 3C1P2F+Dropout, GRU+SVM+Dropout, GoogleNet [41],
recognizing the shapes of objects (i.e., the rectangle or not for AlexNet [3], SqueezeNet-200 [51], MLP 256-128-64, and
the Rectangle and RI benchmarks, and the convex or not for VGG16 [52], which perform the experiments on the raw
the Convex benchmark). Obviously, this category covers the dataset without any preprocessing. The peer competitors on
Rectangle, the RI, and the CS benchmarks that contain 1, 200, other benchmarks are CAE-2 [53], TIRBM [54], PGBM+DN12, 000, and 8, 000 training images, respectively, and all of 1 [55], ScatNet-2 [56], RandNet-2 [57], PCANet-2 (softthem include 50, 000 test images. Compared to the Rectangle max) [57], LDANet-2 [57], SVM+RBF [50], SVM+Poly [50],
benchmark, the RI is generated by adding randomly sampled NNet [50], SAA-3 [50], and DBN-3 [50], which are from
backgrounds from the MNIST for increasing the difficulties the literature [57] recently published and the provider of the
of classification algorithms.
benchmarks3.
In addition, each image in these benchmarks is with the size
28 × 28, and examples from these benchmarks are shown in C. Parameter Settings
Fig. 6 for reference. Furthermore, another reason for using
All the parameter settings are set based on the conventions
these benchmark datasets is that different algorithms have in the communities of evolutionary algorithms [58] and deep
reported their promising results, which is convenient for the learning [59], in addition to the maximal length of each basic
comparisons on the performance of the proposed EvoCNN layer. Specifically, the population size and the total generation
method and these host algorithms (details can be seen Subsec- number are set to be 100. The distribution index of SBX and
tion IV-B).
Polynomial mutation are both set to be 20, and their associated
probabilities are specified as 0.9 and 0.1, respectively. The
maximum lengths of the convolutional layers, the pooling
layers, and the full connection layers are set to be the same
(a) Examples from the first category. From left to right, they are T-shirt, trouser, as 5 (i.e., the maximal depths of CNNs in these experiments
pullover, dress, coat, sandal, shirt, sneaker, bag, and ankle boot.
are 15). Moreover, the proportion of the elitism is set to be
20% based on the Pareto principle. Note that the 20% data
are randomly selected from the training images as the fitness
evaluation dataset. For the implementation of the proposed
(b) Examples from the second category. From left to right, each two images
EvoCNN method, we constrain the same values of the width
are from one group, and each group is from MBi, MRB, MRD, MRDBI, and
MB, respectively. These images refer to the hand-written digits 0, 4, 2, 6, 0, and height for filter, stride, and kernel. The width and height
5, 7, 5, 9, and 6, respectively.
for stride in the convolutional layer are set to be 1, those in
the pooling layer are specified at the same value to its kernel,
and the convolutional type is fixed to “SAME”.
The proposed EvoCNN method is implemented by Tensor(c) Examples from the thrid category, From left to right, the first three images
flow [60], and each copy of the code runs in a computer
are from the Rectangle benchmark, the following four ones are from the RI
benchmark, and the remaingings are from the Convex benchmark. Specifically, equipped with two GPU cards with the identical model number
these examples with the index 1, 2, 6, 7, and 11 are positive samples.
GTX1080. During the final training phase, each individual
Fig. 6. Examples from the benchmarks chosen.
is imposed by the BatchNorm [61] for speeding up and the
weight decay with an unified number for preventing from
B. Peer Competitors
the overfitting. Due to the heuristic nature of the proposed
Ideally, the algorithms for neural network architecture de- EvoCNN method, 30 independent runs are performed on each
sign discussed in Section I should be collected here as the peer benchmark dataset, and the mean results are estimated for
competitors. However, because MetaQNN [21] and LEIC [36] the comparisons unless otherwise specified. Furthermore, the
highly rely on the computational resources, it is impossible experiments take 4 days on the Fashion benchmark for each
to reproduce the experimental results in an academic envi- run, 2–3 days on other benchmarks.
ronment. Furthermore, the benchmark dataset investigated in
2 https://github.com/zalandoresearch/fashion-mnist
these two algorithms employed different data preprocessing
3 http://www.iro.umontreal.ca/∼ lisa/twiki/bin/view.cgi/Public/DeepVsShallowComparisonIC
and augmentation techniques, which will highly enhance the
V. E XPERIMENTAL R ESULTS AND A NALYSIS
classification error rates, where the difference is only 0.2%.
The
proposed EvoCNN method further decreases the error rate
In this section, the experimental results and analysis of the
by
0.83%
to 5.47%. Furthermore, the mean performance of
proposed EvoCNN method against peer competitors are shown
EvoCNN
is
even better than the best performance of eight
in Subsection V-A. Then, the weight initialization method in
competitors,
and
only a little worse than the best of GoogleNet
the proposed EvoCNN method are specifically investigated in
and
VGG16.
However,
EvoCNN has a much smaller number
Subsection V-B.
of connection weights — EvoCNN uses 6.52 million weights
A. Overall Results
while GoogleNet uses 101 million and VGG uses 26 million
Experimental results on the Fashion benchmark dataset are weights. EvoCNN also employs only half of the numbers of
shown in Table II, and those on the MB, MRD, MRB, MBI, training epochs used in the VGG16. The results show that the
MRDBI, Rectangle, RI, and Convex benchmark datasets are proposed EvoCNN method obtains much better performance in
shown in Table III. In Tables II and III, the last two rows the architecture design and connection weight initialization of
denote the best and mean classification errors received from CNNs on the Fashion benchmark, which significantly reduces
the proposed EvoCNN method, respectively, and other rows the classification error of the evolved CNN generated by the
show the best classification errors reported by peer competi- proposed EvoCNN method.
According to Table III, EvoCNN is the best one among all
tors4 . In order to conveniently investigate the comparisons, the
terms “(+)” and “(-)” are provided to denote whether the best the 13 different methods. Specifically, EvoCNN achieves the
result generated by the proposed EvoCNN method is better or best performance on five of the eight datasets, and the second
worse that the best result obtained by the corresponding peer best on the other three datasets — the MB, MRD, and Convex
competitor. The term “—” means there is no available result datasets, where LDANet-2, TIRBM, and PCANet-2(softmax)
reported from the provider or cannot be counted. Most informa- achieve the lowest classification error rate, respectively. Furtion of the number of parameters and number of training epoch thermore, comparing the mean performance of EvoCNN with
from the peer competitors on the Fashion benchmark dataset the best of the other 12 methods, EvoCNN is the best on four
is available, therefore, such information from EvoCNN is also datasets (MRB, MBI, Rectangle and RI), second best on two
shown in Table II for multi-view comparisons. However, for (MRD and Convex), third best and fourth best on the other two
the benchmarks in Table III, such information is not presented datasets (MRDBI and MB). Particularly on MRB and MBI,
because they are not available from the peer competitors. the lowest classification error rates of the others are 6.08%
Note that all the results from the peer competitors and the and 11.5%, and the mean error rates of EvoCNN are 3.59%
proposed EvoCNN method are without any data augmentation and 4.62%, respectively. In summary, the best classification
error of the proposed EvoCNN method wins 80 out of the
preprocessing on the benchmarks.
84 comparisons against the best results from the 12 peer
competitors, and the mean classification error of EvoCNN is
TABLE II
T HE CLASSIFICATION ERRORS OF THE PROPOSED E VO CNN METHOD
better than the best error of the 12 methods on 75 out of the
AGAINST THE PEER COMPETITORS ON THE FASHION BENCHMARK
84 comparisons.
DATASET.
B. Performance Regarding Weight Initialization
It is clearly shown in Table II that by comparing the best
performance, the proposed EvoCNN method outperforms all
the ten peer competitors. The two state-of-the-art algorithms,
GoogleNet and VGG16, achieve respectively 6.5% and 6.3%
4 It is a convention in deep learning community that only the best result is
reported.
60
weight initializaed by EvoCNN
weight initialized by Xavier
50
40
30
0
Fashion
MB
MRD
MRB
MBI
MRDBI
Benchmark datasets
Rectangle
5.39
6.33
10
5.97
8.27
20
0.01
0.23
300
30
—
150
100
—
—
200
25
200
100
100
37.38
41.88
3.27M
100K
—
7.14M
—
101M
60M
500K
41K
26M
6.68M
6.52M
4.62
5.8
8.40(+)
7.50(+)
9.30(+)
7.40(+)
10.30(+)
6.30(+)
10.10(+)
10.00(+)
10.00(+)
6.50(+)
5.47
7.28
3.59
3.93
2C1P2F+Drouout
2C1P
3C2F
3C1P2F+Dropout
GRU+SVM+Dropout
GoogleNet [41]
AlexNet [3]
SqueezeNet-200 [51]
MLP 256-128-64
VGG16 [52]
EvoCNN (best)
EvoCNN (mean)
5.46
6.24
# epochs
1.28
2.45
# parameters
7.28
9.28
error(%)
Classification error (%)
classifier
RI
Convex
Fig. 7. Performance comparison between the proposed EvoCNN method and
the CNN using the evolved architecture and the Xavier weight initializer.
Further experiments are performed to examine the effectiveness of the connection weight initialization method in the
proposed EvoCNN approach. By comparing different weight
initializers, we would also investigate whether the architecture or the connection weight initialization values will affect
the classification performance. To achieve this, we initialize
another group of CNNs with the same architectures as that of
TABLE III
T HE CLASSIFICATION ERRORS OF THE PROPOSED E VO CNN METHOD AGAINST THE PEER COMPETITORS ON THE MB, MRD, MRB, MBI, MRDBI,
R ECTANGLE , RI, AND C ONVEX BENCHMARK DATASETS
classifier
MB
MRD
MRB
MBI
MRDBI
Rectangle
RI
Convex
CAE-2 [53]
TIRBM [54]
PGBM+DN-1 [55]
ScatNet-2 [56]
RandNet-2 [57]
PCANet-2 (softmax) [57]
LDANet-2 [57]
SVM+RBF [50]
SVM+Poly [50]
NNet [50]
SAA-3 [50]
DBN-3 [50]
EvoCNN (best)
EvoCNN (mean)
2.48(+)
—
—
1.27(+)
1.25(+)
1.40(+)
1.05(-)
3.03(+)
3.69(+)
4.69(+)
3.46(+)
3.11(+)
1.18
1.28
9.66(+)
4.20(-)
—
7.48(+)
8.47(+)
8.52(+)
7.52(+)
11.11(+)
15.42(+)
18.11(+)
10.30(+)
10.30(+)
5.22
5.46
10.90(+)
—
6.08(+)
12.30(+)
13.47(+)
6.85(+)
6.81(+)
14.58(+)
16.62(+)
20.04(+)
11.28(+)
6.73(+)
2.80
3.59
15.50(+)
—
12.25(+)
18.40(+)
11.65(+)
11.55(+)
12.42(+)
22.61(+)
24.01(+)
27.41(+)
23.00(+)
16.31(+)
4.53
4.62
45.23(+)
35.50(+)
36.76(+)
50.48(+)
43.69(+)
35.86(+)
38.54(+)
55.18(+)
56.41(+)
62.16(+)
51.93(+)
47.39(+)
35.03
37.38
1.21(+)
—
—
0.01(=)
0.09(+)
0.49(+)
0.14(+)
2.15(+)
2.15(+)
7.16(+)
2.41(+)
2.61(+)
0.01
0.01
21.54(+)
—
—
8.02(+)
17.00(+)
13.39(+)
16.20(+)
24.04(+)
24.05(+)
33.20(+)
24.05(+)
22.50(+)
5.03
5.97
—
—
—
6.50(+)
5.45(+)
4.19(-)
7.22(+)
19.13(+)
19.82(+)
32.25(+)
18.41(+)
18.63(+)
4.82
5.39
the evolved EvoCNN, but their weights are initialized with the
widely used Xavier initializer [38] (details in Subsection II-C).
The comparisons are illustrated in Fig. 7, where the xaxis denotes the benchmark datasets, and y-axis denotes the
classification error rates. Fig. 7 clearly shows that the proposed
connection weight initialization strategy in EvoCNN improves
the classification performance on all the benchmarks over the
widely used Xavier initializer. To be specific, the proposed
weight initialization method reaches ≈1.5% classification accuracy improvement on the Fashion, MB, MRD, MBI, RI, and
Convex benchmarks, and 4.5% on the MRDBI benchmark. By
comparing with the results in Tables II and III, it also can be
concluded that the architectures of CNNs contribute to the
classification performance more than that of the connection
weight initialization. The proposed EvoCNN method gained
promising performance by automatically evolving both the
initial connection weights and the architectures of CNNs.
operation in this context (because crossover operations can
not be easily performed between the genes from different
origins). Therefore, a new crossover operation is introduced
to the proposed EvoCNN method. UC, UA, as well as UR are
designed to complete the crossover operation, which enhances
the communications between the encoded information, and expectedly leverages the performance in pursuing the promising
architectures of CNNs.
The typically used approaches in optimizing the weights of
CNNs are based on the gradient information. It is well known
that the gradient-based optimizers are sensitive to the starting
position of the parameters to be optimized. Without a better
starting position, the gradient-based algorithms are easy to be
trapped into local minima. Intuitively, finding a better starting
position of the connection weights by GAs is intractable due
to the huge numbers of parameters. As we have elaborated, a
huge number of parameters can not be efficiently encoded into
the chromosomes, nor effectively optimized. In the proposed
VI. F URTHER D ISCUSSIONS
EvoCNN method, an indirect encoding approach is employed,
which
only encodes the means and standard derivations of
In this section, we will further discuss the encoding stratthe
weights
in each layer. In fact, employing the standard
egy of the architectures and the weights related parameters,
derivation
and
the mean to initialize the starting position of
and the fitness evaluation in the proposed EvoCNN method.
connection
weights
is a compromised approach and can be
Extra findings from the experimental results are discussed as
ubiquitously
seen
from
deep learning libraries [60], [62], [63],
well, which could provide insights on the applications of the
which
is
the
motivation
of
this design in the proposed EvoCNN
proposed EvoCNN method.
method.
With
this
strategy,
only two real numbers are used
In GAs, crossover operators play the role of exploitation
to
denote
the
hundreds
of
thousands
of parameters, which
search (i.e., the local search), and mutation operators act
could
save
much
computational
resource
for encoding and
the exploration search (i.e., the global search). Because the
optimization.
local search and the global search should complement to each
other, only well designing both of them could remarkably
Existing techniques for searching for the architectures of
promote the performance. Generally, the crossover operators CNNs typically take the final classification accuracy as the
operate on the chromosomes with the same lengths, while fitness of individuals. A final classification accuracy typically
the proposed EvoCNN method employs the variable-length requires many more epochs of the training, which is a very
ones. Furthermore, multiple different basic layers exist in time-consuming process. In order to complete the architecture
CNNs, which improve the difficulties of designing crossover design with this fitness evaluation approach, it is natural
to employ a lot of computation resource to perform this
task in parallel to speed up the design. However, computational resource is not necessarily available to all interested
researchers. Furthermore, it also requires extra professional
assistances, such as the task schedule and synchronization
in the multi-thread context, which is beyond the expertise
of most researchers/users. In fact, it is not necessary to
check individual final classification accuracy, but a tendency
that could predict their future quality would be sufficient.
Therefore, we only employ a small number of epochs to
train these individuals. With this kind of fitness measurement,
the proposed EvoCNN method does not highly reply on the
computational resource. In summary, with the well-designed
yet simplicity encoding strategies and this fitness evaluation
method, researchers without rich domain knowledge could
also design the promising CNN models for addressing specific
tasks in their (academic) environment, where is the computational resources are typically limited.
TABLE IV
T HE CLASSIFICATION ACCURACY AND THE PARAMETER NUMBERS FROM
THE INDIVIDUALS ON THE FASHION BENCHMARK DATASET.
classification accuracy
# parameters
99%
569,250
98%
98,588
97%
7,035
96%
3,205
95%
955
Moreover, interesting findings have also been discovered
when the proposed EvoCNN method terminates, i.e., multiple
individuals are with the similar performance but significant
different numbers of connection weights (See Table IV), which
are produced due to population-based nature of the proposed
EvoCNN method. Recently, various applications have been
developed, such as the auto-driving car and some interesting
real-time mobile applications. Due to the limited processing
capacity and battery in these devices, trained CNNs with
similar performance but fewer parameters are much more
preferred because they require less computational resource
and the energy consumption. In addition, similar performance
of the individuals are also found with different lengths of
the basic layers, and there are multiple pieces of hardware
that have been especially designed to speed the primitive
operations in CNNs, such as the specialized hardware for
convolutional operations. In this regard, the proposed EvoCNN
method could give manufacturers more choices to make a
decision based on their own preferences. If the traditional
approaches are employed here, it is difficult to find out such
models at a single run.
VII. C ONCLUSIONS
AND
F UTURE W ORK
The goal of the paper was to develop a new evolutionary approach to automatically evolving the architectures and weights
of CNNs for image classification problems. This goal has
been successfully achieved by proposing a new representation
for weight initialization strategy, a new encoding scheme
of variable-length chromosomes, a new genetic operator, a
slacked binary tournament selection, and an efficient fitness
evaluation method. This approach was examined and compared with 22 peer competitors including the most state-ofthe-are algorithms on nine benchmark datasets commonly
used in deep learning. The experimental results show that
the proposed EvoCNN method significantly outperforms all
of these existing algorithms on almost all these datasets in
terms of their best classification performance. Furthermore,
the mean classification error rate of the proposed algorithm
is even better than the best of others in many cases. In
addition, the model optimized by EvoCNN is with a much
smaller number of parameters yet promising best classification
performance. Specifically, the model optimized by EvoCNN
employs 100 epochs to reach the lowest classification error
rate of 5.47% on the Fasion dataset, while the state-of-theart VGG16 employs 200 epochs but 6.50% classification error
rate. Findings from the experimental results also suggest that
the proposed EvoCNN method could provide more options
for the manufacturers that are interested in integrating CNNs
into their products with limited computational and battery
resources.
In this paper, we investigate the proposed EvoCNN method
only on the commonly used middle-scale benchmarks. However, large-scale data are also widely exist in the current era
of big data. Because evaluating only one epoch on these largescale data would require significant computational resource
and take a long period, the proposed fitness evaluation method
is not suitable for them unless a huge amount computational
resources are available. In the future, we will put effort on
efficient fitness evaluation techniques. In addition, we will also
investigate evolutionary algorithms for recurrent neural networks, which are powerful tools for addressing time-dependent
data, such as the voice and video data.
R EFERENCES
[1] D. CireşAn, U. Meier, J. Masci, and J. Schmidhuber, “Multi-column
deep neural network for traffic sign classification,” Neural Networks,
vol. 32, pp. 333–338, 2012.
[2] F. Ning, D. Delhomme, Y. LeCun, F. Piano, L. Bottou, and P. E. Barbano,
“Toward automatic phenotyping of developing embryos from videos,”
IEEE Transactions on Image Processing, vol. 14, no. 9, pp. 1360–1371,
2005.
[3] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 1097–1105.
[4] D. H. Hubel and T. N. Wiesel, “Receptive fields, binocular interaction
and functional architecture in the cat’s visual cortex,” The Journal of
Physiology, vol. 160, no. 1, pp. 106–154, 1962.
[5] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard,
W. Hubbard, and L. D. Jackel, “Backpropagation applied to handwritten
zip code recognition,” Neural Computation, vol. 1, no. 4, pp. 541–551,
1989.
[6] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[7] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Cognitive Modeling, vol. 5, no. 3,
p. 1, 1988.
[8] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[9] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 770–778.
[10] H. Bourlard and Y. Kamp, “Auto-association by multilayer perceptrons
and singular value decomposition,” Biological Cybernetics, vol. 59, no.
4-5, pp. 291–294, 1988.
[11] G. E. Hinton and R. S. Zemel, “Autoencoders, minimum description
length, and helmholtz free energy,” Advances in Neural Information
Processing Systems, pp. 3–3, 1994.
[12] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507,
2006.
[13] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13, no. Feb, pp.
281–305, 2012.
[14] C. E. Rasmussen and C. K. Williams, Gaussian processes for machine
learning. MIT press Cambridge, 2006, vol. 1.
[15] J. Močkus, “On bayesian methods for seeking the extremum,” in
Optimization Techniques IFIP Technical Conference. Springer, 1975,
pp. 400–404.
[16] J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl, “Algorithms
for hyper-parameter optimization,” in Advances in Neural Information
Processing Systems, 2011, pp. 2546–2554.
[17] F. Hutter, H. H. Hoos, and K. Leyton-Brown, “Sequential model-based
optimization for general algorithm configuration.” LION, vol. 5, pp. 507–
523, 2011.
[18] J. Beck and T. Fiala, “integer-making theorems,” Discrete Applied
Mathematics, vol. 3, no. 1, pp. 1–8, 1981.
[19] J. Halton and G. Smith, “Radical inverse quasi-random point sequence,
algorithm 247,” Communications of the ACM, vol. 7, p. 701, 1964.
[20] E. I. Atanassov, A. Karaivanova, and S. Ivanovska, “Tuning the generation of sobol sequence with owen scrambling.” in Large-Scale Scientific
Computing. Springer, 2009, pp. 459–466.
[21] B. Baker, O. Gupta, N. Naik, and R. Raskar, “Designing neural network
architectures using reinforcement learning,” International Conference on
Learning Representations, 2017.
[22] D. Ashlock, Evolutionary computation for modeling and optimization.
Springer Science & Business Media, 2006.
[23] T. Back, Evolutionary algorithms in theory and practice: evolution
strategies, evolutionary programming, genetic algorithms.
Oxford
university press, 1996.
[24] M. Mitchell, An introduction to genetic algorithms. MIT press, 1998.
[25] X. Yao, “Evolving artificial neural networks,” Proceedings of the IEEE,
vol. 87, no. 9, pp. 1423–1447, 1999.
[26] G. Miller, “Designing neural networks using genetic algorithms,” Proceedings of ICGA-89, 1989.
[27] K. O. Stanley and R. Miikkulainen, “Evolving neural networks through
augmenting topologies,” Evolutionary computation, vol. 10, no. 2, pp.
99–127, 2002.
[28] K. O. Stanley, D. B. D’Ambrosio, and J. Gauci, “A hypercube-based
encoding for evolving large-scale neural networks,” Artificial life, vol. 15,
no. 2, pp. 185–212, 2009.
[29] K. O. Stanley, “Compositional pattern producing networks: A novel
abstraction of development,” Genetic programming and evolvable machines, vol. 8, no. 2, pp. 131–162, 2007.
[30] J. K. Pugh and K. O. Stanley, “Evolving multimodal controllers with
hyperneat,” in Proceedings of the 15th Annual Conference on Genetic
and Evolutionary Computation. ACM, 2013, pp. 735–742.
[31] M. Kim and L. Rigazio, “Deep clustered convolutional kernels,” in
Feature Extraction: Modern Questions and Challenges, 2015, pp. 160–
172.
[32] C. Fernando, D. Banarse, M. Reynolds, F. Besse, D. Pfau, M. Jaderberg,
M. Lanctot, and D. Wierstra, “Convolution by evolution: Differentiable
pattern producing networks,” in Proceedings of the 2016 on Genetic and
Evolutionary Computation Conference. ACM, 2016, pp. 109–116.
[33] B. Zoph and Q. V. Le, “Neural architecture search with reinforcement
learning,” arXiv preprint arXiv:1611.01578, 2016.
[34] B. Baker, O. Gupta, N. Naik, and R. Raskar, “Designing neural
network architectures using reinforcement learning,” arXiv preprint
arXiv:1611.02167, 2016.
[35] P. Verbancsics and J. Harguess, “Generative neuroevolution for deep
learning,” arXiv preprint arXiv:1312.5355, 2013.
[36] E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu, Q. Le, and
A. Kurakin, “Large-scale evolution of image classifiers,” arXiv preprint
arXiv:1703.01041, 2017.
[37] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural
networks,” in Proceedings of the Fourteenth International Conference
on Artificial Intelligence and Statistics, 2011, pp. 315–323.
[38] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proceedings of the 13th International
Conference on Artificial Intelligence and Statistics, 2010, pp. 249–256.
[39] A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,”
The Journal of physiology, vol. 117, no. 4, pp. 500–544, 1952.
[40] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in European Conference on Computer Vision.
Springer, 2014, pp. 818–833.
[41] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 1–9.
[42] M. N. Omidvar, X. Li, Y. Mei, and X. Yao, “Cooperative co-evolution
with differential grouping for large scale optimization,” IEEE Transactions on Evolutionary Computation, vol. 18, no. 3, pp. 378–393, 2014.
[43] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
review and new perspectives,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
[44] Y. Bengio and O. Delalleau, “On the expressive power of deep architectures,” in International Conference on Algorithmic Learning Theory.
Springer, 2011, pp. 18–36.
[45] O. Delalleau and Y. Bengio, “Shallow vs. deep sum-product networks,”
in Advances in Neural Information Processing Systems, 2011, pp. 666–
674.
[46] A. Blumer, A. Ehrenfeucht, D. Haussler, and M. K. Warmuth, “Occam’s
razor,” Information Processing Letters, vol. 24, no. 6, pp. 377–380, 1987.
[47] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous
search space,” Complex Systems, vol. 9, no. 3, pp. 1–15, 1994.
[48] K. Deb, Multi-objective optimization using evolutionary algorithms.
John Wiley & Sons, 2001, vol. 16.
[49] H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-mnist: a novel image
dataset for benchmarking machine learning algorithms,” arXiv preprint
arXiv:1708.07747, 2017.
[50] H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y. Bengio,
“An empirical evaluation of deep architectures on problems with many
factors of variation,” in Proceedings of the 24th International Conference
on Machine Learning. ACM, 2007, pp. 473–480.
[51] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally,
and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer
parameters and < 0.5 mb model size,” arXiv preprint arXiv:1602.07360,
2016.
[52] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[53] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractive
auto-encoders: Explicit invariance during feature extraction,” in Proceedings of the 28th International Conference on Machine Learning, 2011,
pp. 833–840.
[54] K. Sohn and H. Lee, “Learning invariant representations with local
transformations,” arXiv preprint arXiv:1206.6418, 2012.
[55] K. Sohn, G. Zhou, C. Lee, and H. Lee, “Learning and selecting features
jointly with point-wise gated boltzmann machines,” in International
Conference on Machine Learning, 2013, pp. 217–225.
[56] J. Bruna and S. Mallat, “Invariant scattering convolution networks,”
IEEE transactions on Pattern Analysis and Machine Intelligence, vol. 35,
no. 8, pp. 1872–1886, 2013.
[57] T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “Pcanet: A simple
deep learning baseline for image classification?” IEEE Transactions on
Image Processing, vol. 24, no. 12, pp. 5017–5032, 2015.
[58] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: Nsga-ii,” IEEE Transactions on Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.
[59] G. E. Hinton, “A practical guide to training restricted boltzmann
machines,” in Neural Networks: Tricks of the Trade. Springer, 2012,
pp. 599–619.
[60] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.
Corrado, A. Davis, J. Dean, M. Devin et al., “Tensorflow: Large-scale
machine learning on heterogeneous distributed systems,” arXiv preprint
arXiv:1603.04467, 2016.
[61] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International
Conference on Machine Learning, 2015, pp. 448–456.
[62] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” in Proceedings of the 22nd ACM international
conference on Multimedia. ACM, 2014, pp. 675–678.
[63] T. Chen, M. Li, Y. Li, M. Lin, N. Wang, M. Wang, T. Xiao, B. Xu,
C. Zhang, and Z. Zhang, “Mxnet: A flexible and efficient machine
learning library for heterogeneous distributed systems,” arXiv preprint
arXiv:1512.01274, 2015.
| 1cs.CV
|
Stampery Blockchain
arXiv:1711.04709v1 [cs.CR] 13 Nov 2017
Timestamping Architecture (BTA)
Adán Sánchez de Pedro Crespo
Luis Iván Cuende García
Stampery, CTO
Aragon, CEO
[email protected]
[email protected]
∗
Version 6 – LTS
Published on October 16, 2016
Last reviewed on November 13, 2017
Abstract
A method for timestamping, anchoring and certification of a virtually unlimited amount of
data in one or more blockchains, focusing on scalability and cost-effectiveness while ensuring
existence, integrity and ownership by using cryptographic proofs that are independently verifiable
by anyone in the world without disclosure of the original data and without the intervention of
the certifying party.
I.
B
Introduction
lockchains—understood as immutable, decentralized, distributed and unedited digital
ledgers—have already proved their undeniable capability to serve as a universal support
for a distributed monetary system.
Since the appearance of Bitcoin[1] and its revolutionary approach to digital cash, there have been
many attempts to develop methods allowing anyone to leverage the immutability of the blockchain
technology for purposes other than the transfer of currency.
The Bitcoin protocol itself provides an opcode[2] called OP_RETURN[3] that lets the participants
of the network to embed arbitrary data into the transactions they broadcast to the Bitcoin blockchain
since version 0.9.0 of Bitcoin Core, released in March 2014.
Since OP_RETURN was made available, tools like Proof of Existence[4], Stampery[5] or
Tierion[6] have been making use of the Bitcoin blockchain to timestamp and verify data.
Nevertheless, the OP_RETURN method has several limitations that are intrinsic to the very
specificities of the bitcoin protocol:
∗ Luis
Iván—Stampery’s former CTO and now one of our advisors—is credited here for being the original deviser
and developer of BTA’s very first version back in 2014.
1
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
• In order to keep the average transaction and block size inside acceptable bounds, the maximum
length for a single OP_RETURN operation is 80 bytes.
• It is considered non-standard for a transaction to carry more than one OP_RETURN output.
• Due to the existing block size limit, the bitcoin network can handle up to 7 transactions per
second under ideal circumstances, but the actual value is even lower.
• The fee that must be paid to the network in order to rest assured that all transactions are
accepted and validated in a reasonable time span is high and may keep rising over time as
Bitcoin value fluctuates. Broadcasting a transaction containing a single OP_RETURN with a
certain guarantee that it will be written into the next block and confirmed nearly immediately
can cost up to 120,000 satoshi (0.0012 Bitcoins), which at the time of the last review of this
document roughly amounts to $3.12.
In a very first approach to overcoming some of those limitations, the aforementioned timestamping
services do not actually embed the whole piece of data to be timestamped. Instead, they calculate
a cryptographic hash of the data, which serves as a univocal identifier that has a length between
32 and 64 bytes and therefore fits into a single bitcoin transaction. This is what we refer as data
anchoring, blockchain timestamping or simply stamping.
Nevertheless, there are some use cases which require bulk timestamping of hundreds or thousands
of files; or timestamping the same piece of data over and over in its successive versions. This would
require expending a huge amount of money in transaction fees and would quickly lead to saturation
of the previously mentioned bitcoin block size limit. These consequences would obviously render the
OP_RETURN method unsuitable for such use cases.
It is worth mentioning that other Proof-of-Work blockchains like Ethereum’s or Litecoin’s do not
offer a definitive solution for those problems. Although they may benefit from lower transaction
fees at the moment, given a huge volume of stamping actions, the limitations would be roughly the
sames.
This technical paper proposes a method and architecture that allows (1) carrying out scalable
and cost-effective blockchain data anchoring (overcoming all the already referred limitations), (2)
generating irrefutable proofs of anchoring and (3) verification of the validity of such proofs by any
individual or system in the world.
II.
B
Blockchain Anchoring
lockchain anchoring or blockchain timestamping consists in leveraging the immutability and
security features of public blockchain technologies to certify, ensure and prove that a document,
communication or dataset existed at a certain point in time by generating a digital proof of existence.
Unlike signature-based timestamping and other digital evidentiary tools that existed before the
2
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
emergence of blockchains and their consensus algorithms, blockchain anchoring does not require
relying on a potentially corruptible third party as a source of truth.
Instead, blockchain anchoring relies on mathematically provable digital documents that are
backed by the same consensus algorithms found in the blockchains used to perform the anchoring
process. The conflicting interests of the numerous and heterogeneous participants of blockchain
networks make it impossible for a majority of them to collude with the purpose of altering or
tampering with the timestamps.
Given the historic evolution of the volume of monetary value stored in form of tokens or digital
coins in the main public blockchains, the chance for these to completely disappear in the next decades
or even centuries is very limited, which ensures that proofs generated by blockchain anchoring
methods will be available and enforceable far in the future.
Most blockchain anchoring services make use of one-way “hash functions” to generate unique
fingerprints of the data being anchored and publicly distribute those in such a manner that anyone
holding the original data can verify that they match, but it is impossible for anyone else to retrieve
or reconstruct the original data with the only knowledge of the fingerprint. This property guarantees
that the proofs are completely free of privacy concerns and therefore they can be published safely
on public blockchains.
From a technical standpoint, the stamping or anchoring itself is performed by:
1. Applying a hash function (commonly SHA-256[7]) to the data unit to be anchored,
2. writing the resulting hash into a blockchain transaction,
3. broadcasting the transaction into one or more blockchains, and finally
4. distributing or storing the data of the blockchain transaction along the metadata of the
anchored data unit.
The time assumed to be the timestamp is the one contained in the information of the block in
which the broadcasted transaction gets included, which can be easily looked up in block explorer
websites and services. While the precision and accuracy of these timestamps is not comparable to
that of atomic clocks, it happens to be good enough for most use cases.
Because of the increasing cost of broadcasting transactions in the main public blockchains, there
are some cryptographic constructions that can be applied in order to aggregate several data units
into only one hash and send them in a single blockchain transaction in a similar fashion to how
transactions themselves are aggregated into every block in a blockchain. This makes the cost of the
anchoring process dependent on the period between transactions that the anchoring party decide to
establish rather than on the number of data units being anchored. This way, a virtually unlimited
number of data units can be anchored into a single blockchain transaction, making the unitary cost
of the each stamp approach asymptotically to zero.
3
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
tc(n) = f ee ∈ O (1)
tc(n) = n · f ee ∈ O (n)
tc(n)
= f ee ∈ O (1)
n
cpt(n) =
ppt(n) =
tc(n)
cpt(n)
5
3
1
1
3
5
tc(n)
cpt(n)
5
3
1
tc(n)
f ee
1
=
∈ O( )
n
n
n
1
3
5
Figure 1: Without aggregation, the total cost is
Figure 2: With aggregation, the total cost al-
directly proportional to the number of
ways equals the transacion fee (con-
timestamps (linear) while the cost per
stant) while the cost per timestamp is
timestamp always equals the transaction
inversely proportional to the number of
fee (constant).
timestamps (rectangular hyperbola).
III.
Hashes aggregation
Many single hashes corresponding to different pieces of data or files can be compiled into a single
hash by building a binary tree in which every leaf node is populated with each of the hashes while
every non-leaf node is populated with the hash of the merger of its child nodes. This type of tree is
commonly known as Merkle tree[8].
α
β
γ
δ
& .
& .
αβ
γδ
&
.
αβγδ
Figure 3: Example of Merkle tree with 4 hashes as leaf nodes.
Merkle trees have been used for decades to perform efficient and secure verification of the contents
of large data structures[9]. Indeed, popular blockchain systems like Bitcoin, Ethereum or Litecoin
rely on Merkle trees for gathering together transactions into every block of the chain.
4
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
The tip node in the tree—the hash resulting from aggregation of all the hashes in the leafs—is
called Merkle root or simply root.
The binary hash chain[10] proving that a certain hash belongs to a tree—further discussed in
Section V—is called Merkle proof or simply proof.
α
−→
%
β
αβ
−→
αβγδ
%
γδ
Figure 4: Binary hash chain representing the Merkle proof for the α hash in the tree depicted in Figure 1.
Π(α) = {β, γδ}
By aggregating multiple hashes into a Merkle tree and then publishing only the root, it is possible
to anchor large volumes of data in a blockchain with a single transaction, which dramatically reduces
costs and avoids the aforementioned bottlenecks.
The algorithm followed to calculate the hash in each non-leaf node by joining its two child nodes
is called a mixing fuction, Merkle mixer or simply mixer.
A very common mixer—used by the Stampery API and Chainpoint—can be expressed in the
form SHA256(λ + ρ) and represents the result of calculating the SHA-256 hash of the concatenation
of the left child λ and the right child ρ. However, given that concatenation is a non-commutative
binary operation (λ + ρ 6= ρ + λ), it makes necessary to distribute some additional bits of information
along the Merkle proofs, being them (a) the position (index) in which each leaf hash was added, or
(b) the order in which every pair of child nodes were concatenated.
To avoid the need to deal with those order bits, previous versions of BTA used a different
approach to mixing functions: commutative concatenation.
Commutative concatenation, u, of two values λ and P can be easily achieved by ordering them
from lowest to highest before performing the concatenation. This way, λ u ρ = ρ u λ.
Former versions of BTA proposed the use of the SHA-3 hash algorithm as defined by FIPS-202[11]
with a 512 bit key to ensure resistance to length extension attacks and many other theoretical vulnerabilities affecting SHA-2 and all other algorithms based on the Merkle–Damgård construction[12].
Nevertheless, the current versions of BTA and the Stampery API use SHA-256 by default for the
sake of ease of integration into a wider range of platforms and devices.
Due to the way that binary trees work, the root is always connected to every leaf hash by a
unique path with the exception of the cases when the number of nodes in one of the levels in the
tree is odd. In such cases, in order to find a root that is connected to all leaf hashes, "orphan" nodes
must be (a) promoted to the next level without applying any mixing function to its value; or (b)
mixed with a random string of the same length of the actual hashes. This process is called closing a
5
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
algorithm mixer is
input: left hash λ,
right hash ρ
output: parent node hash π
π ← SHA256(λ + ρ)
return π
Figure 5: The non-commutable mixing function used by the Stampery API, described in pseudocode.
Merkle tree.
α
β
γ
δ
& .
& .
↓
αβ
γδ
&
.
↓
αβγδ
&
.
αβγδ
Figure 6: Example of closing a Merkle tree with an odd number of leaf nodes: the node is promoted to its
immediately upper level until it can be merged into the root.
IV.
Anchorage procedures
As explained in the introduction, Bitcoin anchoring is not perfectly suitable for those cases in which
a) latency is paramount; or b) a really fast or near-immediate confirmation is needed.
In the other hand, newer blockchain projects like Ethereum have still a lower hashing power—
which translates to lower certainty and security—but they are much faster when it comes to the
time needed to put a hash in a block.
Table 1: Average block time comparison between highest capitalization blockchains.
Chain
Average block time
Bitcoin
∼10 minutes
Ethereum
∼17 seconds
Ripple
∼4 seconds
Litecoin
∼2.5 minutes
6
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
BTA is designed to allow anchoring data into more than one blockchain, leveraging the advantages
of each one. A recommended strategy is combining Bitcoin and Ethereum anchoring together to get
certainty and responsiveness at the same time.
Anchoring to multiple chains and different ways to do so will be further discussed in Section VIII.
i. Bitcoin anchoring
Bitcoin anchoring is achieved by using the aforementioned OP_RETURN opcode.
The OP_RETURN data length limit used to be set to 40 bytes before February 2015. Then the
limit was raised to 80 bytes[13], which allowed the use of longer hashing digests.
The hashes resulting from the recommended hashing algorithm in this document—SHA-2 in its
256 bit version—are 32 bytes long, so they fit perfectly into a single OP_RETURN instruction.
Addition of a OP_RETURN operation to a Bitcoin transaction is made by adding a zero-value
output with a scriptPubKey consisting of "6A σ δ", where σ is the size in bytes of the data to attach
to the transaction, and δ is the data itself. Both values are encoded in hexadecimal notation.
6A 20 9D0F5692F0A7CCDBE5554732094CFB52
589F5D8AD762BB54B77A1978462F01C2
Figure 7: Sample scriptPubKey for anchoring a SHA-256 hash. Spaces and newlines must be ignored.
A relatively high fee must be paid in order to rest assured that a transaction gets successfully
accepted and confirmed by the Bitcoin network in a reasonable time even under heavy load
circumstances. Recommended values are between 60,000 and 120,000 satoshi.
ii. Ethereum anchoring
The Ethereum protocol provides a convenient method for anchoring hashes or data into transactions.
This is done by creating a message call transaction T in which the data attribute Td is set to be the
Merkle root.
The protocol expects Td to be an unlimited size byte array[14], so again the hashes resulting from
the recommended hashing algorithm in this document—SHA-2 in its 256 bit version—fit perfectly
into a single transaction.
A gas1 value of 24000 szabo at a 20 Gigawei gas price is typically needed for broadcasting a
message call transaction containing a SHA-256 hash.
1 Gas
is the pricing or fee for running a transaction or contract in Ethereum.
7
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
iii. Prefixing
Following Bitcoin community’s convention for prefixing OP_RETURN operations in order to identify
the generating application, a prefix should be added in front of the data itself, no matter which
blockchain we are anchoring to.
At the time of this writing, there is no standarized format for prefixes, so different BTA
implementations may choose to use it in a different manner.
By way of illustration, Stampery’s BTA implementation uses the prefix format "53 υ φ", where
53 represents the letter "S"—standing for "Stampery"—in ASCII format, υ is one byte telling the
BTA version again in ASCII, and φ is a two-bytes string identifying the server that broadcasted the
transaction.
6A 24 53 35 9B73
9D0F5692F0A7CCDBE5554732094CFB52
589F5D8AD762BB54B77A1978462F01C2
Figure 8: Same scriptPubKey used in Figure 5, this time with proper prefixing and values υ = 5, φ =
”9B73”. Spaces and newlines must be ignored.
The φ value may be calculated by taking the first two bytes from the MD5 hash of the server’s
host name.
V.
Proof of anchorage
As mentioned before, the binary hash chain[10] that proves that a certain hash belongs to a tree is
called Merkle proof or simply proof.
The proof for a single leaf hash contains the sibling hashes that need to be sequentially concatenated and rehashed in pairs in order to reconstruct the path from the leaf to the root.
Π(α) = {β, γδ, }
Π(γ ) = {δ, αβ, }
Π(β ) = {α, γδ, }
Π() = {αβγδ}
Π(δ ) = {γ, αβ, }
Figure 9: Example illustrating the proofs (Π) for all the hashes in the Merkle tree from Figure 4. As it is
noticeable, two contiguous hashes will likely have similar proofs, with the only one or two first
items in the hash chain being different.
In order to allow independent individuals and systems to verify the proofs of anchorage, there
are some other elements that need to be distributed along the hash chain and the Merkle index.
8
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
Those are: (1) the hash of the original data piece, (2) a reference to the chain in which the root was
embedded, and (3) a transaction identifier that unambiguously points to the transaction where the
root was embedded.
i. Proof formatting for internal use
Stampery’s approach to proof formatting is really minimalistic and concise. A BTA proof of
anchorage consists of one array or list containing the following items:
• Π[0] : BTA version (currently 6)
• Π[1] : Merkle index
• Π[2] : Merkle proof (siblings list/hash chain)
• Π[3] : Merkle root
• Π[4] : Anchor tuple
– Π[4][0] : Chain ID
– Π[4][1] : Prefix
– Π[4][2] : Transaction ID
Current BTA version is 6. Prior versions used a different proof format which is not compatible
with the one proposed here.
The Merkle root Π[2] and all the hashes in the proofs list Π[1][n] should be printed using uppercase.
On the contrary, the transaction ID can be printed using either uppercase or lowercase.
Merkle proof lists with zero siblings can be represented by an empty brackets [], empty quotation
marks "" or a null value.
ii. Chainpoint and OpenTimeStamps compatibility
BTA proofs are 100% compatible with Chainpoint[15], a similar technology developed by Tierion,
Inc. This means that the proofs generated by Stampery’s BTA platform and the Stampery API can
be easily presented in the Chainpoint format, as shown in Figure 9.
Similarly, BTA proofs can be also presented in OpenTimeStamps[16] binary format thanks to
the extensibility of such proof format and protocol.
iii. Chain IDs
The chain id is an integer number representing which blockchain is the proof anchored to.
Positive values correspond to production chains—livenets—while negative values correspond to
their test counterparts—testnets.
9
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
[
6,
2,
[
"EFA4BA8F7A66BC3B...D3C358038F5A9C27"
"7DDE76C5E472C9AE...B7B9567DCB3E9551"
],
"1FA2DF8A8ABEA78E...EDC4E25207D2F125",
[
1,
"9B73"
"84ba00d2cebbb4ee...fbb06a053e4fba10"
]
]
Figure 10: A BTA proof of anchorage, represented using JSON format. All the hashes and the transaction
ID have been shortened for fitting this document’s layout.
Table 2: Livenet Chain IDs
ID
Chain
1
Bitcoin livenet
2
Ethereum Classic Livenet
3
Ethereum Fork Livenet
4
Litecoin
Table 3: Testnet Chain IDs
ID
Chain
-1
Bitcoin livenet
-2
Ethereum Rinkeby testnet
-3
Ethereum Morden testnet
-4
Litecoin testnet
ID no. 0 is left unassigned forever and shall only be used for protocol testing purposes.
10
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
{
"@context": "https://w3id.org/chainpoint/v2",
"type": "ChainpointSHA256v2",
"targetHash": "9D0F5692F0A7CCDB...B77A1978462F01C2",
"proof": [
{
"left": "EFA4BA8F7A66BC3B...D3C358038F5A9C27"
},
{
"right": "7DDE76C5E472C9AE...B7B9567DCB3E9551"
}
],
"anchors": [
{
"type": "BTCOpReturn",
"sourceId": "84ba00d2cebbb4ee...fbb06a053e4fba10"
}
]
}
Figure 11: The same BTA proof from Figure 8, in Chainpoint v2 format.
VI.
Proofs verification
Verification of Merkle proofs is performed by reconstructing the path between the original data hash
and the Merkle root. This is done by applying the mixing function to the hash and the first sibling
Π[2][0] in the proof, then applying the same mixer to the resulting hash and the next sibling, and
continuing this process subsequently with Π[2][n] until there are no more hashes left to merge. The
resulting root candidate hash ρ is expected to correspond to the Merkle root Π[3] .
The order in which the hashes are concatenated before the hashing function is applied can be
easily derived from the Merkle index by using this function:
o = ( 2ιn ) mod 2
Figure 12: Given a merkle index ι and merkle proof Π[2] , this function tells the order of concatenation o
for any element Π[2][n] belonging to Π. o = 1 means left side concatenation (P i[2][n] + h) and
o = 0 means right side concatenation (h + P i[2][n] ).
11
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
algorithm prover is
input: original data hash η,
merkle proof π
merkle index ι
output: merkle root ρ
ρ ← η
for n in π
σ ← π [n]
o ← (ι / pow(2, n)) % 2
o = 1 do
if
ρ =
mixer(ρ, σ)
else do
ρ =
mixer(σ, ρ)
return ρ
Figure 13: An iterative proof verification function, described in pseudocode. The mixer function is the one
described in Figure 3.
In the event that the resulting hash and the Merkle root mismatched, either the proof or the
original data hash were modified, corrupted or tampered with at some time after the anchoring took
place. This can be used as a method for checking file integrity over time—time stamp the same file
over and over and keep a history of the data hashes and proofs: if the file gets modified, the hash
will change, will not match the one related to the previous proof and the date of modification will
become evident.
In order to verify the validity of the proof and the anchor, it is also necessary to check if the
root candidate or the Merkle root match the data payload in the transaction. This can be done by
searching the transaction ID Π[4][1] in a block explorer2 , reading the payload and then checking
whether it contains the root as a substring—probably at the end because of prefixing.
η = Π [2]
η∈δ
Figure 14: The candidate root η should match and be a substring of the data payload δ found in the
anchoring transaction.
2 Some
popular block explorers are Blockcypher for Bitcoin and Etherscan for Ethereum.
12
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
VII.
Progressive proofing
Hashing is a quite intensive operation in terms of CPU usage. When dealing with big Merkle
trees—with hundreds, thousands or even millions of leaf nodes—closing the tree and finding a final
Merkle root is not trivial. Specifically, For any number ν of child nodes, the number of hashing
operations needed to find the root is ν − 1.
In addition, calculating the proof for every leaf node once the tree is closed means a lot of time
wasted traversing it and picking the siblings.
Those are the reasons why we recommend to build the Merkle tree and the proofs as the hashes
come and are appended to the leaf level instead of doing it at the time of closing.
Taking this into account, the logic to be carried out every time a new leaf node is added is
described in Figure 11.
VIII.
Anchoring to multiple chains
As mentioned before, it is specially convenient to perform anchoring to multiple blockchains in order
to leverage the best of each one’s feature set and also to have auxiliary anchors just in case some day
one blockchain is somehow hacked or corrupted and the anchoring transaction is reverted, tampered
with or simply disappears.
i. Parallel anchorage
Parallel anchorage is the simplest way to anchor Merkle roots to multiple chains.
It involves closing the Merkle tree every few minutes, calculating the root and then embedding it
into multiple blockchains at the same time.
This way, the proofs for all the chains are identical with the exception of their last element—the
anchor tuple Π[4] —that will have different values for the chain ID Π[3][0] and transaction ID Π[3][2] .
ii. Incremental anchorage
Incremental anchorage is a more complex but also more powerful way to anchor Merkle roots to
multiple chains.
It involves creating several different types of Merkle trees, each type corresponding to a chain we
want to anchor to. Every of them will have its own lifetime—time before tree closing—based on the
average block time of the matching chain.
For example, if we were to perform incremental anchorage with Ethereum and Bitcoin, we could
define the following tree types and tree lifetimes:
13
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
algorithm pusher is
input: merkle tree τ ,
proofs list π,
new hash η,
level λ (default = 0)
output: merkle tree τ ,
proofs list π
if η 6∈ τ [0] do
σ ← τ [λ] or []
σ ← σ + [η ]
if λ = 0 do
π [η ] ← []
if mod(|τ [λ]|) = 1 do
ς ← τ [λ][−1]
$ ← mixer (ς, η )
π ← proof er (τ , π, η, ς, λ)
{τ , π} ← pusher (τ , π, $, λ + 1)
return {τ , π}
Figure 15: A recursive hash appending routine, described in pseudocode. The mixer function is the one in
Figure 3, and the proofer function is described in Figure 11.
Chain
Tree lifetime
Ethereum
1 minute
Bitcoin
10 minutes
A incremental anchorage procedure for those chains and lifetimes could be the following:
1. Create one tree for Ethereum and one for Bitcoin.
2. Receive all the hashes as they come and push them as leafs into Ethereum tree.
3. After 1 minute:
(a) Close Ethereum tree and get its Merkle root.
(b) Anchor the root to Ethereum blockchain.
14
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
algorithm proofer is
input: merkle tree τ ,
proofs list π,
left hash α,
right hash β,
level λ (default = 0)
output: proofs list π
φ ← τ [0]
κ ← τ [λ]
ρ ← b2ˆ(λ + 1)e
σ ← b−ρ + |κ| × (2ˆλ)e
← min((σ + ρ), |φ|) − 1
ι ← σ
for each hash (χ) in φ ∈ [σ, ] do
if ι < bρ / 2e do
π [χ] ← π [χ] + [α ]
else do
π [χ] ← π ]χ] + [β ]
ι ← ι + 1
return π
Figure 16: Auxiliar function for progressive proofing used by the pusher function in Figure 11.
(c) Deliver the Ethereum proofs to the requesting parties.
(d) Push the Ethereum root into the Bitcoin tree as if it were a leaf hash.
(e) Create a new Ethereum tree for the next 1-minute frame.
4. Repeat steps 2 and 3 until minute 10.
5. When minute 10 comes:
(a) Close Bitcoin tree and get its Merkle root.
(b) Anchor the root to Bitcoin blockchain.
(c) Append each Bitcoin proof’s siblings list Π[2] (proof head) to the end of all the Ethereum
proofs’ siblings list Π0[2] (proof tail) having a root Π0[3] that matches the original data
hash from the Bitcoin proof.
15
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
(d) Deliver the whole "merged" proofs to the requesting parties.
(e) Delete all the old Ethereum trees and create a new Bitcoin tree for the next 10-minutes
frame.
6. Start over the whole process.
Because of the way the Merkle proofs are generated and merged together, the resulting Ethereum
proof will be a subset of the Bitcoin proof, but both of them will still be perfectly valid separately.
IX.
Distributed BTA (DBTA)
It is easy to imagine that a computer handling a Merkle tree with thousands or millions of nodes
in its leaf level will need a huge amount of RAM memory. Furthermore, if the computer runs out
of memory while building a Merkle tree, it will never be able to close it and all the hashes will be
simply lost.
This kind of bottlenecks and situations can be easily overcome by using a Distributed BTA
implementation as described hereafter.
DBTA is a computation cluster formed by a number ν of BTA instances called nodes.
A messaging queue system is used to distribute messages and balance workload between all the
nodes.
In addition, a P2P protocol in conjunction to a basic consensus algorithm are used to reduce the
number of Bitcoin transactions made by the cluster.
i. The cluster
We analyzed many different technologies and platforms when we built Stampery’s own DBTA cluster,
and finally opted for Erlang/OTP.
Erlang[17] is a programming language and virtual machine used to build massively scalable
soft real-time systems with requirements on high availability. Some of its uses are in telecoms,
banking, e-commerce, computer telephony and instant messaging. Erlang’s runtime system has
built-in support for concurrency, distribution and fault tolerance.
OTP is a set of Erlang libraries and design principles providing middle-ware to develop these
systems. It includes its own distributed database, applications to interface towards other languages,
debugging and release handling tools.
For convenience, we wrote our BTA implementation using Elixir[18], a dynamic, functional
language designed for building scalable and maintainable applications.
16
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
ii. The messaging queue
The messaging queue will fulfill a variety purposes, most of them related to the incremental anchoring
procedure as described by Section VII:
• Load balancing the distribution of incoming hashes to the nodes.
• Keeping the hashes in memory until they have been successfully anchored to all chains and
the corresponding proofs have been delivered.
• Routing the Ethereum proofs back to their requesting parties.
• Load balancing the distribution of Ethereum roots from one node to another, so that they all
receive an even amount of Ethereum roots to use as leaf hashes in their Bitcoin trees.
• Routing the Bitcoin proof heads to the nodes holding the matching proof tails.
• Routing the whole "merged" proofs to the requesting parties.
• Re-delivering hashes and roots to different nodes in case the ones that were assigned originally
fail to perform their functions.
Therefore, the messaging queue system that we need must fulfill all those load balancing, routing
and re-delivering requirements.
That is the case with RabbitMQ, the most notable implementation of the AMQP protocol[19].
We have been successfully using it in Stampery for DBTA since 2015 and it has proven to be the
perfect fit for the set of features previously described.
To implement DBTA in RabbitMQ, the following queues are needed:
Queue name
eth
btc
In addition, it will be necessary to define the following exchanges3 :
Exchange name
Type
eth
direct
btc
direct
proofs
direct
When a hash is received by a DBTA implementation, it must be published to the eth exchange,
which will route it to a different DBTA node every time4 . At the same time, a binding will be
3 Exchanges
are AMQP entities where messages are sent. Exchanges take a message and route it into zero or more
queues. The routing algorithm used depends on the exchange type and rules called bindings.
4 A direct exchange delivers messages to queues based on the message routing key. Direct exchanges are often used
to distribute tasks between multiple workers (instances of the same application) in a round robin manner.)
17
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
created in the proofs exchange, using the hash as the routing key and pointing to a user-specific
proof delivery queue, named in the form of a user identifier plus the −clnt suffix.
The node will then push the hash into its latest Ethereum tree and will wait 1 minute before
closing it. During that time, it will predictably receive approximately
η
ν
hashes, where η is the
number of hashes sent to the exchange during the 1-minute frame, and ν is the number of DBTA
nodes consuming the eth queue.
Once the 1-minute frame is over, all nodes will close their Ethereum trees, get their roots and
publish them to the btc exchange. Such exchange will then route every root to a different DBTA
node. At the same time, another binding will be created in the proofs exchange, using the root
as the routing key and pointing to the same user-specific proof delivery queue, named as already
stated.
The nodes receiving the roots from the btc exchange will treat them as if they were hashes,
push them into their Bitcoin trees and wait for 10 minutes before closing the trees. During that
time, each node will predictably receive up to a maximum of 10 roots.
Once the 10-minutes frame is over, all nodes must close their Bitcoin trees and get their final
Merkle root.
One important advantage of using AMQP is that if for some reason one or more nodes go offline
or gets isolated from the rest of the architecture—in case of network failure, system maintenance,
manual reboot, etc–it will collect the unacknowledged messages previously sent to such node(s) and
reassign them to a different set of live nodes.
iii. Cluster leadership
At this point, if every DBTA node were to broadcast its own Bitcoin transaction, we would be
incurring in a daily cost of up to 24 × 6 × ν × β USD5 , where ν is the number of nodes in the DBTA
cluster, and β is the price of a Bitcoin transaction in US Dollars.
Because we are focusing on scalability and cost-effectiveness, we have figured out a way to keep
the number of Bitcoin transactions steady so it is always the same regardless of the number of nodes
in the DBTA cluster.
The trick consists of choosing a random leader node for every 10-minutes frame and letting it
join all the roots into one and make a single transaction.
In Stampery’s DBTA, this is made possible by Erlang’s bundled distributed computing system[20].
All nodes must share the same magic cookie6 , so that they will be able to connect to each other
automatically as soon as their EPMD7 is set to discover nodes in the right domain names.
5 At
the time of the last review of this document, this corresponds to ν×225 USD per day.
an Erlang node tries to connect to another node, the magic cookies are compared. If they do not match,
6 When
the connected node rejects the connection.
7 Erlang Port Mapping Daemon
18
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
All nodes must have their clocks synchronized with one another with a maximum lead or lag that
needs to be less than a half the duration of the tree lifetime corresponding to the fastest blockchain
being used. This amount of time is called consensus time.
The time frames are fixed in time so that they can be synchronized across all nodes without need
for signaling. For example, the Ethereum 1-minute frames will start at 00:01:00, 00:02:00, 00:03:00,
etc.; while Bitcoin 10-minutes frames will take place at 00:10:00, 00:20:00, 00:30:00 and so on.
Every time a frame expires, the nodes make a list of their peers. Then they take the peers’
names, append the Epoch time8 for the starting second of the just finished frame, and hash the
resulting string. Finally, the list of hashes is sorted alphabetically and the first one in the list is
considered to be the leader.
Once the leader node has been chosen, the other nodes will report their Bitcoin roots to it,
which will then build a final Merkle tree joining its own Bitcoin root and all the ones from its peers.
After consensus time, it should close the tree, get the final Merkle root and anchor it to the Bitcoin
blockchain.
The last thing the leader needs to do is reporting the Bitcoin transaction ID and the final proof
heads to each of its peers, so that they can join them to their own Bitcoin proofs and report their
proof heads back to the nodes holding the proof tails.
If the leader node failed to fulfill its duty in twice consensus time, the runner-up node in the list
will take over and be the new leader for the unconfirmed frame. If it failed too, it will be the turn of
the third node in the list, and so forth.
iv. Proofs delivery
When an Ethereum proof is ready to be delivered to the requesting party, it just needs to be
published to the proofs exchange using the original data hash as the routing key.
Likewise, when a Bitcoin proof is ready to be delivered to the requesting party, it just needs to
be published to the same proofs exchange, this time using the Ethereum root as the routing key.
Due to the way in which AMQP direct exchanges work, both the Ethereum and Bitcoin proofs
will be routed into the right proof delivery queue thanks to the routing keys and the previously
created bindings.
8 Epoch
time is a system for describing instants in time, defined as the number of seconds that have elapsed since
00:00:00 (UTC), Thursday, 1 January 1970, not counting leap seconds.
19
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
References
[1] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” 2009. https://bitcoin.org/
bitcoin.pdf.
[2] Wikipedia, “Opcode — wikipedia, the free encyclopedia.” https://en.wikipedia.org/w/
index.php?title=Opcode&oldid=728968013.
[3] Bitcoin Wiki, “Op_return — bitcoin wiki.” https://en.bitcoin.it/w/index.php?title=
OP_RETURN&oldid=60872.
[4] M. Aráoz, “What is proof of existence?,” 2014. https://proofofexistence.com/about.
[5] Stampery, “21st century notarization,” 2014. https://stampery.com.
[6] Tierion, “Blockchain proof engine,” 2014. https://tierion.com.
[7] Wikipedia contributors, “SHA-2 — wikipedia, the free encyclopedia.”
https://
en.wikipedia.org/wiki/SHA-2.
[8] Wikipedia, “Merkle tree — wikipedia, the free encyclopedia.” https://en.wikipedia.org/w/
index.php?title=Merkle_tree&oldid=734149217.
[9] R. C. Merkle, A Digital Signature Based on a Conventional Encryption Function. CRYPTO
’87. Lecture Notes in Computer Science. 293. p. 396.
[10] Wikipedia, “Hash chain — wikipedia, the free encyclopedia.” https://en.wikipedia.org/w/
index.php?title=Hash_chain&oldid=731737984.
[11] U.S. National Institute of Standards and Technology, “Sha-3 standard: Permutation-based
hash and extendable-output functions.” Federal Information Processing Standards Publication
no. 202, 2015. https://dx.doi.org/10.6028/NIST.FIPS.202.
[12] Wikipedia, “Length extension attack — wikipedia, the free encyclopedia.”
https://
en.wikipedia.org/w/index.php?title=Length_extension_attack&oldid=733680211.
[13] F. Charlon, “Change the default maximum op_return size to 80 bytes.” https://github.com/
bitcoin/bitcoin/pull/5286.
[14] Dr. Gavin Wood, “Ethereum: A secure decentralised generalised transaction ledger - homestead
revision,” 2016. http://gavwood.com/paper.pdf.
[15] Wayne Vaughan, Jason Bukowski, Shawn Wilkinson, “Chainpoint: A scalable protocol
for anchoring data in the blockchain and generating blockchain receipts.,” 2016. https:
//tierion.com/chainpoint.
20
Stampery Blockchain Timestamping Architecture (BTA) • Version 6 • Oct 2016
[16] Peter Todd, Ricardo Casatta, Luca Vaccaro, Andrew Poelstra, et. al., “Open timestamps: A
timestamping proof standard.,” 2017. https://opentimestamps.org/.
[17] Joe Armstrong, Bjarne Däcker, Thomas Lindgren, Håkan Millroth, et al., “Open-source erlang
- white paper,” 2000. http://www1.erlang.org/white_paper.html.
[18] J. Valim, “Elixir,” 2012. http://elixir-lang.org/.
[19] Sanjay Aiyagari, Matthew Arrot, Mark Atwell, Jason Brome, Alan Conway, Robert Greig, Pieter
Hintjens, John O’Hara, Martin Ritchie, Shahrokh Sadjadi, Rafael Schloming, Steven Shaw,
Gordon Sim, Martin Sustrik, Carl Trieloff, Kim van der Riet and Steve Vinoski, “Amqp advanced message queuing protocol - protocol specification,” 2006. https://www.rabbitmq.com/
resources/specs/amqp0-9.pdf.
[20] Ericcson AB, “Distributed erlang,” 2003.
http://erlang.org/doc/reference_manual/
distributed.html.
21
| 8cs.DS
|
Note on Representing attribute reduction and
concepts in concepts lattice using graphs
arXiv:1711.05509v1 [cs.AI] 15 Nov 2017
Jan Konecny
Dept. Computer Science, Palacky University, Olomouc
17. listopadu 12, CZ-77146 Olomouc, Czech Republic
Abstract. Mao H. (2017, Representing attribute reduction and concepts in concept lattice using graphs. Soft Computing 21(24):7293–7311)
claims to make contributions to the study of reduction of attributes in
concept lattices by using graph theory. We show that her results are either trivial or already well-known and all three algorithms proposed in
the paper are incorrect.
Keywords: Clarification; Reduction; Formal Context; Graph; Formal concept
enumeration
1
Introduction
Mao (2017) claims that she studies attribute reduction and formal concept enumeration in Formal Concept Analysis (FCA) with the aid of Graph Theory.
We show, that the use of the Graph Theory is trivial. Specifically, Mao simply
replaces the notion of formal context by an equivalent notion of pre-weighted
relevant graph and presents well-known and trivial facts as novel and interesting. Additionally, Mao provides a method of enumeration of formal concepts in
a reduced formal context. The method is represented by three algorithms; we
demonstrate, using examples, that they are incorrect.
2
Preliminaries
We use the same notions and the same notations as Mao (2017); however we
need to also recall a few notions from (Ganter and Wille, 1999) to explain our
case. Thus, for the reader’s convenience, we provide full preliminaries.
An input to FCA is a triplet (O, P, I), called a formal context, where O, P are
finite non-empty sets of objects and attributes, respectively, and I is a binary
relation between O and P ; (o, a) ∈ I means that the object o has the attribute
a. Finite formal contexts are usually depicted as tables, in which rows represent
objects, columns represent attributes, and each entry contains a cross if the
corresponding object has the corresponding attribute, and is otherwise left blank
(see top parts of Figures 3–5 for examples).
The formal context induces the following operators:
↑ : 2O → 2P assigns to a set X of objects the set X ↑ of all attributes shared
by all the objects in X.
↓ : 2P → 2O assigns to a set B of attributes the set B ↓ of all objects which
share all the attributes in B.
For singletons we use shortened notation and write o↑ , a↓ instead of {o}↑ , {a}↓ ,
respectively.
Formal concept is a pair (X, B) of sets X ⊆ O, B ⊆ P , s.t. X ↑ = B and
↓
B = X. The first component of a formal concept is called extent, the second
one is called intent. A collection of all formal concepts in (O, P, I) is denoted
B(O, P, I). The collection B(O, P, I) with order ≤ defined by (X1 , B1 ) ≤ (X2 , B2 )
if X1 ⊆ X2 for all formal concepts (X1 , B1 ), (X2 , B2 ) ∈ B(O, P, I), forms a
complete lattice called a concept lattice.
We call a computation of B(O, P, I) a formal concept enumeration.
We consider the following two binary relations on P induced by a formal
context:
– dependency ⊑, defined by a ⊑ b if a↓ ⊆ b↓ for all a, b ∈ P ,
– equivalence ≡, defined by a ≡ b if a↓ = b↓ for all a, b ∈ P .
Formal context (O, P, I) is called clarified if a1 ≡ a2 implies a1 = a2 for any
a1 , a2 ∈ P ; i.e. if it does not contain duplicate columns. Removal of duplicate
columns (keeping one representative of each ≡-class) is called a clarification.
An attribute a is called reducible in formal context (O, P, I) if there is Z ⊆
P \{a} such that a↓ = Z ↓ (equivalently, if B(O, P, I) and B(O, P \{a}, I ∩(O×P \
{a})) are isomorphic). Formal context is reduced if it has no reducible attributes.
Removal of reducible attributes is called a reduction. Ganter and Wille (1999)
provide efficient method for clarification and reduction.
Originally for Rough Set Theory, Pawlak (1982) proposed three types of attributes. These types were introduced into FCA by Zhang et al (2005). We omit
their definitions and just explain their relationship to the reducible and irreducible attributes: absolutely necessary (core) attributes are exactly irreducible
attributes; relatively necessary attributes are those reducible attributes that become irreducible when their duplicates are removed; and absolutely unnecessary
attributes are those which are neither absolutely nor relatively necessary.
3
The use of Graph Theory is trivial and the results are
not novel
Mao (2017) claims to propose a method of attribute reduction which utilizes
Graph Theory and is based on removing vertices from a particular graph; this
should distinguish her method from other known approaches (Ganter and Wille,
1999; Zhang et al, 2005) which are based on the removal of attributes.
Specifically, she introduces a so-called pre-weighted relevant graph G(O, P, I)
(derived from a formal context (O, P, I)) as a graph with vertices being the
attributes in P and edges being given as follows: for a, b ∈ P and a 6= b,
– if x↓ = y ↓ there is a bi-arc joining x and y, i.e. x ⇔ y,
– if x↓ ⊂ y ↓ there is an arc joining y to x, i.e. x → y.
Thus, G(O, P, I) is basically (O, P, I) with the relations ⊑ and ≡ (specifically,
the arcs → correspond to pairs in ⊑ \ ≡ and the bi-arcs ⇔ correspond to non
reflexive pairs in ≡). This represents the entire utilization of Graph Theory in
Mao’s work.
Mao describes the attribute reduction in three steps (see Fig. 1):
(1) For an input formal context (O0 , P0 , I0 ) satisfying
o↑ 6= ∅ and a↓ 6= ∅ for all o ∈ O, a ∈ P
the formal context (O1 , P1 , I1 ) is found by removing full-row objects and
full-column attributes from (O0 , P0 , I0 ).
(2) For (O1 , P1 , I1 ) find context (O2 , P2 , I2 ) by clarification.
(3) For (O2 , P2 , I2 ) find context (O3 , P3 , I3 ) by reduction.
(O0 , P0 , I0 )
(1) remove full-column attributes
and full-row columns
(O1 , P1 , I1 )
(2) remove duplicate attributes
(clarification)
(O2 , P2 , I2 )
(3) remove reducible attributes
(reduction)
(O3 , P3 , I3 )
Fig. 1. Three phase process of attribute reduction in Mao (2017).
At best, Mao’s results could be considered to show how to use precomputed
relations ⊑ and ≡ to achieve better time complexity. However there are two
problems with it.
– First, the relations are used the same way as in (Ganter and Wille, 1999)
(where they are not precomputed), or they present a trivial improvement.
As an example of the latter, Theorem 3.1 characterizes objects with full
row. The theorem states that we need not check whether the object has all
attributes, instead we can just check those that are minimal w.r.t. ⊑.
– Second, if the complexity is taken into account, it should be said that the
computation of ≡ and ⊑ for (O, P, I) requires the same time as entire clarification and reduction using classic methods described in basic literature
(Ganter and Wille, 1999). Mao does not mention the complexity of the computation of ≡ and ⊑ neither does she refer to the classic methods.
Substitution of the basic notions with the equivalent newly introduced notions is Mao’s main resource for results. Instead of removal attributes from
(O, P, I) she removes the corresponding vertices from G(O, P, I). Rephrased,
using basic notions of FCA, the results become trivial or are already well-known.
Specifically, Theorem 3.2 states that duplicate attributes are reducible, which
is obvious. Theorem 3.3 states that in the clarified context (O2 , P2 , I2 ) no attributes are relatively necessary, which is obvious because (O2 , P2 , I2 ) is obtained
by clarification of the context (O1 , P1 , I1 ). Theorem 3.4 is an overcomplicated
characterization of reducible attributes, which after a straightforward simplification becomes the one by Ganter and Wille (1999). Theorem 3.5 states, that
in the reduced context (O3 , P3 , I3 ) all attributes are irreducible, which again is
obvious, because (O3 , P3 , I3 ) is obtained by reduction of (O2 , P2 , I2 ). Finally,
Theorems 3.6 and 3.7 explain trivial relationships between Pawlak’s types of
attributes in P0 , P1 , P2 , and P3 .
4
The proposed algorithms are incorrect
Mao Mao (2017) proposes a method of computing formal concepts in (O, P, I);
more specifically, a subset A containing all formal concepts excluding the top
and the bottom concepts of B(O, P, I) and the attribute concepts.
Her method consists of three algorithms which each compute different portions of A with the aid of a pre-weighted relevant graph; their input is a preweighted relevant graph and a maximal attribute c1 .
1. The algorithm 1 computes a set F containing those concepts (A, B) ∈ A
+
(c1 ) (see Fig. 2 (top)).
which satisfy c1 ∈ B and b ∈
/ B for all b ∈ NG(O,P,I)
2. The algorithm 2 computes a set S containing those concepts (A, B) ∈ A
+
which satisfy c1 ∈ B and b ∈ B for some b ∈ NG(O,P,I)
(c1 ) (see Fig. 2 (middle)).
3. The algorithm 3 computes a set T containing those concepts (A, B) ∈ A
which satisfy c1 ∈
/ B (see Fig. 2 (bottom)).
+
NG(O,P,I)
(c1 ) in items 1. and 2. denotes the lower cone of attribute c1 w.r.t. ⊑,
excluding c1 itself.
All three algorithms are described in a very complicated way which makes
them almost unreadable. More importantly, they are incorrect. In what follows,
we present the examples in which the algorithms fail to deliver correct outputs.
We need to recall some additional notation from (Mao, 2017):
– V →a denotes the upper cone of attribute c w.r.t. ⊑, excluding a itself;
– ω(a) denotes the pre-weight of a, i.e. ω(a) = a↓ .
c1
•
b1
•
b2
•
c1
•
b1
•
b2
•
c1
•
b1
•
b2
•
Fig. 2. Portions of concept lattice computed by the algorithm 1 (top), algorithm 2 (middle), and algorithm 3 (bottom)
c1 c2 b n
1×
2× ×
3 × ××
4
×
c1 {1, 2, 3}
c2 {3, 4}
b
n ∅
{1, 3}
Fig. 3. A formal context (top) and its pre-weighted relevant graph (bottom) as an
example for algorithm 1
4.1
Example for algorithm 1
Consider the formal context depicted in Fig. 3 (top).
We demonstrate that the algorithm 1 fails to deliver correct output for its preweighted relevant graph (Fig. 3 (bottom)) and attribute c1 . We have C = {c1 , c2 }
+
and NG(O,P,I)(c
= {b, n}. The following represents how the algorithm runs
1)
when we exactly follow the steps in Mao (2017).
(step 1)
+
H1 = {x ∈ P \ {c1 } |x ∈
/ NG(O,P,I)
(c1 )
and ω(c1 ) ∩ ω(x) 6= ∅} = {c2 }.
As H1 6= ∅ is the case, we select h1 ∈ H1 . There is only one option, thus we set
h1 := c2 . We compute extent A1 and B1 as
A1 = ω(c1 ) ∩ ω(c2 ) = {3} and B1 = {c1 , c2 }.
As H1 \B1 = ∅, the condition of non-existence of h ∈ H1 \B1 with ω(c1 )∩ω(c2 ) ⊂
ω(h) is trivially satisfied, and (A1 , B1 ) is outputted.
(step 2) As |H1 | = 1 we stop the computation.
We obtained a pair ({3}, {c1, c2 }) as an output, but this pair is not a formal
concept. If we close it, we obtain ({3}, {c1 , c2 , b}) which should not be in F , since
it contains b.
1
2
3
4
5
6
c1 c2 c3 b
×
× ×
× × ×
×
××
×
×
×
c1 {2, 3, 4, 5}
c2 {1, 2, 3}
{3, 4}
c3 {4, 5, 6}
b
Fig. 4. A formal context (top) and its pre-weighted relevant graph (bottom) as an
example for algorithm 2
4.2
Example for algorithm 2
Consider the formal context depicted in Fig. 4 (top).
We demonstrate that algorithm 2 fails to deliver the correct output for its
pre-weighted relevant graph (Fig. 4 (bottom)) and attribute c1 . We have that
+
C = {c1 , c2 , c3 } and NG(O,P,I)
(c1 ) = {b}.
(step 6) We set b1 := b as it is the only option; and we compute
Hb1 = {c2 , c3 }.
(step 7) We have Hb1 6= ∅. We select d1 := c2 . We compute
Ab1 : = ω(b1 ) ∩ ω(d1 ) = ω(b) ∩ ω(c2 ) = {3},
Bb1 : = {b1 , d1 } ∪ V3→b1 ∪ V3→d1 = {b, c2 } ∪ {c1 } ∪ ∅
= {b, c1 , c2 }.
We continue with Case 2 because c3 , the only element in Hb1 \ Bb1 , does not
+
(b) = ∅;
satisfy Ab1 ⊂ ω(c3 ). The rest of this step is trivial since NG(O,P,I)
(Ab1 , Bb1 ) is outputted.
(step 8) We set d11 = c2 , d12 = c3 . And compute
A1d1 = ω(b1 ) ∩ ω(d11 ) ∩ ω(d12 ) = ω(b) ∩ ω(c2 ) ∩ ω(c3 ) = ∅.
Since A1d1 is empty the algorithm stops.
The algorithm terminated without outputting a formal concept ({4}, {c1, c3 , b}),
which belongs to S.
4.3
Example for algorithm 3
The third algorithm uses the previous two algorithms. Even if algorithm 1 and
algorithm 2 were correct, algorithm 3 would be still incorrect, and this is shown
for the formal context depicted in Fig. 5 (top), its pre-weighted relevant graph is
in Fig. 5 (bottom), and its attribute c1 .
c1 c2 c3 b
1
×
2× × ×
3× × ×
4×
×
5
×
c1 {2, 3, 4}
c2 {1, 2, 3}
{2}
c3 {3, 4, 5}
b
Fig. 5. A formal context (top) and its pre-weighted relevant graph (bottom) as an
example for algorithm 3
(step 15) C := C \ {c1 } = {c2 , c3 };
+
G(O, P, I) := G(O, P, I) \ (c1 ∪ NG(O,P,I)
(c1 ));
(1)
i.e. G(O, P, I) now contains only vertices c2 and c3 .
(step 16) We use algorithms 1 and 2 for updated graph G(O, P, I) and c2 . It is
not clear whether (1) includes removal of c1 from P :
– If yes, we get ({3}, {c2 , c3 }) as one of the outputs of algorithm 1 and 2, but
it is not a formal concept
– If no, we get ({3}, {c1 , c2 , c3 }) as one of the outputs of algorithm 1 and 2,
but it does not belong to T .
Either way, we obtain a wrong output.
4.4
Bogus complexity analysis
As the method of formal concept enumeration is incorrect, we could simply
disregard its complexity analysis, however, even then there is a problem which
needs mentioning.
In the description of both, attribute reduction and formal concept enumeration, Mao uses steps which perform intersections and/or comparisons of preweights, i.e. subsets of the set O. In the accompanied complexity analysis, she
claims complexity of these steps to be O(|P |2 ) or O(|P |3 ); i.e independent of
the size of the subsets or the size of O. This could be actually achieved if all the
intersections of pre-weights were precomputed, however, a collection of all the
intersections is, in fact, the collection of all extents. As each extent uniquely determines its formal concept, this assumption means that Mao’s method of formal
concept enumeration requires all formal concepts to be precomputed.
To sum up, either the claimed complexities are incorrect, or Mao works from
an assumption that makes the actual computation superfluous.
5
Conclusion
In (Mao, 2014), Mao substituted the concept-forming operators with seemingly
different operators which produced simple matroids instead of ordinary sets.
Consequently, she presented well-known facts and trivial facts in a complicated
way to make them look novel and interesting. This was exposed by Konecny
(2015).
We have shown that she does the same in (Mao, 2017). Mao provides nothing
new. Her theoretical results are either trivial or well-known and the proposed
algorithms are incorrect.
Acknowledgments
Supported by grant No. 15-17899S, “Decompositions of Matrices with Boolean
and Ordinal Data: Theory and Algorithms”, of the Czech Science Foundation.
Bibliography
Ganter B, Wille R (1999) Formal Concept Analysis – Mathematical Foundations.
Springer
Konecny J (2015) Note on the characterization and reduction of concept lattices
through matroid theory. Inf Sci 307:110–112, DOI 10.1016/j.ins.2015.02.007
Mao H (2014) Characterization and reduction of concept lattices through matroid theory. Inf Sci 281:338–354, DOI 10.1016/j.ins.2014.05.044
Mao H (2017) Representing attribute reduction and concepts in concept lattice using graphs. Soft Computing 21(24):7293–7311, DOI https://doi.org/10.
1007/s00500-016-2441-2
Pawlak Z (1982) Rough sets. International Journal of Computer & Information
Sciences 11(5):341–356
Zhang W, Wei L, Qi J (2005) Attribute reduction theory and approach to concept
lattice. Science in China Series F-Information Sciences 48(6):713, DOI https:
//doi.org/10.1360/122004-104
| 2cs.AI
|
Elliptification of Rectangular Imagery
Chamberlain Fong
[email protected]
Abstract – We present and discuss different algorithms for converting rectangular imagery into elliptical
regions. We will focus primarily on methods that use mathematical mappings with explicit and invertible
equations. We will also present different post-processing effects that could be applied to enhance the
resulting images and minimize distortion.
arXiv:17xx.xxxxv1 [eess.IV] Sep 2017
Keywords – Mapping an Ellipse to a Rectangle, Mapping a Rectangle to an Ellipse, Mapping a Circle to a Square, Mapping a
Square to a Circle, Squircle, Circle and Square Homeomorphism, Spherize filter
Figure 1: The front page of this paper converted to an elliptical region
1
Introduction
Most of the world’s photographs are rectangular in nature. However, for artistic reasons, one might want to
convert them into circular or elliptical images. In this paper, we will follow-up on mathematical work previously
done on disc-to-square mappings and extend it to handle ellipses and rectangles. We will not alter the eccentricity of
the shapes during the mapping. This means that the aspect ratio of the rectangle’s length and width will be the same
as the aspect ratio of the ellipse’s major and minor axis length.
The main application in mind for this paper is photographic art, but the mathematical mappings discussed can
be used in other applications. For example, we have seen the use for similar mappings for diverse applications such
as mathematical modelling [Holhos 2017], computer graphics, geography [Lambers 2016], finite element analysis
[Shterenlikht 2017], and plasma physics [Citrin 2017].
1.1 Organization of this Paper
We will organize this paper into two main parts. The 1st part will describe algorithms for converting
rectangular regions into elliptical regions. These algorithms will be given like a cookbook recipe without any proofs
or derivations. The 2nd part of the paper will focus on the theory behind the mappings and expound on the derivation
of its equations.
There are accompanying presentation slides and C++ implementation of this paper available on these websites:
http://squircular.blogspot.com
1
1.2 Square to Disc Mapping Space
Figure 2: Canonical mapping space from square to disc
Before we discuss mapping between the rectangle to the ellipse (and vice versa), we have to first discuss the
simpler case of mapping between the square and the circle. The author has written a paper about this titled
“Analytical Methods for Squaring the Disc” [Fong 2014] which covers this extensively. This paper is a follow-up
paper and will require the previous paper as a prerequisite. We shall reuse many of the concepts, equations, and
notation from that paper in our exposition here.
The canonical mapping space between the square and disc is shown in figure 2. It consists of a unit disc
centered at the origin with a square circumscribing it. The square is defined as the set 𝒮 = [−1,1] x [−1,1]. In
contrast, the unit disc is defined as the set 𝒟 = {(𝑢, 𝑣)| 𝑢2 + 𝑣 2 ≤ 1}. We shall denote (x,y) as the point in the
interior of the square and (u,v) as the corresponding point in the interior of the disc after the mapping.
1.3 Rectangle to Ellipse Extension
Figure 3: Canonical mapping space from rectangle to ellipse
The canonical mapping space between the rectangular and elliptical regions is shown in figure 3. It consists of
an ellipse centered at the origin with a rectangle circumscribing it. Both shapes have the aspect ratio a/b. The
rectangular region is defined as the set ℛ = [−𝑎, 𝑎] x [−𝑏, 𝑏]. In contrast, the elliptical region is defined as the
set ℰ = {(𝑢, 𝑣)|
𝑢2
𝑎2
+
𝑣2
𝑏2
≤ 1} .
It is also important to review the definition of the signum function as this point.
sgn(x) =
−1 if x < 0,
|x|
= { 0 if x = 0,
x
1 if x > 0.
2
We will cover a bevy of square-to-disc mappings in the next few pages. We will then proceed to extend these
mappings to work with rectangle-to-ellipse mappings. The key idea to this extension is very simple. Simply remove
the eccentricity of the input rectangle and use any of the given square-to-disc mappings to convert it to a circular
disc. Afterwards, reintroduce the eccentricity back to get an ellipse. Of course, removing and introducing
eccentricity is simply a matter of stretching one axis of the shape by a scale factor.
1.4 The Mappings
In the next few pages, we shall present several mappings for converting the circular disc to a square and vice
versa. We include pictures of a radial grid inside the disc converted to a square; and a square grid converted to a
circular disc. This is followed by forward and inverse equations for the mappings.
The first three mappings have been thoroughly covered and derived in the paper “Analytical Methods for Squaring
the Disc” [Fong 2014]. We only provide them here for reference since they will be used and expanded upon by
subsequent mappings.
The subsequent four mappings covered are all variations of the FG-Squircular mapping. They are all radial and
based on the Fernandez-Guasti squircle. These are the
2-Squircular mapping
3-Squircular mapping
Tapered2 mapping
Tapered4 mapping
The next four mappings covered are variations of the Elliptical Grid mapping. These are also based on the
Fernandez-Guasti squircle, but they are not radial.
Squelched Grid mapping
Vertical squelch mapping
Horizontal squelch mapping
Blended Elliptical Grid mapping
Note that for the sake of brevity, we have not singled out cases when there are divisions by zero in the mapping
equations. For these degenerate cases, just equate x=u, y=v and vice versa whenever there is an unwanted division
by zero in the equations. This usually happens when u=0 or v=0 or both.
3
2-Squircular Mapping
Disc to square mapping:
𝑥=
𝑦=
𝑠𝑔𝑛(𝑢𝑣)
𝑣√2
𝑠𝑔𝑛(𝑢𝑣)
𝑢√2
√1 − √1 − 4𝑢2 𝑣 2
√1 − √1 − 4𝑢2 𝑣 2
Square to disc mapping:
𝑢=
𝑥
√1 + 𝑥 2 𝑦 2
𝑣=
𝑦
√1 + 𝑥 2 𝑦 2
5
3-Squircular Mapping
Disc to square mapping:
𝑥=
𝑠𝑔𝑛(𝑢𝑣) 1 − √1 − 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 )
√
𝑣
2(𝑢2 + 𝑣 2 )
𝑦=
𝑠𝑔𝑛(𝑢𝑣) 1 − √1 − 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 )
√
𝑢
2(𝑢2 + 𝑣 2 )
Square to disc mapping:
𝑢=
𝑠𝑔𝑛(𝑥𝑦) −1 + √1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 )
√
𝑦
2(𝑥 2 + 𝑦 2 )
𝑣=
𝑠𝑔𝑛(𝑥𝑦) −1 + √1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 )
√
𝑥
2(𝑥 2 + 𝑦 2 )
6
Tapered2 Mapping
Disc to square mapping:
𝑥=
𝑠𝑔𝑛(𝑢𝑣) −𝑢2 − 𝑣 2 + √(𝑢2 + 𝑣 2 )[𝑢2 + 𝑣 2 + 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 − 2)]
√
𝑣
2(𝑢2 + 𝑣 2 − 2)
𝑦=
𝑠𝑔𝑛(𝑢𝑣) −𝑢2 − 𝑣 2 + √(𝑢2 + 𝑣 2 )[𝑢2 + 𝑣 2 + 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 − 2)]
√
𝑢
2(𝑢2 + 𝑣 2 − 2)
Square to disc mapping:
𝑢 = 𝑥√
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
(𝑥 2 + 𝑦 2 )(1 − 𝑥 2 𝑦 2 )
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
𝑣 = 𝑦√ 2
(𝑥 + 𝑦 2 )(1 − 𝑥 2 𝑦 2 )
7
Tapered4 Mapping
Disc to square mapping:
𝑥=
𝑠𝑔𝑛(𝑢𝑣) √−𝑢2 − 𝑣 2 + √(𝑢2 + 𝑣 2 )[𝑢2 + 𝑣 2 + 2𝑢2 𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)]
𝑣
𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3
𝑦=
𝑠𝑔𝑛(𝑢𝑣) √−𝑢2 − 𝑣 2 + √(𝑢2 + 𝑣 2 )[𝑢2 + 𝑣 2 + 2𝑢2 𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)]
𝑢
𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3
Square to disc mapping:
𝑢=
𝑣=
𝑠𝑔𝑛(𝑥𝑦)
𝑦√𝑥 2 + 𝑦 2
𝑠𝑔𝑛(𝑥𝑦)
𝑥√𝑥 2
+
𝑦2
√1 − √1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
√1 − √1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
8
Squelched Grid Mapping
Disc to square mapping:
𝑥=
𝑦=
𝑢
√1 − 𝑣 2
𝑣
√1 − 𝑢2
Square to disc mapping:
1 − 𝑦2
𝑢= 𝑥√
1 − 𝑥2𝑦2
𝑣 = 𝑦√
1 − 𝑥2
1 − 𝑥2𝑦2
9
Vertical Squelch Mapping
Disc to square mapping:
𝑥=𝑢
𝑦=
𝑣
√1 − 𝑢2
Square to disc mapping:
𝑢=𝑥
𝑣 = 𝑦 √1 − 𝑥 2
10
Horizontal Squelch Mapping
Disc to square mapping:
𝑥=
𝑢
√1 − 𝑣 2
𝑦=𝑣
Square to disc mapping:
𝑢 = 𝑥 √1 − 𝑦 2
𝑣=𝑦
11
Blended Elliptical Grid Mapping (= 0.5)
Disc to square mapping:
𝑥=
𝑦=
𝑠𝑔𝑛(𝑢)
√2𝛽
𝑠𝑔𝑛(𝑣)
√2𝛽
√𝛽 + 1 + 𝛽𝑢2 − 𝑣 2 − √(𝛽 + 1 + 𝛽𝑢2 − 𝑣 2 )2 − 4𝛽(𝛽 + 1)𝑢2
√𝛽 + 1 − 𝑢2 + 𝛽𝑣 2 − √(𝛽 + 1 − 𝑢2 + 𝛽𝑣 2 )2 − 4𝛽(𝛽 + 1)𝑣 2
Square to disc mapping:
𝑦 2 (𝛽 + 1 − 𝛽𝑦 2 ) − (𝛽 + 1 − 𝛽𝑥 2 )(𝛽 + 1 − 𝛽𝑦 2 )
𝑢 = 𝑥√
𝑥 2 𝑦 2 − (𝛽 + 1 − 𝛽𝑥 2 )(𝛽 + 1 − 𝛽𝑦 2 )
𝑣 = 𝑦√
𝑥 2 (𝛽 + 1 − 𝛽𝑥 2 ) − (𝛽 + 1 − 𝛽𝑥 2 )(𝛽 + 1 − 𝛽𝑦 2 )
𝑥 2 𝑦 2 − (𝛽 + 1 − 𝛽𝑥 2 )(𝛽 + 1 − 𝛽𝑦 2 )
12
1.5 Rectangle to Ellipse Mapping
Given the various square-to-disc mappings, we can extend these to handle rectangles and ellipses.
First, let us cover some notation. Let the input rectangular coordinates be (xo,yo) and the output elliptical
coordinates be (uo,vo). Here are the steps to get equations relating (xo,yo) with (uo,vo).
1) Input: rectangular coordinates (xo,yo)
2) Given the rectangular coordinates, remove the eccentricity and squeeze the coordinates into
square coordinates (x,y)
𝑥=
𝑥𝑜
𝑎
𝑦=
𝑦𝑜
𝑏
3) Choose an appropriate square-to-disc mapping from the list previously provided. Denote this
mapping as two functions go and ho such that
𝑢 = 𝑔𝑜 (𝑥, 𝑦)
v = ℎ𝑜 (𝑥, 𝑦)
4) Convert circular coordinates into elliptical coordinates by reintroducing the eccentricity back.
𝑥
𝑦
𝑥
𝑦
𝑢𝑜 = 𝑎𝑢 = 𝑎 𝑔𝑜 ( 𝑎𝑜 , 𝑏𝑜 )
𝑣𝑜 = 𝑏𝑣 = 𝑏 ℎ𝑜 ( 𝑎𝑜 , 𝑏𝑜 )
5) Output: elliptical coordinates (uo,vo) computed from the equations given above.
These steps are summarized in the diagram provided below.
Figure 4: Rectangle to ellipse pipeline
13
1.6 Ellipse to Rectangle Mapping
We can now reverse the mapping using a similar approach and notation. Let the input elliptical
coordinates be (u1,v1) and the output rectangular coordinates be (x1,y1). Here are the steps to get equations
relating (u1,v1) with (x1,y1).
1) Input: elliptical coordinates (u1,v1)
2) Given the elliptical coordinates, remove the eccentricity and squeeze the coordinates into circular
coordinates (u,v)
𝑢=
𝑢1
𝑎
𝑣=
𝑣1
𝑏
3) Choose an appropriate disc-to-square mapping from the list previously provided. Denote this
mapping as two functions g1 and h1 such that
𝑥 = 𝑔1 (𝑢, 𝑣)
𝑦 = ℎ1 (𝑢, 𝑣)
4) Convert square coordinates into rectangular coordinates by reintroducing the eccentricity back.
𝑢
𝑣
𝑢
𝑣
𝑥1 = 𝑎𝑥 = 𝑎 𝑔1 ( 𝑎1, 𝑏1)
𝑦1 = 𝑏𝑦 = 𝑏 ℎ1 ( 𝑎1, 𝑏1)
5) Output: rectangular coordinates (x1, y1) computed from the equations given above.
These steps are summarized in the diagram provided below.
Figure 5: Ellipse to rectangle pipeline
14
2
Results
We provide elliptification results for several images in Figures 6 & 7. Of course, the simplest way to convert a
rectangular image to an elliptical one is by just cropping out the corner regions. For many images, this method can
produce decent results. In fact, this method has been used for centuries on many paintings and photographs.
However, there are several disadvantages to this approach. First, it loses significant detail near the corners. As a
matter of fact, for imagery with plenty of features near the corners, this might not be acceptable. Second, this
approach is not reversible. It is impossible to invert the process to get back a rectangular image because regions are
entirely cropped out.
We have used each of the extended mappings mentioned in the previous section to convert rectangular imagery into
elliptical regions. The mappings produce very similar results near the central regions of the images. It is only
towards the corner regions that the differences are more pronounced. Even though this might not be evident for
some of the resulting images, each of the mappings provided with this paper are unique. They all have distinct
forward and inverse equations.
We would like to mention here that applying our eccentric extension to the Schwarz-Christoffel mapping is
technically not equivalent to the mapping developed by Schwarz and Christoffel in the 1860s. Their work only
covers mapping the upper half-plane or unit disc to a polygonal region in the complex plane. Their work does not
cover the ellipse at all. Therefore, we shall call this eccentric extension of their mapping as the “Stretched SchwarzChristoffel”
2.1
United States flag
The elliptification of flags is a popular form of artistic stylization. In fact, doing an internet search for national
flags depicted as ellipses produces many results for different nations. Most of these results just resort to cropping.
The U.S. flag is not quite amenable to cropping because it contains much detail near the top left corner. As Figure 6
shows, only 41 of 50 stars in the flag are discernable after cropping. On the other hand, the other mappings in the
figure show all 50 stars in the flag. Of course, this comes with a compromise. There are all sorts of size and shape
distortions in these mappings. In our opinion, the Tapered2 and the Tapered4 mappings produce the best results for
this case.
2.2
A Sunday Afternoon on the Island of La Grande Jatte
We believe that the elliptification of paintings is a legitimate and meaningful form of artistic expression. In
figure 7, we show different mappings applied to a world-famous painting. Georges Seurat’s “A Sunday Afternoon
on the Island of La Grande Jatte” is a well-known painting that is considered to be the origin of the Pointillism
movement in Neo-Impressionist art. It is a beautiful painting that took Seurat over 2 years to complete.
The painting depicts a busy park along the banks of the river Seine. It is rich with corner to corner details that will
simply be lost if cropped. Consequently, this painting as a prime example of a painting that requires more
sophisticated methods if converted into an ellipse. The top right image in figure 7 shows much of the detailed
richness of painting gone if simple cropped. These include interesting features such as the lady’s umbrella, the sail
boat, and most of the pipe-smoking man’s legs. On the other hand, the other mappings preserve most of these
details, albeit with some noticeable distortions.
15
Part II. Mathematical Details
3.1
Review of Concepts and Basic Equations
This paper is the spiritual successor to the paper “Analytical Methods for Squaring the Disc” [Fong 2014]. In this
section, we will review some important concepts and equations from that paper.
The Fernandez-Guasti squircle [Fernandez-Guasti 1992] is an intermediate shape between the circle and the square.
It is a quartic curve with the equation
𝑠2
𝑥 2 + 𝑦2 + 2 𝑥 2𝑦2 = 𝑡 2
𝑡
Linear squircular continuum of the square. This follows from setting s=t in the squircle equation with
𝑥 2 + 𝑦2 − 𝑥 2𝑦2 = 𝑡 2
𝑓𝑜𝑟 𝑡 ∈ [0,1]
In other words, the square and its interior can be represented by this equation
𝑥 2 + 𝑦2 − 𝑥 2𝑦2 ≤ 1
Radial constraint of mappings. All radial disc-to-square mappings satisfy the constraint 𝑦 = 𝑢𝑣 𝑥
Radial mapping linear parametric equation. This is the general form of equation for all linear radial square-to-disc
mappings.
𝑢=𝑡
𝑥
√𝑥 2 +𝑦 2
𝑣=𝑡
𝑦
√𝑥 2 +𝑦 2
The parameter t is some suitable function of x and y with 0 ≤ 𝑡 ≤ 1.
3.2
Rampant Functions
There is a family of functions that we consider important in the study of disc-to-square mappings. In fact, they are so
important that we decided to give this type of functions a name.
Definition: We define a rampant function as a monotonic and continuous function f that is valid in the interval [0, 1]
that have this properties
𝑓(0) = 0
𝑎𝑛𝑑
𝑓(1) = 1
Example: The function 𝑓(𝑡) = 𝑡 𝑛 for n>0 is a rampant function. It is easy to check that 0𝑛 = 0 𝑎𝑛𝑑 1𝑛 = 1 for all
n>0. When n≤0, these conditions will no longer hold.
Example: The function 𝑓(𝑡) = 𝑡√2 − 𝑡 2 is a rampant function. It can be verified that 𝑓 ′ (𝑡) = −
2(𝑥 2 −1)
√2−𝑥 2
> 0 in the
interval [0,1] so f is a monotonic function in that interval.
3
1
3−3𝑥 4
2
2
√6−2𝑥 2
Example: The function 𝑓(𝑡) = 𝑡√ − 𝑡 4 is a rampant function. It can be verified that 𝑓 ′ (𝑡) =
> 0 in the
interval [0,1] so f is a monotonic function in that interval
18
4.1
A Simpler Variation of the FG-Squircular Mapping
The FG-Squircular mapping [Fong 2014] is by no means the only radial mapping between the circular disc
and the square that uses the Fernandez-Guasti squircle. In this section, we shall introduce another radial mapping
based on the Fernandez-Guasti squircle which has much simpler equations.
1st recall the radial mapping linear parametric equations for mappings between the circular disc and the square. All
radial mappings between the disc and the square have an equation of this form, where the parameter t can be
expressed as a suitable function of x and y.
𝑢=𝑡
𝑥
𝑣=𝑡
√𝑥 2 +𝑦 2
𝑦
√𝑥 2 +𝑦 2
Next, recall our condition for the linear squircular continuum of the square. We start from the squircle equation
2
𝑥 2 + 𝑦 2 − 𝑠𝑡2𝑥 2 𝑦 2 = 𝑡 2 and set s=t. This gives the square as 𝒮 = {(𝑥, 𝑦) ∈ ℝ2 | 𝑥 2 + 𝑦 2 − 𝑥 2 𝑦 2 = 𝑡 2 , 0 ≤ 𝑡 ≤ 1}
In other words, we have this basic equation for the FG-squircular mapping:
𝑥 2 + 𝑦2 − 𝑥 2𝑦2 = 𝑡 2
We can actually use any rampant function f for the squircular continuum of the square where s=f(t). For example,
consider what happens if we set s = t2 in the squircle equation instead, we get this condition
𝑥 2 + 𝑦2 − 𝑡 2𝑥 2𝑦2 = 𝑡 2
We collect the t2 terms into one side of the equation to isolate t
𝑡 2 (1 + 𝑥 2 𝑦 2 ) = 𝑥 2 + 𝑦 2 ⇒
𝑥 2 +𝑦 2
𝑡 = √1+𝑥 2𝑦2
and plugging back this t to the radial mapping linear parametric equations.
2
𝑥
2
𝑥 +𝑦
𝑢 = √1+𝑥
2 𝑦2
2
2
𝑥 +𝑦
𝑣 = √1+𝑥
2 𝑦2
√𝑥 2 +𝑦 2
𝑦
√𝑥 2 +𝑦 2
This simplifies to
𝒖=
𝒙
√𝟏 +
𝒗=
𝒙𝟐 𝒚 𝟐
𝒚
√𝟏 + 𝒙𝟐 𝒚𝟐
We now solve for the inverse of these mapping equations.
𝑢=
𝑥
√1 + 𝑥 2 𝑦 2
⟹
⇒
𝑢√1 + 𝑥 2 𝑦 2 = 𝑥
𝑥 2𝑦2 =
𝑥2
−1
𝑢2
⇒
⇒
𝑦2 =
𝑢 2 (1 + 𝑥 2 𝑦 2 ) = 𝑥 2
1
1
−
⇒
𝑢2 𝑥 2
⇒
1 + 𝑥 2𝑦2 =
𝑥2
𝑢2
1
1
𝑦= √ 2− 2
𝑢
𝑥
Substitute y into the other equation
𝑣=
𝑦
√1 + 𝑥 2 𝑦 2
1
1
1
1
√ 2− 2
√ 2− 2 𝑢 1
1
𝑢 𝑥 2 − 𝑢2 √𝑥 2 − 𝑢2
𝑢
𝑥
𝑢
𝑥
=
=
= √ 2− 2= √ 2 2 =
2
𝑥 𝑢
𝑥
𝑥
𝑢 𝑥
𝑥2
1
1
√1 + 𝑥 2 ( 2 − 2 )
√𝑥 2
𝑢
𝑥
𝑢
We now find a polynomial equation for x in terms of u and v.
𝑥 2 − 𝑢2
𝑥 2 − 𝑢2
4
𝑣2 =
⇒
𝑥
=
𝑥4
𝑣2
⇒
𝑥 4 𝑣 2 − 𝑥 2 + 𝑢2 = 0
19
This is a biquadratic equation in which we could solve for x2 using the standard quadratic equation with
coefficients: a = v2
b= -1 and c = u2
𝑥2 =
1 ± √1 − 4𝑢 2 𝑣 2
2𝑣 2
We can then get a quadrant-aware expression for x by taking the square root.
𝒙 =
𝒔𝒈𝒏(𝒖𝒗)
𝒗√𝟐
√𝟏 − √𝟏 − 𝟒𝒖𝟐 𝒗𝟐
=
𝑠𝑔𝑛(𝑢𝑣) 1
1
√ − √ − 𝑢2 𝑣 2
𝑣
2
4
√𝟏 − √𝟏 − 𝟒𝒖𝟐 𝒗𝟐
=
𝑠𝑔𝑛(𝑢𝑣) 1
1
√ − √ − 𝑢2 𝑣 2
𝑢
2
4
Similarly, we get an expression for y as
𝒚 =
𝒔𝒈𝒏(𝒖𝒗)
𝒖√𝟐
In summary, we were able to derive another radial mapping based on the Fernandez-Guasti squircle by using
simple algebraic manipulations on the squircular continuum of the square. We were also able to find simple inverse
equations for the mapping. We shall name this mapping as the 2-Squircular mapping.
Note that this case produces mapping equations that are much simpler than the case s=t. The qualitative results
of both mappings are very similar. In fact, it is quite difficult to tell the two mappings apart visually. See the figures
from part 1 of this paper to compare and contrast this mapping with the FG-Squircular mapping.
20
4.2
A Quasi-Symmetric Variation of the FG-Squircular Mapping
In the previous section, we derived a mapping by using the assumption s=t2. We can take this step further
by considering the case when s=t3. The squircle equation then reduces to 𝑥 2 + 𝑦 2 − 𝑡 4 𝑥 2 𝑦 2 = 𝑡 2 or equivalently,
we have this biquadratic polynomial equation in t:
𝑥 2 𝑦 2 𝒕4 + 𝒕2 − 𝑥 2 − 𝑦 2 = 0
We can solve for t2 using the quadratic equation to get
𝑡2 =
±√4𝑥 4 𝑦 2 + 4𝑥 2 𝑦 4 + 1 − 1
2𝑥 2 𝑦 2
or
𝑡=
±√
±√4𝑥 4 𝑦 2 + 4𝑥 2 𝑦 4 + 1 − 1
2𝑥 2 𝑦 2
±√±√1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 ) − 1
=
𝑥𝑦√2
Plugging back into the radial mapping linear parametric equations and accounting for the 4 different quadrants, we
get these square-to-disc mapping equations
𝒖=
𝒔𝒈𝒏(𝒙𝒚) √−𝟏 + √𝟏 + 𝟒𝒙𝟐 𝒚𝟐 (𝒙𝟐 + 𝒚𝟐 )
𝒗=
𝒔𝒈𝒏(𝒙𝒚) √−𝟏 + √𝟏 + 𝟒𝒙𝟐 𝒚𝟐 (𝒙𝟐 + 𝒚𝟐 )
𝒚
𝒙
𝟐 (𝒙𝟐 + 𝒚𝟐 )
𝟐 (𝒙𝟐 + 𝒚𝟐 )
For the inverse of these mapping equations, we can start by manipulating the u equation
𝑢𝑦√2(𝑥 2 + 𝑦 2 ) = √−1 + √1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 )
⇒
2𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) = −1 + √1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 )
⇒
(2𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) + 1)2 = 1 + 4𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 )
⇒
4𝑢4 𝑦 4 (𝑥 2 + 𝑦 2 )2 + 4𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) + 1 = 4𝑥 4 𝑦 2 + 4𝑥 2 𝑦 4 + 1
⇒
4𝑢4 𝑦 4 (𝑥 4 + 2𝑥 2 𝑦 2 + 𝑦 4 ) + 4𝑢4 𝑥 2 𝑦 2 + 4𝑢4 𝑦 4 = 4𝑥 4 𝑦 2 + 4𝑥 2 𝑦 4
We can divide the equation by 4y2 to get
𝑢4 𝑥 4 𝑦 2 + 2𝑢4 𝑥 2 𝑦 4 + 𝑢4 𝑦 6 + 𝑢4 𝑥 2 + 𝑢4 𝑦 2 = 𝑥 4 + 𝑥 2 𝑦 2
Substitute 𝑦 = 𝑢𝑣 𝑥, to get
𝑢2 𝑣 2 𝑥 6 + 2𝑣 4 𝑥 6 +
𝑣6 6
𝑣2 4
2 2
2 2
4
𝑥
+
𝑢
𝑥
+
𝑣
𝑥
=
𝑥
+
𝑥
𝑢2
𝑢2
2
Multiply the equation by 𝑢𝑥2 to get
𝑢4 𝑣 2 𝑥 4 + 2𝑢2 𝑣 4 𝑥 4 + 𝑣 6 𝑥 4 + 𝑢4 + 𝑢2 𝑣 2 = 𝑢2 𝑥 2 + 𝑣 2 𝑥 2
21
Collect the terms to get this biquadratic equation in x
(𝑢4 𝑣 2 + 2𝑢2 𝑣 4 + 𝑣 6 )𝒙4 − (𝑢2 + 𝑣 2 )𝒙2 + 𝑢4 + 𝑢2 𝑣 2 = 0
𝑣 2 (𝑢2 + 𝑣 2 )2 𝒙4 − (𝑢2 + 𝑣 2 )𝒙2 + 𝑢2 (𝑢2 + 𝑣 2 ) = 0
⇒
We can then divide the equation by (𝑢2 + 𝑣 2 ) to get this simple biquadratic equation in x
𝑣 2 (𝑢2 + 𝑣 2 )𝒙4 − 𝒙2 + 𝑢2 = 0
Thus, we can solve for x2 using the quadratic equation.
𝑥2 =
1 ± √1 − 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 )
2𝑣 2 (𝑢2 + 𝑣 2 )
Taking the square root of this and accounting for the 4 quadrants, we get this equation in x.
𝒙=
𝒔𝒈𝒏(𝒖𝒗) 𝟏 − √𝟏 − 𝟒𝒖𝟐 𝒗𝟐 (𝒖𝟐 + 𝒗𝟐 )
√
𝒗
𝟐(𝒖𝟐 + 𝒗𝟐 )
Substitute 𝑦 = 𝑢𝑣 𝑥, we get the equation for y as
𝒚=
𝒔𝒈𝒏(𝒖𝒗) 𝟏 − √𝟏 − 𝟒𝒖𝟐 𝒗𝟐 (𝒖𝟐 + 𝒗𝟐 )
√
𝒖
𝟐(𝒖𝟐 + 𝒗𝟐 )
If we compare these inverse equations with the forward equations derived earlier, we can see that the two sets of
equations look very similar. In fact, they are almost symmetrical if not for a few changes in sign involved in the
addition/subtraction of terms in the numerator. This quasi-symmetrical property of the forward and inverse
equations is noteworthy and surprising.
Since this mapping comes from the relationship s=t3 in the Fernandez-Guasti squircle, we shall denote this
mapping as the 3-Squircular mapping. This mapping is also radial.
22
5
More Variations of the Radial Squircular Mapping
We can further generalize the result in Section 3.1 by setting s=tn for different exponents n. Not all exponents
are amenable to simplification so we have to carefully select those that produce soluble polynomials. Also, we have
to restrict ourselves to n>0 for the exponents.
Note that we will only solve for square-to-disc mapping equations in this section. We have deliberately left out
the inverse equations because these are particularly gnarly, if at all possible to be written in the form of an explicit
equation. For now, we shall relegate the disc-to-square mapping equations as future work.
The 𝟑𝟐-Squircular Mapping (case s=t3/2 )
5.1
In the case s=t3/2, the squircle equation reduces to 𝑥 2 + 𝑦 2 − 𝑡𝑥 2 𝑦 2 = 𝑡 2 or equivalently, we have this
quadratic polynomial equation in t:
𝒕2 + 𝑥 2 𝑦 2 𝒕 − 𝑥 2 − 𝑦 2 = 0
We can solve for t using the quadratic equation to get
1
1
𝑡 = ±2√𝑥 4 𝑦 4+4𝑥2 +4𝑦 2 − 2𝑥 2 𝑦 2
Thus, the square-to-disc mapping equations are
𝒖=
5.2
𝒙
𝟐√𝒙𝟐 + 𝒚𝟐
(√𝒙𝟒 𝒚𝟒 + 𝟒𝒙𝟐 + 𝟒𝒚𝟐 − 𝒙𝟐 𝒚𝟐 )
The ½-Squircular Mapping
𝒗=
𝒚
𝟐√𝒙𝟐 + 𝒚𝟐
(√𝒙𝟒 𝒚𝟒 + 𝟒𝒙𝟐 + 𝟒𝒚𝟐 − 𝒙𝟐 𝒚𝟐 )
(case s=t1/2)
2 2
In the case s=t1/2, the squircle equation reduces to 𝑥 2 + 𝑦 2 − 𝑥 𝑡𝑦 = 𝑡 2 or equivalently, we have this
depressed cubic polynomial equation in t:
𝒕3 − (𝑥 2 +𝑦 2 ) 𝒕 + 𝑥 2 𝑦 2 = 0
We can solve for t using the cubic formula to get
1
𝑡=
3
1 1
( √729𝑥 4 𝑦4 −4(3𝑥 2 +3𝑦 2 )3 −27
𝑥2 𝑦 2 )
2
3 2
𝑥2 + 𝑦2
+
1
2 2 3
(12√729𝑥 4 𝑦4 −4(3𝑥 2 +3𝑦 2 )3 −27
2𝑥 𝑦 )
Thus, the square-to-disc mapping equations are
𝑢=
𝑣=
𝒙
1 1
√𝒙𝟐 + 𝒚𝟐
𝒚
√𝒙𝟐 + 𝒚
1
3
(3 (2√729𝑥 4 𝑦 4−4(3𝑥 2 +3𝑦2 )3 −27
𝑥 2 𝑦2 ) +
2
1
3
1 1
( ( √729𝑥 4 𝑦 4 −4(3𝑥 2 +3𝑦 2 )3 −27
𝑥2 𝑦 2 )
2
𝟐 3 2
+
𝑥2 + 𝑦2
1
)
1
)
2 2 3
(12√729𝑥 4 𝑦 4−4(3𝑥 2 +3𝑦2 )3 −27
2𝑥 𝑦 )
𝑥2 + 𝑦2
2 2 3
(12√729𝑥 4 𝑦4 −4(3𝑥 2 +3𝑦 2 )3 −27
2𝑥 𝑦 )
23
5.3
The 4-Squircular Mapping
(case s=t4)
In the case s=t4, the squircle equation reduces to 𝑥 2 + 𝑦 2 − 𝑡 6 𝑥 2 𝑦 2 = 𝑡 2 or equivalently, we have this soluble
sextic polynomial equation in t:
𝑥 2 𝑦 2 𝒕6 + 𝒕𝟐 − 𝑥 2 − 𝑦 2 = 0
We can solve for t2 using the cubic formula to get
𝑡2 =
(27
𝑥 6 𝑦 4 + 27
𝑥 4 𝑦 6 + 32√12𝑥 6 𝑦 6 + 81(𝑥 6 𝑦 4 + 𝑥 4 𝑦 6 )2 )
2
2
3𝑥 2 𝑦 2
1
3
1
−
(27
𝑥 6 𝑦 4 + 27
𝑥 4 𝑦 6 + 32√12𝑥 6 𝑦 6 + 81(𝑥 6 𝑦 4 + 𝑥 4 𝑦 6 )2 )
2
2
1
3
Thus, the square-to-disc mapping equations are
5.4
1
3
𝑢=
3
4 6
6 6
6 4
4 6 )2 )
(27𝑥 6 𝑦 4 + 27
1
2 𝑥 𝑦 + 2√12𝑥 𝑦 + 81(𝑥 𝑦 + 𝑥 𝑦
√ 2
−
1
2
2
3𝑥 𝑦
3
√𝒙𝟐 + 𝒚𝟐
27 6 4
27 4 6
3
( 2 𝑥 𝑦 + 2 𝑥 𝑦 + 2√12𝑥 6 𝑦 6 + 81(𝑥 6 𝑦 4 + 𝑥 4 𝑦 6 )2 )
𝑣=
3
4 6
6 6
6 4
4 6 )2 )
(27𝑥 6 𝑦 4 + 27
1
2 𝑥 𝑦 + 2√12𝑥 𝑦 + 81(𝑥 𝑦 + 𝑥 𝑦
√ 2
−
1
2
2
3𝑥 𝑦
3
√𝒙𝟐 + 𝒚𝟐
27 6 4
27 4 6
3
( 2 𝑥 𝑦 + 2 𝑥 𝑦 + 2√12𝑥 6 𝑦 6 + 81(𝑥 6 𝑦 4 + 𝑥 4 𝑦 6 )2 )
𝒙
𝒚
1
3
Other Soluble Exponents
One can easily from previous section that the mapping equations get very complicated quickly when we solve
polynomial equations with higher exponents. Moreover, the Abel-Ruffini theorem states that we cannot get explicit
solutions for general polynomials with degree greater than or equal to 5 (i.e. the quintic). In other words, we can
only choose exponents for s=tn where the effective polynomial equation derived from the squircular continuum is
linear, quadratic, cubic, or quartic.
Furthermore, all these complicated mappings probably produce qualitative results not much different from the
FG-Squircular mapping or the 2-Squircular mapping. This means that we are getting diminishing returns and hardly
any improvement for more complicated formulas.
Therefore, we shall not continue examining other exponents and will only list down the different possible
exponents that produce soluble polynomials equations in t. The table on the next page shows a list of these soluble
exponents along with their polynomial equations. All of these exponents except the last entry can be used to produce
explicit square-to-disc mapping equations.
24
function
s=h(t)
squircular continuum
equation
2
𝑥 + 𝑦2 − 𝑥2𝑦2 = 𝑡2
equivalent polynomial equation
𝑥2 + 𝑦2 − 𝑡 3𝑥2𝑦2 = 𝑡2
effective
type
simple
quadratic
simple
quadratic
general
quadratic
depressed
cubic
cubic
𝑥2 + 𝑦2 − 𝑡 4𝑥2𝑦2 = 𝑡2
biquadratic
𝑥 2 𝑦 2 𝝉2 + 𝝉 − 𝑥 2 − 𝑦 2 = 0
s=t
𝑥2 + 𝑦2 − 𝑡 2𝑥2𝑦2 = 𝑡2
s=t2
𝑥 2 + 𝑦 2 − 𝑡𝑥 2 𝑦 2 = 𝑡 2
s=t3/2
s=t1/2
s=t
𝑥2 + 𝑦2 −
𝑥2𝑦2
𝑡
= 𝑡2
s=t
s=t
4
𝒕2 + 𝑥 2 𝑦 2 𝒕 − 𝑥 2 − 𝑦 2 = 0
𝒕3 − (𝑥 2 +𝑦 2 ) 𝒕 + 𝑥 2 𝑦 2 = 0
𝑥 2 𝑦 2 𝒕3 + 𝒕2 − 𝑥 2 − 𝑦 2 = 0
s=t
5
where 𝜏 = 𝑡 2
𝑥2 + 𝑦2 − 𝑡 6𝑥2𝑦2 = 𝑡2
𝑥2 + 𝑦2 − 𝑡 8𝑥2𝑦2 = 𝑡2
s=t
𝒕2 (1 + 𝑥 2 𝑦 2 ) = 𝑥 2 + 𝑦 2
5/2
3
s=t
𝒕2 = 𝑥 2 + 𝑦 2 − 𝑥 2 𝑦 2
5/4
4/3
𝑥 2 + 𝑦 2 − 𝑥 2 𝑦 2 √𝑡 = 𝑡 2
2
3
𝑥2 + 𝑦2 − 𝑡 𝑥2𝑦2 = 𝑡2
𝑥 2 + 𝑦 2 − 𝑡 2𝑛−2 𝑥 2 𝑦 2 = 𝑡 2
s=t
depressed
bi-cubic
depressed
bi-quartic
depressed
quartic
depressed
cubic
polynomial
𝑥 2 𝑦 2 𝝉3 + 𝝉 − 𝑥 2 − 𝑦 2 = 0
where 𝜏 = 𝑡 2
𝑥 2 𝑦 2 𝝉4 + 𝝉 − 𝑥 2 − 𝑦 2 = 0
where 𝜏 = 𝑡 2
𝝉4 + 𝑥 2 𝑦 2 𝝉 − 𝑥 2 − 𝑦 2 = 0
where 𝜏 = √𝑡
𝝉𝟑 + 𝑥 2 𝑦 2 𝝉 − 𝑥 2 − 𝑦 2 = 0
where 𝜏 3 = 𝑡 2
𝑥 2 𝑦 2 𝒕2𝑛−2 + 𝒕2 − 𝑥 2 − 𝑦 2 = 0
n
25
6.1
The Tapered2 Squircular Mapping
Recall the squircle equation:
𝑥 2 + 𝑦2 −
Consider the rampant function
𝑠2 2 2
𝑥 𝑦 = 𝑡2
𝑡2
𝑠 = 𝑡√2 − 𝑡 2 . We can use this relationship for s and t. Substituting, we get
2
(𝑡√2 − 𝑡 2 ) 2 2
𝑥 +𝑦 −
𝑥 𝑦 = 𝑡2
𝑡2
2
2
⇒
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2 + 𝑥 2 𝑦 2 𝑡 2 = 𝑡 2
We can then solve for t in terms of x and y
𝑡=√
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
1 − 𝑥 2𝑦2
We can substitute this t value to the radial mapping linear parametric equations to get these mapping equations
𝒙𝟐 +𝒚𝟐 −𝟐𝒙𝟐 𝒚𝟐
𝒙𝟐 +𝒚𝟐 −𝟐𝒙𝟐 𝒚𝟐
𝒖 = 𝒙√(𝒙𝟐 +𝒚𝟐 )(𝟏−𝒙𝟐 𝒚𝟐 )
𝒗 = 𝒚√(𝒙𝟐 +𝒚𝟐 )(𝟏−𝒙𝟐 𝒚𝟐 )
To invert this mapping, we start by squaring the equation for u
𝑢2 =
(𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2 )𝑥 2
(1 − 𝑥 2 𝑦 2 )(𝑥 2 + 𝑦 2 )
𝑢2 (1 − 𝑥 2 𝑦 2 )(𝑥 2 + 𝑦 2 ) = (𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2 )𝑥 2
⇒
𝑢2 (𝑥 2 + 𝑦 2 − 𝑥 4 𝑦 2 − 𝑥 2 𝑦 4 ) = 𝑥 4 + 𝑥 2 𝑦 2 − 2𝑥 4 𝑦 2
Substitute 𝑦
⇒
⇒
𝑢2 𝑥 2 + 𝑢2 𝑦 2 − 𝑢2 𝑥 4 𝑦 2 − 𝑢2 𝑥 2 𝑦 4 = 𝑥 4 + 𝑥 2 𝑦 2 − 2𝑥 4 𝑦 2
𝑣
= 𝑢 𝑥 to get
𝑢2 𝑥 2 + 𝑣 2 𝑥 2 − 𝑣 2 𝑥 6 −
Multiply the equation by
𝑢2
𝑥2
𝑣4 6
𝑣2 4
𝑣2 6
4
𝑥
=
𝑥
+
𝑥
−
2
𝑥
𝑢2
𝑢2
𝑢2
𝑢 4 + 𝑢2 𝑣 2 − 𝑢2 𝑣 2 𝑥 4 − 𝑣 4 𝑥 4 = 𝑢 2 𝑥 2 + 𝑣 2 𝑥 2 − 2𝑣 2 𝑥 4
Collect all terms as a polynomial in x
0 = (𝑣 4 + 𝑢2 𝑣 2 − 2𝑣 2 )𝑥 4 + (𝑢2 + 𝑣 2 )𝑥 2 − 𝑢2 (𝑢2 + 𝑣 2 )
This is a biquadratic equation in x with coefficients
𝑎 = 𝑣 4 + 𝑢 2 𝑣 2 − 2𝑣 2
𝑏 = 𝑢2 + 𝑣 2
𝑐 = −𝑢2 (𝑢2 + 𝑣 2 )
We can then solve for x2
𝑥2 =
⇒
−𝑢2 − 𝑣 2 ± √(𝑢2 + 𝑣 2 )2 + 4𝑢2 (𝑢 2 + 𝑣 2 )(𝑣 4 + 𝑢2 𝑣 2 − 2𝑣 2 )
2 (𝑣 4 + 𝑢2 𝑣 2 − 2𝑣 2 )
𝑥2 =
−𝑢2 − 𝑣 2 ± √(𝑢2 + 𝑣 2 )(𝑢2 + 𝑣 2 − 4𝑢2 𝑣 2 (𝑢2 + 𝑣 2 − 2))
2𝑣 2 (𝑢2 + 𝑣 2 − 2)
26
We can then simplify to get a quadrant-aware equation for x as
𝒙=
Using
𝒔𝒈𝒏(𝒖𝒗) −𝒖𝟐 − 𝒗𝟐 ± √(𝒖𝟐 + 𝒗𝟐 )(𝒖𝟐 + 𝒗𝟐 − 𝟒𝒖𝟐 𝒗𝟐 (𝒖𝟐 + 𝒗𝟐 − 𝟐))
√
𝒗
𝟐(𝒖𝟐 + 𝒗𝟐 − 𝟐)
𝑣
𝑦 = 𝑥, we can solve for y as
𝑢
𝒚=
𝒔𝒈𝒏(𝒖𝒗) −𝒖𝟐 − 𝒗𝟐 ± √(𝒖𝟐 + 𝒗𝟐 )(𝒖𝟐 + 𝒗𝟐 − 𝟒𝒖𝟐 𝒗𝟐 (𝒖𝟐 + 𝒗𝟐 − 𝟐))
√
𝒖
𝟐(𝒖𝟐 + 𝒗𝟐 − 𝟐)
We know that this mapping converts circular contours inside the circular disc to squircles inside the square.
However, it is important to see what sort of equation comes out of circles after the mapping given the relationship
𝑠 = 𝑡√2 − 𝑡 2 for the mapping.
𝑢2 + 𝑣 2 = 𝑥 2
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
2
2
2
+
𝑦
=
(𝑥
+
𝑦
)
(𝑥 2 + 𝑦 2 )(1 − 𝑥 2 𝑦 2 )
(𝑥 2 + 𝑦 2 )(1 − 𝑥 2 𝑦 2 )
(𝑥 2 + 𝑦 2 )(1 − 𝑥 2 𝑦 2 )
⇒
𝒖𝟐 + 𝒗𝟐 =
𝒙𝟐 + 𝒚𝟐 − 𝟐𝒙𝟐 𝒚𝟐
𝟏 − 𝒙 𝟐 𝒚𝟐
Thus, from the nonlinear relationship of 𝑠 = 𝑡√2 − 𝑡 2, we get an equation relating (u,v) with (x,y). We will
encounter this equation again later, so it is handy to give it a name. We shall call this as the tapered2 squircular
continuum.
27
6.2
The Tapered4 Squircular Mapping
Recall the squircle equation:
𝑥 2 + 𝑦2 −
Consider the rampant function
3
1
2
2
𝑠 = 𝑡√ −
𝑠2 2 2
𝑥 𝑦 = 𝑡2
𝑡2
𝑡 4 . We can use this relationship for s and t. Substituting, we get
2
𝑥 2 + 𝑦2 −
(𝑡√32
−
1 4
𝑡 )
2
𝑡2
𝑥 2𝑦2 = 𝑡 2
This will simplify into a biquadratic equation in t
1 2 2 4
𝑥 𝑦 𝑡
2
3
− 𝑡 2 + 𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2 = 0
We can then solve for t2 using the quadratic equation
𝑡2 =
1 ± √1 − 2𝑥 2 𝑦 2 (𝑥 2 + 𝑦 2 − 32𝑥 2 𝑦 2 )
𝑥 2𝑦2
=
1 ± √1 − 2𝑥4 𝑦2 − 2𝑥2 𝑦4 + 3𝑥4 𝑦4
𝑥 2𝑦2
1 − √1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑡=√
𝑥2𝑦2
⇒
We can then substitute this t value to the radial mapping linear parametric equations to get these mapping equations
𝒖=
𝒗=
𝒔𝒈𝒏(𝒙𝒚)
𝒚√𝒙𝟐 + 𝒚𝟐
𝒔𝒈𝒏(𝒙𝒚)
𝒙√𝒙𝟐
+
𝒚𝟐
√𝟏 − √𝟏 − 𝟐𝒙𝟒 𝒚𝟐 − 𝟐𝒙𝟐 𝒚𝟒 + 𝟑𝒙𝟒 𝒚𝟒
√𝟏 − √𝟏 − 𝟐𝒙𝟒 𝒚𝟐 − 𝟐𝒙𝟐 𝒚𝟒 + 𝟑𝒙𝟒 𝒚𝟒
To invert this mapping, we start by squaring the equation for u
𝑢2 =
1 − √1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑦(𝑥 2 + 𝑦 2 )
𝑢 2 𝑦 2 (𝑥 2 + 𝑦 2 ) = 1 − √1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) + 1 = −√1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
(𝑢 2 𝑦 2 (𝑥 2 + 𝑦 2 ) + 1)2 = 1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑢4 𝑦 4 (𝑥 2 + 𝑦 2 )2 − 2𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) + 1 = 1 − 2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑢4 𝑦 4 (𝑥 4 + 2𝑥 2 𝑦 2 + 𝑦 4 ) − 2𝑢2 𝑦 2 (𝑥 2 + 𝑦 2 ) = −2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
𝑢4 𝑥 4 𝑦 4 + 2𝑢4 𝑥 2 𝑦 6 + 𝑢4 𝑦 8 − 2𝑢2 𝑥 2 𝑦 2 − 2𝑢2 𝑦 4 = −2𝑥 4 𝑦 2 − 2𝑥 2 𝑦 4 + 3𝑥 4 𝑦 4
28
𝑣
𝑦= 𝑥
Substitute
𝑢4 𝑥 4
𝑢
𝑣4 4
𝑣6
𝑣8
𝑣2
𝑣4
𝑣2
𝑣4
𝑣4
𝑥 + 2𝑢4 𝑥 2 6 𝑥 6 + 𝑢4 8 𝑥 8 − 2𝑢2 𝑥 2 2 𝑥 2 − 2𝑢2 4 𝑥 4 = −2𝑥 4 2 𝑥 2 − 2𝑥 2 4 𝑥 4 + 3𝑥 4 4 𝑥 4
4
𝑢
𝑢
𝑢
𝑢
𝑢
𝑢
𝑢
𝑢
Multiply the equation by
𝑢4
𝑣2𝑥4
𝑢4 𝑣 2 𝑥 4 + 2𝑢2 𝑣 4 𝑥 4 + 𝑣 6 𝑥 4 − 2𝑢4 − 2𝑢2 𝑣 2 = −2𝑢2 𝑥 2 − 2𝑣 2 𝑥 2 + 3𝑣 2 𝑥 4
Collect terms as a polynomial in x
0 = 𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)𝑥 4 + 2(𝑢 2 + 𝑣 2 )𝑥 2 − 2𝑢2 (𝑢2 + 𝑣 2 )
This is a biquadratic equation in x with coefficients
𝑎 = 𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)
𝑏 = 2(𝑢2 + 𝑣 2 )
𝑐 = −2𝑢2 (𝑢2 + 𝑣 2 )
We can then solve for x2
𝑥2 =
⇒
−2(𝑢2 + 𝑣 2 ) ± √4(𝑢2 + 𝑣 2 )2 + 8𝑢2 𝑣 2 (𝑢2 + 𝑣 2 )(𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)]
2𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)
𝑥2 =
−𝑢2 − 𝑣 2 ± √(𝑢2 + 𝑣 2 )[𝑢2 + 𝑣 2 + 2𝑢2 𝑣 2 (𝑢4 + 2𝑢2 𝑣 2 + 𝑣 4 − 3)]
𝑣 2 (𝑢4 + 2𝑢 2 𝑣 2 + 𝑣 4 − 3)
We can then simplify to get a quadrant-aware equation for x as
𝒙=
Using
𝒔𝒈𝒏(𝒖𝒗) √−𝒖𝟐 − 𝒗𝟐 + √(𝒖𝟐 + 𝒗𝟐 )[𝒖𝟐 + 𝒗𝟐 + 𝟐𝒖𝟐 𝒗𝟐 (𝒖𝟒 + 𝟐𝒖𝟐 𝒗𝟐 + 𝒗𝟒 − 𝟑)]
𝒗
𝒖𝟒 + 𝟐𝒖𝟐 𝒗𝟐 + 𝒗𝟒 − 𝟑
𝑣
𝑦 = 𝑥, we can solve for y as
𝒚=
𝑢
𝒔𝒈𝒏(𝒖𝒗) √−𝒖𝟐 − 𝒗𝟐 + √(𝒖𝟐 + 𝒗𝟐 )[𝒖𝟐 + 𝒗𝟐 + 𝟐𝒖𝟐 𝒗𝟐 (𝒖𝟒 + 𝟐𝒖𝟐 𝒗𝟐 + 𝒗𝟒 − 𝟑)]
𝒖
𝒖𝟒 + 𝟐𝒖𝟐 𝒗𝟐 + 𝒗𝟒 − 𝟑
29
7.1
A Cursory Review of the Elliptical Grid Mapping
In this section, we shall come up with variations of the Elliptical Grid mapping [Nowell 2005] [Fong 2014]. Before
we do this, let us do a quick cursory review of the Elliptical Grid mapping. There are two main ideas in the
derivation of the mapping.
The 1st idea is to map vertical lines segments inside the square to vertically-oriented elliptical arcs inside the circular
disc. This is illustrated in figure 7 where a vertical line and its corresponding elliptical arc are shown in red.
Figure 7: Vertical constraint of the Elliptical Grid mapping
This mapping relationship can be summarized by an equation which we shall call the vertical constraint of the
mapping.
𝑢2
𝑣2
1= 2+
𝑥
2 − 𝑥2
Essentially, for each vertical line segment of constant x inside the square, we have a corresponding equation of an
ellipse where the semi-major and semi-minor axis lengths that vary with x. The left and right vertex tips of the
ellipse are set equal to x. The top and bottom vertex tips of the ellipse vary as √2 − 𝑥 2 .
The 2nd main idea of the mapping is to map horizontal lines inside the square to sideways-oriented elliptical
arcs inside the circular disc. This is illustrated in figure 8 where a horizontal lines and its corresponding elliptical arc
are shown in red.
Figure 8: Horizontal constraint of the Elliptical Grid mapping
This mapping relationship can also be summarized by an equation which we shall call the horizontal constraint
of the mapping.
𝑢2
𝑣2
1=
+
2 − 𝑦2 𝑦2
30
Essentially, for each horizontal line segment of constant y in the square, we have a corresponding equation of an
ellipse where the semi-major and semi-minor axis lengths that vary with y. The top and bottom vertex tips of the
ellipse are set equal to y. The top and bottom vertex tips of the ellipses vary as √2 − 𝑦 2 .
Basically, we are mapping the grid of perpendicular vertical and horizontal lines in the square to a grid of elliptical
arcs inside the circular disc. This is illustrated in figure 9. Note that the ellipses get more and more eccentric as x or
y approach zero. Also, the ellipses get more circular as x or y approaches ±1.
Figure 9: Vertical and horizontal grids from the constraints of the mapping
We can then superimpose the vertically-oriented and horizontally-oriented elliptical arcs inside the circular disc
together to create a curvilinear grid of elliptical arcs. This is shown in figure 10. This figure is a visual overview of
the how the Elliptical Grid mapping works
Figure 10: Superimposed vertical and horizontal grids
Mathematically, we want to mix the vertical constraint equation with the horizontal constraint equation to get
algebraic expressions for u and v in terms of x and y.
We can do this by starting with the vertical constraint equation and do some manipulations to isolate u2.
1=
𝑢2
𝑣2
+
𝑥2 2 − 𝑥2
⇒
1−
𝑣2
𝑢2
=
2 − 𝑥2
𝑥2
⇒
𝑢2 = 𝑥 2 (1 −
𝑣2
)
2 − 𝑥2
31
We can then plug this u2 value into the horizontal constraint equation
𝑢2
𝑣2
1=
+
2 − 𝑦2 𝑦2
𝑣2
) 𝑣2
2 − 𝑥2
+ 2
2 − 𝑦2
𝑦
𝑥 2 (1 −
⇒
1=
Multiply both sides of the equation by (2 − 𝑥 2 )(2 − 𝑦 2 )𝑦 2 to remove the fractions and get
(2 − 𝑥 2 )(2 − 𝑦 2 )𝑦 2 = 𝑥 2 𝑦 2 (2 − 𝑥 2 − 𝑣 2 ) + 𝑣 2 (2 − 𝑥 2 )(2 − 𝑦 2 )
After which, we can isolate v2 into one side of the equation
(2 − 𝑥 2 )(2 − 𝑦 2 )𝑦 2 − 𝑥 2 𝑦 2 (2 − 𝑥 2 ) = −𝑥 2 𝑦 2 𝑣 2 + (2 − 𝑥 2 )(2 − 𝑦 2 )𝑣 2
𝑣 2 ((2 − 𝑥 2 )(2 − 𝑦 2 ) − 𝑥 2 𝑦 2 ) = (2 − 𝑥 2 )𝑦 2 (2 − 𝑦 2 − 𝑥 2 )
𝑣2 = 𝑦2
(2 − 𝑥 2 )(2 − 𝑦 2 − 𝑥 2 )
(2 − 𝑥 2 )(2 − 𝑦 2 ) − 𝑥 2 𝑦 2
⇒
𝑣2 = 𝑦2
2 − 𝑥2
𝑣 = 𝑦√
2
⇒
⇒
(2 − 𝑥 2 )(2 − 𝑦 2 − 𝑥 2 )
2 − 𝑥2
2
=
𝑦
4 − 2𝑥 2 − 2𝑦 2
2
𝒗 = 𝒚√𝟏 −
⇒
⇒
𝒙𝟐
𝟐
Similarly in a symmetric fashion, we can solve for u as
𝒖 = 𝒙√𝟏 −
𝒚𝟐
𝟐
This completes our derivation of the Elliptical Grid mapping. We will not derive the inverse equations here since
this has already been done in paper “Analytical Methods for Squaring the Disc” [Fong 2014]
32
7.2
A Simple Squelched Version of the Elliptical Grid mapping
Using the same idea of mapping vertical and horizontal lines to elliptical arcs, we can come up with a
simpler version of the Elliptical grid mapping. We shall name this variant mapping as the Squelched Grid mapping
Recall the vertical constraint equation given in the previous section. We can actually make this simpler by fixing the
top and bottom vertex tip locations of the ellipse at coordinates (0, ±1) to get this constraint equation.
1=
𝑢2 𝑣 2
+
𝑥2 1
The effect of this simplified vertical constraint is shown in figure 11. All the ellipses inside the circular disc
effectively have identical top and bottom vertex tips.
Figure 11: Vertical constraint of the Squelched Grid mapping
Similarly, we can do the same with the horizontal constraint equation from the previous section. We can fix the left
and right vertex tip locations of the ellipse at coordinates (±1,0) to get this constraint equation.
1=
𝑢2 𝑣 2
+
1 𝑦2
The effect of this simplified horizontal constraint is shown in figure 12. All the ellipses inside the circular disc
effectively have identical left and right vertex tips.
Figure 12: Horizontal constraint of the Squelched Grid mapping
From these simpler constraint equations we can easily derive the equations for a disc-to-square mapping
𝑢2 𝑣 2
𝑢2
𝑥2
1
𝑢2
𝒖
2
1= 2+
⇒ 2 = 1 − 𝑣2 ⇒
=
⇒
𝑥
=
⇒
𝒙=
2
2
2
𝑥
1
𝑥
𝑢
1−𝑣
1−𝑣
√𝟏 − 𝒗𝟐
Likewise,
1=
𝑢2 𝑣 2
+
1 𝑦2
⇒
𝑣2
= 1 − 𝑢2 ⇒
𝑦2
𝑦2
1
=
2
𝑣
1 − 𝑢2
⇒ 𝑦2 =
𝑣2
1 − 𝑢2
⇒
𝒚=
𝒗
√𝟏 − 𝒖𝟐
33
To invert this mapping, we start with the vertical constraint equation and isolate u2
𝑢2 𝑣 2
𝑢2
1= 2+
⇒
= 1 − 𝑣2
⇒
𝑢2 = 𝑥 2 (1 − 𝑣 2 )
𝑥
1
𝑥2
And plug it into the horizontal constraint equation
𝑣2
1 = 𝑥 2 (1 − 𝑣 2 ) + 2
𝑦
We can then isolate for v to get
1
1
𝑦 2 (1 − 𝑥 2 )
1 = 𝑥 2 + (−𝑥 2 + 2 ) 𝑣 2
⇒
1 − 𝑥2 = ( 2 − 𝑥2) 𝑣2
⇒
𝑣2 =
(1 − 𝑥 2 𝑦 2 )
𝑦
𝑦
𝒗 = 𝒚√
⇒
𝟏 − 𝒙𝟐
𝟏 − 𝒙 𝟐 𝒚𝟐
Similarly in a symmetric fashion, we can solve for u as
𝟏 − 𝒚𝟐
𝒖 = 𝒙√
𝟏 − 𝒙 𝟐 𝒚𝟐
7.3
Singular points in the Squelched Grid Mapping
By looking at the forward and inverse equations of the Squelched Grid mapping, we can see that it is possibly the
simplest disc-square mapping there is when compared to other mappings. Unfortunately, the Squelched Grid
mapping has a serious drawback. Strictly speaking, it is not a bijective mapping. There exists 4 singular points on
the circular disc and 4 singular points on the square for this mapping. These singular points fall on coordinates
(±1,0) and (0, ±1) for the circular disc, and they fall on corner coordinates (±1, ±1) of the square. These singular
points will produce a division by zero in the mapping equations.
In essence, the disc-to-square equations map a whole perimeter arc segment in a quadrant of the circular disc to a
single corner point on the square. Likewise, the square-to-disc equations map a whole line segment in the perimeter
of the square to vertical or horizontal extremum of the circular disc.
7.4
Squircularity of the Squelched Grid Mapping
We already know from the paper “Analytical Methods for Squaring the Disc” [Fong 2014] that the Elliptical Grid
mapping converts circular contours inside the disc to Fernandez-Guasti squircles inside the square. We now pose the
same question for the Squelched Grid mapping.
It turns out that the Squelched Grid mapping also does the same. We shall show this here. Start with the expression
u2+v2 and substitute using the square-to-disc mapping equations for the Squelched Grid mapping.
𝑢2 + 𝑣 2 = 𝑥 2
1 − 𝑦2
1 − 𝑥2
𝑥 2 (1 − 𝑦 2 ) + 𝑦 2 (1 − 𝑥 2 )
𝑥 2 + 𝑦 2 − 2𝑥 2 𝑦 2
2
+
𝑦
=
=
1 − 𝑥 2𝑦2
1 − 𝑥 2𝑦2
1 − 𝑥 2𝑦 2
1 − 𝑥 2𝑦2
We have already encountered this equation before in our derivation of the Tapered2 Squircular mapping in Section
6.1. This equation is the tapered2 squircular continuum. This means that just like the Tapered2 mapping, the
Squelched Grid mapping converts circular contours inside the disc to Fernandez-Guasti squircles inside the square.
34
7.5
Vertical squelch mapping
The simplified vertical constraint and horizontal constraint equations from the Squelched Grid mapping are actually
independent of each other and hence separable. It is quite possible to come up with mappings using the constraint
equations independently.
For example, the simplified horizontal constraint equation from the Squelched Grid mapping
𝑢2 𝑣 2
1=
+
1 𝑦2
can be used in conjunction with the simple relation u=x to produce another square-to-disc mapping, which we shall
call the Vertical squelch mapping.
1=
𝑢2 𝑣 2
+
1 𝑦2
⇒
𝒚=
𝒗
√𝟏 − 𝒖𝟐
It is quite easy to get the inverse equation.
𝑦=
𝑣
√1 − 𝑢2
⇒ 𝑣 = 𝑦 √1 − 𝑢 2
⇒ 𝒗 = 𝒚 √ 𝟏 − 𝒙𝟐
Furthermore, we can show that this mapping also maps circular contours inside the disc to Fernandez-Guasti
squircles inside the square. Using the equations for u and v given above, we show that
𝑢2 + 𝑣 2 = 𝑥 2 + 𝑦 2 (1 − 𝑥 2 ) = 𝑥 2 + 𝑦 2 − 𝑥 2 𝑦 2
This is just the equation for the linear squircular continuum.
7.6
Horizontal squelch mapping
Similarly, the simplified vertical constraint equation
1=
𝑢2 𝑣 2
+
𝑥2 1
can be used in conjunction with the simple relation v=y to produce another square-to-disc mapping, which we shall
call the Horizontal squelch mapping.
1=
𝑢2 𝑣 2
+
𝑥2 1
⇒
𝒙=
𝒖
√𝟏 − 𝒗𝟐
It is also easy to get the inverse equation.
𝑥=
𝑢
√1 − 𝑣 2
⇒
𝑢 = 𝑥 √1 − 𝑣 2
⇒
𝒖 = 𝒙√𝟏 − 𝒚𝟐
Likewise, it is easy to show that this mapping produces the linear squircular continuum.
𝑢2 + 𝑣 2 = 𝑥 2 (1 − 𝑦 2 ) + 𝑦 2 = 𝑥 2 + 𝑦 2 − 𝑥 2 𝑦 2
35
7.7
Generalized Elliptical Grid mapping
We can further generalize the Elliptical Grid mapping by introducing suitable functions B(x) and A(y) into the
vertical and horizontal constraint equations. Specifically, we have this generalized vertical constraint equation
1=
𝑢2 𝑣 2
+
𝑥 2 𝐵2
1=
𝑢2 𝑣 2
+
𝐴2 𝑦 2
and this generalized horizontal constraint equation.
We proceed to do some manipulations on the generalized vertical constraint equation and isolate u2,
1=
𝑢2 𝑣 2
+
𝑥 2 𝐵2
⇒
1−
𝑣2
𝑢2
=
𝐵2
𝑥2
𝑢2 = 𝑥 2 (1 −
⇒
𝑣2
)
𝐵2
We can then plug this u2 value into the generalized horizontal constraint equation
𝑣2
𝑥 2 (1 − 2 ) 𝑣 2
𝑢2 𝑣 2
𝐵
1= 2+ 2
⇒
1=
+ 2
𝐴
𝑦
𝐴2
𝑦
Multiply both sides of the equation by A2B2y2 to remove the fractions and get
𝐴2 𝐵2 𝑦 2 = 𝑥 2 𝑦 2 (𝐵2 − 𝑣 2 ) + 𝑣 2 𝐴2 𝐵2
After which, we can isolate v2 into one side of the equation
𝐴2 𝐵2 𝑦 2 = 𝑥 2 𝑦 2 (𝐵2 − 𝑣 2 ) + 𝑣 2 𝐴2 𝐵2 = 𝑥 2 𝑦 2 𝐵2 − 𝑥 2 𝑦 2 𝑣 2 + 𝑣 2 𝐴2 𝐵2
𝑣 2 𝐴2 𝐵2 − 𝑥 2 𝑦 2 𝑣 2 = 𝐴2 𝐵2 𝑦 2 − 𝑥 2 𝑦 2 𝐵2
⇒
𝑣2 = 𝑦2
⇒
𝐴2 𝐵2 − 𝑥 2 𝐵2
𝑥 2 𝐵2 − 𝐴2 𝐵2
2
=
𝑦
𝐴2 𝐵 2 − 𝑥 2 𝑦 2
𝑥 2 𝑦 2 − 𝐴2 𝐵2
⇒
𝒙 𝟐 − 𝑨𝟐
𝒗 = 𝒚𝑩√ 𝟐 𝟐
𝒙 𝒚 − 𝑨𝟐 𝑩𝟐
Similarly in a symmetric fashion, we can derive an equation for u in terms x, A, and B
𝒖 = 𝒙𝑨√
7.6
𝒚𝟐 − 𝑩𝟐
𝒙𝟐 𝒚𝟐 − 𝑨𝟐 𝑩𝟐
Blended Elliptical Grid mapping
Using the equations for u and v in the previous section, we can blend the Elliptical Grid mapping and the Squelched
Grid mapping to come up with a mapping that generalizes both. Consider setting functions A and B as follows
𝑩(𝑥) = √𝛽 + 1 − 𝛽𝑥 2
𝑨(𝑦) = √𝛽 + 1 − 𝛽𝑦 2
36
This gives us these vertical and horizontal constraint equations
1=
𝑢2
𝑣2
+
𝑥 2 𝛽 + 1 − 𝛽𝑥 2
𝑎𝑛𝑑
1=
𝑢2
𝑣2
+
𝛽 + 1 − 𝛽𝑦 2 𝑦 2
Here is a blend parameter that has a value between 0 and 1. When is zero, the mapping becomes the Squelched
Grid mapping. When is one, the mapping becomes the Elliptical Grid mapping. In between, it is a blended average
of the two mappings.
It is very easy to derive the square-to-disc mapping equations by using the equations provided in the last part of the
previous section and substituting √𝛽 + 1 − 𝛽𝑦 2 and √𝛽 + 1 − 𝛽𝑥 2 for A & B, respectively to get
𝒚𝟐 (𝜷 + 𝟏 − 𝜷𝒚𝟐 ) − (𝜷 + 𝟏 − 𝜷𝒙𝟐 )(𝜷 + 𝟏 − 𝜷𝒚𝟐 )
𝒖 = 𝒙√
𝒙𝟐 𝒚𝟐 − (𝜷 + 𝟏 − 𝜷𝒙𝟐 )(𝜷 + 𝟏 − 𝜷𝒚𝟐 )
𝒗 = 𝒚√
𝒙𝟐 (𝜷 + 𝟏 − 𝜷𝒙𝟐 ) − (𝜷 + 𝟏 − 𝜷𝒙𝟐 )(𝜷 + 𝟏 − 𝜷𝒚𝟐 )
𝒙𝟐 𝒚𝟐 − (𝜷 + 𝟏 − 𝜷𝒙𝟐 )(𝜷 + 𝟏 − 𝜷𝒚𝟐 )
To invert this mapping, we start with the vertical constraint equation and isolate u
1=
𝑢2
𝑣2
+
𝑥 2 𝛽 + 1 − 𝛽𝑥 2
𝑢2 =
(𝛽 + 1)𝑥 2 − 𝛽𝑥 4 − 𝑥 2 𝑣 2
𝛽 + 1 − 𝛽𝑥 2
⇒
𝑢2
𝑣2
=
1
−
𝑥2
𝛽 + 1 − 𝛽𝑥 2
⇒
𝑢2 = 𝑥 2 −
𝑥 2𝑣2
𝛽 + 1 − 𝛽𝑥 2
⇒
⇒
(𝛽 + 1)𝑢2 − 𝛽𝑢2 𝑥 2 = (𝛽 + 1)𝑥 2 − 𝛽𝑥 4 − 𝑥 2 𝑣 2
⇒
𝛽𝑥 4 − (𝛽 + 1 + 𝛽𝑢2 − 𝑣 2 )𝑥 2 + 𝑢2 (𝛽 + 1) = 0
This is a biquadratic equation in x with coefficients
𝑎=𝛽
𝑏 = −(𝛽 + 1 + 𝛽𝑢 2 − 𝑣 2 )
𝑐 = 𝑢2 (𝛽 + 1)
We can then solve for x as
𝒔𝒈𝒏(𝒖)
√𝜷 + 𝟏 + 𝜷𝒖𝟐 − 𝒗𝟐 − √(𝜷 + 𝟏 + 𝜷𝒖𝟐 − 𝒗𝟐 )𝟐 − 𝟒𝜷(𝜷 + 𝟏)𝒖𝟐
𝒙=
√𝟐𝜷
Similarly in a symmetric fashion, we can solve for y as
𝒔𝒈𝒏(𝒗)
√𝜷 + 𝟏 − 𝒖𝟐 + 𝜷𝒗𝟐 − √(𝜷 + 𝟏 − 𝒖𝟐 + 𝜷𝒗𝟐 )𝟐 − 𝟒𝜷(𝜷 + 𝟏)𝒗𝟐
𝒚=
√𝟐𝜷
37
8 Summary
In this paper, we presented a methodology for generalizing square-to-disc mappings into rectangle-to-ellipse
mappings. The key idea for this methodology is to remove the eccentricity first and then reintroduce it back after the
square-to-disc mappings are performed. We mostly focused on invertible mappings and derived many variants to the
FG-Squircular and Elliptical Grid mappings. We also presented applications for converting rectangular imagery to
elliptical ones.
mapping
2-Squircular
3-Squircular
Tapered2
Tapered4
Squelched Grid
Vertical Squelch
Horizontal Squelch
Blended Elliptical
Grid
key property
offshoot from
notes/comments
simple
equations
quasi-symmetric
equations
good corner
behavior
good corner
behavior
simple
equations
vertical lines
preserved
horizontal lines
preserved
parameterized
blend
FG-Squircular
Radial mapping which has considerable
size distortion near the corners
Very similar to 2-Squircular mapping;
Deemphasizes the four corners
Radial mapping which has less distortion in corners;
Reasonably good when applied to Poincare disk.
Very similar to Tapered2 mapping;
puts emphasis on the four corners
Not a bijective mapping because
the mapping has four singular points
Vertical features remain upright;
Vertically-biased equations have 2 singular points
Horizontal features remain sideways;
Horizontally-biased equations have 2 singular points
Has an adjustable parameter to blend
the Elliptical Grid with the Squelched Grid mappings
FG-Squircular
FG-Squircular
FG-Squircular
Elliptical Grid
Elliptical Grid
Elliptical Grid
Elliptical Grid
38
References
Citrin, J., et. al. 2017. “Tractable Flux-Driven Temperature, Density, and Rotation Profile Evolution with
the Quasilinear Gyrokinetic Transport Model QuaLiKiz”. https://arxiv.org/pdf/1708.01224.pdf
Fernandez-Guasti, M. 1992. “Analytic Geometry of Some Rectilinear Figures” International Journal of
Mathematical. Education in Science and Technology. 23, pp. 895-901
Fong, C., 2014 “Analytic Methods for Squaring the Disc” Seoul ICM 2014.
Fong, C., 2016. “The Conformal Hyperbolic Square and Its Ilk” Proceedings of Bridges 2016:
Mathematics, Music, Art, Architecture, Culture. Jyvaskyla, Finland.
Holhos, A., 2017. “Two Area Preserving Maps from the Square to the p-Ball” Mathematical Modelling and
Analysis, volume 22, 2017 – issue 2. Taylor & Francis Online
Lambers, M., 2016. “Mappings between Sphere, Disc, and Square” Journal of Computer Graphics
Techniques, vol. 5, no 2.
Nowell, P. “Mapping a Square to a Circle” (blog)
http://mathproofs.blogspot.com/2005/07/mapping-square-to-circle.html
Shterenlikht, A., Kashani, M., Alexander N., Williams, G., Crewe, A. 2017 “Autotuning of Isotropic
Hardening Constitutive Models on Real Steel Buckling Data with Finite Element Based Multistart Global
Optimisation on Parallel Computers” Modelling and Simulation in Engineering, volume 2017, article ID
6132106. Hindawi
Taimina, D., 2009. “Crocheting Adventures with Hyperbolic Planes” A.K. Peters/CRC Press. ISBN
9781568814520
Weinstein, E., 2017 “Lamé Curve” from
http://mathworld.wolfram.com/LameCurve.html
Mathworld
–
A
Wolfram
Web
Resource.
39
Appendix A: A Poor Man’s Hyperbolic Square Mapping
We can use any of the disc-to-square mappings presented in this paper to convert the Poincare disk to a square. The
best mapping for such a task is the Schwarz-Christoffel mapping because of its conformal nature [Fong 2016].
However, in our experience, the Schwarz-Christoffel mapping is computational expensive and requires many levels
of iterations for convergence on the corners of the square.
In this appendix, we propose using the Tapered2 mapping instead to convert the Poincare disk to a square. This
mapping is not conformal and will produce noticeable shape distortions to the mapped image. However, the
equations for this mapping are relatively simple and converge at an order of magnitude faster than the SchwarzChristoffel mapping. Moreover, the resulting square looks reasonably good.
The figure above shows the hyperbolic soccer ball [Taimina 2009] pattern on the Poincare disc. We used the
Tapered2 mapping to convert it into a hyperbolic square.
40
Appendix B: Mapping Based on the Lamé Curve
Most of the disc-to-square mappings discussed in this paper are based on the Fernandez-Guasti squircle. However, it
is possible to create disc-to-square mappings that are not rooted on the Fernandez-Guasti squircle. For example, we
will present a mapping based on the Lamé curve [Weinstein 2017] here.
The Lamé curve is a shape very similar to the Fernandez-Guasti squircle. The equation for the Lamé curve with no
eccentricity is
|𝑥|𝑛 + |𝑦|𝑛 = 𝑡 𝑛
This closed curve has the characteristic similar to the Fernandez-Guasti squircle in that it blends the circle with the
square. When n=2, the curve is a circle with radius t. As n tends to infinity, the curve becomes more and more
square-like. In theory, the square cannot be explicitly represented by the Lamé curve because this requires an infinite
power value for n, but in practice, any sufficiently large n can fairly approximate the square shape.
Using the radial constraint linear parametric equation in conjunction with the equation for the Lamé curve, we can
come up with a square-to-disc mapping, that is valid for 0<x<1 and 0<y<1.
𝑢=
𝑣=
𝑥
√𝑥 2 +𝑦
𝑦
√𝑥 2 +𝑦
2
2
½(1−|𝑥|)(1−|𝑦|)
2
2
½(1−|𝑥|)(1−|𝑦|)
(|𝑥|(1−|𝑥|)(1−|𝑦|) + |𝑦|(1−|𝑥|)(1−|𝑦|) )
2
(|𝑥|(1−|𝑥|)(1−|𝑦|) + |𝑦|(1−|𝑥|)(1−|𝑦|) )
2
Note that in the interest of brevity, we have not singled-out the degenerate cases that occur when x or y have
numerical values of 0 or 1 in the equations given above. These cases will result in unwanted infinities in the
equations. For these degenerate cases, simply set u=x and v=y.
We were unable to find explicit equations for the inverse mapping. This underlies the difficulty of using the Lamé
curve in lieu of the Fernandez-Guasti squircle. The Fernandez-Guasti is a quartic curve that involves polynomials
with no higher power than 4, whereas the Lamé curve needs an infinite power to express the square. We believe that
there is no closed-form expression for the inverse of this Lamé-based mapping.
41
Appendix C: Another parametric equation for the Fernandez-Guasti Squircle
Using the Squelched Grid mapping, we can come up with a novel parametric equation for the Fernandez-Guasti
Squircle with radius r. Recall the disc-to-square mapping equations for the Squelched Grid mapping
𝑥=
𝑢
√1 −
𝑣2
𝑦=
𝑣
√1 − 𝑢2
and substitute 𝑢 = cos 𝜃 and 𝑣 = sin 𝜃 into the equations. This will give us a parametric equation for the
Fernandez-Guasti squircle. The parameter for the equations is where 0 ≤ 𝜃 ≤ 2𝜋. The constant r is the radius of
the squircle.
𝒙=
𝒚=
𝒓 𝐜𝐨𝐬 𝜽
√𝟏 − 𝐬𝟏 𝐬𝐢𝐧𝟐 𝜽
𝐫 𝐬𝐢𝐧 𝜽
√𝟏 − 𝐬𝟏 𝐜𝐨𝐬 𝟐 𝜽
Here, s1 is a kind of squareness parameter but it is not identical as the s squareness parameter in the FernandezGuasti squircle equation. It is related to s with this equation
𝑠1 = 𝑠√2 − 𝑠 2
or equivalently
𝑠 = √1 − √1 − 𝑠12
42
Appendix D: Asymmetrical parametric equations for the FG-squircle
Just like with the Squelched Grid mapping, we can come up with a new parametric equations for the
Fernandez-Guasti Squircle by using the disc-to-square mapping equations for the Vertical squelch and Horizontal
squelch mappings.
Parametric equation (vertical version)
𝒙 = 𝒓 𝐜𝐨𝐬 𝜽
𝒚=
𝐫 𝐬𝐢𝐧 𝜽
√𝟏 − 𝒔 𝐜𝐨𝐬 𝟐 𝜽
Parametric equation (horizontal version)
𝒙=
𝒓 𝐜𝐨𝐬 𝜽
√𝟏 − 𝒔 𝐬𝐢𝐧𝟐 𝜽
𝒚 = 𝐫 𝐬𝐢𝐧 𝜽
Note that these parametric equations are asymmetrical; i.e. the expressions for x & y do not share a similar form that
interchangeable in terms of cos and sin , unlike those derived from the Squelched Grid mapping.
43
| 1cs.CV
|
Approximate Gradient Coding via Sparse Random Graphs
arXiv:1711.06771v1 [stat.ML] 17 Nov 2017
Zachary Charlesµ , Dimitris Papailiopoulos , and Jordan Ellenbergµ
µ
Department of Mathematics
Department of Electrical and Computer Engineering
University of Wisconsin-Madison
November 21, 2017
Abstract
Distributed algorithms are often beset by the straggler effect, where the slowest compute nodes in
the system dictate the overall running time. Coding–theoretic techniques have been recently proposed
to mitigate stragglers via algorithmic redundancy. Prior work in coded computation and gradient coding
has mainly focused on exact recovery of the desired output. However, slightly inexact solutions can be
acceptable in applications that are robust to noise, such as model training via gradient-based algorithms.
In this work, we present computationally simple gradient codes based on sparse graphs that guarantee fast
and approximately accurate distributed computation. We demonstrate that sacrificing a small amount
of accuracy can significantly increase algorithmic robustness to stragglers.
1
Introduction
Deploying algorithms on distributed systems has become the de facto choice for scaling-out machine
learning on large data sets. Distributed algorithms generally achieve substantial runtime speedups compared to single node algorithms. Unfortunately, these speedup gains often fall short of the theoretically
optimal speedup when scaling beyond a few tens of compute nodes [5, 19]. This commonly observed
behavior is referred to as the speedup saturation phenomenon. One of the causes of speedup saturation
is the presence of stragglers. These are compute nodes whose runtime is substantially higher than the
average runtime among all nodes.
Several methods for mitigating the effect of stragglers have been recently proposed. These approaches
include replicating jobs across redundant nodes [22] and dropping stragglers in the case that the underlying computation is robust to errors [1, 24]. Moreover, redundant compute nodes can help improve the
performance of distributed model training algorithms, as demonstrated in [4].
Recently, tools from coding theory have gained traction in an effort to mitigate stragglers. Lee et
al. [11] proposed the use of techniques from coding theory to compensate for stragglers and communication bottlenecks in machine learning settings, especially for the computation of linear functions. Li et
al. proposed using coding theory to reduce inter-server communication in the shuffling phase of MapReduce [14]. Then, [21] propose another coding-theoretic algorithm for speeding up distributed matrix
multiplication in heterogeneous clusters. The use of codes for distributed matrix multiplication and linear operations on functions was also studied in [6], which analyzes the trade-off between the flexibility
and sparsity of the code.
In [23], the authors propose gradient coding, a technique to exactly recover the sum of gradients
from a subset of compute nodes. The authors show that k gradients can be recovered from any k − s
compute nodes, as long as each node computes s + 1 gradients. In other words, the algorithm is robust
to s stragglers. Gradient coding is particularly relevant to synchronous distributed learning algorithms
that involve computing sums of gradients, such as mini-batch stochastic gradient descent and full-batch
gradient descent. While some of the gradient code constructions in [23] are randomized, the authors
in [20] use deterministic codes based on expander graphs to achieve similar results.
Most of the above results focus on exact reconstruction of a sum of functions. In many practical
distributed settings, we may only require approximate reconstruction of the sum. For example, parallel
model training in machine learning settings has been shown to be robust to noise [16]. In some scenarios,
noisy gradients may even improve the generalization performance of the trained model [18]. By only
1
approximately reconstructing the desired function, we hope to increase the speed and tolerance to stragglers of our distributed algorithm. In fact, expander graphs, particularly Ramanujan graphs, can be used
for such approximate reconstruction [20]. Unfortunately, expander graphs, especially Ramanujan graphs,
can be expensive to compute in practice, especially for large numbers of compute nodes. Moreover, the
desired parameters of the construction may be constrained according to underlying combinatorial rules.
Our Contributions In this work, we use sparse graphs to create gradient codes capable of efficiently
and accurately computing approximate gradients in a distributed manner. More generally, these codes
can be used to approximately compute any sum of functions in a distributed manner. We formally
introduce the approximate recovery problem using a coding–theoretic interpretation, and present and
analyze two decoding techniques for approximate reconstruction: an optimal decoding algorithm that
has polynomial-time complexity, and a fast decoding method that has linear complexity in the sparsity
of the input.
We focus on two different codes that are efficiently computable and require only a logarithmic number
of tasks per compute node. The first is the Fractional Repetition Code (FRC) proposed in [23]. We
show that FRCs can achieve small or zero error with high probability, even if a constant fraction of
compute nodes are stragglers. However, we show that FRCs are susceptible to adversarial stragglers,
where an adversary can force a subset of the nodes to become stragglers. To get around this issue, we
also present the Bernoulli Gradient Code (BGC) and the regularized Bernoulli Gradient Code (rBGC),
whose constructions are based on sparse random graphs. We show that adversarial straggler selection in
general codes is NP-hard, suggesting that these random codes may perform better then FRCs against
polynomial-time adversaries. We give explicit bounds on the error of BGCs and rBGCs that show that
their potential tolerance to adversaries comes at the expense of a worse average-case error than FRCs. We
provide simulations that support our theoretical results. These simulations show that there is a trade-off
between the decoding complexity of a gradient code and its average- and worst-case performance.
2
2.1
Setup
Notation
In the following, we will denote vectors and matrices in bold and scalars in standard script. For a vector
v ∈ Rm , we will let kvk2 refer to its `2 -norm. For any matrix A, we will let Ai,j denote its (i, j) entry
and let aj denote its jth column. We will also let kAk2 denote its spectral norm while σmin (A) will
denote its smallest singular value. We will let 1m denote the m × 1 all ones vector, while 1n×m will
denote the all ones n × m matrix. We define 0m and 0n×m analogously.
2.2
Problem Statement
In this work, we consider a distributed master-worker setup of n compute nodes, each of which is assigned
a maximum of s tasks. The compute nodes can compute locally assigned tasks and they can send messages
to the master node.
The goal of the master node is to compute the sum of k functions
f (x) =
k
X
fi (x) = f T 1k
(2.1)
i=1
in a distributed way, where fi : Rd → Rw , and any of the local functions fi can be assigned to and
computed locally by any of the n compute nodes. Here, f := [f1 (x), . . . , fk (x)]T . In order to keep local
computation manageable, we assign at most s tasks to each compute node. Due to the straggler effect,
we assume that the master only has access to the output of r < n non-straggler compute nodes. If we
wish to exactly recover f (x), then we need r ≥ k − s + 1 [23]. However, in practice we only need to
approximately recover f (x), which we may be able to do with many fewer non-stragglers.
The above setup is relevant to distributed learning algorithms, where we often wish to find some
model x by minimizing
k
X
`(x) =
`(x; zi ).
i=1
Here {zi }ki=1 are our training samples and `(x; z) is a loss function measuring the accuracy of the model
x with respect to the data point z. In order to find a model that minimizes the sum of losses `(x), we
2
master node
...
worker 1
worker 2
worker n
max. degree = s
f1
...
f2
fk
Figure 1: A master-worker architecture of distributed computation with multiple cores per machines. Each of
the k functions is assigned to a subset of the n compute nodes. Each compute node computes the set of assigned
functions, and transmits a linear combination of them to the master node. The master node aims to reconstruct
the sum of the functions.
often use first-order, or gradient-based methods. For example, if we wanted to use gradient descent to
find a good model, we would need to compute the full gradient of `(x), given by
∇`(x) =
k
X
∇`(x; zi ).
i=1
Letting fi (x) = ∇`(x; zi ), we arrive at the setup in 2.1. We focus our discussion on gradient descent for
simplicity, but we note that the same setup applies for mini-batch SGD.
Our main question is whether we can develop gradient coding techniques that are robust to larger
numbers of straggler nodes if we allow for some error in the reconstruction of ∇`(x). We formally describe
the approximate gradient coding problem, first considered in [20].
Approximate Gradient Coding: There are three components of an approximate gradient coding
scheme:
1. The function assignment per compute node.
2. The messages sent from a compute node to the master.
3. The decoding algorithm used by the master to recover an approximate sum of gradients.
First, each compute node is assigned s functions to compute locally. In some versions of the problem,
we allow compute nodes to compute up to Õ(s) functions, where Õ hides poly-log factors.
After assigning tasks to each compute node, we let the compute nodes run local computations for
some maximum amount of time. Afterwards, we may have compute nodes that either failed to compute
some functions or are still running. These are the straggler nodes. During the approximate reconstruction
of the sum, we assume that the master node only has access to the output of the r non-straggler nodes
out of the n total compute nodes. We want to use their output to compute the best approximation
possible to f (x) given in (2.1). We assume that we can only take linear combinations of the outputs of
the non-straggler nodes.
More formally, the task assignments are represented by a function assignment matrix G, a k × n
matrix where the support of column j indexes the functions assigned to compute node j. The entries
of column j correspond to the coefficients of the linear combination of these local functions that the
compute node sends back to the master once the compute node has completed its local computations.
Let A denote the k × r submatrix of G corresponding to the r non-straggler compute nodes. The
minimum recovery error for a given subset matrix A is given by
min (f T Ax − f T 1k )2 .
x
To better analyze this error, we define the optimal decoding error of a matrix A.
3
(2.2)
Definition 1. The optimal decoding error of a non-straggler matrix A is defined as
err(A) := min kAx − 1k k22 .
x
The optimal decoding error quantifies how close 1k is to being in the span of the columns of A.
Taking x = 0r , we see that for any A, 0 ≤ err(A) ≤ k. It is worth noting that err(A) is the absolute
error incurred in our approximation. The multiplicative error is err(A)/k. Note that if err(A) is small,
then the overall minimum recovery error is small relative to kf k22 , since
min (f T Ax − f T 1k )2 ≤ kf k22 err(A),
x
(2.3)
For a given matrix A, let A+ denote its pseudo-inverse. Properties of the pseudo-inverse imply
err(A) = kAA+ 1k − 1k k22 .
In general, we are interested in constructing function assignment matrices G such that submatrices
A have small decoding error. Note that we can either consider the worst-case among all A or consider
the setting where A is chosen uniformly at random. We will refer to these G matrices as approximate
gradient codes.
Approximate gradient codes were constructed in [20] using expander graphs. The authors show in
particular that if G is the adjacency matrix of a Ramanujan graph, then the worst-case decoding error
is relatively small. Unfortunately, such graphs may be expensive to compute in practice. To circumvent
this issue, we use simplified random constructions.
Our main theorem shows that there is an efficiently computable code that has small or zero decoding
error with high probability, even with a constant fraction of stragglers. We state an informal version of
this theorem below. Theorem 8 below will provide a more general and formal statement, along with the
proof.
Theorem 1. We can assign s = O(log k) tasks to each compute node in such a way that with probability
at least 1 − k1 , we can tolerate Θ(k) randomly chosen stragglers within a multiplicative error of where
log(k)
.
=O
k
In other words, for any k and a constant fraction of stragglers r, there is a code with sparsity Θ(log(k))
that can exactly reconstruct the gradient with high probability. While this code achieves smaller error
for most A than previously designed gradient codes, we show that this comes at the expense of the worstcase error, which can be Θ(k). Here, the worst-case is taken over all possible sets of nodes that become
stragglers. Moreover, this worst-case be computed efficiently by an adversary. On the other hand, we
show that in general, adversarial straggler selection is NP-hard. In order to counter polynomial-time
adversaries, we give another approximate gradient code that utilizes randomness. We also bound the
decoding error of this code. We show the following informal theorem.
Theorem 2. We can randomly assign O(log k) tasks to each compute node in such a way that with
probability at least 1 − k1 , we can tolerate δk randomly chosen stragglers within a multiplicative error of
where
1
=O
.
(1 − δ) log(k)
Theorem 24 below gives a more general and formal statement, along with the proof. In comparison,
work in [20] derives worst-case bounds on err(A) when G is the adjacency matrix of an s-regular expander
graph. Given such a G, denotes its eigenvalues as λ1 ≥ . . . ≥ λk . We will let λ(G) := max{|λ2 |, |λk |}.
Then the aforementioned work proves the following theorem.
Theorem 3 ([20]). Suppose G is a log(k)-regular expander. Then we can tolerate any δk stragglers
within a multiplicative error of where
δλ(G)2
=O
.
(1 − δ) log2 (k)
In particular, if G is a Ramanujan graph, then this becomes = O(δk/(1 − δ)s). This comes at
the expense of having a more computationally difficult construction, as expander graphs, especially
Ramanujan graphs, can be difficult to compute.
4
Decoding: We would like to note that the performance of a gradient code depends in part on the decoding
algorithm, i.e., how we use the output from the non-stragglers to approximate the desired output. In our
setting, we want to use the received non-straggler matrix A to approximate 1k . We give two possible
methods below. We will refer to these as decoding methods because of the parallels to coding theory.
Algorithm 1: One-step decoding.
1
2
3
Input : A k × r non-straggler matrix
A, ρ > 0.
Output: An approximation v to 1k
x = ρ1r ;
v = Ax;
return v;
Algorithm 2: Optimal decoding.
1
2
3
Input : A k × r non-straggler matrix
A.
Output: An approximation v to 1k
x = arg min kAx − 1k k22 ;
v = Ax;
return v;
k
For the one-step decoding error, we will generally consider ρ = rs
. If G has s entries in each column
and row, then we would expect A to have roughly rs
entries
in
each
row. If this holds exactly, then
k
k
setting ρ = rs
will allow us to exactly reconstruct the gradient.
A decoding method analogous to the one-step decoding method was previously used in [20]. Note
that the one-step decoding method is more efficient to compute than the optimal decoding, especially
when A is ill-conditioned or k is large. Moreover, we can apply the one-step decoding method even if
we do not have direct access to A but can compute matrix-vector product Ax. The one-step decoding
method allows us to avoid putting the entire matrix A in to memory of the master compute node in
settings where this is not possible.
It is straightforward to see that if v is the optimal decoding vector of A, then kv − 1k k22 = err(A).
On the other hand, if v is the one-step decoding vector of A then kv − 1k k22 ≥ err(A). We define the
one-step decoding error of A as follows.
Definition 2. For a given ρ > 0, the one-step error of A is defined by
err1 (A) := kρA1r − 1k k22 .
3
Fractional Repetition Codes
We would like to devise a code that achieves small error with high probability in the setting that our
stragglers are chosen randomly. In fact, this can be achieved by the fractional repetition code (FRC) used
in [23]. Note that [23] only considers this code for exact reconstruction of the gradient over all subsets
of stragglers. This code can still be used when we only want approximately reconstruct the sum of k
gradients with high probability.
This scheme works by replicating certain tasks between compute nodes. Suppose that we have k tasks
and k compute nodes and we want each compute node to compute s tasks. Without loss of generality,
we suppose that s divides k. The assignment matrix Gfrac for this scheme is then defined by
1s×s 0s×s 0s×s . . . 0s×s
0s×s 1s×s 0s×s . . .
Gfrac = 0s×s 0s×s 1s×s . . . 0s×s .
.
.
.
.
.
..
..
..
..
..
0s×s 0s×s 0s×s . . . 1s×s
We assume that the k × r matrix Afrac of non-stragglers has columns that are sampled uniformly
without replacement from the k columns of Gfrac . We first compute the expected one-step decoding
error. Let ai denote column i of Afrac .
Lemma 4.
s
T
E[ai aj ] = s2
s
k −k
, i=j
, i 6= j
.
Proof. Fix ai . Since ai has s non-zero entries that are all 1, aTi ai = s. Next, suppose j 6= i. By the
construction of Gfrac , there are only s − 1 columns of Gfrac that are not orthogonal to ai . Note that aj
5
is a duplicate of ai with probability
Therefore, for i 6= j,
s−1
.
k
If this holds, then aTi aj = s, and it is 0 if this does not hold.
E[aTi aj ] = s
Theorem 5. Setting ρ =
k
rs
s−1
k
=
s
s2
− .
k
k
in the one-step decoding method, we have
δk
1
s−1
E [err1 (Afrac )] =
−
.
(1 − δ)s
1−δ
s
Proof. Using linearity of expectation, we have
#
"
2
2
k
2k T
k
Afrac 1r − 1k
= E 2 2 1Tr ATfrac Afrac 1r −
1k Afrac 1r + 1Tk 1k
E
rs
r s
rs
2
k2 T
2k T
1r E[ATfrac Afrac ]1r −
s1r 1r + k
r2 s2
rs
2 X
k
= 2 2
E[aTi aj ] − k.
r s i,j
=
Between step 1 and step 2 we used the fact that the columns of Afrac all have s non-zero entries so
1Tk Afrac = s1r . Applying Lemma 4,
#
"
2
k2 X
k
= 2 2
Afrac 1r − 1k
E[aTi aj ] − k
E
rs
r s i,j
2
2
k2
s
s
= 2 2 rs + r(r − 1)
−
−k
r s
k
k
r2 s2
r2 s
rs2
rs
k2
−
−
+
−k
= 2 2 rs +
r s
k
k
k
k
k
k
k
k2
− − +
rs s r
rs
k k
k
k
=
−1 − +
s r
r
rs
δk
1
s−1
=
−
.
(1 − δ)s
1−δ
s
=
Next, we consider the optimal decoding error of Afrac . Note that each column of Afrac must be equal
to one of the following k/s distinct vectors,
1s
0s
0s
0s
1s
0s
v1 = . , v2 = . , . . . , vk/s = . .
..
..
..
0s
0s
1s
There are s copies of each vi in Gfrac and Afrac has a set of columns given by sampling r of these
without replacement. It is straightforward to see that err(Afrac ) = αs, where α is the number of i such
that vi is not a column of Afrac .
Let Y1 , . . . , Yk/s denote the random variables where Yi indicates whether vi is not a column of A.
Note that we then have
k/s
X
err(A) =
sYi .
(3.1)
i=1
Each Yi is 1 iff none of the s columns in the ith block of Gfrac are sampled as part of the r nonstragglers. Therefore,
k−s
P(Yi = 1) =
r−s
k
r
Combining (3.1) and (3.2), we get the following theorem.
6
.
(3.2)
Theorem 6.
E[err(Afrac )] = k
k−s
r−s
k
r
.
We would now like high-probability bounds on err(A). By (3.1), this reduces to bounding how many
of the Yi are non-zero. This can be done via standard techniques concerning with-replacement sampling.
Theorem 7. If Gfrac is k × k, for all α ∈ Z≥0 ,
k/s
α+1
P err(Afrac ) ≤ αs ≥ 1 −
!
k−(α+1)s
r
k
r
.
Proof. Fix T ⊆ {1, 2, . . . , k/s} such that |T | = α + 1. Then
P ∀i ∈ T, vi is not a column of Afrac =
k−(α+1)s
r
k
r
.
Taking a union bound, this implies
[
P ∀i ∈ T, vi is not a column of Afrac ≤
T :|T |=α+1
k/s
α+1
!
k−(α+1)s
r
k
r
.
Note that the probability that we have no more than α of the vi missing from the columns of A is
the probability that err(Afrac ) ≤ αs. Therefore,
!
k−(α+1)s
k/s
r
.
P err(Afrac ) ≤ αs ≥ 1 −
k
α+1
r
While this exact expression is complicated, this result easily shows that if s = Ω(log(k)), then with
probability at least 1− k1 , err(A) is relatively small. Recall that the number of non-stragglers r = (1−δ)k
for δ ∈ (0, 1).
Theorem 8. Suppose
s≥
Then
1+
1
1+α
log(k)
.
1−δ
1
P err(Afrac ) > αs ≤ .
k
Proof. By Theorem 7, we have
P err(Afrac ) > αs ≤
k/s
α+1
!
k−(α+1)s
r
k
r
.
Simple estimates show
k−(α+1)s
r
k
r
(k − (α + 1)s)(k − (α + 1)s − 1) . . . (k − (α + 1)s − r + 1)
k(k − 1) . . . (k − r + 1)
r
(k − (α + 1)s)
≤
.
k
=
Therefore,
P err(Afrac ) > αs ≤ kα+1
(k − (α + 1)s)
k
r
.
1
We wish to show that for s ≥ (1 + 1+α
) log(k)/(1 − δ), the right-hand side of this equation is at most
Manipulating, this is equivalent to s satisfying
(α + 1)s
≥ 1 − k−(α+2)/r .
k
7
1
.
k
(3.3)
Since s ≥ (1 +
1
) log(k)/(1
1+α
− δ), we have
(α + 1)s
(α + 2) log(k)
≥
.
k
r
(3.4)
Letting β = (α + 2) log(k)/r, (3.4) implies that (3.3) holds if
β ≥ 1 − e−β .
Since this occurs for all β ≥ 0, the desired result is shown.
Corollary 9. Suppose
s≥
2 log(k)
.
1−δ
Then
1
P err(Afrac ) > 0 ≤ .
k
Theorem 8 implies that with probability at least 1 − k1 , an FRC will have multiplicative error where
s
=O
.
k
Therefore, this implies Theorem 1. While FRCs have demonstrably small optimal decoding error when
the stragglers are selected randomly, we will later show that it does not perform well when the stragglers
are selected adversarially. In order to improve our tolerance to adversarial stragglers, we will develop a
coding scheme based on random graphs.
4
Adversarial Stragglers
4.1
Adversarial Stragglers and Fractional Repetition Codes
While FRCs have small average-case error, their worst-case error is large. Even worse, it is computationally efficient to find these worst-case straggler sets. In fact, they can be found in linear time in the
number of compute nodes. Recall that the assignment matrix Gfrac for FRC is defined by
1s×s 0s×s 0s×s . . . 0s×s
0s×s 1s×s 0s×s . . .
0s×s 0s×s 1s×s . . . 0s×s
(4.1)
Gfrac =
.
.
..
..
..
..
..
.
.
.
.
0s×s 0s×s 0s×s . . . 1s×s
As previously noted, each column of Afrac has k/s distinct possibilities,
1s
0s
0s
0s
1s
0s
v1 = . , v2 = . , . . . , vk/s = . .
..
..
..
0s
0s
1s
There are s copies of each vi in Gfrac . As long as one instance of vi is a non-straggler, then we will
recover these s entries in 1k . If all s are stragglers, then our optimal decoding vector x̃ = arg minx kAx −
1k k22 will have zeros in the corresponding s entries. This contributes s to the optimal decoding error.
One can therefore see that if we wish to pick s stragglers adversarially, then we would pick all of the s
copies of some vi above.
Furthermore, the optimal decoding error increases by s if and only if all of the columns in one of the
k/s blocks of Gfrac are all stragglers. Therefore, if one were to select k − r stragglers adversarially, they
would pick all of the s columns from one block, then all of the s columns from another block, and continue
until they had selected k − r stragglers. This corresponds to picking r non-stragglers corresponding to
every column of r/s of the blocks in Gf rac. Here, we assume that s divides r for simplicity.
If Gf rac is given as in (4.1), then we can simply select the first r columns of Gfrac to be nonstragglers. If Gfrac is permuted, then we can simply select all columns corresponding to r/s blocks.
There are therefore (k − r)/s blocks missing from A. Each contributes s to the optimal decoding error.
8
This implies that we have an overall error of k − r. The argument above shows that this is the worst-case
error possible.
Note that the adversary can find this set in O(k) operations if they have full-knowledge that an FRC
scheme is being used with presentation as in (4.1). Even if they do not have this knowledge, an adversary
can check for this coding scheme and find the worst-case straggler set in O(k2 ) operations if they only
have access to the matrix Gfrac . This implies the following theorem.
Theorem 10. Suppose that we assign tasks using a Fractional Repetition Code. In the worst-case, the
non-straggler matrix A will satisfy
err(A) = k − r.
Moreover, worst-case straggler sets can be found in quadratic time.
Suppose that r = (1 − δ)k for some constant δ. Then, the adversarial optimal decoding error is Θ(k).
This is in stark contrast to Theorem 8, which shows that if the stragglers are selected randomly and
s = Ω(log k), then with high probability err(A) = O(log k). Also note that since err1 (A) ≥ err(A), the
adversarial one-step decoding error is Ω(k − r).
4.2
Adversarial Straggler Selection is NP-hard
In this section, we show that in general, adversarial straggler selection is NP-hard. This demonstrates
that adversaries with polynomial-time computations may not be able to find a set of stragglers that
maximizes the decoding error. In such cases, the average-case error may be a more useful indicator of
how well a gradient code performs.
To show that general adversarial selection is NP-hard, we first define two problems.
Definition 3 (Densest k-Subgraph Problem). The k-densest subgraph problem (DkS) asks, given a graph
(V, E), what k-vertex subgraph contains the most edges.
As shown in [2], this problem is NP-hard, even if we restrict to regular graphs. We now formally
define the adversarial straggler problem.
Definition 4 (Adversarial Straggler Problem). Fix a constant ρ > 0. The r-adversarial straggler problem
(r-ASP) asks, given a square matrix G ∈n×n , which column-submatrix A maximizes
kρA1r − 1n k22 .
Note that this is the form that one-step decoding takes. We will show that for any ρ ∈ (0, 23 ), this
problem is NP-hard, even when we restrict to G ∈ {0, 1}k×k with at most s non-zero entries in each
column.
Theorem 11. For any ρ ∈ (0, 23 ), the adversarial straggler problem is NP-hard. This holds even if we
restrict to matrices G ∈k×k with entries in {0, 1} and at most s ≥ 2 non-zero entries per column.
In fact, the proof of our theorem also shows that if we instead consider all matrices G ∈k×n where
k ≥ n, then r-ASP is NP-hard for any ρ > 0.
Proof. We will give a reduction from DkS on d-regular graphs to r-ASP where G is boolean with at most
d non-zero entries per column.
Let (V, E) be a d-regular graph on n vertices. Note that |E| = nd. We want to solve DkS for (V, E).
Let M denote the adjacency matrix of (V, E). Note that DkS is equivalent to
max xT Mx s.t. x ∈ {0, 1}n , kxk0 = k.
x
Let B denote the unsigned incidence matrix of (V, E). That is, B is a |E| × |V | boolean matrix where
the row corresponding to edge e has a 1 in column v iff e is incident to v. We will let C denote the
|E| × |E| matrix given by adding |E| − |V | = n(d − 1) zero columns to B. Note that C is a square nd × nd
boolean matrix with at most d non-zero entries in each column since (V, E) is d-regular.
Let r = k + (n − 1)d. Consider r-ASP on C. This is equivalent to finding a vector x ∈ {0, 1}|E| with
kxk0 = r that maximizes
kρCx − 1nd k22 .
Straightforward computations show
kρCx − 1nd k22 = ρ2 xT CT Cx − 2ρ1Tnd Cx + 1Tnd 1nd
= ρ2 xT CT Cx − 2ρ[d1Tn 0Tn(d−1) ]x + nd.
9
y
where y ∈ {0, 1}n , z ∈ {0, 1}n(d−1) . Note that y corresponds to which of the columns
z
of B we select, while z corresponds to columns of 0 we select. Recall that kyk0 + kzk0 = r = t + n(d − 1).
Note that BT B = M + Id. This implies
M + Id 0
.
CT C =
0
0
Let x =
We then get
kρCx − 1nd k22 = ρ2 xT CT Cx − 2ρ[d1Tn 0Tn(d−1) ]x + nd
= ρ2 yT My + dρ2 yT y − 2ρd1Tn y + nd
= ρ2 yT My + dρ2 kyk0 − 2ρdkyk0 + nd.
Therefore, r-ASP in this setting is equivalent to maximizing, over y ∈ {0, 1}n , z ∈ {0, 1}n(d−1) such
that kyk0 + kzk0 = r = k + n(d − 1), the quantity
ρ2 yT My + dρ2 kyk0 − 2ρdkyk0 + nd.
(4.2)
Define
f (y) = ρ2 yT My + (ρ2 − 2ρ)dkyk0 .
Note that if we fix a binary y such that kyk0 = a, then
f (y) = ρ2 yT My + (ρ2 − 2ρ)da.
Therefore, maximizing this quantity corresponds to finding the a-densest subgraph of (V, E). We now
must show that when we maximize this over y and z, then the solution will always have kyk0 = k. Since
r = k + n(d − 1), it suffices to show that y is as sparse as possible.
To show that this is the case, we will show that for ρ ∈ (0, 32 ), increasing the sparsity of y by 1 will
only decrease the objective function. Suppose that kyk0 = a and it has support S. Say that y0 satisfies
ky0 k0 = a + 1 and it has support S 0 where S ⊆ S 0 . Let T, T 0 denote the vertex subgraphs of (V, E)
corresponding to S, S 0 , and let e(S), e(S 0 ) denote the number of edges in these subgraphs. Note that
yT My = 2e(S). We then have
f (y) = 2ρ2 e(S) + (ρ2 − 2ρ)da.
f (y0 ) = 2ρ2 e(S 0 ) + (ρ2 − 2ρ)d(a + 1).
Note that since (V, E) is d-regular, e(S 0 ) ≤ e(S) + d. Therefore,
f (y0 ) − f (y) = 2ρ2 (e(S 0 ) − e(S)) + (ρ2 − 2ρ)d
≤ 2ρ2 d + ρ2 d − 2ρd
= 3ρ2 d − 2ρd.
For ρ ∈ (0, 32 ), this quantity is negative. Therefore, increasing the sparsity of y will decrease the
objective function f (y). Therefore, the maximum of the r-ASP problem applied to C will have y as
sparse as possible. Since r = k + n(d − 1) and kzk0 ≤ n(d − 1), this implies that the maximum occurs at
kyk0 = k. Let S denote the support of y. The objective function in (4.2) is then equal to
2ρ2 e(S) + dρ2 t − 2ρdt + nd.
(4.3)
This is clearly maximized when S is the set of vertices forming the densest k-subgraph.
5
Bernoulli Gradient Codes
In this section we will consider the case that G has entries that are Bernoulli random variables. For
a given s, k, we will refer to the Bernoulli coding scheme as setting, for i ∈ {1, . . . , k}, j ∈ {1, . . . , n},
Gi,j = Bernoulli(s/k). Intuitively, by injecting randomness in to the construction of G, we improve
our tolerance to adversarial stragglers. This comes as the cost of worse average-case error. While [20]
shows that if G is a Ramanujan graph then we have strong bounds on its adversarial decoding error,
such graphs are notoriously tricky to compute. By using Bernoulli coding, we sacrifice a small amount
of error in order to achieve a much simpler, efficiently computable coding scheme.
10
Suppose that the stragglers are selected uniformly at random. Then, the non-straggler submatrix A
also has Bernoulli random entries. Note that the expected number of tasks assigned to each compute
node is s. This construction will allow us to derive high-probability bounds on the decoding error for
s > log(k). We will later show that we can enforce the desired sparsity s of each column and maintain
the same error. Moreover, enforcing this desired sparsity will let us extend these error bounds to the
setting where s < log(k). In order to get a handle on the decoding error, we first develop a method to
bound the optimal and one-step decoding errors.
5.1
Bounding the Decoding Error
Suppose we have a function assignment matrix G such that the sparsity of each column is exactly or
approximately bounded by s. After performing the local computations on each compute node, we have
access to a k × r submatrix A of the r non-stragglers. We assume that r = (1 − δ)k for some δ ≥ 0.
We would like to derive high probability bounds on err(A) in order to bound the optimal decoding
error of A, as in (2.3). Unfortunately, it is not straightforward to directly bound this error for a random
matrix A since it involves the pseudo-inverse of A. Instead, we will use an algorithmic approach to
bound the optimal decoding error with high probability. The following lemma is adapted from [26].
Lemma 12. Let u0 = 1k , and define
ut = ut−1 −
If ν ≥ kAk22 then
AAT
ut−1 .
ν
lim kut k22 = err(A).
t→∞
Moreover, for all t, kut k22 ≥ err(A).
We refer to the ut as the algorithmic decoding error of A. To prove Lemma 12, we will use the
following lemma, adapted from [26].
Lemma 13. If u is in the column span of A and ν ≥ kAk22 then
σmin (A)2
AAT
I−
u ≤ 1−
kuk2 .
ν
ν
2
Proof of Lemma 12. Fix some t ≥ 1. We can decompose 1k as v+w where v is the orthogonal projection
of 1k on to the column span of A and w is in the nullspace of AT . Note that this implies that (I −
AAT /ν)w = w. Therefore,
t
t
AAT
AAT
I−
(v + w) = I −
v + w.
ν
ν
t
T
v is also in the span of A and orthogonal to w. By Lemma 13,
Since v is in the span of A, I − AA
ν
ut =
t
AAT
I−
v + kwk22 .
ν
2
2 2t
σmin (A)
≤ 1−
kvk22 + kwk22 .
ν
kut k22 =
By construction, kwk22 = minx kAx − 1k k22 , completing the proof.
Note that the ut are defined as the iterates of projected gradient descent. Consider the setting where
ν = kAk22 . The matrix P = AAT /kAk22 is a projection operator (ie. P2 = P), and it projects a
vector on to the column-span of A. By letting ut = ut−1 − Put−1 , Lemma 12 one can show that this
eventually converges to u0 − AA+ u0 . In other words, we eventually converge to the component of u0
that is orthogonal to the range of A. Taking u0 = 1k , this eventually converges to err(A).
We can better understand ut by taking a combinatorial view. Note that A encodes a bipartite graph
with k left vertices and r right vertices, where Aij is 1 iff there is an edge between vertex i on the left and
vertex j on the right. Column j of A corresponds to the incidence of the jth right vertex. In particular,
the degree of vertex j on the right equal the number of tasks computed by compute node j. We can
compute kut k22 in terms of walks on this bipartite graph.
11
Lemma 14. 1Tk (AAT )t 1k equals the number of paths of length 2t from a left vertex to a right vertex.
Proof. Note that (AAT )ij is the number of paths of length 2 from the vertex i to vertex j, where i, j
are both left vertices. More generally, (AAT )tij counts the weighted number of paths of length 2t from
vertex i to vertex j. Therefore, 1k (AAT )t 1k is the weighted number of paths of length 2t from a left
vertex to a left vertex.
Lemma 15. Let at denote the weighted number of walks in the associated bipartite graph of A of length
2t starting and ending at a left vertex. Then
!
!
!
2t a1
2t a2
2t a2t
2
kut k2 = a0 −
+
− ... +
.
1 ν
2 ν2
2t ν 2t
Proof. By direct computation we have
kut k22
2
AAT
=
I−
ut−1
ν
2
2
t
T
AA
=
I−
1k
ν
2
!
T
T
AA
1k
2t
1
k
= 1Tk 1k −
+
1
ν
!
2t 1Tk (AAT )2 1k
− ... +
2
ν2
!
2t 1Tk (AAT )2t 1k
.
ν 2t
2t
By Lemma 14, the result follows.
While ut may be difficult to bound for sufficiently large t, we can handle u1 more directly. Moreover,
as theory and simulations will show, even u1 will give us good bounds on err(A).
5.2
One-step Error of Bernoulli Gradient Codes
Recall that our function assignment matrix G has entries that are Bernoulli random variables with
probability s/k. The non-straggler matrix A is a column submatrix and therefore also has Bernoulli
random entries. We will view A as encoding a bipartite graph with k left vertices and r right vertices.
We say that there is an edge between left vertex i and right vertex j iff Ai,j = 1. Note that E[A] = ks 1k×r .
Therefore, the expected degree of any vertex in the associated bipartite graph is at most s. We will bound
kA − EAk2 for various s and use this to bound err(A). We will use the following lemma.
Lemma 16. Suppose kA − EAk2 ≤ γ. If ρ =
k
rs
then
err1 (A) ≤
γ2k
.
(1 − δ)s2
Proof. By standard norm properties,
k
A1r − 1k
rs
2
=
2
k
k
A1r − EA1r
rs
rs
2
2
k2
≤ 2 2 kA − EAk22 k1r k22
r s
γ 2 k2
≤
rs2
γ2k
=
.
(1 − δ)s2
This approach is analogous to bounding u1 , as the following lemma shows.
p
Lemma 17. Suppose that kA − EAk2 ≤ γ where γ ≤ (1 − δ)s. Then for ν =
u1 ≤
5γ 2 k
.
(1 − δ)s2
12
rs2
,
k
Proof. Recall that for a given ν, u1 is given by
u1 =
AAT
I−
1k
ν
2
.
2
We then have
u1 = 1k −
2
AAT
1k
ν
2
T
= 1k −
A(EA + AT − EAT )
1k
ν
T
≤ 1k −
Note that EA =
s
1
.
k k×r
AEA
1k
ν
Therefore, 1k =
u1 ≤ 1k −
2
+
2
2
A(A − EAT )
1k
ν
2
2
+
2
2
.
2
Using this fact and taking ν =
A(AT − EAT )
1k
ν
+
k
sA
EA1r −
1r
≤
rs
ν
2
T
k
EA1r .
rs
AEAT
1k
ν
2
rs2
,
k
we get
2
1
kAk22 kA − EAk22 k1k k22
ν2
k2
k3
≤ 2 kA − EAk22 + 2 4 kAk22 kA − EAk22 .
rs
r s
Note that since EA =
s
1
,
k k×r
we have
√
s√
s
k1k×r k2 =
kr = 1 − δs.
k
k
p
Using our assumption that kA − EAk2 ≤ γ ≤ (1 − δ)s, we find
kEAk2 =
kAk22 ≤ (kEAk2 + γ)2
p
≤ (2 (1 − δ)s)2
≤ 4(1 − δ)s2 .
Finally, this implies
k2
k3
kA − EAk22 + 2 4 kAk22 kA − EAk22
2
rs
r s
4k3 (1 − δ)s2 γ 2
k2 γ 2
+
≤
rs2
r2 s2
5γ 2 k
=
.
(1 − δ)s2
u1 ≤
Therefore, to bound err1 (A) or u1 , it suffices to bound kA − EAk2 . Moreover, the bounds we get by
either method are within a constant factor of each other. By [7], if s log4 (k), then A will concentrate
well around EA. This bound was later improved by the following result from [13].
Lemma 18. Suppose we have a random Erdős-Rényi graph G(n, p) with adjacency matrix B where
np ≥ log(n). For any α ≥ 1 there exists a universal constant C1 = C1 (α) such that with probability at
least 1 − n−α ,
√
kB − EBk2 ≤ C1 np.
More generally, assume that B is a n×n adjacency matrix where Bi,j is Bernoulli with probability pi,j .
This is sometimes referred to as the inhomogeneous Erdős-Rényi model G(n, (pi,j )). Let p = maxi,j pi,j .
As discussed in [10], Lemma 18 extends to this setting using this definition of p (see section 1.1). While
this result applies directly to n × n adjacency matrices, we can easily extend this to A. This will first
require a basic lemma about the spectral norm of a structured block matrix.
13
Lemma 19. Let D be a n1 × n2 matrix. Suppose that C is a n × n block matrix of the form
0
D
C=
.
DT 0
Then kCk2 = kDk2 .
Proof. Standard properties of singular values imply that kDT Dk2 = kDDT k2 = kDk22 . Moreover,
DDT
0
T
CC =
.
0
DT D
Since the eigenvalues of a block diagonal matrix are given by the eigenvalues of all the blocks,
kCk22 = kCCT k2 = max{kDT Dk2 , kDDT k2 } = kDk22 .
Theorem 20. Let A be a k × r matrix where k ≥ r and Ai,j is Bernoulli with probability s/k. Then for
all α ≥ 1, there exists a universal constant C2 = C2 (α) such that with probability at least 1 − (k + r)−α ,
√
kA − EAk2 ≤ C2 s.
Proof. A encodes the structure of a bipartite graph with k + r vertices. After relabeling, we can denote
these vertices as v1 , . . . , vk , vk+1 , . . . , vk+r where the bipartite blocks are given by {v1 , . . . , vk }, {vk+1 , . . . , vk+r }.
The adjacency matrix B is therefore of the form
0
A
B=
.
AT 0
Note that B comes from an inhomogeneous Erdős-Rényi graph G(k + r, (pi,j )) where pi,j is zero if i and
j are both in {1, . . . , k} or {k + 1, . . . , k + r}, and s/k otherwise. Therefore, p = maxi,j pi,j = s/k. By
Lemma 18 (and the discussion following it), for all α > 0 there exists some universal constant C1 = C1 (α)
such that with probability at least 1 − (k + r)−α ,
r
√
√
(k + r)s
≤ 2C1 s.
kB − EBk2 ≤ C1
k
Here, we used the fact that r ≤ k. Note that EB satisfies
0
EA
EB =
.
EAT
0
Therefore,
kB − EBk2 =
0
(A − EA)T
A − EA
0
2
= kA − EAk2 .
This last equality holds by Lemma 19. Taking C2 =
√
2C1 we conclude the proof.
Combining this with Lemma 16, we get the following theorem.
Theorem 21. Suppose that s ≥ log(n). Then for any α ≥ 1, there is a universal constant C2 = C2 (α)
k
, with probability at least 1 − (k + r)−α ,
such that for ρ = rs
err1 (A) ≤
C22 k
.
(1 − δ)s
Remark 1. Empirically, the same bound holds for other methods of generating G. If we choose the
non-zero support of each column by selecting s indices with or without replacement from {1, . . . , k}, then
we conjecture that the same theorem holds. Unfortunately, standard concentration inequalities are not
enough to prove this result in such settings.
14
5.3
Regularized Bernoulli Gradient Codes
Bernoulli codes have two issues when s < log(k). First, each column only computes s tasks in expectation,
but may have columns with degree up to s + log(k). If s ≥ log(k), this is not an issue as this gives us
sparsity that is O(s). When s < log(k), however, this may be an issue. Second, if s log(k), A may
not concentrate around EA. For example, if s = O(1) then by [9],
s
log k
kAk2 = (1 + o(1))
.
log log k
√
√
On the other hand, EA = p1k×r so kEAk2 = p kr = 1 − δs. For s log(k), this implies that A does
not concentrate as well. Therefore we cannot use Lemma 16 to bound err(A).
Both of these issues have the same cause: vertices whose degree is too large. Fortunately, this issue
of enforcing concentration of sparse graphs has been studied and partially resolved in [10]. They show
that by appropriate regularization of graphs, we can improve their concentration in the sparse setting.
Theorem 22 ([10]). Let B be a random graph from the inhomogeneous Erdős-Rényi model G(n, (pi,j ))
and let p = maxi,j pi,j . For any α ≥ 1, the following holds with probability at least 1 − n−α . Take all
vertices of B with degree larger than 2np and reduce the weights of the edges incident to those vertices in
any way such that they have degree at most np. Let B0 denote the resulting graph. Then
√
kB0 − EBk2 ≤ C3 r3/2 np.
Here C3 is a universal constant. Note that this regularization can be performed analogously on A.
To form A0 , we simply look at all columns with degree more than 2s and change entries in those columns
from 1 to 0 until these columns have degree s. This satisfies the criterion in the above theorem. We can
then use an almost identical version of the proof of Theorem 20 to prove the following theorem.
Theorem 23. There is a universal constant C4 = C4 (α) such that for any α ≥ 1, s ≥ 1, with probability
at least 1 − (k + r)−α ,
√
kA0 − EAk2 ≤ C3 α3/2 s.
We can combine this with Lemma 16 to derive the following theorem concerning err(A0 ). As before,
this bound applies for both the one-step decoding and the optimal decoding.
Theorem 24. For any α ≥ 1, s ≥ 1, and letting ρ =
err1 (A0 ) ≤
k
,
rs
with probability at least 1 − (k + r)−α ,
C32 α3 k
.
(1 − δ)s
By regularizing in the above manner, we ensure that each compute node computes at most 2s tasks
and that our error bound works for all s. Note that in practice, we cannot form A0 from A as we don’t
know A a priori. Therefore we cannot tell the compute nodes to compute the functions corresponding
to A0 . Instead, we can regularize G in the same manner to obtain G0 in the following way such that we
can apply Theorem 24 above. We refer to this code as the regularized Bernoulli Gradient Code, (rBGC).
The construction is simple. We initialize G with each entry Bernoulli(s/k). For each column j with
more than 2s non-zero entries, we randomly set entries to 0 until it has s non-zero entries. A detailed
algorithm is provided below.
Note that the error incurred by an rBGC corresopnds to a multiplicative error where
1
=O
.
(1 − δ)s
This therefore implies Theorem 2.
15
Algorithm 3: Regularized Bernoulli Gradient Code.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
6
6.1
Input : n, k, s.
Output: A k × n regularized function assignment matrix G0 with max degree ≤ 2s
G0 = 0k×n .;
for j = 1 to n do
d = 0;
for i = 1 to k do
G0i,j = Bernoulli(s/k);
d = d + G0i,j ;
end
if d > 2s then
while d > s do
remove a random edge from column j;
d = d − 1;
end
end
end
return G0 ;
Simulations
Decoding Error of Various Coding Schemes
In this section we compare the empirical decoding error of Fractional Repetition Codes (FRCs) and
Bernoulli Gradient Codes (BGC). Recall that we gave two decoding methods, one that corresponding to
the optimal decoding error
err(A) = min kAx − 1k k22
x
and one corresponding to the one-step decoding error
err1 (A) = kρA1r − 1k k22 .
We compare FRCs and BGCs to a coding scheme proposed in [20]. There, Raviv et al. consider the
scheme where G is the adjacency matrix of an s-regular expander graph with k vertices. They show that
for all k × r submatrices A,
λ(G)2 δk
.
err1 (A) ≤
s2 (1 − δ)
Here, λ(G) = max{|λ2 |, |λk |}, where the eigenvalues of G are given by
λ1 ≥ λ2 ≥ . . . ≥ λk .
We would like to construct G to have λ(G) as small as possible. This is achieved by Ramanujan graphs.
In practice, constructing expander graphs with small values of λ is difficult. By taking a random s-regular
graph, however, we can obtain can expander graph with high probability [15]. As k → ∞, λ tends to
the optimal value. In order to generate empirical data, we consider the setting where G is the adjacency
matrix of a random s-regular graph.
Below, we plot the one-step and optimal decoding error err(A) and err1 (A) for these three schemes
when k = 100 and the fraction of stragglers δ varies. In order to normalize the error, we plot err(A)/k
k
in the one-step decoding.
and err1 (A)/k. We take ρ = rs
16
1
0.45
Bernoulli
Random s-regular
Fractional Repetition
0.8
Bernoulli
Random s-regular
Fractional Repetition
0.4
0.35
Error
Error
0.3
0.6
0.4
0.25
0.2
0.15
0.1
0.2
0.05
0
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
0.1
Fraction of stragglers
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Fraction of stragglers
(a) s = 5
(b) s = 10
Figure 2: A plot of the average one-step error err1 (A)/k over 5000 trials. We take k = 100, r = (1 − δ)k for
varying δ. The figure on the left has s = 5 while the figure on the right has s = 10.
We see that under one-step decoding, FRCs and s-regular expanders perform extremely comparably.
In this setting, BGCs seem to sacrifice some accuracy for simplicity. However, FRCs are also computationally simple and perform as well as taking s-regular expanders in the average case under one-step
decoding. For optimal decoding, FRCs perform significantly better than s-regular expanders or BGCs,
as the following plots show.
0.3
0.45
Bernoulli
Random s-regular
Fractional Repetition
0.4
0.35
0.3
0.2
Error
Error
Bernoulli
Random s-regular
Fractional Repetition
0.25
0.25
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
Fraction of stragglers
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Fraction of stragglers
(a) s = 5
(b) s = 10
Figure 3: A plot of the average optimal decoding error err(A)/k over 5000 trials. We take k = 100, r = (1 − δ)k
for varying δ. The figure on the left has s = 5 while the figure on the right has s = 10.
These plots show that if we instead consider optimal decoding, then FRCs greatly outperform the
other two methods. In particular, FRCs can achieve zero optimal decoding error even with a non-trivial
fraction of stragglers. If s = 10, then we can achieve close to zero error even with half of the compute
nodes being stragglers.
Finally, we compare the one-step and optimal decoding error for the BGCs, FRCs, and s-regular
graphs. The results are plotted below.
17
0.8
1
One-step Decoding
Optimal Decoding
0.8
One-step Decoding
Optimal Decoding
0.7
One-step Decoding
Optimal Decoding
0.7
0.4
0.6
0.6
0.5
0.5
Error
0.6
Error
Error
0.8
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.2
0
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
0
0.1
Fraction of stragglers
0.3
0.4
0.5
0.6
0.7
0.8
0
0.25
0.2
0.3
0.3
0.25
0.25
0.2
0.15
0.15
0.1
0.05
0.2
0.3
0.4
0.5
0.6
0.7
0.1
0.05
0.8
0.7
0.8
0.7
0.8
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
0.1
0.2
Fraction of stragglers
Fraction of stragglers
0.3
0.4
0.5
0.6
Fraction of stragglers
(b) s-Regular Graphs
(a) Bernoulli Gradient Code
0.6
0.2
0.1
0
0.1
0.5
0.15
0.05
0
0.4
One-step Decoding
Optimal Decoding
0.35
Error
Error
0.3
0.3
0.4
One-step Decoding
Optimal Decoding
0.35
0.35
0
0.2
Fraction of stragglers
0.4
One-step Decoding
Optimal Decoding
0.4
0.1
Fraction of stragglers
0.45
Error
0.2
(c) Fractional Repetition Code
Figure 4: A plot of the average optimal decoding error err(A)/k and one-step decoding error err1 (A)/k over
5000 trials. We take k = 100, r = (1 − δ)k for varying δ. We compare these two decoding errors for BGCs (left),
s-regular graphs (middle), and FRCs (right). The top plots have s = 5, while the bottom have s = 10.
6.2
Algorithmic Decoding Error of Bernoulli Gradient Codes
In this section we give empirical one-step error rates of BGCs for varying sizes of r and sparsity s of the
matrix A . Recall that we defined the algorithmic decoding error of A by a sequence of vectors ut . Here,
ku1 k22 corresponds to the one-step decoding error, while kut k22 converges to the optimal decoding error.
Let G be constructed via a BGC. We fix k = 100 and take varying values of δ and s, letting r = (1−δ)k.
We then calculate the average value of kut k22 /k based on a Monte Carlo simulation for increasing values
of t. We set ρ = kAk22 . The results are below.
1
1
/ = 0:8
/ = 0:5
/ = 0:1
0.5
kut k2 =k
kut k2 =k
0.5
/ = 0:8
/ = 0:5
/ = 0:1
0.25
0.25
0.1
0.1
0.05
0.05
1
3
5
7
9
11
13
15
17
19
1
t
3
5
7
9
11
13
15
17
19
t
(a) s = 5
(b) s = 10
Figure 5: The average value of kut k22 /k of a BGC for δ ∈ {0.1, 0.2, 0.3, 0.5, 0.8} and varying t for 5000 trials. The
figure on the left plots the algorithmic error for sparsity s = 5, while the figure on the right plots the algorithmic
error for sparsity s = 10.
18
7
Conclusion and Open Problems
In this work, we formally described the approximate gradient coding problem and gave two different
decoding methods for such codes. We analyzed two efficiently computable gradient codes, FRCs and
BGCs, and gave explicit bounds for their decoding error. While FRCs exhibit extremely low averagecase error, they are susceptible to adversaries, even polynomial-time adversaries. This is in contrast
to our result that adversarial straggler selection is NP-hard. BGCs are constructed via sparse random
graphs and are potentially less susceptible to polynomial-time adversaries.
While the problem of approximate gradient coding can be stated in relatively simple terms, our work
shows that the problem is not simple. Bounding the error of gradient codes, even relatively straightforward codes such as BGCs, can be an arduous task. This is especially true in the context of optimal
decoding error. Still, approximate gradient codes have exciting connections to coding theory, concentration of random graphs, expander graphs, and many other interesting combinatorial and algebraic
objects. This work is intended to be a starting point towards understanding the approximate gradient
coding problem. There are many remaining open problems, some directly continuing this work, others
more tangentially related.
Tighter bounds on the optimal decoding error: For both BGCs and s-regular expander graphs,
our empirical results show that there is a significant gap between the one-step and the optimal decoding
error. So far, the only known results on the error of these codes concern their one-step decoding error.
Better bounds on their optimal decoding error and, more generally, better methods for bounding the
optimal decoding error would have great implications for the analysis and design of gradient codes.
Unfortunately, such bounds are not straightforward. They require greater care and may require analysis
of second- and higher-order moments or a better understanding of the pseudo-inverse of random matrices.
Connections to the Restricted Isometry Property: Bounding the decoding error under the
presence of r non-stragglers amounts to bounding kGx − 1k k22 when x is r-sparse and G is sparse. If
we were instead interested in kGx − 1k k1 , then this would bear a strong resemblance to the Restricted
Isometry Property under the `1 norm (RIP-1). This connection is further bolstered by the fact that
expander graphs, well known to have the RIP-1 property, appear to perform well as gradient codes.
Further exploration of the connection between gradient codes and the RIP-1 property could lead to
better techniques for bounding the decoding error and better gradient codes.
Algorithmic error and weighted walk counting: One potential method towards better bounds
on the optimal decoding error comes from Lemma 12 above. As we show, one can bound the optimal
decoding error by a weighted alternating sum involving the number of walks on a bipartite graph. Tight
bounds on the number of walks in a random bipartite graph or deterministic constructions for bipartite
graphs with more explicit fomrulas for the number of walks could lead to much better bounds on the
error.
Approximate gradient coding capacity: Our work gives deterministic and randomized codes
that achieve varying worst-case and average-case errors. In general, we know very little about how these
error rates behave. One can ask: what is the information theoretic limit of an approximate gradient
code, e.g., what is the optimal tradeoff between the recovery error, the probability of recovery, and the
sparsity of the code?
Acknowledgements
The first author was supported in part by the National Science Foundation grant DMS-1502553.
References
[1] Ganesh Ananthanarayanan, Ali Ghodsi, Scott Shenker, and Ion Stoica. Effective straggler mitigation: Attack of the clones. In NSDI, volume 13, pages 185–198, 2013.
[2] Yuichi Asahiro, Refael Hassin, and Kazuo Iwama. Complexity of finding dense subgraphs. Discrete
Applied Mathematics, 121(1):15–26, 2002.
19
[3] Manmohan Chaubey and Erik Saule. Replicated data placement for uncertain scheduling. In Parallel
and Distributed Processing Symposium Workshop (IPDPSW), 2015 IEEE International, pages 464–
472. IEEE, 2015.
[4] Jianmin Chen, Rajat Monga, Samy Bengio, and Rafal Jozefowicz. Revisiting distributed synchronous sgd. arXiv preprint arXiv:1604.00981, 2016.
[5] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior,
Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in
Neural Information Processing Systems, pages 1223–1231, 2012.
[6] Sanghamitra Dutta, Viveck Cadambe, and Pulkit Grover. Short-dot: Computing large linear transforms distributedly using coded short dot products. In Advances In Neural Information Processing
Systems, pages 2100–2108, 2016.
[7] Zoltán Füredi and János Komlós. The eigenvalues of random symmetric matrices. Combinatorica,
1(3):233–241, 1981.
[8] Kristen Gardner, Samuel Zbarsky, Sherwin Doroudi, Mor Harchol-Balter, and Esa Hyytia. Reducing latency via redundant requests: Exact analysis. ACM SIGMETRICS Performance Evaluation
Review, 43(1):347–360, 2015.
[9] Michael Krivelevich and Benny Sudakov. The largest eigenvalue of sparse random graphs. Comb.
Probab. Comput., 12(1):61–72, January 2003.
[10] Can M Le, Elizaveta Levina, and Roman Vershynin. Concentration and regularization of random
graphs. Random Structures & Algorithms, 2017.
[11] Kangwook Lee, Maximilian Lam, Ramtin Pedarsani, Dimitris Papailiopoulos, and Kannan Ramchandran. Speeding up distributed machine learning using codes. In Information Theory (ISIT),
2016 IEEE International Symposium on, pages 1143–1147. IEEE, 2016.
[12] Kangwook Lee, Ramtin Pedarsani, Dimitris Papailiopoulos, and Kannan Ramchandran. Coded computation for multicore setups. In Information Theory (ISIT), 2017 IEEE International Symposium
on, pages 2413–2417. IEEE, 2017.
[13] Jing Lei, Alessandro Rinaldo, et al. Consistency of spectral clustering in stochastic block models.
The Annals of Statistics, 43(1):215–237, 2015.
[14] Songze Li, Mohammad Ali Maddah-Ali, and A Salman Avestimehr. Coded mapreduce. In Communication, Control, and Computing (Allerton), 2015 53rd Annual Allerton Conference on, pages
964–971. IEEE, 2015.
[15] Alexander Lubotzky. Expander graphs in pure and applied mathematics. Bulletin of the American
Mathematical Society, 49(1):113–162, 2012.
[16] Horia Mania, Xinghao Pan, Dimitris Papailiopoulos, Benjamin Recht, Kannan Ramchandran, and
Michael I Jordan. Perturbed iterate analysis for asynchronous stochastic optimization. arXiv
preprint arXiv:1507.06970, 2015.
[17] Jorma K Merikoski, Ravinder Kumar, and Ram Asrey Rajput. Upper bounds for the largest eigenvalues of a bipartite graph. Electronic Journal of Linear Algebra, 26(1):13, 2013.
[18] Arvind Neelakantan, Luke Vilnis, Quoc V Le, Ilya Sutskever, Lukasz Kaiser, Karol Kurach, and
James Martens. Adding gradient noise improves learning for very deep networks. arXiv preprint
arXiv:1511.06807, 2015.
[19] Hang Qi, Evan Sparks, and Ameet Talwalkar. Paleo: A performance model for deep neural networks.
In International Conference on Learning Representations, 2017.
[20] Netanel Raviv, Itzhak Tamo, Rashish Tandon, and Alexandros G Dimakis. Gradient coding from
cyclic mds codes and expander graphs. arXiv preprint arXiv:1707.03858, 2017.
[21] Amirhossein Reisizadehmobarakeh, Saurav Prakash, Ramtin Pedarsani, and Salman Avestimehr.
Coded computation over heterogeneous clusters. arXiv preprint arXiv:1701.05973, 2017.
[22] Nihar B Shah, Kangwook Lee, and Kannan Ramchandran. When do redundant requests reduce
latency? IEEE Transactions on Communications, 64(2):715–722, 2016.
[23] Rashish Tandon, Qi Lei, Alexandros G Dimakis, and Nikos Karampatziakis. Gradient coding. arXiv
preprint arXiv:1612.03301, 2016.
20
[24] Da Wang, Gauri Joshi, and Gregory Wornell. Efficient task replication for fast response times in
parallel computation. In ACM SIGMETRICS Performance Evaluation Review, volume 42, pages
599–600. ACM, 2014.
[25] Matei Zaharia, Andy Konwinski, Anthony D Joseph, Randy H Katz, and Ion Stoica. Improving
mapreduce performance in heterogeneous environments. In Osdi, volume 8, page 7, 2008.
[26] Anastasios Zouzias and Nikolaos M Freris. Randomized extended kaczmarz for solving least squares.
SIAM Journal on Matrix Analysis and Applications, 34(2):773–793, 2013.
21
| 7cs.IT
|
1
Localization of Multiple Targets with Identical
Radar Signatures in Multipath Environments with
Correlated Blocking
arXiv:1711.01293v1 [cs.IT] 3 Nov 2017
Sundar Aditya, Student Member, IEEE, Andreas F. Molisch, Fellow, IEEE, Naif Rabeah, Member, IEEE and
Hatim Behairy
Abstract—This paper addresses the problem of localizing an
unknown number of targets, all having the same radar signature,
by a distributed MIMO radar consisting of single antenna
transmitters and receivers that cannot determine directions of
departure and arrival. Furthermore, we consider the presence
of multipath propagation, and the possible (correlated) blocking
of the direct paths (going from the transmitter and reflecting
off a target to the receiver). In its most general form, this
problem can be cast as a Bayesian estimation problem where
every multipath component is accounted for. However, when the
environment map is unknown, this problem is ill-posed and hence,
a tractable approximation is derived where only direct paths are
accounted for. In particular, we take into account the correlated
blocking by scatterers in the environment which appears as a
prior term in the Bayesian estimation framework. A sub-optimal
polynomial-time algorithm to solve the Bayesian multi-target
localization problem with correlated blocking is proposed and
its performance is evaluated using simulations. We found that
when correlated blocking is severe, assuming the blocking events
to be independent and having constant probability (as was done
in previous papers) resulted in poor detection performance, with
false alarms more likely to occur than detections.
Index Terms—Multi-target localization, Correlated blocking,
distributed MIMO radar, data association, matching
I. I NTRODUCTION
Over the past decade, there has been an increasing demand for accurate indoor localization solutions to support a
multitude of applications, ranging from providing locationbased advertising to users in a shopping mall [1], to better
solutions for tracking inventory in a warehouse [2], to assistedliving applications [3]. In many of these applications, the
availability of location information enhances communication
protocols (e.g., in a warehouse, communication with a tag
can take place once it is within range of a reader) and the
importance of localization to such communications systems
has led to standardization efforts such as IEEE 802.15.4a [4],
which is a standard for joint localization and communication.
Sundar Aditya and Andreas F. Molisch are with the Wirless Devices and
Systems (WiDeS) Lab, Ming Hsieh Dept. of Electrical Engineering, USC,
Los Angeles, CA 90089, USA (email:{sundarad,molisch}@usc.edu).
Naif Rabeah and Hatim Behairy are with the King Abdulaziz City for
Science and Technology (KACST), P. O. Box 6086, Riyadh 11442, Saudi
Arabia (email: {nbnrabeah,hbehairy}@kacst.edu/sa) Revised: November 7,
2017.
This work was supported by KACST under grant number 33-878. Part of
this work was presented at the International Conference on Communications
(ICC) Workshop on Advances in Network Localization and Navigation
(ANLN), held at London in June, 2015.
While localization of cellular devices falls under the category
of active localization (where the target transmits a ranging
signal), the current paper will focus on the important case
of passive localization, where the target only reflects radiofrequency (RF) signals (i.e., radar). In the above-mentioned
applications, there are usually multiple targets present that
have nearly identical radar signatures and hence, cannot be
distinguished on that basis alone.
The typical localization architecture involves the deployment of multiple transmitters (TXs) and receivers (RXs) and
hence, can be modeled using the distributed MIMO (multipleinput multiple-output) radar framework [5]. Due to cost and
space constraints (e.g., in sensor applications), each TX and
RX may be equipped with a single antenna only1 . Hence, the
radar cannot exploit the information contained in the angles of
arrival/departure from the target-reflected signal as direction
finding requires not just multiple antennas, but also careful
calibration of the antenna elements, which might be costprohibitive in the case of sensor nodes and similar devices.
This motivates the study of multi-target localization (MTL)
without angular information in an indoor distributed MIMO
radar setting.
In such a setting, a direct path (DP) is one that propagates
directly from TX to target to RX. Each DP gives rise to
an ellipse-shaped ambiguity region passing through the target
location, with the TX and RX at the foci, and the intersection
of three or more such curves unambiguously determine the
target location. For indoor localization, the following additional challenges arise - (i) targets can be blocked by nontarget scatterers such as walls, furniture etc., (ii) the scatterers
can also give rise to indirect paths (IPs) which need to be
distinguished from DPs, (iii) in the presence of multiple
targets, yet another challenge is to match the DPs to the right
targets2 . An incorrect matching would result in ghost targets
[8].
There has been a considerable amount of literature on
MIMO radar over the last decade. The fundamental limits
of localization in MIMO radar networks were studied in [9].
A number of works have dealt with MTL using co-located
antenna arrays at the TX and RX [10]–[22]. The single1 Since a major contributor to the cost is the amplifier in a RF chain, a
cheaper distributed MIMO architecture can be realized by deploying a single
RF chain each for the TX and RX antennas and switching between them.
2 This process is also referred to as data association [6], [7]. Throughout
this work, we shall use the term matching to refer to data association.
2
Fig. 1: Correlated blocking: An example
target localization problem using widely-spaced antenna arrays
was investigated in [23], [24] and the multi-target case in
[25]. None of these works address the issues of blocking
and multipath common in an indoor environment. The works
closest to ours are [26] and [8], where MTL in a distributed
MIMO radar setting is addressed. The experiments and the
system model in [26] do not consider the effect of blocking
and a brute-force method is used for matching the DPs to
targets, which is computationally infeasible for a large number
of targets, as shown in Section IV. On the other hand, [8]
considers the effect of blocking, but relies on the assumption
of a constant and independent blocking probability for all DPs.
In reality however, the DP blocking events in any environment
are not mutually independent. As shown in Fig. 1, the location
of the two TXs is such that if one of them has line-of-sight
(LoS) to the target, it is highly likely that the other does as
well. Similarly, if one of them is blocked with respect to the
target, it is highly likely that the other is as well. In other
words, the DP blocking events are, in general, correlated and
the extent of correlation depends on the network geometry.
In this work, we investigate how correlated blocking can be
exploited to obtain better location estimates for the targets.
Intuitively, our approach works as follows: when three or
more ellipses intersect at a point, we first assume that they are
DPs. We then compute the joint probability that LoS exists
to the TXs and RXs in question at the point of intersection.
If the probability is sufficiently high, then we conclude that a
target is present.
The main contributions of this work are as follows:
• The general problem of localizing all targets and scatterers in an unknown environment is cast as a Bayesian
estimation problem. Such a fundamental formulation of
the problem goes beyond the description in [8] (and other
papers, to the best of our knowledge).
• We show this problem to be ill-posed and proceed to
derive a tractable approximation called the Bayesian MTL
problem, where the objective is to localize only the
targets, but not the scatterers. This is also a Bayesian
estimation problem where the joint DP blocking distribution plays the role of a prior.
• We propose a polynomial time approximate algorithm to
solve the Bayesian MTL problem, which can be used
even when only empirical blocking statistics, obtained
via measurements or simulations, are available.
This paper consists of six sections. In the system model in
Section II, we define decision variables to decide if a multipath
component (MPC) is a DP, IP or a noise peak. The generalized
problem of localizing all targets and scatterers is formulated
as a Bayesian estimation problem in Section III, where along
with the target and scatterer locations, the aforementioned
decision variables are the estimation parameters. Furthermore,
this problem is shown to be ill-posed and a more tractable
approximation called the Bayesian MTL problem is derived,
where the objective is to localize only the targets. In Section
IV, the brute force solution to the Bayesian MTL problem is
shown to have exponential complexity in the number of targets
and TX-RX pairs (TRPs). As a result, a sub-optimal polynomial time algorithm taking correlated blocking into account is
proposed instead. Simulation as well as experimental results
for the proposed algorithm are presented in Section V and
finally, Section VI concludes the paper.
Notation: Vectors and scalars are represented using bold
(e.g., x) and regular (e.g., x) lowercase letters, respectively.
In particular, 1 denotes the all-one vector. For a collection of
scalars {aij : i ∈ J1 , j ∈ J2 }, where J1 and J2 are discrete
index sets, vec(aij ) denotes the column vector containing all
aij , ordered first
Q to index i, followed by j and so
P according
on. Similarly,
and
respectively denote summation and
i,j
i,j
product over index i, followed by j and so on. For positive
integers a and b, a mod b denotes the modulo operator, i.e.,
the remainder of the operation a/b. The set of real numbers
is denoted by R. For x ∈ R, ⌊x⌋ denotes the greatest integer
lesser than or equal to x. For continuous random variables
X and Y , f (X, Y ) denotes their joint probability density
function (pdf), f (X) the marginal pdf of X, and f (X|Y )
the conditional pdf of X, given Y . P(.) and E[.] denote the
probability and expectation operators, respectively.
II. S YSTEM M ODEL
In this section, we introduce the formal notation to decide
the identity of a MPC and model the correlated blocking of
DPs. This lays the groundwork for the fundamental problem
formulation that will then be solved in Sections III and IV.
Consider a distributed MIMO radar with MTX TXs and
MRX RXs, each equipped with a single omni-directional
antenna and deployed in an unknown environment. An unknown number of stationary point targets3 are present and
the objective is to localize all of them. We assume that the
environment has non-target scatterers too, which can either
block some target(s) to some TX(s) and/or RX(s), and/or give
rise to IPs. All TX and RX locations are assumed to be known.
The number of TRPs, denoted by I, equals MTX MRX . Unless
otherwise mentioned, the convention throughout this work is
that the i-th TRP (i = 1, · · · , I) comprises of the iT -th
TX and iR -th RX, where iT = 1 + (i − 1) mod MTX and
iR = ⌊(i − 1)/MTX ⌋ + 1 (Table I).
For ease of notation, we restrict our attention to the twodimensional (2D) case, where all the TXs, RXs, scatterers
and targets are assumed to lie in the horizontal plane. The
extension to the 3D case is straightforward. Let the TX and
3 The point target assumption simplifies the analysis of the problem, as each
target can give rise to at most one DP. The impact of real (i.e., non-point)
targets can be seen in our experimental results, presented in Section V-D.
3
TRP No.
TX No.
RX No.
1
1
1
2
2
1
3
3
1
4
1
2
5
2
2
6
3
2
7
1
3
8
2
3
9
3
3
TABLE I: TRP indexing notation for the 3 TX, 3 RX case
the RX for the i-th TRP be located at (ci , di ) and (ai , bi ),
respectively. We assume that the TXs use orthogonal signals
so that the RXs can distinguish between signals sent from
different TXs. For each TRP, the RX extracts the channel
impulse response from the measured receive signal; an MPC
is assumed to exist at a particular delay (within the resolution limit of the RX) when the amplitude of the impulse
response at that delay bin exceeds a threshold; alternately, a
maximum likelihood (ML) estimator or other high-resolution
algorithms can be used to extract the amplitudes and delays
of all the MPCs [27], [28]. All MPCs that do not involve a
reflection off a target (e.g., TX→scatterer→RX) are assumed
to be removed by a background cancellation technique. For
stationary or even slow-moving targets, a simple way to
achieve this is to measure the impulse responses for all
the TRPs when no targets are present. This set of template
signals can then be subtracted from the signals obtained when
the target(s) are introduced, which would remove MPCs of
the form TX→scatterer→RX since they appear twice [29].
Other background subtraction techniques for localization and
tracking applications are described in [30], [31]. An MPC
involving more than two reflections is assumed to be too weak
to be detected. Finally, two or more MPCs could have their
delays so close to one another that they can be unresolvable
due to finite bandwidth. Under this model, each extracted MPC
could be one or more of the following:
1. A DP to one or more targets, which occurs when a target
has LoS to both the TX and RX in question.
2. An IP of the first kind, which is of the form
TX→target→scatterer→RX.
3. An IP of the second kind, having the form
TX→scatterer→target→RX.
4. A noise peak.
Each MPC gives rise to a time-of-arrival (ToA) estimate
which, in turn, corresponds to a range estimate. If only
additive white Gaussian noise (AWGN) is present at the RXs,
then each ToA estimate is approximately perturbed by zeromean Gaussian errors whose variance depends on the SNR
via the Cramer-Rao lower bound (CRLB) and the choice
of estimator [32]. For simplicity, it is assumed that all ToA
estimation errors have the same variance σ̂ 2 . The extension to
the general case where the variance is different for each MPC
is straightforward. Thus, for a DP, the true range of the target
from its TRP is corrupted by AWGN of variance σ 2 = c2 σ̂ 2 ,
where c is the speed of light in the environment.
Suppose the i-th TRP has Ni MPCs extracted from its
received signal. Let rij denote the range of the j-th extracted
MPC at the i-th TRP and let ri = [ri1 , ri2 , · · · , riNi ] ∈
RNi ×1 denote the vector of range estimates from the i-th
TRP. Similarly, let r = [r1 , r2 , · · · , rI ] ∈ RN1 N2 ...NI ×1
denote the stacked vector of range estimates from all TRPs.
If rij is a DP corresponding to a target at (xt , yt ), then the
conditional pdf of rij , given (xt , yt ), is Gaussian and denoted
by fDP (rij |xt , yt ) and has the following expression:
(rij − ri (xt , yt ))2
1
exp −
(1)
fDP (rij |xt , yt ) = √
2σ 2
2πσ
p
where, ri (xt , yt ) = (xt − ai )2 + (yt − bi )2
p
+ (xt − ci )2 + (yt − di )2
ri (xt , yt ) denotes the range of a target at (xt , yt ) from
the i-th TRP. Similarly, let fIP,1 (rij |xt , yt , um , vm ) and
fIP,2 (rij |xt , yt , um , vm ) denote the conditional IP pdfs of the
first and second kind, respectively, given a target at (xt , yt )
and a scatterer at (um , vm ). These pdfs are also Gaussian,
fIP,1 (rij |xt , yt , um , vm ) =
(rij − li (xt , yt , um , vm ))2
1
√
exp −
2σ 2
2πσ
(2)
fIP,2 (rij |xt , yt , um , vm ) =
(rij − mi (xt , yt , um , vm ))2
1
√
exp −
2σ 2
2πσ
(3)
p
where li (xt , yt , um , vm ) = (ci − xt )2 + (di − yt )2
p
p
+ (xt − um )2 + (yt − vm )2 + (um − ai )2 + (vm − bi )2
p
mi (xt , yt , um , vm ) = (ci − um )2 + (di − vm )2
p
p
+ (um − xt )2 + (vm − yt )2 + (xt − ai )2 + (yt − bi )2
li (xt , yt , um , vm ) and mi (xt , yt , um , vm ) respectively denote
the path length between the i-th TRP, a target at (xt , yt ) and
a scatterer at (um , vm ) for an IP of the first and second kind.
Finally, the range of a noise peak is modelled as a uniform
random variable in the interval [0, Robs ], where Robs denotes
the maximum observable range in the region of interest.
Let the number of targets and scatterers be denoted by T and
M , respectively. To determine all the unknowns, every MPC
needs to be accounted for. Hence, we define the following
variables,
(
1, if t-th target is NOT blocked to the i-th TRP
kit =
0, else
(4)
gimt
himt
1, if ∃ an IP of the first kind between the i-th
=
TRP, m-th scatterer and t-th target
0, else
(5)
1, if ∃ an IP of the second kind between the i-th
=
TRP, m-th scatterer and t-th target
0, else
(6)
The values of kit , gimt and himt (i ∈ {1, · · · , I}, m ∈
{1, · · · , M }, t ∈ {1, · · · , T }) capture the ground truth regarding the existence of DPs and IPs and depend on the
map of the environment, which is unknown. Therefore, these
quantities need to be estimated from r. To do this, we define
4
the following decision variables to determine if an MPC rij
is a DP, IP or noise peak,
(
1, if rij is a DP to the t-th target
k̃ijt =
(7)
0, else
1, if rij is an IP of the first kind between the
g̃ijmt =
m-th scatterer and t-th target
0, else
(8)
1, if rij is an IP of the second kind between the
h̃ijmt =
m-th scatterer and t-th target
0, else
(9)
(
1, if rij is a noise peak
z̃ij =
(10)
0, else
Since two or more resolvable MPCs cannot be DPs to the
same target or IPs of a particular kind between a given targetscatterer pair, it follows that the estimates of kit , gimt and
himt , denoted by k̂it , ĝimt and ĥimt , respectively, are given
by:
k̂it =
Ni
X
k̃ijt
(11)
g̃ijmt
(12)
h̃ijmt
(13)
j=1
ĝimt =
Ni
X
j=1
ĥimt =
Ni
X
j=1
Before concluding this section, we define the following vectors
which shall be useful when the Bayesian MTL problem is
defined in the next section
Ground truth: k = vec(kit )
(14)
g = vec(gimt )
h = vec(himt )
(15)
(16)
DP/IP/noise peak decisions: k̃ = vec(k̃ijt )
g̃ = vec(g̃ijmt )
(17)
(18)
h̃ = vec(h̃ijmt )
(19)
z̃ = vec(z̃ij )
(20)
Estimates of ground truth: k̂ = vec(k̂it )
(21)
ĝ = vec(ĝimt )
(22)
ĥ = vec(ĥimt )
(23)
III. BAYESIAN MTL
Using the notation from the previous section, the MTL
problem in multipath environments with correlated blocking is
formulated as a Bayesian estimation problem in this section.
We first show that the scatterer locations cannot be determined
uniquely, in general, as they are not point objects. Then, we
show that that the distribution of k in (14) captures correlated
blocking in its entirety and acts as a prior. We also assume
a single error at most between the entries of k̂t and kt in
order to obtain a tractable algorithm for the MTL problem in
Section IV.
Let Θtar = {(xt , yt ) : t = 1, · · · , T } and Θsc =
{(um , vm ) : 1, · · · , M } denote the collection of target and
scatterer locations, respectively, and let p̃dec = [k̃, g̃, h̃, z̃]
denote the vector of decision variables. Using the terminology
defined in Section II, determining the location of all targets and
scatterers can be formulated as a Bayesian estimation problem
in the following manner,
maximize f (r|p̃dec , Θtar , Θsc , k, g, h) × f (Θtar , Θsc )
T,M,Θtar ,Θsc ,
p̃dec ,k,g,h
×P(k̂, ĝ, ĥ|Θtar , Θsc , k, g, h) × P(k, g, h|Θtar , Θsc )
(24)
subject to (11), (12), (13)
X
X
X
z̃ij ≥ Ni , ∀i (25)
(g̃ijmt +h̃ijmt ) +
k̃ijt +
j,t
j
j,m,t
k̃ijt , g̃ijmt , h̃ijmt , k̂it , ĝimt ,ĥimt ∈ {0, 1}, ∀i, j, t, m
(26)
where, the first term in the objective (24) denotes the likelihood function and the remaining three terms denote the prior.
A detailed explanation of all the terms and constraints in (24)(26) is provided below:
(a) The term f (Θtar , Θsc ) denotes the prior joint distribution of the target and scatterer locations. It is reasonable to assume that the target and scatterer locations
are independent of each other. Hence, f (Θtar , Θsc ) =
f (Θtar )f (Θsc ). In addition, f (Θtar ) and f (Θsc ) are both
assumed to be uniform pdfs over the region of interest.
(b) The discrete distribution P(k, g, h|Θtar , Θsc ) represents
the geometry of the environment, such as the blocked DPs
for each TRP, the IPs (if any) between a target-scatterer
pair etc. Let ΘTX = {(ci , di ) : i = 1, · · · , I} and
ΘRX = {(ai , bi ) : i = 1, · · · , I} denote the collection
of TX and RX locations, respectively. ΘTX and ΘRX
are known quantities and for a given set of values for
Θtar and Θsc , the set Θenv = {ΘTX , ΘRX , Θtar , Θsc }
completely describes all the propagation paths in the environment and the values of k, g and h are deterministic
functions of Θenv , denoted by k(det) (Θenv ), g(det) (Θenv )
and h(det) (Θenv ), respectively4. Hence,
P(k, g, h|Θtar , Θsc ) = 1k(det) (Θenv ) (k) × 1g(det) (Θenv ) (g)
× 1h(det) (Θenv ) (h)
(27)
where 1y (x) equals 1 if x = y and 0, otherwise.
(c) The estimates k̂it , ĝimt and ĥimt may differ from their
respective ground truths, kit , gimt and himt due to noise
or IPs. Assuming that k̂it (or ĝimt , ĥimt ) is conditionally
independent of other estimates, given kit (or gimt , himt ),
we get
P(k̂, ĝ, ĥ|Θtar , Θsc , k, g, h)
= P(k̂, ĝ, ĥ|k(det) (Θenv ), g(det) (Θenv ), h(det) (Θenv ))
(28)
4 This
is akin to ray-tracing
5
=
Y
i,t,m
(det)
P(k̂it |kit
(det)
(Θenv )) × P(ĝimt |gimt (Θenv ))
(det)
× P(ĥimt |himt (Θenv ))
(29)
where (28) follows from (27).
(d) p̃dec is a sufficient statistic for estimating
k, g and h. Hence, the likelihood function,
f (r|p̃dec , Θtar , Θsc , k, g, h), equals f (r|p̃dec , Θtar , Θsc ).
Further, f (r|p̃dec , Θtar , Θsc ) decomposes into product
form as the noise terms on each rij are mutually
independent. Thus,
f (r|p̃dec , Θtar , Θsc , k, g, h) = f (r|p̃dec , Θtar , Θsc )
Y
f (rij |p̃dec , Θtar , Θsc )
=
i,j
where, f (rij |p̃dec , Θtar , Θsc ) =
Y
(fDP (rij |xt , yt ))k̃ijt
t,m
×(fIP,1 (rij |xt , yt , um , vm ))g̃ijmt
×(fIP,2 (rij |xt , yt , um , vm ))h̃ijmt
z̃ij
1
×
(30)
Robs
(e) Finally, constraint (25) ensures that the number of DPs,
IPs and noise peaks received at the i-th TRP is at least
Ni , the number of resolvable MPCs extracted at the i-th
TRP.
After taking natural logarithms, (24) can be re-written as
follows to obtain problem P 1, where k, g and h are no longer
unknowns due to (27):
1 X
P1 :
minimize
k̃ijt (rij − ri (xt , yt ))2
T,M,p̃dec ,Θtar ,Θsc σ 2
i,j,t
1 X
g̃ijmt (rij − li (xt , yt , um , vm ))2
+ 2
σ i,j,t,m
1 X
+ 2
h̃ijmt (rij − mi (xt , yt , um , vm )2
σ i,j,t,m
X
X
√
(g̃ijmt + h̃ijmt ) log 2πσ
k̃ijt +
+
i,j,m,t
i,j,t
+
−
−
X
i,j
X
i,m,t
X
i,m,t
z̃ij log Robs −
(det)
X
i,t
(det)
log P(k̂it |kit
(Θenv ))
log P(ĝimt |gimt (Θenv ))
(det)
log P(ĥimt |himt (Θenv ))
need not be satisfied in all circumstances. Hence, P 1 is illposed if the map of the environment is unknown5. To make P 1
tractable, we restrict ourselves to localizing only the targets by
retaining those terms and constraints involving just the DPs in
(24)-(26). This gives rise to the following approximation, P 2,
which is also a Bayesian estimation problem that accounts for
all the DPs
X
√
k̃ijt log 2πσ − log P(k|Θtar )
P 2 : minimize
T,k̃,Θtar
i,j,t
X
1 X
+ 2
log P(k̂it |kit )
k̃ijt (rij − ri (xt , yt ))2 −
σ i,j,t
i,t
(32)
subject to k̃ijt , k̂it ∈ {0, 1}, ∀i, j, t, m
X
k̃ijt = k̂it
(33)
(34)
j
The joint DP blocking distribution P(k|Θtar ) in P 2 is no
longer a discrete-delta function, like (27). Instead, it depends
on the distribution of scatterer locations in the environment.
From (4), kit = 0 if either the TX or the RX of the i-th TRP
does not have LoS to (xt , yt ); hence, kit can be expressed as
a product of two terms in the following manner:
kit = viT ,t × wiR ,t ,
(35)
(
1, if the iT -th TX has LoS to (xt , yt )
where, viT ,t =
0, else
(
1, if the iR -th RX has LoS to (xt , yt )
wiR ,t =
0, else
kit can be interpreted as a Bernoulli random variable when
considering an ensemble of settings in which the scatterers
are placed at random. For vectors kt = [k1t , · · · , kIt ],
vt = [v1,t , · · · , vMTX ,t ] andN
wt = [w1,t ,N
· · · , wMRX ,t ],
denotes the
it can be seen that kt = wt vt , where
Kronecker product. kt is a vector of dependent Bernoulli
random variables (Fig. 1) and shall henceforth be referred to
as the blocking vector at (xt , yt ). Note that k = vec(kit ) =
[k1 , · · · , kT ] and therefore, P(k|Θtar ) = P(k1 ; · · · ; kT ). In
general, two or more blocking vectors may also be dependent
as nearby targets can experience similar blocking. Thus, the
joint distribution P(k1 ; · · · ; kT ) captures correlated blocking
in its entirety. Consequently, target-by-target localization is not
optimal, in general. However, for ease of computation, we
resort to such an approach in this paper, thereby implicitly
assuming independent
Y blocking vectors at distinct locations,
i.e., P(k|Θtar ) ≈
P(kt ). The generalization to joint-target
t
(31)
subject to (11), (12), (13), (25), (26)
Typically, Θsc represents a finite collection of points belonging to distributed non-point objects (e.g., a wall), where
reflection takes place. A minimum of three reflections are
needed at each (um , vm ) for uniquely determining Θsc , which
localization will be described in a future work.
Among the 2I possible values, kt can only take on
MTX
realizable values,
− 1)(2MRX − 1) + 1 physically N
(2
vt (e.g., kt =
which can be expressed in the
form
w
t
N
[1 1 0 1 1 0 1 1 0] = [1 1 1] [1 1 0] for the TRP indexing
notation in Table I). These are referred to as consistent
5 If the map of the environment is known, then P 1 is not ill-posed and the
IPs can be re-cast as virtual DPs, obtained from virtual TXs and RXs [33].
6
blocking vectors while the remaining values are inconsistent
(e.g., kt = [1 1 0 1 1 1 1 0 0]). If kt is inconsistent, then
P(kt ) = 0.
To characterize P(k̂it |kit ), a distinction between two kinds
of estimation errors needs to be made:
a) The DP corresponding to the t-th target at the i-th TRP
may not detected if the noise pushes the range estimate far
away from the true value. As a result, k̂it = 0 when kit =
1. If the noise is independent and identically distributed
(i.i.d) for all TRPs, we may assume that P(k̂it = 0|kit =
1) = ρ01 (∀ i, t), where ρ01 is determined by the SNR
(signal-to-noise ratio) and the ToA estimator.
b) If the DP for the t-th target at the i-th TRP is blocked,
but a noise peak or IP is mistaken for a DP because
it has the right range, then k̂it = 1 and kit = 0.
P(k̂it = 1|kit = 0) depends on the scatterer distribution
and varies according to TX, RX and target locations.
However, in the absence of IP statistics, we make the
simplifying assumption that P(k̂it = 1|kit = 0) = ρ10 ,
for all i, t. The availability of empirical IP statistics would
obviously improve localization performance.
Let k̂t = [k̂1t , · · · , k̂It ] denote the estimated blocking vector
at (xt , yt ). While k̂t can, in principle, take on all 2I values,
a false alarm is less likely if k̂t is a short Hamming distance
away from a consistent vector having high probability. Let K
denote the set of consistent blocking vectors. We restrict k̂t to
be at most a unit Hamming distance from some element in K.
This assumption is reasonable when the number of scatterers
is small and the SNR at all RXs is sufficiently high. Given
k̂t , let Kt ⊆ K denote the set of consistent vectors that are at
most a unit Hamming distance away from k̂t . Then,
X
P(k̂t |kt )P(kt )
P(k̂t ) =
kt ∈Kt
=
X
Y
i
kt ∈Kt
≈
!
P(k̂it |kit ) P(kt )
(36)
X
ρη01 (1 − ρ01 )η11 ρη1010 (1 − ρ10 )η00 P(kt ),
k ∈K 01
t
t
if Kt is non-empty
otherwise
0,
(37)
where, η01 = |{i : k̂it = 0; kit = 1}|
η11 = |{i : k̂it = 1; kit = 1}|
η10 = |{i : k̂it = 1; kit = 0}|
η00 = |{i : k̂it = 0; kit = 0}|
Using (36) and assuming independent blocking vectors at
distinct points (i.e., target-by-target detection), P 2 can be
reduced to the Bayesian MTL problem P 3, given below:
X
1
k̃ijt (rij − ri (xt , yt ))2
P 3 : minimize 2
σ
T,k̃,Θtar
i,j,t
X
X
√
−
k̃ijt log 2πσ −
log P(k̂t )
(38)
i,j,t
t
subject to (33), (34)
A matching qt (r) = {rij ∈ r|k̃ijt = 1} is the set of DPs
corresponding to the t-th target. Given qt (r) and a point
(xt , yt ), the term in square parentheses in (38) determines if
the ellipses corresponding to the MPCs in qt (r) pass through
(xt , yt ) or not. The other term in (38) plays the role of a
prior by determining the probability of the blocking vector,
k̂t , obtained from qt (r), at (xt , yt ). The objective in (38) is
minimized only when both these quantities are small. To solve
P 3, a mechanism for detecting DPs is required. Since the IP
distribution is unknown, none of the conventional tools such
as Bayesian, minimax or Neyman-Pearson hypothesis testing
can be used for this purpose. In the next section, we describe
our DP detection technique and propose a polynomial-time
algorithm to solve P 3.
IV. MTL A LGORITHM
USING
B LOCKING S TATISTICS
In this section, we define a likelihood function for identifying DPs that enable us to obtain the matchings required for
solving P 3 in a tractable manner.
The number of matchings
possible
M scat
for T targets,
terers and I TRPs is ( I3 N 3 + I4 N 4 · · · + II N I )T , where
N = (2M + 1)T is an upper bound on the number of
MPCs extracted at each TRP, ignoring noise peaks. The
computational complexity of a brute-force search over all
possible matchings for solving P 3 is O(N IT ), which is
intractable for a large number of TRPs and/or targets. To
obtain accurate matchings in a tractable manner, we employ
an iterative approach. Consider, without loss of generality, a
(i−1)
matching qt
(r) for the t-th target consisting of MPCs from
(i−1)
(r) is at
the first i − 1 TRPs (3 ≤ i ≤ I). The size of qt
(i−1) (i−1)
most i − 1. Let (x̂t
, ŷt
) denote the estimate of the
(i−1)
(r) (e.g., using the twotarget location obtained from qt
step estimation method [32]). For an MPC riji from the i-th
(i)
(i)
(i−1)
TRP, let qt,temp (r) = qt
(r) ∪ riji and let k̂t denote the
(i−1) (i−1)
, ŷt
), obtained
i-length partial blocking vector at (x̂t
(i)
(i)
from qt,temp (r). If qt,temp (r) consists entirely of DPs from
(xt , yt ), then (i) the ellipses corresponding to its constituent
MPCs should pass close to (xt , yt ), and (ii) the blocking
(i)
vector k̂t should have high probability. This motivates the
definition of a blocking-aware vector likelihood function,
(i)
LB (qt,temp(r)), defined as follows:
!
(i)
LE (qt,temp (r))
(i)
(i)
LB (qt,temp (r)) =
, − log P(k̂t )
(i)
σ(qt,temp (r))
(39)
(i)
(i−1)
where, LE (qt,temp (r)) = riji − ri (x̂t
(i)
(i−1)
, ŷt
)
(40)
(i)
and σ(qt,temp (r)) is the standard deviation of LE (qt,temp (r)).
(i)
If the ellipses corresponding to the MPCs in qt,temp (r)
(i)
pass through the vicinity of (xt , yt ), then LE (qt,temp (r))
should be very small in magnitude. Under this condition,
it can be shown by a Taylor’s series approximation that
(i)
LE (qt,temp (r)) is a zero-mean Gaussian random variable [8].
(i)
(i)
Hence, if |LE (qt,temp (r))/σ(qt,temp (r))| ≤ δ, where δ is an
7
ellipse intersection threshold, then we conclude that riji passes
(i−1) (i−1)
through (x̂t
, ŷt
).
If the above ellipse intersection condition is satisfied, then
(i)
the term − log P(k̂t ), which denotes the blocking likelihood
(i)
(i−1) (i−1)
of qt,temp (r) at (x̂t
, ŷt
), needs to be small as well.
The following cases are of interest:
(i)
(i)
1. If k̂t is consistent and − log P(k̂t ) ≤ µ, where
(i)
µ(> 0) is a blocking threshold, then we define qt (r) =
(i)
qt,temp (r) and compute a refined target location estimate
(i) (i)
(i)
(x̂t , ŷt ) from qt (r).
(i)
(i)
2. If k̂t is inconsistent, then let Kt denote the set of
consistent i-length partial blocking vectors that are at
(i)
most a unit Hamming distance away from k̂t . The
following cases are of interest then:
(i)
(i)
(a) If Kt is empty, then P(k̂t ) = 0 (from (36) and
(37), which hold for partial blocking vectors as well)
(i)
and − log P(k̂t ) = ∞. Hence, we conclude that a
target is not present at the estimated location.
(i)
(i)
(b) If Kt is not empty, then each element of Kt is
a feasible ground truth. In particular, an element in
(i)
Kt whose Hamming weight is lower than that of
(i)
k̂t represents a ground truth where exactly one
(i)
MPC in qt,temp(r) is not a DP. For each such
element, a new matching can be derived by removing
(i)
the corresponding non-DP from qt,temp (r) and evaluated a new blocking likelihood. On the other hand,
(i)
an element of Kt with a higher Hamming weight
(i)
compared to k̂t represents a ground truth where one
(i)
DP is absent from qt,temp(r) due to noise. Unlike
the previous case, no modification of the matching
(i)
is possible and the blocking likelihood of k̂t is
computed according to (36)-(37). In this manner, it
is possible that multiple matchings may exist for a
single potential target location, each corresponding
to a different ground truth. All the matchings whose
blocking likelihood satisfies the threshold µ are retained, since it is premature to determine the most
likely ground truth until all TRPs are considered.
After the I-th TRP has been processed, if multiple
matchings still exist for the t-th target, then the
one that minimizes the objective function in (38)
is declared the true matching and the corresponding
(I) (I)
(x̂t , ŷt ) is the location estimate for the t-th target.
Otherwise, if no MPC from the i-th TRP satisfies the ellipse
(i)
(i)
intersection condition (i.e., |LE (qt,temp (r))/σ(qt,temp (r))| >
(i)
(i−1)
(i) (i)
(r) and (x̂t , ŷt ) =
δ, for all riji ), then qt (r) = qt
(i−1) (i−1)
(i)
(x̂t
, ŷt
). For the resulting k̂t , inconsistencies are
(i)
handled as stated above in point 2. If − log P(k̂t ) > µ,
then we conclude that a target is not present at the estimated
location.
This motivates an algorithmic approach that is divided
into stages, indexed by i. In general, let (z1 , z2 , ..., zI ), a
permutation of (1, 2, · · · , I), be the order in which TRPs are
processed. At the beginning of the i-th stage (3 ≤ i ≤ I),
(i−1)
each qt
(r) has at most i − 1 entries. During the i-th
stage, all the DPs among the MPCs of the zi -th TRP are
(i)
identified to obtain a set of matchings {qt (r)} for each target
t. A matching is consistent (inconsistent) if the corresponding
blocking vector is consistent (inconsistent). By construction,
the only inconsistent matchings are due to missing DPs (see
bullet point 2(b) in previous paragraph). A finite value of µ
(i)
ensures that an inconsistent k̂t is always a unit Hamming
distance away from consistency, due to (37). Since the block(i)
ing likelihood − log P(k̂t ) is non-decreasing in i, a matching
and its corresponding target location can be removed from
consideration if at any stage its blocking likelihood exceeds
the blocking threshold, µ.
Let P (a, b, ja , jb ) denote the points of intersection of the
ellipses corresponding to the ja -th MPC of the a-th TRP
and the jb -th MPC of the b-th TRP. For the initial set of
matchings (i.e., i = 3), P (z1 , z2 , jz1 , jz2 ) is computed for
all jz1 , jz2 (1 ≤ jz1 ≤ Nz1 , 1 ≤ jz2 ≤ Nz2 ). There can
be at most four points in any P (z1 , z2 , jz1 , jz2 ) and each
such point is an ML estimate of the target location for
(2)
the matching qt (r) = {rz1 ,jz1 , rz2 ,jz2 }. Hence, the target
(2) (2)
location estimate (x̂t , ŷt ) need not be unique. Furthermore,
in the i-th stage (i ≥ 4) we also compute P (zu , zi , jzu , jzi )
for all jzu , jzi (∀u < i) to identify previously blocked targets.
In summary, any intersection of two ellipses is a potential
target location to begin with. At each such location, the
likelihood of a target being present is updated depending
on the number of other ellipses passing around its vicinity.
Unlikely target locations, corresponding to matchings whose
likelihood (given by LB (.)) does not satisfy the thresholds δ
and µ, are eliminated at each stage. The number of targets that
remain at the end is the estimate of T . Algorithm 1 lists the
pseudocode of the Bayesian MTL algorithm.
A. Complexity of Bayesian MTL Algorithm
Let T̂ (i) denote the number of targets identified at the end
of stage i. The following relation holds,
T̂ (i) ≤ T̂ (i − 1) +
and, T̂ (3) ≤ N 3
i−1
N 3 , (i = 4, · · · , I) (41)
2
(42)
At the end of i−1-th stage, each target can have at most (i−1)
matchings. Hence, O(iT̂ (i − 1)N ) likelihood computations
are carried out in the i-th stage due to existing targets. The
second term in (41) is an upper bound on the number of
new targets that can be identified in the i − th stage and the
number of likelihood computations due to these is O(i2 N 3 ).
At each stage, the number of potential targets increases at
most polynomially in N and I (41). Hence, the number
of likelihood computations is also polynomial in N and I.
The reduction in complexity occurs because target locations
are determined by ‘grouping’ pair-wise ellipse intersections
that are close together. Since there are only O(I 2 N 2 ) ellipse
intersections to begin with, it is intuitive that the proposed
algorithm terminates in polynomial-time.
8
Algorithm 1 Bayesian MTL algorithm
B. Limitations of Bayesian MTL Algorithm
The Bayesian MTL algorithm assumes complete knowledge
of the distribution of kt at all locations (xt , yt ). This would
have to be obtained either from very detailed theoretical
models or exhaustive measurements, neither of which might
be feasible in practice. A sub-optimal, but more practical,
alternative could involve the use of second-order statistics
of kt . In particular, the Mahalanobis distance, defined as
q
(k̂t − mt )T C−1
t (k̂t − mt ), where mt and Ct respectively
denote the mean vector and covariance matrix of kt and (.)T
and (.)−1 denote the matrix transpose and inverse operations,
respectively, can be compared to a threshold µ2 as the basis for
a blocking likelihood decision. Even in this simplified case,
one still needs the mean blocking vector and the covariance
matrix at each point. In practice, these can be measured at
only at a fixed set of grid points. Hence, the accuracy of the
algorithm would depend on the grid resolution of the measured
data.
V. S IMULATION
AND
E XPERIMENTAL R ESULTS
In this section, we present our simulation and experimental results for the Bayesian MTL algorithm introduced in
Section IV. In Section V-A, the algorithm is validated by
reproducing the results described in prior art for independent
blocking, which is a special instance of P 3. The importance of
considering correlated blocking and the accuracy of the matchings obtained by the Bayesian MTL algorithm are discussed
in Sections V-B and V-C, respectively. Finally, experimental
results which provide insights into the impact of non-point
targets and imperfect background subtraction are presented in
Section V-D.
Unless otherwise mentioned, we use the following settings
for our simulation results: G = [−10m, 10m] × [−10m, 10m]
is the region of interest. Scatterers are modelled as balls of
diameter L; obviously, the blocking correlation increases with
L. The standard deviation of the ranging error, σ is assumed
to be 0.01m. Two or more MPCs that are within a distance
of 2σ apart are considered to be unresolvable; in that case,
the earliest arriving peak is retained and the other peaks are
discarded. For a given δ, ρ01 = ρ10 = 2Q(δ) was assumed,
Z∞
√
2
where Q(x) = e−x /2 / 2πdx. A target is considered to be
x
missed if there is no location estimate lying within a radius
of 3σ from the actual coordinates. Similarly, a false alarm
is declared whenever there is no target within a radius of
3σ from an estimated target location. For a given network
realization, let T̂D and T̂F denote the number of detections
and false alarms, respectively. Then, the detection and false
alarm probabilities, denoted by PD and PF , respectively, are
calculated as follows,
PD = E[T̂D /T ]
(43)
PF = E[T̂F /(T̂D + T̂F )]
(44)
where the expectation is over the ensemble of network realizations
Obtain the TRP processing order (z1 , z2 , · · · , zn ) [8]
t=0
⊲ (Initial set of matchings)
for each jz1 , jz2 do
for each ellipse intersection (x, y) corresponding to
rz1 ,jz1 and rz2 ,jz2 do
if LB ({rz1 ,jz1 , rz2 ,jz2 }) ≤ (δ, µ) then
t=t+1
(2)
qt (r) = {rz1 ,jz1 , rz2 ,jz2 }
(2) (2)
(x̂t , ŷt ) = (x, y)
end if
end for
end for
T̂ (2) = t
⊲ T̂ (i) denotes the number of estimates at the
end of the i-th stage
for i = 3 to I do
for t = 1 to T̂ (i − 1) do
⊲ (Updating existing
matchings)
(i−1)
(r) ∪ rzi ,jzi ) ≤ δ
if ∃ any rzi ,jzi such that LE (qt
then
(i)
(i−1)
qt,temp (r) = qt
(r) ∪ rzi ,jzi
else
(i)
(i−1)
qt,temp (r) = qt
(r)
end if
(i)
(i)
Derive Kt from qt,temp (r)
(i)
(i)
for each k̂t ∈ Kt do
(i)
if − log k̂t ≤ µ then
(i)
(i)
Derive qt (r) from qt,temp (r) according to
(i)
k̂t
end if
end for
end for
Update T̂ (i) and set t = T̂ (i)
for each jzi , jzu (u = 1, · · · , i − 1) do ⊲ (New targets,
previously unidentified due to blocking)
for each ellipse intersection (x, y) corresponding to
rzi ,jzi and rzu ,jzu do
if LB ({rzi ,jzi , rzu ,jzu }) ≤ (δ, µ) then
t= t+1
(i)
qt (r) = {rzi ,jzi , rzu ,jzu }
(i) (i)
(x̂t , ŷt ) = (x, y)
end if
end for
end for
T̂ (i) = t
end for
A. Comparison to Prior Art
In [8], the probability of any DP being blocked was assumed
to be constant throughout G and independent of other blocking
probabilities. Target detection was achieved if there existed
a matching of size at least I − Φ, where Φ denotes the
maximum number of undetected DPs permitted, regardless of
consistency. We now proceed to demonstrate how this criterion
is a special case of the Bayesian MTL Algorithm, obtained
9
I.c.b. assumption is a special case of the
Bayesian MTL algotrithm
10
1
δ = 1, 2, 3
8
δ=3
0.9
δ=2
6
0.8
4
0.7
TXs
RXs
Targets
0
Φ=1
µ (Φ = 1)
Φ=3
µ (Φ = 3)
Φ=6
µ ( Φ = 6)
δ=1
0.6
PD
y (m)
2
δ=3
0.5 δ = 2
-2
0.4
-4
0.3
-6
0.2
-8
0.1
δ=1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
PF
-10
-10
-8
-6
-4
-2
0
2
4
6
8
x (m)
Fig. 2: A network consisting of 3 TXs at (−8m, 7m),
(−7m, 8m) and (7m, 7m), 3 RXs at (−7m, 7m), (8m, 7m)
and (7m, 8m) (i.e., I = 9 TRPs) and 2 targets at (0m, 0m)
and (0m, 5m). The TX and RX locations are such that the
LoS blocking probabilities are independent only if L is very
small. For L = 0.001m, the independent blocking assumption
holds.
by assuming independent blocking with constant blocking
probabilities (henceforth referred to as the i.c.b assumption).
Let plos denote the probability that LoS exists between any
two points in G. The probability that a DP is blocked is then
given by pb = 1 − p2los . Taking into account both blockage
and missed detection by noise, the probability that a DP is
undetected (denoted by pdp ) is given by
pdp = (1 − pb ).2Q(δ) + pb
(45)
The blocking likelihood of a matching with Φ undetected
DPs equals − log((1 − pdp )I−Φ pΦ
dp ). If pdp < 1/2, then the
blocking likelihood monotonically increases with Φ. Hence,
for a given Φ, the corresponding blocking threshold, µ(Φ),
can be set as follows
µ(Φ) = − log((1 − pdp )I−Φ pΦ
dp )
(46)
which ensures that the detected targets have matchings of size
at least I − Φ.
To validate the Bayesian MTL algorithm, we compared it
with the prior art proposed in [8], under the i.c.b assumption.
The comparison was done on the network shown in Fig. 2.
To model the i.c.b condition, the values for L and plos were
chosen to be 0.001m and 0.9, respectively. With probability
1 − plos , a scatterer was placed independently and uniformly
along each line segment between a node (TX/RX) and a target.
The two algorithms were evaluated over 100 realizations for
three values of δ(= 1, 2 and 3) and Φ (= 1, 3 and 6).
For each value of Φ, the threshold µ(Φ) for the Bayesian
10
Fig. 3: Region of Convergence (ROC) curves plotting PD
versus PF for the i.c.b condition. The prior art in [8] is a
special case of the Bayesian MTL algorithm
MTL algorithm was chosen according to (46). The region
of convergence (ROC) curves, plotting PD versus PF , for
both algorithms are shown in Fig. 3. As expected, they yield
identical missed-detection and false alarm rates.
Increasing δ loosens the compactness constraint on the
ellipse intersections around a potential target location, while
increasing µ relaxes the constraint on the probability of a
blocking vector/matching. Hence, both PF and PD are nondecreasing in δ and µ, as seen in Fig. 3. In the special case
where only three ellipse intersections are sufficient to declare
the presence of a target (Φ = 6), the false alarm rates are very
high. This is in agreement with the results reported in [34].
B. Effect of Correlated Blocking
To highlight the effect of correlated blocking, the value
of L was increased to 5m and the scatterer centers were
distributed according to a homogeneous Poisson point process
(PPP) of intensity λ = 0.0075m−2, which amounts to three
scatterers in G, on average, per realization. The blocking
distribution for the PPP scatterer model is derived in Appendix
A. A total of 100 network realizations were considered, with
MTX = MRX = 3 and T = 2, which corresponds to
N = T (2M + 1) = 2(2 × 3 + 1) = 14 MPCs per TRP,
on average. Let Ssc ⊆ G denote the region occupied by
scatterers in a given realization. The TX, RX and target
locations were uniformly and independently distributed over
the region G \ Ssc , where ‘\’ denotes the set difference
operator. Under the i.c.b assumption for the above settings,
plos = exp(−λLdavg ), where davg = 10.1133m is the average
distance between a target and a node. Hence, from (45),
pdp = 0.5329 > 1/2 for δ = 3. The distribution of the
average number of DPs at a point is tabulated in Table II for
both the true blocking distribution and the i.c.b assumption.
As per the true blocking distribution, a target has LoS to all
1
10
ROC curve for L = 5m, δ = 3
1
0.9
0.8
0.8
0.7
0.7
P(missed detection)
0.9
PD
0.6
Correlated blocking
i.c.b assumption
0.5
L = 5m, δ = 3
1
0.4
0.6
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
Bayesian MTL algorithm
Genie-aided method
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
PF
Fig. 4: Ignoring correlated blocking can result in false alarms
being more likely to occur than detections.
TXs and RXs (i.e., 9 DPs) over 66% of the time and the
probability that a target has only 3 DPs is a little over 1%.
As a result, a matching of size 3 is more likely to be a false
alarm. However, since pdp > 1/2, a matching of size 3 is
more probable than a matching of size 9 (which occurs with
less than 1% probability) under the i.c.b assumption. As a
result, false alarms are identified first, followed by detections,
as the value of µ increases. This is reflected in the ROC curves
plotted in Fig. 4, where the i.c.b assumption gives rise to very
high false alarm rates.
C. Comparison with genie-aided method
In many radar applications, a missed-detection is more
costly than a false alarm. As a benchmark, the misseddetection probability of the Bayesian MTL algorithm is compared with that of a genie-aided method, which involves
running the Bayesian MTL Algorithm on the true target
matchings, in Fig. 5. It can be seen that the proposed algorithm
performs as well as the genie-aided method.
D. Experimental Results
We now present some experimental results that further
validate the performance of the Bayesian MTL algorithm.
We chose a portion of UltraLab at USC, a cluttered indoor
environment, for our measurements. The floor was paved with
square tiles of side 0.61m which provided a natural Cartesian
coordinate system, as shown in Fig. 6a. The measurement
setup is shown in Fig. 6b. For the i-th TRP, the frequency
response of the ultrawideband (UWB) channel over 6-8 GHz
was measured twice at 1601 frequency points - once without
the targets (i.e., the background measurement, denoted by
1
0
1
2
3
4
5
6
7
8
µ
Fig. 5: Comparison with genie-aided method.
Hiback (f )) and then with the targets present (denoted by
Hitar (f )) - using a pair of horn antennas with beamwidth
60o , connected to a vector network analyzer (VNA). This
corresponds to σ = 0.15m. Horn antennas were preferred over
omnidirectional antennas to restrict the background clutter to
a narrow sector. The antennas were maintained at the same
height from the ground in order to create a 2D localization
scenario, and were oriented to face the targets. Two identical,
foil-wrapped cylindrical poles were chosen as the targets.
Although the height of the cylinders exceeded that of the TX
and RX antennas, the portion of the cylinder that was in the
plane of the antennas was wrapped in foil to maintain the 2D
nature of the problem.
Let hi (t) denote the channel impulse response for the
i-th TRP due to the targets alone (i.e., after background
subtraction). Then, hi (t) is given by the following expression,
hi (t) = IFFT(Hitar (f ) − Hiback (f ))
(47)
The noise floor corresponding to the i-th TRP was determined
by computing the average power in the last 100 samples of
hi (t). These delay bins correspond to a signal run length
in excess of 200m, which is well in excess of the ranges
encountered in our measurement scenario (less than 10m).
Hence, it is reasonable to assume that the energy in these
delay bins is due to thermal noise alone. After determining
the noise power, MPCs were extracted from hi (t) whenever
the SNR was greater than 10dB. A distributed virtual MIMO
radar was implemented by moving the TX and RX antennas
to different locations, as shown in Fig. 7. Six TRPs were
considered, which are indexed in Table III. LoS was present
between all TXs, RXs and targets (i.e., P(kt ) = 1, t ∈ {1, 2}).
The estimated target locations are plotted in Fig. 8, from
which the following inferences can be drawn:
11
Blocking Distribution
True
i.c.b assumption
<3
0.0700
0.3367
3
0.0150
0.1961
4
0.0750
0.2578
5
0
0.2259
6
0.1750
0.1320
7
0
0.0496
8
0
0.0109
9
0.6650
0.0011
TABLE II: Distribution of the average number of DPs at a point for L = 5m and λ = 0.0075m−2.
Layout of nodes in cluttered environment
TX 5:
(0, 3.66)
3.5
3
RX 3:
(1.22, 3.05)
TX 4:
(0, 3.05)
Target 2:
(0.915, 2.44)
y (m)
2.5
RX 2:
(1.22, 1.83)
2
TX 3:
(0, 1.83)
1.5
Target 1:
(0.61, 1.22)
1
0.5
TX 2:
(1.22, 0.61)
TX 1:
(0, 0.61)
RX 1:
(1.22, 0)
0
0
0.2
0.4
(a) The cluttered indoor measurement environment.
0.6
0.8
1
1.2
x (m)
Fig. 7: Layout of TXs, RXs and targets in the cluttered
environment of Fig. 6a.
δ =2, µ =1
3.5
(b) Measurement setup using VNA
Actual location
Estimated location
3
Fig. 6: The experimental setup.
1
1
1
2
2
2
3
3
2
4
3
3
5
5
3
6
4
3
TABLE III: Look-up table mapping the TRP index with the
corresponding TX and RX IDs, corresponding to Figs. 6a and
7.
2.5
y (m)
TRP
TX
RX
2
1.5
Steel pillar
1
(i) Setting µ = 1 ensures that only those points at which
all six ellipses (corresponding to the six TRPs) intersect
are detected as target locations. It can be seen that both
targets are localized.
(ii) The Bayesian MTL algorithm was formulated under the
assumption of point targets. Since the targets are not
point objects, multiple DPs are possible, in general, for
each target. As a result, we obtain a cluster of location
estimates for each target location.
(iii) The DPs to the targets are at least 10 dB above the noise
0.5
0
0
0.2
0.4
0.6
0.8
1
x (m)
Fig. 8: Position estimates for the targets obtained from the
Bayesian MTL algorithm in the Measurement Scenario 2.
1.2
12
floor post background cancellation, even in a cluttered
environment, which can be attributed to the targets being
strong reflectors and having a sufficiently large radar
cross-section, due to the foil wrapping.
(iv) The steel pillar to the left of Target 1, which is a part
of the clutter, was still ‘localized’ in spite of background
subtraction. In terms of range, Target 1 and the pillar are
closely separated for all the TRPs. Hence, some of the
energy from the DP to Target 1 spills over into the delay
bin corresponding to the pillar location. As a result, the
pillar cannot be perfectly canceled out during background
subtraction. The residual energy manifests itself as a DP
to the pillar, leading to its localization. An implication of
this is that the background in the immediate vicinity of
a target cannot be subtracted completely.
VI. S UMMARY
AND
C ONCLUSIONS
In this paper, we considered the impact of environmentinduced correlated blocking on localization performance. We
first provided a theoretical framework for MTL using a distributed MIMO radar by formulating the general problem of
localizing all the targets and scatterers in an unknown environment as a Bayesian estimation problem. We then proceeded to
derive a more tractable approximation, known as the Bayesian
MTL problem, where the objective was to localize only the
targets, but not the scatterers. We then proposed a polynomialtime approximation algorithm - at the heart of which was a
blocking-aware vector likelihood function that took correlated
blocking into account - to solve this problem. The algorithm
relies on two thresholds, δ and µ, to detect targets and works
with either theoretical or empirical blocking statistics that may
be obtained via measurements or simulations. Our simulations
showed that ignoring correlated blocking can be lead to very
poor detection performance, with false alarms being more
likely to occur than detections, and our experiments yielded
encouraging results, even in the presence of non-idealities such
as improper background subtraction and non-point targets.
VII. ACKNOWLEDGMENTS
The authors would like to thank C. Umit Bas, O. Sangodoyin and R. Wang for their assistance in carrying out the
measurements.
A PPENDIX A
B LOCKING MODEL
Let the scatterers be represented by balls of diameter L,
whose centers are distributed according to a homogeneous PPP
with intensity λ. For LoS to exist between two points separated
by a distance d, no scatterer center should lie within a rectangle
of sides L and d (Fig. 9). Therefore, the LoS probability is
exp(−λLd).
N
Consider a consistent blocking vector kt = wt vt at
(xt , yt ). The set of nodes that are blocked/unblocked at (xt , yt )
is determined by vt and wt . For each unblocked node, there
exists a rectangle which cannot contain any scatterer center.
The LoS polygon, Slos , is the union of such rectangles (shaded
grey in Fig. 10). In contrast, for each blocked node n, there
Fig. 9: LoS is obstructed if there exists at least one scatterer
center within a distance of L/2 from the LoS path
13
[12] H. Yan, J. Li, and G. Liao, “Multitarget identification and localization
using bistatic MIMO radar systems,” EURASIP J. Adv. Signal
Process, vol. 2008, pp. 48:1–48:8, Jan. 2008. [Online]. Available:
http://dx.doi.org/10.1155/2008/283483
[13] F. Liu and J. Wang, “An effective virtual ESPRIT algorithm for multitarget localization in bistatic MIMO radar system,” in Computer Design
and Applications (ICCDA), 2010 International Conference on, vol. 4,
June 2010, pp. V4–412–V4–415.
[14] A. Gorji, R. Tharmarasa, and T. Kirubarajan, “Tracking multiple unresolved targets using MIMO radars,” in 2010 IEEE Aerospace Conference, march 2010, pp. 1 –14.
[15] C. Duofang, C. Baixiao, and Q. Guodong, “Angle estimation using
ESPRIT in MIMO radar,” Electronics Letters, vol. 44, no. 12, pp. 770–
771, 2008.
[16] C. Jinli, G. Hong, and S. Weimin, “Angle estimation using ESPRIT
without pairing in MIMO radar,” Electronics Letters, vol. 44, no. 24,
pp. 1422–1423, Nov. 2008.
[17] M. Jin, G. Liao, and J. Li, “Joint DoD and DoA estimation for bistatic
MIMO radar,” Signal Processing, vol. 89, no. 2, pp. 244–251, 2009.
[18] J. L. Chen, H. Gu, and W. M. Su, “A new method for joint DoD and
DoA estimation in bistatic MIMO radar,” Signal Processing, vol. 90,
no. 2, pp. 714–718, 2010.
N
[19] H. L. Miao, M. Juntti, and K. Yu, “2-D unitary-ESPRIT based joint
Fig. 10: kt = [0, 0, 0, 1, 0, 0, 1, 0, 0] = [0, 1, 1] [1, 0, 0]
AoA and AoD estimation for MIMO systems,” in IEEE PIMRC, 2006.
[20] A. A. Gorji, R. Tharmarasa, W. Blair, and T. Kirubarajan, “Multiple unresolved target localization and tracking using colocated MIMO radars,”
exists an NLoS polygon Sn - the portion of its rectangle not
IEEE Transactions on Aerospace and Electronic Systems, vol. 48, no. 3,
pp. 2498–2517, 2012.
contained in Slos - which must contain at least one scatterer
[21] W. Xia, Z. He, and Y. Liao, “Subspace-based method for multiple-target
center. Let Nbl denote the number of blocked nodes. Then,
localization using MIMO radars,” in IEEE International Symposium on
Signal Processing and Information Technology, 2007, pp. 715–720.
Nbl
Y
Xu and J. Li, “Iterative generalized-likelihood ratio test for MIMO
P(kt ) ≥ exp(−λAr(Slos ))
(1 − exp(−λAr(Sn ))) (48) [22] L.
radar,” IEEE Transactions on Signal Processing, vol. 55, no. 6, pp. 2375–
n=1
2385, 2007.
He, R. Blum, and A. Haimovich, “Noncoherent MIMO radar for
where Ar(.) denotes the area operator, acting on sets in R2 . [23] Q.
location and velocity estimation: More antennas means better perforThe expression in (48) is a lower bound since it ignores
mance,” Signal Processing, IEEE Transactions on, vol. 58, no. 7, pp.
3661–3680, July 2010.
overlapping NLoS polygons which may share scatterer centers
Niu, R. Blum, P. Varshney, and A. Drozd, “Target localization and
(e.g., TX 2 and TX 3 in Fig. 10). The bound is met with [24] R.
tracking in noncoherent multiple-input multiple-output radar systems,”
equality when none of the NLoS polygons overlap.
Aerospace and Electronic Systems, IEEE Transactions on, vol. 48, no. 2,
pp. 1466–1489, APRIL 2012.
[25] Y. Ai, W. Yi, G. Cui, and L. Kong, “Multi-target localization for
R EFERENCES
noncoherent MIMO radar with widely separated antennas,” in Radar
[1] S. Steineger, M. Neun, A. Edwardes, and B. Lenz, “Foundations
Conference, 2014 IEEE, May 2014, pp. 1267–1272.
of
location
based
services,”
2006.
[Online].
Available:
[26] C. K. Kim and J. Y. Lee, “ToA-based multi-target localization and
http://www.e-cartouche.ch/content reg/cartouche/LBSbasics/en/text/LBSbasics.pdfrespiration detection using UWB radars,” EURASIP Journal on Wireless
[2] “SELECT: Smart and efficient location, identification and cooperation
Communications and Networking, vol. 2014, no. 1, 2014.
techniques.” [Online]. Available: http://www.selectwireless.eu/home.asp
[27] J. A. Hogbom, “Aperture synthesis with a non-regular distribution
[3] K. Witrisal et al., “High-accuracy localization for assisted living: 5G
of interferometer baselines,” Astronomy and Astrophysics Supplement,
systems will turn multipath channels from foe to friend,” IEEE Signal
vol. 15, pp. 417–426, June 1974.
Processing Magazine, vol. 33, no. 2, pp. 59–70, March 2016.
[28] A. Richter, “Estimation of Radio Channel Parameters: Models and Algorithms,” Ph. D. dissertation, Technischen Universität Ilmenau, Germany,
[4] E. Karapistoli, F. N. Pavlidou, I. Gragopoulos, and I. Tsetsinas, “An
May 2005, ISBN 3-938843-02-0.
overview of the IEEE 802.15.4a standard,” IEEE Commun. Mag.,
[29] J. Salmi and A. F. Molisch, “Propagation Parameter Estimation, Modelvol. 48, no. 1, pp. 47–53, January 2010.
ing and Measurements for Ultrawideband MIMO Radar,” IEEE Trans[5] A. Haimovich, R. Blum, and L. Cimini, “MIMO radar with widely
actions on Antennas and Propagation, vol. 59, no. 11, pp. 4257–4267,
separated antennas,” IEEE Signal Process. Mag., vol. 25, no. 1, pp.
nov. 2011.
116–129, 2008.
[30] X. Chen, H. Leung, and M. Tian, “Multitarget detection and tracking
[6] Y. Bar-Shalom and X. Li, Multitarget-multisensor Tracking: Principles
for through-the-wall radars,” IEEE Transactions on Aerospace and
and Techniques. Storrs, CT, USA: Yaakov Bar-Shalom, 1995.
Electronic Systems, vol. 50, no. 2, pp. 1403–1415, April 2014.
[7] R. Mahler, Statistical Multisource-Multitarget Information Fusion. Nor[31] R. Zetik, M. Eschrich, S. Jovanoska, and R. S. Thoma, “Looking behind
wood, MA, USA: Artech House, 2007.
a corner using multipath-exploiting UWB radar,” IEEE Transactions on
[8] J. Shen and A. F. Molisch, “Estimating multiple target locations in multiAerospace and Electronic Systems, vol. 51, no. 3, pp. 1403–1415, July
path environments,” IEEE Trans. Wireless Commun., vol. 13, no. 8, pp.
2015.
4547–4559, Aug 2014.
[32] J. Shen, A. F. Molisch, and J. Salmi, “Accurate Passive Location
[9] H. Godrich, A. Haimovich, and R. Blum, “Cramer-Rao bound on target
Estimation Using TOA Measurements,” IEEE Transactions on Wireless
localization estimation in MIMO radar systems,” in Information Sciences
Communications, vol. 11, no. 6, pp. 2182–2192, June 2012.
and Systems, 2008. CISS 2008. 42nd Annual Conference on, March
[33] P. Setlur, G. Smith, F. Ahmad, and M. Amin, “Target Localization with a
2008, pp. 134–139.
single sensor via Multipath Exploitation,” IEEE Trans. Aerosp. Electron.
[10] Z. Zheng, J. Zhang, and Y. Wu, “Multi-target localization for bistatic
Syst., vol. 48, no. 3, pp. 1996–2014, 2012.
MIMO radar in the presence of unknown mutual coupling,” Systems
[34] S. Aditya, A. F. Molisch, and H. Behairy, “Bayesian multi-target
Engineering and Electronics, Journal of, vol. 23, no. 5, pp. 708–714,
localization using blocking statistics in multipath environments,” in the
Oct 2012.
IEEE ICC 2015 Workshop on Advances in Network Localization and
[11] J. Li, H. Li, L. Long, G. Liao, and H. Griffiths, “Multiple target
Navigation (ANLN), June 2015.
three-dimensional coordinate estimation for bistatic MIMO radar with
uniform linear receive array,” EURASIP Journal on Advances in Signal
Processing, vol. 2013, no. 1, pp. 1–11, 2013.
| 7cs.IT
|
arXiv:cs/0408045v1 [cs.PL] 19 Aug 2004
On computing the fixpoint of a set of boolean equations
Viktor Kuncak
K. Rustan M. Leino
MIT
[email protected]
Microsoft Research
[email protected]
30 December 2003
Technical Report
MSR-TR-2003-08
This paper presents a method for computing a least fixpoint of a system of equations over booleans. The resulting computation can be
significantly shorter than the result of iteratively evaluating the entire
system until a fixpoint is reached.
Microsoft Research
Microsoft Corporation
One Microsoft Way
Redmond, WA 98052
http://www.research.microsoft.com
KRML 118 / VK0099 / MSR-TR-2003-08, p.1
0 Introduction
Many problems in computer science, in particular those arising in the context of
program analysis, involve the computation of a least (or, dually, greatest) fixpoint
of a system of equations. In this paper, we consider a way to compute a least
fixpoint when the equations involved are over the booleans. In some important
cases, the resulting computation can be significantly shorter than the computation
that iteratively evaluates the entire system until a fixpoint is reached.
Let us begin with an overview of our result. We restrict our attention to a
finite lattice. A finite lattice is a complete lattice and has no infinite ascending
chains, and any monotonic function on such a lattice is also continuous. Hence,
the Kleene Fixpoint Theorem [2] states that the least fixpoint of any monotonic
function F is the lattice join of the sequence of elements
F 0 (⊥), F 1 (⊥), F 2 (⊥), . . .
where exponentiation denotes successive function applications and ⊥ denotes the
bottom element of the lattice. Because this sequence is ascending and because the
lattice is finite, there exists a natural number K such that
F K (⊥)
is the least fixpoint of F . We call the least such K the fixpoint depth of F .
If we are able to evaluate function F and if we are able to determine whether
two given lattice elements are equal, then we can compute the least fixpoint of
F : starting from the value ⊥ , repeatedly apply F until the application of F
leaves the value unchanged. The existence of a fixpoint depth guarantees that
this process terminates. In this paper, we consider the problem of computing an
expression for the least fixpoint, without computing the value of the expression.
By first computing a small expression for the least fixpoint, we can relegate the
computation of the value of the expression to an external tool such as a SAT solver
[3]. In the sequel we therefore do not assume that we are able to compute the value
of an expression into a particular lattice element.
The fixpoint depth of a function F on a lattice is bounded by the height of the
lattice. Therefore, for the 2-element lattice B of the booleans (which has height
1), the least fixpoint of F is given by F (⊥) , and for the 2n -element lattice Bn
that is the Cartesian product space of n booleans (which has height n ), the least
fixpoint of F is given by F n (⊥) .
KRML 118 / VK0099 / MSR-TR-2003-08, p.2
Any function F : Bn → Bn can be represented isomorphically by n functions
fi : Bn → B . We write
F = (f1 , . . . , fn )
where the tuple of functions is itself defined to be a function, as follows, for any
n -tuple X of booleans:
(f1 , . . . , fn )(X ) = (f1 (X ), . . . , fn (X ))
For example, let n = 3 and let F = (f , g, h) . Then, the least fixpoint of F
equals F 3 (⊥, ⊥, ⊥) , as we have argued above. In terms of the functions f , g, h ,
this expands to:
(f (f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
g(f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
h(f (f (⊥, ⊥, ⊥),
g(f (⊥, ⊥, ⊥),
h(f (⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
g(⊥, ⊥, ⊥),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥)),
h(⊥, ⊥, ⊥))))
We refer to this closed form of the fixpoint as the Expanded Closed Form. A
different way to write down the Expanded Closed Form, which shares common
subexpressions, is:
let a1 = ⊥,
a2 = ⊥,
a3 = ⊥
let b1 = f (a1 , a2 , a3 ), b2 = g(a1 , a2 , a3 ), b3 = h(a1 , a2 , a3 )
let c1 = f (b1 , b2 , b3 ), c2 = g(b1 , b2 , b3 ), c3 = h(b1 , b2 , b3 )
let d1 = f (c1 , c2 , c3 ), d2 = g(c1 , c2 , c3 ), d3 = h(c1 , c2 , c3 )
(d1 , d2 , d3 )
in
in
in
in
This representation is cubic in n , which means that computing it may take time
and space that is cubic in n .∗
If we allow ourselves to write functions of n arguments as functions over n -tuples, then we
can obtain a quadratic representation. For example, with n = 3 , we have let a1 = . . . , a2 =
. . . , a3 = . . . in let a = (a1 , a2 , a3 ) in let b1 = f (a), b2 = g(a), b3 = h(a) in let b =
(b1 , b2 , b3 ) in . . . .
∗
KRML 118 / VK0099 / MSR-TR-2003-08, p.3
Let us consider another closed form, which we call the Pruned Closed Form.
In the Pruned Closed Form, an application of a function fi is replaced by ⊥ if
it occurs in another application of the same function fi . For the example above,
where n = 3 , the Pruned Closed Form is:
(f (⊥,
g(⊥,
h(⊥,
g(f (⊥,
⊥,
h(⊥,
h(f (⊥,
g(⊥,
⊥
⊥,
h(⊥, ⊥, ⊥)),
g(⊥, ⊥, ⊥), ⊥
)),
⊥,
h(⊥, ⊥, ⊥)),
⊥,
⊥
g(⊥, ⊥, ⊥), ⊥
⊥,
⊥
)),
),
),
))
If we do not have any interpretation for the functions fi —in other words, if
each fi is just a symbolic name for an uninterpreted function—then the cubicsized Expanded Closed Form may be a reasonably small closed-form representation of the fixpoint. The Pruned Closed Form is generally much larger than cubic
in n : for every subset S of f2 , . . . , fn , function f1 appears expanded in a context where the set of enclosing functions is S . (A smaller Pruned Closed Form
can be obtained by taking advantage of common subexpressions.) However, there
are cases where the Pruned Closed Form can be significantly smaller than the Expanded Closed Form, for example when the fixpoint computation is dominated
by the computation of local fixpoints, meaning fixpoints that involve only a small
number of the functions. An important situation in program analysis where this
case applies is when each function represents a control point in a given program,
a function is defined in terms of the functions corresponding to the successor (or
predecessor) control points, and the given program contains many local loops.
For example, suppose
f (x , y, z ) = f(x , y)
g(x , y, z ) = g(x )
h(x , y, z ) = h(y, z )
for some functions f , g , and h . Then the Expanded Closed Form is
let a1 = ⊥,
a2 = ⊥,
a3 = ⊥
let b1 = f(a1 , a2 ), b2 = g(a1 ), b3 = h(a2 , a3 )
let c1 = f(b1 , b2 ), c2 = g(b1 ), c3 = h(b2 , b3 )
let d1 = f(c1 , c2 ), d2 = g(c1 ), d3 = h(c2 , c3 )
(d1 , d2 , d3 )
in
in
in
in
KRML 118 / VK0099 / MSR-TR-2003-08, p.4
In contrast, the Pruned Closed Form yields the much shorter expression
(f( f(⊥, g(⊥)),
g(⊥)),
g(f(⊥, ⊥)),
h(g(f(⊥, ⊥)),
⊥))
More generally, for an even n , suppose fi (x1 , . . . , xn ) is fi (xi , xi+1 ) when i is
odd and fi (xi−1 , xi ) when i is even. Then the Expanded Closed Form is still
cubic, whereas the Pruned Closed Form is the linear-sized expression
( f1 (⊥, f2 (⊥, ⊥)), f2 (f1 (⊥, ⊥), ⊥),
...
fi (⊥, fi+1 (⊥, ⊥)), fi+1 (fi (⊥, ⊥), ⊥),
...
fn−1 (⊥, fn (⊥, ⊥)), fn (fn−1 (⊥, ⊥), ⊥)
)
In the rest of this paper, we define the Pruned Closed Form more precisely and
prove that it yields the same value as the Expanded Closed Form.
1 Using the Bekić-Leszczyłowski Theorem
In this section, we sketch how to obtain the Pruned Closed Form by applications
of the Bekić-Leszczyłowski Theorem [1, 4].
We write
( ↓ x • R(x ) )
for the lattice meet of all values for x that satisfy the predicate R(x ) . For any
monotonic function F , we then write
( ↓ x • x = F (x ) )
(0)
to denote the least fixpoint of F , because the Tarski Fixpoint Theorem [5] says
that the meet of all fixpoints is itself a fixpoint. Using for a function F : Bn →
Bn the isomorphic representation of n functions fi : Bn → B , we can write (0)
equivalently as:
( ↓ x1 , . . . , xn •
x1 = f1 (x1 , . . . , xn )
..
.
xn = fn (x1 , . . . , xn ) )
∧
∧
KRML 118 / VK0099 / MSR-TR-2003-08, p.5
We can now state the Bekić-Leszczyłowski Theorem [1, 4], for any monotonic
functions F and G (possibly over different lattices):
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
=
( ↓ a, b • a = F (a, b) ∧ b = ( ↓ b • b = G(a, b) ))
Note that each side of the equality expresses a fixpoint in the lattice Bn if F and
G are functions of types Bp × Bq → Bp and Bp × Bq → Bq , respectively, for p
and q such that p + q = n .
A consequence of the Bekić-Leszczyłowski Theorem and the Kleene Fixpoint
Theorem for a known fixpoint depth is the following lemma:
Lemma 0 For any lattice domain A and monotonic functions F : A × B → A
and G: A × B → B ,
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
=
( ↓ a, b • a = F (a, b) ∧ b = G(a, ⊥) )
Proof.
=
=
( ↓ a, b • a = F (a, b) ∧ b = G(a, b) )
{ Bekić-Leszczyłowski Theorem }
( ↓ a, b • a = F (a, b) ∧ b = ( ↓ b • b = G(a, b) ))
{ ( λ b • G(a, b) ) is a function on B , and
therefore its fixpoint depth is at most 1, and
therefore ( ↓ b • b = G(a, b) ) = G(a, ⊥) }
( ↓ a, b • a = F (a, b) ∧ b = G(a, ⊥) )
Using Lemma 0, we now show that the Pruned Closed Form is indeed the least
fixpoint in B2 . For any monotonic boolean functions f and g :
=
=
=
( ↓ a, b • a = f (a, b) ∧ b = g(a, b) )
{ Lemma 0 with F , G := f , g }
( ↓ a, b • a = f (a, b) ∧ b = g(a, ⊥) )
{ substitute equals for equals }
( ↓ a, b • a = f (a, g(a, ⊥)) ∧ b = g(a, ⊥) )
{ Lemma 0 with
G, F := ( λ a, b • f (a, g(a, ⊥)) ), ( λ a, b • g(a, ⊥) ) }
( ↓ a, b • a = f (⊥, g(⊥, ⊥)) ∧ b = g(a, ⊥) )
(1)
KRML 118 / VK0099 / MSR-TR-2003-08, p.6
This calculation shows that an expression for the least solution of a in equation
(1) is
f (⊥, g(⊥, ⊥))
By a symmetric argument, an expression for the least solution of b in equation
(1) is
g(f (⊥, ⊥), ⊥)
That is, an expression for (1) is
( f (⊥, g(⊥, ⊥)), g(f (⊥, ⊥), ⊥) )
which is the Pruned Closed Form.
Using the result for B2 , we can show that the Pruned Closed Form is also the
least fixpoint in B3 . For any monotonic boolean functions f , g , and h :
=
=
=
=
=
=
=
=
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, c) ∧ c = h(a, b, c) )
(2)
{ Lemma 0 with G := h (and with F as the isomorphic
representation of functions f and g ) }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, c) ∧ c = h(a, b, ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, b, h(a, b, ⊥)) ∧ c = h(a, b, ⊥) )
{ Lemma 0 with G := ( λ a, b, c • g(a, b, h(a, b, ⊥)) ) }
( ↓ a, b, c • a = f (a, b, c) ∧ b = g(a, ⊥, h(a, ⊥, ⊥)) ∧ c = h(a, b, ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, h(a, ⊥, ⊥)) ∧ c = h(a, b, ⊥) )
{ the first 3 steps of this calculation, in reverse order }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, b, c) ∧ c = h(a, b, c) )
{ Lemma 0 with G := g }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, b, c) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, c), c) )
{ Lemma 0 with G := ( λ a, b, c • h(a, g(a, ⊥, c), c) ) }
KRML 118 / VK0099 / MSR-TR-2003-08, p.7
=
=
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), c) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
{ substitute equals for equals }
( ↓ a, b, c • a = f (a, g(a, ⊥, h(a, ⊥, ⊥)), h(a, g(a, ⊥, ⊥), ⊥)) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
{ Lemma 0 with G :=
( λ a, b, c • f (a, g(a, ⊥, h(a, ⊥, ⊥)), h(a, g(a, ⊥, ⊥), ⊥)) ) }
( ↓ a, b, c • a = f (⊥, g(⊥, ⊥, h(⊥, ⊥, ⊥)), h(⊥, g(⊥, ⊥, ⊥), ⊥)) ∧
b = g(a, ⊥, c) ∧ c = h(a, g(a, ⊥, ⊥), ⊥) )
This calculation shows that an expression for the least solution of a in (2) is
f (⊥, g(⊥, ⊥, h(⊥, ⊥, ⊥)), h(⊥, g(⊥, ⊥, ⊥), ⊥))
and similarly for b and c .
Our main result is that the Pruned Closed Form is the least fixpoint in Bn for
any n . In the next section, we prove this result directly, not using Lemma 0.
2 The theorem
We are given n ≥ 1 monotonic functions f1 , . . . , fn : Bn → B , where B is the
boolean domain {0, 1} ordered by ≤ (with 0 ≤ 1 ). To represent an indexed
→
n -tuple of things, like a list of booleans x1 , . . . , xn , we write −
x . The fact that
the given functions are monotonic is written as follows, for any index i and any
→
→
tuples of booleans −
x and −
y :
→→
−
→
−
→
→
x ≤−
y ⇒ f i .−
x ≤ fi .−
y
where an infix dot (with the highest operator precedence) denotes function appli→
−
cation, and the order ≤ is the component-wise ordering of tuples:
→→
−
→
−
x ≤−
y ≡ ( ∀ i • xi ≤ yi )
We are interested in viewing the functions as specifying a system of equations,
namely:
x1 , . . . , xn : x1 = f1 .(x1 , . . . , xn )
...
xn = fn .(x1 , . . . , xn )
(3)
KRML 118 / VK0099 / MSR-TR-2003-08, p.8
where the variables to the left of the colon show the unknowns. We take a tuple of
→
−
functions (f1 , . . . , fn ) , which we can also write as f , to itself be a function, one
which produces a tuple from the results of applying the given argument to each of
→
the functions. For example, for the functions given above and an argument −
x ,
we have:
→
→
→
(f1 , . . . , fn ).−
x = (f1 .−
x , . . . , fn .−
x)
Thus, we can write the system (3) of equations as:
→→
−
→
−
→
x : −
x = f .−
x
→
−
→
We are interested in the least (in the sense of the ordering ≤ ) solution −
x that
satisfies this equation. That is, we are interested in the least fixpoint of the function
→
−
→
−
f . Because the lattice of boolean n -tuples has height n , the least fixpoint of f
−
→
can be reached by applying f n times starting from the bottom element of the
→
−
lattice. That is, the least fixpoint of f is given by:
−n −
→
→
f .0
→
−
where exponentiation denotes successive function applications and 0 is the tuple
of n 0’s.
To precisely specify the Pruned Closed Form, we introduce a notation that
keeps track of which functions have been applied in the enclosing context. In
particular, we use a set that contains the indices of the functions already applied.
Formally, we define the following family of functions, for any index i and set S
of indices:
fi ◦ (gS ∪{i},1 , . . . , gS ∪{i},n ) if i 6∈ S
gS ,i =
−
(λ→
x • 0)
if i ∈ S
Taking advantage of our previous notation and using 0̇ to denote the function that
always returns 0 (that is, the boolean 0 extended pointwise to a boolean function),
we can write the definition of g as follows:
−→
fi ◦ −
g−
S ∪{i} if i 6∈ S
gS ,i =
0̇
if i ∈ S
Our goal is now to prove the following:
Theorem 1
−
→n −
→
−
→
−
→
g∅ . 0 = f . 0
KRML 118 / VK0099 / MSR-TR-2003-08, p.9
3 Proof
We start by proving some lemmas that we use in the proof of this theorem.
Lemma 2 For any index i and for any S $ {1, . . . , n} ,
→n−|S |−1 −
−
→
−
→
gS ,i . 0 ≤ (fi ◦ f
). 0
Proof. By induction on n − |S | . Let T denote S ∪ {i } . We consider three cases.
C ASE i ∈ S :
→
−
gS ,i . 0
=
{ definition of g , since i ∈ S }
→
−
0̇. 0
=
{ definition of 0̇ }
0
≤
{ 0 is bottom element of ≤ }
→n−|S |−1 −
−
→
(fi ◦ f
). 0
C ASE i 6∈ S ∧ |T | < n :
→
−
gS ,i . 0
=
{ definition of g , since i 6∈ S }
→
−
−
→
(fi ◦ gT ). 0
=
→
−
(fi ◦ (gT ,1, . . . , gT ,n )). 0
→
−
=
{ distribute . 0 }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
≤
{ for each index j , induction hypothesis with i , S := j , T , since
|S | + 1 = |T | < n ; and monotonicity of fi }
→n−|S |−2 −
−
→n−|S |−2 −
−
→
→
fi .((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
→
−
=
{ distribute . 0 }
→n−|S |−2
−
→n−|S |−2 −
−
→
(fi ◦ (f1 ◦ f
, . . . , fn ◦ f
)). 0
→n−|S |−2
−
=
{ distribute ◦ f
}
→n−|S |−2 −
−
→
(fi ◦ (f1 , . . . , fn ) ◦ f
). 0
=
{ exponentiation }
→n−|S |−1 −
−
→
(fi ◦ f
). 0
KRML 118 / VK0099 / MSR-TR-2003-08, p.10
C ASE i 6∈ S ∧ |T | = n :
→
−
gS ,i . 0
=
{ see first 3 steps of previous case }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
=
{ for each index j , j ∈ T , so gT ,j = 0̇ }
→
−
fi . 0
=
{ |S | = n − 1 , so f n−|S |−1 is the identity function }
→
−
(fi ◦ f n−|S |−1 ). 0
The following corollary of Lemma 2 proves one direction of Theorem 1.
Corollary 3
→n −
→−
−
→
→−
−
→
g∅ . 0 ≤ f . 0
Proof.
→
−
−
→
g∅ . 0
=
−
→
≤
=
=
→
−
→
−
(g∅,1 . 0 , . . . , g∅,n . 0 )
{ for each index j , Lemma 2 with i , S := j , ∅ }
→n−1 −
−
→n−1 −
−
→
→
((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
→n−1
−
→
−
{ distribute . 0 and ◦ f
}
→n−1 −
−
→
((f1 , . . . , fn ) ◦ f
). 0
{ exponentiation }
→n −
−
→
f .0
To support the remaining lemmas, we define one more family of functions.
For any index i and set S of indices,
fi if i 6∈ S
hS ,i =
0̇ if i ∈ S
Lemma 4 For any index i , monotonic function H : Bn → Bn , and m ≥ 0 ,
→
−
→
0 ≤ p ≤ m • (fi ◦ H p ).−
(fi ◦ H m ). 0 = 0 ⇒ ( ∀ p
0 = 0)
Proof. We prove the term of the quantification as follows:
KRML 118 / VK0099 / MSR-TR-2003-08, p.11
≤
=
=
→
−
(fi ◦ H p ). 0
− −
→
→
→
−
{ monotonicity of fi and H , since 0 ≤ H m−p . 0 }
→
−
(fi ◦ H p ).(H m−p . 0 )
→
−
(fi ◦ H m ). 0
{ antecedent }
0
Lemma 5 For any index i , set S of indices, m ≥ 0 , and T = S ∪ {i } ,
−
→m −
−
→p →
−
→p −
→
→
0 ≤ p ≤ m • hS .−
(fi ◦ hS ). 0 = 0 ⇒ ( ∀ p
0 = hT . 0 )
Proof. If i ∈ S , then S = T and the consequent follows trivially. For i 6∈ S ,
we prove the term of the quantification by induction on p .
C ASE p = 0 : Trivial—exponentiation with 0 gives identity function.
C ASE p > 0 :
−
→p −
→
hS . 0
=
{ exponentiation, since p > 0 }
−
→ −
→p−1 −
→
hS .(hS . 0 )
−
→p−1 −
→
=
{ distribute .(hS . 0 ) }
−
→p−1 −
−
→p−1 −
→
→
(hS ,1.(hS . 0 ), . . . , hS ,n .(hS . 0 ))
−
→p−1 −
−
→p−1 −
→
→
=
{ for any index j , hS ,j .(hS . 0 ) = hT ,j .(hS . 0 ) , see below }
−
→p−1 −
−
→p−1 −
→
→
(hT ,1.(hS . 0 ), . . . , hT ,n .(hS . 0 ))
−
→p−1 −
→
=
{ distribute .(hS . 0 ) }
−
→ −
→p−1 −
→
hT .(hS . 0 )
=
{ induction hypothesis with p := p − 1 }
−
→ −
→p−1 −
→
hT .(hT . 0 )
=
{ exponentiation }
−
→p −
→
hT . 0
Now for the proof of the third step in the calculation above. If j 6= i , then
j ∈ S ≡ j ∈ T , so hS ,j = hT ,j . If j = i , then:
=
−
→p−1 −
→
hS ,i .(hS . 0 )
{ definition of h , since i 6∈ S }
KRML 118 / VK0099 / MSR-TR-2003-08, p.12
=
=
=
−
→p−1 −
→
fi .(hS . 0 )
−
→
{ Lemma 4 with H , p := hS , p − 1 , using the antecedent of
Lemma 5 to fulfill the antecedent of Lemma 4 }
0
{ definition of 0̇ }
−
→p−1 −
→
0̇.(hS . 0 )
{ definition of h , since i ∈ T }
−
→p−1 −
→
hT ,i .(hS . 0 )
We need one more lemma.
Lemma 6 For any index i , set S of indices, and m satisfying 0 ≤ m ≤ n −|S | ,
−
→m −
→
→
−
(hS ,i ◦ hS ). 0 ≤ gS ,i . 0
Proof. By induction on m . We consider three cases.
C ASE i ∈ S :
−
→m
hS ,i ◦ hS
=
{ definition of h , since i ∈ S }
−
→m
0̇ ◦ hS
=
{ 0̇ is left zero element of ◦ }
0̇
=
{ definition of g , since i ∈ S }
gS ,i
C ASE i 6∈ S ∧ m = 0 :
−
→m −
→
(hS ,i ◦ hS ). 0
=
{ exponentiation, since m = 0 }
→
−
hS ,i . 0
=
{ definition of h , since i 6∈ S }
→
−
fi . 0
→−
−
→
→ →−
≤
{ monotonicity of fi , since 0 ≤ −
gT . 0 }
→
−
fi .(−
g→
T. 0 )
=
{ definition of g , since i 6∈ S }
→
−
gS ,i . 0
(4)
KRML 118 / VK0099 / MSR-TR-2003-08, p.13
C ASE i 6∈ S ∧ m > 0 : It suffices to prove that the left-hand side of (4) is 0
whenever the right-hand side is 0. Therefore, we assume the latter to be 0:
→
−
gS ,i . 0 = 0
(5)
and prove the former to be 0:
−
→m −
→
(hS ,i ◦ hS ). 0
=
{ definition of h , since i 6∈ S }
−
→m −
→
(fi ◦ hS ). 0
=
{ exponentiation, since m > 0 }
−
→m−1 −
→
(fi ◦ (hS ,1 , . . . , hS ,n ) ◦ hS
). 0
−
→m−1
→
−
=
{ distribute ◦ hS
and . 0 }
−
→m−1 −
−
→m−1 −
→
→
fi .((hS ,1 ◦ hS
). 0 , . . . , (hS ,n ◦ hS
). 0 )
≤
{ (6), see below; and monotonicity of fi }
→
−
→
−
fi .(gT ,1. 0 , . . . , gT ,n . 0 )
=
→
−
(fi ◦ −
g→
T ). 0
=
{ definition of g , since i 6∈ S }
→
−
gS ,i . 0
=
{ assumption (5) }
0
In this calculation, we used the following fact: for every index j ,
−
→m−1 −
→
→
−
(hS ,j ◦ hS
). 0 ≤ gT ,j . 0
(6)
which we now prove. We divide the proof of (6) up into two sub-cases.
−
→m−1 −
→
S UB - CASE (hS ,j ◦ hS
). 0 = 0 : Formula (6) follows immediately.
−
→m−1 −
→
S UB - CASE (hS ,j ◦ hS
). 0 6= 0 : First, we derive some consequences of
assumption (5):
→
−
gS ,i . 0 = 0
⇒
{ induction hypothesis with S , i , m := S , i , m − 1 }
−
→m−1 −
→
(hS ,i ◦ hS
(7)
). 0 = 0
=
{ definition of h , since i 6∈ S }
−
→m−1 −
→
(fi ◦ hS
). 0 = 0
⇒
{ Lemma 5 with m, p := m − 1, m − 1 }
−
→m−1 −
−
→m−1 −
→
→
hS
(8)
. 0 = hT
.0
KRML 118 / VK0099 / MSR-TR-2003-08, p.14
Now, calculating from the assumption we made in this sub-case:
−
→m−1 −
→
(hS ,j ◦ hS
). 0 6= 0
=
{ (7) }
−
→m−1 −
→
(hS ,j ◦ hS
). 0 6= 0 ∧ i 6= j
⇒
{ i 6= j , so j ∈ S ≡ j ∈ T , so hS ,j = hT ,j }
−
→m−1 −
→
(hT ,j ◦ hS
). 0 6= 0
=
{ (8) }
−
→m−1 −
→
(hT ,j ◦ hT
). 0 6= 0
⇒
{ induction hypothesis with S , i , m := T , j , m − 1 }
→
−
gT ,j . 0 6= 0
⇒
(6)
This concludes the proof of Lemma 6.
And finally, the proof of the theorem:
Proof of Theorem 1. The proof is a ping-pong argument.
→
−
→
−
g∅ . 0
→
−
≤
{ Corollary 3 }
→n −
−
→
f .0
—ping!
=
{ exponentiation, since n ≥ 1 }
→n−1 −
−
→
((f1 , . . . , fn ) ◦ f
). 0
→n−1
−
→
−
=
{ distribute ◦ f
and . 0 }
→n−1 −
−
→n−1 −
−
→
→
((f1 ◦ f
). 0 , . . . , (fn ◦ f
). 0 )
=
{ by definition of h , h∅,i = fi for each index i ; and
→ −
−
→
thus also h∅ = f }
→n−1 −
−
→n−1 −
−
→
→
((h∅,1 ◦ h∅ ). 0 , . . . , (h∅,n ◦ h∅
). 0 )
→
−
≤
{ Lemma 6 with S , i , m := ∅, i , n − 1 for each i }
→
−
→
−
(g∅,1 . 0 , . . . , g∅,n . 0 )
→
−
=
{ distribute . 0 }
→
−
→
−
g∅ . 0
—pong!
4 Related Work and Acknowledgments
Our theorem has already found a use, namely in the translation of boolean programs into satisfiability formulas [3].
KRML 118 / VK0099 / MSR-TR-2003-08, p.15
Before we knew of the Bekić-Leszczyłowski Theorem, one of us (Kuncak)
proved the theorem as detailed in Section 3. Tony Hoare then proposed a way to
prove the theorem in a way that would eliminate recursive uses of variables, one
by one. In doing this, Hoare also proved what essentially amounts to the BekićLeszczyłowski Theorem, appealing only to the Tarski Fixpoint Theorem [5]. We
elaborated this format in Section 1, to whose formulation Carroll Morgan also
contributed. We learnt about the Bekić-Leszczyłowski Theorem from Patrick
Cousot. The theorem is often called simply the Bekić Theorem, but de Bakker [0]
traces an independent proof thereof to Leszczyłowski. Finally, we are grateful for
feedback from the Eindhoven Tuesday Afternoon Club and the participants of the
IFIP WG 2.3 meeting in Biarritz, France (March 2003).
References
[0] J. de Bakker. Mathematical Theory of Program Correctness. Prentice-Hall,
1980.
[1] Hans Bekić. Definable operation in general algebras, and the theory of automata and flowcharts. In Cliff B. Jones, editor, Programming Languages and
Their Definition—Hans Bekić (1936–1982), volume 177 of Lecture Notes in
Computer Science, pages 30–55. Springer, 1984.
[2] Stephen Cole Kleene. Introduction to Metamathematics. D. Van Nostrand,
New York, 1952.
[3] K. Rustan M. Leino. A SAT characterization of boolean-program correctness.
In Thomas Ball and Sriram K. Rajamani, editors, Model Checking Software,
volume 2648 of Lecture Notes in Computer Science, pages 104–120. Springer,
May 2003.
[4] Jacek Leszczyłowski. A theorem on resolving equations in the space of languages. Bull. Acad. Polon. Sci., 19:967–970, 1971.
[5] Alfred Tarski. A lattice-theoretical fixpoint theorem and its applications. Pacific Journal of Mathematics, 5:285–309, 1955.
| 2cs.AI
|
An Overview of Particle Methods for Random Finite
Set Models
arXiv:1602.03945v1 [cs.SY] 12 Feb 2016
Branko Ristica,∗, Michael Beardb , Claudio Fantaccic
a RMIT
University, Australia
University, Australia
Università di Firenze, DINFO, Florence, Italy
b Curtin
c
Abstract
This overview paper describes the particle methods developed for the implementation of the a class of Bayes filters formulated using the random finite set
formalism. It is primarily intended for the readership already familiar with the
particle methods in the context of the standard Bayes filter. The focus in on the
Bernoulli particle filter, the probability hypothesis density (PHD) particle filter
and the generalised labelled multi-Bernoulli (GLMB) particle filter. The performance of the described filters is demonstrated in the context of bearings-only
target tracking application.
Keywords: Stochastic nonlinear filtering, Monte Carlo estimation, particle
methods, random set models, target tracking, bearings-only measurements
∗ Corresponding author: B. Ristic, RMIT University, School of Electrical and Computer
Engineering, Rm 10.8.14, 376-392 Swanston street, Melbourne, VIC 3000, Australia; email:
[email protected]; tel: +61 3 9925 3768
Preprint submitted to Information Fusion
February 15, 2016
Contents
1 Introduction
3
2 Particle methods for the standard Bayes filter
2.1 Problem formulation and the standard Bayes filter
2.2 A primer on the particle method . . . . . . . . . .
2.3 Calibration of system parameters . . . . . . . . . .
2.4 Demonstration: Bearings-only filtering . . . . . . .
4
5
6
8
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Background: FISST
10
3.1 Random finite sets . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Some common RFS variables . . . . . . . . . . . . . . . . . . . . 14
4 Particle methods for the RFS Bayes-optimal filter
4.1 Formulation of the RFS Bayes-optimal filter . . . . .
4.2 Particle method approximations . . . . . . . . . . . .
4.3 Bernoulli particle filter . . . . . . . . . . . . . . . . .
4.4 Demonstration: Bearings-only detection and tracking
5 PHD particle filters
5.1 Formulation of the PHD filter . . . . . . . . . . . .
5.2 The particle method applied to PHD filtering . . .
5.3 Calibration of tracking algorithms . . . . . . . . .
5.4 Demonstration: Bearings-only multi-object filtering
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
17
20
21
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22
23
24
29
30
6 Labelled RFS Bayes tracking filters
6.1 Labelled RFS . . . . . . . . . . . . . . . . . . . . . . .
6.2 Labelled multi-object transition and likelihood models
6.3 δ-GLMB particle filter . . . . . . . . . . . . . . . . . .
6.4 Demonstration: Bearings-only multi-object tracking .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
32
34
35
42
7 Summary and further reading
.
.
.
.
43
2
Notation
X
F (X )
L
x
X
X
Z
z
Z
k
p(x)
p̃(x, ℓ)
f (X)
f˜(X)
the single object state space
the multiple object state space
the space of labels
the state of a single object (a random vector)
the state of multiple-objects (a random finite set, RFS)
the state of multiple labelled objects (a labelled RFS)
measurement space
a measurement of a single object (a random vector)
a detector output measurement (a RFS)
discrete-time index
a probability density function (PDF) of x ∈ X
a PDF of a labelled random vector (x, ℓ) ∈ X × L
a PDF of an RFS variable X ∈ F (X )
a PDF of a labelled RFS variable X ∈ (F (X ) × L)
1. Introduction
In many areas of science and engineering there is a need to infer the behaviour
of a stochastic dynamic system, using its partial and indirect observations. By
combining (typically nonlinear) mathematical models of system evolution and
sensor measurements, one can formulate the optimal sequential estimator in
the Bayesian framework. This estimator, commonly referred to as the Bayesoptimal (or simply Bayes) filter, provides a recursive formula for the complete
probabilistic characterisation of the dynamic system in the form of a timevarying posterior probability density of its state [1].
For most nonlinear/non-Gaussian formulations, analytic closed-from solutions of the Bayes filter are intractable. Practical solutions, therefore, need
to be based on approximations. Particle filters are a class of Monte Carlo
simulation based methods which can provide very accurate approximations of
the Bayes filter. Despite being computationally expensive, particle filters have
become universally popular, primarily due to their accuracy, relatively simple
implementation and the ever increasing speed of computers. As a result of their
widespread application, a few good tutorials and books have been published on
the subject of particle filters [2], [3], [4], [5], [6], [7].
Particle filters have been introduced and traditionally applied as the approximate solutions of the standard Bayes filter, formulated during the 1960s
[1] under the following assumptions: (i) the stochastic dynamic system (object,
3
phenomenon) is permanently active (or present); (ii) observations are noisy, but
collected with perfect detection (i.e. there are no false or missed detections). All
the aforementioned tutorials and books discuss the particle filters in this context
only. However, in many practical applications, one may have to deal with multiple stochastic dynamic systems (objects), which can be simultaneously active
(present), and which can randomly switch on and off (appear/disappear). In
addition, perfect detection using surveillance sensors (e.g. radar, sonar, video
cameras) is rarely possible [8]. Until recently, particle filters have been applied
to this class of problems using a clever combination of Bayesian estimation theory with ad-hoc logic. However, the recent advances in Bayesian estimation
using random finite set (RFS) models [9] resulted in elegant and rigorous mathematical formulations of the Bayes-optimal and principled Bayes-suboptimal
filters, applicable to multiple interacting on/off switching systems with possibly
imperfect detection of measurements.
This overview paper describes the particle methods developed for the implementation of the new class of RFS-Bayes filters. It is primarily intended for
the readership already familiar with the particle methods in the context of the
standard Bayes filter. One of the most popular and convincing applications of
particle filters, versus standard approximation methods, such as the Extended
Kalman filter (EKF) [10] and unscented Kalman filter (UKF) [11], has been
for bearings-only tracking problems [5, Ch.6]. Hence, this application has been
chosen to demonstrate throughout the paper different RFS-Bayes particle filters
and their performance. The paper is organised as follows. Sec. 2 reviews the
particle method for the standard Bayes filter. The elements of mathematics for
random finite set models are presented in Sec. 3. The particle method for the
RFS Bayes-optimal filter and its special case, the Bernoulli filter, are discussed
in Sec. 4. A multi-target particle filter, referred to as the PHD particle filter,
is presented in Sec. 5. The labelled RFS Bayes tracking filters implemented
using the particle method are discussed in Sec. 6. The summary and pointers
to advanced research topics are given in Sec. 7.
2. Particle methods for the standard Bayes filter
In order to familiarise with the notation, let us start with a quick review of
the standard Bayes filter and the corresponding particle methods. The problem
is by no means simple and it is still an active and highly relevant research topic.
4
2.1. Problem formulation and the standard Bayes filter
Suppose the state vector xk ∈ X provides the complete specification of the
state of a dynamic system (object, phenomenon) at time tk . Here X ⊆ Rnx is
the state space, while k is the discrete-time index corresponding to tk . Let us
adopt the discrete-time additive-noise formulation, specified by two equations:
xk = fk−1 (xk−1 ) + vk−1 ,
(1)
zk =
(2)
hk (xk ) + wk ,
referred to as the dynamics equation and the measurement equation, respectively.
Function fk−1 : Rnx → Rnx in (1) is a nonlinear transition function defining the
temporal evolution of the state vector as a first-order Markov process. Random
disturbances vk ∈ Rnx , also known as process noise, are assumed to be independent identically distributed (IID) according to the probability density function
(PDF) pv . Function hk : Rnx → Rnz in (2) defines the relationship between
the state xk and the measurement zk ∈ Z, where Z ⊆ Rnz is the measurement
space. Random disturbances wk ∈ Rnz , also known as measurement noise, are
assumed independent of vk , and modelled as an IID process with the PDF pw .
Typically nz < nx , giving rise to the term partial observations of the system.
In the formulation specified by (1)-(2), the functions fk and hk , the probability distributions pv and pw , and the PDF of the state vector at initial time
k = 0, (i.e. p0 (x0 )), are all assumed known. Equations (1) and (2) effectively
define two probability functions: the transitional density πk|k−1 (xk |xk−1 ) =
pv (xk − fk−1 (xk−1 )) and the likelihood function gk (zk |xk ) = pw (zk − hk (xk )).
The problem is to compute recursively the posterior PDF of the state, denoted
as pk|k (xk |z1:k ) at discrete-time k, where the notation z1:k stands for the sequence z1 , z2 , · · · , zk .
The solution is usually presented as a two step procedure. Let pk−1|k−1 (xk−1 |z1:k−1 )
denote the posterior PDF at k − 1. The first step predicts the density of the
state to time k via the Chapman-Kolmogorov equation [1]:
Z
pk|k−1 (xk |z1:k−1 )= πk|k−1 (xk |x′ )p(x′ |z1:k−1 )dx′ .
(3)
The second step applies Bayes rule to update p(xk |z1:k−1 ) using measurement
zk :
gk (zk |xk ) pk|k−1 (xk |z1:k−1 )
.
(4)
pk|k (xk |z1:k ) = R
gk (zk |x)pk|k−1 (x|z1:k−1 )dx
Knowing the posterior pk|k (xk |z1:k ), one can compute a point estimate of the
state x̂k (e.g. as the mean or the mode of the posterior) and a confidence (or
credible) interval.
5
The closed-form analytic solution to (3)-(4) can be found only in some special cases. One important case is when fk and hk are linear functions and
PDFs pv , pw and p0 are Gaussian; the solution in this case is the Kalman
filter. In general, however, stochastic filtering via (3)-(4) can be solved only
numerically. Many algorithms have been proposed for this purpose, including
analytic approximations (e.g. Extended Kalman filter and its variants), gridbased methods (where the posterior PDF is evaluated at a finite and fixed set
of points), Gaussian sum filters (where the posterior PDF is approximated by
a Gaussian mixture), unscented transforms [11] and particle filters [5], [12].
2.2. A primer on the particle method
Suppose the posterior density at discrete-time k − 1 is approximated by
(i)
(i)
(i)
a set of random samples (particles) {wk−1 , xk−1 }1≤i≤N , where xk−1 is the
(i)
state of particle i and wk−1 is its weight. The weights are normalized, that
PN
(i)
is i=1 wk−1 = 1. This approximation of the posterior improves as N → ∞.
(i)
(i)
Given {wk−1 , xk−1 }1≤i≤N and using the measurement zk at time k, the key
question is how to form the particle approximation of the posterior at k, i.e.
(i)
(i)
pk|k (xk |z1:k ), denoted {wk , xk }1≤i≤N .
The computation of the weights and particles at time k is based on the
concept of importance sampling [13]. Let us introduce a proposal or importance
density qk (xk |xk−1 , zk ), whose support contains the support of the posterior
PDF at time k. Then the (preliminary) particles at time k are drawn from the
importance density:
(i)
(i)
x̃k ∼ qk (xk |xk−1 , zk ),
(5)
whose weights are computed as follows:
(i)
w̃k
=
(i)
=
wk
(i)
(i) (i)
(i) gk (zk |x̃k ) πk|k−1 (x̃k |xk−1 )
wk−1
(i) (i)
qk (x̃k |xk−1 , zk )
(6)
(i)
w̃k
PN
j=1
(7)
(j)
w̃k
for i = 1, . . . , N . This recursive procedure starts at time k = 0 by sampling N
times from the initial PDF p0 .
The described particle method, also known as sequential importance sampling (SIS), inevitably fails after many iterations, because all particle weights,
except a few, become zero (a poor approximation of the posterior PDF due
to particle degeneracy). The collapse of the SIS scheme can be prevented
by resampling the particles. The resampling step chooses N particles from
6
(i)
(i)
{wk , x̃k }1≤i≤N , where the selection of particles is based on their weights: the
(i)
probability of particle i being selected during resampling equals wk . After
resampling, all particle weights are equal to 1/N . While resampling avoids degeneracy of particles, it leads to the loss of diversity among the particles, because
the particles with large weights are selected (repeated) many times. In order to
increase the particle diversity, it is usually recommended to perform a Markov
chain Monte Carlo (MCMC) move step after resampling [14], [5], [6].
The choice of the importance density qk plays an important role in the implementation of the particle filter (PF). The simplest choice is to select qk as
the transitional density, i.e. qk ≡ πk|k−1 . This PF, referred to as the bootstrap
filter [15], can lead to poor performance because many of the particles could be
sampled from the region of the state space which is not in the support of the
posterior (and therefore wasted). A better strategy is to use the information
contained in the latest measurement zk in the design of the importance density. Research into good importance densities have resulted in many versions
of the particle filter, such as: the optimal importance density (OID) PF [16],
the auxiliary PF [17], the local-linearisation PF [16], [18], exact particle flow
nonlinear filters [19], particle filters with progressive correction or tempering
[20] and particle filters using Laplace approximation [21].
The bootstrap filter, being the simplest and possibly the most popular PF,
is described in more detail and its pseudo code (for a single processing cycle at
time k) is given in Alg. 1. Note that, since resampling is performed at the end
of every cycle, there is no need to input/output particle weights. Recall also
that after resampling it is recommended to carry out an MCMC move step.
Algorithm 1 Pseudo-code of the bootstrap filter
1: function Bootstrap Filter
(i)
2:
Input: {xk−1 }1≤i≤N ; zk
3:
for i = 1, . . . , N do
(i)
(i)
4:
Draw a sample: x̃k ∼ πk|k−1 (xk |xk−1 )
5:
6:
7:
8:
9:
(i)
(i)
Calculate weight w̃k = gk (zk |x̃k )
end for
(i)
(i) P
(j)
wk = w̃k / N
j=1 w̃k , for i = 1, . . . , N
for i = 1, . . . , N do
(i)
Select index j i ∈ {1, . . . , N } with probability wk
(i)
⊲ (Resampling)
(j )
10:
xk = x̃k i
11:
end for
(i)
12:
Apply MCMC move and output {xk }1≤i≤N
13: end function
A point estimate of the state at time k, denoted x̂k , can be computed from
7
(i)
(i)
the particle approximation {wk , xk }1≤i≤N of the true posterior pk|k (xk |z1:k )
either: (a) as the weighted mean of the particles, referred to as the expected
a posterior (EAP) estimate, or (b) as the mode of the density estimated from
the particles (using for example the kernel density estimation method, see [22]),
referred to as the maximum a posterior (MAP) estimate.
2.3. Calibration of system parameters
In many practical applications of nonlinear stochastic filtering, the transitional density and/or the likelihood function are dependent on a static parameter
vector θ ∈ Θ ⊆ Rnθ . This is indicated in notation as πk|k−1 (xk |xk−1 , θ) and
gk (zk |xk , θ). The problem is to estimate the posterior density p(θ|z1:k ) (i.e. to
calibrate the system) given its prior p(θ) and observations z1:k .
Note that according to Bayes rule p(θ|z1:k ) ∝ ̺(z1:k |θ) p(θ), where ̺(z1:k |θ)
is the likelihood function. If we knew ̺(z1:k |θ), we could apply standard
Bayesian parameter estimation techniques (e.g. MCMC, population Monte
Carlo [13]). In general, however, ̺(z1:k |θ) cannot be expressed in closed-form.
One natural option is to augment the state vector and carry out sequential
Bayesian estimation on the joint space X × Θ. Unfortunately, this is not a good
idea, because the direct use of particle filtering in this case is inefficient [6] due
to the absence of stochastic evolution for θ (i.e. the particles in Θ space are
sampled effectively only once).
The key idea of particle MCMC methods is to estimate the likelihood ̺(z1:k |θ)
using the particle filter [23]. Note first the following decomposition of the likelihood:
k
Y
̺(z1:k |θ) = ̺(z1 |θ)
̺(zt |z1:t−1 , θ)
(8)
t=2
where the terms ̺(zt |z1:t−1 , θ) can be estimated from the un-normalised weights
of the particles, see (6), at time step t [6]:
Z
X (i,θ )
̺(zt |z1:t−1 , θ) = gt (zt |x)pt|t−1 (x|z1:t−1 )dx ≈
w̃t
.
(9)
1≤i≤N
Using for example the Metropolis-Hasting MCMC algorithm, the parameter
vector values θ∗ are iteratively generated from the proposal distribution q(θ|θ′ )
in order to be accepted or rejected. Note that for each proposed θ∗ , one has to
run the PF to estimate its likelihood ̺(z1:k |θ∗ ).
Suppose next that the parameter vector is also time varying. For example,
its evolution may also be modeled by a first-order Markov process. In this case,
although the inference needs to be made on the joint space X × Θ, the particle
8
filter can be executed on sub-spaces Θ and X |Θ because the posterior PDF
can be factorized as pk|k (xk , θk |z1:k ) = pk|k (xk |θk , z1:k )pk|k (θk |z1:k ). In some
applications, the conditional posterior pk|k (xk |θ k , z1:k ) is analytically tractable
(e.g. linear/Gaussian case), which leads to the Rao-Blackwellised formulation
of the marginalised PF [6], [24].
2.4. Demonstration: Bearings-only filtering
The problem of bearings-only filtering (or tracking) arises in a variety of
important applications, including submarine tracking, using a passive sonar,
and aircraft surveillance, using a radar in a passive mode [5]. The objective is to
sequentially estimate the kinematics of a moving object using noise-corrupted
bearing measurements. Two features of the problem are noteworthy [5],[25]:
(a) the observation platform needs to manoeuvre in order to estimate the target
range; (b) the problem becomes particulary difficult during the time steps when
the bearings-rate is high. The particle filter, as a universal method for nonlinear
filtering, has been demonstrated to outperform both EKF and UKF in this
application1 [5],[25].
Let us adopt the state vector of the moving object as
i⊺
h
m
m
m
m
(10)
xm
=
x
y
ẋ
ẏ
k
k
k
k
k
m
m m
where (xm
k , yk ) and (ẋk , ẏk ) are its position and velocity in Cartesian coordinates, respectively. The observer state vector xok , which is known, is similarly
defined. The dynamic (motion) model is written for the relative state vector,
and is given by:
i⊺
h
o
(11)
xk := xm
yk ẋk ẏk .
k − xk = xk
We adopt a nearly constant velocity (CV) motion model, as a linear version of
(1):
xk+1 = Fxk − Uk+1,k + Γvk
(12)
where
"
1
F=
0
#
T
⊗ I2 ,
1
Uk+1,k
xok+1 − xok − T ẋok
y o − y o − T ẏ o
= k+1o k o k ,
ẋk+1 − ẋk
o
o
ẏk+1 − ẏk
1 Notice
"
#
T 2 /2
Γ=
⊗ I2 . (13)
T
that the shifted Rayleigh filter [25], which is also a very accurate bearings-only
filter, is not a universal nonlinear filter.
9
Explanation: ⊗ is the Kroneker product; In is identity matrix of dimension n;
T = tk+1 − tk is the (constant) sampling interval; F is the transition matrix;
Uk+1,k is a known deterministic matrix taking into account the effect of observer
accelerations; vk ∼ N (0, Q) is white Gaussian process noise with Q = σv2 I2 .
The available measurement at time k is the bearing measurement from the
observer platform to the target, referenced clockwise positive to the y-axis and
specified by a scalar version of (2):
zk = h(xk ) + wk ,
(14)
where h(xk ) = atan2(xk , yk ) is the four-quadrant inverse tangent function and
2
wk is zero-mean white Gaussian noise with variance σw
.
Fig. 1 illustrates the scenario and the performance of the particle filter.
Fig. 1.(a) shows the top-down observer-target geometry, together with the EAP
estimates from the PF. This scenario was copied from [26], but considers only
one of the targets (the full scenario with four targets that appear/disappear at
different times, will be used later). The circles indicate the starting points of
the two trajectories. Red dots indicate the cloud of particles at t = 1200[s],
t = 1800[s] and t = 2400[s]. Figs. 1.(b) and (c) display the positional and
velocity RMS error (over time), obtained by averaging over 100 Monte Carlo
runs. The parameters used in simulations were as follows: σw = 0.3o , T =
20[s], σv = 0.005 [m/s2 ], N = 5000 particles. Nonlinear filtering was carried
out during the period when the target was present, that is from t = 200[s] to
t = 2400[s], which corresponds to 111 discrete-time steps. Note that Fig. 1.(a)
also shows the clouds of particles at time instances 1200[s], 1800[s] and 2400[s].
The particle filter is initalised by drawing samples from the initial or birth
density designed using the first bearing measurement z1 , as well as prior knowledge of (i) the sensor range rmax and (ii) the maximum speed of the target vmax .
The birth density in target position is the sector of a circle (“pizza slice”), centred at the observer position at k = 1, i.e. (xo1 , y1o ), determined by the angles
z1 ± 3σw and the radius rmax = 10000[m]. The target birth density in velocity
m
is uniform PDF U[−vmax , +vmax ] in both ẋm
1 and ẏ1 , with vmax = 7.5 [m/s].
3. Background: FISST
The prerequisite for the random finite set formulation of the Bayes-optimal
filter is advanced mathematics, referred to as finite set statistics (FISST) [27],
developed by Ron Mahler. This section will introduce only the bare minimum
necessary to follow the rest of the paper. The full theoretical details are beyond
10
6000
(a)
Observer
y [m]
4000
Estimated
trajectory
2000
t=1800s
Target
t=2400s
t=1200s
0
Particles
−5000
0
5000
10000
x [m]
6
(b)
3000
RMS error in VELOCITY [m/s]
RMS error in POSITION [m]
3500
2500
2000
1500
1000
500
0
0
500
1000
1500
2000
(c)
5
4
3
2
1
0
0
2500
500
Time [s]
1000
1500
2000
2500
Time [s]
Figure 1: Particle filter applied to bearings-only filtering problem (σw = 0.3o ): (a) top-down
view of the scenario and the estimated trajectory (red dots are particle clouds); (b) RMS error
in position; (c) RMS error in velocity
the scope of this paper, however the proofs and derivations can be found in
Mahler’s excellent books [27, Part II] and [9, Part I].
3.1. Random finite sets
A random finite set is a convenient probabilistic model for the representation of multiple stochastic dynamic systems (objects) and sensor measurements.
Suppose that at discrete-time k there are nk objects with states xk,1 , . . . , xk,nk ,
taking values in the state space X ⊆ Rnx . Both the number of dynamic objects nk and their individual states in X are random and time-varying. The
multi-object state at k is a finite set
Xk = {xk,1 , . . . , xk,nk } ∈ F (X ),
11
(15)
which can be modelled as a random finite set on X . Here F (X ) is the set of
finite subsets of X .
Typically, the detection process is imperfect, meaning that not all of the
objects in Xk are detected, while spurious (false) detections may be created due
to background noise and interferences. Suppose that Zk is a measurement set
from such a detector, reported at time k. Then
Zk = {zk,1 , . . . , zk,mk } ∈ F (Z),
(16)
can also be modelled by a random finite set on the observation space Z ⊆ Rnz .
Both the cardinality mk = |Zk | and the individual states in Z are random. Here
F (Z) is the set of finite subsets of Z.
A random finite set (RFS) variable is a random variable that takes values as
unordered finite sets. The cardinality of an RFS variable X is random and modelled by a discrete distribution ρ(n) = P r{|X| = n}, where n is a non-negative
integer, n ∈ N ∪ {0}. An RFS X is completely specified by its cardinality distribution ρ(n) and a family of symmetric joint distributions2 pn (x1 , . . . , xn ), that
characterise the distribution of its elements over the state space, conditioned on
cardinality n.
Since an RFS is nothing but a finite-set valued random variable, the usual
probabilistic descriptors of a random variable, such as the PDF and its statistical
moments, can be defined for it. FISST provides the tools for mathematical
representation and manipulation of random finite sets3 . The FISST PDF of an
RFS variable X is denoted f (X) and defined as:
f ({x1 , . . . , xn }) = n! · ρ(n) · pn (x1 , . . . , xn ).
(17)
For example, f (∅) = ρ(0), f ({x}) = ρ(1)p(x), f ({x1 , x2 }) = 2ρ(2) p2 (x1 , x2 ),
etc. Note that f (∅), f ({x}), f ({x1 , x2 }), etc, have different units. Throughout
the text, we will refer to the FISST densities as to the multi-object PDFs.
Being a PDF, f (X) should integrate to one. However, the standard integral
cannot be applied; instead we need to introduce the set integral, defined as
2A
joint distribution function pn (x1 , . . . , xn ) is said to be symmetric if its value remains
unchanged for all of the n! possible permutations of its variables.
3 Note that while the FISST densities are not probability densities, they have been shown
to be equivalent to probability densities on F (X) relative to some reference measure [28].
Subsequently, we do not distinguish between FISST densities and probability densities of
random finite sets.
12
follows:
Z
Z
∞
X
1
f (X) δX := f (∅) +
f ({x1 , · · · , xn })dx · · · dxn
n!
n=1
(18)
Now it is straightforward to verify that indeed f (X) integrates to 1:
Z
f (X) δX =
ρ(0) +
∞
X
ρ(n)
n=1
=
∞
X
Z
pn (x1 , . . . , xn )dx1 · · · dxn
{z
}
|
=1, being a standard PDF
ρ(n) = 1
(since ρ(n) is a discrete distribution)
n=0
The cardinality distribution ρ(n) of an RFS variable X can be obtained from
the multi-object PDF f (X) as:
Z
1
ρ(n) =
f ({x1 , . . . , xn })dx1 , . . . dxn
(19)
n!
The intensity function (also known as the probability hypothesis density
or PHD) is an important characterisation of an RFS X on X , defined as its
first-order statistical moment. In order to define the PHD function, let us first
introduce the set Dirac delta function
X
δX (x) =
δw (x)
w∈X
with δw (x) being the standard Dirac delta function concentrated at w. Now we
can express the cardinality of an RFS X as follows:
Z
δX (x) dx .
(20)
|X| =
X
We would like to define the PHD function D(x) of X in such a way that the
expected cardinality of X over the state space X is obtained as the integral
Z
E{|X|} =
D(x) dx .
(21)
X
Since
E{|X|}
,
=
=
Z
|X|f (X) δX
Z Z
δX (x) dx f (X) δX
X
Z Z
δX (x) f (X) δX dx ,
X
13
(22)
(23)
(24)
comparing (21) with (24), notice that
Z
D(x) , E{δX (x)} =
δX (x) f (X) δX.
(25)
Here E{δX (x)} is the expected value (the first statistical moment) of the RFS
X. Note that D(x) is a density function on the state space X .
In general, the PHD function D(x) and the cardinality distribution ρ(n) do
not completely characterize the multi-object PDF. However, we will see in the
next section that for some specific RFSs, the characterization is complete.
3.2. Some common RFS variables
Bernoulli RFS. This RFS can either be empty (with probability 1 − r) or
have one element (with probability r), spatially distributed over X according to
the (standard) PDF p(x). The FISST PDF of the Bernoulli RFS X is given by:
1 − r,
if X = ∅
(26)
f (X) =
r · p(x), if X = {x}.
The intensity function of the Bernoulli RFS X is D(x) = r · p(x).
IID Cluster RFS. Let the cardinality distribution of independent identically
distributed (IID) cluster RFS X be ρ(n). For a given cardinality, the elements of
X are each IID random variables distributed according to the (standard) PDF
p(x) on X . The multi-object PDF of X is:
Y
f (X) = |X|! · ρ(|X|)
p(x)
(27)
x∈X
The main simplification in comparison with (17) is that due to the IID property,
the symmetric joint distribution is replaced by a product. The intensity function
of an IID cluster RFS is:
D(x) = p(x)
∞
X
n ρ(n)
(28)
n=1
Poisson RFS. If the cardinality distribution of an IID cluster RFS is Poisson
with parameter λ > 0, i.e.
e−λ λn
,
n = 0, 1, 2, . . .
(29)
n!
then this RFS is referred to as the Poisson RFS X. Its multi-object PDF follows
from (27) and is given by:
Y
f (X) = e−λ
λ p(x)
(30)
ρ(n) =
x∈X
14
while its intensity function is
D(x) = λ p(x).
(31)
The Poisson RFS is the only RFS which is completely specified by its intensity
function, because knowledge of λ and p(x) fully determines the multi-object
PDF in (30). Both λ and p(x) can be worked out from D(x): λ is the expected
R
cardinality of X, i.e. E{|X|} = D(x)dx = λ, while p(x) = D(x)/λ.
Multi-Bernoulli RFS. A multi-Bernoulli RFS X is a union of independent
Bernoulli RFSs Xi , each characterized by existence probability ri and the spatial
PDF pi (x), for i = 1, . . . , M :
M
[
Xi .
X=
i=1
The multi-object PDF of X = {x1 , . . . , xn } is given by:
M
X
Y
ri1 pi1 (x1 )
ri pi (xn )
f (X) = n! (1 − rj ) ·
··· n n
.
1 − ri1
1 − rin
j=1
(32)
1≤i1 <···<in ≤M
Q
P
ri pi (x)
For example, f (∅) = M
(1 − rj ), f ({x}) = f (∅) M
i=1 1−ri , etc. Note that
j=1
the pairs ri , pi (x) , i = 1, . . . , M , fully specify the multi-object PDF (32).
Consequently, they also determine the intensity function of the multi-Bernoulli
PM
RFS, which is given by: D(x) = i=1 ri pi (x).
4. Particle methods for the RFS Bayes-optimal filter
4.1. Formulation of the RFS Bayes-optimal filter
The goal of the RFS Bayes-optimal filter is to estimate the posterior density
of a multi-object state, represented by the RFS variable Xk . Evolution of Xk
is modelled by a Markov process, characterized by its initial FISST density
f0 (X0 ) and the FISST transitional density Πk|k−1 (Xk |Xk−1 ). The standard
case of Πk|k−1 (Xk |Xk−1 ) has been derived in Chapter 13 of [27] as the union
of a multi-Bernoulli object survival RFS and a Poisson birth RFS, under the
assumption that the constituent RFSs are mutually independent. It can be
expressed in a compact form as:
′
Πk|k−1 (X|X′ ) = fb (X)(1 − pS )|X |
X Y
θ
i:θ(i)>0
pS · πk|k−1 (xθ(i) |x′i )
(1 − pS ) · µ0 bk (xθ(i) )
(33)
where θ : {1, . . . , |X′ |} → {0, 1, . . . , |X|} represents a list of all possible assignments of elements from the set X′ to the elements of the set X (here the
15
assumptions are that |X′ | > 0 and |X| ≥ 0); pS is the probability of object
survival from time k − 1 to k; fb (X) is the object birth multi-object PDF, which
assuming a Poisson birth process with the mean rate µ0 and distribution bk (x),
according to (30) is given by:
Y
fb (X) = e−µ0
µ0 bk (x).
x∈X
Finally, πk|k−1 (x|x′ ) is the (standard) transitional density, defined in Sec. 2.1.
The multi-object state Xk is not observed directly, but through the observation process, assumed to be conditionally independent given the multi-object
state process, and fully specified by the (multi-object) likelihood function. Many
different observation models have been considered in the literature, such as the
models for intensity measurements versus detector-output measurements4 , extended versus point target measurements, finite resolution measurements, fuzzy
or imprecise measurements, see for details [29], [30], [72], [9], [26].
The detector-output measurement model for a point-size target, referred to
as the standard model [27],[9], will be adopted throughout this tutorial. Recall
from (16) that the standard measurement is modelled by an RFS variable Zk .
The likelihood function ϕk (Zk |Xk ) for this model has been derived in Chapter
12 of [27] as the superposition of a multi-Bernoulli object-detection RFS and a
Poisson clutter (false detection) RFS. Under the assumption that the constituent
RFSs are mutually independent, the multi-object likelihood function can be
expressed by:
ϕk (Z|X) = fc (Z)(1 − pD )|X|
X Y
θ
i:θ(i)>0
pD · gk (zθ(i) |xi )
.
(1 − pD ) · λ c(zθ(i) )
(34)
Explanation: θ : {1, . . . , |X|} → {0, 1, . . . , |Z|} represents a list of all associations of elements from set X to the elements of set Z (here the assumption is
that if xi ∈ X is not detected, then θ(i) = 0; also, a target x ∈ X can generate
at most one measurement z ∈ Z); pD is the probability of detection; fc (Z) is
the multi-object PDF of clutter, which is assumed to be a Poisson RFS with
the mean rate λ and distribution c(z) over Z. According to (30), the clutter
4 An intensity measurement at time k represents a raw sensor measurement, prior to detection thresholding. In tracking literature this is also referred to as track-before-detect approach.
Using intensity measurements (rather than the detector-output measurements) can lead to the
better error performance, but at increased computational load due to the large data flow. An
intensity measurement at time k is a vector, rather than a random finite set. A detailed
explanation with examples of intensity measurement models can be found in [29].
16
multi-object PDF is:
fc (Z) = e−λ
Y
λ c(z).
(35)
z∈Z
Finally, gk (z|x) in (34) is the conventional likelihood function, introduced in
Sec. 2.1.
Given Xk , measurement Zk is assumed to be statistically independent of
Zℓ , where ℓ 6= k. The Bayes-optimal filtering problem can now be cast in
the random finite set framework. Suppose that at time k − 1 the posterior
FISST PDF of the multi-object state, fk−1|k−1 (Xk−1 |Z1:k−1 ) is known. Here
Z1:k−1 ≡ Z1 , . . . , Zk−1 is the sequence of all previous measurements. Then
the predicted and updated posterior multi-object densities can be expressed as
follows [27]:
Z
fk|k−1 (Xk |Z1:k−1 ) =
Πk|k−1 (Xk |X′ )fk−1|k−1 (X′ |Z1:k−1 )δX′ (36)
fk|k (Xk |Z1:k ) =
R
ϕk (Zk |Xk )fk|k−1 (Xk |Z1:k−1 )
,
ϕk (Zk |X)fk|k−1 (X|Z1:k−1 )δX
(37)
respectively.
The recursion (36)-(37) is a non-trivial generalisation of (3)-(4), because the
integrals in (36)-(37) are set integrals and the expressions for Πk|k−1 (Xk |Xk−1 )
and ϕk (Zk |Xk ) are quite involved. Computing the exact multi-object posterior density is numerically intractable and all practical algorithms are based on
approximations. An important feature of the RFS Bayes-optimal filter is that
it performs multi-object filtering, as opposed to target (object) tracking. The
difference is significant: a point estimate from the multi-object filter, at each
time step, is a collection of (unlabelled and unordered) object state estimates;
a multi-target tracker, on the other hand, produces labelled state trajectory
estimates, or tracks.
4.2. Particle method approximations
Implementation of the RFS Bayes-optimal filter using the particle method
has been considered in the past, both assuming the intensity measurement model
(typically without a reference to RFS models, e.g. [31]) and the standard measurement model [32], [33], [28], [34], [35], [36]. Since the filter is defined on the
set of subsets F (X ), it is computationally very demanding and practical only for
a small number of objects. A particle in the state space F (X ) can be expressed
17
as:
(i)
Xk
∅,
(i)
[xk ],
(i)
⊺
= [(x(i)
(xk,2 )⊺ ]⊺ ,
k,1 )
. . .
(i)
(i)
(i) ⊺
[(xk,1 ) (xk,2 )⊺ . . . (xk,νmax )⊺ ]⊺ ,
(i)
if |Xk | = 0
(i)
if |Xk | = 1
(i)
if |Xk | = 2
(38)
(i)
if |Xk | = νmax
where νmax denotes the maximum number of targets (a design parameter). Note
that particle i = 1, 2, . . . , N , for a given cardinality, is represented in (38) by a
vector. While this was done to simplify computer implementation, one should
keep in mind that the multi-object particle essentially represents a random finite
set, that is, any permutation of its elements (objects, targets) results in the
equivalent particle. By ignoring the permutation equivalence, the multi-object
state particle filter can be affected by the mixed labelling problem, see [37]. It
has been noted that mixed labeling is typically resolved after few time steps due
to resampling step in the particle filter [37].
The pseudo-code of the bootstrap-type particle filter which implements a
cycle of the RFS Bayes-optimal filter at time k is given in Alg. 2. The proposed
multi-object particles at time k are constructed based on the transitional density,
the survival probability and the birth distribution of new targets, ignoring the
association events in (33). The average number of newborn targets is µ0 , with
birth distribution bk (x) in line 14 typically designed using the measurements
(i)
from time k − 1. The computation of the multi-object likelihood ϕk (Zk |Xk )
in line 17, defined in (34), requires the evaluation all association hypotheses θ
(i)
between objects in Xk and detections in Zk . The number of these hypotheses
grows exponentially with the number of targets. For example, if the cardinalities
(i)
are: |Xk | = 2 and |Zk | = 3, the number of association hypothesis 5 is 13. This
exponentially growing computational cost is the major limitation of the RFS
Bayes-optimal filter in practice.
Estimation of the multi-object state Xk from the particle approximation of
the posterior fk|k (Xk |Z1:k ) is discussed briefly next. The cardinality distribution
5 In this case θ : {x , x } → {∅, z , z , z }, because targets can be undetected. The list of
1
2
1 2 3
possible assignments is: θ1 : x1 → ∅; x2 → ∅; θ2 : x1 → ∅; x2 → z1 ; θ3 : x1 → ∅; x2 → z2 ;
θ4 : x1 → ∅; x2 → z3 ; θ5 : x1 → z1 ; x2 → ∅; θ6 : x1 → z1 ; x2 → z2 ; θ7 : x1 → z1 ; x2 → z3 ;
θ8 : x1 → z2 ; x2 → ∅; θ9 : x1 → z2 ; x2 → z1 ; θ10 : x1 → z2 ; x2 → z3 ; θ11 : x1 → z3 ; x2 → ∅;
θ12 : x1 → z3 ; x2 → z1 ; θ13 : x1 → z3 ; x2 → z2 .
18
Algorithm 2 Pseudo-code of the RFS bootstrap filter
1: function RFS bootstrap filter
(i)
2:
Input: {Xk−1 }1≤i≤N , Zk
3:
for i = 1, . . . , N do
(i)
4:
X̃k = [ ]
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
(i)
for j = 1, . . . , |Xk−1 | do
u∗ ∼ U[0,1]
if u∗ < pS then
(i)
x∗ ∼ πk|k−1 (x|xk−1,j )
h
⊺
i⊺
(i)
(i)
X̃k = X̃k
x⊺∗
end if
end for
µ ∼ Poisson(µ0 )
for j = 1, . . . , µ do
x∗ ∼ bkh
(x)
i
(i) ⊺
(i)
⊺
X̃k = X̃k
x⊺∗
end for
(i)
(i)
Calculate weight: w̃k = ϕk (Zk |X̃k )
end for
(i)
(i) P
(j)
wk = w̃k / N
j=1 w̃k , for i = 1, . . . , N
for i = 1, . . . , N do
(i)
Select index j i ∈ {1, . . . , N } with probability wk
(i)
(j )
22:
Xk = X̃k i
23:
end for
(i)
24:
Output: {Xk }1≤i≤N
25: end function
can be estimated as:
ρ̂k|k (n|Z1:k ) =
N
1 X h (i) i
δ |Xk |, n ,
N i=1
n = 0, 1, 2, . . . , νmax ,
(39)
where δ[ℓ, n] is the Kronecker delta function which equals 1 if ℓ = n and zero
otherwise. Practically, for each n = 0, 1, . . . , νmax , it is necessary to count
the number of multi-target particles whose cardinality equals n, and then to
divide this count with the total number of particles N . Then one can estimate
the number of targets in Xk , that is n̂k|k , either as the expected value or the
maximum of the posterior ρ̂k|k (n|Z1:k ). Given n̂k|k , the multi-target state can
(i)
(i)
be estimated from the particle system {wk , Xk }1≤i≤N as
i
h
PN
(i) (i)
(i)
w
X
δ
|X
|,
n̂
k
i=1 k
k
k
i ,
h
X̂k = P
(40)
(i)
(i)
N
w
δ
|X
|,
n̂
k
i=1 k
k
which represents the mean of the multi-target particles characterized by cardinality n̂k|k .
19
4.3. Bernoulli particle filter
The Bernoulli filter is the special case of the RFS Bayes-optimal filter, derived by Mahler [27] under the assumption that Xk is a Bernoulli RFS. Recall
that the Bernoulli RFS X can have zero or one element and its PDF f (X)
is completely specified by: (1) the probability of object existence r and (2)
the PDF p(x) on X . Hence, the Bernoulli filter is the Bayes-optimal filter for
joint detection and tracking of a single object, where detection is carried out by
monitoring the posterior probability of object existence.
A detailed tutorial on Bernoulli filters, their formulation for different measurement models, their numerical implementation and various applications can
be found in [29]. Here we present only the prediction and update equations for
the standard measurement model. Suppose the posterior fk−1|k−1 (X|Z1:k−1 ) is
known and specified by the pair rk−1|k−1 , pk−1|k−1 (x) . The prediction equations of the Bernoulli filter are given by:
rk|k−1
=
pk|k−1 (x)
=
pb (1 − rk−1|k−1 ) + pS rk−1|k−1
pb (1 − rk−1|k−1 )bk|k−1 (x)
+
rk|k−1
R
pS rk−1|k−1 πk|k−1 (x|x′ )pk−1|k−1 (x′ )dx′
rk|k−1
(41)
(42)
where pb is the probability of object “birth” from the previous time k − 1 to k
and bk|k−1 (x) is the object birth density. The update equations are as follows:
rk|k
=
pk|k (x)
=
1 − ∆k
rk|k−1
1 − rk|k−1 ∆k
P
k (z|x)
1 − pD + pD z∈Zk gλc(z)
1 − ∆k
(43)
pk|k−1 (x)
where, assuming pD is independent of the state,
!
R
X gk (z|x)pk|k−1 (x)dx
.
∆k = pD 1 −
λc(z)
(44)
(45)
z∈Zk
It can be verified that if rk−1|k−1 = 1, pb = 0, pD = 1 and λ = 0, the Bernoulli
filter equations reduce to the standard Bayes filter equations (3)-(4).
The Bernoulli particle filter propagates through time the scalar value of
(i)
(i)
rk|k and the weighted particle system {wk , xk }1≤i≤N . The pseudo-code of
the Bernoulli bootstrap-type particle filter is given in Alg. 3. It implements
R
the birth density in (42) as bk|k−1 (x) = πk|k−1 (x|x′ )bk−1 (x′ )dx′ , where (see
line 3 in Alg. 3) the birth density at k − 1, i.e. bk−1 (x′ ) is designed using the
20
measurements in Zk−1 . This is done by drawing Nm particles for each z ∈ Zk−1 ,
hence Bk−1 = |Zk−1 | · Nm . The weights of these “new-born” particles are set to
1/Bk−1 . The number of particles from step 5 onwards is N + Bk−1 , and reduces
back to N only by the resampling step (lines 13-16).
Algorithm 3 Pseudo-code of a Bernoulli bootstrap-type particle filter
1: function Bernoulli Particle Filter
(i)
2:
Input: rk−1|k−1 , {xk−1 }1≤i≤N , Zk , Zk−1
(i)
3:
4:
Draw: xk−1 ∼ bk−1 (x) designed based on Zk−1 for i = N + 1, . . . , N + Bk−1
Compute the predicted existence rk|k−1 using (41)
5:
Draw a sample: xk|k−1 ∼ πk|k−1 (xk |xk−1 ) for i = 1, . . . , N + Bk−1
6:
7:
8:
9:
10:
11:
(i)
(i)
Weights: wk|k−1 =
Weights:
(i)
wk|k−1
=
(i)
pS rk−1|k−1
for
N rk|k−1
pb (1−rk−1|k−1 )
Bk−1 rk|k−1
i = 1, . . . , N
⊲ 2nd term in (42)
for i = N + 1, . . . , N + Bk−1
⊲ 1st term in (42)
PN+Bk−1 (i)
(i)
For every z ∈ Zk compute Ik (z) = i=1
wk|k−1 · gk (z|xk|k−1 )
P
Ik (z)
⊲ Eq.(45)
Compute δk ≈ pD 1 − z∈Zk λ c(z)
Compute the updated existence rk|k using (43)
"
#
(i)
)
P gk (z|xk|k−1
(i)
(i)
Weights: w̃k|k ≈ 1 − pD + pD
· wk|k−1 , for i = 1, . . . , N + Bk−1
λ c(z)
z∈Zk
(i)
12:
Normalise weights:
(i)
wk|k
=
w̃k|k
PN +Bk−1
j=1
13:
14:
(j)
w̃k|k
, for i = 1, . . . , N + Bk−1
for i = 1, . . . , N do
(i)
Select index j i ∈ {1, . . . , N + Bk−1 } with probability wk|k
(i)
⊲ (Resampling)
(j )
i
15:
xk = xk|k−1
16:
end for
(i)
17:
Apply MCMC move to particles and output: rk|k , {xk }1≤i≤N
18: end function
4.4. Demonstration: Bearings-only detection and tracking
Let us demonstrate the Bernoulli particle filter in action. The observertarget scenario is the same as described in Sec. 2.4, however, this time we
consider the full observation interval from t = 0 to t = 3000[s]. Recall that the
target exists only for 200[s] ≤ t ≤ 2400[s]. Furthermore, detection probability is
pD = 0.95 and the average number of false bearings-only measurements is λ = 1
per scan. Clutter distribution is uniform, that is c(z) = (2π)−1 [rad−1 ]. The
parameters of the Bernoulli PF were set to: Nm = 2500, pb = 0.01, pS = 0.98.
The remaining parameters were the same as in Sec. 2.4: σw = 0.3o , T = 20[s],
σv = 0.005 [m/s2 ], rmax = 10000[m], N = 5000 particles.
Birth density bk−1 (x) is designed adaptively using Zk−1 , to form a mixture
21
density:
bk−1 (x|Zk−1 ) =
1
|Zk−1 |
X
β(x|z)
(46)
z∈Zk−1
Here β(x|z) is the birth density created using the standard technique for particle
filter initialisation when tracking with bearings-only measurements, explained
already in Sec. 2.4.
Fig. 2 shows the results obtained by averaging over 100 Monte Carlo runs of
the Bernoulli particle filter. Fig. 2.(a) displays the average estimated probability
of existence rk|k , versus time. The true existence is indicated by the dashed blue
line. Figs. 2.(b) and (c) present the RMS errors in estimated target position
(i)
and velocity, respectively. The EAP estimates were computed from {xk }1≤i≤N
only at time steps when rk|k > 0.2. The dashed blue lines in Figs. 2.(b)
and (c) are the copies of the corresponding RMS errors, obtained in Sec. 2.4
(see Fig. 1) using the standard particle filter under ideal conditions (i.e. with
perfect knowledge of target existence and measurements obtained with perfect
detection, that is with pD = 1, λ = 0). Overall, we can observe a remarkably
accurate performance of the Bernoulli particle filter: it is only marginally worse
than the performance under ideal conditions.
5. PHD particle filters
The RFS Bayes-optimal filter propagates the multi-object PDF fk|k (Xk |Z1:k ),
defined on the space F (X ) of finite subsets of the single-object space X . Since
even for a relatively small number of objects it becomes cumbersome to work on
F (X ), several principled approximations of the RFS Bayes-optimal filter have
been proposed in the context of various measurement models. The most popular among them is the probability hypothesis density (PHD) filter, derived by
Mahler for the standard measurement model in [38]. Instead of propagating
the posterior multi-object PDF of Xk over time, the PHD filter propagates its
first-order statistical moment: the intensity function or PHD (see (25)), which is
defined on the space X . As a simple and fast multi-object Bayes (sub-optimal)
filter, whose computational complexity grows (only) linearly with the number
of objects, the PHD filter has quickly become very popular among researchers.
This resulted in numerous practical applications, such as passive radar [39],
sonar [40], computer vision [41], traffic monitoring and road mapping [42, 43],
robotic navigation and mapping [45, 46], cell microscopy [47], to name a few.
The intensity function is in general a very crude approximation of the multiobject PDF. (Recall, the multi-object PDF can be recovered from its intensity
22
Existence probability
1
(a)
0.8
0.6
0.4
average r
0.2
truth
k|k
0
0
500
1000
1500
2000
2500
3000
Time [s]
5
(b)
3000
RMS error in VELOCITY [m/s]
RMS error in POSITION [m]
3500
2500
2000
1500
1000
500
0
0
500
1000
1500
2000
2500
3
2
1
0
0
3000
(c)
4
500
1000
1500
2000
2500
3000
Time [s]
Time [s]
Figure 2: Bernoulli particle filter for joint detection and tracking using bearings-only measurements (pD = 0.95; uniformly distributed clutter with λ = 1 per scan; σw = 0.3o ): (a) average
probability of existence rk|k (dashed line is ideal); (b) RMS error in position; (c) RMS error
in velocity (the dashed lines in (b) and (c) are the copies from Fig. 1, obtained with pD = 1
and λ = 0).
function only for a Poisson RFS). Hence Mahler subsequently introduced the
Cardinalised PHD filter [48], which propagates jointly the intensity function and
the cardinality distribution of the multi-object PDF. The Cardinalised PHD
filter improves both the estimate of the number of objects and the accuracy of
the individual state estimates [49], but comes at an increased complexity. In this
paper we focus only on the PHD particle filter for the standard measurement
model.
5.1. Formulation of the PHD filter
The posterior intensity function (the posterior PHD) at time k, Dk|k (x|Z1:k ),
is abbreviated for simplicity to Dk|k (x). Recall from Sec. 3.1 that the expected
number of objects present at time k can be estimated from Dk|k (x) as νk|k =
R
Dk|k (x) dx, where νk|k ∈ R.
X
Assuming that the PHD function at k−1 is available, the prediction equation
23
of the PHD filter is given by [38]:
Dk|k−1 (x) = γk|k−1 (x) + pS
Z
πk|k−1 (x|x′ ) Dk−1|k−1 (x′ ) dx′
(47)
where γk|k−1 (x) is the PHD of the RFS of object births between time k − 1 and
k. Following the measurement-driven design of the birth density in Sec. 4.3,
R
we model γk|k−1 (x) = pS πk|k−1 (x|x′ ) γk−1 (x′ ) dx′ , where assuming a Poisson
birth RFS, γk−1 (x) = νb bk−1 (x|Zk−1 ). Here νb is the expected number of object
births between time k − 1 and k (a design parameter, typically small, e.g. 0.1)
and bk−1 (x|Zk−1 ) is the birth density (46).
Upon receiving the measurement set Zk at time k, the update step of the
PHD filter is computed according to:
"
#
X
pD gk (z|x)
R
Dk|k (x) = 1 − pD +
Dk|k−1 (x) (48)
κk (z) + pD gk (z|x)Dk|k−1 (x)dx
z∈Z
k
where κk (z) is the PHD of the clutter RFS at time k. Assuming clutter is a
Poisson RFS, see (35), its PHD is κk (z) = λ c(z). Note that in (48) for simplicity
we assume that the probability of detection pD is independent of the state. The
case pD (x) is a straightforward generalisation.
It can be easily verified that in the absence of clutter and target birth, with
Zk = {z}, pS = pD = 1, the PHD filter equations (47)-(48) reduce to the
standard Bayes filter equations (3)-(4).
The PHD filter recursion (47)-(48) is initialised with D0|0 (x). In the absence
of any prior, we can set D0|0 (x) = 0, meaning that initially there are no objects
in the surveillance volume.
5.2. The particle method applied to PHD filtering
In the general nonlinear/non-Gaussian context, the PHD filter cannot be
solved analytically, and is typically implemented using the particle method.
The conceptual framework for an efficient particle PHD filter implementation
has been cast in [28], where the proposal (importance) densities for drawing
particles need to depend on the latest measurement set Zk . How to construct
these importance densities has been a topic of intensive research in the last
decade, see [50], [49], [51], [52]. The method described below mainly follows
[53].
The PHD particle filter propagates through time the particle system Pk ≡
P k (i)
(i)
(i)
(i)
{wk , xk }1≤i≤Nk , which approximates Dk|k (x). Note that wk ≥ 0 and N
i=1 wk =
R
νbk|k ≥ 0 is an estimate of the expected number of objects νk|k = X Dk|k (x) dx.
24
The pseudo-code of a PHD particle filter algorithm is given in Alg. 4. This
code is general enough to be used with any type of a particle filter (e.g. the
auxiliary PF, the exact particle flow filter, the PF with progressive correction,
see Sec. 2.2), naturally including the bootstrap filter. In addition, at each time
b k and, if required, the
step it produces an estimate of the multi-object state X
b
probability of existence for each x ∈ Xk .
Algorithm 4 Pseudo-code of a PHD particle filter
1: function PHD Particle Filter
(i)
(i)
2:
Input: Pk−1 ≡ {wk−1 , xk−1 }1≤i≤Nk−1 , Zk , Zk−1
(i)
3:
Draw: xk−1 ∼ bk−1 (x|Zk−1 ) for i = Nk−1 + 1, . . . , Nk−1 + Bk−1
4:
Set weights: wk−1 = νb /Bk−1 for i = Nk−1 + 1, . . . , Nk−1 + Bk−1 .
5:
Draw a sample: xk|k−1 ∼ πk|k−1 (xk |xk−1 ) for i = 1, . . . , Nk−1 + Bk−1
6:
(i)
(i)
Predict weights: wk|k−1 = pS wk−1 for i = 1, . . . , Nk−1 + Bk−1 .
(i)
(i)
Partition {wk|k−1 , xk|k−1 }1≤i≤Nk−1 +Bk−1 to form clusters Ck|k−1 (z),
(i)
(i)
(i)
7:
bk = ∅
8:
Initialise: Pk = ∅, X
9:
for every z ∈ Zk do
10:
if Ck|k−1 (z) 6= ∅ then
11:
Ck (z) = P F U [Ck|k−1 (z), z]
12:
Pk = Pk ∪ Ck (z)
13:
Compute existence prob. pe (z) for Ck (z)
14:
if pe (z) > η then,
15:
Compute EAP estimate x̂k from Ck (z)
bk = X
b k ∪ {x̂k }
16:
X
17:
end if
18:
end if
19:
end for
20:
for every pair (wk|k−1 , xk|k−1 ) ∈ Ck|k−1 (∅) do
21:
if wk|k−1 > ξ then
22:
Pk = Pk ∪ {((1 − pD )wk|k−1 , xk|k−1 )}
23:
end if
24:
end for
(i)
(i)
bk
25:
Output: Pk ≡ {wk , xk }1≤i≤Nk ; X
26: end function
∀z ∈ Zk ∪ ∅
⊲ See Alg. 5
⊲ η is reporting threshold
⊲ ξ is threshold
In line 3, Bk−1 = Nm · |Zk−1 | particles, for potentially newborn targets,
are drawn from the birth density bk−1 (x|Zk−1 ) defined in (46). These particles, indexed by i = Nk−1 + 1, . . . , Nk−1 + Bk−1 , are given equal weights,
PNk−1 +Bk−1 (i)
(i)
wk−1 = νb /Bk−1 (see line 4), so that their sum is
i=Nk−1 +1 wk−1 = νb .
Implementation of (47) then follows in lines 4-6: both persistent and newborn
target particles are propagated through the transitional density in line 5, while
their weights are multiplied by pS (line 6).
Line 7 in Alg. 4 is very unusual and needs a detailed explanation. Previous
25
approaches to PHD particle filtering [28], [49] treated the term in the square
brackets on the RHS of (48) as a pseudo-likelihood. Then the weights of all
(i)
(i)
predicted particles, obtained after line 6, i.e. {wk|k−1 , xk|k−1 }1≤i≤Nk−1 +Bk−1
were updated directly using (48) as follows:
(i)
X
)
p
g
(z|x
D k
k|k−1
(i)
w(i)
w̃k|k = 1 − pD +
PNk−1 +Bk−1
k|k−1
(i)
(i)
gk (z|xk|k−1 )wk|k−1
z∈Zk κk (z) + pD
i=1
(49)
for i = 1, . . . , Nk−1 + Bk−1 . This would be followed by resampling (using
normalised weights) and possibly particle diversification using for example the
MCMC move step. This approach, referred to as the pseudo-likelihood update
(PLU), will be demonstrated for the sake of comparison. It suffers from two
problems. First, the particles representing the targets that were undetected at
time k would not survive the resampling step. As a result, these targets would
have to be re-initailsed when detected at a future time, which is very inefficient
for applications such as bearings-only filtering (due to the un-observability of
b k , from the
target range). Second, estimation of the multi-target state, i.e. X
particle system after the described update step would be cumbersome. The
standard approach [54], [27, p.623], [9, Sec.9.6] is to first estimate the number
of targets as the nearest integer value corresponding to the sum of the weights
in (49), followed by partitioning of the particles using a data clustering algorithm. Finally, the mean values of the clusters of particles become the EAP
b k . Various clustering algorithms have been compared,
estimates included in X
such as hierarchical clustering, k-means and the EM algorithm. Clustering in
this context is completely ad-hoc and destroys the elegance of the PHD particle
filter.
Alg. 4 avoids the described shortcomings. In line 7 particles are partitioned
in a principled manner using the update equation (48). Note that the RHS
of (48) contains |Zk | + 1 additive terms. For each particle-measurement pair,
an additive term in the sum on the RHS of (48) can be interpreted as the
probability that measurement zj , j = 1, . . . , |Zk | is due to an object in the state
(i)
xk|k−1 , i = 1, . . . , Nk−1 + Bk−1 [30, p.78]:
(i)
Pij =
(i)
pD gk (zj |xk|k−1 ) wk|k−1
.
PNk−1 +Bk−1
(ℓ)
(ℓ)
κ(zj ) + pD ℓ=1
gk (zj |xk|k−1 ) wk|k−1
(50)
The remaining additive term in (48) can be interpreted as the probability that
(i)
an object in the state xk|k−1 has not been detected. We index this case with
26
j = 0, i.e. its probability is [30, p.78]:
(i)
Pi0 = (1 − pD )wk|k−1 .
(51)
It can be easily verified that 0 ≤ Pij ≤ 1 for all j = 0, . . . , |Zk |.
(i)
(i)
Partitioning of Pk|k−1 ≡ {wk|k−1 , xk|k−1 }1≤i≤Nk−1 +Bk−1 now proceeds based
on (50) and (51). For a particle with index i in Pk|k−1 , one can compute the
probability distribution over the elements of Zk , plus the empty set, as follows
Pij
pi (j) = P|Z |
k
ℓ=0
Piℓ
, j = 0, 1, . . . , |Zk |.
(52)
Partitioning of particles in Pk|k−1 is next carried out in a probabilistic manner as
follows. For each i = 1, . . . , Nk−1 + Bk−1 we select an index j i ∈ {0, 1, . . . , |Zk |}
(i)
(i)
with probability pi (j). The weighted particle (wk|k−1 , xk|k−1 ) from Pk|k−1 is
then assigned to cluster Ck+1|k (zj i ). When this procedure is complete we end
up with |Zk | + 1 clusters Ck+1|k (z), ∀z ∈ Zk ∪ ∅. Note that some of the clusters
may end up being empty (for example, if z ∈ Zk is a false detection, Ck|k−1 (z)
is likely to be empty).
Processing steps from line 9 to line 24 in Alg. 4 are self-explanatory, except
for line 11. Let us denote a component of Dk|k−1 (x) which is approximated by
j
(x). According to (48), the Bayes
the particles in cluster Ck|k−1 (zj ) by Dk|k−1
j
update of Dk|k−1
(x) using the assigned measurement zj ∈ Zk is as follows:
Dkj (x) =
j
(x)
pD gk (zj |x) Dk|k−1
R
j
κ(zj ) + pD gk (zj |x) Dk|k−1 (x)dx
(53)
A slight modification of any standard particle filter can implement (53). This
step is carried out in line 11 where PFU stands for particle-filter update (to
be explained later). The output of the PFU based on z ∈ Zk is a cluster
(ℓ)
(ℓ)
Ck (z) = {wk , xk }1≤ℓ≤L , which is added to the particle system Pk in line 12.
The sum of the updated weights in cluster Ck (z) is less than or equal to 1 and
represents the probability of existence pe (i.e. the probability that a target,
whose posterior PDF is approximated by Ck (z), exists). The probability of
existence is computed in line 13, and then in line 14 compared to the reporting
threshold η. If pe is above this threshold, an EAP estimate is computed using
cluster Ck (z), see line 15, to be subsequently included in the multi-object state
estimate X̂k , see line 16.
The particles of cluster Ck|k−1 (∅) have not been assigned any measurement
for the update, and therefore are treated differently (see the loop between lines
20 and 24). Recall that every measurement induces the so-called newborn target
27
particles. If unchecked, this could potentially result in an ever-growing number
of particles over time. The if-then clause in line 21 of Alg. 4 is introduced to prevent that from happening. The particles in cluster Ck|k−1 (∅) whose weights are
smaller than threshold ξ are eliminated and therefore cannot propagate further
in time (e.g. those initially induced on false detections, or those corresponding
to targets that ceased to exist). The particle elimination threshold ξ (line 21)
must be chosen so that the particles on undetected but currently present (existing) objects are not eliminated. This is particularly important in applications
such as bearings-only tracking, where target range is initially unobservable.
The bootstrap-type PFU (line 11 in Alg. 4) is described by pseudo-code in
Alg. 5. A PFU routine based on progressive correction is discussed in [53]. Let
(m)
(m)
the cluster Ck|k−1 (z) consist of M weighted particles, {wk|k−1 , xk|k−1 }1≤m≤M .
(m)
According to (53), the weight of a predicted particle xk|k−1 from Ck|k−1 (z) is
updated as:
(m)
(m)
pD gk (z|xk|k−1 )wk|k−1
(m)
(54)
w̃k =
P
(j)
(j)
κ(z) + pD M
j=1 gk (z|xk|k−1 )wk|k−1
This step is carried out in line 4 of Alg. 5, followed by resampling (using
normalised weights), see the loop between lines 7 and 11. While the number
of particles M in the input cluster Ck|k−1 (z) varies from cluster to cluster,
resampling (line 7) is always carried out L times, where L is a user-defined
parameter. Note from line 10 (Alg. 5) that the probability of cluster existence
pe remains unchanged by the PFU routine.
Algorithm 5 Bootstrap-type PFU (line 11 of Alg. 4)
1: function Bootstrap-type PFU
(m)
(m)
2:
Input: Ck|k−1 (z) = {wk|k−1 , xk|k−1 }M
m=1 , z
3:
for m = 1, . . . , M do
(m)
4:
Compute w̃k according to (54)
5:
end for
P
(m)
6:
Compute pe = M
m=1 w̃k
7:
for ℓ = 1, . . . , L do
(c)
8:
Select c ∈ {1, . . . , M } with prob. w̃k /pe
9:
(ℓ)
(c)
xk ← xk|k−1
(ℓ)
10:
wk = pe /L
11:
end for
(ℓ)
(ℓ))
12:
Output: Apply MCMC move and output Ck = {wk , xk }L
ℓ=1
13: end function
28
⊲ Resampling loop
5.3. Calibration of tracking algorithms
Tracking algorithms are based on mathematical models, in particular the dynamic model (for target birth/surival and motion) and the sensor measurement
model. These models typically include many parameters, such as the target
birth rate νb , survival probability pS , the false alarm rate λ, the probability
of detection pD , measurement noise variance, but also sensor biases, various
factors such as the propagation losses, receiver gains, etc. Calibration of tracking algorithms, through estimation of their model parameters, is an important
prerequisite for their operational deployment.
In accordance with Sec. 2.3, let a random vector θ ∈ Θ ⊂ Rnθ represent
the static parameter vector of interest for estimation/calibration. Again we
will indicate this in notation as Πk|k−1 (Xk |Xk−1 , θ), for the FISST transitional
density and ϕk (Zk |Xk , θ), for the FISST likelihood function. The problem is to
estimate the posterior density p(θ|Z1:k ) ∝ ̺(Z1:k |θ) p(θ), given the prior p(θ).
Similarly to the arguments presented in Sec. 2.3, the complicating factor is that
the likelihood function ̺(Z1:k |θ) cannot be expressed in closed-from. However,
note that, similar to (8), we can apply a decomposition:
̺(Z1:k |θ) = ̺(Z|θ)
k
Y
̺(Zt |Z1:t−1 , θ)
t=2
where
̺(Zt |Z1:t−1 , θ) =
Z
ϕk (Zk |Xk , θ)f (Xk |Z1:k−1 , θ)δXk
The key observation here is that the conditional likelihood ̺(Zt |Z1:t−1 , θ) can
be computed as a by-product of the PHD filter, see [38, Eq.(116)]:
Z
̺(Zk |Z1:k−1 , θ) ∝ exp −pD (θ)
Dk|k−1 (x|Z1:k−1 , θ)dx ×
Z
Y
κ(z|θ) + pD (θ) gk (z|x, θ)Dk|k−1 (x|Z1:k−1 , θ)dx
z∈Zk
(55)
Using the PHD particle filter, (55) can be estimated from the particle system
(i)
(i)
Pk|k−1 ≡ {wk|k−1 , xk|k−1 }1≤i≤Nk−1 +Bk−1 , available after line 6 in Alg. 4, as
29
follows:
Nk−1 +Bk−1
X
(i)
b k |Z1:k−1 , θ) ∝ exp −pD (θ)
wk|k−1 ×
ℓ(Z
i=1
Nk−1 +Bk−1
Y
X
(i)
(i)
κ(z|θ) + pD (θ)
gk (z|xk|k−1 , θ)wk|k−1
i=1
z∈Zk
(56)
Now one can apply any of the standard Monte Carlo parameter estimation
algorithms (e.g. MCMC) to estimate the posterior p(θ|Z1:K ). Note that in
every MCMC iteration, for the proposed value of θ, it is necessary to run the
PHD particle filter in order to compute (56). The factor of proportionality in
(56) cancels out and hence is irrelevant.
The described method has been demonstrated in the context of sensor bias
calibration, see [55]. However, its impact is much broader. Suppose the calibration parameter is a stochastic dynamic process whose evolution is independent of Xk . Then based on the factorization of the posterior PHD function:
Dk|k (xk , θk |Z1:k ) = Dk|k (xk |θ k , Z1:k ) pk|k (θk |Z1:k ), one can apply a hierarchical particle method as follows: the standard particle filter on the Θ space, and
the PHD particle filter on the X |Θ space. This approach was demonstrated in
the context of joint tracking and time-varying bias estimation in [56]. Moreover, if Dk|k (xk |θk , Z1:k ) is analytically tractable (e.g. linear/Gaussian target
tracking), then one can effectively apply a Rao-Blackwellised formulation. This
is the essence of the algorithms proposed for: tracking groups of targets [57],
tracking an extended target [58] and simultaneous localisation and mapping
(SLAM) [44, 46, 59].
5.4. Demonstration: Bearings-only multi-object filtering
The performance of the described PHD particle filter is demonstrated using
the simulated scenario taken from [26], which consist of four targets and lasts
3000[s]. The geometry is shown in Fig. 3.(a). The circles on each trajectory
indicate the starting points. One target is present at the beginning, with another
three arriving during the first 400[s]. Three targets are terminating in the final
900[s]. The bearings of all four targets cross one another in the middle of the
observation interval. We assume an ideal sensor with infinite resolution.
In order to measure the statistical error of multi-object filtering at each time
step we need a distance metric between two finite sets of objects: the ground
b k . The optimal subpattern assignment (OSPA)
truth Xk and its estimate X
30
distance [60] has become a widespread metric for this purpose as it captures
both the difference in the cardinality of the two finite sets and the positional
error of their respective elements. The OSPA distance has two parameters: (1)
the cut-off value, which we adopt as c = 5[km]; (2) the order parameter of the
Euclidian base distance (in position), which we adopt as p = 2.
Fig. 3.(b) shows the OSPA error averaged over 500 Monte Carlo runs, for
two PHD particle filters: the one described by Algs. 4 and 5, and the alternative,
based on the pseudo-likelihood update step [49]. The parameters used in simulations were: pD = 0.95; Poisson clutter with λ = 1 and c(z) = (2π)−1 [rad−1 ];
Nm = 2500, pb = 0.01, pS = 0.98; σw = 1o , T = 20[s], σv = 0.005 [m/s2 ],
rmax = 10000[m], and N = 5000 particles. Clearly, the PHD particle filter described by Algs. 4 and 5, which includes the partitioning of particles, performs
much better than the alternative based on PLU.
6. Labelled RFS Bayes tracking filters
Since the introduction of RFS Bayes filters (optimal and suboptimal), there
have been numerous attempts to apply them to multi-target tracking problems,
e.g. [54], [61], [41], [62]. These heuristic methods typically label the targets and
apply some form of data association to keep track of the labels as time progresses. Mahler’s view [27] was that the most natural and therefore convenient
RFS for multi-target tracking is the multi-Bernoulli RFS, introduced in Sec.
3.2. His initial formulation led to the development of an approximation of the
RFS Bayes-optimal filter referred to as the CBMeMBer filter [63]. Assuming
the multi-object PDF fk|k (Xk |Z1:k ) is multi-Bernoulli, the CBMeMBer filter
j
propagates the pairs rk|k
, pjk|k (x) , j = 1, . . . , Mk over time via the prediction
and update steps. While the prediction step (36) for the CBMeMBer filter can
be derived in closed-form exactly, the update step (37) formulation is based on
a (fairly un-intuitive) approximation.
The breakthrough was made in [64] with the formulation of labelled random
finite sets. While preserving the mathematical rigor of the RFS framework, the
labelled RFS concept allows the assignment of a distinct label to each object
(target) that appears in the surveillance volume during the observation time.
The history of the state evolution for each object (i.e. its trajectory) can then be
naturally identified. More importantly, this framework allows the formulation
of an exact closed-from Bayes-optimal multi-target tracking filter [65]. Related
relevant recent papers include [66, 67].
31
(a)
7000
Observer
y [m]
5000
3000
1000
−1000
−3000
−5000
0
5000
10000
x [m]
5000
Algorithm 4 & 5
PLU approach
Mean OSPA Error [m]
4000
3000
2000
1000
0
0
500
1000
1500
T i m e [s]
2000
2500
3000
Figure 3: Bearings-only multi-target filtering results: (a) the simulated scenario (observer and
four targets, circles indicating the starting points); (b) The mean OSPA error (which captures
both the cardinality error and the positional error) for two PHD particle filters.
6.1. Labelled RFS
Suppose the elements of an RFS X on X are uniquely tagged by an (unobserved) label drawn from a discrete countable space of labels L. A labelled RFS,
denoted X, is an RFS on the product space X × L, such that each realisation
of X has distinct labels. We express distinctness of labels mathematically by
introducing the distinct label indicator function ∆(X). Specifically, if we assume
that X = {(x1 , ℓ1 ), · · · , (xn , ℓn )}, then the distinct label indicator is defined as
∆(X) = δ[ |X|, |LX |],
where LX = {ℓ1 , · · · , ℓn } is the set of unique labels in X and δ[a, b] is the
Kroneker delta function, defined earlier. By definition, a realisation of a labelled
RFS X always satisfies ∆(X) = 1. Labelled RFSs include, for example, the
32
labelled Poisson RFS, and the labelled multi-Bernoulli RFS, see [64]. In order to
distinguish the distributions and statistics of labelled versus unlabelled random
variables (both vectors and finite sets), functions of labelled random variables
will be denoted with a tilde sign. For example, a PDF on the space X × L is
denoted p̃, while a PDF on F (X × L) is denoted f˜.
For an analytical formulation of the RFS Bayes optimal tracking filter, of
particular interest is a general class of labelled RFS, referred to as the generalised
labelled multi-Bernoulli (GLMB) RFS [64]. The PDF of a GLMB RFS X, with
single object state space X and discrete-label space L is given by:
X
Y
p̃j (x, ℓ)
(57)
f˜(X) = ∆(X)
φj (LX )
j∈J
(x,ℓ)∈X
where J is an arbitrary index set, and φj and p̃j satisfy
Z
XX
j
p̃j (x, ℓ)dx = 1, ∀ℓ ∈ L.
φ (L) = 1 and
x∈X
L⊆L j∈J
This RFS has a remarkable property that its PDF is a conjugate prior for the
standard multi-object likelihood function (34) and is closed under the ChapmanKolmogorov prediction with the standard multi-object transitional model.
The GLMB is a very general and flexible RFS model, but its application
to multi-target tracking problems is more easily demonstrated by considering
a more specific type of GLMB RFS, which was also proposed in [64]. A δgeneralised labelled multi-Bernoulli (δ-GLMB) RFS with state space X and
label space L, is a GLMB RFS with density of the same form as (57), with the
following substitutions
J = F (L) × Ξ,
j
I,ξ
φ (L) = φ
(58)
I,ξ
(L) = φ
δ [I, L] ,
p̃j = p̃I,ξ = p̃ξ ,
(59)
(60)
where δ[X, Y ] is a generalisation of the Kroneker delta for sets (i.e. δ[X, Y ] = 1
if X = Y and zero otherwise), and Ξ is a discrete space. Thus, a δ-GLMB RFS
is distributed according to
Y
X
p̃ξ (x, ℓ) .
(61)
φI,ξ δ [I, LX ]
f˜ (X) = ∆ (X)
(I,ξ)∈F (L)×Ξ
(x,ℓ)∈X
In a δ-GLMB, the sum is taken over the Cartesian product between the
space of finite subsets of L, and the discrete space Ξ. In tracking applications,
each subset of L represents a set of target labels, while Ξ represents the space of
33
measurement-to-label association histories. Hence, a particular element (I, ξ) ∈
F (L) × Ξ can be considered to be the hypothesis that the set of currently
existing targets are those with labels I and association history ξ. The weight
φI,ξ is the probability of this hypothesis, and p̃ξ (x, ℓ) is the PDF of the target
with label ℓ under the association history ξ.
Like the GLMB, the δ-GLMB has also been shown to be closed under the
standard multi-object prediction, and a conjugate prior with respect to the
standard multi-object measurement likelihood [64]. This makes the δ-GLMB
a highly useful tool for deriving analytical Bayes recursions for multi-object
tracking problems.
6.2. Labelled multi-object transition and likelihood models
The labelled version of the likelihood model ϕk (Z|X) essentially has the same
form as (34), with only a few notes of caution. First, the probability of detection
is assumed to depend on the state of the target and its label, and therefore
pD in (34) should be replaced with p̃D (x, ℓ). Second, a target-to-measurement
association is defined as a mapping θ : LX → {0, 1, . . . , |Z|}. Finally, in the
context of labelled RFSs, we need to substitute xi and X in (34) with (xi , ℓi )
and X, respectively.
The labelled version of the standard multi-target dynamic model is more
involved. Let X be the labelled RFS of objects at the current time with label
space L. A particular object (x, ℓ) ∈ X has probability p̃S (x, ℓ) of surviving
to the next time with state (x+ , ℓ+ ) and probability density π (x+ |x, ℓ) δ [ℓ, ℓ+ ]
(where π (x+ |x, ℓ) is the single target transition kernel), and probability, 1 −
p̃S (x, ℓ) of being terminated. Thus, the set S of surviving objects at the next
time is distributed according to
Y
Φ̃ (S; x, ℓ) ,
(62)
Π̃S (S|X) = ∆ (S) ∆ (X) 1LX (LS )
(x,ℓ)∈X
where 1A (B) is a generalisation of the indicator function for sets, such that
1A (B) = 1 if B ⊆ A and zero otherwise, and
p̃ (x, ℓ) π(x |x, ℓ), if ℓ ∈ L
S
+
S
(63)
Φ̃ (S; x, ℓ) =
1 − p̃S (x, ℓ) ,
if ℓ 6∈ LS
Let B be the labelled RFS of newborn objects with label space B, where
L ∩ B = ∅. Since the births have distinct labels, and assuming that their states
are independent, we model B as a labelled multi-Bernoulli (LMB) RFS, which
34
is distributed according to
f˜b (B) = ∆ (B) φb (LB )
Y
b̃ (x, ℓ) ,
(64)
(x,ℓ)∈B
where b̃ (x, ℓ) is the single object birth density corresponding to label ℓ, and
φb (L) is the birth weight defined by
Y
Y
(1 − rb (ℓ))
1B (ℓ) rb (ℓ) ,
(65)
φb (L) =
ℓ∈L
ℓ∈B\L
in which rb (ℓ) is the existence probability of the newborn object with label ℓ.
The overall prediction of the multi-object state at the next time step is the
union of survivals and new births, i.e. X+ |X = S ∪ B. The label spaces L and
B are disjoint, and the states of newborn objects are independent of surviving
objects, hence S and B are independent. It can be shown [64] that the multiobject transition can be expressed as a product of the transition density for
surviving objects and the density of newborn objects:
Π̃ (X+ |X) = Π̃S (X+ ∩ (X × L) |X) · f˜b (X+ \ (X × L)) .
(66)
Both the GLMB and δ-GLMB families are closed under the Chapman-Kolmogorov
prediction with this transition [64].
6.3. δ-GLMB particle filter
Suppose that the posterior multi-object density at time k − 1 is a δ-GLMB
with the label space L, given by (61), i.e.
Y
X
p̃ξk−1 (x, ℓ)
(67)
φI,ξ
f˜k−1 (X) = ∆(X)
k−1 δ[I, LX ]
(x,ℓ)∈X
(I,ξ)∈F (L)×Ξ
in which the distribution p̃ξk−1 (x, ℓ) of the target with label ℓ and measurement
association history ξ is represented by a weighted set of particles
(i),ξ
(i),ξ
{wk−1 (ℓ), xk−1 (ℓ)}1≤i≤N ξ .
ℓ
The particle weights are normalised, i.e.
ξ
p̃ξ (x, ℓ) ≈
Nℓ
X
i=1
P
Nℓξ
i=1
(i),ξ
wk−1 (ℓ) = 1 and we can write:
(i),ξ
(i),ξ
wk−1 (ℓ) δ x − xk−1 (ℓ) .
(68)
The goal of the δ-GLMB particle filter is to recursively compute the posterior
δ-GLMB density, conditioned on new sets of measurements as they are received.
35
As usual this is done by first applying the Chapman-Kolmogorov equation with
the multi-object transition model, which yields the predicted δ-GLMB density
at the next observation time k. Bayes rule is then applied to this prediction
using the multi-object likelihood model and the received measurement set. The
result is the posterior δ-GLMB density at time k. The prediction and update
steps are described next.
Prediction
Let us assume that within the LMB birth model (64), the single-object density b̃(x, ℓ) corresponding to the target with label ℓ is approximated by particles
(i)
(i)
{wb (ℓ), xb (ℓ)}1≤i≤Nℓb . Under the labelled multi-object transition model defined in (66), the predicted multi-object density at time k is a δ-GLMB with
the label space L+ = L ∪ B, given by [64]
Y ξ
X
I ,ξ
(69)
p̃+ (x, ℓ),
φ++ δ[I+ , LX ]
f˜k|k−1 (X) = ∆(X)
(I+ ,ξ)∈F (L+ )×Ξ
(x,ℓ)∈X
where, assuming the bootstrap-type particle filter with the importance density
equal the (single-target) transitional density,
I ,ξ
φ++ (L) = φb (I+ ∩ B) · φξS (I+ ∩ L),
P ξ
Nℓ wS(i),ξ (ℓ) δ x − x(i),ξ
(ℓ)
, if ℓ ∈ L
S
i=1
p̃ξ+ (x, ℓ) ≈ PN b (i)
(i)
ℓ
if ℓ ∈ B
i=1 wb (ℓ) δ x − xb (ℓ) ,
(i),ξ
(i),ξ
xS (ℓ) ∼ πk|k−1 x|xk−1 (ℓ) , i = 1, . . . , Nℓξ
(i),ξ
(i),ξ
(i),ξ
wS (ℓ) ∝ wk−1 (ℓ) p̃S xk−1 (ℓ), ℓ , i = 1, . . . , Nℓξ
Nℓξ
ηSξ (ℓ) =
X
i=1
ξ
φS (M ) =
(i),ξ
(i),ξ
wk−1 (ℓ) p̃S xk−1 (ℓ), ℓ ,
Y
ℓ∈M
ξ
!
ηS (ℓ)
X
1I (M ) φξk−1 (I)
I⊆L
(70)
(71)
(72)
(73)
(74)
Y
ℓ∈I\M
1 − ηSξ (ℓ) .
(75)
In principle, the δ-GLMB prediction involves generating a new predicted
component (hypothesis) for every combination of target birth, death and survival, for each component in the current posterior density. In practice however,
doing so is usually infeasible because the number of combinations becomes prohibitively large. To improve the efficiency, approximations can be made that
avoid computing unlikely components. One approach is to use a “shortest paths”
algorithm to generate the most likely combinations of targets.
36
This can be carried out separately for births and survivals, using the graph
structures shown in Fig. 4. Each row in the graphs in Fig. 4 corresponds to
a target label: the left-hand column represents target survival/birth, and the
right-hand column represents target death/non-birth. The goal is to maximise
the predicted component weights, which are expressed in (70), (75) and (65) as
products of single-target weights. Since the shortest paths algorithm is based
on minimising a set of additive path costs, negative logarithms are used to
change the problem from maximum-product to minimum-sum form. Paths are
generated from top-to-bottom, and only those targets corresponding to rows in
which the left-hand column was visited are included in the prediction.
start
start
ξ
ξ
-log[1-ηs(l 1 )]
ξ
-log[1-ηs(l 2)]
-log[ηs(l1 )]
ξ
-log[ηs(l 2 )]
ξ
-log[1-ηs(l n )]
ξ
-log[ηs(ln )]
-log[rb(l 1)]
-log[1-rb (l 1)]
-log[r b(l 2)]
-log[1-r (l 2)]
-log[rb(l m)]
-log[1-rb(l m)]
b
end
end
(a)
(b)
Figure 4: Graph structures for δ-GLMB prediction: (a) survival prediction graph; (b) birth
prediction graph. Note that the labels in (a) and (b) are disjoint sets.
Pseudo-code of the prediction step of the δ-GLMB particle filter is given in
Alg. 6. The input to this function in line 2 represents f˜k−1 (X) of (67), with
Lhk−1 being the set of target labels according to hypothesis h = 1, . . . , H. The
input also includes the number of predicted components to be generated for each
component in the current density, that is K h for h = 1, . . . , H. The input in
line 3 represents f˜b (B) of (64). Note that the ShortestPaths function returns
a collection of pairs Lj , cj 1≤j≤K , where each Lj is a list of surviving target
labels, and cj is the corresponding path cost (the negative logarithm of the
37
weight).
Algorithm 6 The prediction step of the δ-GLMB particle filter
1: function δ-GLMB
particle filter: Prediction step
n
o
(i),h
(i),h
wk−1 (ℓ), xk−1 (ℓ)
, Kh
2:
Input: Lh
,
, φh
k−1
k−1
h
1≤i≤Nℓ
ℓ∈Lh
k−1
1≤h≤Hk−1
n
o
(i)
(i)
3:
Input: rb (ℓ), wb (ℓ), xb (ℓ)
,
K
b
b
1≤i≤Nℓ
4:
5:
6:
7:
8:
9:
10:
11:
12:
1≤j≤Kb
m=1
for h = 1, . . . , Hk−1 do
h
Construct graph
n
o GS according to Figure 4.(a)
h,j
h
h
= ShortestPaths Lh
Lh,j
,
c
S
S
k−1 , GS , K
h
1≤j≤K
13:
14:
15:
16:
17:
18:
19:
20:
21:
ℓ∈B
Construct
ograph Gb according to Figure 4.(b)
n
= ShortestPaths (B, Gb , Kb )
Ljb , cjb
for j = 1, . . . , K h do
for β = 1, . . . , Kb do
h,j
β
Lm
+ = Lb ∪ LS
h,j
m
h
φ+ = φk−1 exp −cβ
b exp −cS
n
o
(i),m
(i),m
Compute w+
(ℓ), x+
(ℓ)
m (ℓ)
1≤i≤N+
, ∀ℓ ∈ Lm
+
m=m+1
end for
end for
end for
H+ = m − 1
n o
Normalise weights φj+
1≤j≤H+
(
)
n
o
(i),h
(i),h
h,
w
(ℓ),
x
(ℓ)
Output: Lh
,
φ
+
+
h (ℓ)
+ +
1≤i≤N+
ℓ∈Lh
+
22: end function
⊲ Eqn (71)
1≤h≤H+
Update
The prior multi-object density at time k − 1 is a δ-GLMB given by (69)
and represented by line 21 of Alg. 6. Next we wish to compute the posterior
multi-object density f˜k|k (X), conditioned on a set of received measurements Zk .
Let Θ be the set of all mappings of target labels to measurement indices,
θ : L → {0, 1, ..., |Z|}, such that [θ (i) = θ (j) > 0] ⇒ [i = j]. Note that a given
θ can map multiple labels to 0, meaning that there can be multiple misdetected
targets. However, only one label can be mapped to each non-zero measurement
index, which means that each measurement can only be assigned to one target.
Under the standard multi-object observation likelihood ϕk (Z|X) (see the first
paragraph of Sec. 6.2), the posterior multi-object density is a δ-GLMB given
38
by [64]
f˜k|k (X) = ∆(X)
X
X
I ,ξ,θ
φZ+
δ[I+ , LX ]
Y
p̃ξ,θ (x, ℓ|Zk ) ,
(x,ℓ)∈X
(I+ ,ξ)∈F (L+ )×Ξ θ∈Θ(LX )
(76)
where
I ,ξ
φ++ (L)
I ,ξ,θ
φZ+
(L) =
P
P
Q
ℓ∈I+
ξ,θ
ηZ
(ℓ)
ω J,α (J)
ℓ∈J
(J,α)∈F (L+ )×Ξ θ∈Θ(J)
Nℓξ
p̃ξ,θ (x, ℓ|Z) ≈
X
(i),ξ
w+ (ℓ)ψ̃Z
ξ
ξ,θ
ηZ
(ℓ) =
(i),ξ
x+ (ℓ), ℓ; θ
ξ,θ
ηZ
(ℓ)
i=1
Nℓ
X
(i),ξ
w+
i=1
Q
α,θ
ηZ
(ℓ)
,
(i),ξ
δ x − x+ (ℓ) ,
(i),ξ
(ℓ)ψ̃Z x+ (ℓ), ℓ; θ ,
p̃D (x,ℓ) gk (zθ(ℓ) |x,ℓ) , θ (ℓ) > 0
λc(zθ(ℓ) )
ψ̃Z (x, ℓ; θ) =
.
1 − p̃D (x, ℓ) ,
θ (ℓ) = 0
(77)
(78)
(79)
(80)
In principle, the update involves generating a new δ-GLMB component for
all possible associations of measurements to targets. This is not usually feasible
in practice, since the number of associations increases combinatorially with the
number of targets and measurements. An effective method of reducing the number of posterior δ-GLMB components is to use a ranked assignment algorithm
to generate a requested number of components with highest weights. For a set
of targets with labels {ℓ1 , . . . , ℓn }, and a set of measurements Z = {z1 , . . . , zm },
this can be achieved by constructing the following cost matrix,
ϑξ (ℓ1 , z1 ) · · · ϑξ (ℓ1 , zm ) ϑξ (ℓ1 , ∅) · · ·
0
..
..
..
..
..
..
, (81)
− log
.
.
.
.
.
.
ϑξ (ℓn , z1 ) · · · ϑξ (ℓn , zm )
0
· · · ϑξ (ℓn , ∅)
where
ξ
ϑξ (ℓ, z) =
(i),ξ
Nℓ
w+
X
i=1
ξ
ϑξ (ℓ, ∅) =
Nℓ
X
i=1
(i),ξ
w+
(i),ξ
(i),ξ
(ℓ) p̃D x+ , ℓ gk z|x+ , ℓ
λc (z)
(i),ξ
(ℓ) 1 − p̃D x+ , ℓ
39
(82)
(83)
In (81), each row represents a target, the first m columns represent detections, and the last n columns represent misdetections. Note that if measurement
gating is carried out beforehand, many of the detection terms in the first m
columns will be zero, indicating that the corresponding measurement-to-target
association is infeasible. The matrix (81) is processed using a ranked assignment algorithm. For example, Murty’s algorithm [68], yields a sorted list of
the cheapest one-to-one assignments of rows to columns, in increasing order of
cost. Note that using this formulation, each solution must assign every row to
a column, but there may be columns that are not assigned to a row. In generating the list of ranked assignments, whenever a row is assigned to a column
index greater than m by Murty’s algorithm, the assignment is reported as 0, to
indicate that the target was misdetected. Each element in the list is then used
to construct a component in the posterior δ-GLMB density.
Pseudo-code for the δ-GLMB particle filter update step is given in Alg. 7.
Similarly to the prediction, the input includes the components of the prior δGLMB, and the number of posterior components M h for h = 1, . . . , H+ to
generate for each prior component.
Estimation of labelled target states
The final step in the filtering recursion is to estimate the labelled target
states based on the posterior δ-GLMB density. One method is to approximate
the posterior δ-GLMB in the form of a labelled multi-Bernoulli PDF, where each
target label has an associated existence probability. Assuming the posterior δm
k
GLMB density is specified by line 24 in Alg. 7, for each ℓ ∈ ∪H
m=1 Lk the
existence probability is computed as
rk,ℓ =
Hk
X
(ℓ).
φm
k 1Lm
k
(84)
m=1
The corresponding spatial PDF pk,ℓ (x) is approximated by a weighted particle
set obtained as a union:
∅,
if ℓ 6∈ Lm
k
n
o
k
∪H
(i),m
(i),m
m=1
m
φm
·
w
(ℓ),
x
(ℓ)
,
if
ℓ
∈
L
k
k
k
k
m
1≤i≤Nℓ
A threshold can then be applied to rk,ℓ determine which tracks are to be reported.
A simpler alternative method, which we use here, is to first compute the
40
Algorithm 7 The update step of the δ-GLMB particle filter
1: function δ-GLMB
Filter Update
Particle
(
)
n
o
(i),h
(i),h
h,
w
(ℓ),
x
(ℓ)
2:
Input: Zk , Lh
,
φ
+
+
h (ℓ)
+ +
1≤i≤N+
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
, Mh
ℓ∈Lh
+
m=1
for h = 1, . . . , H+ do
h according to (81)
Compute the cost matrix CZ
h,j h,j
h
h
= RankedAssignments Lh
θ ,c
+ , CZ , M
1≤j≤M h
1≤h≤H+
for j = 1, . . . , M h do
h
Lm
k = L+
h exp −ch,j
=
φ
φm
+
k
m
for ℓ ∈ Lk do
(i),m
(i),h
h (ℓ)
(ℓ), i = 1, . . . , N+
(i),h
(i),h
h (ℓ)
= w+ (ℓ) ψ̃Z (x+ (ℓ), ℓ; θ h,j ), i = 1, . . . , N+
h
PN+
(ℓ) (j),m
(i),m
= w̃k
(ℓ) j=1 w̃k
(ℓ)
⊲ (Normalisation)
⊲ (Resampling)
for i = 1, . . . , Nℓh do
(i),m
i
h
Select index j ∈ {1, . . . , N+ (ℓ)} with probability wk
(ℓ)
(i),m
(j i ),m
xk
(ℓ) = x̃k
(ℓ)
x̃k
(ℓ) = x+
(i),m
w̃k
(ℓ)
(i),m
wk
(ℓ)
end for
end for
m=m+1
end for
end for
Hk = m − 1
Normalise weights φm
k 1≤m≤Hk
n
o
(i),m
(i),m
wk
(ℓ), xk
(ℓ)
Output: Lm
,
, φm
k
k
m
1≤i≤Nℓ
ℓ∈Lm
k
25: end function
1≤m≤Hk
posterior cardinality distribution of the δ-GLMB as follows:
ρ(n) =
Hk
X
m
φm
k · δ[n, |Lk |],
n = 0, 1, 2, . . .
m=1
The maximum a posteriori estimate of the target cardinality is then
n̂ = arg max [ρ(n)] .
n
The next step is to select a component m∗ as the highest weighted δ-GLMB
component representing cardinality n̂. The set of reported target states consist
of the labels and weighted particle sets in the component m∗ .
41
6.4. Demonstration: Bearings-only multi-object tracking
The performance of the described δ-GLMB particle filter is demonstrated by
application to bearings-only multi-target tracking. The scenario geometry and
parameters are the same as those used in Sec. 5.4. The δ-GLMB filter retains
the top 100 hypotheses after each scan, and uses 5000 particles per target.
The number of hypotheses generated in the prediction is 1000, and the number
generated in the update is 4000. A single run output tracks of the δ-GLMB
particle filter are shown in Fig. 5. The ground truth tracks are shown by thick
solid lines (compare with Fig. 3.a), while the estimated tracks are represented
by the same coloured thin lines. Because all four targets in the scenario are
moving from left to right, the accuracy of estimated tracks also improves from
left to right.
8000
Oberver
6000
y [m]
4000
2000
0
-2000
-5000
0
5000
10000
x [m]
Figure 5: A single run of the δ-GLMB particle filter applied to bearings-only multi-target
tracking.
The average tracking performance is studied next. For comparison sake, we
also present the tracking results of the Bernoulli particle filter (Sec. 4.3) which
has been modified to track multiple targets via the use of the “linear multitarget” (LM) technique proposed in [69]. The LM technique is an approximation
in which the detections from the nearby targets are considered as false detections
that affect only the clutter density c(z). This filter is referred to as the LMBernoulli particle filter. Fig. 6 shows the mean OSPA errors, obtained by
averaging over 500 Monte Carlo runs of the δ-GLMB PF and the LM-Bernoulli
PF. One can observe that the average performance of the two tracking filters
is almost equivalent when the targets are well separated. However, when the
42
targets approach each other (from around 1400s to 1900s), the δ-GLMB PF
exhibits better performance. This is to be expected since the δ-GLMB PF is an
implementation of the principled Bayes-optimal multi-target tracker. We can
also comment on the performance of the two Bernoulli-type tracking particle
filters versus the PHD particle filters, whose OSPA errors were shown in Fig.
3.(b). First, recall that the tracking filters produce tracks (as opposed to the
PHD filters, which produce only multi-target state estimates). Comparing Figs.
3 and 6 we also observe that both Bernoulli-type tracking particle filters achieve
somewhat lower mean OSPA errors than the PHD particle filter described in
Sec. 5.
5000
δ-GLMB PF
LM-Bernoulli PF
Mean OSPA error [m]
4000
3000
2000
1000
0
0
500
1000
1500
T i m e [s]
2000
2500
3000
Figure 6: Bearings-only multi-target tracking results: the mean OSPA error of the δ-GLMB
particle filter (blue) and the LM-Bernoulli particle filter (red).
7. Summary and further reading
This overview paper is an attempt to describe the particle method based
implementation of the new class of Bayes-optimal and suboptimal filters, derived
using random finite set models. The primary focus was on the Bernoulli PF, the
probability hypothesis density PF and the δ-generalised labelled multi-Bernoulli
PF. Each of these particle filters has been studied in detail and subsequently
demonstrated in the context of bearings-only filtering/tracking.
Particle filters for random finite set models is a growing research field and
there are many topics that this overview paper could not address due to the
space limitation. For an interested reader we briefly list some of these topics: different (non-standard) measurement models (e.g. intensity measurements
43
[70], [71], extended versus point target measurements [72],[73],[74],[75], group
tracking [76], finite sensor resolution or merged measurements [77], [78],[26]);
multi-sensor fusion aspects [79]; smoothing (as opposed to filtering) [80]; sensor
control aspects [35] and handling the switching dynamic models [81].
References
[1] A. H. Jazwinski. Stochastic processes and filtering theory. Academic press, 1970.
[2] A. Doucet, J. F. G. de Freitas, and N. J. Gordon, editors. Sequential Monte Carlo
Methods in Practice. Springer, 2001.
[3] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle filters for non-linear/non-Gaussian Bayesian tracking. IEEE Trans. Signal
Processing, 50(2):174–188, Feb. 2002.
[4] P. Djuric, J. H. Kotecha, J. Zhang, Y. Huang, T. Ghirmai, M. Bugallo, and
J. Miguez. Particle filtering. IEEE Signal Processing Magazine, pages 19–38,
Sept. 2003.
[5] B. Ristic, S. Arulampalam, and N. Gordon. Beyond the Kalman filter: Particle
filters for tracking applications. Artech House, 2004.
[6] O. Cappé, S. J. Godsill, and E. Moulines. An overview of existing methods and
recent advances in sequential Monte Carlo. Proc. IEEE, 95(5):899–924, 2007.
[7] A. Doucet and A. M. Johansen. A tutorial on particle filtering and smoothing:
Fifteen years later. Handbook of Nonlinear Filtering, 12:656–704, 2009.
[8] S. M. Kay. Fundamentals of statistical signal processing, Vol. II: Detection theory.
Prentice Hall, 1998.
[9] R. P. S. Mahler. Advances in Statisical Multisource-multitarget information fusion. Artech House, 2014.
[10] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan. Estimation with Applications to
Tracking and Navigation. John Wiley & Sons, 2001.
[11] S. J. Julier and J. K. Uhlmann. Unscented filtering and nonlinear estimation.
Proceedings of the IEEE, 92(3):401–422, 2004.
[12] S. Challa, M. R. Morelande, D. Mušicki, and R. J. Evans. Fundamentals of Object
Tracking. Cambridge Univ. Press, 2011.
[13] C. P. Robert and G. Casella. Monte Carlo statistical methods. Springer, 2nd
edition, 2004.
44
[14] C. Berzuini and W. R. Gilks. Resample-move filtering with cross-model jumps.
In A. Doucet, N. De Freitas, and N. Gordon, editors, Sequential Monte Carlo
Methods in Practice, chapter 6. Springer, 2001.
[15] N. J. Gordon, D. J. Salmond, and A. F. M. Smith. Novel approach to
nonlinear/non-Gaussian Bayesian state estimation. IEE Proc.-F, 140(2):107–113,
1993.
[16] A. Doucet, S. Godsill, and C. Andrieu. On sequential Monte Carlo sampling
methods for Bayesian filtering. Statistics and Computing, 10(3):197–208, 2000.
[17] M. Pitt and N. Shephard. Filtering via simulation: Auxiliary particle filters.
Journal of the American Statistical Association, 94(446):590–599, 1999.
[18] R. Van der Merwe, A. Doucet, N. De Freitas, and E. Wan. The unscented particle
filter. In Advances in Neural Information Processing Systems, volume 13. 2000.
[19] F. Daum, J. Huang, and A. Noushin. Exact particle flow for nonlinear filters. In
Proc. SPIE, volume 7697, 2010.
[20] C. Musso, N. Oudjane, and F. LeGland. Improving regularised particle filters.
In A. Doucet, N. deFreitas, and N. J. Gordon, editors, Sequential Monte Carlo
methods in Practice. Springer, 2001.
[21] P. B. Quang, C. Musso, and F. Le Gland. Particle filtering and the laplace method
for target tracking. IEEE Trans. Aerospace and Electronic Systems, 2015. (to
appear).
[22] B. W. Silverman. Density estimation for statistical and data analysis. Chapman
and Hall, 1986.
[23] C. Andrieu, A. Doucet, and R. Holenstein. Particle Markov chain Monte Carlo
methods. Journ. Royal Statistical Soc. B, 72(Part 3):269–342, 2010.
[24] T. Schon, F. Gustafsson, and P. J. Nordlund. Marginalized particle filters
for mixed linear/nonlinear state-space models. IEEE Trans Signal Processing,
53(7):2279–2289, 2005.
[25] S. Arulampalam, M. Clark, and R. Vinter. Performance of the shifted Rayleigh
filter in single-sensor bearings-only tracking. In Proc. 10th IEEE Int. Conf. Information Fusion, July 2007.
[26] M. Beard, B.-T. Vo, and B.-N. Vo. Bayesian multi-target tracking with merged
measurements using labelled random finite sets. IEEE Trans. Signal Processing,
63(6):1433–1447, 2015.
45
[27] R. Mahler. Statistical Multisource Multitarget Information Fusion. Artech House,
2007.
[28] B. N. Vo, S. Singh, and A. Doucet. Sequential Monte Carlo methods for multitarget filtering with random finite sets. IEEE Trans. Aerospace & Electronic
Systems, 41(4):1224–1245, Oct. 2005.
[29] B. Ristic, B.-T. Vo, B.-N. Vo, and A. Farina. A tutorial on Bernoulli filters: Theory, implementation and applications. IEEE Trans. Signal Processing,
61(13):3406–3430, 2013.
[30] B. Ristic. Particle filters for random set models. Springer, 2013.
[31] Y. Boers and J. N. Driessen. A particle filter multi target track before detect
application. IEE Proc. Radar, Sonar and Navigation, 151(6):351–357, 2004.
[32] H. Sidenbladh and S. L. Wirkander. Tracking random sets of vehicles in terrain.
In Proc. 2nd IEEE Workshop on Multi-Object Tracking, Madison, WI, USA, June
2003.
[33] T. Zajic and R. Mahler. A particle-systems implementation of the PHD multitarget tracking filter. In Proc. SPIE, volume 5096, pages 291–299, April 2003.
[34] M. Vihola. Rao-Blackwellised particle filtering in random set multitarget tracking.
IEEE Trans. Aerospace & Electronic Systems, 43(2):689–705, 2007.
[35] B. Ristic and B.-N. Vo. Sensor control for multi-object state-space estimation
using random finite sets. Automatica, 46:1812–1818, 2010.
[36] S. Reuter and K. Dietmayer. Pedestrian tracking using random finite sets. In
Proc. Int. Conf. Information Fusion, Chicago, USA, July 2011.
[37] Y. Boers, E. Sviestins, and H. Driessen. Mixed labelling in multitarget particle
filtering. IEEE Trans Aerospace and Electronic Systems, 46(2):792–802, 2010.
[38] R. P. S. Mahler. Multi-target Bayes filtering via first-order multi-target moments.
IEEE Trans. Aerospace & Electronic Systems, 39(4):1152–1178, 2003.
[39] M. Tobias and A.D. Lanterman. Probability hypothesis density-based multitarget
tracking with bistatic range and Doppler observations. IEE Proc.-Radar Sonar
Navig, 152(3):195–205, 2005.
[40] D. Clark, I. T. Ruiz, Y. Petillot, and J. Bell. Particle PHD filter multiple target
tracking in sonar image. IEEE Trans. Aerospace & Electronic Systems, 43(1):409–
416, 2007.
46
[41] E. Maggio, M. Taj, and A. Cavallaro. Efficient multitarget visual tracking using random finite sets. IEEE Trans. Circuits & Systems for Video Technology,
18(8):1016–1027, 2008.
[42] C. Lundquist, L. Hammarstrand, and F. Gustafsson. Road intensity based mapping using radar measurements with a probability hypothesis density filter. IEEE
Trans. on Signal Processing, 59(4):1397–1408, 2011.
[43] M. Canaud, L. Mihaylova, J. Sau, and N.-E. El Faouzi. Probabilty hypothesis
density filtering for real-time traffic state estimation and prediction. Networks
and Heterogeneous Media (NHM), 8(3):825–842, 2013.
[44] J. Mullane, B.-N. Vo, M. D. Adams and B.-T. Vo. A Random-Finite-Set Approach to Bayesian SLAM. IEEE Transactions on Robotics, 27(2):268-282, 2011.
[45] M. Adams, B.-N. Vo, R. Mahler, and J. Mullane. SLAM gets a PHD: New
concepts in map estimation. IEEE Robotics & Automation Magazine, 21(2):26–
37, 2014.
[46] C. S. Lee, S. Nagappa, N. Palomeras, D. E. Clark, and J. Salvi. Slam with SCPHD filters: An underwater vehicle application. IEEE Robotics & Automation
Magazine, 21(2):38–45, 2014.
[47] S. H. Rezatofighi, S. Gould, B.-N. Vo, K. Mele, and W. E. Hughesand R. Hartley. A multiple model probability hypothesis density tracker for time-lapse cell
microscopy sequences. In Information Processing in Medical Imaging, pages 110–
122. Springer, 2013.
[48] R. P. S. Mahler. PHD filters of higher order in target number. IEEE Trans.
Aerospace & Electronic Systems, 43(4):1523–1543, 2007.
[49] B. Ristic, D. Clark, B.-N. Vo, and B.-T. Vo. Adaptive target birth intensity
for PHD and CPHD filters. IEEE Trans. on Aerospace and Electronic Systems,
48(2):1656–1668, 2012.
[50] N. P. Whiteley, S. S. Singh, and S. J. Godsill. Auxiliary particle implementation of
the probability hypothesis density filter. IEEE Trans. on Aerospace & Electronic
Systems, 46(3):1437–1454, July 2010.
[51] J. H. Yoon, D. Y. Kim, and K.-Y. Yoon. Efficient importance sampling function
design for sequential Monte Carlo PHD filter. Signal Processing, 92:2315–2321,
2012.
[52] T. Li, S. Sun, and T. P. Sattar. High speed sigma-gating SMC-PHD filter. Signal
Processing, 93:2586–2593, 2013.
47
[53] B. Ristic. Efficient update of persistent particles in the SMC-PHD filter. In
Proc. IEEE Int. Conf. Acoustic Speech Signal Processing (ICASSP), Brisbane,
Australia, 2015.
[54] D. E. Clark, J. Bell, Y. de Saint-Pern, and Y. Petillot. PHD filter for multi-target
tracking in 3D sonar. In Proc. IEEE OCEANS-05-Europe, pages 265–270, Brest,
France, 2005.
[55] B. Ristic, D. E. Clark, and N. Gordon. Calibration of multi-atrget tracking
algorithms using non-cooperative targets. IEEE Journal of Selected Topics in
Signal Processing, 7(3):390–398, 2013.
[56] B. Ristic and D. Clark. Particle filter for joint estimation of multi-object dynamic
state and multi-sensor bias. In Proc. IEEE Int. Conf. Acoustics, Speech & Signal
Proc. (ICASSP), pages 3877–3880, Kyoto, Japan, March 2012.
[57] A. Swain and D. E. Clark. First-moment filters for spatial independent cluster
processes. Proc. of SPIE 7697, 2010.
[58] A. Swain and D. Clark. Extended object filtering using spatial independent cluster
processes. Proc. Int. Conf. Information Fusion, 2010.
[59] C. S. Lee, D. E. Clark, and J. Salvi. SLAM with dynamic targets via singlecluster PHD filtering. IEEE Journal of Selected Topics in Signal Processing,,
7(3):543–552, 2013.
[60] D. Schuhmacher, B.-T. Vo, and B.-N. Vo. A consistent metric for performance
evaluation of multi-object filters. IEEE Trans. Signal Processing, 56(8):3447–
3457, 2008.
[61] L. Lin, Y. bar Shalom, and T. Kirubarajan. Track labelling and PHD filter for
multitarget tracking. IEEE Trans Aerospace and Electronic Systems, 42(3):778–
795, 2006.
[62] J. Yang and H. Ji. A novel track maintenance algorithm for PHD/CPHD filter.
Signal Processing, 92:2371–2380, 2012.
[63] B.-T. Vo, B. N. Vo, and A. Cantoni. The cardinality balanced multi-target
multi-Bernoulli filter and its implementations. IEEE Trans. on Signal Processing,
57(2):409–423, 2009.
[64] B.-T. Vo and B.-N. Vo. Labeled random finite sets and multi-object conjugate
priors. IEEE Trans. on Signal Processing, 61(13):3460–3475, 2013.
[65] B.-N. Vo, B.-T. Vo, and D. Phung. Labeled random finite sets and the Bayes
multi-target tracking filter. IEEE Trans. on Signal Processing, 62(24):6554–6567,
2014.
48
[66] J. L. Williams. An efficient, variational approximation of the best fitting multiBernoulli filter. IEEE Trans. Signal Processing, 63(1):258–273, 2015.
[67] J. Correa, M. Adams, and C. Perez. A Dirac delta mixture-based random finite
set filter. Proc. IEEE Int. Conf. Control, Automation and Information Sciences
(ICCAIS), 2015.
[68] K. G. Murty. An algorithm for ranking all the assignments in order of increasing
cost. Operations Research, 16(3):682687, 1968.
[69] D. Musicki and B. La Scala. Multi-target tracking in clutter without measurement
assignment. IEEE Trans. Aerospace and Electronic Systems, 44(3):877–896, July
2008.
[70] B.-N. Vo, B.-T. Vo, N.-T. Pham, and D. Suter. Joint detection and estimation
of multiple objects from image observations. IEEE Trans. Signal Processing,
58(10):5129–5141, 2010.
[71] F. Papi, B.-N. Vo, B.-T. Vo, C. Fantacci, and M. Beard. Generalized labeled
multi-Bernoulli approximation of multi-object densities. IEEE Trans. Signal Processing, 63(20):5487–5497, Oct. 2015.
[72] K. Granström, C. Lundquist, and O. Orguner. Extended target tracking using a
Gaussian-mixture PHD filter. IEEE Trans. on Aerospace and Electronic Systems,
48(4):3268–3286, 2012.
[73] B. Ristic and J. Sherrah. Bernoulli filter for joint detection and tracking of an
extended object in clutter. IET Radar, Sonar & Navigation, 7(1):26–35, 2013.
[74] Meiqin Liu, Tongyang Jiang, and Senlin Zhang. The sequential Monte Carlo
multi-Bernoulli filter for extended targets. In 18th Intern. Conf. on Information
Fusion, pages 984–990, 2015.
[75] M. Beard, S. Reuter, K. Granström, B.-T. Vo, B.-N. Vo, and A. Scheel. Multiple
extended target tracking with labelled random finite sets. IEEE Trans. Signal
Processing, 2016. In print.
[76] A. Swain and D. Clark. The single-group PHD filter: an analytic solution. In
Proc. 14th Int. Conf. Information Fusion, Chicago, USA, July 2011.
[77] R. Mahler. PHD filters for nonstandard targets, II: unresolved targets. In Proc.
12th Int. Conf. Information Fusion, pages 922–929, Seattle, USA, 2009.
[78] F. Lian, C. Han, W. Liu, J. Liu, and J. Sun. Unified cardinalized probability
hypothesis density filters for extended targets and unresolved targets. Signal
Processing, 92(7):1729–1744, 2012.
49
[79] M. Uney, S. Julier, D. Clark, and B. Ristic. Monte carlo realisation of a distributed multi-object fusion algorithm. In IET Sensor Signal Processing for Defence (SSPD 2010), 2010.
[80] R. P. S. Mahler, B.-T. Vo, and B.-N. Vo. Forward-backward probability hypothesis density smoothing. IEEE Trans. Aerospace and Electronic Systems, 48(1):707–
728, Jan 2012.
[81] D. Dunne and T. Kirubarajan. Multiple model multi-Bernoulli filters for manoeuvering targets. IEEE Trans. Aerospace and Electronic Systems, 49(4):2679–2692,
2013.
50
| 3cs.SY
|
A DIMENSION-INDEPENDENT DISCRIMINANT BETWEEN DISTRIBUTIONS
Salimeh Yasaei Sekeh, Brandon Oselio, Alfred O. Hero III
arXiv:1802.04497v1 [cs.IT] 13 Feb 2018
Department of Electrical Engineering and Computer Science
University of Michigan
1301 Beal Ave, Ann Arbor, MI, 48109, USA
ABSTRACT
Henze-Penrose divergence is a non-parametric divergence measure that can be used to estimate a bound on the Bayes error in a
binary classification problem. In this paper, we show that a crossmatch statistic based on optimal weighted matching can be used to
directly estimate Henze-Penrose divergence. Unlike an earlier approach based on the Friedman-Rafsky minimal spanning tree statistic, the proposed method is dimension-independent. The new approach is evaluated using simulation and applied to real datasets to
obtain Bayes error estimates.
Index Terms— Bayes error rate, classification, Henze-Penrose
divergence, Cross-match test statistic, Optimal weighted matching,
Friedman-Rafsky statistic.
1. INTRODUCTION
Many information theoretic measures have been applied to measure
the discrimination between probability density functions. They have
been used in various applications in signal processing, classification,
image registration, clustering and structure learning, see [1, 2, 3, 4].
A special class of divergence measures, called f -divergences have
the property that the divergence functional f is convex and f (1) =
0. Among the different divergence functions belonging to the f divergence family, [5, 6] the Henze-Penrose (HP) divergence has
been of great interest due to its application to binary classification,
in particular to bound the Bayes error rate.
Let x1 , x2 , ..., xN ∈ Rd be realizations of random vector X
and class labels y ∈ {0, 1}, with prior probabilities c0 = P (y = 0)
and c1 = P (y = 1), such that c0 + c1 = 1. Given conditional
distributions p0 (x) and p1 (x), the Bayes error rate is given by
Z
ǫ=
min c0 p0 (x), c1 p1 (x) dx.
(1)
Rd
The Bayes error rate is the expected risk for the Bayes classifier,
which assigns a given feature vector x to the class with the highest posterior probability, and is the lowest possible error rate of any
classifier for a particular joint distribution. It is thus a reasonable
measure for assessing the intrinsic difficulty of a particular classification problem. By estimating and bounding this value, we can then
have a better understanding of the problem difficulty, which allows
the user to make more informed decisions.
We define the HP-divergence between p0 and p1 , Dc (p0 , p1 ) by
"Z
#
2
c0 p0 (x) − c1 p1 (x)
1
2
(2)
dx − (c0 − c1 ) .
4c0 c1
c0 p0 (x) + c1 p1 (x)
Rd
This work was partially supported by ARO grant W911NF-15-1-0479.
Note that for all c0 and c1 , 0 ≤ Dc (p0 , p1 ) ≤ 1 and when p0 = p1
the HP-divergence becomes zero.
The authors of [7] showed that HP-divergence yields tighter
bounds on the Bayes error rate ǫ, given in (1), than those based on
the Bhattacharya distance, [8]. In particular, the following bound on
the Bayes error rate holds:
1p
1
1
1
−
(3)
uc (p0 , p1 ) ≤ ǫ ≤ − uc (p0 , p1 ),
2
2
2
2
where uc (p0 , p1 ) = 4c0 c1 Dc (p0 , p1 ) + (c0 − c1 )2 .
In this paper we propose a new direct estimator for HPdivergence using a statistic based on optimal weighted matching
[9]. Matching for general graphs is a combinatorial optimization
problem that can be solved in polynomial time. In [9], the optimal
weighted matching was used to find a statistical test for equal posterior distributions using the cross match statistic. We demonstrate
that the same statistic described in that series of papers can be utilized to estimate HP-divergence. We emphasize that the proposed
weighted matching estimator is completely different from weighted
K-NN graph estimators.
The rest of the paper is organized as follows. Section 2 briefly
describes related work on HP-divergence and optimal weighted
matching. Section 3 defines the cross-match statistic, and in Section 4 we prove that the cross-match statistic approximately tends
to the HP-divergence when samples sizes of two classes increases
simultaneously in a specific regime. Section 5 shows sets of simulations for our proposed method and compares the Friedman-Rafsky
(FR) and cross-match estimators experimentally, and we estimate the
Bayes error rate on a few real datasets. Finally, Section 6 concludes
the paper.
2. RELATED WORK
Several estimators for HP-divergence have been proposed in the literature: Plug-in estimates were introduced in [10] and later have
been studied more in [11, 12, 13]. Plug-in approaches estimate the
underlying distribution function and then plug this value into the divergence function. The drawback with the plug-in estimates is that
these methods are not accurate near support boundaries and are also
more computationally complex. There have been a number of attempts to non-parametrically approximate divergence measures using graph-based algorithms such as minimal spanning tree (MST),
[14, 15] and k-nearest neighbors graphs (k-NNG), [16].
One of the most common direct estimators is based on FriedmanRafsky (FR) multivariate test statistic [17]. This approach is constructed from the MST on the concatenated data set drawn from
sufficiently smooth probability densities. Henze and Penrose [18]
showed that the FR test is consistent against all alternatives. Therefore, the HP-divergence has the appealing property that there exists
4
2
2
label
y
0
label
0
y
0
1
1
0
−2
−2
−4
−2
0
2
−2
0
x
2
x
Fig. 1. An example of the cross-match statistics for two cases p0 = p1 (left-generated from standard Gaussian distributions) and p0 6= p1
(right-Generated from Gaussian distributions with means [0, 0], [2, 2]). The total number of blue edges is the cross match statistics.
an asymptotically consistent direct estimator in terms of the FR test
statistic, see [18, 19, 7]. The variance of the FR test statistic under
the assumption of equal distributions depends on the dimension of
the data d, which may be unknown, especially when the support of
the densities is a common but unknown lower dimensional manifold.
Optimal weighted matching is a well studied combinatorial optimization problem [20]. It has been used extensively in operations
engineering. Previous statistical work using weighted matching have
derived useful applications of the cross-match test statistic in fields
like biological networks [9, 21].
3. THE CROSS-MATCH TEST STATISTIC
Consider N i.i.d. samples XN = {x1 , . . . , xN }, xi ∈ Rd and
corresponding labels yi ∈ {0, 1}. Define y = (y1 , . . . , yN ), and
N
P
yl , and n = N − m, so that m is the number
further m =
l=1
of samples in x with class 1, and n is the number with class label
0. Further, we create D, a N × N Euclidean distance matrix, with
Dij = ||xi − xj ||. Without loss of generality, we assume N is
even, as we can always add a ‘ghost point’ xN+1 , where DiN+1 =
0, ∀i. In the following, we consider a complete weighted graph G =
(V, E, D), with the vertices V = 1, . . . , N representing the sample
points x1 , . . . , xN , edges E = {{i, j}, i, j ∈ V }, and weights for
each edge {i, j} as Dij .
A complete matching M ⊂ E on a weighted graph is a set of
edges such that no two edges in M share a common vertex, and every
vertex is used in the matching. The complete minimum weighted
∗
∗
matching M
P is defined as the matching on G such that M =
arg minM i,j∈M Dij . We note that this is similar to the FR test
[17], which uses the same matrix D to find the minimal spanning
tree. The FR test statistic is the total number of edges in the D-based
MST connecting different labeled nodes.
Using this matching, we find the cross-match statistic, A(XN )
which is the number of edges that match dichotomous samples, i.e.
samples with different class labels, that is
X
yi (1 − yj ) + (1 − yi )yj .
(4)
A(XN ) =
{i,j}∈M ∗
In Figure 1 we show two numerical examples. The left plot
shows samples from two equal distributions, and right plot shows
samples from differing distributions. Qualitatively, we see that A
is much greater for the equivalent distributions than for the differing distributions, because the optimal matching tries to reduce long
distances, which will reduce the number of edges between differing
distributions.
In Proposition 1 in [9], under the assumption of equal distributions, the expectation and variance of A(x) are derived:
E[A] =
mn
2n(n − 1)m(m − 1)
, V ar[A] =
.
N −1
(N − 3)(N − 1)2
(5)
We note that the mean and variance of the cross-match statistic under
equal distributions are dimension-independent, but this is not true for
the FR statistic, whose variance is dependent on the degrees of the
MST. The maximal degrees of the MST is in fact dependent on the
dimension d of the underlying samples, e.g., the MST has maximal
degree 4 in d = 2 dimensions while its maximal degree is known to
be between 13 or 14 in 3 dimensions [22]. This dependence causes
the FR statistic to perform poorly in higher dimensions. In Section 5
we perform a set of experiments where dimension varies to demonstrate the advantage of the cross-match statistic over the FR statistic.
4. HP-DIVERGENCE ESTIMATION
Here we introduce the cross-match statistic as an estimate of the HPdivergence given in (2). Assume that we have two sets of samples
Xm = {X1 , . . . , Xm } and Un = {U1 , . . . , Un } with two different
labels. In order to show asymptotic convergence to HP-divergence,
we make the following assumption regarding the cross-match statistic (similar to Lemma 1 in [18]).
Assumption 1: For disjoint sets Xm , Un and {s, t} we have
A(Xm ∪ {s, t} ∪ Un ) − A(Xm ∪ Un ) ≤ kd .
(6)
where kd is a constant that may depend on d. This means that even
if the optimal matching changes a great deal, the number of edges
that are between the two samples is still approximately the same.
We empirically check this assumption in Figure 2. We generate two sets of d-dimensional samples from standard Gaussian
with mean µ0 = [0]d , µ1 = [1]d and Σ0 = Σ1 = Id for d =
2, 4, 6, 8. We plot the difference in cross-match statistic when adding
two points (labeled by Adiff ), and perform this test over varying sample size. We see that A does not vary significantly when adding a
new sample in the tested cases.
N→∞
1
2
Z
6
8
−1
Rd
200
300
400
500
n
(7)
N
−1/d
g(x, ZN
)
j ) − g(x, x) 1 Zj ∈ B(x, N
j=2
Z
= (N − 1)
g(x, y) − g(x, x) pN (y) dy
B(x,N −1/d )
Z
= (N − 1)
g(x, y)pN (y) − h(x, x)pN (x)
B(x,N −1/d )
+g(x, x)(pN (x) − pN (y) dy,
(8)
Since x is a Lebesgue point of pN and g(x, .)PN (.) then (8) tends to
x
zero. Note that the degree of vertex in M ∗ (ZN
) is one. For almost
all x,
j=2
4
100
N
X
N
X
2
g(x, x) p(x).
Proof: For given x in a subset S ∈ Rd , the degree of vertex x in
M ∗ (S) is one. Let x be a Lebesgue point of p(.) and p(.)g(x, .)
x
N
N
N
and
ZN be the point process {x, Z2 , Z3 , . . . , ZN }. Let B(x, r) =
y : ky − xk ≤ r . Therefore, we can write
E
d
0
1≤i<j≤N
=
E
1
A_diff
Lemma 1 Let g : Rd × Rd → [0, 1] be a symmetric and measurable function, such that for almost every x ∈ Rd , g(x, .) is measurable with x a Lebesgue point of the functions p(.)g(x, .) and p(.).
N
N
For each N , let ZN
1 , Z2 , . . . , ZN be independent d-dimensional
variables with common density function pN convergent to p as N →
N
∞ and set ZN = {ZN
1 , . . . , ZN }. Consider the complete minimum
∗
weighted matching M on ZN . Then
XX
N N
N
∗
g(ZN
lim N −1 E
i , Zj )1 (Zi , Zj ) ∈ M (ZN )
N
∗
x
g(x, ZN
j )1 (x, Zj ) ∈ M (ZN ) = g(x, x) + o(1). (9)
The function g has range [0, 1] so the left hand side of (9) is bounded
by one. By the dominated convergence theorem
N N
PP
N
N
∗
N −1 E
1≤i<j≤N g(Zi , Zj )1 (Zi , Zj ) ∈ M (ZN )
N
N N
1 X
N
∗
E
g(ZN
1 , Zj )1 (Z1 , Zj ) ∈ M (ZN )
2 j=2
Z
N
X
1
N
∗
=
pN (x)E
g(x, ZN
i )1 (x, Zj ) ∈ M (ZN ) .
2 x
j=2
=
(10)
The last line in (10) tends to right hand side of (8).
The following theorem proves the direct estimate of HPdivergence based on A(XN ). Due to space limitations only an
outline of the proof is given.
Theorem 1 As m → ∞ and n → ∞ such that m/N → c1 and
n/N → c0 , where N = m + n. Denote Am,n := A(Xm ∪ Un ) the
cross-match statistic given by the optimal weighted matching over
Xm and Un . Then under Assumption 1 we have
N
Am,n → Dc (p0 , p1 ), a.s.
(11)
1−
mn
Proof: The proof shares some similarity with the FR convergence
proof of the HP-divergence in [18]. The primary difference lies
Fig. 2. The cross-match statistics difference with error bars at the
standard deviation from 50 trials for the Gaussian samples by adding
two points.
in handling the difference between the cross-match statistic when
nodes are added, i.e. (6). We use Lemma 1 and Poissonization to
prove (11).
Let Mm and Nn be Poisson variables with mean m and n such
that m + n is even and independent of one another
and of Xi and
′
Uj . Let Xm
and Un′ be the Poisson processes X1 , . . . , XMm
′
and U1 , . . . , UNn , respectively. Set A′m.n = A(Xm
∪ Un′ ), the
cross-match statistic. By (6), we have
(12)
A′m,n − Am,n ≤ kd |Mm − m| + |Nn − n| .
Note that (m + n)−1 E A′m,n − Am,n | → 0. Poissonization makes
′
the identities of the points of Xm
∪ Un′ conditionally independent,
given their positions. For each m and n let Zm,n
, Zm,n
, . . . be
1
2
independent discrete variables with common density pm,n (x) =
(mp0 (x) + np1 (x))/(m + n). Let Wm,n be an independent
′
Poisson variable with even valued mean (m + n). Let Zm,n
=
m,n
m,n
{Z1 , . . . , ZWm,n } be a non-homogeneous Poisson process of
rate mp0 + np1 . Following the same arguments in [18], assign a
′
mark from the set {1, 2} to each point of Zm,n
. Specifically, a point
x is assigned mark 1 with probability mp0 (x) mp0 (x) + np1 (x)
′
and mark 2 otherwise. Let Xem and Uen be the set of points of Zm,n
e
marked 1 and 2 respectively. Also denote Am,n the cross match
statistic given from optimal weighted matching over Xem ∪ Uen . De′
fine the probability of two points in Zm,n
having different marks by
gm,n (x, y):
gm,n (x, y) =
mp0 (x)np1 (y) + np1 (x)mp0 (y)
.
(mp0 (x) + np1 (x))(mp0 (y) + np1 (y))
(13)
We know that m/N → c0 and n/N → c1 , hence gm,n (x, y) →
g(x, y) where
c0 c1 p0 (x)p1 (y) + p1 (x)p0 (y)
. (14)
g(x, y) =
c0 p0 (x) + c1 p1 (x) c0 p1 (y) + c1 p1 (y)
′
So, the conditional expectation E Aem,n |Zm,n
] becomes:
XX
′
gm,n (Zm,n
, Zm,n
)1 (Zm,n
, Zm,n
) ∈ M ∗ (Zm,n
) .
i
j
i
j
1≤i<j≤Wm,n
(15)
By taking expectations in (15), one yields E Aem,n .
(m+n)
be the original nonLet Zm,n := Zm,n
, Zm,n
, . . . , Zm,n
1
2
Poissonized set of points. By the fact that
E |Mm + Nn − (m + n)| = o(m + n),
the Poissonized limit of E Ãm,n . Set p(x) = c0 p0 (x) + c1 p1 (x),
then pm,n (x) → p(x). Using Lemma 1, we get
Z
E Aem,n
1
g(x, x)p(x)
→
(m + n)
2 Rd
Z
p0 (x)p1 (x)
= c0 c1
.
Rd c0 p0 (x) + c1 p1 (x)
0.6
(16)
HP
type
FR
0.4
Monte Carlo
OWM
0.2
This completes the proof of Theorem 1.
5
10
15
20
d
5. EXPERIMENTS
0.37
0.015
MSE
We perform multiple experiments to demonstrate the utility of the
proposed direct estimator of HP-divergence in terms of dimension
and sample size. We subsequently apply our estimator to determine
empirical bounds on the Bayes error rate for various datasets.
For the following simulations, the sample sizes for each class
were equal (m = n). Each simulation used a multivariate Normal
distribution for each class.
We first analyze the estimator’s performance as the sample size
N = m + n increases. For each value of N , the simulation was run
50 times, and the results were averaged. Samples from each class
were i.i.d. 2-dimensional Normal random variables, with µ0 = [0, 0]
and µ1 = [1, 1], Σ0 = Σ1 = I2 .
type
0.010
FR
OWM
0.005
0.000
5
10
15
20
d
Fig. 4. HP-divergence (top) and empirical MSE (bottom) vs. dimension. The empirical MSE of both estimators increases for larger
dimensional data sets. The MSE is better for the proposed (OWM)
estimator.
0.36
HP
type
Actual Value
0.35
FR
OWM
0.34
0.33
250
500
750
Data set
Breast cancer [23]
Mines vs. Rocks [24]
Pima diabetes [24]
Hyper thyroid [24]
Bayes Error Bounds
bc
A(XN )
n
D
33
0.791
488
7
0.864
97
67
0.641
549
37
0.743 3012
m
241
111
283
151
ǫ
0.093
0.067
0.161
0.023
1000
n
Fig. 3. HP-divergence estimation vs. sample size n. Error bars
denote the standard deviation over 50 trials. The proposed estimator
and the FR estimator perform approximately equivalently over this
range of sample sizes.
We see that as N increases the performance of the FR estimator
and our proposed estimator (labeled OWM) are comparable for N up
to 1000. The observed variance of our estimators are slightly higher
than the FR estimator. For dimension d = 2 this is not surprising as
we would expect the FR estimator to perform the best in this case.
Figure 4 (top) shows the averaged estimates of the HP-divergences
over increasing dimension. Here we see that the proposed crossmatching estimator shows improvement with respect to the FR
estimator, as expected. For each dimension evaluated in Figure 4,
N = 1000, and µ0 = [0]d and µ1 = [0.5]d , Σ0 = Σ1 = Id .
The proposed cross-matching estimator is slightly less biased as
dimension increases, and as shown in Figure 4 (bottom) we improve
in empirical MSE.
Next we show the results of applying the HP-divergence estimator to 4 different real data sets. Table 1 shows the cross match
statistics and estimated upper bounds for Bayes Error (denoted by
the column labeled ǫ).
b c , n, m and ǫ are the cross-match statistics, HPTable 1. A(XN ), D
divergence estimates using A(XN ), sample sizes and upper bounds
for Bayes Error respectively.
6. CONCLUSION
We proposed a new dimension-independent direct estimator of HPdivergence using a statistic derived from optimal weighted matching. The estimator is more accurate than the FR approach and its
variance is independent of the dimension of the support of the distributions. This translates to improved MSE performance as compared
to other HP-divergence estimation methods, especially for high dimension. We validated our proposed estimator using simulations,
and illustrated the approach for the meta-learning problem of estimating Bayes classification error for four real-world data sets.
7. ACKNOWLEDGEMENTS
We thank the UCI machine learning repository for the use of their
various datasets used in this paper [24].
8. REFERENCES
[1] Xuan Guorong, Chai Peiqi, and Wu Minhui, “Bhattacharyya
distance feature selection,” in In Pattern Recognition, Proceedings of the 13th International Conference on IEEE, 1996,
vol. 2, pp. 195–199.
[2] Paul Viola and William M Wells III, “Alignment by maximization of mutual information,” International journal of computer
vision, vol. 24, no. 2, pp. 137–154, 1997.
[3] A B Hamza and H Krim, “Image registration and segmentation by maximizing the jensen-rényi divergence,” in Energy
Minimization Methods in Computer Vision and Pattern Recognition, 2003, pp. 147–163.
[4] Kevin R Moon, Morteza Noshad, Salimeh Yasaei Sekeh, and
Alfred O Hero, “Information theoretic structure learning with
confidence,” in in Proc. IEEE Int. Conf. Acoust Speech Signal
Process, 2017.
[5] S Ali and S D Silvey, “A general class of coefficients of divergence of one distribution from another,” J. Royal Statist. Soc.
Ser. B (Methodology.), pp. 131–142, 1996.
[6] I Csiszár and P C Shields, “Information theory and statistics:
A tutorial,” J. Royal Statist. Soc. Ser. B (Methodology.), vol. 1,
no. 4, pp. 417–528, 2004.
[7] Visar Berisha, Alan Wisler, Alfred O. Hero, and Andreas
Spanias, “Empirically estimable classification bounds based
on a nonparametric divergence measure,” IEEE Trans. on Signal Process., vol. 64, no. 3, pp. 580–591, 2016.
[8] A Battacharyya, “On a measure of divergence between two
multinomial populations,” Sankhy ā: The Indian Journal of
Statistics, pp. 401–406, 1946.
[9] Paul R Rosenbaum, “An exact distribution-free test comparing
two multivariate distributions based on adjacency,” Journal of
Royal Statistics Society B, vol. 67, no. 4, pp. 515–530, 2005.
[10] Kumar Scricharan, Raviv Raich, and Alfred O Hero, “Estimation of nonlinear functionals of densities with confidence,”
IEEE Trans. Inf. Theory, vol. 57, no. 7, pp. 4135–4159, 2012.
[11] Kevin R Moon and Alfred O Hero, “Multivariate f -divergence
estimation with confidence,” in Advances in Neural Information Processing Systems, 2014, pp. 2420–2428.
[12] Kevin R Moon and Alfred O Hero, “Ensemble estimation of
multivariate f -divergence,” in IEEE International Symposium
on Information Theory, 2016, pp. 356–360.
[13] Kevin R Moon, Kumar Sricharan, Kristjan Greenewald, and
Alfred O Hero, “Non-parametric ensemble estimation of distributional functionals,” arXiv preprint arXiv:1601.06884v2.
[14] J E Yukish, Probability theory of classical Euclidean optimization, Vol. 1675 of lecture notes in Mathematics, SpringerVerlag, Berlin, 1998.
[15] D Aldous and J M Steele, “Asymptotic for euclidean minimal spanning trees on random points,” Probab. Theory Related
Fields, vol. 92, pp. 247–258, 1992.
[16] Jillian Beardwood, J H Halton, and J M Hammersley, “The
shortest path through many points,” in Mathematical Proceedings of the Cambridge Philosophical Society, 1959, pp. 299–
327.
[17] J H Friedman and L C Rafsky, “Multivariate generalizations
of the wald-wolfowitz and smirnov two-sample tests,” Ann.
Statist., pp. 697–717, 1979.
[18] Norbert Henze and Mathew D Penrose, “On the multivariate
runs test,” Ann. Statist., vol. 27, no. 1, pp. 290–298, 1999.
[19] Visar Berisha and Alfred O Hero, “Empirical non-parametric
estimation of the fisher information,” IEEE Signal Process.
Lett., vol. 22, no. 7, pp. 988–992, 2015.
[20] C H Papadimitriou and K Steiglitz, Combinatorial Optimization: Algorithms and Complexity, Englewood Cliffs, Prentice
Hall, 1982.
[21] Bo Lu and Paul R Rosenbaum, “Optimal pair matching with
two control groups,” Journal of Computational and Graphical
Statistics, vol. 13, no. 2, pp. 422–434, 2004.
[22] Gabriel Robins and Jeffrey S Salowe, “On the maximum degree of minimum spanning trees,” in Proceedings of the tenth
annual symposium on Computational geometry. ACM, 1994,
pp. 250–258.
[23] W H Wolberg and O L Mangasarian, “Multisurface method
of pattern separation for medical diagnosis applied to breast
cytology,” Proceedings of the National Academy of Sciences,
vol. 87, no. 23, pp. 9193–9196, 1990.
[24] M. Lichman, “UCI machine learning repository,” 2013.
| 10math.ST
|
n
-NEIGHBOURLY
3
MOMENT-ANGLE COMPLEXES AND THEIR UNSTABLE
SPLITTINGS
arXiv:1510.05283v2 [math.AT] 24 Feb 2016
PIOTR BEBEN AND JELENA GRBIĆ
Abstract. Given an n
-neighbourly simplicial complex K on vertex set [n], we show that the
3
moment-angle complex ZK is a co-H-space if and only if K satisfies a homotopy analogue of the
Golod property.
1. Introduction
A ring R = k[v1 , . . . , vn ]/I for I a homogeneous ideal is said to be Golod if the multiplication and
higher Massey products are trivial in Tor+k[v1 ,...,vn ] (R, k). The Poincaré series of the ring TorR (k, k)
of R represents a rational function whenever R is Golod [8]. In combinatorics, one says a simplicial
complex K on vertex set [n] = {1, . . . , n} is Golod over k if its Stanley-Reisner ring k[K] is Golod.
Equivalently, when k to be a field or Z, K is Golod over k if all cup products and Massey products
(of positive degree elements) are trivial in the cohomology ring H ∗ (ZK ; k) of the moment-angle
complex ZK . This follows from isomorphisms of graded commutative algebras given in [6, 7, 2, 13]
(1)
H ∗ (ZK ; k) ≅ Tork[v1 ,...,vn ] (k[K], k) ≅ ⊕ H̃ ∗ (Σ∣I∣+1 ∣KI ∣; k)
I⊆[n]
where KI is the restriction of K to vertex set I ⊆ [n]. The multiplication on the right hand side
is induced by maps ιI,J ∶ ∣KI∪J ∣ Ð→ ∣KI ∗ KJ ∣ ≅ ∣KI ∣ ∗ ∣KJ ∣ ≃ Σ∣KI ∣ ∧ ∣KJ ∣ that realize the canonical
inclusions KI∪J Ð→ KI ∗ KJ whenever I and J are non-empty and disjoint, and are defined zero
otherwise. Thus, the Golod condition is equivalent to ιI,J inducing trivial maps on k-cohomology
for all disjoint non-empty I, J ⊊ [n], as well as Massey products vanishing.
Since the cohomology ring H ∗ (ZK ) is in its simplest algebraic form when K is Golod, it is natural
to ask what the homotopy type of ZK is in this case. Quite a lot of work has been done in this
direction [10, 11, 12, 15, 9, 3, 14], and it has been conjectured that K is Golod if and only if ZK is a
co-H-space. One can then determine the homotopy type of ZK since, by [15, 1], ZK is a co-H-space
if and only if ZK ≃ ⋁I⊆[n] Σ∣I∣+1 ∣KI ∣. This conjecture has been verified for flag complexes [9, 5], and
rationally or localized at large primes [4, 3].
A simplicial complex K on vertex set [n] is said to be k-neighbourly if every subset of k or less
vertices in [n] is a face of K. That is, KI is a simplex for each I ⊆ [n], ∣I∣ ≤ k. If K is k-neighbourly
then it is k ′ -neighbourly for all k ′ ≤ k. When K is n3 -neighbourly, we show that a homotopy analogue
of Golodness is equivalent to the moment-angle complex ZK being a co-H-space.
Theorem 1.1. Let K be n3 -neighbourly. Then ZK is a co-H-space (equivalently, ZK ≃ ⋁ Σ∣I∣+1 ∣KI ∣)
I⊆[n]
if and only if
Σ∣I∪J∣ ιI,J ∶ Σ∣I∪J∣ ∣KI∪J ∣ Ð→ Σ∣I∪J∣ ∣KI ∗ KJ ∣
is nullhomotopic for each disjoint non-empty I, J ⊊ [n].
2010 Mathematics Subject Classification. Primary 55P15, 55U10, 13F55.
Key words and phrases. polyhedral product, moment-angle complex, toric topology, Stanley-Reisner ring, Golod
ring, neighbourly simplicial complexes.
1
2
PIOTR BEBEN AND JELENA GRBIĆ
Remark 1.2. The necessity part of Theorem 1.1 is true in general (see Proposition 2.4 below,
or Proposition 5.4 in [3]). Sufficiency generally holds up to a certain coherence condition ([3],
Theorem 1.2). Since this coherence condition involves joins ∣KI1 ∗ ⋯ ∗ KIm ∣ over disjoint non-empty
I1 , . . . , Im ⊊ [n] and m ≥ 3, coherence should be given to us for free when K is n3 -neighbourly, as
3-or-more-fold joins are contractible in this case. This is the heuristic behind Theorem 1.1.
Example 1.3. Suppose K is n2 -neighbourly. Then each ∣KI ∣ is contractible when ∣I∣ ≤ n2 , is at least
(⌊ n2 ⌋−1)-connected, and has dimension at most n−2 when K is not a simplex. From isomorphism (1),
or by Theorem 6.33 [6], one sees ZK is contractible when K is a simplex or has the homotopy type
of an n-connected CW -complex of at most dimension 2n − 1. Thus ZK is a co-H-space by the
Freudenthal suspension theorem. This also follows from distinct arguments found in [3, 14]. On the
other hand, it is an easy application of Theorem 1.1. Since for each disjoint non-empty I, J ⊊ [n]
ιI,J
at least one of KI or KJ must be a simplex, each ∣KI∪J ∣ Ð→ ∣KI ∗ KJ ∣ must be nullhomotopic since
∣KI ∗ KJ ∣ ≃ Σ∣KI ∣ ∧ ∣KJ ∣ is contractible, so ZK is a co-H-space by Theorem 1.1.
The first examples outside the stable range occur when K is ( n2 − 1)-neighbourly.
Example 1.4. Suppose K is ( n2 −1)-neighbourly and n is even. Then either (1) K = ∂∆ 2 −1 ∗∂∆ 2 −1
n
n
and ZK is a sphere product; (2) K is ∂∆ 2 −1 ∗ ∂∆ 2 −1 with some faces removed; or else (3) K has
only one minimal missing face with n2 vertices. In the third case, each ιI,J for disjoint I and
J is nullhomotopic since at least one of ∣KI ∣ or ∣KJ ∣ is contractible. In the second case, ιI,J is
nullhomotopic since ∣K∣ is at most (n − 4)-dimensional and ∣KI ∗ KJ ∣ is at least (n − 4)-connected.
Then in the last two cases, ZK is a co-H-space by Theorem 1.1.
n
n
Massey products do not factor into Theorem 1.1. In view of the isomorphism (1) only cup products
are relevant. This is as might be expected when localized at the rationals, since k-connected CW complexes of dimension ≤ 3k + 1 are rationally formal by [17]. Therefore they have trivial Massey
products in their rational cohomology.
2. Some Background
2.1. Conventions. We take −1 to be the basepoint of D1 = [−1, 1]. The suspension ΣX of a space
X is taken to be the reduced suspension D1 × X/({−1, 1} × X ∪ D1 × {∗}) whenever X is basepointed
with basepoint ∗. Otherwise it is the unreduced suspension, in other words, the quotient space of
D1 × X under identifications (−1, x) ∼ ∗−1 and (1, x) ∼ ∗1 . In any case, ΣX is always basepointed,
in the unreduced case the basepoint taken to be ∗−1 .
2.2. Moment-Angle Complexes. Let K be a simplicial complex on n vertices, and (X, A) a
CW -pair. The polyhedral product (X, A)K is the subspace of X ×n defined as the union
(X, A)K = ⋃ Y1σ × ⋯ × Ynσ ,
σ∈K
where Yiσ = X if i ∈ σ, or else Yiσ = A if i ∉ σ. The moment-angle complex is ZK = (D2 , S 1 )K and
the real moment-angle complex is RZK = (D1 , S 0 )K , where S 1 = ∂D2 and S 0 = {−1, 1} = ∂D1 .
Given I ⊆ [n], we can think of ZKI and RZKI as the subspace of ZK and RZK consisting of those
ℓ
ℓ
points whose coordinates not indexed by I are basepoints. In this case we define ZK
and RZK
to be
the unions of ZKI and RZKI respectively over all I ⊆ [n] such that ∣I∣ = ℓ, i.e. the subspaces of ZK
and RZK consisting of those points that have at least n−ℓ coordinates the basepoint. The quotiented
∣I∣−1
n−1
n−1 ̂
moment-angle complexes are defined as ẐK = ZK /ZK
, RẐK = RZK /RZK
, ZKI = ZKI /ZKI ,
∣I∣−1
RẐK = RZK /RZ
. Alternatively, they are the images of ZK , RZK , ZK , and RZK under
I
I
KI
I
I
the quotient maps (D2 )×n Ð→ (D2 )∧n , (D1 )×n Ð→ (D1 )∧n , (D2 )×∣I∣ Ð→ (D2 )∧∣I∣ , and (D1 )×∣I∣ Ð→
(D1 )∧∣I∣ , respectively. Then RẐK = ⋃σ∈K Y1σ ∧ ⋯ ∧ Ynσ , where Yiσ = D1 if i ∈ σ, or Yiσ = S 0 if i ∉ σ.
n
-NEIGHBOURLY
3
MOMENT-ANGLE COMPLEXES AND THEIR UNSTABLE SPLITTINGS
3
We say that a sequence (I1 , . . . , Im ) of subsets of a set I is an (ordered) partition of I if they are
mutually disjoint, non-empty, and I1 ∪⋯∪Im = I. Given a partition (I1 , . . . , Im ) of I with ∣I∣ = k and
Ij = {ij1 , . . . , ijkj } in increasing order, by rearranging coordinates, we can, and usually will, think of
the smash product RẐKI1 ∧ ⋯ ∧ RẐKIm as the subspace of (D1 )∧k given by
(2)
{(x1 , . . . , xk ) ∈ (D1 )∧k ∣ (xij1 , . . . , xijkj ) ∈ RẐKIj for each j}
(the order of the smash product therefore being irrelevant). Then RẐKI1 ∪⋯∪Im is a subspace of
RẐKI1 ∧ ⋯ ∧ RẐKIm , and we let
ι̂I1 ,...,Im ∶ RẐKI1 ∪⋯∪Im Ð→ RẐKI1 ∧ ⋯ ∧ RẐKIm
denote the inclusion. This carries over similarly for moment-angle complexes ZK . By rearranging
coordinates there is a homeomorphism ẐKI ≅ Σ∣I∣ RẐKI , and so we see that the inclusion
ι̂1I1 ,...,Im ∶ ẐKI1 ∪⋯∪Im Ð→ ẐKI1 ∧ ⋯ ∧ ẐKIm
is homeomorphic to the k-fold suspension of ι̂I1 ,...,Im above.
The homotopy type of a quotiented moment angle complex RẐK is not mysterious, at least not
with regards to the combinatorics of K. The existence of a homeomorphism RẐKI ≅ Σ∣KI ∣ was
given in [1, 3], and this is obtainable as a restriction of a homeomorphism Σ∣∆k−1 ∣ ≅ (D1 )∧k where
∣I∣ = k. More generally, Σ∣∆k−1 ∣ ≅ (D1 )∧k restricts to a homeomorphism
Σ∣KI1 ∗ ⋯ ∗ KIm ∣ ≅ RẐKI1 ∧ ⋯ ∧ RẐKIm
where (I1 , . . . , Im ) is a partition of I. Consequently, ι̂I1 ,...,Im is homeomorphic to the suspended
inclusion ΣιI1 ,...,Im ∶ Σ∣KI1 ∪⋯∪Im ∣ Ð→ Σ∣KI1 ∗ ⋯ ∗ KIm ∣.
2.3. A Few Splitting Conditions. If I is a subset of [n] with k = ∣I∣ elements, and R = (s1 , . . . , sk )
is any sequence of real numbers, let iℓ denote the ℓth smallest element in I, s′j denote the j th smallest
element in the set S = {s1 , . . . , sk }, and m = ∣S∣ be the number of distinct elements in R. Then assign
to R the ordered partition IR = (I1 , . . . , Im ) of I where Ij = {iℓ ∈ I ∣ sℓ = s′j }. For example, if I = [4]
and R = (−1, π, −1, 0), then [4]R = ({1, 3}, {4}, {2}). Take the diagonal
△n = {(x1 , . . . , xn ) ∈ Rn ∣ x1 = ⋯ = xn }
and consider the following subspace of the smash product Pn = (Rn − △n ) ∧ Σ∣∆n−1 ∣:
QK =
⋃
y∈(Rn −△n )
(I1 ,...,Im )=[n]y
{y} ∧ Σ∣KI1 ∗ ⋯ ∗ KIm ∣,
where [n]y is the ordered partition of [n] for the given sequence y, as defined above. Here we took
Rn −△n to be without a basepoint, so Pn is the half-smash ((Rn −△n )×Σ∣∆n−1 ∣)/(Rn −△n )×{∗}, and
is itself basepointed, as is QK . Given any ω = (t1 , . . . , tn−1 , t, z) ∈ Σn ∣K∣ (z ∈ ∣K∣ and t1 , . . . tn−1 , t ∈
[−1, 1]), let β = max{∣t1 ∣, . . . , ∣tn−1 ∣, 0}, and define the map
ΦK ∶ Σn ∣K∣ Ð→ ΣQK
by
⎧
⎪
if β ∈ {0, 1};
⎪∗
ΦK (ω) = ⎨
⎪
⎪
⎩(2β − 1, (t1 , . . . , tn−1 , 0), (t, z)) if 0 < β < 1.
Remark: Since (t1 , . . . , tn−1 , 0) cannot be on the diagonal △n , one cannot generally define a nullhomotopy of ΦK by trying to shift the parameter β to 1.
Definition 2.1. K is weakly coherently homotopy Golod if K is a single vertex, or (recursively)
the vertex deletion K/{i} is weakly coherently homotopy Golod for each i ∈ [n], and the map
ΦK ∶ Σn ∣K∣ Ð→ ΣQK is nullhomotopic.
4
PIOTR BEBEN AND JELENA GRBIĆ
Remark 2.2. Equivalently, K is weakly coherently homotopy Golod if ΦKI ∶ Σ∣I∣ ∣KI ∣ Ð→ ΣQKI is
nullhomotopic for each I ⊆ [n].
Proposition 2.3 (Theorem 4.3 in [3]). If K is weakly coherently homotopy Golod, then ZK is a
co-H-space.
◻
On the other hand, a somewhat weaker looking condition is a necessary one.
Proposition 2.4 (Proposition 4.4 in [3]). If ZK is a co-H-space, then each of the suspended inclusions
Σ∣I∪J∣ ιI,J ∶ Σ∣I∪J∣ ∣KI∪J ∣ Ð→ Σ∣I∪J∣ ∣KI ∗ KJ ∣
is nullhomotopic for every disjoint non-empty I, J ⊊ [n].
We include the proof since it is short and illustrative.
Proof of Proposition 2.4. Recall that Y is a co-H-space with comultiplication Y Ð→ Y ∨ Y if and
ψ
only if the diagonal map Y Ð→ Y × Y is homotopic to the composite Y Ð→ Y ∨ Y Ð→ Y × Y . In
△
≅
¯ ∶ Y Ð→
particular, this implies that the reduced diagonal map △
Y ×Y Ð→ (Y ×Y )/(Y ∨Y ) Ð→ Y ∧Y
is nullhomotopic whenever Y is a co-H-space.
Suppose ZK is a co-H-space. Since ZKI∪J is a retract of ZK , then ZKI∪J is a co-H-space for any
ψ
△
include
disjoint non-empty I, J ⊊ [n], and so ZKI∪J Ð→ ZKI∪J ∧ZKI∪J is nullhomotopic. Then the composite
¯
△
△
ZKI∪J Ð→ ZKI∪J ∧ ZKI∪J Ð→ ZKI ∧ ZKJ Ð→ ẐKI ∧ ẐKJ
¯
is nullhomotopic (where the second last map is the smash of the coordinate-wise projection maps
onto ZKI and ZKJ , and the last map is the smash of quotient maps), and since it is equal to
ZKI∪J
quotient
Ð→
I,J
ẐKI∪J Ð→ ẐKI ∧ ẐKJ ,
ι̂1
this last composite is nullhomotopic as well. But ZKI∪J is a co-H-space, so by Proposition 2.11 in [3],
the quotient map ZKI∪J Ð→ ẐKI∪J has a right homotopy inverse, implying ι̂1I,J is nullhomotopic.
Since ι̂1I,J is up to homeomorphism the (∣I ∪ J∣)-fold suspension of ι̂I,J ∶ RẐKI∪J Ð→ RẐKI ∧ RẐKJ ,
and since ι̂I,J is ΣιI,J up to homeomorphism, then Σ∣I∪J∣+1 ιI,J is nullhomotopic.
Now Σ∣I∪J∣ ∣KI∪J ∣ is at most (2(∣I∣ + ∣J∣) − 1)-dimensional and Σ∣I∪J∣ ∣KI ∗ KJ ∣ ≃ Σ∣I∪J∣+1 ∣KI ∣ ∧ ∣KJ ∣
is at least (∣I∣ + ∣J∣)-connected, so the suspension homomorphism
[ Σ∣I∪J∣ ∣KI∪J ∣ , Σ∣I∪J∣ ∣KI ∗ KJ ∣ ] Ð→ [ Σ∣I∪J∣+1 ∣KI∪J ∣ , Σ∣I∪J∣+1 ∣KI ∗ KJ ∣ ]
is an isomorphism by the Freudenthal suspension theorem ([16], Theorem 4.10). Thus Σ∣I∪J∣ ιI,J is
nullhomotopic since Σ∣I∪J∣+1 ιI,J is.
Remark 2.5. Proposition 2.3 is a consequence of the following. Given any map f ∶ ΣX Ð→ Y , f
Σf ad
factors as ΣX Ð→ ΣΩY Ð→ Y where f ad is the adjoint of f and ev the adjoint of the identity 1ΩY .
n−1
Taking f to be an attaching map Σn ∣K∣ Ð→ ZK
whose cofiber is ZK , it was shown in [3] that Σf ad
ev
ΦK
factors through Σn ∣K∣ Ð→ ΣQK . Then Proposition 2.3 follows from the fact that f is nullhomotopic
if and only if Σf ad is.
2.4. A Reformulation of QK . It will be convenient to redefine QK and ΦK in terms of quotiented
moment-angle complexes. Take the following subspace of (Rn − △n ) ∧ (D1 )∧n :
Q′K ≅
⋃
y∈(Rn −△n )
(I1 ,...,Im )=[n]y
{y} ∧ RẐKI1 ∧ ⋯ ∧ RẐKIm .
n
-NEIGHBOURLY
3
MOMENT-ANGLE COMPLEXES AND THEIR UNSTABLE SPLITTINGS
5
As remarked before, there is a homeomorphism Σ∣∆n−1 ∣ ≅ (D1 )∧n restricting for each partition
(I1 , . . . , Im ) of [n] to a homeomorphism Σ∣KI1 ∗ ⋯ ∗ KIm ∣ ≅ RẐKI1 ∧ ⋯ ∧ RẐKIm , such the inclusion
ι̂I1 ,...,Im is up to homeomorphism the suspended inclusion ΣιI1 ,...,Im . Thus, RẐK ≅ Σ∣K∣ and there
is a homeomorphism (Rn − △n ) ∧ Σ∣∆n−1 ∣ ≅ (Rn − △n ) ∧ (D1 )∧n restricting to a homeomorphism
QK ≅ Q′K ,
with respect to which ΦK is the map
Φ′K ∶ Σn−1 RẐK Ð→ ΣQ′K
defined by: Φ′K (t1 , . . . , tn−1 , x) = (2β − 1, (t1 , . . . , tn−1 , 0), x) if β = max{∣t1 ∣, . . . , ∣tn−1 ∣, 0} is between
0 and 1, otherwise Φ′K maps to the basepoint. We will work with Q′K and Φ′K instead of QK and
ΦK from now on.
2.5. Clusters and Comultiplication. Given integers m ≤ n, a metric space M , a fixed point
xn ∈ M , and a function f ∶ M ×(n−1) Ð→ R, consider the configuration space A ⊆ M ×(n−1) consisting
of all points y = (x1 . . . xn−1 ) ∈ M ×(n−1) such that each xi (including xn ) has a subset of m of the
xj ’s (j ≠ i) clustered around it within distance f (y). We describe the topology of A when M is the
interval (0, 1), m = ⌊ n3 ⌋, xn = 0, and f (y) is proportional to the maximal distance between the xi ’s.
As a convention, min ∅ = 0 and max ∅ = 0. If y = (t1 , . . . , tn−1 ) ∈ Rn−1 , let ẙ = (t1 , . . . , tn−1 , 0) ∈ Rn .
Given I ⊆ [n], n ≥ 2, and z = (t1 , . . . , tn ) ∈ Rn , define the following non-negative values
n
z
νI,i
= min {max {∣ti − tj ∣ ∣ j ∈ S} ∣ S ⊆ I − {i} and ∣S∣ = ⌊ ⌋} ,
3
z
∣ i ∈ I} ,
νIz = max {νI,i
1
(max{t1 , . . . , tn } − min{t1 , . . . , tn }).
n
z
z
z
In other words, νI,i
is the smallest value such that the interval [ti − νI,i
, ti + νI,i
] contains at least
n
⌊ 3 ⌋ of the tj ’s over j ∈ I − {i}, and is zero if no such value exists. Consider the subspace A of the
interior int(Dn−1 ) of the (n − 1)-disk Dn−1 = [−1, 1]×(n−1) given by
δz =
ẙ
A = {y ∈ int(Dn−1 ) ∣ ν[n]
< δ ẙ } .
Take the subset of ordered partitions of [n]
O = {(I, J) ∣ I ∪ J = [n], I ∩ J = ∅, and ∣I∣, ∣J∣ >
n
},
3
and for each (I, J) ∈ O, consider the subspace of int(Dn−1 )
ẙ
ẙ
AI,J = {y = (t1 , . . . , tn−1 ) ∈ int(Dn−1 ) ∣ tj − ti > δ ẙ , νI,i
< δ ẙ , νJ,j
< δ ẙ for i ∈ I, j ∈ J, where tn = 0} .
Let
ÂI,J = cl(AI,J )/∂AI,J
where cl(AI,J ) is the closure of AI,J , and ∂AI,J is its boundary.
Lemma 2.6. (i) The AI,J ’s are mutually disjoint, and A is a union of AI,J over all partitions
(I, J) ∈ O; and (ii) each AI,J is an open subspace of int(Dn−1 ) homeomorphic to the interior of
Dn−1 . Hence
Dn−1 /(Dn−1 − A)
where ÂI,J ≅ S
n−1
for each (I, J) ∈ O.
≅
⋁
(I,J)∈O
ÂI,J
≅
⋁
(I,J)∈O
S n−1
6
PIOTR BEBEN AND JELENA GRBIĆ
Proof of (i). Take any distinct (I, J), (I ′ , J ′ ) ∈ O, and let tn = 0 and y = (t1 , . . . , tn−1 ) ∈ Rn−1 .
Since we have ti < tj when y ∈ AI,J , i ∈ I and j ∈ J; or when y ∈ AI ′ ,J ′ , i ∈ I ′ and j ∈ J ′ ; then
AI,J ∩AI ′ ,J ′ = ∅ whenever I ∩J ′ ≠ ∅ and J ∩I ′ ≠ ∅. Suppose I ∩J ′ = ∅ or J ∩I ′ = ∅. Since (I, J) and
(I ′ , J ′ ) are distinct partitions of [n], then one of I ′ ⊊ I, I ⊊ I ′ , J ′ ⊊ J, or J ⊊ J ′ holds. Without loss
ẙ
of generality, suppose I ′ ⊊ I. If y ∈ AI,J , we have νI,i
< δ ẙ for any i ∈ I, so given that i ∈ I − I ′ ⊆ J ′ ,
n
there must be an S ⊆ I such that ∣S∣ = ⌊ 3 ⌋ and max {∣ti − tj ∣ ∣ j ∈ S} < δ ẙ . Since ∣I∣, ∣J∣, ∣I ′ ∣ > n3 and
I ′ ⊊ I, then ∣I − I ′ ∣ < n3 , so S must contain at least one element i′ in I ′ . But if y ∈ AI ′ ,J ′ , then
max {∣ti − tj ∣ ∣ j ∈ S} ≥ ∣ti − ti′ ∣ ≥ δ ẙ since i ∈ J ′ and i′ ∈ I ′ , a contradiction. Thus AI,J ∩ AI ′ ,J ′ = ∅
here as well. In summary, AI,J and AI ′ ,J ′ are disjoint for any distinct (I, J), (I ′ , J ′ ) ∈ O.
Suppose that y ∈ A. Let a = min{t1 , . . . , tn } and b = max{t1 , . . . , tn }. Since there are at most
n − 2 t′i s in the interval (a, b), there is an l and an r in [n] such that there are no ti ’s in the interval
(tl , tr ), and tr − tl > n1 (b − a) = δ ẙ . Then we have disjoint non-empty sets I = {i ∈ [n] ∣ ti ≤ tl } and
J = {j ∈ [n] ∣ tj ≥ tr } partitioning [n] with tj − ti > δ ẙ for every i ∈ I and j ∈ J. Notice that this last
fact implies for every k ∈ I there cannot be an S ⊆ [n] − {k} such that S contains an element in J
ẙ
< δ ẙ (as y ∈ A), there is for each
and: (∗) ∣S∣ = ⌊ n3 ⌋ and max {∣tk − tj ∣ ∣ j ∈ S} < δ ẙ . But since ν[n]
k ∈ I at least one such S ⊆ [n] − {k} such that (∗) holds, and we must have S ⊆ I since S cannot
ẙ
ẙ
< δ ẙ for every k ∈ I. Similarly, ∣J∣ > n3 and νJ,k
< δ ẙ
contain elements in J. Therefore ∣I∣ > n3 and νI,k
for every k ∈ J. It follows that (I, J) is in O, and y ∈ AI,J . Thus A ⊆ ⋃(I,J)∈O AI,J .
ẙ
ẙ
On the other hand, if y ∈ AI,J , since either νI,i
< δ ẙ or νJ,i
< δ ẙ for any i ∈ [n], depending on
ẙ
ẙ
whether i ∈ I or i ∈ J, then ν[n],i
< δ ẙ for each i, implying that ν[n]
< δ ẙ . Therefore y ∈ A. Thus
⋃(I,J)∈O AI,J ⊆ A.
Proof of (ii). Notice that A is an open subspace of int(Dn−1 ) since it does not contain any points
ẙ
in its boundary {y ∈ int(Dn−1 ) ∣ ν[n]
= δ ẙ }. Thus AI,J is also open by part (i).
Given (I, J) ∈ O, consider the point b = (b1 , . . . , bn−1 ) in AI,J whose coordinates are defined for
indices i ∈ I and j ∈ J by either: (1) bi = − 12 and bj = 0 if n ∈ J; or else (2) bi = 0 and bj = 21 if
ẙt
ẙ
1
), νI,i
= (1 − t)νI,i
,
n ∈ I. Given y ∈ AI,J , let yt = (1 − t)y + tb. Notice that δ ẙt = (1 − t)δ ẙ + t( 2n
ẙt
ẙ
νJ,j
= (1 − t)νJ,j
and ((1 − t)tj + tbj ) − ((1 − t)ti + tbi ) = (1 − t)(tj − ti ) + t( 21 ) for any i ∈ I and j ∈ J,
where tn = 0. It follows that yt is in AI,J for each t ∈ [0, 1] whenever y is, and AI,J deformation
retracts onto the point b via the homotopy sending y ↦ yt at each time t. Moreover, this is a linear
homotopy with yt lying on the line through b and y, and AI,J is an open subspace of int(Dn−1 ). We
conclude that AI,J is homeomorphic to int(Dn−1 ).
Since A is in the interior of Dn−1 , ∂Dn−1 ⊆ Dn−1 − A. Of use to us will be the comultiplication
ψ ∶ S n−1 ≅ Dn−1 /∂Dn−1 Ð→ Dn−1 /(Dn−1 − A) Ð→
≅
⋁
S n−1
(I,J)∈O
the first map being the quotient map, the last map the wedge of homeomorphisms ÂI,J Ð→ S n−1
restricting away from the basepoint to the homeomorphism
h ∶ AI,J Ð→ int(Dn−1 )
from Lemma 2.6. On S n−1 ∧ RẐK = Σn−1 RẐK , this gives the comultiplication
ψK = ψ ∧ 1 ∶ Σn−1 RẐK Ð→
where 1∶ RẐK Ð→ RẐK is the identity.
⋁
(I,J)∈O
Σn−1 RẐK
n
-NEIGHBOURLY
3
MOMENT-ANGLE COMPLEXES AND THEIR UNSTABLE SPLITTINGS
7
3. Proof of Theorem 1.1
Suppose K is n3 -neighbourly. Notice that RZKI ′ = (D1 )×∣I ∣ whenever ∣I ′ ∣ ≤ n3 , in which case we
′
have RẐKI ′ = (D1 )∧∣I ∣ . Thinking of RẐKI ∧ RẐKJ as the subspace of (D1 )∧n with coordinates
rearranged as in (2), we can define for any (I, J) ∈ O a map
′
λI,J ∶ Σn−1 RẐKI ∧ RẐKJ Ð→ ΣQ′K
that is given for any ω = (t1 , . . . , tn−1 , (x1 , . . . , xn )) by
⎧
⎪
if β ∈ {0, 1};
⎪∗
λI,J (ω) = ⎨
′
′
⎪
⎪
⎩(2β − 1, (t1 , . . . , tn−1 , 0), (ϕ1 (x1 ), . . . , ϕn (xn ))) if 0 < β < 1,
where β = max{∣t1 ∣, . . . , ∣tn−1 ∣, 0}, y ′ = (t′1 , . . . , t′n−1 ) = h−1 (t1 , . . . , tn−1 ) for h−1 ∶ int(Dn−1 ) Ð→ AI,J
the inverse of the homeomorphism h in the definition of ψK , and
⎫
⎧
ẙ ′
ẙ ′
⎪
⎛
⎪
⎪
⎪⎞
⎪ δ − ν[n],i ⎪
⎬⎟ − 1.
ϕi (xi ) = (1 + xi ) ⎜max ⎨ 0 ,
′
ẙ
⎪
⎪
δ
⎪
⎪
⎝
⎪
⎪
⎭⎠
⎩
The t′i ’s cannot all be 0 = t′n since by definition of AI,J , we have t′i < t′j when i ∈ I and j ∈ J, thus
λI,J (ω) is in Σ(Rn − △n ) ∧ (D1 )∧n at the very least. By this last fact, and since we have ϕi (xi ) = xi
for any i such that there is an I ′ ⊆ [n] with i ∈ I ′ , ∣I ′ ∣ > n3 , and t′i = t′j for any j ∈ I ′ , it follows that
′
λI,J (ω) ∈ ΣQ′K . This is necessary since RẐKI ′ is not necessarily (D1 )∧∣I ∣ when ∣I ′ ∣ > n3 . Otherwise,
′
when ∣I ′ ∣ ≤ n3 , we have RẐKI ′ = (D1 )∧∣I ∣ , meaning ϕi (xi ) is free to take on any value in D1 = [−1, 1].
n
-neighbourly,
3
Lemma 3.1. Suppose K is
the composite
ζ ∶ Σn−1 RẐK Ð→
ψK
⋁
(I,J)∈O
≅
n ≥ 2. Then Φ′K ∶ Σn−1 RẐK Ð→ ΣQ′K is homotopic to
Σn−1 RẐK Ð→
⋁
(I,J)∈O
Σn−1 RẐKI ∧ RẐKJ Ð→ ΣQ′K
Σ
ι̂I,J
where the second map is the wedge of the suspended inclusions Σn−1 RẐK Ð→ Σn−1 RẐKI ∧ RẐKJ
over (I, J) ∈ O, and the last map restricts to λI,J on a summand Σn−1 RẐKI ∧ RẐKJ .
n−1
Proof. Define the homotopy Ht ∶ Σn−1 RẐK Ð→ ΣQ′K for each ω = (t1 , . . . , tn−1 , (x1 , . . . , xn )) in
Σn−1 RẐK , and at each time t ∈ [0, 1], by mapping to the basepoint if β = max{∣t1 ∣, . . . , ∣tn−1 ∣, 0}
is 0 or 1, otherwise
Ht (ω) = (2β − 1, (t1 , . . . , tn−1 , 0), (ϕ1,t (x1 ), . . . , ϕn,t (xn ))),
where ϕi,t (xi ) = (1 − t)xi + tϕi (xi ) and ϕi is as defined before for y = (t1 , . . . , tn−1 ) in place of y ′ .
We see that Ht (ω) ∈ Σ(Rn − △n ) ∧ (D1 )∧n is in ΣQ′K for similar reasons as outlined for λI,J .
Clearly H0 = Φ′K . By definition of A, ψK , and ϕi , we have ψK (ω) = ∗ and ϕi,1 (xi ) = ϕi (xi ) = −1
(the basepoint in D1 ) for some i when y ∉ A, thus ζ(ω) = H1 (ω) = ∗ in this case. On the other hand
A is the union of disjoint sets ⋃(I,J)∈O AI,J , and if y ∈ AI,J for some (I, J) ∈ O, then by definition of
ψK , ψK (ω) = (h(y), (x1 , . . . , xn )) in the summand Σn−1 RẐK corresponding to (I, J), in which case
ζ(ω) = λI,J ○ Σn−1 ι̂I,J (ψK (ω))
= λI,J (h(y), (x1 , . . . , xn ))
= (2β − 1, (h−1 (h(y)), 0), (ϕ1 (x1 ), . . . , ϕn (xn )))
= (2β − 1, (t1 , . . . , tn−1 , 0), (ϕ1 (x1 ), . . . , ϕn (xn )))
= H1 (ω).
Thus, H1 is the composite ζ.
8
PIOTR BEBEN AND JELENA GRBIĆ
Proof of Theorem 1.1. Necessity follows from Proposition 2.4. The theorem is trivial to check when
n = 1 since ZK = D2 here, so assume n ≥ 2. Since K is n3 -neighbourly, then so is KI , or ∣I∣
3
neighbourly in particular. Since the maps Φ′K and ι̂I,J ∶ RẐK Ð→ RẐKI ∧ RẐKJ are homeomorphic
to ΦK and ΣιI,J ∶ Σ∣K∣ Ð→ Σ∣KI ∗ KJ ∣ respectively, sufficiency follows from Proposition 2.3 and
applying Lemma 3.1 for each KI , ∣I∣ ≥ 2, in place of K.
References
1. A. Bahri, M. Bendersky, F. R. Cohen, and S. Gitler, The polyhedral product functor: a method of decomposition for moment-angle complexes, arrangements and related spaces, Adv. Math. 225 (2010), no. 3, 1634–1668.
MR 2673742 (2012b:13053)
2. I. V. Baskakov, V. M. Bukhshtaber, and T. E. Panov, Algebras of cellular cochains, and torus actions, Uspekhi
Mat. Nauk 59 (2004), no. 3(357), 159–160. MR 2117435 (2006b:57045)
3. Piotr Beben and Jelena Grbić, Configuration spaces and polyhedral products, preprint, arxiv:1409.4462.
4. Alexander Berglund, Homotopy invariants of davis-januszkiewicz spaces and moment-angle complexes.
5. Alexander Berglund and Michael Jöllenbeck, On the Golod property of Stanley-Reisner rings, J. Algebra 315
(2007), no. 1, 249–273. MR 2344344 (2008f:13036)
6. Victor M. Buchstaber and Taras E. Panov, Torus actions and their applications in topology and combinatorics, University Lecture Series, vol. 24, American Mathematical Society, Providence, RI, 2002. MR 1897064
(2003e:57039)
7. M. Franz, The integral cohomology of toric manifolds, Tr. Mat. Inst. Steklova 252 (2006), no. Geom. Topol.,
Diskret. Geom. i Teor. Mnozh., 61–70. MR 2255969 (2007f:14050)
8. E. S. Golod, Homologies of some local rings, Dokl. Akad. Nauk SSSR 144 (1962), 479–482. MR 0138667 (25
#2110)
9. Jelena Grbić, Taras Panov, Stephen Theriault, and Jie Wu, Homotopy types of moment-angle complexes for flag
complexes, preprint, arXiv:1211.0873, to appear in Trans. Amer. Math. Soc. http://dx.doi.org/10.1090/tran/6578.
10. Jelena Grbić and Stephen Theriault, Homotopy type of the complement of a configuration of coordinate subspaces
of codimension two, Uspekhi Mat. Nauk 59 (2004), no. 6(360), 203–204. MR 2138475 (2005k:55023)
11.
, The homotopy type of the complement of a coordinate subspace arrangement, Topology 46 (2007), no. 4,
357–396. MR 2321037 (2008j:13051)
, The homotopy type of the polyhedral product for shifted complexes, Adv. Math. 245 (2013), 690–715.
12.
MR 3084441
13. Melvin Hochster, Cohen-Macaulay rings, combinatorics, and simplicial complexes, Ring theory, II (Proc. Second
Conf., Univ. Oklahoma, Norman, Okla., 1975), Dekker, New York, 1977, pp. 171–223. Lecture Notes in Pure and
Appl. Math., Vol. 26. MR 0441987 (56 #376)
14. Kouyemon Iriye and Daisuke Kishimoto, Fat wedge filtrations and decomposition of polyhedral products, preprint,
arxiv:1412.4866.
15.
, Topology of polyhedral products and the golod property of stanley-reisner rings.
16. John McCleary, A user’s guide to spectral sequences, second ed., Cambridge Studies in Advanced Mathematics,
vol. 58, Cambridge University Press, Cambridge, 2001. MR 1793722 (2002c:55027)
17. James Stasheff, Rational Poincaré duality spaces, Illinois J. Math. 27 (1983), no. 1, 104–109. MR 684544
(85c:55012)
School of Mathematics, University of Southampton,Southampton SO17 1BJ, United Kingdom
E-mail address: [email protected]
School of Mathematics, University of Southampton,Southampton SO17 1BJ, United Kingdom
E-mail address: [email protected]
| 0math.AC
|
DOTA: A Large-scale Dataset for Object Detection in Aerial Images∗
arXiv:1711.10398v2 [cs.CV] 27 Jan 2018
Gui-Song Xia1∗ , Xiang Bai2†, Jian Ding1 , Zhen Zhu2 , Serge Belongie3 ,
Jiebo Luo4 , Mihai Datcu5 , Marcello Pelillo6 , Liangpei Zhang1
1
State Key Lab. LIESMARS, Wuhan University, China
{guisong.xia, jding, zlp62}@whu.edu.cn
2
EIS, Huazhong Univ. Sci. and Tech., China
{xbai, zzhu}@hust.edu.cn
3
Computer Science Depart., Cornell University, USA
[email protected]
4
Computer Science Depart., Rochester University, USA
[email protected]
5
German Aerospace Center (DLR), Germany
[email protected]
6
DAIS, University of Venice, Italy
[email protected]
January 30, 2018
Abstract
Object detection is an important and challenging problem in computer vision. Although the
past decade has witnessed major advances in object detection in natural scenes, such successes
have been slow to aerial imagery, not only because of the huge variation in the scale, orientation
and shape of the object instances on the earth’s surface, but also due to the scarcity of wellannotated datasets of objects in aerial scenes. To advance object detection research in Earth
Vision, also known as Earth Observation and Remote Sensing, we introduce a large-scale Dataset
for Object deTection in Aerial images (DOTA). To this end, we collect 2806 aerial images from
different sensors and platforms. Each image is of the size about 4000 × 4000 pixels and contains
objects exhibiting a wide variety of scales, orientations, and shapes. These DOTA images are
then annotated by experts in aerial image interpretation using 15 common object categories. The
fully annotated DOTA images contains 188, 282 instances, each of which is labeled by an arbitrary
(8 d.o.f.) quadrilateral. To build a baseline for object detection in Earth Vision, we evaluate
state-of-the-art object detection algorithms on DOTA. Experiments demonstrate that DOTA well
represents real Earth Vision applications and are quite challenging.
∗
†
DOTA dataset is available at http://captain.whu.edu.cn/DOTAweb or https://captain-whu.github.io/DOTA.
Equal contributions
1
1
Introduction
Object detection in Earth Vision refers to localizing objects of interest (e.g., vehicles, airplanes) on the
earth’s surface and predicting their categories. In contrast to conventional object detection datasets,
where objects are generally oriented upward due to gravity, the object instances in aerial images often
appear with arbitrary orientations, as illustrated in Fig. 1, depending on the perspective of the Earth
Vision platforms.
Extensive studies have been devoted to object detection in aerial images [3, 14, 17–19, 21, 23, 30,
31, 38], drawing upon recent advances in Computer Vision and accounting for the high demands of
Earth Vision applications. Most of these methods [3, 18, 31, 38] attempt to transfer object detection
algorithms developed for natural scenes to the aerial image domain. Recently, driven by the successes
of deep learning-based algorithms for object detection, Earth Vision researchers have pursued approaches based on fine-tuning networks pre-trained on large-scale image datasets (e.g., ImageNet [6]
and MSCOCO [13]) for detection in the aerial domain, see e.g. [2, 3, 18, 29].
While such fine-tuning based approaches are a reasonable avenue to explore, images such as Fig. 1
reveals that the task of object detection in aerial images is distinguished from the conventional object
detection task in the following respects:
- The scale variations of object instances in aerial images are huge. This is not only because of the
spatial resolutions of sensors, but also due to the size variations inside the same object category.
- Many small object instances are crowded in aerial images, for example, the ships in a harbor
and the vehicles in a parking lot, as illustrated in Fig. 1. Moreover, the frequencies of instances
in aerial images are unbalanced, for example, some small-size (e.g. 1k × 1k) images contain
1900 instances, while some large-size images (e.g. 4k × 4k) may contain only a handfull of small
instances.
- Objects in aerial images often appear in arbitrary orientations. There are also some instances
with an extremely large aspect ratio, such as a bridge.
Besides these distinct difficulties, the studies of object detection in Earth Vision are also challenged
by the well-known dataset bias problem [28], i.e.the degree of generalizability across datasets is often
low. In order to alleviate such biases, the dataset should be annotated to reflect the demands of real
world applications.
Therefore, it is not surprising that the object detectors learned from natural images are not suitable
for aerial images. However, existing annotated datasets for object detection in aerial images, such as
UCAS-AOD [40] and NWPU VHR-10 [2], tend to use images in ideal conditions (clear backgrounds
and without densely distributed instances), which cannot adequately reflect the problem complexity.
To advance the object detection research in Earth Vision, this paper introduces a large-scale
Dataset for Object deTection in Aerial images (DOTA). We collect 2806 aerial images from different
sensors and platforms with crowdsourcing. Each image is of the size about 4000 × 4000 pixels and
contains objects of different scales, orientations and shapes. These DOTA images are annotated
by experts in aerial image interpretation, with respect to 15 common object categories. The fully
annotated DOTA dataset contains 188,282 instances, each of which is labeled by an oriented bounding
box, instead of an axis-aligned one, as is typically used for object annotation in natural scenes. The
main contributions of this work are:
2
(a)
(b)
(c)
ship
harbor
roundabout
(e)
#Instances
car
(d)
#pixels
#Instances
(f)
radians
Figure 1: An example taken from DOTA. (a) Typical image in DOTA consisting of many instances
across multiple categories. (b) Illustration of the variety in instance orientation and size. (c),(d)
Illustration of sparse instances and crowded instances, respectively. Here we show four out of fifteen
of the possible categories in DOTA. Examples shown in (b),(c),(d) are cropped from source image
(a). The histograms (e),(f) exhibit the distribution of instances with respect to size and orientation
in DOTA.
3
- To our knowledge, DOTA is the largest annotated object dataset with a wide variety of categories
in Earth Vision. It can be used to develop and evaluate object detectors in aerial images. We
will continue to update DOTA, to grow in size and scope and to reflect evolving real world
conditions.
- We also benchmark state-of-the-art object detection algorithms on DOTA, which can be used
as the baseline for future algorithm development.
In addition to advancing object detection studies in Earth Vision, DOTA will also pose interesting
algorithmic questions to conventional object detection in computer vision.
2
Motivations
Datasets have played an important role in data-driven research in recent years [6, 13, 32, 35, 37, 39].
Large datasets like MSCOCO [13] are instrumental in promoting object detection and image captioning
research. When it comes to the classification task and scene recognition task, the same is true for
ImageNet [6] and Places [39], respectively.
However, in aerial object detection, a dataset resembling MSCOCO and ImageNet both in terms of
image number and detailed annotations has been missing, which becomes one of the main obstacles to
the research in Earth Vision, especially for developing deep learning-based algorithms. Aerial object
detection is extremely helpful for vehicle counting, remote object tracking and unmanned driving.
Therefore, a large-scale and challenging aerial object detection benchmark, being as close as possible
to real-world applications, is imperative for promoting research in this field.
We argue that a good aerial image dataset should possess four properties, namely, 1) a large
number of images, 2) many instances per categories, 3) properly oriented object annotation, and
4) many different classes of objects, which make it approach to real-world applications. However,
existing aerial image datasets [15, 17, 24, 40] share in common several shortcomings: insufficient data
and classes, lack of detailed annotations, as well as low image resolution. Moreover, their complexity
is inadequate to be considered as a reflection of the real world.
Dataset
NWPU VHR-10 [2]
SZTAKI-INRIA [1]
TAS [9]
COWC [20]
VEDAI [24]
UCAS-AOD [40]
HRSC2016 [17]
3K Vehicle Detection [15]
DOTA
Annotation way
horizontal BB
oriented BB
horizontal BB
one dot
oriented BB
oriented BB
oriented BB
oriented BB
oriented BB
#main categories
10
1
1
1
3
2
1
2
14
#Instances
3651
665
1319
32716
2950
14,596
2976
14,235
188,282
#Images
800
9
30
53
1268
1510
1061
20
2806
Image width
∼1000
∼800
792
2000∼19,000
512, 1024
∼1000
∼1100
5616
800∼4000
Table 1: Comparison among DOTA and object detection datasets in aerial images. BB is short
for bounding box. One-dot refers to annotations with only the center coordinates of an instance
provided. Fine-grained categories are not taken into account. For example, DOTA consist of 15
different categories but only 14 main categories, because small vehicle and large vehicle are both
sub-categories of vehicle.
Datasets like TAS [9], VEDAI [24], COWC [20] and DLR 3K Munich Vehicle [15] only focus on
vehicles. UCAS-AOD [40] contains vehicles and planes while HRSC2016 [17] only contains ships even
4
though fine-grained category information are given. All these datasets are short in the number of
classes, which restricts their applicabilities to complicated scenes. In contrast, NWPU VHR-10 [2] is
composed of ten different classes of objects while its total number of instances is only around 3000.
Detailed comparisons of these existing datasets are shown in Tab. 1. Compared to these aerial datasets,
as we shall see in Section 4, DOTA is challenging for its tremendous object instances, arbitrary but
well-distributed orientations, various categories and complicated aerial scenes. Moreover, scenes in
DOTA is in coincidence with natural scenes, so DOTA is more helpful for real-world applications.
When it comes to general objects datasets, ImageNet and MSCOCO are favored by researchers due
to the large number of images, many categories and detailed annotations. ImageNet has the largest
number of images among all object detection datasets. However, the average number of instances per
image is far smaller than MSCOCO and our DOTA, plus the limitations of its clean backgrounds and
carefully selected scenes. Images in DOTA contain an extremely large number of object instances,
some of which have more than 1,000 instances. PASCAL VOC Dataset [7] is similar to ImageNet in
instances per image and scenes but the inadequate number of images makes it unsuitable to handle
most detection needs. Our DOTA resembles MSCOCO in terms of the instance numbers and scene
types, but DOTA’s categories are not as many as MSCOCO because objects which can be seen clearly
in aerial images are quite limited.
Besides, what makes DOTA unique among the above mentioned large-scale general object detection
benchmarks is that the objects in DOTA are annotated with properly oriented bounding boxes (OBB
for short). OBB can better enclose the objects and differentiate crowded objects from each other.
The benefits of annotating objects in aerial images with OBB are further described in Section 3. We
draw a comparison among DOTA, PASCAL VOC, ImageNet and MSCOCO to show the differences
in Tab. 2.
Dataset
PASCAL VOC
(07++12)
MSCOCO
(2014 trainval)
ImageNet
(2017train)
DOTA
Category
Image
quantity
BBox
quantity
Avg. BBox
quantity
20
21,503
62,199
2.89
80
123,287
886,266
7.19
200
349,319
478,806
1.37
15
2,806
188,282
67.10
Table 2: Comparison among DOTA and other general object detection datasets. BBox is short for
bounding boxes, Avg. BBox quantity indicates average bounding box quantity per image. Note that
for the average number of instances per image, DOTA surpasses other datasets hugely.
3
3.1
Annotation of DOTA
Images collection
As mentioned in [5], in aerial images, the resolution and variety of sensors being used are factors
to produce dataset biases. To eliminate the biases, images in our dataset are collected from multiple
sensors and platforms (e.g. Google Earth) with multiple resolutions. To increase the diversity of data,
we collect images shot in multiple cities carefully chosen by experts in aerial image interpretation. We
5
record the exact geographical coordinates of the location and capture time of each image to ensure
there are no duplicate images.
3.2
Category selection
Fifteen categories are chosen and annotated in our DOTA dataset, including plane, ship, storage tank,
baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small
vehicle, helicopter, roundabout, soccer ball field and basketball court.
The categories are selected by experts in aerial image interpretation according to whether a kind
of objects is common and its value for real-world applications. The first 10 categories are common
in the existing datasets, e.g., [2, 15, 20, 40], We keep them all except that we further split vehicle
into large ones and small ones because there is obvious difference between these two sub-categories in
aerial images. Others are added mainly from the values in real applications. For example, we select
helicopter considering that moving objects are of significant importance in aerial images. Roundabout
is chosen because it plays an important role in roadway analysis.
It is worth discussing whether to take “stuff” categories into account. There are usually no clear
definitions for the ”stuff” categories (e.g. harbor, airport, parking lot), as is shown in the SUN dataset
[33]. However, the context information provided by them may be helpful for detection. We only adopt
the harbor category because its border is relatively easy to define and there are abundant harbor
instances in our image sources. The final extended category is soccer field.
In Fig.2, we compare the categories of DOTA with NWPU VHR-10 [2], which has the largest
number of categories in previous aerial object detection datasets. Note that DOTA surpass NWPU
VHR-10 not only in category numbers, but also the number of instances per category.
3.3
Annotation method
We consider different ways of annotating. In computer vision, many visual concepts such as region
descriptions, objects, attributes, and relationships, are annotated with bounding boxes, as shown in
[12]. A common description of bounding boxes is (xc , yc , w, h), where (xc , yc ) is the center location,
w, h are the width and height of the bounding box, respectively.
Objects without many orientations can be adequately annotated with this method. However,
bounding boxes labeled in this way cannot accurately or compactly outline oriented instances such as
text and objects in aerial images. In an extreme but actually common condition as shown in Fig. 3 (c)
and (d), the overlap between two bounding boxes is so large that state-of-the-art object detection
methods cannot differentiate them. In order to remedy this, we need to find an annotation method
suitable for oriented objects.
An option for annotating oriented objects is θ-based oriented bounding box which is adopted in
some text detection benchmarks [36], namely (xc , yc , w, h, θ), where θ denotes the angle from the
horizontal direction of the standard bounding box. A flaw of this method is the inability to compactly
enclose oriented objects with large deformation among different parts. Considering the complicated
scenes and various orientations of objects in aerial images, we need to abandon this method and choose
a more flexible and easy-to-understand way. An alternative is arbitrary quadrilateral bounding boxes,
which can be denoted as {(xi , yi ), i = 1, 2, 3, 4}, where (xi , yi ) denotes the positions of the oriented
bounding boxes’ vertices in the image. The vertices are arranged in a clockwise order. This way is
6
soccer ball field
helicopter
swimming pool
roundabout
large vehicle
small vehicle
bridge
harbor
ground track field
basketball court
tennis court
baseball diamond
storage tank
ship
plane
100
1000
NWPU VHR 10
10000
100000
DOTA
Figure 2: Comparison between DOTA and NWPU VHR-10 in categories and responding quantity of
instances.
(a)
(b)
(c)
(d)
Figure 3: Visualization of adopted annotation method. The yellow point represents the starting point,
which refers to: (a) top left corner of a plane, (b) the center of sector-shaped baseball diamond, (c)
top left corner of a large vehicle. (d) is a failure case of the horizontal rectangle annotation, which
brings high overlap compared to (c).
widely adopted in oriented scene text detection benchmarks [11]. We draw inspiration from these
researches and use arbitrary quadrilateral bounding boxes to annotate objects.
To make a more detailed annotation, as illustrated Fig. 3, we emphasize the importance of the
first point (x1 , y1 ), which normally implies the “head” of the object. For helicopter, large vehicle,
small vehicle, harbor, baseball diamond, ship and plane, we carefully denote their first point to enrich
potential usages. While for soccer-ball field, swimming pool, bridge, ground track field, basketball
7
court and tennis court, there are no visual clues to decide the first point, so we normally choose the
top-left point as the starting point.
Some samples of annotated patches (not the whole original image) in our dataset are shown in
Fig. 4.
It is worth noticing that, Papadopoulos et al. [22] have explored an alternative annotation method
and verify its efficiency and robustness. We assume that the annotations would be more precise
and robust with more elaborately designed annotation methods, and alternative annotation protocols
would facilitate more efficient crowd-sourced image annotations.
Figure 4: Samples of annotated images in DOTA. We show three samples per each category, except
six for large-vehicle.
3.4
Dataset splits
In order to ensure that the training data and test data distributions approximately match, we randomly
select half of the original images as the training set, 1/6 as validation set, and 1/3 as the testing set.
8
We will publicly provide all the original images with ground truth for training set and validation set,
but not for the testing set. For testing, we are currently building an evaluation server.
4
4.1
Properties of DOTA
Image size
Aerial images are usually very large in size compared to those in natural images dataset. The original
size of images in our dataset ranges from about 800 × 800 to about 4000 × 4000 while most images
in regular datasets (e.g. PASCAL-VOC and MSCOCO) are no more than 1000 × 1000. We make
annotations on the original full image without partitioning it into pieces to avoid the cases where a
single instance is partitioned into different pieces.
4.2
Various orientations of instances
As shown in Fig.1 (f), our dataset achieves a good balance in the instances of different directions,
which is significantly helpful for learning a robust detector. Moreover, our dataset is closer to real
scenes because it is common to see objects in all kinds of orientations in the real world.
(a)
(b)
(c)
Figure 5: Statistics of instances in DOTA. AR denotes the aspect ratio. (a) The AR of horizontal
bounding box. (b) The AR of oriented bounding box. (c) Histogram of number of annotated instances
per image.
4.3
Spatial resolution information
We also provide the spatial resolution for each image in our dataset, which implies the actual size of
an instance and plays a significant role in aerial object detection. The importance of spatial resolution
for detection task are two folds. First, it allows the model to be more adaptive and robust for varieties
of objects of the same category. It’s known that objects appear smaller when seen from a distance.
The same object with different sizes will trouble the model and hurt classification. However, a model
can pay more attention to the shape with resolution information provided instead of objects’ size.
Second, it’s better for fine-grained classification. For example, it will be simple to distinguish a small
boat from a large warship.
Spatial resolution can also be used to filter mislabeled outliers in our dataset because intra-class
varieties of actual sizes for most categories are limited. Outliers can be found by selecting the objects
whose size is far different from those of the same category in a small range of spatial resolution.
9
4.4
Various pixel size of categories
Following the convention in [34], we refer to the height of a horizontal bounding box, which we call
pixel size for short, as a measurement for instance size. We divide all the instances in our dataset
into three splits according to their height of horizontal bounding box: small for range from 10 to 50,
middle for range from 50 to 300, and large for range above 300. Tab. 3 illustrates the percentages of
three instance splits in different datasets. It is clear that the PASCAL VOC dataset, NWPU VHR-10
dataset and DLR 3K Munich Vehicle dataset are dominated by middle instances, middle instances and
small instances, respectively. However, we achieve a good balance between small instances and middle
instances, which is more similar to real-world scenes and thus, helpful to better capture different size
of objects in practical applications.
It’s worth noting that pixel size varies in different categories. For example, a vehicle may be as
small as 30, however, a bridge can be as large as 1200, which is 40 times larger than a vehicle. The
huge differences among instances from different categories make the detection task more challenging
because models have to be flexible enough to handle extremely tiny and huge objects.
Dataset
PASCAL VOC
MSCOCO
NWPU VHR-10
DLR 3K Munich Vehicle
DOTA
10-50 pixel
0.14
0.43
0.15
0.93
0.57
50-300 pixel
0.61
0.49
0.83
0.07
0.41
above 300 pixel
0.25
0.08
0.02
0
0.02
Table 3: Comparison of instance size distribution of some datasets in aerial images and natural images.
4.5
Various aspect ratio of instances
Aspect ratio is an essential factor for anchor-based models, such as Faster RCNN [26] and YOLOv2 [25].
We count two kinds of aspect ratio for all the instances in our dataset to provide a reference for better model design: 1) Aspect ratio of minimally circumscribed horizontal rectangle bounding box, 2)
Aspect ratio of original quadrangle bounding box. Fig. 5 illustrates these two types of distribution
of aspect ratio for instances in our dataset. We can see that instances varies greatly in aspect ratio.
Moreover, there are a large number of instances with a large aspect ratio in our dataset.
4.6
Various instance density of images
It is common for aerial images to contain thousands of instances, which is different from natural
images. For example, images in ImageNet [6] contain on the average 2 categories and 2 instances,
while MSCOCO contains 3.5 categories and 7.7 instances, respectively. Our dataset is much richer in
instances per image, which can be up to 2000. Fig. 5 illustrates the number of instances in our DOTA
dataset.
With so many instances in a single image, it is unavoidable to see areas densely crowded with
instances. For COCO, instances are not annotated one by one because occlusion makes it difficult to
distinguish an instance from its neighboring instances. In these cases, the group of instances is marked
as one segment with attribute named “crowd”. However, this is not the case for aerial images because
there are rarely occlusion due to the perspective from the above. Therefore, we can annotate all the
10
instances in a dense area one by one. Fig. 4 shows examples of densely packed instances. Detecting
objects in these cases poses an enormous challenge for the current detection methods.
5
Evaluations
We evaluate the state of the art object detection methods on DOTA. For horizontal object detection,
we carefully select Faster R-CNN1 [26], R-FCN2 [4], YOLOv23 [25] and SSD2 [16] as our benchmark
testing algorithms for their excellent performance on general object detection. For oriented object
detection, we modify the original Faster R-CNN algorithm such that it can predict properly oriented
bounding boxes denoted as {(xi , yi ), i = 1, 2, 3, 4}.
Note that, the backbone networks are ResNet-101 [8] for R-FCN and Faster R-CNN, InceptionV2 [10] for SSD and customized GoogLeNet [27] for YOLOv2, respectively.
5.1
Tasks
To comprehensively evaluate the state of the art deep learning based detection methods on DOTA,
we propose two tasks, namely detection on horizontal bounding boxes (HBB for short) and detection
on oriented bounding boxes (OBB for short). To be more specific, we evaluate those methods on two
different kinds of ground truths, HBB or OBB, no matter how those methods were trained.
5.2
Evaluation prototypes
Images in DOTA are so large that they cannot be directly sent to CNN-based detectors. Therefore, we
crop a series of 1024 × 1024 patches from the original images with a stride set to 512. Note that some
complete objects may be cut into two parts during the cropping process. For convenience, we denote
the area of the original object as Ao , and the area of divided parts Pi , (i = 1, 2) as ai , (i = 1, 2). Then
we compute the parts areas over the original object area:
Ui =
ai
.
Ao
Finally, we label the part Pi with Ui < 0.7 as difficult and for the other one, we keep it the same as
the original annotation. For the vertices of the newly generated parts, we need to ensure they can be
described as an oriented bounding box with 4 vertices in the clockwise order with a fitting method.
In the testing phase, first we send the cropped image patches to obtain temporary results and then
we combine the results together to restore the detecting results on the original image. Finally, we
use non-maximum suppression (NMS) on these results based on the predicted classes. We keep the
threshold of NMS as 0.3 for the HBB experiments and 0.1 for the oriented experiments. In this way,
we indirectly train and test CNN-based models on DOTA.
For evaluation metrics, we adopt the same mAP calculation as for PASCAL VOC.
1
https://github.com/msracver/Deformable-ConvNets
https://github.com/tensorflow/models/tree/master/research/object_detection
3
https://github.com/pjreddie/darknet
2
11
Plane
BD
Bridge
GTF
SV
LV
Ship
TC
BC
ST
SBF
RA
Harbor
SP
HC
Avg.
YOLOv2 [25]
76.9
33.87
22.73
34.88
38.73
32.02
52.37
61.65
48.54
33.91
29.27
36.83
36.44
38.26
11.61
39.2
R-FCN [4]
79.33
44.26
36.58
53.53
39.38
34.15
47.29
45.66
47.74
65.84
37.92
44.23
47.23
50.64
34.9
47.24
FR-H [26]
80.32
77.55
32.86
68.13
53.66
52.49
50.04
90.41
75.05
59.59
57
49.81
61.69
56.46
41.85
60.46
SSD [16]
44.74
11.21
6.22
6.91
2
10.24
11.34
15.59
12.56
17.94
14.73
4.55
4.55
0.53
1.01
10.94
Table 4: Numerical results (AP) of baseline models evaluated with HBB ground truths. The short
names for categories are defined as: BD–Baseball diamond, GTF–Ground field track, SV–Small vehicle,
LV–Large vehicle, TC–Tennis court, BC–Basketball court, SC–Storage tank, SBF–Soccer-ball field,
RA–Roundabout, SP–Swimming pool, and HC–Helicopter. FR-H means Faster R-CNN [26] trained
on Horizontal bounding boxes. FR-O means Faster R-CNN [26] trained on Oriented bounding boxes.
Plane
BD
Bridge
GTF
SV
LV
Ship
TC
BC
ST
SBF
RA
Harbor
SP
HC
Avg.
YOLOv2 [25]
39.57
20.29
36.58
23.42
8.85
2.09
4.82
44.34
38.25
34.65
16.02
37.62
47.23
25.5
7.45
21.39
R-FCN [4]
37.8
38.21
3.64
37.26
6.74
2.60
5.59
22.85
46.93
66.04
33.37
47.15
10.6
25.19
17.96
26.79
SSD [16]
39.83
9.09
0.64
13.18
0.26
0.39
1.11
16.24
27.57
9.23
27.16
9.09
3.03
1.05
1.01
10.59
FR-H [26]
47.16
61
9.8
51.74
14.87
12.8
6.88
56.26
59.97
57.32
47.83
48.7
8.23
37.25
23.05
36.29
FR-O [26]
79.09
69.12
17.17
63.49
34.2
37.16
36.2
89.19
69.6
58.96
49.4
52.52
46.69
44.8
46.3
52.93
Table 5: Numerical results (AP) of baseline models evaluated with OBB ground truths. The short
names are defined the same as depicted in Tab. 4. Note that only FR-O [26] is trained with OBB.
5.3
Baselines with horizontal bounding boxes
Ground truths for HBB experiments are generated by calculating the axis-aligned bounding boxes
over original annotated bounding boxes. To make it fair, we keep all the experiments’ settings and
hyper parameters the same as depicted in corresponding papers [4, 16, 25, 26].
The experimental results of HBB prediction are shown in Tab. 4. Note that results of SSD is
much lower than other models. We suspect it should be attributed to the random crop operation in
SSD’s data augmentation strategies, which is quite useful in general object detection while degrades
in aerial object detection for tremendous small training instances. The results further indicate the
huge differences between aerial and general objects with respect to instance sizes.
12
5.4
Baselines with oriented bounding boxes
Prediction of OBB is difficult because the state of the art detection methods are not designed for
oriented objects. Therefore, we choose Faster R-CNN as the base framework for its accuracy and
efficiency and then modify it to predict oriented bounding boxes.
RoIs (Region of Interests) generated by RPN (Region Proposal Network) are rectangles which can
be written as R = (xmin , ymin , xmax , ymax ), for a more detailed interpretation, R = {(xi , yi ), i = 1, 2, 3, 4},
where x1 = x4 = xmin , x2 = x3 = xmax , y1 = y2 = ymin , y3 = y4 = ymax . In R-CNN procedure, each
RoI is attached to a ground truth oriented bounding box written as G = {(gxi , gyi ), i = 1, 2, 3, 4}.
Then R-CNN’s output target T = {(txi , tyi ), i = 1, 2, 3, 4} is calculated by following equations,
txi = (gxi − xi )/w,
(1)
tyi = (gyi − yi )/h,
(2)
where i = 1, 2, 3, 4, w = xmax − xmin , and h = ymax − ymin .
Other settings and hyper parameters are kept the same as depicted in Faster R-CNN [26]. The
numerical results are shown in Tab. 5. To make a comparison to our implemented Faster R-CNN for
OBB, we evaluate YOLOv2, R-FCN, SSD and Faster R-CNN trained on HBB with the OBB ground
truth. As shown in Tab.5, the results of those methods trained on HBB are much lower than Faster
R-CNN trained on OBB, indicating that for oriented object detection in aerial scenes, those methods
should be adjusted accordingly.
5.5
Experimental analysis
When analyzing the results exhibited in Table. 4, performances in categories like small vehicle, large
vehicle and ship are far from satisfactory, which attributes to their small size and densely crowded
locations in aerial images. As a contrast, large and discrete objects, like planes, swimming pools and
tennis courts, the performances are rather fair.
In Fig. 6, we compare the results between object detection experiments of HBB and OBB. For
densely packed and oriented objects shown in Fig. 6 (a) and (b), location precision of objects in HBB
experiments are much lower than OBB experiments and many results are suppressed through postprogress operations. So OBB regression is the correct way for oriented object detection that can be
really integrated to real applications. In Fig. 6 (c), large aspect ratio objects annotated in OBB style
like (harbor, bridge) are hard for current detectors to regress. But in HBB style, those objects usually
have normal aspect ratios and as a consequence, results seem to be fairly good as shown in Fig. 6 (d).
However in extremely dense scenes, e.g in Fig. 6 (e) and (f), results of HBB and OBB are all not
satisfying which implies the defects of current detectors.
6
Cross-dataset validations
The cross dataset generalization [28] is an evaluation for the generalization ability of a dataset. We
choose the UCAS-AOD dataset [40] to do cross-dataset generalization for its comparatively large
number of data comparing to other aerial object detection datasets. For there are no official data
splits for UCAS-AOD, we randomly select 1110 for training and 400 for testing. We choose YOLOv2
as the testing detector for all experiments described below and HBB-style annotations for all ground
13
(a)
(c)
(e)
(b)
(d)
(f)
Figure 6: Visualization results of testing on DOTA using well-trained Faster R-CNN. TOP and
Bottom respectively illustrate the results for HBB and OBB in cases of orientation, large aspect
ratio, and density.
truths. Input image size is changed to 960 × 544 around the original image sizes in UCAS-AOD while
other setting kept unchanged.
Results are shown in Tab. 6. The performance difference across two datasets is 35.8 for YOLOv2-A
and 15.6 for YOLOv2-D models, respectively. It suggests that DOTA hugely covers UCAS-AOD and
furthermore has more patterns and properties that are not shared in UCAS-AOD. And both models
get a low results on DOTA which reflects that DOTA is much more challenging.
Testing set
UCAS-AOD
DOTA
Detector
YOLOv2-A
YOLOv2-D
YOLOv2-A
YOLOv2-D
Plane
90.66
87.18
62.92
74.83
Small-vehicle
88.17
65.13
44.17
46.18
Avg.
89.41
76.15
53.55
60.51
Table 6: Results of cross-dataset generalization. Top: Detection performance evaluated on UCASAOD. Bottom: Detection performance evaluated on DOTA. YOLOv2-A and YOLOv2-D are
trained with UCAS-AOD and DOTA, respectively.
14
7
Conclusion
We build a large-scale dataset for oriented objects detection in aerial images which is much larger
than any existing datasets in this field. In contrast to general object detection benchmarks, we
annotate a huge number of well-distributed oriented objects with oriented bounding boxes. We assume
this dataset is challenging but very similar to natural aerial scenes, which are more appropriate for
practical applications. We also establish a benchmark for object detection in aerial images and show
the feasibility to produce oriented bounding boxes by modifying a mainstream detection algorithm.
Detecting densely packed small instances and extremely large instances with arbitrary orientations
in a large picture would be particularly meaningful and challenging. We believe DOTA will not only
promote the development of object detection algorithms in Earth Vision, but also pose interesting
algorithmic questions to general object detection in computer vision.
8
Acknowledgement
We thank Fan Hu, Pu Jin, Xinyi Tong, Xuan Hu, Zhipeng Dong, Liang Wu, Jun Tang, Linyan Cui,
Duoyou Zhou, Tengteng Huang, and all the others who involved in the annotations of DOTA.
References
[1] C. Benedek, X. Descombes, and J. Zerubia. Building development monitoring in multitemporal remotely
sensed image pairs with stochastic birth-death dynamics. IEEE TPAMI, 34(1):33–50, 2012.
[2] G. Cheng, P. Zhou, and J. Han. Learning rotation-invariant convolutional neural networks for object
detection in VHR optical remote sensing images. IEEE Trans. Geosci. Remote Sens., 54(12):7405–7415,
2016.
[3] G. Cheng, P. Zhou, and J. Han. Rifd-cnn: Rotation-invariant and fisher discriminative convolutional neural
networks for object detection. In CVPR, pages 2884–2893, 2016.
[4] J. Dai, Y. Li, K. He, and J. Sun. R-FCN: object detection via region-based fully convolutional networks.
In NIPS, pages 379–387, 2016.
[5] A.-M. de Oca, R. Bahmanyar, N. Nistor, and M. Datcu. Earth observation image semantic bias: A
collaborative user annotation approach. IEEE J. of Selected Topics in Applied Earth Observations and
Remote Sensing, 2017.
[6] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and F. Li. Imagenet: A large-scale hierarchical image database.
In CVPR, pages 248–255, 2009.
[7] M. Everingham, L. V. Gool, C. Williams, J. Winn, and A. Zisserman. The pascal visual object classes
(VOC) challenge. IJCV, 88(2):303–338, 2010.
[8] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, June 2016.
[9] G. Heitz and D. Koller. Learning spatial context: Using stuff to find things. In ECCV, pages 30–43, 2008.
[10] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal
covariate shift. CoRR, abs/1502.03167, 2015.
[11] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. K. Ghosh, A. D. Bagdanov, M. Iwamura, J. Matas,
L. Neumann, V. R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. ICDAR 2015 competition
on robust reading. In Proc. ICDAR, 2015.
[12] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y. Kalantidis, L.-J. Li, D. A.
Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123(1):32–73, 2017.
15
[13] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft
COCO: common objects in context. In ECCV, pages 740–755, 2014.
[14] Y. Lin, H. He, Z. Yin, and F. Chen. Rotation-invariant object detection in remote sensing images based
on radial-gradient angle. IEEE Geosci.Remote Sensing Lett., 12(4):746–750, 2015.
[15] K. Liu and G. Máttyus. Fast multiclass vehicle detection on aerial images. IEEE Geosci. Remote Sensing
Lett., 12(9):1938–1942, 2015.
[16] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, and A. C. Berg. SSD: single shot multibox
detector. In ECCV, pages 21–37, 2016.
[17] Z. Liu, H. Wang, L. Weng, and Y. Yang. Ship rotated bounding box space for ship extraction from
high-resolution optical satellite images with complex backgrounds. IEEE Geosci. Remote Sensing Lett.,
13(8):1074–1078, 2016.
[18] Y. Long, Y. Gong, Z. Xiao, and Q. Liu. Accurate object localization in remote sensing images based on
convolutional neural networks. IEEE Trans. Geosci. Remote Sens., 55(5):2486–2498, 2017.
[19] T. Moranduzzo and F. Melgani. Detecting cars in uav images with a catalog-based approach. IEEE Trans.
Geosci. Remote Sens., 52(10):6356–6367, 2014.
[20] T. N. Mundhenk, G. Konjevod, W. A. Sakla, and K. Boakye. A large contextual dataset for classification,
detection and counting of cars with deep learning. In ECCV, pages 785–800, 2016.
[21] A. Ö. Ok, Ç. Senaras, and B. Yüksel. Automated detection of arbitrarily shaped buildings in complex
environments from monocular VHR optical satellite imagery. IEEE Trans. Geosci. and Remote Sens.,
51(3-2):1701–1717, 2013.
[22] D. P. Papadopoulos, J. R. R. Uijlings, F. Keller, and V. Ferrari. Extreme clicking for efficient object
annotation. CoRR, abs/1708.02750, 2017.
[23] J. Porway, Q. Wang, and S. C. Zhu. A hierarchical and contextual model for aerial image parsing. IJCV,
88(2):254–283, 2010.
[24] S. Razakarivony and F. Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. J
Vis. Commun. Image R., 34:187–203, 2016.
[25] J. Redmon and A. Farhadi. YOLO9000: better, faster, stronger. CoRR, abs/1612.08242, 2016.
[26] S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN: towards real-time object detection with region
proposal networks. IEEE TPAMI, 39(6):1137–1149, 2017.
[27] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In CVPR, pages 1–9, 2015.
[28] A. Torralba and A. A. Efros. Unbiased look at dataset bias. In CVPR, pages 1521–1528, 2011.
[29] M. Vakalopoulou, K. Karantzalos, N. Komodakis, and N. Paragios. Building detection in very high resolution multispectral data with deep learning features. In IGARSS, pages 1873–1876, 2015.
[30] L. Wan, L. Zheng, H. Huo, and T. Fang. Affine invariant description and large-margin dimensionality
reduction for target detection in optical remote sensing images. IEEE Geosci. Remote Sensing Lett., 2017.
[31] G. Wang, X. Wang, B. Fan, and C. Pan. Feature extraction by rotation-invariant matrix representation
for object detection in aerial image. IEEE Geosci.Remote Sensing Lett., 2017.
[32] G. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu. AID: A benchmark data set for
performance evaluation of aerial scene classification. IEEE Trans. Geosci. Remote Sens., 55(7):3965–3981,
2017.
[33] J. Xiao, J. Hays, K. Ehinger, A. Oliva, and A. Torralba. SUN database: Large-scale scene recognition from
abbey to zoo. In CVPR, pages 3485–3492, 2010.
[34] S. Yang, P. Luo, C. C. Loy, and X. Tang. WIDER FACE: A face detection benchmark. In CVPR, pages
5525–5533, 2016.
[35] B. Yao, X. Yang, and S.-C. Zhu. Introduction to a large-scale general purpose ground truth database:
Methodology, annotation tool and benchmarks. In EMMCVPR 2007, pages 169–183, 2007.
16
[36] C. Yao, X. Bai, W. Liu, Y. Ma, and Z. Tu. Detecting texts of arbitrary orientations in natural images. In
CVPR, 2012.
[37] Q. You, J. Luo, H. Jin, and J. Yang. Building a large scale dataset for image emotion recognition: The
fine print and the benchmark. In AAAI, pages 308–314, 2016.
[38] F. Zhang, B. Du, L. Zhang, and M. Xu. Weakly supervised learning based on coupled convolutional neural
networks for aircraft detection. IEEE Trans. Geosci. Remote Sens., 54(9):5553–5563, 2016.
[39] B. Zhou, À. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. Learning deep features for scene recognition
using places database. In NIPS, pages 487–495, 2014.
[40] H. Zhu, X. Chen, W. Dai, K. Fu, Q. Ye, and J. Jiao. Orientation robust object detection in aerial images
using deep convolutional neural network. In ICIP, pages 3735–3739, 2015.
17
| 1cs.CV
|
1
A Comparison of SC-FDE and UW DFT-s-OFDM
for Millimeter Wave Communications
Alphan Şahin, Rui Yang, Frank La Sita, Robert L. Olesen
{Alphan.Sahin, Rui.Yang, Frank.LaSita, Robert.Olesen}@InterDigital.com
Abstract—In this study, we compare the single-carrier (SC)
waveform adopted in IEEE 802.11ad and unique word discrete
Fourier transform spread orthogonal frequency division
multiplexing (UW DFT-s-OFDM) waveform. We provide
equivalent representations of up-sampling and down-sampling
operations of the SC waveform by using discrete Fourier transform
(DFT) and inverse DFT to enable explicit comparison of these two
similar waveforms. By using this representation, we discuss why the
IEEE 802.11ad SC waveform can cause suboptimal performance in
multipath channel and discuss how to improve it with UW DFT-sOFDM. With comprehensive link-level simulations, we show that
replacing the 802.11ad SC waveform with UW DFT-spread OFDM
can result in 1 dB gain in peak throughput without affecting the
IEEE 802.11ad packet structure. We also evaluate the cross links
where the transmitter is UW-DFT-s-OFDM and the receiver is
traditional SC-FDE or vice versa. We demonstrate that UW DFTs-OFDM receiver can decode an IEEE 802.11ad SC waveform with
a slight SNR loss while IEEE 802.11ad SC receiver can decode a
UW DFT-spread OFDM waveform with an interference floor.
Keywords—5G, 802.11ay, 802.11ad, U W D F T - s - O F D M ,
FDE, millimeter wave, PAPR, SC, waveform
I. INTRODUCTION
Millimeter wave (mmWave) communication is considered
one of the key components of the Fifth Generation (5G) cellular
and Wi-Fi networks. With the availability of wide bandwidth at
higher gigahertz frequencies, mmWave communications can
achieve tens of gigabits per second data throughputs [1].
However, radio communications in high frequency bands may
suffer from several physical limitations such as severe path loss
and penetration loss. With narrow analog beams using advanced
beamforming technologies, the coverage range may be extended
depending on the maximum transmit power level. Hence, the
energy efficiency has become a key design criterion for the
mmWave communication networks.
At the physical layer, a waveform not only needs to enable
data to be transmitted with high spectral efficiency, but also
needs to address hardware challenges such as non-linear
distortion from the power amplifier (PA) and energy efficiency.
A waveform that has high peak-to-average power ratio (PAPR)
requires a large amount of power back-off at the input to the PA
to avoid non-linear signal distortion, which in turn reduces PA
efficiency and coverage area. Therefore, single carrier (SC)
schemes, which have a PAPR advantage over multicarrier
schemes, are some of the most appealing waveforms for
mmWave communications.
In the literature, there have been numerous studies that discuss
SC waveforms, particularly from the perspective of frequency
domain (FD) processing techniques at receiver (e.g., [2], [3] and
the references therein). For SC waveforms, low-complexity FD
processing techniques are typically enabled by the methods that
convert the linear convolution of the channel to a circular
convolution. For example, extending the data symbol block with
a sequence (i.e., known as unique word (UW) [2], pseudo noise
extension [3], or training sequence [4]), or cyclically extending
the data blocks before pulse shaping are the traditional
approaches that convert the impact of channel on data blocks to
a circular convolution at the receiver for SC systems. Among
these methods, it has been shown that extending the data block
with a sequence can be more beneficial than the cyclic extension
as it enables various non-linear FD equalization techniques [3],
time-frequency synchronization [4], channel estimation, phase
tracking, and noise estimation [5] for SC waveforms. Due to
these advantages, SC with UW, adopted in IEEE 802.11ad [6],
is a promising waveform for mmWave communications. It is
worth noting that the concept of UW is also applied to OFDM at
the expense of higher transmitter and receiver complexity as
compared to cyclic prefix (CP) OFDM [7].
Historically, a major innovation for SC systems occurred
through the introduction of a DFT precoder in FD to an OFDM
waveform [8] [9], known as cyclic prefix discrete Fourier
transform spread orthogonal frequency division multiplexing
(CP DFT-s-OFDM). One can show that the data symbols are
convolved with a Dirichlet sinc function circularly in time with
CP DFT-s-OFDM [10], which is similar to SC waveforms.
However, since CP DFT-s-OFDM is also a precoded OFDM
symbol, it enables efficient frequency domain multiple access
(FDMA) by eliminating the guard bands required for the one
with traditional SC waveforms. Due to its PAPR advantage
compared to OFDM and its excellent efficieny in FDMA, it was
adopted for the use in the LTE uplink [11] (The reader is also
refered to the discussions on CP DFT-s-OFDM and OFDM in
[12]). Since its adoption, there have been various notable studies
on DFT-s-OFDM and its variants. For example, the spacefrequency block code (SFBC) scheme which preserves the PAPR
advantage of DFT-s-OFDM is proposed in [13]. In [14] [15] [16],
the CP extension of CP DFT-s-OFDM is eliminated by padding
zeros to the beginning and the last part of the data block. In [17]
and [18], a scheme which is called UW DFT-s-OFDM is
proposed by showing that the zeros can be replaced by a set of
fixed symbols, i.e., UW sequence, as done in the traditional SC
waveform. In this scheme, the last parts of the UW DFT-sOFDM symbols, i.e., tails, are approximately the same. After the
UW DFT-s-OFDM symbols are concatenated in time, the tail
part of the 𝑖th symbol serves as the CP for the (𝑖 + 1)th UW
2
DFT-s-OFDM symbol. Hence, no additional CP is required for
the (𝑖 + 1)th UW DFT-s-OFDM symbol. In addition, various
cancellation signals are proposed to improve the tail
characteristics of UW DFT-s-OFDM by mitigating the leakage
from the data symbols at the tail in [17] and [18]. In [10], a lowcomplexity suppression is proposed by applying a windowing to
the output of DFT block in FD.
In this paper, in contrast to previous studies on SC and UW
DFT-s-OFDM, we compare the SC waveform adopted in IEEE
802.11ad [6] and UW DFT-s-OFDM by expressing equivalent
operations for the up-sampling and down-sampling blocks of the
SC transmitter and receiver by using DFT and inverse DFT. By
using this representation, we show the followings:
1) The synthesis of the SC waveform is shown using blockbased operations, indicative of the compatibility of the
numerology for UW DFT-s-OFDM to the numerology of
the SC adopted in IEEE 802.11ad. Hence, UW DFT-sOFDM does not change the packet structure which in turn
leads to a framework without less standardization effort.
2) The down-sampling at the SC receiver causes incoherent
combinations in the FD under a fading channel, which
causes degradation at the receiver. By providing peak
throughput, we show that UW DFT-s-OFDM provides 1
dB gain as compared to the SC waveform.
3) We discuss FD windowing in [10] for UW DFT-s-OFDM
and develop a two-stage equalization to address the noncoherent additions to improve error rate performance.
4) We demonstrate that the UW DFT-s-OFDM receiver can
decode the IEEE 802.11ad SC waveform with a slight
SNR loss while the IEEE 802.11ad SC receiver can
decode UW DFT-s-OFDM with an interference floor.
The rest of the paper is organized as follows. In Section II, we
provide system models for SC and UW DFT-s-OFDM. In
Section III, we discuss the representation of the SC transmitter
and receiver while comparing them to that for UW DFT-sOFDM. In Section IV, we present numerical results performed
under IEEE 802.11ad framework. We conclude the paper in
Section V with final remarks.
Notations: Matrices [columns vectors] are denoted with upper
[lower] case boldface letters (e.g., 𝐀 and [𝐚]). The Hermitian and
the transpose operations are denoted by (⋅)H and (⋅)T ,
respectively. The conjugate of 𝐚 is denoted by 𝐚̅. The symbols ∗
and ⨂ denote linear convolution and Kronecker product,
respectively. The ℓ2 -norm of 𝐚 is denoted by ‖𝐚‖2 . The
operation flip{𝐚}, the operation circ{𝐚, 𝜏}, and the operation
diag{𝐚} reverse the order of the elements of 𝐚, circularly shift 𝐚
by 𝜏, and create a diagonal matrix where its diagonal is 𝐚,
respectively. The set of complex, real, and integer numbers are
shown as ℂ, ℝ, and ℤ, respectively. ℤ+ symbolizes the set of
positive integers. The multivariate complex Gaussian
distribution is denoted by 𝒞𝒩(𝛍, 𝐂), where 𝛍 is the mean vector
and 𝐂 is the covariance matrix. 𝐈𝑁 , 𝟎𝑁×𝑀 , 𝟏𝑁×𝑀 are the 𝑁×𝑁
identity, 𝑁×𝑀 zero, and 𝑁×𝑀 one matrices, respectively.
II. SYSTEM MODEL
A. Single Carrier Waveform
We consider an SC packet based on IEEE 802.11ad [6] which
consists of multiple SC blocks. Each SC block is generated from
a group of data and fixed symbols via up-sampling, filtering, and
down-sampling operations. Let data and fixed symbols
transmitted within the 𝑖th SC block be the elements of vector
𝐝𝑖 ∈ ℂ𝑀d ×1 and 𝐬 ∈ ℂ𝑀s ×1 , respectively, where 𝑀d and 𝑀s are
the number of data symbols and fixed symbols, respectively. The
fixed symbols are transmitted before the data symbols to enable
frequency domain equalization (FDE). After up-sampling,
filtering, and down-sampling, the 𝑖th SC block of the SC packet
𝐱 𝑖 ∈ ℂ𝑇tx×1 can be represented as
𝐬
𝐱 𝑖 = √𝑏 ↓𝑏 𝐏tx ↑𝑎 [𝐝 ],
𝑖
(1)
where 𝑎 ∈ ℤ+ is the up-sampling factor, ↑𝑎 ∈ ℝ𝐾×𝑀 with 𝑀 =
𝑀s + 𝑀d and 𝐾 = 𝑎×𝑀 is the up-sampling matrix, 𝐏tx ∈ ℂ𝑃tx×𝐾
with 𝑃tx = 𝐾 + 𝐿tx − 1 is the convolution matrix which applies
an 𝐿tx -tap pulse shaping vector denoted by 𝐟 ∈ ℂ𝐿tx×1 with
‖𝐟‖2 = 1, and ↓𝑏 ∈ ℝ𝑇tx×𝑃tx is the down-sampling matrix, where
𝑏 ∈ ℤ+ is the down-sampling factor. Without loss of generality,
we assume that 𝑏 ≤ 𝑎 and 𝑇tx = 𝑃tx /𝑏 = (𝑎×𝑀 + 𝐿tx − 1)/𝑏 ∈
ℤ+ , i.e., 𝑃tx is an integer multiple of 𝑏. Typically, 𝑎 and 𝑏 are
chosen such that the radio operates above the critical sample rate
while the choice of 𝑎 and 𝑏 may be used to address hardware
issues. For example, IEEE 802.11ad [6] suggests 𝑎/𝑏 to be 3/2
so that both OFDM and SC can operate at the same sample rate.
To model the packet based on the block-based description
above, we adopt the overlap-add (OA) method without loss of
generality, which is an efficient way to calculate the convolution
of a long signal [19]. After the SC blocks are generated via (1),
the last 𝑂tx samples of the 𝑖th SC block are overlapped with the
first 𝑂tx samples of the (𝑖 + 1)th SC block and summed, as
illustrated in Fig. 1a. To maintain the alignment between the
overlapping samples of SC blocks and simplify the notation, we
assume that 𝑁 ≜ 𝑎/𝑏×𝑀 ∈ ℤ+ , which leads that 𝑂tx = 𝑇tx −
𝑁 = (𝐿tx − 1)/𝑏 ∈ ℤ+ and 𝑀 is the integer multiple of 𝑏.
Let the channel impulse response (CIR) between the
transmitter and the receiver be a vector 𝐡 = [ℎ0 ℎ1 ⋯ hℒ ] where
ℒ + 1 is the number of taps for the multipath channel (MPC).
After the SC packet passes through the MPC, we assume that the
SC receiver discards the first 𝑁g samples from the received SC
packet for the sake of FDE. The receiver then prepares the epochs
with the length of 𝑇rx = 𝑁 + 𝐿rx for block processing, where 𝐿rx
is the length of receive filter denoted by 𝐠 ∈ ℂ𝐿rx×1 . The 𝑖th SC
epoch 𝐞𝑖 ∈ ℂ𝑇rx×1 can be expressed as
𝐞𝑖 = [𝟎𝑇rx ×𝑁g
𝐈𝑇tx +ℒ−𝑁g
𝟎𝑁−𝑁g×𝑇rx−𝑁+𝑁g
𝐲𝑖
𝟎𝑇rx ×ℒ−𝑞+𝑁−𝑁g ] [𝐲 ],
𝟎𝑞−ℒ+𝑁g×𝑇tx +ℒ−𝑁g 𝐈 𝑇rx −𝑁+𝑁g
𝑖+1
(2)
and 𝐲𝑙 = 𝐇𝐿 𝐱 𝑖 where 𝐇L ∈ ℂ(𝑇tx+ℒ)×𝑇tx is the linear channel
convolution matrix, 𝐲𝑖 is the 𝑖th SC block 𝐱 𝑖 alone after passing
through the channel, 𝑞 = 𝑇rx − 𝑇tx . To ensure that 𝐞𝑖 does not
3
Data
𝐬 𝑀s
/
↑𝑎
𝐝𝑖𝑀d
/
𝐾
𝐏tx
𝑃tx
↓𝑏
𝑇tx
𝐱𝑖
ℒ 𝐱2
𝑂tx
𝑁g
𝑀h 𝐡
𝐲2
𝑇rx
𝑂rx
𝐞1
↑𝑏
𝑃rx
H
𝐏rx
𝐾
↓𝑎
𝑀
𝐞
𝐅𝑀
𝑀
Equalization
Seq.
𝐲1
𝐱1
𝑀
𝑀d 𝐝𝑖
𝐅𝑀H
𝑀t 𝐭
Seq. H
/
Data
/
Seq. T
/
a) SC waveform.
Data
Seq. T
/
/
/
𝐡 𝑀h
𝐝𝑖𝑀d
𝑀h 𝐡
Tail Head Data
𝐅𝑀
𝑀
𝐌
𝐭 𝑀t
𝑁
𝐅𝑁H
𝑁
𝐱𝑖
𝑁
𝑁c
𝐞1
𝐅𝑁
𝑁
𝐌H
Equalization
Seq. H
𝑀
𝐞
𝑀
𝐅𝑀H
/
𝑀d 𝐝𝑖
/
𝑀t 𝐭
/
Seq. H
Data
Seq. T
b) UW DFT-s-OFDM.
Fig. 1. The transmit and receive block diagrams for SC waveform with FDE support and UW DFT-s-OFDM.
contain samples related to 𝐝𝑖−1 and 𝐝𝑖+1 , 𝑁g satisfies the
following constraint:
𝑂tx + ℒ ≤ 𝑁g ≤ 𝑎/𝑏×𝑀s − 𝑂rx ,
(3)
where 𝑂rx = (𝐿rx − 1)/𝑏 ∈ ℤ+ . The upper bound in (3) is
because the SC receiver utilizes the first 𝑂rx samples of the 𝑖th
SC epoch 𝐞𝑖 for receive filtering. However, this constraint
reduces the room for the MPC by 𝑂rx samples. To avoid intersymbol interference (ISI), 𝑀s needs to satisfy the condition given
by
𝑀s 𝑎 ≥ 𝐿tx + 𝐿rx + (ℒ𝑏 + 1) − 3.
(4)
It is also worth emphasizing that the 𝑖th SC epoch 𝐞𝑖 is a
cyclically extended block where the block size is 𝑁 and the cyclic
extension size is 𝑂rx samples. This is because the identical fixed
sequences are transmitted before and after the data vector 𝐝𝑖 . To
maintain the same structure for the last SC epoch, an additional
fixed sequence 𝐬 also needs to be transmitted at the end of the
packet, as shown in Fig. 1a. This is why IEEE 802.11ad SC PHY
attaches an extra Golay sequence to the end of the payload [6].
Based on the discussions above, the overall FDE operations on
the 𝑖th SC epoch can be represented as
̃H
[𝐡
𝐝̃H
𝑖
H
↑𝑏 (𝐞𝑖 + 𝐧𝑖 ),
𝐭̃ H ]H = √𝑏 𝐅
⏟𝑀H 𝐄𝐅𝑀 ↓𝑎 𝐏rx
(5)
FDE
where 𝐧𝑖 ∈ ℂ𝑇×1 ∼ 𝒞𝒩(𝟎 𝑇rx×1 , 𝜎n2 𝐈 𝑇rx ) is the additive white
Gaussian noise (AWGN) with variance 𝜎n2 , 𝐅𝑀 is the 𝑀-point
normalized discrete Fourier transform (DFT) matrix, 𝐄 is a
̃ ∈ ℂ𝑀h×1 and 𝐭̃ ∈
diagonal matrix for the FDE operation, and 𝐡
𝑀t ×1
ℂ
are the estimates of the head and tail sequences, i.e., 𝐭 and
𝐡, respectively, where 𝐬 = [𝐭 T 𝐡T ]T and 𝑀s = 𝑀h + 𝑀t . Based
on the minimum mean square error (MMSE) criterion, the
equalizer matrix 𝐄 can be derived as (𝐐H 𝐐 + 𝜎n2 𝐈𝑀 )−1 𝐐H where
𝐡effective
𝐐 = diag {√𝑀𝐅𝑀 ↓𝑎 circ {[𝟎
] , 𝜏0 }}
𝑀𝑎−𝐿tx −𝐿rx +2−𝑏ℒ−𝑏×1
determines the position of data symbols after the equalization in
(5) by circularly shifting the original vector [𝐬 T 𝐝T ]T by −𝑀t .
The variable 𝑀h will be utilized to identify the similarities
between SC and UW DFT-s-OFDM in the following sections.
B. UW DFT-s-OFDM
The block diagram for UW DFT-s-OFDM [10] is given in
Fig. 1b. At the transmitter, the data sequence 𝐝 and the fixed
sequences, i.e., 𝐡 and t, are first mapped to the inputs of a DFT
matrix 𝐅𝑀 . To generate fixed samples at the tail part of each UW
DFT-s-OFDM block, the upper-end and the lower-end of DFTspread, i.e., the first 𝑀h and last 𝑀t columns of 𝐅𝑀 , respectively,
are allocated for 𝐡 and 𝐭 since the footprints of 𝐡 and 𝐭 in time
corresponds to the head and tail of each UW DFT-s-OFDM
symbol with this allocation, respectively [10]. The output of 𝐅𝑀
is then mapped to a set of 𝑀 contiguous subcarriers in the FD via
a mapping matrix 𝐌 ∈ ℝ𝑁×𝑀 . The spread data and fixed symbols
in frequency are converted to the time domain via 𝐅𝑁H as
𝐱 𝑖 = 𝐅𝑁H 𝐌𝐅𝑀 [𝐡H
𝐝H
𝑖
𝐭 H ]H ,
(7)
where 𝐱 𝑖 ∈ ℂ𝑁×1 is 𝑖th UW DFT-s-OFDM symbol. 𝑁CP samples
are added only for the first symbol to be able to equalize it with
FDE.
Providing that 𝑀t ≥ 𝑀/𝑁×ℒ [18], the 𝑖th epoch 𝐞𝑖 after the
MPC can approximately be represented as
𝐞𝑖 ≅ 𝐇C 𝐱 𝑖 ,
(8)
𝑁×𝑁
where 𝐇C ∈ ℂ
is a circular channel convolution matrix. The
receiver operation can finally be expressed as
̃H
[𝐡
𝐝̃H
𝑖
𝐭̃ H ]H = 𝐅𝑀H 𝐌 H 𝐄𝐅𝑁 (𝐞𝑖 + 𝐧𝑖 ),
(9)
where 𝐧𝑖 ∈ ℂ𝑁×1 ∼ 𝒞𝒩(𝟎𝑁×1 , 𝜎n2 𝐈𝑁 ) and 𝐄 can be derived
based on the MMSE criterion as (𝐐H 𝐐 + 𝜎n2 𝐈𝑀 )−1 𝐐H and 𝐐 =
diag{√𝑁𝐅𝑁 [𝐡T 𝟎1×𝑁−ℒ−1 ]T }.
(6)
in which 𝜏0 = (𝑀s − 𝑀h )𝑎 − 𝑁g 𝑏 − 𝐿rx + 1 and 𝐡effective =
{𝒇} ∗ {𝑮} ∗ {↑𝑏 𝐡} . Note that 𝜏0 in (6) is a function of 𝑀h which
III. SINGLE CARRIER VERSUS UW DFT-S-OFDM
In this section, we shed light on the structural differences and
similarities, and the performance differences between the SC and
4
Seq. H
Data
Seq.T
(a)
(b)
(c)
/
𝐾
↑𝑎
𝐝𝑖𝑀d
𝐏tx
𝑃tx
𝐞𝟏
𝑇rx
𝑇tx
𝐱𝑖
↓𝑏
𝑁t
𝐡 𝑀h
/
𝐝𝑖𝑀d
/
𝐅𝑀
𝑀
w
w
(d)
w
w
w
H
𝐏rx
𝐾
↓𝑎
𝑀
𝑁
𝑁
𝑀
𝑁
𝑁
𝑀
(b)
(c)
(b)
𝚲
𝑁
(c)
w
𝐅H
(d) 𝑁
𝑁
z
z
+
𝑁
𝐦
𝑃rx
𝑀
𝑀
z
𝑁
𝐦𝟏
↑𝑏
𝑀
(a)
𝐭 𝑀t
/
𝑀h 𝐡
𝐞
z
z
z
𝑁
𝐱𝑖
𝑁
𝐅𝑁
𝑁
(a)
(b)
(c)
(d)
𝑀
𝑀
(d)
𝑀
w
w
w
𝑀
𝐅𝑀H
𝑀d 𝐝𝑖
𝑀t 𝐭
𝑀h 𝐡
(a)
𝑁
bin
bin
bin
bin
bin
𝐅𝑀
Equalization
Data
/
𝑁g
𝐬 𝑀s
Equalization
Seq.
𝑀
𝐅𝑀H
𝑀t 𝐭
z
w
z
+
w
𝑀d 𝐝𝑖
z
z
z
+
/
/
/
/
/
/
Seq. H
Data
Seq. T
Seq. H
Data
Seq. T
bin
bin
bin
bin
bin
Fig. 2. Equivalent representation of SC waveform with DFT and inverse DFT blocks, and an example for 𝑎 = 3 and 𝑏 = 2.
UW DFT-s-OFDM waveforms. To this end, we represent the upsampling and down-sampling operations at the SC transmitter
and receiver by using DFT and inverse DFT. We show why SC
is suboptimal in MPC as compared to UW DFT-s-OFDM under
MMSE FDE. We also discuss the windowing in FD and twostage equalization for UW DFT-s-OFDM.
A. Up-sampling and Down-sampling at SC Transmitter
Let the consecutive epochs of length 𝑁 at the output of the SC
transmitter after 𝑁t samples from the beginning of the SC packet
that are discarded be 𝐦𝑖 ∈ ℂ𝑁×1 , where 𝑂tx ≤ 𝑁t ≤ 𝑎/𝑏×𝑀s .
The 𝑖th epoch 𝐦𝑖 ∈ ℂ𝑁×1 can also be calculated as
𝐦𝑖 = [𝟎𝑁×𝑁t
𝐈 𝑇tx−𝑁t
𝟎𝑁−𝑁t×𝑁t
𝐱𝑖
𝟎𝑁×Ttx−𝑁t ] [𝐱 ].
𝟎𝑁−𝑇tx+𝑁t×𝑇tx−𝑁𝑡
𝐈𝑁t
𝑖+1
(10)
Since the last 𝑁t 𝑏 elements of the corresponding vectors that
generate 𝐦𝑖 before linear filtering and down-sampling
operations are identical when 𝑂tx ≤ 𝑁t ≤ 𝑎/𝑏×𝑀s , one can
express linear pulse shaping as a circular pulse shaping
operation. Hence, 𝐦𝑖 can be expressed as
𝐦𝑖 = √𝑏 ↓𝑏 𝐂tx ↑𝑎 [𝐡H
𝐝H
𝑖
𝐭 H ]H ,
(11)
where 𝐂tx ∈ ℂ𝑎𝑀×𝑎𝑀 is a circular convolution matrix where the
T
first column of 𝐂tx is circ {[𝐟 H 𝟎T𝐾−𝐿tx+1×1 ] , 𝜏1 } and 𝜏1 =
(𝑀s − 𝑀h )𝑎 − 𝑁t 𝑏. We then exploit the following properties
related to the up-sampling and down-sampling operations to
represent 𝐦𝑖 with DFT-based operations:
Property 1 (Up-sampling): Let 𝐱 ∈ ℂ𝑀×1 be a vector of size 𝑀 ∈
ℤ and let 𝐲 ∈ ℂ𝑎𝑀×1 be another vector of size 𝑎𝑀 ∈ ℤ where 𝑚th
element of 𝐲 is equal to 𝑘th element of 𝐱 where 𝑘 = (𝑚/𝑎) ∈ ℤ,
otherwise it is 0. Then,
𝐲 =↑𝑎 𝐱 =
1
√𝑎
H (𝟏
𝐅𝑎M
𝑎×1 ⨂𝐈𝑀 )𝐅𝑀 𝐱.
(12)
Property 2 (Down-sampling): Let 𝐱 ∈ ℂ𝑏𝑁×1 be a vector of size
𝑏𝑁 ∈ ℤ and let 𝐲 ∈ ℂ𝑁×1 be another vector of size 𝑁 ∈ ℤ where
𝑚th element of 𝐲 is equal to 𝑘th element of 𝐱 where 𝑘 = 𝑚𝑏 ∈
ℤ. Then,
1
𝐅𝑁H (𝟏1×𝑏 ⨂𝐈𝑁 )𝐅𝑏𝑁 𝐱.
(13)
√𝑏
The proofs for Property 1 and Property 2 can be found in [20].
By considering the fact that any circulant matrix can be
H
decomposed
as
𝐂tx = 𝐅𝑎𝑀
𝚲𝐅𝑎𝑀
where
𝚲=
𝐲 =↓𝑏 𝐱 =
diag {√𝑎𝑀𝐅𝑎𝑀 circ {[𝐟 H
T
𝟎T𝑀𝑎−𝐿tx+1×1 ] , 𝜏1 }}, and employing
the identities given in Property 1 and Property 2, (11) can be
rewritten as
𝐦𝑖 = 𝐅𝑁H
(𝟏1×𝑏 ⨂𝐈𝑁 )𝚲(𝟏𝑎×1 ⨂𝐈𝑀 )
√𝑎
𝐅𝑀 [𝐡H
𝐝H
𝑖
𝐭 H ]H . (14)
By using (14), the following interpretations of the SC transmitter
can be made:
1) Block-based Implementation of SC packet: (14) shows that
the SC packet can be generated by simply concatenating {𝐦𝑖 |𝑖 =
1,2, … } synthesized through DFT-based operations. Note that
this method is simpler than the overlap-save method which
would require (𝑁 + 𝐿tx − 1)-point inverse DFT [19] [20].
However, (14) allows for an 𝑁-point inverse DFT by exploiting
the existence of the UW transmitted before and after 𝐝𝑖 .
2)Difference between SC and UW DFT-s-OFDM Transmitter:
(14) reveals the connection between the SC and UW DFT-sOFDM waveforms. 𝚲 corresponds to a windowing operation in
frequency, where the window is determined by the frequency
response of filter 𝐟 and which windows the repeated output of the
𝑭M with a factor of 𝑎. Since 𝐟 is typically a low-pass filter, the
filter primarily selects one of the repetitions. As illustrated in Fig.
2, the pass bandwidth of the filter typically is large enough to
cover all information (shown as 𝑤 and 𝑧 in the illustration, where
we consider 𝑁 frequency bins due to 𝐅𝑁H ) with some roll-off in
spectrum. However, since the impulse response of the filter in
practice is finite in the time domain, the frequency response of
the filter is spread across the bins. On the other hand, if one
designs 𝚲 such that it applies a rectangular function, the
5
corresponding transmit diagram in Fig. 2 is identical to the UW
DFT-s-OFDM transmitter described in Section II.B. To remove
phase rotation in frequency (which also means a symmetric and
real filter), the condition on 𝑁t given by
Condition 1 (Transmitting SC packet with UW DFT-s-OFDM):
𝜏1 = (𝑀s − 𝑀h )𝑎 − 𝑁t 𝑏 = −(𝐿tx − 1)/2
(15)
must also hold true. For example, if we consider 802.11ad
numerology as 𝑀s = 64, 𝑎 = 3, and 𝑏 = 2, and assume that
𝐿tx = 67, 𝑁t must equal to (225 − 3𝑀h )/2 to remove the phase
rotation in frequency. When (15) is satisfied, the diagonal
elements of 𝚲 in (14) become real numbers and 𝐦𝑖 is
approximately a UW DFT-s-OFDM symbol, but not exactly the
same due to the additional shaping in frequency. This also
implies that the SC receiver can demodulate the UW DFT-sOFDM packet, but that the data symbols are slightly interfered
due to the SC approximation. Note that it is straightforward to
limit the support for the diagonal elements of 𝚲 with the FD
interpretation of SC. However, it is not possible to achieve to the
same 𝐦𝑖 with the time-domain implementation of SC since
limiting the support for the diagonal elements of 𝚲 results in a
filter length of 𝑀𝑎, which causes ISI due to (4).
3) Interference due to Down-sampling at Transmitter: (14)
shows the distortion at the SC transmitter due to down-sampling
when 𝑎/𝑏 ∉ ℤ+ . The operation after shaping in (14) divides the
resulting vector into 𝑏 subgroups and overlaps them. The
overlapping is not harmful as 𝑎/𝑏 ≥ 1, i.e., above the Nyquist
rate. However, the frequency response of the filter 𝐟 is spread
across the bins in practice and the components on the side lobes
may still cause interference to the components on the main lobe.
Therefore, when 𝑎/𝑏 ∉ ℤ+ , the receiver may not perfectly
recover the symbols with single-tap MMSE FDE due to the small
aliasing components even under a noiseless condition. For
example, Fig. 2 illustrates the aliasing due to the non-zero
components (from 𝑤 to 𝑧 or vice versa) on side lobes after downsampling for 𝑎 = 3 and 𝑏 = 2 in step (d) at the transmitter side.
Note that this imperfection can be mitigated with a well-designed
filter, and the down-sampling operation can still be considered
as not harmful since it reduces the sample rate of the device.
B. Down-sampling and Up-sampling at SC Receiver
At the receiver side, we exploit the fact that the 𝑖th SC epoch
𝐞𝑖 is a cyclically extended vector by 𝑂rx samples, which
maintains circular convolution of the receive filtering on the last
𝑁 samples of 𝑖th SC epoch 𝐞𝑖 , to modify (5) as
̃T
[𝐡
𝐝̃T𝑖
H
↑𝑏 𝐞̇ 𝑖 .
𝐭̃ T ]T = √𝑏𝐅𝑀H 𝐄𝐅𝑀 ↓𝑎 𝐂rx
(16)
where 𝐞̇ 𝑖 is the last 𝑁 samples of 𝑖th SC epoch 𝐞𝑖 , 𝐂rx ∈ ℂ𝑎𝑀×𝑎𝑀
is a circular convolution matrix and the first column of 𝐂rx is
circ {[𝐠 H
T
H
𝟎T𝐾−𝐿rx+1×1 ] , }. Since 𝐂rx = 𝐅𝑎𝑀
𝐅𝑎𝑀 , where
diag {√𝑎𝑀𝐅𝑎𝑀 circ {[𝐠 H
rewritten as
T
=
𝟎T𝑀𝑎−𝐿rx+1×1 ] , }}, (16) can then be
̃H
[𝐡
𝐝̃H
𝑖
𝐭̃ H ]H = 𝐅𝑀H 𝐄
(𝟏1×a ⨂𝐈𝑀 ) (𝟏𝑏×1 ⨂𝐈𝑁 )
√𝑎
𝐅N 𝐦𝑖 ,
(17)
by using Property 1 and Property 2. By inferring (17), the
following interpretations on the SC receiver can be made:
1) Non-coherent additions in Fading Channel: Assuming that
the filter is a Nyquist filter and the filter is known at the receiver
side, one may apply a matched filter to the SC epochs. In that
case, it is well-known that (16) is the optimum receiver in an
AWGN channel in the sense that it maximizes SNR. The
equivalent operation in FD is that the receiver applies an optimal
weight to each bin before changing the dimension of the upsampled vector of the DFT of 𝐞̇ 𝑖 before the overlapping
operation. On the other hand, (17) loses its optimality in a fading
channel. Since the DFT of 𝐞̇ 𝑖 includes the impact of both
frequency responses of the transmit filter and MPC via point-topoint multiplications in frequency, the frequency response of the
receive filter clearly does not provide the optimal weights as the
channel coefficients in frequency are arbitrary complex numbers.
Thus, the combination in (17), i.e., 𝟏1×a ⨂𝑰𝑀 , cause incoherent
combinations as the phase of each bin is not aligned. For UW
DFT-s-OFDM, there is no incoherent combination before
equalization due to the perfect rectangular filter in FD. This
observation leads to two-stage equalization given in Section III.C
when FD windowing applied to UW DFT-s-OFDM.
2) Decoding SC packet with UW DFT-s-OFDM Receiver:
(17) shows the relationship between the SC receiver and the UW
DFT-s-OFDM receiver. When a UW DFT-s-OFDM receiver
receive an SC block, the UW DFT-s-OFDM receiver equalizes
one of the aliases (weighted by the transmit filter) and the discard
the rests of the information based on (9). On the other hand, the
traditional SC receiver combines the aliases before FDE as in
(17). This means that the UW DFT-s-OFDM receiver can receive
an SC waveform without any interference, but with some
degradation due to discarding other weighted aliases.
C. Two-stage FDE: First-Phase-Then-Amplitude
The SC waveform yields better PAPR results than the UW
DFT-s-OFDM waveform when it employs an FIR filter with high
roll-off factor [10]. However, because of the finite support of the
filter in time, the SC waveform spreads the information in
frequency. On the other hand, the information is localized in the
FD with UW DFT-s-OFDM. Therefore, UW DFT-s-OFDM
maintains the orthogonality between the frequency-domain
resources, which may be important in certain scenarios, e.g.,
accessing different channels with different beamforming gains in
the uplink. To keep the benefit of both SC and UW DFT-sOFDM, one approach is to apply FD shaping [11] to UW DFTs-OFDM, so called UW DFT-s-W-OFDM [10]. In this scheme,
the support of the diagonal elements of 𝚲 in (14) is still limited
as in UW DFT-s-OFDM, but it is extended, i.e., larger than 𝑀,
and smoothed with a FD shaping function. This operation
modifies the Dirichlet sinc kernel of UW DFT-s-OFDM to
reduce the PAPR of UW DFT-s-OFDM. In addition, the UW
DFT-s-OFDM packet approximates an SC waveform further
6
Fig. 3. Temporal characteristics (MSE: -21 dB)
Fig. 4. PAPR performance.
with this operation under the condition of (12). To take the
energy in the aliases into account, we propose a simple
equalization strategy so called two-staged equalization. In the
first stage of the equalization, the phase distortion due to the
channel is equalized in the FD (i.e., phase-first). The aliases are
then combined based on their SNRs (note that SNR for each
frequency bin differs due to the channel frequency response and
the windowing in the FD at transmitter). Since the phases are
corrected before the combination, a coherent combination is
achieved. In the second stage, the amplitude distortion is
equalized before the DFT de-spread operation (i.e., then
amplitude). Two-stage FDE can also be applied to the SC
waveform as long as one considers the FD interpretation of the
SC receiver in the implementation. Note that UW DFT-s-OFDM
removes the need of prior knowledge of the transmit filter at the
receiver side (i.e., the filter is basically a rectangular function in
frequency) and a two-stage FDE.
IV. NUMERICAL RESULTS
In this section, we compare the SC and UW DFT-s-OFDM
waveforms numerically. We consider the numerology adopted in
IEEE 802.11ad for the SC waveform, where 𝑀 = 512, 𝑀s = 64,
𝑎 = 3, and 𝑏 = 2. We assume that the SC transmitter employs
an 𝐿tx = 67 tap root-raised cosine (RRC) filter with roll-off
Fig. 5. BER performance in AWGN channel.
Fig. 6. BER performance in fading channel.
factor of 𝜌 = { , .2, .3} and the receiver uses a matched filter,
i.e., 𝐿rx = 67. For UW DFT-s-OFDM, we assume that 𝑀ℎ = 9.
For MPC, we consider the IEEE 802.11ad channel model [21],
where the scenario is a conference room (STA-AP as subscenario). We assume that the transmitter and receiver employ
8×1 phase antenna arrays (PAAs), unless otherwise stated.
In Fig. 3, we provide temporal characteristics of the UW DFTs-OFDM and SC waveforms where 𝜌 = for BPSK. By using
the condition given in (15), we find that 𝑁t = (225 − 3𝑀h )/2 =
99. By offsetting 𝑁t = 99 samples from the beginning of the SC
packet, we plot the UW DFT-s-OFDM waveform on top of SC
waveform. As shown in Fig. 3, the UW DFT-s-OFDM and SC
waveforms are approximately the same, where the sample MSE
between two waveforms is -21 dB. The difference is due to the
imperfect rectangular shape of the SC filter as given in Fig. 3.
In Fig. 4, we provide the PAPR results for QPSK and 16QAM.
The SC waveform is 2 dB and 1 dB better than UW DFT-sOFDM for QSPK and 16QAM, respectively, when 𝜌 = .2.
However, when the FD windowing is applied to UW DFT-sOFDM, where the amount of the extensions are 51 bins on left
and right side of the signal bandwidth and an RRC filter is
employed (i.e., corresponds to 0.2 roll-off for UW DFT-s-WOFDM), the PAPR performance significantly improved and it is
almost identical to that of the SC waveform.
7
Fig. 7. Throughput with SC and UW DFT-s-OFDM.
In Fig. 5, we investigate the BER performance in AWGN
when there is only one antenna element at both transmit and
receive PAAs. As expected, UW DFT-s-OFDM, DFT-s-WOFDM, and SC waveforms achieve the Gaussian BER bound.
On the other hand, when there is a mismatch between the
transmitter and receiver structures, the BER performance
degrades. If the transmitter is for SC and the receiver is for UW
DFT-s-OFDM, the performance degrades because of not
accounting for the energy on the aliases. When the transmitter is
for UW DFT-s-OFDM and the receiver is for SC, the degradation
in BER is because the SC receiver cannot perfectly model the
Dirichlet sinc kernel.
In Fig. 6 and Fig. 7, we investigate the BER and the peak
throughput performance for MCS indices 4, 8, 10, 11, and 12
under an 802.11ad channel model. We consider two-stage FDE
for UW DFT-s-W-OFDM. Fig. 6 shows that the SC waveform
loses its optimality in a fading channel due to the incoherent
additions in FD, as discussed in Section III.B. The performance
degradation for SC is about 1 dB, as compared to UW DFT-sOFDM. Fig. 6 also shows the cross-link performances; while
UW DFT-s-OFDM decodes the SC waveform successfully with
the degradation of 0.1 dB, SC receiver cannot perfectly decode
UW DFT-s-OFDM and suffers from an error floor due to the
Dirichlet sinc function as investigated in Section III.A and
Section III.B. The peak throughput results in Fig. 7 shows that
UW DFT-s-OFDM achieves 1 dB gain as compared to the SC
waveform, which indicates that UW DFT-s-OFDM is superior to
SC in peak throughput performance in mmWave channels.
V. CONCLUDING REMARKS
In this study, we give insights on the similarities and
differences between the SC and UW DFT-s-OFDM waveforms
by re-expressing the up-sampling and down-sampling operations
of the IEEE 802.11ad SC waveform by using DFT and inverse
DFT. This representation reveals that IEEE 802.11ad SC
waveform and UW DFT-s-OFDM are structure-wise similar to
each other and the IEEE 802.11ad packet structure does not
change for UW DFT-s-OFDM. On the other hand, it shows that
the SC receiver with down-sampling operation yields suboptimal
error rate performance in a fading channel since the downsampling operation before equalization causes incoherent
additions in frequency. The peak throughput results also show
that the use of the UW DFT-s-OFDM waveform provides a 1dB
gain over the current IEEE 802.11ad SC waveform and the UW
DFT-s-OFDM receiver achieves this gain while providing
backwards compatibility when receiving IEEE 802.11ad SC
waveforms. With cross-link analysis, we show that a UW DFTs-OFDM receiver can decode an SC waveform with only a slight
performance degradation whereas an SC receiver suffers from an
error floor due to the Dirichlet Sinc kernel of UW DFT-s-OFDM.
We also investigate FD windowing for UW DFT-s-OFDM and
two-stage equalization which equalizes the phase and amplitude
in separate stages to achieve coherent additions in FD for UW
DFT-s-W-OFDM.
REFERENCES
[1]
NGMN Alliance, "5G
https://www.ngmn.org.
White
Paper,"
17
Feb.
2015.
[Online].
Available:
[2]
D. Falconer, S. L. Ariyavisitakul, A. Benyamin-Seeyar and B. Eidson, "Frequency Domain
Equalization for Single-Carrier Broadband Wireless Systems," IEEE Communications
Magazine, vol. 40, no. 4, pp. 58-66, Apr. 2002.
[3]
N. Benvenuto, R. Dinis, D. Falconer and S. Tomasin, "Single Carrier Modulation With
Nonlinear Frequency Domain Equalization: An Idea Whose Time Has Come - Again,"
Proceedings of the IEEE, vol. 98, no. 1, pp. 69-96, 2010.
[4]
L. Deneire, B. Gyselinckx and M. Engels, "Training sequence versus cyclic - A new look
on single carrier communication," IEEE Commun. Lett., vol. 5, pp. 292-294, 2001.
[5]
J. Coon, M. Beach, M. McGeehan and M. Sandell, "Channel and Noise Variance Estimation
and Tracking Algorithms for Unique-Word Based Single-Carrier Systems," 2003.
[6]
"Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications," 2016.
[7]
M. Huemer, C. Hofbauer, A. Onic and J. B. Huber, "Design and analysis of UW-OFDM
signals," Elsevier International Journal of Electronics and Communications, vol. 68, no.
10, pp. 958-968, 2014.
[8]
H. Sari, G. Karam and I. Jeanclaud, "Frequency-domain equalization of mobile radio and
terrestrial broadcast channels," in IEEE GLOBECOM, San Francisco, 1994.
[9]
H. Sari, G. Karam and I. Jeanclaude, "Transmission techniques for digital terrestrial TV
broadcasting," IEEE Communications Magazine, vol. 33, no. 2, pp. 100-109, 1995.
[10] A. Sahin, R. Yang, E. Bala, M. C. Beluri and R. L. Olesen, "Flexible DFT-S-OFDM:
Solutions and Challenges," IEEE Communications Magazine, vol. 54, no. 11, pp. 106-112,
11 2016.
[11] S. Sesia, I. Toufik and M. Baker, LTE - The UMTS Long Term Evolution: From Theory to
Practice, 2nd Edition, Wiley, 2011.
[12] C. Ciochina and H. Sari, "A Review of OFDMA and Single-Carrier FDMA," in IEEE
European Wireless Conference (EW), Lucco, 2010.
[13] C. Ciochina, D. Castelain, D. Mottier and H. Sari, "New PAPR-preserving mapping
methods for single-carrier FDMA with space-frequency block codes," IEEE Transactions
on Wireless Communications, vol. 8, no. 10, pp. 5176-5186, Oct. 2009.
[14] G. Berardinelli, F. Frederiksen, T. B. Sorensen, P. Mogensen and K. Pajukoski, "Zero-tail
DFT-spread-OFDM signals," in IEEE Globecom Workshops (GC Wkshps), Atlanta, 2013.
[15] G. Berardinelli, K. I. Pedersen, T. B. Sorensen and P. Mogensen, "Generalized DFTSpread-OFDM as 5G Waveform," IEEE Communications Magazine, vol. 54, no. 11, pp.
99-105, 11 2016.
[16] T. S. Rappaport, R. W. Heath, R. C. Daniels and J. N. Murdock, Millimeter Wave Wireless
Communications, Prentice Hall, 2017.
[17] F. Hasegawa, S. Shinjo and A. Okazaki, "Static sequence assisted out-of-band power
suppression for DFT-s-OFDM," in IEEE Personal, Indoor, and Mobile Radio
Communications (PIMRC), Hong Kong, 2015.
[18] A. Sahin, R. Yang, M. Ghosh and R. L. Olesen, "An Improved Unique Word DFT-Spread
OFDM Scheme for 5G Systems," in EEE Globecom Workshops (GC Wkshps), San Diego,
2015.
[19] J. G. Proakis and D. G. Manolakis, Digital Signal Processing Principles, Algorithms, and
Applications, Fourth edition ed., Pearson Education, Inc., 2007.
[20] K. R. Rao, D. N. Kim and J. -J. Hwang, Fast Fourier Transform - Algorithms and
Applications, 1st ed., Springer Publishing Company, Incorporated, 2010.
[21] A. Maltsev, et al. "Channel Models for 60 GHz WLAN Systems," May 2010.
| 7cs.IT
|
Reverse Curriculum Generation
for Reinforcement Learning
arXiv:1707.05300v2 [cs.AI] 17 Oct 2017
Carlos Florensa
UC Berkeley
[email protected]
David Held
UC Berkeley
[email protected]
Markus Wulfmeier
Oxford Robotics Institute
[email protected]
Michael Zhang
UC Berkeley
[email protected]
Pieter Abbeel
OpenAI
UC Berkeley
ICSI
[email protected]
Abstract: Many relevant tasks require an agent to reach a certain state, or to
manipulate objects into a desired configuration. For example, we might want
a robot to align and assemble a gear onto an axle or insert and turn a key in a
lock. These goal-oriented tasks present a considerable challenge for reinforcement
learning, since their natural reward function is sparse and prohibitive amounts
of exploration are required to reach the goal and receive some learning signal.
Past approaches tackle these problems by exploiting expert demonstrations or by
manually designing a task-specific reward shaping function to guide the learning
agent. Instead, we propose a method to learn these tasks without requiring any
prior knowledge other than obtaining a single state in which the task is achieved.
The robot is trained in “reverse”, gradually learning to reach the goal from a set
of start states increasingly far from the goal. Our method automatically generates a curriculum of start states that adapts to the agent’s performance, leading
to efficient training on goal-oriented tasks. We demonstrate our approach on difficult simulated navigation and fine-grained manipulation problems, not solvable
by state-of-the-art reinforcement learning methods.
Keywords: Reinforcement Learning, Robotic Manipulation, Automatic Curriculum Generation
1
Introduction
Reinforcement Learning (RL) is a powerful learning technique for training an agent to optimize a
reward function. Reinforcement learning has been demonstrated on complex tasks such as locomotion [1], Atari games [2], racing games [3], and robotic manipulation tasks [4]. However, there are
many tasks for which it is hard to design a reward function such that it is both easy to maximize
and yields the desired behavior once optimized. An ubiquitous example is a goal-oriented task; for
such tasks, the natural reward function is usually sparse, giving a binary reward only when the task
is completed [5]. This sparse reward can create difficulties for learning-based approaches [6]; on the
other hand, non-sparse reward functions for such tasks might lead to undesired behaviors [7].
For example, suppose we want a seven DOF robotic arm to learn how to align and assemble a gear
onto an axle or place a ring onto a peg, as shown in Fig. 1c. The complex and precise motion required
to align the ring at the top of the peg and then slide it to the bottom of the peg makes learning highly
impractical if a binary reward is used. On the other hand, using a reward function based on the
distance between the center of the ring and the bottom of the peg leads to learning a policy that
places the ring next to the peg, and the agent never learns that it needs to first lift the ring over the
top of the peg and carefully insert it. Shaping the reward function [8] to efficiently guide the policy
towards the desired solution often requires considerable human expert effort and experimentation
to find the correct shaping function for each task. Another source of prior knowledge is the use of
demonstrations, but it requires an expert intervention.
1st Conference on Robot Learning (CoRL 2017), Mountain View, United States.
In our work, we avoid all reward engineering or use of demonstrations by exploiting two key insights.
First, it is easier to reach the goal from states nearby the goal, or from states nearby where the agent
already knows how to reach the goal. Second, applying random actions from one such state leads
the agent to new feasible nearby states, from where it is not too much harder to reach the goal. This
can be understood as requiring a minimum degree of reversibility, which is usually satisfied in many
robotic manipulation tasks like assembly and manufacturing.
We take advantage of these insights to develop a “reverse learning” approach for solving such difficult manipulation tasks. The robot is first trained to reach the goal from start states nearby a given
goal state. Then, leveraging that knowledge, the robot is trained to solve the task from increasingly
distant start states. All start states are automatically generated by executing short random walk from
the previous start states that got some reward but still require more training. This method of learning
in reverse, or growing outwards from the goal, is inspired by dynamic programming methods like
value iteration, where the solutions to easier sub-problems are used to compute the solution to harder
problems.
In this paper, we present an efficient and principled framework for performing such “reverse learning.” Our method automatically generates a curriculum of initial positions from which to learn to
achieve the task. This curriculum constantly adapts to the learning agent by observing its performance at each step of the training process. Our method requires no prior knowledge of the task other
than providing a single state that achieves the task (i.e. is at the goal). The contributions of this paper
include:
• Formalizing a novel problem definition of finding the optimal start-state distribution at
every training step to maximize the overall learning speed.
• A novel and practical approach for sampling a start state distribution that varies over the
course of training, leading to an automatic curriculum of start state distributions.
• Empirical experiments showing that our approach solves difficult tasks like navigation or
fine-grained robotic manipulation, not solvable by state-of-the-art learning methods.
2
Related Work
Curriculum-based approaches with manually designed schedules have been explored in supervised
learning [9, 10, 11, 12] to split particularly complex tasks into smaller, easier-to-solve sub-problems.
One particular type of curriculum learning explicitly enables the learner to reject examples which it
currently considers too hard [13, 14]. This type of adaptive curriculum has mainly been applied
for supervised tasks and most practical curriculum approaches in RL rely on pre-specified task
sequences [15, 16]. Some very general frameworks have been proposed to generate increasingly
hard problems [17], although few implementations concretize the idea and only tackle preliminary
tasks [18]. A similar line of work uses intrinsic motivation based on learning progress to obtain
“developmental trajectories” that focus on increasingly harder tasks [19]. Nevertheless, their method
requires iteratively partitioning the full task space, which strongly limits the application to fine-grain
manipulation tasks like the ones presented in our work (see detailed analysis on easier tasks in [5]).
More recent work in curriculum for RL assumes baseline performances for several tasks are given,
and it uses them to gauge which tasks are the hardest (furthest behind the baseline) and require more
training [20]. However, this framework can only handle finite sets of tasks and requires each task to
be learnable on its own. On the other hand, our method trains a policy that generalizes to a set of
continuously parameterized tasks, and it is shown to perform well even under sparse rewards by not
allocating training effort to tasks that are too hard for the current performance of the agent.
Closer to our method of adaptively generating the tasks to train on, an interesting asymmetric selfplay strategy has recently been proposed [21]. Contrary to our approach, which aims to generate
and train on all tasks that are at the appropriate level of difficulty, the asymmetric component of
their method can lead to biased exploration concentrating on only a subset of the tasks that are at
the appropriate level of difficulty, as the authors and our own experiments suggests. This problem
and their time-oriented metric of hardness may lead to poor performance in continuous state-action
spaces, which are typical in robotics. Furthermore, their approach is designed as an exploration
bonus for a single target task; in contrast, we define a new problem of efficiently optimizing a policy
across a range of start states, which is considered relevant to improve generalization [22].
2
Our approach can be understood as sequentially composing locally stabilizing controllers by growing a tree of stabilized trajectories backwards from the goal state, similar to work done by Tedrake
et al. [23]. This can be viewed as a “funnel” which takes start states to the goal state via a series of
locally valid policies [24]. Unlike these methods, our approach does not require any dynamic model
of the system. An RL counterpart, closer to our approach, is the work by Bagnell et al. [25], where
a policy search algorithm in the spirit of traditional dynamic programming methods is proposed to
learn a non-stationary policy: they learn what should be done in the last time-step and then “back
it up” to learn the previous time-step and so on. Nevertheless, they require the stronger assumption
of having access to baseline distributions that approximate the optimal state-distribution at every
time-step.
The idea of directly influencing the start state distribution to accelerate learning in a Markov Decision Process (MDP) has already drawn attention in the past. Kakade and Langford [26] studied the
idea of exploiting the access to a ‘generative model’ [27] that allows training the policy on a fixed
‘restart distribution’ different from the one originally specified by the MDP. If properly chosen, this
is proven to improve the policy training and final performance on the original start state distribution.
Nevertheless, no practical procedure is given to choose this new distribution (only suggesting to use
a more uniform distribution over states, which is what our baseline does), and they don’t consider
adapting the start state distribution during training, as we do. Other researchers have proposed to
use expert demonstrations to improve learning of model-free RL algorithms, either by modifying the
start state distribution to be uniform among states visited by the provided trajectories [7], or biasing
the exploration towards relevant regions [28]. Our method works without any expert demonstrations,
so we do not compare against these lines of research.
3
Problem Definition
We consider the general problem of learning a policy that leads a system into a specified goal-space,
from any start state sampled from a given distribution. In this section we first briefly introduce the
general reinforcement learning framework and then we formally define our problem statement.
3.1
Preliminaries
We define a discrete-time finite-horizon Markov decision process (MDP) by a tuple M =
(S, A, P, r, ρ0 , T ), in which S is a state set, A an action set, P : S × A × S → R+ is a transition probability distribution, r : S × A → R is a bounded reward function, ρ0 : S → R+ is an
start state distribution, and T is the horizon. Our aim is to learn a stochastic policy πθ : S ×A → R+
parametrized by θ that maximizes objective the expected return, ηρ0 (πθ ) = Es0 ∼ρ0 R(π, s0 ). We
PT
denote by R(π, s0 ) := Eτ |s0 [ t=0 r(st , at )] the expected reward starting when starting from a
s0 ∼ ρ0 , where τ = (s0 , a0 , , . . . , aT −1 , sT ) denotes a whole trajectory, with at ∼ πθ (at |st ), and
st+1 ∼ P(st+1 |st , at ). Policy search methods iteratively collect trajectories on-policy (i.e. sampling
from the above distributions ρ0 , πθ , P) and use them to improve the current policy [29, 30, 31].
In our work we propose to instead use a different start-state distribution ρi at every training iteration
i such that the learning speed. Learning progress is still evaluated based on the original distribution
ρ0 . Convergence of ρi to ρ0 is desirable but not required as an optimal policy πi? under a start
distribution ρi is also optimal under any other ρ0 , as long as their support coincide. In the case of
approximately optimal policies under ρi , bounds on the performance under ρ0 can be derived [26].
3.2
Goal-oriented tasks
We consider the general problem of reaching a certain goal space S g ⊂ S from any start state in
S 0 ⊂ S. This simple, high-level description can be translated
into an MDP without further domain
knowledge by using a binary reward function r(st ) = 1 st ∈ S g and a uniform distribution over
the start states ρ0 = Unif(S 0 ). We terminate the episode when the goal is reached. This implies
that the return R(π, s0 ) associated with every start state s0 is the probability of reaching the goal at
some time-step t ∈ {0 . . . T }.
T
T
n[
o
[
R(π, s0 ) = Eπ(·|st ) 1
st ∈ S g |s0 = P
st ∈ S g π, s0
(1)
t=0
t=0
3
As advocated by Rajeswaran et al. [22], it is important to be able to train an agent to achieve the goal
from a large set of start states S 0 . An agent trained in this way would be much more robust than an
agent that is trained from just a single start state, as it could recover from undesired deviations from
the intended trajectory. Therefore, we choose the set of start states S 0 to be all the feasible points
in a wide area around the goal. On the other hand, the goal space S g for our robotics fine-grained
manipulation tasks is defined to be a small set of states around the desired configuration (e.g. key in
the key-hole, or ring at the bottom of the peg, as described in Sec. 5).
As discussed above, the sparsity of this reward function makes learning extremely difficult for RL
algorithms [6, 32, 33], and approaches like reward shaping [8] are difficult and time-consuming to
engineer for each task. In the following subsection we introduce three assumptions, and the rest of
the paper describes how we can leverage these assumptions to efficiently learn to achieve complex
goal-oriented tasks directly from sparse reward functions.
3.3
Assumptions for reverse curriculum generation
In this work we study how to exploit three assumptions that hold true in a wide range of practical
learning problems (especially if learned in simulation):
Assumption 1 We can arbitrarily reset the agent into any start state s0 ∈ S at the beginning of all
trajectories.
Assumption 2 At least one state sg is provided such that sg ∈ S g .
Assumption 3 The Markov Chain induced by taking uniformly sampled random actions has a communicating class1 including all start states S 0 and the given goal state sg .
The first assumption has been considered previously (e.g. access to a generative model in Kearns
et al. [27]) and is deemed to be a considerably weaker assumption than having access to the full
transition model of the MDP. The use of Assumption 1 to improve the learning in MDPs that require
large exploration has already been demonstrated by Kakade and Langford [26]. Nevertheless, they
do not propose a concrete procedure to choose a distribution ρ from which to sample the start states
in order to maximally improve on the objective in Eq. (1). In our case, combining Assumption 1
with Assumption 2 we are able to reset the state to sg , which is critical in our method to initialize
the start state distribution to concentrate around the goal space at the beginning of learning.For the
second assumption, note that we only assume access to one state sg in the goal region; we do not
require a description of the full region nor trajectories leading to it. Finally, Assumption 3 ensures
that the goal can be reached from any of the relevant start states, and that those start states can
also be reached from the goal; this assumption is satisfied by many robotic problems of interest,
as long as there are no major irreversibilities in the system. In the next sections we detail our
automatic curriculum generation method based on continuously adapting the start state distribution
to the current performance of the policy. We demonstrate the value of this method for challenging
robotic manipulation tasks.
4
Methodology
In a wide range of goal-oriented RL problems, reaching the goal from an overwhelming majority
of start states in S 0 requires a prohibitive amount of on-policy or undirected exploration. On the
other hand, it is usually easy for the learning agent (i.e. our current policy πi ) to reach the goal S g
from states nearby a goal state sg ∈ S g . Therefore, learning from these states will be fast because
the agent will perceive a strong signal, even under the indicator reward introduced in Section 3.2.
Once the agent knows how to reach the goal from these nearby states, it can train from even further
states and bootstrap its already acquired knowledge. This reverse expansion is inspired by classical
RL methods like Value Iteration or Policy Iteration [34], although in our case we do not assume
knowledge of the transition model and our environments have high-dimensional continuous actionstate spaces. In the following subsections we propose a method that leverages the Assumptions
from the previous section and the idea of reverse expansion to automatically adapt the start state
distribution, generating a curriculum of start state distributions that can be used to tackle problems
unsolvable by standard RL methods.
1
A communicating class is a maximal set of states C such that every pair of states in C communicates with
each other. Two states communicate if there is a non-zero probability of reaching one from the other.
4
Algorithm 1: Policy Training
Input : π0 , sg , ρ0 , Nnew , Nold , Rmin , Rmax , Iter
Output: Policy πN
startsold ← [sg ];
starts, rews ← [sg ], [1];
for i ← 1 to Iter do
starts ← SampleNearby(starts, Nnew );
starts.append[sample(startsold , Nold )];
ρi ← Unif(starts);
πi , rews ← train pol(ρi , πi−1 );
starts ← select(starts, rews, Rmin , Rmax );
startsold .append[starts];
evaluate(πi , ρ0 );
end
4.1
Procedure 2: SampleNearby
Input : starts, Nnew , Σ, TB , M
Output: startsnew
while len(starts) < M do
s0 ∼ Unif(starts);
for t ← 1 to TB do
at = t , t ∼ N (0, Σ);
st ∼ P(st |st−1 , at );
starts.append(st );
end
end
startsnew ←
sample(starts, Nnew )
Policy Optimization with modified start state distribution
Policy gradient strategies are well suited for robotic tasks with continuous and high dimensional
action-spaces [35]. Nevertheless, applying them directly on the original MDP does poorly in tasks
with sparse rewards and long horizons like our challenging manipulation tasks. If the goal is not
reached from the start states in S 0 , no reward is received, and the policy cannot improve. Therefore,
we propose to adapt the distribution ρi from where start states s0 are sampled to train policy πi .
Analogously to Held et al. [5], we postulate that in goal-oriented environments, a strong learning
signal is obtained when training on start states s0 ∼ ρi from where the agent reaches the goal
sometimes, but not always. We call these start states “good starts”. More formally, at training
iteration i, we would like to sample from ρi = Unif(Si0 ) where Si0 = {s0 : Rmin < R(πi , s0 ) <
Rmax }. The hyper-parameters Rmin and Rmax are easy to tune due to their interpretation as bounds
on the probability of success, derived from Eq. (1). Unfortunately, sampling uniformly from Si0
is intractable. Nevertheless, at least at the beginning of training, states nearby a goal state sg are
more likely to be in Si0 . Then, after some iterations of training on these start states, some will
0
be completely mastered (i.e. Rmax > Rmax and no longer in Si+1
), but others will still need
more training. To find more “good starts”, we follow the same reasoning: the states nearby these
0
0
remaining s ∈ Si+1
are likely to also be in Si+1
. In the rest of the section we describe an effective
way of sampling feasible nearby states and we layout the full algorithm.
4.2
Sampling “nearby” feasible states
For robotic manipulation tasks with complex contacts and constraints, applying noise in state-space
s0 = s + , ∼ N may yield many unfeasible start states s0 . For example, even small random
perturbations of the joint angles of a seven degree-of-freedom arm generate large modifications to
the end-effector position, potentially placing it in an infeasible state that intersects with surrounding
objects. For this reason, the concept of “nearby” states might be unrelated to the Euclidean distance
ks0 − sk2 between these states. Instead, we have to understand proximity in terms of how likely it
is to reach one state from the other by taking actions in the MDP.
Therefore, we choose to generate new states s0 from a certain seed state s by applying noise in action
space. This means we exploit Assumption 1 to reset the system to state s, and from there we execute
short “Brownian motion” rollouts of horizon TB taking actions at+1 = t with t ∼ N (0, Σ). This
method of generating “nearby” states is detailed in Procedure 2. The total sampled states M should
be large enough such that the Nnew desired states startsnew , obtained by subsampling, extend in all
directions around the input states starts. All states visited during the rollouts are guaranteed to be
feasible and can then be used as start states to keep training the policy.
4.3
Detailed Algorithm
Our generic algorithm is detailed in Algorithm 1. We first initialize the policy with π0 and the
“good start” states list starts with the given goal state sg . Then we perform Iter training iterations
of our RL algorithm of choice train pol. In our case we perform 5 iterations of Trust Region
Policy Optimization (TRPO) [36] but any on-policy method could be used. At every iteration, we
set the start state distribution ρi to be uniform over a list of start states obtained by sampling Nnew
5
start states from nearby the ones in our “good starts” list starts (see SampleNearby in previous
section), and Nold start states from our replay buffer of previous “good starts” startsold . As already
shown by Held et al. [5], the replay buffer is an important feature to avoid catastrophic forgetting.
Technically, to check which of the states s0 ∈ starts are in Si0 (i.e. the “good starts”) we should
execute some trajectories from each of those states to estimate the expected returns R(s0 , πi−1 ),
but this considerably increases the sample complexity. Instead, we use the trajectories collected by
train pol to estimate R(πi−1 , s0 ) and save it in the list rews. These are used to select the
“good” start states for the next iteration - picking the ones with Rmin ≤ R(πi−1 , s0 ) ≤ Rmax . We
found this heuristic to give a good enough estimate and not drastically decrease learning performance
of the overall algorithm.
Our method keeps expanding the region of the state-space from which the policy can reach the goal
reliably. It samples more heavily nearby the start states that need more training to be mastered and
avoiding start states that are yet too far to receive any reward under the current policy. Now, thanks to
Assumption 3, the Brownian motion used to generate further and further start states will eventually
reach all start states in S 0 , and therefore our method improves the metric ηρ0 defined in Sec. 3.1 (see
also Sec. A.2 for a practical implementation).
5
Experimental Results
We investigate the following questions in our experiments:
• Does the performance of the policy on the target start state distribution ρ0 improve by training
on distributions ρi growing from the goal?
• Does focusing the training on “good starts” speed up learning?
• Is Brownian motion a good way to generate “good starts” from previous “good starts”?
We use the below task settings to explore these questions. All are implemented in MuJoCo [37] and
the hyperparameters used in our experiments are described in Appendix A.1.
(a) Point-mass maze task
(c) Ring on Peg task
(b) Ant maze task
(d) Key insertion task
Figure 1: Task images. Source code and videos of the performance obtained by our algorithm are
available here: http://bit.ly/reversecurriculum
Point-mass maze: (Fig. 1a) A point-mass agent (orange) must navigate within 30cm of the goal
position (4m, 4m) at the end of a G-shaped maze (red). The target start state distribution from
which we seek to reach the goal is uniform over all feasible (x, y) positions in the maze.
Ant maze: (Fig. 1b) A quadruped robot (orange) must navigate its Center of Mass to within 50cm
of the goal position (0m, 4m) at the end of a U-shaped maze (red). The target start state distribution
from which we seek to reach the goal is uniform over all feasible ant positions inside the maze.
Ring on Peg: (Fig. 1c) A 7 DOF robot must learn to place a “ring” (actually a square disk with a
hole in the middle) on top of a tight-fitting round peg. The task is complete when the ring is within
3 cm of the bottom of the 15 cm tall peg. The target start state distribution from which we seek to
reach the goal is uniform over all feasible joint positions for which the center of the ring is within
40 cm of the bottom of the peg.
Key insertion: (Fig. 1d) A 7 DOF robot must learn to insert a key into a key-hole. The task is
completed when the distance between three reference points at the extremities of the key and its
corresponding targets is below 3cm. In order to reach the target, the robot must first insert the key
at a specific orientation, then rotate it 90 degrees clockwise, push forward, then rotate 90 degrees
counterclockwise. The target start state distribution from which we seek to reach the goal is uniform
over all feasible joint positions such that the tip of the key is within 40 cm of key-hole.
6
5.1
Effect of start state distribution
In Figure 2, the Uniform Sampling (baseline) red curves show the average return of policies learned
with TRPO without modifying the start state distribution. The green and blue curves correspond
to our method and an ablation, both exploiting the idea of modifying the start state distribution at
every learning iterations. These approaches perform consistently better across the board. In the
case of the point-mass maze navigation task in Fig. 2a, we observe that Uniform Sampling has a
very high variance because some policies only learn how to perform well from one side of the goal
(see Appendix B.2 for a thorough analysis). The Ant-maze experiments in Fig. 2b also show a
considerable slow-down of the learning speed when using plain TRPO, although the effect is less
drastic as the start state distribution ρ0 is over a smaller space.
In the more complex manipulation tasks shown in Fig. 2c-2d, we see that the probability of reaching
the goal with Uniform Sampling is around 10% for the ring task and 2% for the key task. These
success probabilities correspond to reliably reaching the goal only from very nearby positions: when
the ring is already on the peg or when the key is initialized very close to the final position. None of
the learned policies trained on the original ρ0 learn to reach the goal from more distant start states.
On the other hand, our methods do succeed at reaching the goal from a wide range of far away start
states. The underlying RL training algorithm and the evaluation metric are the same. We conclude
that training on a different start state distribution ρi can improve training or even allow it at all.
(a) Point-mass Maze task
(b) Ant Maze task
(c) Ring on Peg task
(d) Key insertion task
Figure 2: Learning curves for goal-oriented tasks (mean and variance over 5 random seeds).
5.2
Effect of “good starts”
In Figure 2 we see how applying our Algorithm 1 to modify the start state distribution considerably
improves learning (Brownian on Good Starts, in green) and final performance on the original MDP.
Two elements are involved in this improvement: first, the backwards expansion from the goal, and
second, the concentration of training efforts on “good starts”. To test the relevance of this second
element, we ablate our method by running our SampleNearby Procedure 2 on all states from
which the policy was trained in the previous iteration. In other words, the select function in
Algorithm 1 is replaced by the identity, returning all starts independently of the rewards rews they
obtained during the last training iteration. The resulting algorithm performance is shown as the
Brownian from All Starts blue curve in Figures 2. As expected, this method is still better than
not modifying the start state distribution but has a slower learning than running SampleNearby
around the estimated good starts.
7
Now we evaluate an upper bound of the benefit provided by our idea of sampling “good starts”. As
mentioned in Sec. 4.1, we would ideally like to sample start states from ρi = Unif(Si0 ), but it is
0
intractable. Instead, we evaluate states in Si−1
, and we use Brownian motion to find nearby states,
0
to approximate Si . We can evaluate how much this approximation hinders learning by exhaustively
sampling states in the lower dimensional point-mass maze task. To do so, at every iteration we can
samples states s0 uniformly from the state-space S, empirically estimates their return R(s0 , πi ), and
rejects the ones that are not in the set Si0 = {s0 : Rmin < R(πi , s0 ) < Rmax }. This exhaustive
sampling method is orders of magnitude more expensive in terms of sample complexity, so it would
not be of practical use. In particular, we can only run it in the easier point-mass maze task. Its
performance is shown in the brown curve of Fig. 2a, called “Oracle (rejection sampling)”; training
on states sampled in such a manner further improves the learning rate and final performance. Thus
0
we can see that our approximation of using states in Si−1
to find states in Si0 leads to some loss in
performance, at the benefit of a greatly reduced computation time.
Finally, we compare to another way of generating start states based on the asymmetric self-play
method of Sukhbaatar et al. [38]. The basic idea is to train another policy, “Alice”, that proposes start
states to the learning policy, “Bob”. As can be seen, this method performs very poorly in the pointmass maze task, and our investigation shows that “Alice” often gets stuck in a local optimum, leading
to poor start states suggestions for “Bob”. In the original paper, the method was demonstrated only
on discrete action spaces, in which a multi-modal distribution for Alice can be maintained; even in
such settings, the authors observed that Alice can easily get stuck in local optima. This problem is
exacerbated when moving to continuous action spaces defined by a unimodal Gaussian distribution.
See a detailed analysis of these failure modes in Appendix B.3.
5.3
Brownian motion to generate “nearby” states
Here we evaluate if running our Procedure 2 SampleNearby around “good starts” yields more
good starts than running SampleNearby from all previously visited states. This can clearly be
seen in Figs. 3a-3b for the robotic manipulation tasks.
(a) Key insertion task
(b) Ring on Peg task
Figure 3: Fraction of Good Starts generated during training for the robotic manipulation tasks
6
Conclusions and Future Directions
We propose a method to automatically adapt the start state distribution on which an agent is trained,
such that the performance on the original problem is efficiently optimized. We leverage three assumptions commonly satisfied in simulated tasks to tackle hard goal-oriented problems that state of
the art RL methods cannot solve.
A limitation of the current approach is that it generates start states that grow from a single goal
uniformly outwards, until they cover the original start state distribution Unif(S 0 ). Nevertheless, if
the target set of start states S 0 is far from the goal and we have some prior knowledge, it would
be interesting to bias the generated start distributions ρi towards the desired start distribution. A
promising future line of work is to combine the present automatic curriculum based on start state
generation with goal generation [5], similar to classical results in planning [39].
It can be observed in the videos of our final policy for the manipulation tasks that the agent has
learned to exploit the contacts instead of avoiding them. Therefore, the learning based aspect of the
presented method has a huge potential to tackle problems that classical motion planning algorithms
could struggle with, such as environments with non-rigid objects or with uncertainties in the task
geometric parameters. We also leave as future work to combine our curriculum-generation approach
with domain randomization methods [40] to obtain policies that are transferable to the real world.
8
References
[1] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel. High-Dimensional continuous
control using generalized advantage estimation. In International Conference on Learning Representation, 2015.
[2] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,
M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
[3] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra.
Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015.
[4] S. Levine, C. Finn, T. Darrell, and P. Abbeel. End-to-end training of deep visuomotor policies.
Journal of Machine Learning Research, 17(39):1–40, 2016.
[5] D. Held, X. Geng, C. Florensa, and P. Abbeel. Automatic goal generation for reinforcement
learning agents. arXiv preprint arXiv:1705.06366, abs/1705.06366, 2017.
[6] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel. Benchmarking deep reinforcement learning for continuous control. International Conference on Machine Learning, 2016.
[7] I. Popov, N. Heess, T. Lillicrap, R. Hafner, G. Barth-Maron, M. Vecerik, T. Lampe, Y. Tassa,
T. Erez, and M. Riedmiller. Data-efficient deep reinforcement learning for dexterous manipulation. arXiv preprint arXiv: 1704.03073, 2017.
[8] A. Y. Ng, D. Harada, and S. Russell. Policy invariance under reward transformations: Theory and application to reward shaping. In International Conference in Machine Learning,
volume 99, pages 278–287, 1999.
[9] Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In International
Conference on Machine Learning, pages 41–48. ACM, 2009.
[10] W. Zaremba and I. Sutskever. Learning to execute. CoRR, abs/1410.4615, 2014.
[11] S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer. Scheduled sampling for sequence prediction
with recurrent neural networks. In Advances in Neural Information Processing Systems, 2015.
[12] A. Graves, M. G. Bellemare, J. Menick, R. Munos, and K. Kavukcuoglu. Automated Curriculum Learning for Neural Networks. arXiv preprint, arXiv:1704.03003, 2017.
[13] M. P. Kumar, B. Packer, and D. Koller. Self-paced learning for latent variable models. In
Advances in Neural Information Processing Systems, pages 1189–1197, 2010.
[14] L. Jiang, D. Meng, Q. Zhao, S. Shan, and A. G. Hauptmann. Self-paced curriculum learning.
In AAAI, volume 2, page 6, 2015.
[15] M. Asada, S. Noda, S. Tawaratsumida, and K. Hosoda. Purposive behavior acquisition for a
real robot by Vision-Based reinforcement learning. Machine Learning, 1996.
[16] A. Karpathy and M. Van De Panne. Curriculum learning for motor skills. In Canadian Conference on Artificial Intelligence, pages 325–330. Springer, 2012.
[17] J. Schmidhuber. POWER P LAY : Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem. Frontiers in Psychology, 2013.
[18] R. K. Srivastava, B. R. Steunebrink, M. Stollenga, and J. Schmidhuber. Continually adding
self-invented problems to the repertoire: First experiments with POWERPLAY. In IEEE International Conference on Development and Learning and Epigenetic Robotics, 2012.
[19] A. Baranes and P.-Y. Oudeyer. Active learning of inverse models with intrinsically motivated
goal exploration in robots. Robotics and Autonomous Systems, 61(1), 2013.
[20] S. Sharma and B. Ravindran. Online Multi-Task Learning Using Biased Sampling. arXiv
preprint arXiv: 1702.06053, 2017.
9
[21] S. Sukhbaatar, I. Kostrikov, A. Szlam, and R. Fergus. Intrinsic Motivation and Automatic
Curricula via Asymmetric Self-Play. arXiv preprint, arXiv: 1703.05407, 2017.
[22] A. Rajeswaran, K. Lowrey, E. Todorov, and S. Kakade. Towards generalization and simplicity
in continuous control. arXiv preprint, arXiv:1703.02660, 2017.
[23] R. Tedrake, I. R. Manchester, M. Tobenkin, and J. W. Roberts. Lqr-trees: Feedback motion
planning via sums-of-squares verification. The International Journal of Robotics Research, 29
(8):1038–1052, 2010.
[24] R. R. Burridge, A. A. Rizzi, and D. E. Koditschek. Sequential composition of dynamically
dexterous robot behaviors. The International Journal of Robotics Research, 1999.
[25] J. A. Bagnell, S. Kakade, A. Y. Ng, and J. Schneider. Policy search by dynamic programming.
Advances in Neural Information Processing Systems, 16:79, 2003.
[26] S. Kakade and J. Langford. Approximately Optimal Approximate Reinforcement Learning.
International Conference in Machine Learning, 2002.
[27] M. Kearns, Y. Mansour, and A. Y. Ng. A Sparse Sampling Algorithm for Near-Optimal Planning in Large Markov Decision Processes. Machine Learning, 49(2/3):193–208, 2002.
[28] K. Subramanian, C. L. Isbell, Jr., and A. L. Thomaz. Exploration from demonstration for interactive reinforcement learning. In Proceedings of the International Conference on Autonomous
Agents & Multiagent Systems, 2016.
[29] D. P. Bertsekas, D. P. Bertsekas, D. P. Bertsekas, and D. P. Bertsekas. Dynamic programming
and optimal control, volume 1. Athena scientific Belmont, MA, 1995.
[30] J. Peters and S. Schaal. Reinforcement learning of motor skills with policy gradients. Neural
networks, 21(4):682–697, 2008.
[31] I. Szita and A. Lörincz. Learning tetris using the noisy cross-entropy method. Learning, 18
(12), 2006.
[32] I. Osband, B. Van Roy, and Z. Wen. Generalization and exploration via randomized value
functions. In International Conference on Machine Learning, 2016.
[33] S. D. Whitehead. Complexity and Cooperation in Q-Learning. Machine Learning Proceedings
1991, pages 363–367, 1991.
[34] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT press Cambridge,
1998.
[35] M. P. Deisenroth, G. Neumann, J. Peters, et al. A survey on policy search for robotics. Foundations and Trends in Robotics, 2(1–2):1–142, 2013.
[36] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz. Trust region policy optimization.
In International Conference on Machine Learning, pages 1889–1897, 2015.
[37] E. Todorov, T. Erez, and Y. Tassa. Mujoco: A physics engine for model-based control. In
IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012.
[38] S. Sukhbaatar, I. Kostrikov, A. Szlam, and R. Fergus. Intrinsic motivation and automatic
curricula via asymmetric self-play. arXiv preprint arXiv:1703.05407, 2017.
[39] J. Kuffner and S. LaValle. RRT-connect: An efficient approach to single-query path planning.
In IEEE International Conference on Robotics and Automation, volume 2, pages 995–1001.
IEEE, 2000.
[40] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain Randomization
for Transferring Deep Neural Networks from Simulation to the Real World. arXiv preprint,
arXiv:1703.06907, 2017.
10
A
Experiment Implementation Details
A.1
Hyperparameters
Here we describe the hyperparemeters used for our method. Each iteration, we generate new start
states, (as described in Section 4.2 and Procedure 2), which we append to the seed states until we
have a total of M = 10000 start states. We then subsample these down to Nnew = 200 new start
states. These are appended with Nold = 100 sampled old start states (as described in Section 4.3
and Procedure 1), and these states are used to initialize our agent when we train our policy. The
“Brownian motion” rollouts have a horizon of TB = 50 timesteps, and the actions taken are random,
sampled from a standard normal distribution (e.g. a 0-mean Gaussian with a covariance Σ = I).
For our method as well as the baselines, we train a (64, 64) multi-layer perceptron (MLP) Gaussian
policy with TRPO [36], implemented with rllab [6]. We use a TRPO step-size of 0.01 and a (32, 32)
MLP baseline. For all tasks, we train with a batch size of 50,000 timesteps. All experiments use a
maximum horizon of T = 500 time steps but the Ant maze experiments that use a maximum horizon
of T = 2000. The episode ends as soon as the agent reaches a goal state. We define the goal set S g
to be a ball around the goal state, in which the ball has a radius of 0.03m for the ring and key tasks,
0.3m for the point-mass maze task and 0.5m for the ant-maze task. In our definition of Si0 , we use
Rmin = 0.1 and Rmax = 0.9. We use a discount factor γ = 0.998 for the optimization, in order to
encourage the policy to reach the goal as fast as possible.
A.2
Performance metric
The aim of our tasks is to reach a specified goal region S g from all start states s0 ∈ S 0 that are
feasible and within a certain distance of that goal region. Therefore, to evaluate the progress on
ηρ0 (πi ) we need to collect trajectories starting at states uniformly sampled from S 0 . For the pointmass maze navigation task this is straight forward as the designer can give a concrete description
of the feasible (x, y) space, so we can uniformly sample from it. Nevertheless, it is not trivial to
uniformly sample from all feasible start states for the robotics tasks. In particular, the state space is in
joints angles and angular velocities of the 7 DOF arm, but the physical constraints of these contactrich environments are given by the geometries of the task. Therefore, uniformly sampling from
the angular bounds mostly yields unfeasible states, with some part of the arm or the end-effector
intersecting with other objects in the scene. In order to approximate uniformly sampling from S 0 ,
we use Assumption 2: from the provided feasible goal state sg . We simply run our SampleNearby
procedure initialized with starts = [sg ] with a very large M and long time horizons TB . This
large aggregated state data-set is saved and samples from it are used as proxy to S 0 to evaluate the
performance of our algorithm. Figures 4a and 4b show six sampled start states from the data sets
used to evaluate the ring task and the key task. These data sets are available at the project website2
for future reproducibility and benchmarking.
Given the quasi-static nature of the tasks considered, we generate only initial angle positions, and
we set all initial velocities to zero. Generating initial velocities is a fairly simple extension of our
approach that we leave for future work.
B
Other methods
B.1
Distance reward shaping
Although our method is able to train policies with sparse rewards, the policy optimization steps
train pol can use any kind of reward shaping available. Up to an extend, we already do that by
using a discount factor γ, which motivates the policies to reach the goal as soon as possible. Similar
reward modulations could be included to take into account energy penalties or reward shaping from
prior knowledge. For example, in the robotics tasks considered in this paper, the goal is defined
in terms of a reference state, and hence it seems natural to try to use the distance to this state as an
additional penalty to guide learning. However, we have found that this modification does not actually
improve training. For the start states near to the goal, the policy can learn to reach the goal simply
from the indicator reward introduced in Section 3.2. For the states that are further away, the distance
2
Videos, data sets and code available at: bit.ly/reversecurriculum
11
(a) Uniformly sampled start states for ring task. There are 39,530 states in the data-set, of which 5,660 have
the ring with its hole already in the peg
(b) Uniformly sampled start states for key task. There are 544,575 states in the data-set, of which 120,784
have the key somewhere inside the key-hole
Figure 4: Samples from the test distribution for the manipulation tasks
to the goal is actually not a useful metric to guide the policy; hence, the distance reward actually
guides the policy updates towards a suboptimal local optimum, leading to poor performance. In
Fig. 5 we see that the ring task is not much affected by the additional reward, whereas the key task
suffers considerably if this reward is added.
(a) Ring on Peg task
(b) Key insertion task
Figure 5: Learning curves for the robotics manipulation tasks
B.2
Failure cases of Uniform Sampling for maze navigation
In the case of the maze navigation task, we observe that applying TRPO directly on the original
MDP incures in very high variance across learning curves. We have observed that some policies
only learned how to perform well from a certain side of the goal. The reason for this is that our
learning algorithm (TRPO) is a batch on-policy method; therefore, at the beginning of learning,
uniformly sampling from the state-space might give a batch with very few trajectories that reach
the goal and hence it is more likely that they all come from one side of the goal. In this case, the
algorithm will update the policy to go in the same direction from everywhere, wrongly extrapolating
from these very few successful trajectories it received. This is less likely to happen if the trajectories
for the batch are collected with a different start state distribution that concentrates more uniformly
around the goal, as the better learning progress of the other curves show.
B.3
Failure cases of Asymmetric Self-play
In Section 5.2, we compare the performance of our method to the asymmetric self-play approach
of Sukhbaatar et al. [38]. Although such an approach learns faster than the uniform sampling baseline, it gets stuck in a local optimum and fails to learn to reach the goal from more than 40% of
start-states in the point-mass maze task.
12
As explained above, part of the reason that this method gets stuck in a local optimum is that “Alice”
(the policy that is proposing start-states) is represented with a unimodal Gaussian distribution, which
is a common representation for policies in continuous action spaces. Thus Alice’s policy tends to
converge to moving in a single direction. In the original paper, this problem is somewhat mitigated
by using a discrete action space, in which a multi-modal distribution for Alice can be maintained.
However, even in such a case, the authors of the original paper also observed that Alice tends to
converge to a local optimum [38].
A further difficulty for Alice is that her reward function can be sparse, which can be inherently
difficult to optimize. Alice’s reward is defined as rA = max(0, tB − tA ), where tA is the time that
it takes Alice to reach a given start state from the goal (at which point Alice executes the “stop”
action), and tB is the time that it takes Bob to return to the goal from the start state. Based on this
reward, the optimal policy for Alice is to find the nearest state for which Bob does not know how
to return to the goal; this will lead to a large value for tB with a small value for tA . In theory, this
should lead to an automatic curriculum of start-states for Bob.
However, in practice, we find that sometimes, Bob’s policy might improve faster than Alice’s. In
such a case, Bob will have learned how to return to the goal from many start states much faster than
Alice can reach those start states from the goal. In such cases, we would have that tB < tA , and
hence rA = 0. Thus, Alice’s rewards are sparse (many actions that Alice takes result in 0 reward)
and hence it will be difficult for Alice’s policy to improve, leading to a locally optimal policy for
Alice. For these reasons, we have observed Alice’s policy often getting “stuck,” in which Alice is
unable to find new start-states to propose for Bob that Bob does not already know how to reach the
goal from.
We have implemented a simple environment that illustrates these issues. In this environment, we
use a synthetic “Bob” that can reach the goal from any state within a radius rB from the goal. For
states within rB , Bob can reach the goal in a time proportional to the distance between the state
and the goal; in other words, for such states s0 ∈ {s : |s − sg | < rB , s ∈ S 0 }, we have that
tB = |s0 − sg |/vB , where |s0 − sg | is the distance between state s0 and the goal sg , and vB is Bob’s
speed. For states further than rB from the goal, Bob does not know how to reach the goal, and thus
tB is the maximum possible value.
This setup is illustrated in Figure 6. The region shown in red designates the area within rB from the
goal, e.g. the set of states from which Bob knows how to reach the goal. On the first iteration, Alice
has a random policy (Figure 6a). After 10 iterations of training, Alice has converged to a policy that
reaches the location just outside of the set of states from which Bob knows how to reach the goal
(Figure 6b). From these states, Alice receives a maximum reward, because tB is very large while
tA is low. Note that we also observe the unimodal nature of Alice’s policy; Alice has converged to
a policy which proposes just one small set of states among all possible states for which she would
receive a similar reward.
(a) Iteration 1
(b) Iteration 10
(c) Iteration 32
Figure 6: Simple environment to illustrate asymmetric self-play [38]. The red areas indicate the
states from which Bob knows how to reach the goal. The blue points are the start-states proposed
by Alice at each iteration (i.e. the states from which Alice performed the stop action)
.
At this point we synthetically increase rB , corresponding to the situation in which Bob learns how to
reach the goal from a larger set of states. However, Alice’s policy has already converged to reaching
a small set of states which were optimal for Bob’s previous policy. From these states Alice now
receives a reward of 0, as described above: Bob can return from these states quickly to the goal, so
13
we have that tB < tA and rA = 0. Thus, Alice does not receive any reward signal and is not able to
improve her policy. Hence, Alice’s policy remains stuck at this point and she is not able to find new
states to propose to Bob (Figure 6c).
In this simple case, one could attempt to perform various hacks to try to fix the situation, e.g. by
artificially increasing Alice’s variance, or by resetting Alice to a random policy. However, note that,
in a real example, Bob is learning an increasingly complex policy, and so Alice would need to learn
an equally complex policy to find a set of states that Bob cannot succeed from; hence, these simple
fixes would not suffice to overcome this problem. Fundamentally, the asymmetric nature of the selfplay between Alice and Bob creates a situation in which Alice has a difficult time learning and often
gets stuck in a local optimum from which she is unable to improve.
14
| 2cs.AI
|
Metric Embedding via Shortest Path Decompositions
Ittai Abraham1 , Arnold Filtser2 , Anupam Gupta3 , and Ofer Neiman2
1
arXiv:1708.04073v1 [cs.DS] 14 Aug 2017
2
VMWare. Email: [email protected]
Ben-Gurion University of the Negev. Email: {arnoldf,neimano}@cs.bgu.ac.il
3
Carnegie Mellon University. Email: [email protected]
August 15, 2017
Abstract
We study the problem of embedding weighted graphs of pathwidth k into `p spaces. Our
main result is an O(k min{1/p,1/2} )-distortion embedding. For p = 1, this is a super-exponential
improvement over the best previous bound of Lee and Sidiropoulos. Our distortion bound is
asymptotically tight for any fixed p > 1.
Our result is obtained via a novel embedding technique that is based on low depth decompositions of a graph via shortest paths. The core new idea is that given a geodesic shortest path
P , we can probabilistically embed all points into 2 dimensions with respect to P . For p > 2 our
embedding also implies improved distortion on bounded treewidth graphs (O((k log n)1/p )). For
asymptotically large p, our results also implies improved distortion on graphs excluding a minor.
1
Introduction
Low-distortion metric embeddings are a crucial component in the modern algorithmist toolkit. Indeed,
they have applications in approximation algorithms [LLR95], online algorithms [BBMN11], distributed
algorithms [KKM+ 12], and for solving linear systems and computing graph sparsifiers [ST04]. Given
a (finite) metric space (V, d), a map φ : V → RD , and a norm k · k, the contraction and expansion of
the map φ are the smallest τ, ρ ≥ 1, respectively, such that for every pair x, y ∈ V ,
1
kφ(x) − φ(y)k
≤
≤ρ .
τ
d(x, y)
The distortion of the map is then τ · ρ In this paper we will investigate embeddings into `p norms;
the most prominent of which are the Euclidean norm `2 and the cut norm `1 ; the former for obvious
reasons, and the latter because of its close connection to graph partitioning problems, and in particular
the Sparsest Cut problem.
We focus on embedding of metrics arising from certain graph families. Indeed, since general n-point
metrics require Ω(log n/p)-distortion to embed into `p -norms, much attention was given to embeddings
of restricted graph families that arise in practice. (Embedding an (edge-weighted) graph is short-hand
for embedding the shortest path metric of the graph generated by these edge-weights.) Since the class
of graphs embeddable with some distortion into some target normed space is closed under taking
minors, it is natural to focus on minor-closed graph families. A long-standing conjecture in this area
is that all non-trivial minor-closed families of graphs embed into `1 with distortion depending only on
the graph family and not the size n of the graph.
While this question remains unresolved in general, there has been some progress on special classes of
graphs. The class of outerplanar graphs (which exclude K2,3 and K4 as a minor) embeds isometrically
into `1 ; this follows from results of Okamura and Seymour [OS81]. Following [GNRS04], Chakrabarti
et al. [CJLV08] show that every graph with treewidth-2 (which excludes K4 as a minor) embeds into
`1 with distortion 2 (which is tight, as shown by [LR10]). Lee and Sidiropoulos [LS13] showed that
3
every graph with pathwidth k can be embedded into `1 with distortion (4k)k +1 . See Section 1.3 for
additional results.
We note that `2 is a potentially more natural and useful target space than `1 (in particular, `2 embed isometrically into `1 ). Alas, there are only few (natural) families of metrics that admit constant
distortion embedding into Euclidean space, such as “snowflakes” of a doubling metrics [Ass83], doubling trees [GKL03] and graphs of bounded bandwidth [BCMN13]. All these families have bounded
doubling dimension. (For definitions, see Section 2.)
1.1
Our Results
In this paper we develop a new technique for embedding certain graph families into `p spaces with
constant distortion. Our main result is an improved embedding for bounded pathwidth graphs (see
Section 2 for the definition of pathwidth and other graph families).
Theorem 1 (Pathwidth Theorem). Any graph with pathwidth k embeds into `p with distortion O(k 1/p ).
3
Note that this is a super-exponential improvement over the best previous distortion bound of O(k)k ,
by Lee and Sidiropoulos. Their approach was based on probabilistic embedding into trees, which
implies embedding only into `1 . Such an approach cannot yield distortion better than O(k), due to
known lower bounds for the diamond graph [GNRS04], that has pathwidth k + 1. Our embedding
1
holds for any `p space, and we can overcome the barrier
√ of O(k) (since `2 embeds into `p isometrically,
the distortion of Theorem 1 is never larger than O( k)). √In particular, we obtain embeddings of
pathwidth-k graphs into both `2 and `1 with distortion O( k). Moreover, an embedding with this
distortion can be found efficiently via semidefinite-programming; see, e.g., [LLR95], even without access to the actual path decomposition (which is NP-hard even to approximate [BGHK92]). We remark
that graphs of bounded pathwidth can have arbitrarily large doubling dimension (exhibited by star
graphs that have pathwidth 1), and thus our result is a noteworthy example of a non-trivial Euclidean
embedding with constant distortion for a family of metrics with unbounded doubling dimension.
Since graphs of treewidth k have pathwidth O(k log n) (see, e.g., [GTW13]), Theorem 1 provides
an embedding of such graphs into `p with distortion O((k log n)1/p ). This strictly improves the
best previously known bound, which follows from a theorem in [KLMN05] (who obtained distortion O(k 1−1/p log1/p n) ), for any p > 2, and matches it for 1 ≤ p ≤ 2. While [KK16] obtained recently
a distortion bound with improved dependence on k, their result O((log(k log n))1−1/p (log1/p n)) has
sub-optimal dependence on n.
A General Embedding Framework. The embedding of Theorem 1 follows as a special case of a more
general theorem: we devise embeddings for any graph family which admits “shortest path decompositions” (SPDs) of “low depth”. Every (weighted) path graph has an SPD of depth 1. A graph G has
an SPD of depth k if after removing some shortest path P , every connected component in G \ P has
an SPD of depth k − 1. (An alternative definition appears in Definition 1.) Our main technical result
is the following.
Theorem 2 (Embeddings for SPD Families). Let G = (V, E) be a weighted graph with an SPD of
depth k. Then there exists an embedding f : V → `p with distortion O(k 1/p ).
Since bounded-pathwidth graphs admit SPDs of low depth, we get Theorem 1 as a simple corollary
of Theorem 2. Moreover, we derive several other results, which are summarized in Figure 1.1; these
results either improve on the state-of-the-art, or provide matching bounds using a completely different
approach.
Graph Family
Our results.
1/p
Pathwidth k
Treewidth k
O(k )
O((k log n)1/p )
Planar
H-minor-free
O(log /p n)
O((g(H) log n)1/p )
1
Previous results
3
(4k)k +1 into `1
O(k 1−1/p · log1/p n)
O((log(k log n))1−1/p (log1/p n))
1
O(log /p n)
1
O(|H|1−1/p log /p n)
[LS13]
[KLMN05]
[KK16]
[Rao99]
[AGG+ 14]+[KLMN05]
Figure 1: Our and previous results for embedding certain graph families into `p . (For H-minor-free
graphs, g(H) is some function of |H|.)
A Matching Lower Bound. Our result of Theorem 1 (and thus also Theorem 2) is asymptotically tight
for any fixed p > 1. The family exhibiting this fact is the diamond graphs.
Theorem 3 ([LN04, JS09]). For any fixed p > 1 and every k ≥ 1, there exists a graph G = (V, E)
with pathwidth-k, such that every embedding f : V → `p has distortion Ω(k min{1/p,1/2} ).
The bound in Theorem 3 was proven first for p = 2 in [NR03], generalized to 1 < p ≤ 2 in [LN04] and
for p ≥ 2 by [JS09] (see also [MN13, JLM11]). For completeness, we provide a direct proof of the case
p ≥ 2 in Appendix A. We note that for `1 only the trivial Ω(log k) lower bound is known.
2
1.2
Technical Ideas
Many known embeddings [Bou85, Rao99, KLMN05, ABN11] are based on a collection of 1-dimensional
embeddings, where we embed each point to its distance from a given subset of points. We follow this
approach, but differ in two aspects. Firstly, the subset of points we use is not based on random
sampling or probabilistic clustering. Rather, inspired by the works of [And86] and [AGG+ 14], the
subset used is a geodesic shortest path. The second is that our embedding is not 1-dimensional but
2-dimensional: this seemingly small change crucially allows us to use the structure of the shortest
paths to our advantage.
The SPD induces a collection of shortest paths (each shortest path lies in some connected component).
A natural initial attempt is to embed a vertex v relative to a geodesic path P using two dimensions:1
• The first coordinate ∆1 is the distance to the path d(v, P ).
• The second coordinate ∆2 is the distance d(v, r) to the endpoint of the path, called its “root”.
v
∆2
(v , r
=d
)
∆1 = d(v, P )
P
r
Unfortunately, this embedding may have unbounded expansion: If two vertices u, v are separated by
some shortest path, in future iterations v may have a large distance to the root of a path P in its
component, while u has zero in that coordinate (because it’s not in that component), incurring a large
stretch. The natural fix is to enforce a Lipschitz condition on every coordinate: for v in cluster X, we
truncate the value v can receive at O(dG (v, V \ X)). I.e., a vertex close to the boundary of X cannot
get a large value. Using the fact that the SPD has depth k, each vertex will have only O(k) nonzero
coordinates, which implies expansion O(k 1/p ).
To bound the contraction, for each pair u, v we consider the first path P in the SPD that lies “close”
to {u, v} or separates them to different connected components. Then we show that at least one of the
two coordinates should give sufficient contribution. (See, e.g., Figure 2.)
P
C2
C1
u
(a)
P
v
r
v
P
C2
C1
C1
v
C2
u
u
(b)
r
(c)
r
Figure 2: A shortest path P , rooted at r, partitions cluster X into clusters C1 and C2 . In cases (a) and (b),
the first coordinate (the distance to P ) provides sufficient contribution. In case (c) the second coordinate (the
distance to root r) provides the contribution.
But what about the effect of truncation on contraction? A careful recursive argument shows that the
contribution to u, v from the first coordinate (the distance from the path P ) is essentially not affected
by this truncation. Hence the argument in cases (a) and (b) of Figure 2 still works. However, the
1
In fact, we use different dimensions for each connected components.
3
argument using the distance to the root of P , case (c), can be ruined. Solving this issue requires some
new non-trivial ideas. Our solution is to introduce a probabilistic sawtooth function that replaces the
simple truncation. The main technical part of the paper is devoted to showing that a collection of
these functions for all possible distance scales, with appropriate random shifts, suffices to control the
expected contraction in case (c), for all relevant pairs simultaneously.
1.3
Other Related Work
There has been work on embedding several other graph families into normed spaces: Chekuri et
al. [CGN+ 06] extend the Okamura and Seymour bound for outerplanar graphs to k-outerplanar
graphs, and showed that these embed into `1 with distortion 2O(k) . Rao [Rao99] (see also [KLMN05])
embed planar graphs into `p with distortion O(log1/p n). For
√ graphs with genus g, [LS10] showed an
embedding into Euclidean space with distortion O(log g + log n). Finally, for H-minor-free graphs,
combining the results of [AGG+ 14, KLMN05] give `p -embeddings with O(|H|1−1/p log1/p n) distortion.
Following [And86, Mil86], the idea of using geodesic shortest paths to decompose the graph has been
used for many algorithmic tasks: MPLS routing [GKR04], directed connectivity, distance labels and
compact routing [Tho04], object location [AG06], and nearest neighbor search [ACKW15]. However, to
the best of our knowledge, this is the first time it has been used directly for low-distortion embeddings
into normed spaces.
2
Preliminaries and Notation
For k ∈ Z, let [k] := {1, . . . , k}. For p ≥ 1, the `p -norm of a vector x = (x1 , . . . , xd ) ∈ Rd is
P
kxkp := ( di=1 |xi |p )1/p , where kxk∞ := maxi |xi |.
Doubling dimension. The doubling dimension of a metric is a measure of its local “growth rate”.
Formally, a metric space (X, d) has doubling dimension λX if for every x ∈ X and radius r, the ball
B(x, r) can be covered by 2λX balls of radius 2r . A family is doubling if the doubling dimension of all
metrics in it is bounded by some universal constant.
Graphs. We consider connected undirected graphs G = (V, E) with edge weights w : E → R>0 . Let
dG denote the shortest path metric in G; we drop subscripts when there is no ambiguity. For a vertex
x ∈ V and a set A ⊆ V , let dG (x, A) := mina∈A d(x, a), where dG (x, ∅) := ∞. For a subset of vertices
A ⊆ V , let G[A] denote the induced graph on A, and let dA := dG[A] be the shortest path metric in
the induced graph. Let G \ A := G[V \ A] be the graph after deleting the vertex set A from G.
Special graph families. Given a graph G = (V, E), a tree decomposition of G is a tree T with nodes
B1 , . . . , Bs (called bags) where each Bi is a subset of V such that the following properties hold:
• For every edge {u, v} ∈ E, there is a bag Bi containing both u and v.
• For every vertex v ∈ V , the set of bags containing v form a connected subtree of T .
The width of a tree decomposition is maxi {|Bi | − 1}. The treewidth of G is the minimal width of a
tree decomposition of G.
A path decomposition of G is a special kind of tree decomposition where the underlying tree is a path.
The pathwidth of G is the minimal width of a path decomposition of G.
A graph H is a minor of a graph G if we can obtain H from G by edge deletions/contractions, and
vertex deletions. A graph family G is H-minor-free if no graph G ∈ G has H as a minor.
4
2.1
The Sawtooth function
Another component in our embeddings will be the following sawtooth function. For t ∈ N, we define
gt : R+ → R the sawtooth function w.r.t. 2t as follows. For x ≥ 0, if qx := bx/2t+1 c then
gt (x) = 2t − x − qx · 2t+1 + 2t .
x2
2t
x1
0
2t+1
x3
2 · 2t+1
3 · 2t+1
4 · 2t+1
5 · 2t+1
Figure 3: The graph of the “sawtooth” function gt . The points x1 = 5 · 2t−1 and x3 = 15 · 2t−1 are mapped
to 2t−1 , while x2 = 10 · 2t−1 is mapped to 2t .
See Figure 3 for an illustration. The following observation is straightforward.
Observation 1. The Sawtooth function gt is 1-Lipschitz, bounded by 2t , periodic with period 2t+1 .
The proof of the following lemma appears in Appendix C.
Lemma 2 (Sawtooth Lemma). Let x, y ∈ R+ . Let α ∈ [0, 1], β ∈ [0, 4] be drawn uniformly and
independently. The following properties hold:
1. Eα,β gt (βx + α · 2t+1 ) = 2t−1 .
2. If |x − y| ≤ 2t−1 , then Eα,β gt (βx + α · 2t+1 ) − gt (βy + α · 2t+1 ) = Ω(|x − y|).
3. If |x − y| > 2t−1 , then Eα,β gt (βx + α · 2t+1 ) − gt (βy + α · 2t+1 ) = Ω(2t ).
3
Shortest Path Decompositions
Our embeddings will crucially depend on the notion of shortest path decompositions. In the introduction we provided a recursive definition for SPD. Here we show an equivalent definition which will
be more suitable for our purposes.
Definition 1 (Shortest Path Decomposition (SPD)). Given a weighted graph G = (V, E, w), a SPD
of depth k is a pair {X , P}, where X is a collection X1 , . . . , Xk of partial partitions of V 2 , and P is
a collection of sets of paths P1 , . . . , Pk , where X1 = {V }, Xk = Pk , and the following properties hold:
1. For every 1 ≤ i ≤ k and every subset X ∈ Xi , there exist a unique path PX ∈ Pi such that PX
is a shortest path in G[X].
2. For every 2 ≤ i ≤ k, Xi consists of all connected components of G[X \ PX ] over all X ∈ Xi−1 .
2
i.e. for every X ∈ Xi , X ⊆ V , and for every different subsets X, X 0 ∈ Xi , X ∩ X 0 = ∅.
5
S
In other words, ki=1 Pk is a partition of V into paths, where each path PX is a shortest path in the
component X it belongs to at the point it is deleted.
For a given graph G let SPDdepth(G) be the minimum k such that G admits an SPD of depth k. For
a given family of graphs G let SPDdepth(G) := maxG∈G {SPDdepth(G)}. In the following we consider
the SPDdepth of some graph families.
3.1
The SPD Depth for Various Graph Families
One advantage of defining the shortest path decomposition is that several well-known graph families
have bounded depth SPD.
• Pathwidth. Every graph G = (V, E, w) with pathwidth k has an SPDdepth of k + 1. Indeed, let
T = hB1 , . . . , Bs i be a path decomposition of G, where B1 , Bs are the two bags at the end of this
path. Choose arbitrary vertices x ∈ B1 and y ∈ Bs , and let P be a shortest path in G from x to
y. By the definition of a path decomposition, the path P contains at least one vertex from every
bag Bi . Hence, deleting the vertices of P would reduce the size of each bag by one; consequently
each connected component of G \ P has pathwidth k − 1, and by induction SPDdepth k. Finally,
a connected component of pathwidth 0 is necessarily a singleton, which has SPDdepth 1.
• Treewidth. Since every tree has pathwidth O(log n), we can show that an n-vertex treewidth-k
graph has pathwidth O(k log n). Hence, treewidth-k graphs have SPDdepth O(k log n).
• Planar. Using cycle separators [Mil86] as in [Tho04, GKR04], every planar graph has SPDdepth
O(log n); this follows as each cycle separator can be constructed as union of two shortest paths.
• Minor-free. Finally, every H-minor-free graph admits a balanced separator consisting of g(H)
shortest paths (for some function g) [AG06], and hence has an SPDdepth O(g(H) · log n).
Combining these observation with Theorem 2, we get the following set of results:
Corollary 1. Consider an n-vertex weighted graph G, Theorem 2 implies the following:
•
•
•
•
If G has pathwidth k, it embeds into `p with distortion O(k 1/p ).
If G has treewidth k, it embeds into `p with distortion O((k log n)1/p ).
1
If G is planar, it embeds into `p with distortion O(log /p n).
1
For every fixed H, if G excludes H as a minor, it embeds into `p with distortion O(log /p n),
where the constant in the big-O depends on H.
As mentioned in Section 1, we get a substantial improvement for the pathwidth case. Our result
for treewidth improves upon that from [KLMN05] for p > 2; they got O(k 1−1/p (log n)1/p ) distortion
compared to our O((k log n)1/p ). Our result appears to be closer to the truth, since the distortion
tends to O(1) as p → ∞.
Finally, our results for planar graphs match the current state-of-the-art.
Our results for minor-free graphs depend on the Robertson-Seymour decomposition, and hence are
currently better only for large values of p. (It remains an open question to improve the SPDdepth of
H-minor-free graphs to have a poly(|H|) log n dependence, perhaps using the ideas from [AGG+ 14].)
In general, we hope that our results will be useful in getting other embedding results, and will spur
further work on understanding shortest path separators.
6
4
The Embedding Algorithm
Let G = (V, E) be a weighted graph, and let {X , P} = {{X1 , , . . . , Xk } , {P1 , . . . , Pk }} be an SPD of
depth k for G. By scaling, we can assume that the minimum weight of an edge is 1; let M ∈ N be
the minimal such that the diameter of G is strictly bounded by 2M . Pick α ∈ [0, 1] and β ∈ [0, 4]
uniformly and independently.
For every i ∈ [k], and X ∈ Xi , we now construct an embedding fX : V → RD (for some number of
dimensions D ∈ N). This map fX consists of two parts.
First coordinate: Distance to the Path. The first coordinate of the embedding implements the distance
path
to the path PX , and is denoted by fX
. Let X1 , . . . , Xs ∈ Xi+1 be the connected components of
G [X \ PX ] (note that it is also possible that s = 0). We use a separate coordinate for each Xj , and
path
hence fX
: V → Rs . Moreover, for v ∈ X we truncate at 2 dG (v, V \ X) in order to guarantee
Lipschitz-ness. In particular, the coordinate corresponding to Xj is set to
(
min {dX (v, PX ), 2dG (v, V \ X)}
if v ∈ Xj ,
path
(v) =
fX
Xj
0
otherwise.
See Figure 4 for an illustration.
PX
(0, 0)
(0, 0)
(7, 0)
3
(6, 0)
(0, 0)
X2
7
3
X1
(0, 0)
2
(0, 4)
(0, 3)
2
(2, 0)
Figure 4: The set X ∈ Xi surrounded by a closed curve. The path PX partitions X into X1 , X2 ∈ Xi+1 . The
path
embedding fX
consists of two coordinates, and represented in the figure by a horizontal vector next to each
vertex, where the first entry is w.r.t. X1 and the second w.r.t. X2 . Each point on PX , or not in X maps to 0
in both the coordinates. Each point in X1 maps to min {dX (v, PX ) , 2dG (v, V \ X)} in the first coordinate and
to 0 in the second.
root , which is intended to
Second coordinate: Distance to the Root. The second part is denoted fX
capture the distance from the root r of the path. Again, to get the Lipschitz-ness, we would like to
path
truncate the value at 2 dG (v, V \ X) as we did for fX
. However, a problem with this idea is that the
root r can be arbitrarily far from some pair u, v that needs contribution from this coordinate. And
hence, even if |dG (u, r) − dG (v, r)| ≈ dG (u, v), there may be no contribution after the truncation. So
we use the sawtooth function.
Specifically, we replace the ideal contribution dG (v, r) by the sawtooth function gt (dG (v, r)), where
the scale t for the function is chosen such that 2t ≈ dG (v, V \ X). To avoid the case that two
7
nearby points use two different scales (and hence to guarantee Lipschitz-ness), we take an appropriate
linear combination of the two distance scales closest to 2 dG (v, V \ X). Recall that the sawtooth
function does not guarantee contribution for u, v due to its periodicity: we may be unlucky and
have gt (dG (v, r)) = gt (dG (u, r)) even when dG (v, r) and dG (u, r) are very different. To guarantee a
large enough contribution for all relevant pairs simultaneously, we add a random shift α, and apply a
random “stretch” β to dG (v, r) before feeding it to gt . Lemma 2 then shows that many of the choices
of α and β give substantially different values for u, v.
root consists of M + 1 coordinates, one for each
Formally, the mapping is as follows. The function fX
root . Let r be an
distance scale t ∈ {0, 1, . . . , M }. The coordinate corresponding to t is denoted by fX,t
arbitrary endpoint of PX ; we will call r the “root” of PX . Let tv ∈ N be such that 2 dG (v, V \ X) ∈
tv
[2tv , 2tv +1 ). Set λv = 2 dG (v,V2t\X)−2
. Note that 0 ≤ λv < 1. For v ∈ X, we define
v
t+1
if t = tv + 1,
λv · gt (β · dX (v, r) + α · 2 )
root
t+1
fX,t (v) = (1 − λv ) · gt (β · dX (v, r) + α · 2 ) if t = tv ,
(1)
0
otherwise.
root (v) = ~
0.
For all nodes v ∈
/ X, we set fX
path
root , and the final embedding is
⊕ fX
Define the map fX = fX
f=
k M
M
fX ,
i=1 X∈Xi
i.e., the concatenation of all the constructed embeddings. Before we start the analysis, let us record
some simple observations.
Observation 3. For the map f defined above, the following hold:
• The number of coordinates in f does not depend on α, β.
• For every X ∈ Xi and v ∈
/ X, the map fX (v) is the constant vector ~0.
• For every X ∈ Xi and v ∈ X, the map fX is nonzero in at most 3 coordinates.
Hence, since Xi is a partial partition of V and the depth of the SPD is k, we get that f (v) is nonzero
in at most 3k coordinates for each v ∈ V .
5
The Analysis
The main technical lemmas now show that the per-coordinate expansion is constant, and that for
every pair, there exists a coordinate for which the expected contraction is constant. (The proofs of
these lemmas appear in Section 5.1 and Section 5.2, respectively.)
Lemma 4 (Expansion Bound). For any vertices u, v, every coordinate j, and every choice of α, β,
|fj (v) − fj (u)| = O(dG (u, v)).
Lemma 5 (Contraction Bound). For any vertices u, v, there exists some coordinate j such that
Eα,β [|fj (v) − fj (u)|] = Ω(dG (u, v)).
Given these two lemmas, we can combine them together to show that the entire embedding has small
distortion. (Proof of the composition lemma can be found in Appendix B.)
8
Lemma 6 (Composition Lemma). Let (X, d) be a metric space. Suppose that there are constants ρ, τ
and a function f : X → Rs , drawn from some probability space such that:
1. For every u, v ∈ X and every j ∈ [s], |fj (v) − fj (u)| ≤ ρ · d(v, u).
2. For every u, v ∈ X, there exists j ∈ [s] such that E[ |fj (v) − fj (u)| ] ≥
1
τ
· d(v, u).
3. For every v ∈ X, there is a subset of indices Iv ⊆ [s] of size |Iv | ≤ k, such that for every j ∈
/ Iv ,
fj (v) = 0. In other words, for every v ∈ X, f (v) has support of size at most k.
Then, for every p ≥ 1, there is an embedding of (X, d) into `p with distortion O(k 1/p ). Moreover,
if there is an efficient algorithm for sampling such an f , then there is a randomized algorithm that
constructs the embedding efficiently (in expectation).
Now Theorem 2, our embedding for graphs with low depth SPDs, immediately follows by applying
the Composition Lemma (Lemma 6) to Lemma 4, Lemma 5, and Observation 3.
5.1
Bounding the Expansion: Proof of Lemma 4
In this section we bound the expansion of any coordinate in our embedding. Recall that the embedding
of v lying in some component X consists of two sets of coordinates: its distance from the path, and
its distance from the root. As mentioned in the introduction, since points outside X are mapped to
zero, maintaining Lipschitz-ness requires us to truncate the contribution of v of any coordinate to its
distance from the boundary. This truncation (either via taking a minimum with dG (v, V \ X), or via
the sawtooth function), means that our proofs of expansion require more care. The details appear
below.
Consider any level i, any set X ∈ Xi , and any pair of vertices u, v. It suffices to show that kfX (v) −
fX (u)k∞ = O(dG (u, v)). To begin, we may assume that both u, v ∈ X. Indeed, if both u, v ∈
/ X, then
~
fX (v) = fX (u) = 0 and we are done. If one of them, say v belongs to X while the other u ∈
/ X, then
fX (u) = ~0 while fX (v) is bounded by 2tv +1 ≤ 4dG (v, V \ X) ≤ 4 dG (u, v) in each coordinate.
Moreover, we may also assume that the shortest u-v path in G contains only vertices from X. Indeed,
suppose their shortest path in G uses vertices from V \X, then dG (u, V \X)+dG (v, V \X) ≤ dG (u, v).
But since both fX (v), fX (u) are bounded in each coordinate by 4 · max {dG (u, V \ X), dG (v, V \ X)},
we have constant expansion. Henceforth, we can assume that dG (u, v) = dX (u, v). We now bound the
expansion in each of the two parts of fX separately.
path
Expansion of fX
. Let Xv , Xu be the connected components in G [X \ PX ] such that v ∈ Xv and
u ∈ Xu . Consider the first case Xv 6= Xu , then PX intersects the shortest path between v and u. In
particular,
path
path
kfX
(v) − fX
(u)k∞ ≤ min {dX (v, PX ) , 2dG (v, V \ X)} + min {dX (u, PX ) , 2dG (u, V \ X)}
≤ dX (v, PX ) + dX (u, PX ) ≤ dX (v, u) = dG (v, u) .
path
path
Otherwise, Xv = Xu and the two vertices lie in the same component. Now kfX
(v) − fX
(u)k∞
equals
min {dX (v, PX ) , 2dG (v, V \ X)} − min {dX (u, PX ) , 2dG (u, V \ X)}
9
Assuming (without loss of generality) that the first term is at least the second, we can drop the
absolute value signs. Now the bound on the expansion follows from a simple case analysis. Indeed,
suppose dX (u, PX ) ≤ 2dG (u, V \ X). Then we get
path
path
kfX
(v) − fX
(u)k∞ = min {dX (v, PX ) , 2dG (v, V \ X)} − dX (u, PX )
≤ dX (v, PX ) − dX (u, PX ) ≤ dX (u, v) = dG (u, v).
The other case is that dX (u, PX ) > 2dG (u, V \ X), and then
path
path
kfX
(v) − fX
(u)k∞ = min {dX (v, PX ) , 2dG (v, V \ X)} − 2dG (u, V \ X)
≤ 2dG (v, V \ X) − 2dG (u, V \ X) ≤ 2dG (u, v).
Hence the expansion is bounded by 2.
root . Let r be the root of P . For t ∈ {0, 1, . . . , M }, let p (respectively, q )
Expansion of fX
t
t
X
be the “weight” of v (respectively, u) on gt —in other words, pt is the constant in (1) such that
root (v) = p · g (β · d (v, r) + α · 2t+1 ). Note that p ∈ {0, λ , 1 − λ } is chosen deterministically, and
fX,t
t
v
v
t
t
X
is nonzero for at most two indices t.
First, observe that for every t,
root
root
fX,t
(v) − fX,t
(u) = pt · gt (β · dX (v, r) + α · 2t+1 ) − qt · gt (β · dX (u, r) + α · 2t+1 )
≤ min {pt , qt } · gt (β · dX (v, r) + α · 2t+1 ) − gt (β · dX (u, r) + α · 2t+1 ) + |pt − qt | · 2t
≤ min {pt , qt } · β · |dX (v, r) − dX (u, r)| + |pt − qt | · 2t
≤ O(dG (u, v)) + |pt − qt | · 2t .
(2)
The first inequality used that gt is bounded by 2t , and the second inequality that gt is 1-Lipschitz; both
follow from Observation 1. The last inequality follows by the triangle inequality (since we assumed
that the shortest path from v to u is contained within X).
Hence, it suffices to show that |pt − qt | = O(dG (u, v)/2t ). Indeed, for indices t ∈
/ {tu , tu + 1, tv , tv + 1},
pt = qt = 0, hence |pt − qt | = 0. Let us consider the other cases. W.l.o.g., assume that dG (v, V \ X) ≥
dG (u, V \ X) and hence tv ≥ tu .
• tu = tv : In this case, |ptv − qtv | = |(1 − λv ) − (1 − λu )| = λv − λu = |ptv +1 − qtv +1 |. Moreover,
this quantity is
2dG (v, V \ X) − 2t 2dG (u, V \ X) − 2t
−
2t
2t
2 (dG (v, V \ X) − dG (u, V \ X))
2dG (u, v)
=
≤
.
t
2
2t
λv − λu =
Hence, we get that |pt − qt | = O(dG (u, v)/2t ) for both t ∈ {tv , tv + 1}.
• tu = tv − 1 : It holds that
2dG (v, V \ X) − 2tv
2tu +1 − 2dG (u, V \ X)
+
2tv
2tu
2dG (v, V \ X) − 2dG (u, V \ X)
2dG (u, v)
=
≤
.
2tu
2tu
λv + (1 − λu ) ≤ 2 ·
If we define χ := λv + (1 − λu ), we conclude that
10
|ptv +1 − qtv +1 |
|ptv − qtv |
|ptu − qtu |
=
=
=
λv
|1 − λv − λu |
1 − λu
≤
≤
≤
χ
χ
χ
=
=
=
O(dG (u, v)/2tv +1 )
O(dG (u, v)/2tv )
O(dG (u, v)/2tu ) .
• tu < tv − 1 : By the definition of tv and tu ,
2dG (v, u) ≥ 2 (dG (v, V \ X) − dG (u, V \ X)) ≥ 2tv − 2tu +1 ≥ 2tv −1 .
dG (u,v)
In particular, for every t ≤ tv + 1, |pt − qt | ≤ 1 ≤ 2d2Gtv(u,v)
=
O
.
−1
2t
5.2
Bounding the Contraction: Proof of Lemma 5
Now we need to bound the contraction. A natural proof idea would be to say that vertices u, v would
eventually be separated and then either they are far from the separating path P , or they lie close to
the path. And hence either d(v, P ) or the distance from one end of the path (the “root”) would give a
large contribution. But the value of the v’s embedding in any single coordinate cannot be more than
v’s distance to the boundary, and this causes problems. Indeed, if u, v fall very close to the path P at
some step of the algorithm, they must get most of their contribution at this level, since future levels
will not give much contribution. How can we do it, without assigning large values? This is where we
use the sawtooth function: it gives a good contribution between points without assigning any vertex
too large a value in any coordinate.
To bound the contraction and prove Lemma 5, for nodes u, v we need to show that there exists a
coordinate j such that Eα,β [|fj (v) − fj (u)|] = Ω(dG (u, v)). For brevity, define
∆uv := dG (u, v).
(3)
Fix c = 12. Let i be the minimal index such that there exists X ∈ Xi with u, v ∈ X, and at least one
of the following holds:
1. min {dX (v, PX ), dX (u, PX )} ≤ ∆uv /c (i.e., we choose a path close to {v, u}).
2. v and u are in different components of X \ PX .
By the definition of SPD such index i indeed exists. By the minimality of i, for every X 0 ∈ Xi0 such
that i0 < i and u, v ∈ X 0 , necessarily min{dX 0 (v, PX 0 ), dX 0 (u, PX 0 )} > ∆uv /c. Therefore, the ball with
radius ∆uv /c around each of v, u is contained in X. In particular, min {dG (v, V \ X), dG (u, V \ X)} >
∆uv /c.
path
Suppose first that (2) occurs but not (1). Let j be the coordinate in fX
created for the connected
component of v in X \ PX . Then
∆uv
∆uv ∆uv
path
path
,2
=
.
(fX )j (v) − (fX )j (u) = min {dX (v, PX ) , 2dG (v, V \ X)} − 0 ≥ min
c
c
c
Next assume that (1) occurs. W.l.o.g., dX (v, PX ) ≤ dX (u, PX ), so that dX (v, PX ) ≤ ∆uv /c. Suppose
path
first that dX (u, PX ) ≥ 2∆uv /c. Then in the coordinate j in fX
created for the connected component
of u in X \ PX , we have
path
path
(fX
)j (v) − (fX
)j (u) ≥ |min {dX (u, PX ) , 2dG (u, V \ X)} − min {dX (v, PX ) , 2dG (v, V \ X)}|
∆uv ∆uv
∆uv
∆uv
≥ min 2
,2
−
=
.
c
c
c
c
11
u
PX
2∆uv
c
∆uv
≥
0
u
r
v
∆uv
c
≥
v0
(resp. 2∆cuv ) from v 0
As u0 , v 0 lay on the same
(resp u ), it’s closest vertex on PX . By triangle inequality dX (v , u ) ≥ (1 −
0
0
0
0
shortest path starting at r, |dX (v , r) − dX (u , r)| = dX (v , u ). Using the triangle inequality again we conclude
|dX (v, r) − dX (u, r)| ≥ |dX (v 0 , r) − dX (u0 , r)| − 3c ∆uv ≥ (1 − 6c )∆uv .
Figure 5: PX is a shortest path with root r. v (resp. u) is at distance at most
0
0
0
3
c )∆uv .
∆uv
c
(It does not matter whether v, u are in the same connected component or not.) Thus it remains to
consider the case dX (u, PX ) < 2∆uv /c. Let r be the root of PX . Let v 0 (resp. u0 ) be the closest vertex
on PX to v (resp. u) in G[X]. Then by the triangle inequality
dX (v 0 , u0 ) ≥ dX (v, u) − dX (v, v 0 ) − dX (u, u0 ) ≥
c−3
∆uv .
c
In particular,
c−6
1
∆uv = ∆uv ,
(4)
c
2
where we used the fact that PX is a shortest path in G[X] (implying |dX (v 0 , r) − dX (u0 , r)| =
dX (v 0 , u0 )). See Figure 5 for illustration.
|dX (v, r) − dX (u, r)| ≥ dX (v 0 , r) − dX (u0 , r) − dX (v, v 0 ) − dX (u, u0 ) ≥
Set x = dX (v, r) and y = dX (u, r). Assume first that dG (v, V \ X) ≥ dG (u, V \ X). In particular,
tv ≥ tu . By the definition of tv , 2dG (v, V \ X) ≤ 2tv +1 . Thus
2tv ≥
∆uv
= Ω(∆uv )
c
Claim 1. Let t ≥ tv , then there is a constant φ such that
Eα,β gt (β · x + α · 2t+1 ) − gt (β · y + α · 2t+1 ) ≥ ∆uv /φ.
Proof. If |x − y| ≤ 2t−1 , then using Property 2 of Lemma 2
(4)
Eα,β gt (β · x + α · 2t+1 ) − gt (β · y + α · 2t+1 ) = Ω (|x − y|) = Ω(∆uv ) .
Otherwise, using Property 3 of Lemma 2
(5)
Eα,β gt (β · x + α · 2t+1 ) − gt (β · y + α · 2t+1 ) = Ω 2t ≥ Ω(∆uv ) .
Set S = max 8φ, 8c
2 . We consider two cases:
• If for some t ∈ {0, 1, . . . , M }, |pt − qt | · 2t > ∆Suv , then
root
root
Eα,β fX,t
(v) − fX,t
(u) = Eα,β pt · gt (β · x + α · 2t+1 ) − qt · gt (β · y + α · 2t+1 )
≥ pt · Eα,β gt (β · x + α · 2t+1 ) − qt · Eα,β gt (β · y + α · 2t+1 )
= |pt − qt | · 2t = Ω(∆uv ) .
Where the equality follows by Property 1 of Lemma 2.
12
(5)
• The other case is that for every t, |pt − qt | · 2t ≤ ∆Suv . Note that ptv + ptv +1 = (1 − λv ) + λv = 1.
Let t ∈ {tv , tv + 1} be such that pt ≥ 12 . Using (5), qt ≥ pt − 2∆tvuv·S ≥ 12 − 2∆2cuv · ∆Suv ≥ 41 . In
particular,
root
root
Eα,β fX,t
(v) − fX,t
(u) = Eα,β pt · gt (β · x + α · 2t+1 ) − qt · gt (β · y + α · 2t+1 )
≥ min {pt , qt } · Eα,β gt (β · x + α · 2t+1 ) − gt (β · y + α · 2t+1 ) − |pt − qt | · 2t
1 ∆uv
∆uv
≥ ·
−
= Ω(∆uv ) ,
4
φ
S
where in the first inequality we used Property 1 of Lemma 2, and in the second inequality we
used Claim 1.
Finally, recall that we assumed dG (v, V \ X) ≥ dG (u, V \ X) for the proof above. The other case
(dG (v, V \ X) < dG (u, V \ X)) is completely symmetric.
6
Conclusions
In this paper we introduced the notion of shortest path decompositions with low depth. We showed
how these can be used to give embeddings into `p spaces. Our techniques give optimal embeddings
of bounded pathwidth graphs into `2 , and also new embeddings for graphs with bounded treewidth,
as well as planar and excluded-minor families of graphs. Our embedding into `p admits a tight lower
bound for fixed p > 1. We hope that our techniques will be useful for other embedding results.
Our work raises several open questions. While our embeddings are tight for fixed p > 1, can we
improve the bounds for `1 embedding of bounded pathwidth graphs?
Can we give better results for
√
the SPDdepth of H-minor-free graphs? Our approach gives a O( log n)-distortion embedding of planar
graphs into `1 , which is quite different from the previous known results via padded decompositions:
can these ideas be used to make progress towards the planar graph embedding conjecture?
References
[ABN11]
Ittai Abraham, Yair Bartal, and Ofer Neiman. Advances in metric embedding theory. Advances in
Mathematics, 228(6):3026 – 3126, 2011.
[ACKW15] Ittai Abraham, Shiri Chechik, Robert Krauthgamer, and Udi Wieder. Approximate nearest neighbor search in metrics of planar graphs. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, APPROX/RANDOM 2015, August 24-26, 2015, Princeton,
NJ, USA, pages 20–42, 2015.
[AG06]
Ittai Abraham and Cyril Gavoille. Object location using path separators. In Proceedings of the
Twenty-Fifth Annual ACM Symposium on Principles of Distributed Computing, PODC 2006, Denver, CO, USA, July 23-26, 2006, pages 188–197, 2006.
[AGG+ 14] Ittai Abraham, Cyril Gavoille, Anupam Gupta, Ofer Neiman, and Kunal Talwar. Cops, robbers,
and threatening skeletons: padded decomposition for minor-free graphs. In Symposium on Theory
of Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 79–88, 2014.
[And86]
Thomas Andreae. On a pursuit game played on graphs for which a minor is excluded. Journal of
Combinatorial Theory, Series B, 41(1):37 – 47, 1986.
[Ass83]
P. Assouad. Plongements lipschitziens dans Rn . Bull. Soc. Math. France, 111(4):429–448, 1983.
13
[BBMN11] Nikhil Bansal, Niv Buchbinder, Aleksander Madry, and Joseph Naor. A polylogarithmiccompetitive algorithm for the k-server problem. In Foundations of Computer Science (FOCS),
2011 IEEE 52nd Annual Symposium on, pages 267 –276, oct. 2011.
[BCMN13] Yair Bartal, Douglas E. Carroll, Adam Meyerson, and Ofer Neiman. Bandwidth and low dimensional embedding. Theor. Comput. Sci., 500:44–56, 2013.
[BGHK92] Hans L. Bodlaender, John R. Gilbert, Hjálmtýr Hafsteinsson, and Ton Kloks. Approximating
treewidth, pathwidth, and minimum elimination tree height, pages 1–12. Springer Berlin Heidelberg,
Berlin, Heidelberg, 1992.
[Bou85]
Jean Bourgain. On Lipschitz embedding of finite metric spaces in Hilbert space. Israel J. Math.,
52(1-2):46–52, 1985.
[Car04]
N. L. Carothers. A Short Course on Banach Space Theory. London Mathematical Society Student
Texts. Cambridge University Press, 2004.
[CGN+ 06] Chandra Chekuri, Anupam Gupta, Ilan Newman, Yuri Rabinovich, and Alistair Sinclair. Embedding k-outerplanar graphs into `1 . SIAM J. Discrete Math., 20(1):119–136, 2006.
[CJLV08]
Amit Chakrabarti, Alexander Jaffe, James R. Lee, and Justin Vincent. Embeddings of topological
graphs: Lossy invariants, linearization, and 2-sums. In 49th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2008, October 25-28, 2008, Philadelphia, PA, USA, pages
761–770, 2008.
[GKL03]
Anupam Gupta, Robert Krauthgamer, and James R. Lee. Bounded geometries, fractals, and
low-distortion embeddings. In FOCS ’03: Proceedings of the 44th Annual IEEE Symposium on
Foundations of Computer Science, page 534, Washington, DC, USA, 2003. IEEE Computer Society.
[GKR04]
Anupam Gupta, Amit Kumar, and Rajeev Rastogi. Traveling with a Pez dispenser (or, routing
issues in MPLS). SIAM J. Comput., 34(2):453–474, 2004.
[GNRS04] Anupam Gupta, Ilan Newman, Yuri Rabinovich, and Alistair Sinclair.
embeddings of graphs. Combinatorica, 24(2):233–269, 2004.
Cuts, trees and `1 s-
[GTW13]
Anupam Gupta, Kunal Talwar, and David Witmer. On the non-uniform sparsest cut problem on
bounded treewidth graphs. In Annual ACM Symposium on Theory of Computing, pages 281–290,
Jun 2013.
[JLM11]
Alexander Jaffe, James R. Lee, and Mohammad Moharrami. On the optimality of gluing over
scales. Discrete & Computational Geometry, 46(2):270–282, Sep 2011.
[JS09]
William b. Johnson and Gideon Schechtman. Diamond graphs and super-reflexivity. Journal of
Topology and Analysis, 01(02):177–189, 2009.
[KK16]
Lior Kamma and Robert Krauthgamer. Metric decompositions of path-separable graphs. Algorithmica, pages 1–9, 2016.
[KKM+ 12] Maleq Khan, Fabian Kuhn, Dahlia Malkhi, Gopal Pandurangan, and Kunal Talwar. Efficient
distributed approximation algorithms via probabilistic tree embeddings. Distributed Computing,
25(3):189–205, 2012.
[KLMN05] Robert Krauthgamer, James R. Lee, Manor Mendel, and Assaf Naor. Measured descent: a new
embedding method for finite metrics. Geometric and Functional Analysis, 15(4):839–858, 2005.
[LLR95]
N. Linial, E. London, and Y. Rabinovich. The geometry of graphs and some of its algorithmic
applications. Combinatorica, 15(2):215–245, 1995.
[LN04]
James R. Lee and Assaf Naor. Embedding the diamond graph in `p and dimension reduction in `1 .
Geometric & Functional Analysis GAFA, 14(4):745–747, 2004.
[LR10]
James R. Lee and Prasad Raghavendra. Coarse differentiation and multi-flows in planar graphs.
Discrete & Computational Geometry, 43(2):346–362, 2010.
14
[LS10]
James R. Lee and Anastasios Sidiropoulos. Genus and the geometry of the cut graph. In Proceedings
of the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2010, Austin,
Texas, USA, January 17-19, 2010, pages 193–201, 2010.
[LS13]
James R. Lee and Anastasios Sidiropoulos. Pathwidth, trees, and random embeddings. Combinatorica, 33(3):349–374, 2013.
[Mil86]
Gary L. Miller. Finding small simple cycle separators for 2-connected planar graphs. J. Comput.
Syst. Sci., 32(3):265–279, 1986.
[MN13]
Manor Mendel and Assaf Naor. Markov convexity and local rigidity of distorted metrics. Journal
of the European Mathematical Society, 15(1):287–337, 2013.
[NR03]
Ilan Newman and Yuri Rabinovich. A lower bound on the distortion of embedding planar metrics
into Euclidean space. Discrete Comput. Geom., 29(1):77–81, 2003.
[OS81]
Haruko Okamura and P.D. Seymour. Multicommodity flows in planar graphs. Journal of Combinatorial Theory, Series B, 31(1):75 – 81, 1981.
[Rao99]
Satish Rao. Small distortion and volume preserving embeddings for planar and Euclidean metrics.
In Proceedings of the Fifteenth Annual Symposium on Computational Geometry, Miami Beach,
Florida, USA, June 13-16, 1999, pages 300–306, 1999.
[ST04]
Yuval Shavitt and Tomer Tankel. Big-bang simulation for embedding network distances in Euclidean
space. IEEE/ACM Trans. Netw., 12(6):993–1006, 2004.
[Tho04]
Mikkel Thorup. Compact oracles for reachability and approximate distances in planar digraphs. J.
ACM, 51(6):993–1024, November 2004.
A
Lower Bound: Proof of Theorem 3
We start with the definition of the diamond graphs Dk .
Definition 2 (Diamond Graphs). Let D0 , D1 , D2 , . . . be a sequence of graphs defined as follows: D0
is a single edge, and for i ≥ 1, Di is obtained from Di−1 by replacing every edge of Di−1 with a square
with two new vertices. See Figure 6 for illustration.
For each of the new squares created at level i, call the two new vertices a diagonal at level-i.
Consider the graph Dk . For 1 ≤ i ≤ k, denote by Di the set of all level-i diagonals, and let Ei be the
set of pairs of vertices which were edges in Di . It holds that |Ek | = 4k and Dk = 4k−1 . Moreover, D0
has pathwidth 1, and it can be verified by induction that Dk has pathwidth k + 1.
√
Newman and Rabinovich [NR03] proved that every embedding of Dk into `2 requires distortion k + 1.
Lee and Naor [LN04]
p generalized it for 1 < p ≤ 2, by proving that every embedding of Dk into `p
requires distortion 1 + (p − 1)k. We will generalize further, specifically, we will prove that for p ≥ 2,
1/p
every embedding of Dk into `p requires distortion at least 2k+1
.
p−2
The following claim will be essential.
Claim 2 (`p Quadrilateral Inequality). For p ≥ 2, and every four vectors a, b, c, d ∈ `p , it holds that
ka − ckpp + kb − dkpp ≤ 2p−2 ka − bkpp + kb − ckpp + kc − dkpp + kd − akpp
(6)
15
v
a1
b1
v
s
t
s
t
a2
b2
c2
d2
s
t
u
d1
c1
u
D0
D1
D2
Figure 6: The first 3 diamond graphs. {s, t} is the level 0 diagonal, {u, v} is the level 1 diagonal, {a1 , a2 },
{b1 , b2 }, {c1 , c2 }, {d1 , d2 } are the level 2 diagonals. E0 = {{s, t}}, E1 = {{s, u}, {t, u}, {t, v}, {s, v}} .
Proof. The proof of the following inequality can be found at [Car04, Theorem 11.12],
∀x, y ∈ `p , kx + ykpp + kx − ykpp ≤ 2p−1 kxkpp + kykpp .
Define x1 = b − a, y1 = a − d and x2 = b − c, y2 = c − d. We get
kb − dkpp + kb − 2a + dkpp ≤ 2p−1 ka − bkpp + kd − akpp .
kb − dkpp + kb − 2c + dkpp ≤ 2p−1 kb − ckpp + kc − dkpp .
By summing up and dividing by 2,
kb −
dkpp
+
kb − 2a + dkpp + kb − 2c + dkpp
2
≤ 2p−2 ka − bkpp + kb − ckpp + kc − dkpp + kd − akpp . (7)
The claim now follows by convexity.
Fix some p ≥ 2, and embedding f : Dk → `p . We will assume w.l.o.g that f is non-contractive, and
1
1
denote by ρ its expansion, i.e. distortion. Set α0 = 2k(p−2)
and for i > 0, αi = 2(k−i+1)(p−2)
. Note that
p−2
for i ≥ 1, αi · 2
= αi+1 . Our proof will be based on the following Poincaré-type inequality
Claim 3 (Diamond graph `p Poincaré-type inequality).
k
X
i=0
αi ·
X
{x,y}∈Di
kf (x) − f (y)kpp ≤ αk+1 ·
16
X
{x,y}∈Ek
kf (x) − f (y)kpp .
(8)
Proof. For edge {x, y} ∈ Ei−1 , denote by {x0 , y 0 } ∈ Di the diagonal created by it. We have
X
X
kf (x) − f (y)kpp +
kf (x) − f (y)kpp
{x,y}∈Di
{x,y}∈Ei−1
X
=
{x,y}∈Ei−1
p
kf (x) − f (y)kpp + f x0 − f y 0 p
(6)
≤ 2p−2 ·
= 2p−2 ·
X
{x,y}∈Ei−1
X
{x,y}∈Ei
f (x) − f x0
p
p
+ f x0 − f (y)
p
p
+ f (y) − f y 0
p
p
+ f y 0 − f (x)
p
p
kf (x) − f (y)k22 .
Summing over 1 ≤ i ≤ k’s, with appropriate scaling,
k
X
X
X
kf (x) − f (y)kpp
αi ·
kf (x) − f (y)kpp +
i=1
{x,y}∈Ei−1
{x,y}∈Di
≤
k
X
i=1
p−2
αi · 2
·
X
{x,y}∈Ei
(y)kpp
kf (x) − f
=
k
X
i=1
αi+1 ·
X
{x,y}∈Ei
kf (x) − f (y)kpp .
Hence
k
X
i=1
αi ·
X
{x,y}∈Di
kf (x) − f (y)kpp + α1 ·
X
kf (x) − f (y)kpp ≤ αk+1 ·
{x,y}∈E0
X
{x,y}∈Ek
kf (x) − f (y)kpp .
As E0 = D0 and α0 = α1 , the claim follows.
Next, we calculate
k
X
i=0
αi ·
X
{x,y}∈Di
k
p X
p
d (x, y)p =α0 · 2k +
|Di | · αi · 2k−i+1
i=1
= 22k +
k
X
i=1
2
4i−1 · 2k−i+1 = (k + 1) · 4k .
Recall that f is non-contractive and has expansion ρ. Consequently,
(9)
(k + 1) · 4k =
≤
k
X
i=0
k
X
i=0
αi ·
αi ·
(8)
≤ αk+1 ·
≤ αk+1 ·
X
{x,y}∈Di
X
{x,y}∈Di
X
{x,y}∈Ek
X
{x,y}∈Ek
We conclude
ρ≥
d (x, y)p
4k k + 1
·
|Ek | αk+1
kf (x) − f (y)kpp
kf (x) − f (y)kpp
(ρ · d(x, y))p
1/p
=
17
k+1
2p−2
= αk+1 · (ρ)p · |Ek | .
1/p
= Ω(k /p ) .
1
(9)
B
Proof of Lemma 6
We restate the lemma for convenience:
Lemma 6 (Composition Lemma). Let (X, d) be a metric space. Suppose that there are constants ρ, τ
and a function f : X → Rs , drawn from some probability space such that:
1. For every u, v ∈ X and every j ∈ [s], |fj (v) − fj (u)| ≤ ρ · d(v, u).
2. For every u, v ∈ X, there exists j ∈ [s] such that E[ |fj (v) − fj (u)| ] ≥
1
τ
· d(v, u).
3. For every v ∈ X, there is a subset of indices Iv ⊆ [s] of size |Iv | ≤ k, such that for every j ∈
/ Iv ,
fj (v) = 0. In other words, for every v ∈ X, f (v) has support of size at most k.
Then, for every p ≥ 1, there is an embedding of (X, d) into `p with distortion O(k 1/p ). Moreover,
if there is an efficient algorithm for sampling such an f , then there is a randomized algorithm that
constructs the embedding efficiently (in expectation).
Proof. Fix n = |X|, and set m = 48ρτ · ln n. Let f (1)L
, f (2) , . . . , f (m) : X → Rs be functions chosen i.i.d
1/p
m
(i)
−
according to the given distribution. Set g = m
i=1 f . We argue that with high probability, g
1/p
has distortion O(k ) in `p .
Fix some pair of vertices v, u ∈ V . Set d(v, u) = ∆. The upper bound follows from Property 1 and
Property 3 of the lemma:
kg(v) − g(u)kpp =
≤
m
X
X
i=1 j∈Iv ∪Iu
m
X
X
i=1 j∈Iv ∪Iu
(i)
(i)
m− /p · fj (v) − fj (u)
1
p
1
· (ρ · ∆)p ≤ 2k · (ρ · ∆)p ,
m
thus kg(v) − g(u)kp ≤ O(k 1/p · ∆).
Next, for the contraction bound, let j be the index of Property 2 w.r.t v, u. Set F = {f : |fj (v) − fj (u)| ≥
∆/2τ } to be the event that we draw a function with significant contribution to v, u. Then using Property 1 and Property 2,
∆
∆
∆
≤ E [|fj (v) − fj (u)|] ≤ Pr F ·
+ Pr [F] · ρ∆ ≤
+ Pr [F] · ρ∆ ,
τ
2τ
2τ
(i)
1
which implies that Pr [F] ≥ 2ρτ
. Let Qu,v be an indicator random variable for the event f (i) ∈ F, and
Pm
(i)
m
set Qu,v = i=1 Qu,v . By linearity of expectation, E[Qu,v ] ≥ 2ρτ
= 24 · ln n. By a Chernoff bound
1
1
Pr [Qu,v ≤ 12 · ln n] ≤ Pr Qu,v ≤ · E [Qu,v ] ≤ exp(− E [Qu,v ]) ≤ exp(−3 ln n) = n−3 .
2
8
By taking a union bound over the n2 pairs, with probability at least 1 − n1 , for every u, v ∈ V ,
Qu,v > 12 ln n = Ω(m). (Recall that both ρ, τ are universal constants.) If this event indeed occurs,
18
then the contraction is indeed bounded:
kg(v) − g(u)kpp ≥
≥
m
X
i=1
1
m
(i)
(i)
m− /p · fj (v) − fj (u)
1
X
(i)
(i)
p
fj (v) − fj (u)
(i)
p
p
i:Qu,v =1
Qu,v
≥
·
m
∆
2ρτ
=Ω
∆
2ρτ
p
.
In particular, for every u, v, kg(v) − g(u)kp ≥ Ω(∆).
C
Proof of the Sawtooth Lemma (Lemma 2)
We restate Lemma 2 for convenience:
Lemma 2 (Sawtooth Lemma). Let x, y ∈ R+ . Let α ∈ [0, 1], β ∈ [0, 4] be drawn uniformly and
independently. The following properties hold:
1. Eα,β gt (βx + α · 2t+1 ) = 2t−1 .
2. If |x − y| ≤ 2t−1 , then Eα,β gt (βx + α · 2t+1 ) − gt (βy + α · 2t+1 ) = Ω(|x − y|).
3. If |x − y| > 2t−1 , then Eα,β gt (βx + α · 2t+1 ) − gt (βy + α · 2t+1 ) = Ω(2t ).
Property 1 is straightforward,
as by Observation
1 gt is
with period length 2t+1 . Indeed, for
periodic
t+1
t+1
t−1
every fixed β, Eα gt (βx + α · 2 ) = Eα gt (α · 2 ) = 2 . The following claim will be useful in
the proofs of Property 2 and Property 3.
(2t+1 −z )z
Claim 4. For z ∈ [0, 2t+1 ], Eα∈[0,1] gt (z + α · 2t+1 ) − gt (α · 2t+1 ) = 2t+1 .
Proof. Set (∗) = Eα∈[0,1] gt (z + α · 2t+1 ) − gt (α · 2t+1 ) . By substituting the variable of integration,
R 2t+1
1
· 0
|gt (z + α) − gt (α)| dα. First assume that z ≤ 2t , then there are 5 “phase changes”
(∗) = 2t+1
in |gt (z + α) − gt (α)| from 0 to 2t+1 , at 2t − z, 2t − z2 , 2t , 2t+1 − z, 2t+1 − z2 . (see Figure 7 for
illustration). We calculate
t+1
2
· (∗) =
Z
2t −z
Z
zdα +
0
=2·
0
Z
0
z
2
(z − 2α)dα +
2t −z
zdα + 2 ·
Z
z
2
0
Z
z
2
Z
2αdα +
0
2t −z
Z
zdα +
0
0
z
2
(z − 2α)dα +
Z
z
2
2αdα
0
zdα = 2t+1 − z z .
For z > 2t , set w = 2t+1 − z. Then using that g t is periodic,
Eα∈[0,1] gt (w + α · 2t+1 ) − gt (α · 2t+1 ) = Eα∈[0,1] gt (w + z + α · 2t+1 ) − gt (z + α · 2t+1 )
= Eα∈[0,1] gt (2t+1 + α · 2t+1 ) − gt (z + α · 2t+1 )
= Eα∈[0,1] gt (z + α · 2t+1 ) − gt (α · 2t+1 ) .
Hence by the first case, (∗) =
(2t+1 −w)w
2t+1
=
(2t+1 −z )z
2t+1
19
.
2t
2t
2t
2t+1
2t+1
2t
2t
α = 2t
2t+1
α = 2t − z2
α = 2t − z
α=0
2t
2t+1
2t+1
2t
2t+1
α = 2t+1 − z2
α = 2t+1 − z
α = 2t+1
2t+1
Figure 7: α is going from 0 to 2t+1 . z ≤ 2t . In each of the figures the leftmost red point represents α while
the rightmost red point represents z + α. Each of the middle figures represent a moment when gt (z + α) − gt (z)
changes its derivative.
For the proofs
of Property 2 and Propertyt+13 assume w.l.o.g x > y. Set z = x − y, and
(∗) = Eα,β gt (βx + α · 2t+1 ) − gt (βy + α · 2
) . As gt is a periodic function, we have that
t+1
t+1
(∗) = Eβ Eα gt (βz + α · 2 ) − gt (α · 2 ) .
Proof of Property 2. Using Claim 4, we have
"
#
t+1
2t+1 − βz βz
1
1
2 z 2 z2 3 4
= t+1 · ·
β −
· β |0
(∗) = Eβ
2t+1
2
4
2
3
1
16
z 2 26
1
64
2
= ·
· z − t+1 ·
≥ · 8−
· z = · |x − y| ,
4
2
2
3
4
3·4
3
where in the inequality we used that z ≤ 2t−1 .
Proof of Property 3. As gt is periodic function, Claim 4 implies that for every w ≥ 0 it holds that
(2t+1 −(w mod 2t+1 ))(w mod 2t+1 )
Eα gt (w + α · 2t+1 ) − gt (α · 2t+1 ) =
. Let a ∈ [0, 4] such that a · z =
2t+1
t+1
t−1
2
(such a exists as |x − y| > 2 ). The claim follows as,
(∗) · 2t+1 = Eβ∈[0,4] 2t+1 − (βz mod 2t+1 ) (βz mod 2t+1 )
≥
4
−1
bX
ac
i=0
bX
4
a
=
c−1
i=0
1
·
4
(i+1)a
Z
t+1
2
2t+1
− (β ·
a
mod 2
t+1
2t+1
) (β ·
a
mod 2t+1 )dβ
ia
1
·
4
Za
0
2t+1 − β ·
2t+1
a
·β·
2t+1
dβ
a
2Zt+1
4
1
a
· · t+1 ·
=
2t+1 − γ · γdγ
a
4 2
0
3
2
2
2t+1
2t+1
4 1
a
γ 3 2t+1
1
1
t+1 γ
≥
· ·
· 2
−
|
= · t+1 ·
=
.
2a 4 2t+1
2
3 0
2 2
6
12
20
| 8cs.DS
|
arXiv:1404.5772v3 [cs.IR] 28 Jul 2014
Sequential Click Prediction for Sponsored Search
with Recurrent Neural Networks
Yuyu Zhang∗
Hanjun Dai∗
Chang Xu∗
Jun Feng
Institute of Computing Technology
Chinese Academy of Sciences
[email protected]
Fudan University
[email protected]
Nankai University
[email protected]
Tsinghua University
[email protected]
Taifeng Wang
Jiang Bian
Bin Wang
Tie-Yan Liu
Microsoft Research
[email protected]
Microsoft Research
[email protected]
Institute of Computing Technology
Chinese Academy of Sciences
[email protected]
Microsoft Research
[email protected]
Abstract
Click prediction is one of the fundamental problems
in sponsored search. Most of existing studies took advantage of machine learning approaches to predict ad
click for each event of ad view independently. However,
as observed in the real-world sponsored search system,
user’s behaviors on ads yield high dependency on how
the user behaved along with the past time, especially
in terms of what queries she submitted, what ads she
clicked or ignored, and how long she spent on the landing pages of clicked ads, etc. Inspired by these observations, we introduce a novel framework based on Recurrent Neural Networks (RNN). Compared to traditional
methods, this framework directly models the dependency on user’s sequential behaviors into the click prediction process through the recurrent structure in RNN.
Large scale evaluations on the click-through logs from
a commercial search engine demonstrate that our approach can significantly improve the click prediction accuracy, compared to sequence-independent approaches.
Introduction
Sponsored search has been a major business model for modern commercial Web search engines. Along with organic
search results, it presents to users with sponsored search results, i.e., advertisements (ads) targeting to the search query.
Sponsored search accounts for the overwhelming majority
of income for three major search engines: Google, Yahoo
and Bing. Even in the US search market alone, it generates over 20 billion dollars per year, the amount of which
still keeps rising1 . According to the common cost-per-click
(CPC) model for sponsored search, advertisers are only
charged once their advertisements are clicked by users. In
this mechanism, to maximize the revenue for search engine
and maintain a desirable user experience, it is crucial for
search engines to estimate the click-through rate (CTR) of
ads.
∗
This work was performed when the first three authors were
visiting Microsoft Research Asia.
Copyright c 2014, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
1
Source: eMarketer, June 2013.
Recently, click prediction has received much attention
from both industry and academia (Fain and Pedersen 2006;
Jansen and Mullen 2008). State-of-the-art sponsored search
systems typically employ machine learning approaches to
predict the click probability by using the feature extracted
based on: 1) historical CTR for ad impressions 2 with respect to different elements, e.g., CTR of query, ad, user, and
their combinations (Graepel et al. 2010; Richardson, Dominowska, and Ragno 2007); 2) semantic relevance between
query and ad (Radlinski et al. 2008; Shaparenko, Çetin, and
Iyer 2009; Hillard et al. 2011).
However, most of previous works take single ad impression as the input instance to predict the click probability,
without considering dependency between different ad impressions. Some recent research work (Xu, Manavoglu, and
Cantu-Paz 2010; Xiong et al. 2012) pointed out that they
could achieve more accurate click prediction by modeling
spatial relationship between ad slots in the same query session. Inspired by them, we conduct further data analysis to
study other types of dependency between user’s behaviors
in sponsored search, through which we find that user’s behaviors also yield explicit temporal dependency. For example, if a user clicks an ad, comes to the ad landing page,
but closes it very quickly, the click probability of her next
view of this ad will become fairly low; moreover, if a user
has previously submitted a query on booking flight, he/she
will be with higher probability to click the ads under flight
booking. These findings motivate us to advance the state-ofthe-art of click prediction by modeling the important temporal dependency into the click prediction process. Although
some kinds of dependency can be modeled as features, it’s
still hard to identify all of them explicitly. Thus, it is necessary to empower the model with the ability to extract and
leverage various kinds of dependency automatically.
In real-world sponsored search system, the event of any
ad impression, click, and corresponding context information
(e.g. user query, ad text, click dwell time, etc.) is recorded
with the time stamp in search logs. Thus, it is natural to
employ time series analysis methods to model sequential
dependency between user’s behaviors. Previous studies on
2
We refer to a certain ad shown to a particular user in a specific
search result page as an ad impression.
• We investigate the sequential dependency among particular user’s ad impressions, and identify several important
sequential dependency relationships.
• We use Recurrent Neural Networks to model user’s click
sequence, and successfully incorporate sequential dependency into enhancing the accuracy of click prediction.
• We conduct large scale experiments to validate the RNN
model’s effectiveness for modeling sequential data in
sponsored search.
In the following parts of this paper, we will first present
our data analysis results to verify the potential dependency
which might affects click prediction. Then, we propose our
RNN model for the task of sequence-based click prediction.
After that, we describe experimental settings followed by the
experimental results and further performance study. At last,
we summarize this paper and discuss some future work.
Data Analysis on Sequential Dependency
To gain more understanding on why sequential information
is important in click prediction, in this section, we will discuss the effects of sequential dependency from multiple per-
35%
30%
25%
CTR
time series analysis (Kirchgassner, Wolters, and Hassler
2012; Box, Jenkins, and Reinsel 2013) usually focused on
modeling trends or periodic patterns in data series. However,
the sequential dependency between ad impressions is so
complex and dynamic that time series analysis approaches
is not capable enough to model it effectively. On the other
hand, a few most recent studies leverage Recurrent Neural
Networks (RNN) to model the temporal dependency in data.
For example, RNN language model (Mikolov et al. 2010;
Mikolov et al. 2011a; Mikolov et al. 2011b) successfully
leverages long-span sequential information among the massive language corpus, which results in better performance
than traditional neural networks language model (Bengio
et al. 2006). Moreover, RNN based handwriting recognition (Graves et al. 2009), speech recognition (Kombrink et
al. 2011), and machine translation (Auli et al. 2013) systems
have also led to much improvement in the corresponding
tasks. Compared to traditional feedforward neural networks,
RNN has demonstrated its strong capability to exploit dependencies in the sequence due to its specific recurrent network structure.
In this work, we propose to leverage RNN to model sequential dependency into predicting ad click probability. We
consider each user’s ad browsing history as one sequence
which yields the intrinsic internal dependency. In the training process of RNN model, features of each ad impression
will be feedforwarded into the hidden layer, together with
previously accumulated hidden state. In this way, the dependency among impressions will be embedded into the recurrent network structure. Our experiments on the large scale
data from a commercial search engine reveal that, such RNN
structure can give rise to a significant improvement on the
click prediction accuracy compared with the state-of-the-art
dependency-free models such as Neural Networks and Logistic Regression.
The main contributions of this paper are in three folds:
20%
15%
10%
5%
0%
0
50
100
150
200
Last Click Dwell Time (seconds)
250
300
Figure 1: Correlation between last click dwell time and current click-through rate.
spectives. We collect data for analysis from the logs of a
commercial sponsored search system.
Once a user clicks an ad, she will enter into the corresponding ad landing page and stay for a certain period of
time, which is referred to as the click dwell time. Generally,
longer dwell time implies better user experience. For a particular user, all the ad impressions can be organized as an ordered sequence along with the time. To explore the sequential effect of click dwell time, we first pick up all the “first
clicks” in each user’s sequence, and track whether each ad
will be clicked again in its consecutively next impression.
The correlation between the previous click dwell time and
the current click-through rate3 is shown in Figure 1. From
this figure, it is clear to observe an obvious positive correlation, i.e., the longer a user stays on an ad’s landing page, the
more likely she will click this ad right at the next time.
When the click dwell time is less than 20 seconds, we call
such ad click as a “quick back” one. From the data analysis above, we can observe that “quick back” can give rise to
rather lower click-through rate in the next impression, which
indicates that users tend to avoid clicking the ad once they
had an unsatisfied experience. However, it is not clear how
users behave if they experienced a “quick back” click long
time ago (e.g., half a month). To explore this, we collect all
the “quick back” clicks and calculate the click-through rate
after different time intervals, i.e., the time elapsed since the
“quick back” click, respectively. Figure 2 shows the result,
where the time interval is binning by half-days. As conjectured, along with increasing elapsed time, the overall clickthrough rate grows significantly and then stay steadily in
a certain level, which implies that users tend to gradually
forget the unsatisfied experience with the time passing.
Besides studies on the sequential effect of dwell time,
we further analyze such effects from the aspect of query. In
sponsored search systems, ads are automatically selected by
matching with the user submitted query. Queries also form
a time ordered sequence, binding on that of ad impressions.
After categorizing the queries into topics based on our proprietary query taxonomy, we calculate users’ click-through
rate when they submit each query topic for the first time, and
compare with their future CTR on subsequent queries of the
same query topic. Analysis results, as shown in Figure 3, il3
In this paper, we present relative click-through rate to preserve
the proprietary information.
35%
i(t)
Feedforward
30%
U
V
CTR
25%
R
15%
10%
h(t-1)
i(t-2)
5%
0%
0
h(t)
i(t-1)
20%
y(t)
5
10
15
20
25
30
35
40
TimefElapsedfSincefLastf)QuickfBack)fClickf(half−days)
45
Figure 2: Click-through rate right after a “quick back” click
on the same ad.
h(t-2)
Backpropagation
h(t-3)
3.5%
Figure 4: RNN training process with BPTT algorithm. Unfolding step is set to 3 in this figure.
3.0%
2.5%
2.0%
1.5%
y(t) = σ(h(t)V T ),
1.0%
−2z
0.5%
0%
Retail
Flight
First Submission
Health
Movie
Subsequent Submission
Figure 3: Click-through rate on users’ first and subsequent
submissions of a certain type of query.
lustrate that if a user has submitted a query belonging to a
certain query topic, he/she will become more likely to click
the ads under the same topic.
All the analysis results above indicate that user’s previous
behaviors in sponsored search may cause strong but quite
dynamic impact her subsequent behaviors. As long as we
can identify such sequential dependency between user behaviors, we can design features accordingly to enhance the
click prediction. However, since a big challenge to enumerate such dependency in the data manually, it is necessary to
let the model have the ability to learn such kind of dependency by itself. To this end, we propose to leverage a widely
used framework, Recurrent Neural Network, as our model,
as it naturally embeds dependencies in the sequence.
The Proposed Framework
Model
The architecture of a recurrent neural network is shown in
Figure 4. It consists of an input layer i, an output unit, a
hidden layer h, as well as inner weight matrices. Here we
use t ∈ N to represent the time stamp. For example, h(t)
denotes the hidden state in time t. Specifically, the recurrent
connections R between h(t − 1) and h(t) can propagate
sequential signals. The input layer consists of a vector i(t)
that represents the features of current user behaviors, and
the vector h(t − 1) represents the values in the hidden layer
computed from the previous step.
The activation values of the hidden and output layers are
computed as
h(t) = f (i(t)U T + h(t − 1)RT ),
where f (z) = 1−e
1+e−2z is the tanh function we use for nonlinear activation, and σ(z) = 1+e1−z is the sigmoid function
for predicting the click probability.
The hidden layer can be considered as an internal memory which records dynamic sequential states. The recurrent
structure is able to capture a long-span history context of
user behaviors. This makes RNN applicable to the tasks related to sequential prediction.
In our framework, i(t) represents the features correlated
to user’s current behavior and h(t) represents sequential information of user’s previous behaviors. Thus, our prediction
depends on not only the current input features, but also the
sequential historical information.
Feature Construction
In our study, we take ad impressions as instances for both
model training and testing. Based on the rich impressioncentric information, we construct the input features that
can carry crucial information to achieve accurate CTR prediction for a given impression. All these features can be
grouped into several general categories: 1) Ad features consist of information about ad ID, ad display position, and ad
text relevance with query. 2) User features include user ID
and query (submitted by user) related semantic information.
3) Sequential features include time interval since the last impression, dwell time on the landing page of the last click
event, and whether the current impression is the head of sequence, i.e., whether it is the first impression for this user.
With all these diverse types of features, each ad impression can be described in a large and complex feature space.
In this paper, all of the click prediction models will follow this input feature setting, so that we can fairly compare
their capabilities of predicting whether an impression will
be clicked by user.
Data Organization
To effectively harness the sequential information for modeling temporal dependencies, the training process requires
large quantity of data. Luckily, the data in computational
advertising is big enough to make RNN tractable. To obtain the data for sequential prediction, we re-organize the
input features along with the user dimension (i.e., reordered
each user’s historical behaviors according to the timeline).
In particular, for the ads in the same search session, we rank
them by their natural display orders in the mainline and then
sidebar.
i(t)
R
t+1
t
V
U
Test Samples (user behavior sequence)
y(t)
h(t)
t-1
Feedforward
h(t-1)
Learning
Loss Function In our work, the loss function is defined
as an averaged cross entropy, which aims at maximizing the
likelihood of correct prediction,
L=
M
1 X
(−yi log(pi ) − (1 − yi )log(1 − pi )),
M i=1
where M is the number of training samples. The ith sample is labeled with yi ∈ {0, 1} and pi is the predicted click
probability of the given ad impression.
Learning Algorithm (BPTT) RNN can be trained in the
same way as normal feedforward network using backpropagation algorithm. In this way, basically, the state of the hidden layer from previous time step is simply regarded as an
additional input. With only one hidden layer, the network
tries to optimize prediction of the next sample given the
previous sample and previous hidden state. However, no effort is directly devoted towards longer context information,
which may hurt the performance of RNN.
A simple extension of the training algorithm is to unfold
the network and backpropagate errors even further. This is
called Back Propagation Through Time (BPTT) algorithm.
BPTT was proposed in (Rumelhart, Hinton, and Williams
2002), and has been used in the practical application of RNN
language model (Mikolov 2012).
We illustrate the overall training pipeline that applies
BPTT to the RNN based click prediction models in Figure 4.
Such unfolded RNN can be viewed as a deep neural network
with T hidden layers where the recurrent weight matrices are
shared and identical. In this approach, the hidden layer can
actually exploit the information of the most recent inputs and
put more importance to the latest input, which is essentially
coherent with the the sequential dependency. In the following part, we use T to denote the number of unfolding steps
in the BPTT algorithm.
The network is trained by Stochastic Gradient Descent
(SGD). The gradient of the output layer is computed as
Figure 5: RNN testing process with sequential input samples. The hidden state of previous test sample will be used
as input, together with the current sample features.
where ∗ represents the element-wise product, and ~1 is a vector with all elements equal to one.
Errors are also recursively propagated from the hidden
layer h(t − τ ) to the hidden layer from previous step
h(t − τ − 1), that is
eh (t−τ −1) = eh (t−τ )R∗(~1−h(t−τ −1)∗h(t−τ −1)),
where τ ∈ [0, T ). The weight matrix U and the recurrent
weights R are then updated as
"T −1
#
X
U (t + 1) = U (t) − α
eh (t − z)T i(t − z) ,
z=0
R(t + 1) = R(t) − α
"T −1
X
#
T
eh (t − z) h(t − z − 1) .
z=0
Note that, in our practical experiments, we add the bias
terms and L2 penalty of weights to the model, and the gradients can still be computed easily based on a slight modification to the equations above.
Inference
eo (t) = y(t) − l(t),
In contrast to traditional neural networks, RNN has a recurrent layer to store the previous hidden state. In the inference
phase, we still need to store the hidden state of previous test
sample, and feedforward it with the recurrent weights R.
Figure 5 illustrates the testing process. The test data is also
organized as ordered user behavior sequences. We feedforward current sample features, together with the hidden state
of previous sample to get the current hidden state. Then we
make the prediction and replace the stored hidden state with
current values. Here we only record the hidden state of the
last test sample, no matter how many unfolding steps there
are in the BPTT training process.
where y(t) is the predicted click probability, and l(t) is the
binary true label according to the ad is clicked or not. The
weights V between the hidden layer h(t) and output unit
y(t) are updated as
This section first describes the settings of our experiments,
and then reports the experimental results.
V (t + 1) = V (t) − α × eo (t) × h(t),
where α is the learning rate. Then, gradients of errors are
propagated from the output layer to the hidden layer as
eh (t) = eo (t)V ∗ (~1 − h(t) ∗ h(t)),
Experiments
Data Setting
To validate whether the RNN model we proposed can really help enhance the click prediction accuracy, we conduct
a series of experiments based on the click-through logs of
a commercial search engine. In particular, we collect half a
Table 1: Statistics of the dataset for training and testing click
prediction models.
Training
Testing
Ad Impressions
3,740,980
3,741,500
Ads
1,363,687
1,379,581
Users
235,215
235,215
Table 2: Overall performance of different models in terms of
AUC and RIG.
Model AUC
RIG
LR
87.48% 22.30%
NN
88.51% 23.76%
RNN
88.94% 26.16%
month logs from November 9th to November 22nd in 2013
as our experimental dataset. And, we randomly sample a
set of search engine users (fully anonymized) and collect
their corresponding events from the original whole traffic.
Finally, we collect over 7 million ad impressions in this period of time. After that, we use the first week’s data to train
click prediction models, and apply those models to the second week’s data for testing. Detailed statistics of the dataset
can be found in Table 1.
Evaluation Metrics
In our work, there are multiple models to be applied to predict the click probability for ad impressions in the testing
dataset. We use recorded user actions, i.e., click or non-click,
in logs as the true labels. To measure the overall performance
of each model, we follow the common practice in previous
click prediction research in sponsored search and employ
Area Under ROC Curve (AUC) and Relative Information
Gain (RIG) as the evaluation metrics (Graepel et al. 2010;
Xiong et al. 2012; Wang et al. 2013).
Compared Methods
In order to investigate the model effectiveness, we compare the performance of our RNN model with other classical click prediction models, including Logistic Regression
(LR) and Neural Networks (NN), with identical feature set
as described aforementioned. We set LR and NN as baseline
models due to the following reasons: 1) Quite a few previous studies (Richardson, Dominowska, and Ragno 2007;
McMahan et al. 2013; Wang et al. 2013) have demonstrated
that they are state-of-the-art models for click prediction in
sponsored search. 2) LR and NN models ignore the sequential dependency among the data, while our RNN based
framework is able to model such information. Through the
comparison with them, we will see whether RNN can successfully leverage dependencies in the data sequence to help
improve the accuracy of click prediction.
Experimental Results
Overall Performance For fair model comparison, we
carefully select the parameters of each model with cross validation, and ensure every model achieve its best performance
respectively. To be more specific, parameters for grid search
include: the coefficient of L2 penalty, the number of training
epochs, the hidden layer size for RNN and NN models, and
the number of unfolding steps for RNN. Finally, we get the
best settings of parameters as follows: the coefficient of L2
penalty is 1e − 6, the number of training epochs is 3, the
hidden layer size is 13, and the number of unfolding steps
should be 3 (more details will be provided later).
Table 2 reports the overall AUC and RIG of all three methods on test dataset. It demonstrates that our proposed RNN
model can significantly improve the accuracy of click prediction, compared with baseline approaches. In particular,
in terms of RIG, there is about 17.3% relative improvement
over LR, and about 10% relative improvement over NN. As
for the metric of AUC, we can find there is about 1.7% relative gain over LR, and about 0.5% relative gain over NN.
In real sponsored search system, such improvement in click
prediction accuracy will lead to a significant revenue increment.
The overall performance above shows the effectiveness
of our RNN model, which clearly transcends sequence independent models. Next, we will conduct detailed analysis
on how the sequential information help to get more accurate
click prediction.
Performance on Specific Ad Positions It is well-known
that the click-through rate on different ad positions varies a
lot, which is often referred to as the position bias. To further
check the performance of models within specific positions,
in our experiments, we separately analyze the performance
of RNN model and two baseline algorithms on different ad
positions: top first, mainline and sidebar.
Figure 6 shows the evaluation results on different positions. In Figure 6(a), RNN outperforms NN and LR measured by AUC on all positions. In Figure 6(b), in terms of
RIG, RNN achieves impressive relative gain over NN and
LR, especially on mainline positions, where RNN beats NN
by 3.12%. According to our statistics on daily traffic data,
most of revenue comes from the mainline ad clicks, where
RNN can achieve significantly better performance. While
for sidebar positions, the ads shown there are easily to be
ignored by users, so that the clicks or positive instances are
very rare. This may drastically hurt the performance of LR
model. Nevertheless, the RNN model still performs the best,
which indicates that even in rare cases, sequential information can still help.
Effect of Recurrent State for Inference We have shown
the inference method of RNN in Figure 5. To further verify the importance of utilizing historical sequences in the
inference phase, we remove the recurrent part of the RNN
model after training, and feedforward the testing samples as
a normal NN, which means that we just ignore the sequential dependencies in the testing phase. Finally, the AUC is
88.25% and RIG is 18.95%. Compared to Table 2, we can
observe a severe drop of the performance of RNN model,
which shows that the sequential information is indeed embedded in the recurrent structure and significantly contribute
to the prediction accuracy.
91.00%
30.00%
90.00%
28.00%
RIG
AUC
AUC
78%
76%
74%
72%
70%
68%
66%
64%
62%
60%
58%
56%
89.00%
24.00%
88.00%
Top First
LR
Mainline
NN
RNN
Sidebar
(a) AUC on specific ad positions
12%
87.00%
26.00%
100 200 300 400 500
LR
Sequence Length Threshold
22.00%
NN
100 200 300 400 500
RNN
Sequence Length Threshold
Figure 7: Performance with different history length T .
10%
RIG
8%
6%
4%
2%
0%
Top First
LR
Mainline
NN
RNN
Sidebar
(b) RIG on specific ad positions
Figure 6: Performance on specific ad positions.
Performance with Long vs. Short History In this part,
we conduct experiments to check the model performance
with different length of history. We first collect all available
user sequences whose length is larger than a threshold T .
In these sequences, the first T samples in each sequences
are fed into model to serve as the “accumulation period”.
Then, we continue feeding and testing samples for the rest
part of each sequence, and calculate the AUC and RIG on
all those rest parts. In such setting, the user sequences which
are selected with a larger threshold T have longer history
to feed as “accumulation period”. By doing so, we aim to
verify whether our RNN model can maintain more robust
sequential information in longer sequences.
Figure 7 shows the results. Just as expected, it turns out
that our RNN model performs the best in all settings. Moreover, when the “accumulation period” gets longer, our RNN
model tends to achieve even more relative gain compared to
baseline models. This result validates the capability of RNN
to capture and accumulate sequential information, especially
for long sequences, and help further improve the accuracy of
click prediction.
Effect of RNN Unfolding Step As described in the framework section, the unfolding structure plays an import role in
RNN training process with BPTT algorithm. Since unfolding can directly incorporate several previous samples, and
the depth of such explicit sequence modeling is determined
by the steps of unfolding, it is necessary to delve into the
effect of various RNN unfolding steps. This further analysis can help us better understand the property of the RNN
model.
According to our experimental results, the prediction ac-
curacy surges along with the increasing unfolding steps at
the beginning. The best AUC and RIG can be achieved when
unfolding 3 steps, after which the performance drops. By
checking the error terms during the process of BPTT, we
discover that the backpropagated error vanishes after 3 steps
of unfolding, which explains why larger unfolding step is
detrimental.
With these observations, intuitively, our RNN based
framework can model sequential dependency in two ways:
short-span dependency by explicitly learning from current
input and several of its leading inputs through unfolding;
long-span dependency by implicitly learning from all previous input, accumulated or embedded in the weights of the recurrent part. Meanwhile, in sponsored search, user’s behavior is also affected by both the very recent events as explicit
factor and the long-run history as implicit (background) factor. This reveals the intrinsic reason why RNN works so well
for the sequential click prediction.
Conclusion and Future Work
In this paper, we propose a novel framework for click prediction based on Recurrent Neural Networks. Different from
traditional click prediction models, our method leverages the
temporal dependency in user’s behavior sequence through
the recurrent structure. A series of experiments show that our
method outperforms state-of-the-art click prediction models
in various settings. In this future, we will continue this direction in several aspects: 1) The sequence is currently built on
user level. We will study different kinds of sequence building methods, e.g. by (user, ad) pair, (user, query) pair, advertiser, or even merge all users on the level of whole system. 2) We are going to deduce the meaning of dependency
learnt by RNN via deep understanding of RNN structure.
This may help up better utilize the property of the recurrent part. 3) Recently, some research work (Hermans and
Schrauwen 2013) has been done on Deep Recurrent Neural Networks (DRNN) and shows good results. We plan to
study whether “deep” structure can also help in click prediction, together with the “recurrent” structure.
References
[Auli et al. 2013] Auli, M.; Galley, M.; Quirk, C.; and
Zweig, G. 2013. Joint language and translation modeling
with recurrent neural networks. EMNLP.
[Bengio et al. 2006] Bengio, Y.; Schwenk, H.; Senécal, J.-S.;
Morin, F.; and Gauvain, J.-L. 2006. Neural probabilistic language models. In Innovations in Machine Learning.
Springer. 137–186.
[Box, Jenkins, and Reinsel 2013] Box, G. E.; Jenkins, G. M.;
and Reinsel, G. C. 2013. Time series analysis: forecasting
and control. Wiley. com.
[Fain and Pedersen 2006] Fain, D. C., and Pedersen, J. O.
2006. Sponsored search: A brief history. Bulletin of the
American Society for Information Science and Technology
32(2):12–13.
[Graepel et al. 2010] Graepel, T.; Candela, J. Q.; Borchert,
T.; and Herbrich, R. 2010. Web-scale bayesian clickthrough rate prediction for sponsored search advertising in
microsoft’s bing search engine. In Proceedings of the 27th
International Conference on Machine Learning (ICML-10),
13–20.
[Graves et al. 2009] Graves, A.; Liwicki, M.; Fernández, S.;
Bertolami, R.; Bunke, H.; and Schmidhuber, J. 2009. A
novel connectionist system for unconstrained handwriting
recognition. Pattern Analysis and Machine Intelligence,
IEEE Transactions on 31(5):855–868.
[Hermans and Schrauwen 2013] Hermans,
M.,
and
Schrauwen, B.
2013.
Training and analysing deep
recurrent neural networks. In NIPS, 190–198.
[Hillard et al. 2011] Hillard, D.; Manavoglu, E.; Raghavan,
H.; Leggetter, C.; Cantú-Paz, E.; and Iyer, R. 2011. The
sum of its parts: reducing sparsity in click estimation with
query segments. Information Retrieval 14(3):315–336.
[Jansen and Mullen 2008] Jansen, B. J., and Mullen, T. 2008.
Sponsored search: An overview of the concept, history, and
technology. International Journal of Electronic Business
6(2):114–131.
[Kirchgassner, Wolters, and Hassler 2012] Kirchgassner, G.;
Wolters, J.; and Hassler, U. 2012. Introduction to modern
time series analysis. Springer.
[Kombrink et al. 2011] Kombrink, S.; Mikolov, T.; Karafiát,
M.; and Burget, L. 2011. Recurrent neural network based
language modeling in meeting recognition. In INTERSPEECH, 2877–2880.
[McMahan et al. 2013] McMahan, H. B.; Holt, G.; Sculley,
D.; Young, M.; Ebner, D.; Grady, J.; Nie, L.; Phillips, T.;
Davydov, E.; Golovin, D.; Chikkerur, S.; Liu, D.; Wattenberg, M.; Hrafnkelsson, A. M.; Boulos, T.; and Kubica, J.
2013. Ad click prediction: a view from the trenches. In
Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD).
[Mikolov et al. 2010] Mikolov, T.; Karafiát, M.; Burget, L.;
Cernockỳ, J.; and Khudanpur, S. 2010. Recurrent neural
network based language model. In INTERSPEECH, 1045–
1048.
[Mikolov et al. 2011a] Mikolov, T.; Kombrink, S.; Burget,
L.; Cernocky, J.; and Khudanpur, S. 2011a. Extensions
of recurrent neural network language model. In Acoustics,
Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on, 5528–5531. IEEE.
[Mikolov et al. 2011b] Mikolov, T.; Kombrink, S.; Deoras,
A.; Burget, L.; and Černockỳ, J. 2011b. Rnnlm-recurrent
neural network language modeling toolkit. In Proc. IEEE
workshop on Automatic Speech Recognition and Understanding, 16.
[Mikolov 2012] Mikolov, T. 2012. Statistical Language
Models Based on Neural Networks. Ph.D. Dissertation, Ph.
D. thesis, Brno University of Technology.
[Radlinski et al. 2008] Radlinski, F.; Broder, A.; Ciccolo, P.;
Gabrilovich, E.; Josifovski, V.; and Riedel, L. 2008. Optimizing relevance and revenue in ad search: a query substitution approach. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development
in information retrieval, 403–410. ACM.
[Richardson, Dominowska, and Ragno 2007] Richardson,
M.; Dominowska, E.; and Ragno, R. 2007. Predicting
clicks: estimating the click-through rate for new ads. In
Proceedings of the 16th international conference on World
Wide Web, 521–530. ACM.
[Rumelhart, Hinton, and Williams 2002] Rumelhart, D. E.;
Hinton, G. E.; and Williams, R. J. 2002. Learning representations by back-propagating errors. Cognitive modeling
1:213.
[Shaparenko, Çetin, and Iyer 2009] Shaparenko, B.; Çetin,
Ö.; and Iyer, R. 2009. Data-driven text features for sponsored search click prediction. In Proceedings of the Third
International Workshop on Data Mining and Audience Intelligence for Advertising, 46–54. ACM.
[Wang et al. 2013] Wang, T.; Bian, J.; Liu, S.; Zhang, Y.; and
Liu, T.-Y. 2013. Exploring consumer psychology for click
prediction in sponsored search. In Proceedings of the 19th
ACM SIGKDD international conference on Knowledge discovery and data mining. ACM.
[Xiong et al. 2012] Xiong, C.; Wang, T.; Ding, W.; Shen, Y.;
and Liu, T.-Y. 2012. Relational click prediction for sponsored search. In Proceedings of the fifth ACM international
conference on Web search and data mining, 493–502. ACM.
[Xu, Manavoglu, and Cantu-Paz 2010] Xu, W.; Manavoglu,
E.; and Cantu-Paz, E. 2010. Temporal click model for sponsored search. In Proceedings of the 33rd international ACM
SIGIR conference on Research and development in information retrieval, 106–113. ACM.
| 9cs.NE
|
IMAG
arXiv:cs/0412012v1 [cs.PL] 3 Dec 2004
Institut d’Informatique et de
Mathématiques Appliquées
de Grenoble
LSR
Laboratoire Logiciels, Systèmes, Réseaux
RAPPORT DE RECHERCHE
Jartege: a Tool for Random Generation of
Unit Tests for Java Classes
Catherine Oriat
RR 1069
Juin 2004
B.P. 72 – 38402 SAINT MARTIN D’HERES CEDEX – France
Centre National de la Recherche Scientifique
Institut National Polytechnique de Grenoble
Université Joseph Fourier Grenoble I
Jartege: a Tool for Random Generation of Unit Tests for
Java Classes
Catherine Oriat
LSR-IMAG, Grenoble
email: [email protected]
Résumé
Ce rapport présente Jartege, un outil qui permet la génération aléatoire de tests unitaires
pour des classes Java spécifiées en JML. JML (Java Modeling Language) est un langage
de spécification pour Java qui permet d’écrire des invariants pour des classes, ainsi que
des pré- et des post-conditions pour des opérations. Comme dans l’outil JML-JUnit, nous
utilisons les spécifications JML d’une part pour éliminer des cas de test non pertinents,
et d’autre part comme oracle de test. Jartege génère de façon aléatoire des cas de test,
qui consistent en une séquence d’appels de constructeurs et de méthodes des classes sous
test. L’aspect aléatoire de l’outil peut être paramétré en associant des poids aux classes
et aux opérations, et en contrôlant le nombre d’instances créées pour chaque classe sous
test. L’utilisation pratique de Jartege est illustrée par une petite étude de cas.
Mots-clés
Test, test unitaire, génération aléatoire de cas de test, Java, JML
Abstract
This report presents Jartege, a tool which allows random generation of unit tests for Java
classes specified in JML. JML (Java Modeling Language) is a specification language for
Java which allows one to write invariants for classes, and pre- and postconditions for
operations. As in the JML-JUnit tool, we use JML specifications on the one hand to
eliminate irrelevant test cases, and on the other hand as a test oracle. Jartege randomly
generates test cases, which consist of a sequence of constructor and method calls for the
classes under test. The random aspect of the tool can be parameterized by associating
weights to classes and operations, and by controlling the number of instances which are
created for each class under test. The practical use of Jartege is illustrated by a small case
study.
Keywords
Testing, unit testing, random generation of test cases, Java, JML
Jartege: a Tool for Random Generation
of Unit Tests for Java Classes
Catherine Oriat
LSR-IMAG, Grenoble
email: [email protected]
1
Introduction
Main validation technique in software engineering, program testing aims at ensuring that
the program is correct, i.e. conforms to its specifications. As the input domain of a
program is usually very large or infinite, exhaustive testing, which consists in testing the
program for all its possible inputs, is in general impossible. The objective of testing is
thus rather to improve the software quality by finding faults in it. Testing is an important
activity of software development, whose cost is usually estimated to about 40% of the total
cost of software development, exceeding the cost of code writing.
A test campaign for a program requires several steps: design and development of test
sets, execution and results examination (or oracle). Considering the cost of testing, it is
interesting to automate some of these steps.
For Java programs, the JUnit framework [JUn, BG98] allows the developer to write
an oracle for each test case, and to automatically execute test sets. JUnit in particular
permits to automatically regression test several test sets.
If a formal specification is available, it can be translated into assertions which can be
checked at runtime, and thus serve as a test oracle. For instance, the DAISTS system
[GMH81] compiles algebraic axioms of an abstract data type into consistency checks;
Rosenblum’s APP pre-processor allows the programmer to write assertions for C programs
[Ros92]; the Eiffel design and contract approach integrates assertions in the programming
language [Mey88, Mey92].
It is also interesting to automate the development of tests. We can distinguish between
two groups of strategies to produce test sets: random and systematic strategies. Systematic strategies, such as functional testing or structural testing, consist in decomposing the
input domain of the program in several subdomains, often called “partitions”.
Many systematic strategies propose to derive test cases from a formal specification
[Gau95]. For instance, the Dick and Faivre method [DF93], which consists in constructing
a finite state automaton from the formal specification and in selecting test cases as paths
in this automaton, has been used as a basis by other approaches, in particular Casting
[ABM97] or BZTT [LPU02]. BZTT uses B or Z specifications to generate test cases which
consist in placing the system in a boundary state and calling an operation with a boundary
value.
In contrast to systematic methods, random testing generally does not use the program
nor the specification to produce test sets. It may use an operational profile of the program,
which describe how the program is expected to be used. The utility of random testing is
controversial in the testing community. It is usually presented as the poorest approach for
selecting test data [Mye94]. However, random testing has a few advantages which make
us think that it could be a good complement to systematic testing:
• random testing is cheap and rather easy to implement. In particular, it can produce
large or very large test sets;
3
• it can detect a substantial number of errors at a low cost [DN84, HT90, Nta01].
Moreover, if an operational profile of the program is available,
• random testing allows early detection of the failures that are most likely to appear
when using the program [FHLS98];
• it can be used to evaluate the program reliability [HT90, Ham94].
However, partition testing can be much more effective at finding failures, especially when
the strategy defines some very small subdomains with a high probability to cause failures
[WJ91].
Among the best practices of extreme programming, or XP [Bec99, Bec00], are continuous testing and code refactoring. While they are writing code, developers should write
corresponding unit tests, using a testing framework such as JUnit. Unit testing is often
presented as a support to refactoring: it gives the developer confidence that the changes
have not introduced new errors. However, code refactoring often requires to change some
of the corresponding unit tests as well. If a class has to be tested intensively, the amount
of code corresponding to the tests often exceeds the amount of code of the class. Both
practices can therefore be hard to conciliate.
This report proposes to use random generation of tests to facilitate countinuous testing
and code refactoring in the context of extreme programming. We presents Jartege, a tool
for random generation of unit tests for Java classes specified in JML, which aims at easily
producing numerous test cases, in order to detect a substantial number of errors at a low
cost.
The rest of the report is organized as follows. Section 2 introduces the approach.
Section 3 presents a case study which consists in modeling bank accounts. This case study
will serve to illustrate the use of our testing tool. Section 4 presents our tool Jartege
and how it can be used to test the bank account case study. Section 5 introduces more
advanced features of Jartege, which allow one to parameterize its random aspect. Section
6 shows the errors which are detected by test cases generated by Jartege in the case study.
Section 7 presents and compares related approaches. Section 8 discusses some points about
random generation of tests and draw future work we intend to undertake around Jartege.
2
Approach
JML (Java Modeling Language) is a specification language for Java inspired by Eiffel,
VDM and Larch, which was designed by Gary Leavens and his colleagues [JML, LBR03,
LPC+ 03]. Several teams are currently still working on JML design and tools around JML
[BCC+ 03]. JML allows one to specify various assertions in particular invariants for classes
as well as pre- and postconditions for methods.
The JML compiler (jmlc) [CL02a] translates JML specifications into assertions checked
at runtime. If an assertion is violated, then a specific exception is raised. In the context
of a given method call, the JML compiler makes a useful difference between an entry
precondition which is a precondition of the given method, and an internal precondition,
which is a precondition of an operation being called, at some level, by the given method.
The JML-JUnit tool [CL02b] generates JUnit test cases for a Java program specified
in JML, using the JML compiler to translate JML specifications into test oracles. Test
cases are produced from a test fixture and parameter values supplied by the user.
4
Our approach is inspired by the JML-JUnit tool: we propose to generate random tests
for Java programs specified in JML, using this specification as a test oracle, in the JMLJUnit way. Our aim is to produce a big number of tests at a low cost, in order to facilitate
unit testing.
To implement these ideas, we started developing a prototype tool, called Jartege for
Java Random Test Generator. Jartege is designed to generate unit tests for Java classes
specified with JML. By unit tests, we here mean tests for some operations in a single class
or a small cluster of classes. In our context, a test case is a Java method which consists of
a sequence of operation (constructor or method) calls.
As in the JML-JUnit tool, we use the JML specification to assist test generation in
two ways:
1. It permits the rejection of test sequences which contain an operation call which
violates the operation entry precondition. We consider that these test sequences are
not interesting because they detect errors which correspond to a fault in the test
program.
(Although such sequences could be used to detect cases when a precondition is too
strong, this goal would prevent us from producing long sequences of calls. Thus,
we choose to trust the specification rather than the code. We can note that if an
operation uses a method whose precondition is too strong, this will produce an
internal precondition error and the sequence will not be rejected.)
2. The specification is also used as a test oracle: the test detects an error when another
assertion (e.g. an invariant, a postcondition or an internal precondition) is violated.
Such an error corresponds to a fault in the Java program or in the JML specification.
Our work has been influenced by the Lutess tool [Par96, dBORZ99], which aims at deriving test data for synchronous programs, with various generation methods, in particular
a purely random generation and a generation guided by operational profiles. The good
results obtained by Lutess, which won the best tool award of the first feature interaction
detection contest [dBZ99], have encouraged us to consider random generation of tests as
a viable approach.
3
Case Study
In this section, we present a case study to illustrate the use of Jartege. This case study
defines bank accounts and some operations on these accounts.
3.1
Informal Specification of the Bank Accounts
We describe here an informal specification of bank accounts:
1. An account contains a certain available amount of money (its balance), and is associated with a minimum amount that this account may contain (the minimum
balance).
2. It is possible to credit or debit an account. A debit operation is only possible if there
is enough money on the account.
3. One or several last credit or debit operations may be cancelled.
5
4. The minimum balance of an account may be changed.
3.2
Bank Account Modeling
To represent accounts, we define a class Account with two attributes: balance and min
which respectively represent the balance and the minimum balance of the account, and
three methods: credit, debit and cancel. In order to implement the cancel operation, we
associate with each account an history, which is a linked list of the previous balances of
the account. The class History has one attribute, balance, which represents the balance
of its associated account before the last credit or debit operation. With each history is
associated its preceding history. Figure 1 shows the UML class diagram of bank accounts.
hist
Account
History
0..1
balance
min
balance
0..1
credit(amount: int)
debit(amount: int)
cancel()
prec
Figure 1: UML diagram of the case study
3.3
JML Specification and Java Implementation
We implement both class Account and History in Java, and specify them with JML. We
choose to write lightweight public specifications to emphasize that these specifications are
destined for clients and need not be complete.
In order to forbid uncontrolled modifications of the attributes, we declare them as
private and define associated access methods: getBalance, getMin and getHist in class
Account and getBalance and getPrec in class History. These methods are specified as pure
methods in JML because they are side effect free, which allows us to use them in JML
assertions.
The class Account has an invariant which specifies that the balance of an account must
always be greater than the minimum balance.
/* Class of bank accounts. */
public class Account {
/* Invariant of class Account. */
/*@ public invariant getBalance( ) >= getMin( ); */
private int balance; // The balance of this account
private int min;
// The minimum balance
private History hist; // The history list of this account
/* The balance of this account. */
public /*@ pure */ int getBalance( ) { return balance; }
6
/* The history list of this account. */
public /*@ pure */ History getHist( ) { return hist; }
/* The minimum balance of this account. */
public /*@ pure */ int getMin( ) { return min; }
The constructor of class Account constructs an account with the specified balance and the
specified minimum balance. Its precondition asserts that the specified balance is greater
than the specified minimum balance.
/* Constructs an account with the specified balance and
* minimum balance. */
/*@ requires balance >= min; */
public Account (int balance, int min) {
this.balance = balance;
this.min = min;
this.hist = null ;
}
As the minimum balance min is a private attribute, we have to introduce a method to
change its value. The method setMin (int min) sets the minimum balance to the specified
value. Its precondition asserts that the balance is greater than the specified minimum
value.
/* Sets the minimum balance to the specified value. */
/*@ requires getBalance ( ) >= min; */
public void setMin (int min) { this.min = min; }
The method credit (int amount) credits the account with the specified amount. Its precondition requires the amount to be positive. Its postcondition asserts that the new balance is
the former balance augmented by the specified amount, that a new history is created with
balance the former balance of the account and with previous history the former history of
the account. Its exceptional postcondition asserts that the method should never terminate
abruptly.
/* Credits this account with the specified amount. */
/*@ requires amount >= 0;
*@ ensures getBalance ( ) == \old (getBalance ( )) + amount &&
*@
\fresh (getHist ( )) &&
*@
getHist ( ).getBalance ( ) == \old (getBalance ( )) &&
*@
getHist ( ).getPrec ( ) == \old (getHist ( ));
*@ signals (Exception e) false;
*/
public void credit(int amount) {
hist = new History (balance, getHist ( ));
balance = balance + amount;
}
The debit operation, which is very similar to the credit operation, is not detailed here.
It has the additional precondition that the balance decreased by the specified amount is
greater than the minimum balance.
7
The method cancel cancels the last credit or debit operation. Its precondition requires
that the history is not null, which means that at least one operation of credit or debit has
taken place since the account was created. Its postcondition ensures that the balance and
the history of the account have been updated with their former values.
/* Cancels the last credit or debit operation. */
/*@ requires getHist ( ) != null ;
*@ ensures getHist ( ) == \old (getHist ( ).getPrec ( )) &&
*@
getBalance ( ) == \old (getHist ( ).getBalance ( ));
*@ signals (Exception e) false;
*/
public void cancel ( ) {
balance = hist.getBalance ( );
hist = hist.getPrec ( );
}
} // End of class Account
We do not define any JML assertion for the class History.
/* Class of histories. */
public class History {
private int balance; // The balance of this history.
private History prec; // The preceding history.
/* Constructs a history with the specified balance and preceding history. */
public History (int balance, History prec) {
this.balance = balance; this.prec = prec;
}
/* The balance of this history. */
public /*@ pure */ int getBalance ( ) { return balance; }
/* The preceding history. */
public /*@ pure */ History getPrec ( ) { return prec; }
} // End of class History
4
Jartege
Jartege (Java Random Test Generator) is a framework for automatic random generation
of unit tests for Java classes specified with JML. This approach consists in producing test
programs which are composed of test cases, each test case consisting of randomly chosen
sequences of method calls for each class under test. Each generated test program can be
executed to test the classes, and re-executed later on either after having corrected some
faults or for regression test.
The tool is designed to produce unit tests, i.e. tests composed of calls of some methods
which belong to a few classes. As noticed in [LTFWD00], because of complex dependences
that exist between classes in object-oriented programs, it is usually not possible to test a
method or a class in complete isolation. Jartege thus is able to generate test cases which
allow the integration of several classes.
8
4.1
Practical Use of Jartege
Suppose we wish to generate tests for the classes Account and History. We write the
following Java program:
import jartege.*;
/** Jartege test cases generator for classes Account and History. */
class TestGen {
public static void main (String[ ] args) {
// Creates a class tester
ClassTester t = new ClassTester ( );
// Adds the specified classes to the set of classes under test
t.addClass ("Account");
t.addClass ("History");
// Generates a test class TestBank, made of 100 test cases.
// For each test case, the tool tries to generate 50 method calls.
t.generate ("TestBank", 100, 50);
}
}
The main class of the Jartege framework is ClassTester. This class must be instantiated
to allow the creation of test programs.
The method addClass (String className) adds the class className to the set of classes
under test. In this example, we wish to generate tests for the classes Account and History.
The method generate (String className, int numberOfTests, int numberOfMethodCalls)
generates a file className.java which contains a class called className. This class is composed of numberOfTests test cases. For each test case, the tool makes numberOfMethodCalls attempts to generate a method call of one of the classes under test. Using the
accessible constructors, the tool constructs objects which serve as parameters for these
method calls.
When the program TestGen is executed, it produces a file TestGen.java which contains
a main program. This main program calls successively 100 test methods test1 ( ), test2 ( )
... test100 ( ). Each test method contains about 50 method calls.
While the program is generated, Jartege executes on the fly each operation call, which
allows it to eliminate calls which violate the operation precondition. When this precondition is strong, it may happen that the tool does not succeed in generating a call for a
given method, which explains that about 50 method calls are generated.
4.2
Test Programs Produced by Jartege
A test program produced by Jartege is a class with a main method which consists in calling
sequentially all generated test cases. Each test case consists of a sequence of constructor
and method calls of the classes under test. Here is an example of such a test case:
// Test case number 1
public void test1 ( ) throws Exception {
try {
Account ob1 = new Account (1023296578, 223978640);
ob1.debit (152022897);
9
History ob2 = new History (1661966075, (History) null);
History ob3 = new History (-350589348, ob2);
History ob4 = ob2.getPrec ( );
int ob5 = ob3.getBalance ( );
ob1.cancel ( );
// ...
} catch (Throwable except) {
error ( except, 1);
}
}
For each test method, if a JML exception is raised, then an error message (coming from
jmlc) is printed. The test program terminates by printing an assessment of the test. As an
example, here is an excerpt of what is printed with a generated program TestBank.java:
1) Error detected in class TestBank by method test2:
org.jmlspecs.jmlrac.runtime.JMLInvariantError:
By method ”[email protected]:79:18i” of class ”Account”
for assertions specified at Account.java:11:32 [...]
at TestBank.test2(TestBank.java:138)
[...]
Number of tests: 100
Number of errors: 71
Number of inconclusive tests: 0
The program has detected 71 errors. The first error detected comes from a violation of
the invariant of class Account (specified line 11), which happened after a credit operation.
The test program also indicates the number of inconclusive tests. A test case is inconclusive when it does not allow one to conclude whether the program behaviour is correct
or not. A test program generated by Jartege indicates that a test case is inconclusive
when it contains an operation call whose entry precondition is violated. As Jartege is
designed to eliminate such operation calls, this situation may only arise when the code or
the specification of one of the classes under test has been modified after the test file was
generated. A high number of inconclusive tests indicates that the test file is no longer
relevant.
5
Controlling Random Generation
If we leave everything to chance, Jartege might not produce interesting sequences of calls.
Jartege thus provides a few possibilities to parameterize its random aspect. These features
can be useful for stress testing, for instance if we want to test more intensively a given
method. More generally, they allow us to define an operational profile for the classes under
test, which describe how these classes are likely to be used by other components.
5.1
Weights
With each class and operation of a class is associated a weight, which defines the probability
that a class will be chosen, and that an operation of this class will be called. In particular,
it is possible to forbid to call some operation by associating a null weight with it. By
10
default, all weights are equal to 1. A weight can be modified by a weight change method.
In particular:
• changeAllMethodsWeight (String className, double weight) changes the weight of all
methods in the class className to the specified weight.
• changeMethodWeight (String className, String methodName, double weight) changes
the weight of the specified method(s) in the class className to the specified weight.
• changeMethodWeight (String className, String methodName, String [ ] signature, double weight) changes the weight of the method by its name and its signature to the
specified weight.
5.2
Creation of Objects
Objects creation is commanded by creation probability functions, which define the probability of creating a new object according to the number of existing objects of the class
against that of reusing an already created object. If this probability is low, Jartege is more
likely to reuse an already created object than to construct a new one. This allows the user
either to create a predefined number of instances for a given class, or on the opposite, to
create numerous instances for a class.
In the example of bank accounts, it is not very interesting to create many accounts.
It is possible to test the class Account more efficiently for example by creating a unique
account and by applying numerous method calls to it.
The function changeCreationProbability(String className, CreationProbability creationProbabilityFunction) changes the creation probability function associated with the specified
class to the the specified creation probability function.
The interface CreationProbability contains a unique method
double theFunction(int nbCreatedObjects)
which must satisfy the condition
theFunction(0) = 1;
theFunction(n) ∈ [0, 1], ∀n ≥ 1.
The class ThresholdProbability allows one to define threshold probability functions whose
value is 1 under some threshold s and 0 above.
theFunction(n) = 1, if n < s;
theFunction(n) = 0, otherwise.
A threshold probability function with threshold s allows one to define at most s instances
of a given class. We can for instance forbid the creation of more than one instance of
Account by adding the following statement in the test generator:
t.changeCreationProbability ("Account", new ThresholdProbability (1));
11
5.3
Parameter Generation of Primitive Types
When a method has a strong precondition, the probability that Jartege, without any further indication, will generate a call to this method which does not violate this precondition
is low. For primitive types, Jartege provides the possibility to define generators for some
parameters of a given method.
For example, the precondition of the debit operation requires the parameter amount to
be in range
[0, getBalance ( ) − getMin ( )].
Let us suppose the range is small, in other words that the balance is closed to the minimum
balance. If Jartege chooses an amount to be debited entirely randomly, this amount is not
likely to satisfy the method precondition.
Jartege provides a way of generating parameter values of primitive types for operations.
For this, we define a class JRT Account as follows.
import jartege.RandomValue;
public class JRT Account {
private Account theAccount; // The current account
/* Constructor. */
public JRT Account (Account theAccount) { this.theAccount = theAccount; }
/** Generator for the first parameter of operation debit (int). */
public int JRT debit int 1 ( ) {
return RandomValue.intValue (0,
theAccount.getBalance ( ) - theAccount.getMin ( ));
}
}
The class JRT Account must contain a private field of type Account which will contain
the current object on which an operation of class Account is applied. A constructor allows
Jartege to initialize this private field. The class also contains one parameter generation
method for each parameter for which we specify the generation of values. In the example,
to specify the generation of the first parameter of operation debit (int amount), we define
the method int JRT debit int 1 ( ). We use the signature of the operation in the name of
the method to allow overloading. The method RandomValue.intValue (int min, int max)
chooses a random integer in range [min, max].
5.4
Fixtures
If we want to generate several test cases which operate on a particular set of objects, we
can write a test fixture, in a similar way to JUnit. A test fixture is a class which contains:
• attributes corresponding to the objects a test operates on;
• an optional setUp method which defines the preamble of a test case (which typically
constructs these objects);
• an optional tearDown method which defines the postamble of a test case.
12
6
Applying Jartege to the Case Study
The 100 test cases generated by Jartege, showing 71 failures, revealed three different errors:
one error caused by a credit operation, and two errors caused by a cancel operation. We
extracted the shorter sequence of calls which resulted in each failure and obtained the
following results. We also changed the parameter values and added some comments for
more readability.
Error 1. The credit operation can produce a balance inferior to the previous balance
because of an integer overflow.
public void test1 ( ) {
Account ob1 = new Account (250000000, 0);
ob1.credit (2000000000); // Produces a negative balance,
}
// below the minimum balance.
Error 2. The cancel operation can produce an incorrect result if it is preceded by a
setMin operation which changes the minimum balance of the account to a value which is
superior to the balance before cancellation.
public void test11 ( ) {
Account ob1 = new Account (-50, -100);
ob1.credit (100);
ob1.setMin (0);
ob1.cancel ( ); // Restores the balance to a value
}
// inferior to the minimum balance.
Error 3. The third error detected is a combination of an overflow on a debit operation
(similar to Error 1, which comes from an overflow on a credit operation) and of the second
error.
public void test13 ( ) {
Account ob1 = new Account (-1500000000, -2000000000);
ob1.debit (800000000); // Produces a positive balance.
ob1.setMin (0);
ob1.cancel ( ); // Restores the balance to a value
}
// inferior to the minimum balance.
We have the feeling that the three errors detected with test cases generated by Jartege
are not totally obvious, and could have easily been forgotten in a manually developed test
suite. Errors 2 and 3 in particular require three method calls to be executed in a specific
order and with particular parameter values.
It must be noted that the case study was originally written to show the use of JML to
undergraduate students, without us being aware of the faults.
7
Comparison with Related Work
Our work has been widely inspired by the JML-JUnit approach [CL02b]. The JML-JUnit
tool generates test cases for a method which consist of a combination of calls of this method
13
with various parameter values. The tester must supply the object invoking the method
and the parameter values. With this approach, interesting values could easily be forgotten
by the tester. Moreover, as a test case only consists of one method call, it is not possible
to detect errors which result of several calls of different methods. At last, the JML-JUnit
approach compels the user to construct the test data, which may require the call of several
constructors. Our approach thus has the advantage of being more automatic, and of being
able to detect more potential errors.
Korat [BKM02] is a tool also based on the JML-JUnit approach, which allows exhaustive testing of a method for all objects of a bounded size. The tools automatically
construct all non isomorphic test cases and execute the method on each test case. Korat
therefore has the advantage over JML-JUnit of being able to construct the objects which
invoke the method under test. However, test cases constructed by Korat only consist of
one object construction and one method invocation on this object.
Tobias [Led02, MLBdB02] is a combinatorial testing tool for automatic generation of
test cases derived from a “test pattern”, which abstractly describes a test case. Tobias was
first designed to produce test objectives for the TGV tool [JM99] and was then adapted
to produce test cases for Java programs specified in JML and for programs specified in
VDM. The main problem of Tobias is the combinatorial explosion which happens if one
tries to generate test cases which consist of more than a couple of method calls. Jartege
was designed to allow the generation of long test sequences without facing the problem of
combinatorial explosion.
8
Discussion and Future Work
Jartege is only in its infancy and a lot of work remains to be done.
Primitive values generation for methods parameters is currently done manually by
writing primitive parameters generating methods. Code for these methods could be automatically constructed from the JML precondition of the method. This could consist in
extracting range constraints from the method precondition and automatically produce a
method which could generate meaningful values for the primitive parameters.
Jartege easily constructs test cases which consist of hundreds of constructors and methods calls. It would be useful to develop a tool for extracting a minimum sequence of calls
which results in a given failure.
We developed Jartege in Java, and we specified some of its classes with JML. We
applied Jartege to these classes to produce test cases, which allowed us to experiment
our tool on a larger case study and to detect a few errors. We found much easier to
produce tests with Jartege than to write unit tests with JUnit or JML-JUnit. We intend
to continue our work of specifying Jartege in JML and testing its classes with itself. We
hope that this real case study will help us to evaluate the effectiveness and scalability of
the approach.
A comparison of our work with other testing strategies still remains to be done. We can
expect systematic methods, using for instance boundary testing such as BZTT [LPU02],
to be able to produce more interesting test cases that ours. Our goal is certainly not to
pretend that tests produced randomly can replace tests produced by more sophisticated
methods, nor a carefully designed test set written by an experienced tester.
Our first goal in developing Jartege was to help the developer to write unit tests for
unstable Java classes, thus for “debug unit testing”. It would also be interesting to use
Jartege to evaluate the reliability of a stable component before it is released. Jartege
14
provides some features to define an operational profile of a component, which should allow
statistical testing. However, the definition of a correct operational profile, especially in
the context of object-oriented programming, is a difficult task. Moreover, the relation
between test sets generated by Jartege and the reliability of a component requires more
theoretical work, one difficult point being to take into account the state of the component.
9
Conclusion
This report presents Jartege, a tool for random generation of unit tests for Java classes
specified in JML. The aim of the tool is to easily produce numerous test cases, in order to
detect a substantial number of errors without too much effort. It is designed to produce
automated tests, which can in part replace tests written by the developer using for instance JUnit. We think that the automatic generation of such unit tests should facilitate
continuous testing as well as code refactoring in the context of extreme programming.
The JML specifications are used on the one hand to eliminate irrelevant test cases,
and on the other hand as a test oracle. We think that the additional cost of specification
writing should be compensated by the automatic oracle provided by the JML compiler, as
long as we wish to intensively test the classes. Moreover, this approach has the advantage
of supporting the debugging of a specification along with the corresponding program.
This allows the developer to increase his confidence in the specification and to use this
specification in other tools.
Most test generation methods are deterministic, while our approach is statistical. We
do not wish to oppose both approaches, thinking that they both have their advantages
and drawbacks, and that a combination of both could be fruitful.
At last, we found JML to be good language to start learning formal methods. Its
Java-based syntax makes it easy to learn for Java programmers. As JML specifications
are included in Java source code as comments, it is easy to develop and debug a Java
program along with its specification. Moreover, automatic test oracles as well as automatic
generation of test cases are good reasons of using specification languages such as JML.
References
[ABM97]
Lionel Van Aertryck, Marc Benveniste, and Daniel Le Métayer. Casting: a formally based software test generation method. In Proceedings of
the First IEEE Internatinal Conference on Formal Engineering Methods —
ICFEM’97, Hiroshima, Japan, pages 101–111, November 1997.
[BCC+ 03]
Lilian Burdy, Yoonsik Cheon, David R. Cok, Michael Ernst, Joe Kiniry,
Gary T. Leavens, K. Rustan, M. Leino, and Erik Poll. An overview of JML
tools and applications. Technical Report NIII-R0309, Department of Computer Science, University of Nijmegen, March 2003.
[Bec99]
Kent Beck. Embracing change with extreme programming. IEEE Computer,
32(10):70–77, October 1999.
[Bec00]
Kent Beck. Extreme Programming Explained. Addison Wesley, 2000.
[BG98]
Kent Beck and Erich Gamma. Test infected: Programmers love writing tests.
Java Reports, 3(7):51–56, 1998.
15
[BKM02]
Chandrasekhar Boyapati, Safraz Khurshid, and Darko Marinov. Korat: Automated testing based on Java predicates. In Proceedings of the International
Symposium on Software Testing and Analysis — ISSTA’02, Rome, pages
123–133, July 2002.
[CL02a]
Yoonsik Cheon and Gary T. Leavens. A runtime assertion checker for the Java
Modeling Language (JML). In Hamid R. Arabnia and Youngsong Mun (eds.),
International Conference on Software Engineering Research and Practice —
SERP’02, Las Vegas, Nevada, pages 322–328. CSREA Press, 2002.
[CL02b]
Yoonsik Cheon and Gary T. Leavens. A simple and practical approach to unit
testing: The JML and JUnit way. In Boris Magnusson (ed.), 16th European
Conference on Object-Oriented Programming — ECOOP’02, Malaga, Spain,
number 2374 in Lecture Notes in Computer Science, pages 231–255. SpringerVerlag, June 2002.
[dBORZ99] Lydie du Bousquet, Farid Ouabdesselam, Jean-Luc Richier, and Nicolas
Zuanon. Lutess: a specification-driven testing environment for synchronous
software. In International Conference on Software Engineering — ICSE’99,
Los Angeles, USA. ACM Press, May 1999.
[dBZ99]
Lydie du Bousquet and Nicolas Zuanon. An overview of Lutess: a
specification-based tool for testing synchronous software. In 14th IEEE International Conference on Automated Software Engineering — ASE’99, pages
208–215, October 1999.
[DF93]
Jeremy Dick and Alain Faivre. Automating the generation and sequencing
of test cases from model-based specifications. In Proceedings of FME’93,
number 670 in LNCS, pages 268–284. Springer-Verlag, April 1993.
[DN84]
Joe W. Duran and Simeon C. Ntafos. An evaluation of random testing. IEEE
Transactions on Software Engineering, 10(4):438–444, 1984.
[FHLS98]
Phyllis G. Frankl, Richard G. Hamlet, Bev LittleWood, and Lorenzo Strigini.
Evaluating testing methods by delivered reliability. IEEE Transactions on
Software Engineering, 24(8):586–601, August 1998.
[Gau95]
Marie-Claude Gaudel. Testing can be formal too. In Proceedings of TAPSOFT’95, Aarhus, Denmark, number 915 in LNCS, pages 82–96. SpringerVerlag, May 1995.
[GMH81]
John Gannon, Paul McMullin, and Richard Hamlet. Data-abstraction implementation, specification, and testing. ACM Transactions on Programming
Languages and Systems, 3(3):211–233, July 1981.
[Ham94]
Richard Hamlet. Random testing. In J. Marciniak, editor, Encyclopedia of
Software Engineering, pages 970–978. Wiley, 1994.
[HT90]
Dick Hamlet and Ross Taylor. Partition testing does not inspire confidence.
IEEE Transactions on Software Engineering, 16(12), December 1990.
16
[JM99]
Thierry Jéron and Pierre Morel. Test generation derived from modelchecking. In Proceedings of the 11th International Conference on Computer
Aided Verification — CAV’99, Trento, Italy, number 1633 in LNCS, pages
108–121. Springer-Verlag, July 1999.
[JML]
The Java Modeling Language (JML) Home Page,
http://www.cs.iastate.edu/˜leavens/JML.
[JUn]
The JUnit Home Page, http://www.junit.org.
[LBR03]
Gary T. Leavens, Albert L. Baker, and Clyde Ruby. Preliminary Design of
JML: A Behavioral Interface Specification Language for Java. Department
of Computer Science, Iowa State University, 1998–2003.
[Led02]
Yves Ledru. The TOBIAS test generator and its adaptation to some ASE
challenges (position paper). In Workshop on the State of the Art in Automated Software Engineering, ICS Technical Report UCI-ICS-02-17, University of California, Irvine, USA, 2002.
[LPC+ 03]
Gary T. Leavens, Erik Poll, Curtis Clifton, Yoonsik Cheon, and Clyde Ruby.
JML Reference Manual. (Draft), April 2003.
[LPU02]
Bruno Legeard, Fabien Peureux, and Mark Utting. Automated boundary
testing from Z and B. In Proceedings of FME’02, Formal Methods Europe, Copenhaguen, Denmark, number 2391 in LNCS, pages 21–40. SpringerVerlag, July 2002.
[LTFWD00] Yvan Labiche, Pascale Thévenod-Fosse, Hélène Waeselynck, and M.-H. Durand. Testing levels for object-oriented software. In Proceedings of the 22nd
International Conference on Software Engineering — ICSE’00, Limerick,
Ireland, pages 136–145. ACM, June 2000.
[Mey88]
Bertrand Meyer. Object-Oriented Software Construction. Prentice Hall, 1988.
[Mey92]
Bertrand Meyer. Applying design by contract. IEEE Computer, 25(10):40–
51, October 1992.
[MLBdB02] Olivier Maury, Yves Ledru, Pierre Bontron, and Lydie du Bousquet. Using
TOBIAS for the automatic generation of VDM test cases. In Third VDM
Workshop (at FME’02), Copenhaguen, Denmark, July 2002.
[Mye94]
G. J. Myers. The Art of Software Testing. John Wiley and Sons, New York,
1994.
[Nta01]
Simeon C. Ntafos. On comparisons of random, partition, and proportional
partition testing. IEEE Transactions on Software Engineering, 27(10):949–
960, July 2001.
[Par96]
Ioannis Parissis. Test de logiciels synchrones spécifiés en Lustre. PhD thesis,
Grenoble, France, Septembre 1996.
[Ros92]
David S. Rosenblum. Towards a method of programming with assertions. In
International Conference on Software Engineering — ICSE’92. IEEE Computer Society Press, 1992.
17
[WJ91]
Elaine J. Weyuker and Bingchiang Jeng. Analyzing partition testing strategies. IEEE Transactions on Software Engineering, 17(7):703–711, July 1991.
18
| 6cs.PL
|
Clustering of Complex Networks and Community Detection Using Group Search
Optimization
G. Kishore Kumar†, Dr. V. K. Jayaraman
Evolutionary Computing and Image Processing Group, Centre for Development of Advanced Computing(CDAC),
Pune University Campus, Ganeshkhind, Pune-411007, Maharashtra, India.
Email: [email protected], [email protected]
Abstract—Group Search Optimizer(GSO) is one of the
best algorithms, and is very new in the field of
Evolutionary Computing. It is very robust and efficient
algorithm, which is inspired by animal searching
behaviour. The paper describes an application of GSO
to clustering of networks. We have tested GSO against
five standard benchmark datasets and the algorithm is
proven to be very competitive in terms of accuracy and
convergence speed.
Index
Terms—Group
Search Optimizer(GSO),
Evolutionary Computing, Optimization, Network
Clustering, Candidate Solution, Modules Matrix,
Newman's Modularity, Community Detection.
I. INTRODUCTION
We are in the world of networks- our computers,
mobiles and even we ourselves are part of complex social
networks. The ability to measure the relationship between
networks and different organizational attributes will
enable us to create better working environments [1]. Once
very least preferred research area like complex network
analysis is now attracting much attention of
mathematicians, computer scientists and even physicists
[2]. Since it is at the amateur level, many aspects of the
networks are not clearly understood, one needs to explore
all the properties of the network to understand the
dynamics of these networks.
By the very nature, these networks are complex to
understand and formalise. Few empirical properties of
networks are defined in the course of time, with the help
of graph theoretical and statistical physics aspects like
Degree distribution, Clustering coefficient, Community
structure, Entropy etc [2]. Effort has to be placed in the
direction of identification and integration of very crucial,
sensitive and mutually consistent properties of the
networks. One of those identifiedand well-studied
properties is community structure. The structure detection
is closely connected to graph partitioning in theory
ofgraphs [19]. One can getbasic understanding of these
aspects referring to Newman [2-4].
Newman and Girvan have proposed a measure of
quality of a particular division of network, which is now
famous Newman's modularity [2]. A particular network
can be divided in to K communities, and this can be
represented by a K×K symmetric matrix, we call it as
modules matrix (e), eij is the fraction of all network edges
that link vertices in the group from i to j. Modularity
measure is as follows:
𝑄 = (𝑒𝑖𝑖 − 𝑎𝑖2 ) = 𝑇𝑟 𝑒 − 𝑒 2
(1)
where 𝑥 is the sum of matrix x elements. Q is the best
measure of community structure. Values of Q that are
close to 1 represent better community structure. Since
community structure identification is a graph partitioning
problem, which is in turn an optimization problem, we
need effective algorithms to tackle this problem. The
potential candidates are evolutionary algorithms [5].
In recent years, various animal behaviour-inspired
algorithms, which are known as swarm intelligence(SI)
algorithms are coming up [7]. Most popular ones are Ant
Colony Optimization [8] and Particle Swarm Optimizer
[9]. One of the most recent entries in to SI algorithms
family is GSO,proposed by He and Wu, based on
animals foraging behaviour, which is primarily designed
to handle continuous optimization problems [10].
Foraging tactics are very basic requirements, should be
possessed by animals. Search for food is very common
part of animal behaviour, which consumes most part of
their life time. Efficacy of animal searching mechanism
depends up on the group behaviour [10].
The GSO is a process of obtaining optimum solution
in a search space, which is analogous to search for better
clustering of a network to obtain best Newman's
modularity. In the rest of the paper, we will first discuss
the basics of GSO and later we get in to the details of
implementation of the algorithm for network clustering.
Even though many algorithms in literature addressed the
clustering problem, still there is a need for efficient
algorithms. Here in this paper we try toprove the
superiority of the GSO over other available algorithms
when applied to clustering problems.
II. GROUP SEARCH OPTIMIZER
Group is the name given to population in GSO, and
each individual is a member. Here we just want to give a
few details of GSO, which are required to proceed with
the analysis and application of algorithm. For more details
on this algorithm, one needs to refer Hu [10], where it is
very nicely and thoroughly explained. In GSO, the group
consists of three types of members: producers, scroungers
and rangers. Basically, there is only one producer at each
searching bout and remaining members are scroungers
and rangers (who perform random walk).
a) Producers
At each iteration, the candidate solution(Group
member) conferring the best fitness value is chosen as
producer. That member scans the search space for
optimum position. Soon the producer will find a better
position(Resources) with the best fitness value. If that
position has a better resource than the current position,
then producer will fly to this position or it will stay in
current position and search for other optimal position. If
the producer cannot find a better position after ’a’no. of
iterations, it will retain back to its original position.
b) Scroungers
The Scroungers do take the best obtained by the
producers, they always try to keep track of the best fitness
values obtained by the producers. In case, the current
producer can't find better fitness value, it will be replaced
with one of those scroungers, which is next to producer,
processing better fitness values. In case if a scrounger
findsabetter optima in the course of time, in the next bout
it will be made as a producer and all other members
including the producer in the previous searching bout,
will be kept to perform scrounging.
c) Rangers
The rest of the population comprises a group called
rangers; they perform random walks in search of better
resources. They always save the total population by not
letting them get in to local optima. The potential of the
algorithm lies in proper integration of different search
strategies: systematic and random.
III. IMPLEMENTATION DETAILS
Consider a population of 'm' members, whom we call
as candidate solutions. Each candidate solution is an ndimensional vector. Here 'n' represents number of nodes
in the network. The ith element of the vector gives the
details of cluster(module) number, where ith node is
placed. TheFig.1 given below shows the details of the
data structure used.
4
2
8
-
1
2
3
-
2
7
10
- -
1
11
5
-
3
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
c
c
-
-
n
12
2
9
3
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1
13
-
-
-
-
-
-
-
3
m
Fig. 1.Each member in a group is an n-dimensional vector, and
each dimension possesses a cluster no. There are c no. of
clusters in the complete network and m no of candidate
solutions(members) in total population.
We randomly generate 'm' no of candidate solutions
and each candidate solution in turn produces a Modules
Matrix(MM), which gives the sum of all node
connections(edges) between every two clusters. Using
MM, Newman's modularity for each candidate solution
is calculated. The candidate solution holding best fitness
value(i.e., Newman's modularity value) is chosen as
producer and remaining all candidate solutions are
selected as scroungers and rangers.
Perform GSO for 'j' no of iterations. Basically,
each candidate solution(member) corresponds to a
different way of clustering the nodes.During each
iteration, the members of the group try to attain new
positions in accordance with GSO. After each iteration,
fitness value of each candidate solution on its new
positionis calculated. The current producer is replaced
with the candidate solution which attains best fitness
value. The best fitness value and corresponding
candidate solution is recorded for further use. The Fig. 2
given below shows the flow chart for implementation of
GSO.
IV. EVALUATION
The code for the algorithm has been written in GNU
Octave and tested on five social network datasets:
Zachary karate club, Dolphins network, Jazz Musicians
network, American football network and Les Miserables
novel characters network.
Algorithms
Modularity
Newman
DA
GN
EO
GSO
0.381
0.419
0.401
0.424
0.613
Table 1: Results of GSO on karate club data
Start
Generate Adjacency matrix from the network data set
Considering ‘c’ no of clusters, generate ‘m’ no of candidate solutions
of n-dimensions each
Generate Modulus matrix for each candidate solution and evaluate its
fitness(modularity)
Choose candidate solution having best fitness measure as producer. It
performs production
Choose 80% of the total population as scroungers to perform
scrounging
Disperse the rest of the population to perform ranging
After each bout, evaluate all candidate solutions and replace the
producer with candidate solution possessing best modularity
No
Termination
condition satisfied
Yes
Stop
Fig.3. The friendships among individuals in the karate club
network are represented. Nodes 1 and 33 representthe
administrator and instructor, respectively. Green colour circles
represent individuals who are joined with the club's
administrator, and pink circles are individuals who joined
hands with the instructor after fission.
Fig. 2.Flow chart for the implementation of GSO.
a) Zachary karate club network
This is one of the well-studied network datasets, in
which, relations between the members of karate clubs are
described [11]. Prior to internal conflict, the club had a
friends group of 34 people. Network has 34 nodes and 78
edges [21]. Table 1 shows the results of modularity
obtained by GN [12], Newman [13], DA [14], EO [15]
and GSO. One can easily notice how GSO is much
superior to other methods.
b) Jazz musicians’ network
This is collected from the Red Hat Jazz archive [18].
The network gives the details of collaborations between
early Jazz musicians of Gleiser and Danon. It has 196
nodes and 2742 edges [21]. Table 3 compares the
modularity obtained by GN [12], Newman [13], DA [14]
EO [15] and GSO.
Algorithms
Modularity
Newman
DA
GN
EO
GSO
0.438
0.445
0.405
0.468
0.520
Table 2: Results of GSO on Jazz musicians’ network
c) American football network
This network was drawn from the schedule of games
played between 115 NCAA divisions. I-A American
college football teams in the year 2000 [12]. It has 115
nodes and 615 edges [21]. Table 4 shows the results of
modularity obtained by Newman [20], spectral [20] and
GSO.
Algorithms
Modularity
Newman
Spectral-2
GSO
0.556
0.553
0.604
Table 3: Results of GSO on American football network
d) Les Miserables
This network gives the details of interactions between
major characters in Victor Hugo's novel of crime and
redemption in post-restoration France [22]. The network
has 77 nodes and 254 edges [21]. Table 5 compares the
modularity values obtained by Newman [20], CNM
heuristic [23] and GSO.
Fig.4. The social structure of the bottlenose dolphin
community. The network is divided in to two groups as shown
above. Pink circles represent one group and green circles
represent the other.
Algorithms
Modularity
V. ACKNOWLEDGMENT
Newman
CNM heuristic
GSO
0.540
0.500
0.630
The authors thank the Centre for Development of
Advanced Computing(CDAC), Pune, India, for
providing computational facilities, software and great
working environment. Thanks also to M.E.J. Newman
for keeping datasets available to public for free of cost,
and to Dr.Sanjay Kadam, Shameek Ghosh, Kalpesh,
B.Rajasekhar, Vinay Nair and Utkarsh Gupta for helpful
discussions.
Table 4: Results of GSO on Les Miserables characters data
e) Dolphins network
An undirected social network of frequent associations
between 62 dolphins in a community listing of doubtful
sound, New Zealand [16,17]. Network has 62 nodes and
159 edges [21]. Table 2 show the results of GSO
compared with modularity obtained by Newman [13], GN
[12], EO [15] and GSO.
Algorithms
Modularity
Newman
GN
EO
GSO
0.52
0.52
0.53
0.623
Table 5: Results of GSO on Dolphins network data
VI. REFERENCES
1. Tanzeen Khalid Choudhury, "Sensing and modelling
human networks", PhD Thesis (2004), MIT.
2. Mark Newman, Networks: An Introduction, Oxford
University Press.
3. S. N. Dorogovtsev and J. F. F. Mandes, Evolution of
Networks: From biological nets to the Internet and www,
Oxford University Press, Oxford, (2003).
4. R. Albert and A-L. Barabasi, Rev. Mod. Phys. 74, 47
(2002).
5. Thomas Back and H-P.Schwefel, Evolutionary
Computing, Springer, Vol.1, No.1, pp. 1-23, (1993).
6. Herbert Spence, Principles of biology, Williams and
Norgate, (1864).
7. Eric Bonabeau and Guy Theraulaz, Swarm
Intelligence: From natural to artificial systems, Oxford
University Press, (1999).
8. March Dorigo, Mauro Birathari and Thomas Stutzle,
IEEE Computational Intelligence Magazine, Nov.,(2006).
9. Kennedy, J and Eberhart, R, "Particle Swarm
Optimization", Proceeding of IEEE International
Conference on Neural Networks, (1995).
10. S.He, Q.H.Wu, Group Search Optimizer, IEEE
Transactions on Evolutionary Computing, Vol.13, No.5,
Oct.,(2009).
11. W. W. Zachaary, J. Anthoropol. Res., 33, 452,(1977).
12. M. Girvan, M. E. J. Newman, "Community structure
in social and biological networks", Proc. Natl. Acad. Sci.
USA 99, pp. 7821-7826, (2002).
13. M. E. J. Newman, "Fast algorithm for detecting
community structure in networks", Physical Review E 69,
066133, (2004).
14. J. Duch, A. Areans, "Community detection in
complex networks using external optimization" Physical
review E 72, 027104, (2005).
15. N. Azizifard, M. Mahadavi and B. Nasersharif,
"Modularity optimization for clustering in social
networks, ICETCIP' 2011, pp. 52-55.
16. D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase,
E.Slooten and S. M. Dawson, Behav. Ecol. Sociobiol. 54,
396 (2003).
17. D. Lusseau, Proc. R. Soc. London, Ser. B(Suppl.) 270,
S186 (2003).
18. P. Gleiser, L. Danon, "Community structure in jazz",
Advances in complex systems, Vol. 6, pp. 565-573,
(2003).
19. M. E.J. Newman and M. Girvan, "Finding and
evaluating community structure in networks", Physical
Review E 69, 026113 (2004).
20. Scott White and Padhraic Smyth, "A structural
clustering approach to finding communities in graphs,
Proc. of 5th SIAM International Conf. on data mining. pp.
274-286.
21. www-personal.umich.edu/~mejn/netdata/
22. D. E. Knuth, The Stanford GraphBase: A platform for
combinatorial computing, Addison-Wesley, Reading,
MA, (1993).
23. Andrea Bettinelli, Pierre Hansen, Leo Liberti, "An
algorithm for parametric communities detection in
networks", Phys. Rev. E 86, 016107 (2012)
| 9cs.NE
|
Hindsight policy gradients
arXiv:1711.06006v1 [cs.LG] 16 Nov 2017
Paulo Rauber
IDSIA, USI, SUPSI
Manno-Lugano, Switzerland
[email protected]
Filipe Mutz∗
IFES
Serra, Brazil
[email protected]
Jürgen Schmidhuber
IDSIA, USI, SUPSI
NNAISENSE
Manno-Lugano, Switzerland
[email protected]
Abstract
Goal-conditional policies allow reinforcement learning agents to pursue specific
goals during different episodes. In addition to their potential to generalize desired
behavior to unseen goals, such policies may also help in defining options for arbitrary subgoals, enabling higher-level planning. While trying to achieve a specific
goal, an agent may also be able to exploit information about the degree to which it
has achieved alternative goals. Reinforcement learning agents have only recently
been endowed with such capacity for hindsight, which is highly valuable in environments with sparse rewards. In this paper, we show how hindsight can be introduced
to likelihood-ratio policy gradient methods, generalizing this capacity to an entire
class of highly successful algorithms. Our preliminary experiments suggest that
hindsight may increase the sample efficiency of policy gradient methods.
1
Introduction
In a traditional reinforcement learning setting, an agent interacts with an environment in a sequence
of episodes, observing states and acting according to a policy that ideally maximizes expected
cumulative reward [1]. If the agent is required to achieve a specific goal during an episode, then
such goal may be encoded as part of the states [2], possibly allowing the generalization of desired
behavior to goals that were never encountered before. Recent examples of learning goal-conditional
behavior include the works of Da Silva et al. [3], Schmidhuber [4], Srivastava et al. [5], Kupcsik et al.
[6], Deisenroth et al. [7], Fabisch and Metzen [8], Schaul et al. [9], Zhu et al. [10], Held et al. [11].
An interesting application for goal-conditional policies is hierarchical reinforcement learning [12–30],
where they may help in defining options for arbitrary subgoals [22]. In that case, instead of planning
to perform a sequence of actions, an agent could plan to achieve a sequence of subgoals, abstracting
the details involved in lower-level decisions. Recent examples of this approach include the works of
Oh et al. [31], Vezhnevets et al. [32], Kulkarni et al. [33].
While trying to achieve a specific goal, an agent may also be able to exploit information about the
degree to which it has achieved alternative goals. For example, we may expect a traveller to learn how
to arrive at his eventual destination, whether or not that was his intended destination. This capacity for
hindsight was introduced by Andrychowicz et al. [34] to off-policy reinforcement learning algorithms
that rely on experience replay [35], and has proven to be particularly important in environments with
sparse rewards.
In this paper, we show how importance sampling can be used to introduce hindsight to likelihoodratio policy gradient methods [36–39] (Sec. 2), generalizing this idea to a highly successful class of
reinforcement learning algorithms that achieve state-of-the-art results in many tasks [40]. Importance
sampling has previously been applied to policy gradient methods in order to efficiently reuse information obtained by earlier policies [41]. In contrast, our approach attempts to efficiently learn about
∗
Work performed while at IDSIA.
Accepted to NIPS 2017 Hierarchical Reinforcement Learning Workshop, Long Beach, CA, USA.
different goals using information obtained by the current policy for a specific goal. Our preliminary
experiments suggest that hindsight may indeed increase the sample efficiency of policy gradient
methods (Sec. 3).
2
Hindsight policy gradients
Policy gradients. Consider an agent that interacts with its environment in a sequence of episodes,
each of which lasts for exactly T time steps. The agent receives a goal g at the beginning of each
episode. At each time step t, the agent observes a state st , receives a reward r(st , g), and chooses an
action at . In this setting, a policy gradient method may represent a policy by a probability distribution
over actions given state and goal. The objective is finding parameters for such a policy that achieve
maximum expected return (cumulative reward). For simplicity of notation, we consider finite state,
action, and goal spaces. We denote random variables by upper case letters and assignments to these
variables by lower case letters.
Let τ = s1 , a1 , s2 , a2 , . . . , sT −1 , aT −1 , sT denote a trajectory. We assume that the probability
p(τ | g, θ) of trajectory τ given goal g and a policy parameterized by θ is given by
p(τ | g, θ) = p(s1 )
TY
−1
p(at | st , g, θ)p(st+1 | st , at ).
(1)
t=1
The expected return η(θ) of a policy parameterized by θ is defined as
η(θ) =
X
g
p(g)
X
p(τ | g, θ)
τ
T
X
r(st , g).
(2)
t=1
Employing the likelihood-ratio trick and noting that At ⊥
⊥ St0 | St , G, Θ for t ≥ t0 [39], the gradient
∇η(θ) of the expected return with respect to the policy parameters is given by
∇η(θ) =
X
p(g)
X
g
p(τ | g, θ)
τ
T
−1
X
∇ log p(at | st , g, θ)
T
X
r(st0 , g).
(3)
t0 =t+1
t=1
Therefore, if τ is a trajectory obtained by following a policy parameterized by θ to achieve a goal g
chosen by the environment, an unbiased estimate δ PG of the gradient ∇η(θ) is given by
δ PG =
T
−1
X
∇ log p(at | st , g, θ)
T
X
r(st0 , g).
(4)
t0 =t+1
t=1
We will refer to this estimate by the term policy gradient (PG), which may be used for gradient ascent.
Hindsight policy gradients. Alternatively, suppose that we are easily able to compute r(s, g) for
any state s and goal g, which is equivalent to the assumption made by Andrychowicz et al. [34].
In that case, it is possible to evaluate a trajectory obtained while trying to achieve goal g 0 for an
alternative goal g. For an arbitrary goal g 0 , importance sampling [42] allows rewriting the gradient of
the expected return as
T
−1
T
X p(τ | g 0 , θ)
X
X
p(τ
|
g,
θ)
∇
log
p(a
|
s
,
g,
θ)
r(st0 , g)
(5)
t
t
p(τ | g 0 , θ)
g
τ
t=1
t0 =t+1
"T −1
# −1
T
X
X
Y p(at | st , g, θ) TX
X
=
p(τ | g 0 , θ)
p(g)
∇
log
p(a
|
s
,
g,
θ)
r(st0 , g).
t
t
p(at | st , g 0 , θ) t=1
0
τ
g
t=1
∇η(θ) =
X
p(g)
t =t+1
Therefore, if τ is a trajectory obtained by following a policy parameterized by θ to achieve an
arbitrary goal g 0 , another unbiased estimate δ HPG of the gradient ∇η(θ) is given by
"T −1
# −1
T
X
Y p(at | st , g, θ) TX
X
δ HPG =
p(g)
∇
log
p(a
|
s
,
g,
θ)
r(st0 , g).
(6)
t
t
p(at | st , g 0 , θ) t=1
0
g
t=1
t =t+1
2
We will refer to this estimate by the term hindsight policy gradient (HPG), as it evaluates a trajectory
on goals other than the intended. Although it would be possible to sample goals to compute an
estimate of δ HPG , in this paper we focus on two variants of this estimate that are particularly simple
to compute for the environments that we chose for the experiments presented in Sec. 3.
The first variant is the uniform hindsight policy gradient δ UHPG given by
# T −1
" −1
T
X TY
X
p(at | st , g, θ) X
δ UHPG =
∇
log
p(a
|
s
,
g,
θ)
r(st0 , g),
t
t
p(at | st , g 0 , θ) t=1
0
t=1
(7)
t =t+1
g∈G
where G = {g | exists a t such that r(st , g) 6= 0}. This variant is named after the fact that δ UHPG ∝
δ HPG when the goal distribution is uniform.
The second variant is the average hindsight policy gradient δ AHPG given by
δ AHPG =
X
g
q(g)
T
−1
X
∇ log p(at | st , g, θ)
T
X
r(st0 , g),
(8)
t0 =t+1
t=1
where q is the empirical probability function over states in trajectory τ , which comes with the
additional assumption that goals represent states. This variant also ignores the likelihood-ratio,
performing updates as if the trajectory τ were equally likely given any goal. Perhaps surprisingly,
this variant achieves excellent results in our preliminary experiments.
We emphasize that these two variants are not in general unbiased estimates of the hindsight policy
gradient, which will be studied further in future work.
3
Experiments
Environments. We performed preliminary experiments in two simple environments, which allowed
systematic hyperparameter search and aggregating results across a large number of runs.
In the empty grid environment, the agent starts each episode in the upper left corner of a 6 × 6 grid,
and its goal is to achieve a randomly chosen position in this grid. The actions allow the agent to
move in the four cardinal directions or to stay in the same position (which also happens if an action is
invalid). A state or goal is represented by a pair of integers between 0 and 5. The reward is zero if the
agent is not at the goal, and one otherwise. Each episode lasts for 36 time steps.
In the bit flipping environment, the agent starts each episode in the same state (0) represented by 6
bits, and its goal is to achieve a randomly chosen state. The actions allow the agent to flip each of
the bits individually, or to remain in the same state. The reward is zero if the agent is not at the goal
state, and one otherwise. Each episode lasts for 16 time steps. Andrychowicz et al. [34] use a similar
environment to evaluate their hindsight approach.
These environments have two important characteristics in common. Firstly, the optimal behavior corresponds to reaching the goal state as fast as possible and staying there. Secondly, the UHPG/AHPG
can be computed using only the states (goals) observed during a trajectory.
Policy optimization. The policy is represented by a feedforward neural network with a single
hyperbolic tangent hidden layer and a softmax output layer. Parameters are updated using Adam [43].
The UHPG is scaled by the inverse of the cumulative likelihood-ratio across goals in G (or a large
constant if the denominator is close to zero) to avoid large discrepancies between step sizes.
Hyperparameter search. Each technique was evaluated in 21 runs lasting 10000 episodes for every
combination of environment, hidden layer size in {8, 16, 32, 64}, and learning rate in {10−1 , 5 ×
10−2 , 10−2 , 5 × 10−3 , 10−3 , 5 × 10−4 , 10−4 , 5 × 10−5 }. The average return is computed for each
individual run, and its average (average performance) is used to select the best combination for
each technique after the corresponding standard deviation is subtracted. This criterion discourages
hyperparameter combinations with large fluctuations in performance [40]. For the empty grid, all
techniques settle on a hidden layer of size 8 and a learning rate of 5 × 10−3 , except for the AHPG
with a learning rate of 10−2 . For the bit flipping environment, learning rate is 5 × 10−3 for PG and
UHPG, 10−3 for AHPG; hidden layer size is 32 for PG and AHPG, and 8 for UHPG.
3
10
20
PG
AHPG
UHPG
10
0
0
2500
5000
episode
7500 10000
(a) Empty grid results
PG
AHPG
UHPG
5
0
0
2500
5000
episode
7500 10000
(b) Bit flipping results
offgoalcumulative
likelihoodratio
return
(moving average)
return
(moving average)
Emptygrid
30
itflipping
104
100
10−4
10−8
10−12
0
5000
episode
10000
(c) Off-goal cumulative LR
Figure 1: (a) Empty grid results. PG: 16.07 ± 2.31, AHPG: 23.51 ± 3.78, UHPG: 17.06 ± 2.15.
(b) Bit flipping results. PG: 5.76 ± 1.49, AHPG: 7.34 ± 1.38, UHPG: 7.77 ± 2.83. (c) Cumulative
likelihood-ratio across goals other than the intended (UHPG, single run, values clipped at 10−12 ).
Analysis. The results of 51 additional runs using the hyperparameters described above are summarized in Figures 1a and 1b, which also include the average performance for each technique. The
learning curves represent the average smoothed returns across runs. These smoothed returns are
obtained using a moving average across episodes with a window of size 10.
These results indicate that the AHPG achieves better sample efficiency than the PG, specially in the
empty grid. Besides substituting the distribution over goals by the empirical distribution of states in a
trajectory, this HPG variant ignores the fact that a trajectory obtained while trying to achieve goal g 0
may be unlikely when trying to achieve goal g. Although this could potentially disrupt the behavior
of the policy on trajectories that are more likely to occur when trying to achieve g, this risk appears
to be outweighed by the benefits of hindsight in the environments that we considered.
The UHPG achieves comparable performance to the PG in the empty grid. This is explained by the
fact that the likelihood-ratio becomes very small for goals other than the intended as the policy begins
to exhibit distinct behaviors for distinct goals (see Fig. 1c). If the likelihood-ratio were always zero
for goals other than the intended, the UHPG would become equivalent to the PG (cf. Eqs. 4 and 7).
The off-goal likelihood-ratio no longer vanishes in the bit flipping environment (see Fig. 1c), likely
due to shorter episodes. In this environment, the UHPG also appears to outperform the AHPG during
early training. More experiments are required to investigate this behavior.
4
Conclusion
We introduced techniques that enable optimizing goal-conditional policies using hindsight. In
this context, hindsight refers to the capacity to exploit information about the degree to which an
arbitrary goal has been achieved while another goal was intended. Prior to this work, hindsight has
been limited to off-policy reinforcement learning algorithms that rely on experience replay [34].
Our preliminary experiments suggest that hindsight may increase the sample efficiency of policy
gradient methods. In future work, we will further study the properties of the proposed techniques,
include improvements commonly found in policy gradient methods (such as baselines), and perform
comprehensive experiments on more challenging environments.
Acknowledgments
We would like to thank Sjoerd van Steenkiste and Klaus Greff for their valuable feedback. This
research was supported by the Swiss National Science Foundation (grant 200021_165675/1) and
CAPES (Filipe Mutz, PSDE, 88881.133206/2016-01). We are also grateful to NVIDIA Corporation
for donating a DGX-1 machine and to IBM for donating a Minsky machine.
References
[1] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. Bradford Book, 1998. ISBN
9780262193986.
4
[2] J. Schmidhuber and R. Huber. Learning to generate artificial fovea trajectories for target detection.
International Journal of Neural Systems, 2(1 & 2):135–141, 1991. (Based on TR FKI-128-90, TUM,
1990).
[3] B. C. Da Silva, G. Konidaris, and A. G. Barto. Learning parameterized skills. In Proceedings of
International Conference of Machine Learning, 2012.
[4] J. Schmidhuber. P OWER P LAY: Training an increasingly general problem solver by continually searching
for the simplest still unsolvable problem. Frontiers in Psychology, 2013. (Based on arXiv:1112.5309v1
[cs.AI], 2011).
[5] R. K. Srivastava, B. R. Steunebrink, and J. Schmidhuber. First experiments with PowerPlay. Neural
Networks, 41(0):130 – 136, 2013. ISSN 0893-6080. Special Issue on Autonomous Learning.
[6] A. G. Kupcsik, M. P. Deisenroth, J. Peters, and G. Neumann. Data-efficient generalization of robot skills
with contextual policy search. In Proceedings of the 27th AAAI Conference on Artificial Intelligence, AAAI
2013, pages 1401–1407, 2013.
[7] M. P. Deisenroth, P. Englert, J. Peters, and D. Fox. Multi-task policy search for robotics. In Robotics and
Automation (ICRA), 2014 IEEE International Conference on, pages 3876–3881. IEEE, 2014.
[8] A. Fabisch and J. H. Metzen. Active contextual policy search. The Journal of Machine Learning Research,
15(1):3371–3399, 2014.
[9] T. Schaul, D. Horgan, K. Gregor, and D. Silver. Universal value function approximators. In Proceedings of
the International Conference on Machine Learning, pages 1312–1320, 2015.
[10] Y. Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. Fei-Fei, and A. Farhadi. Target-driven visual
navigation in indoor scenes using deep reinforcement learning. In Robotics and Automation (ICRA), 2017
IEEE International Conference on, pages 3357–3364. IEEE, 2017.
[11] D. Held, X. Geng, C. Florensa, and P. Abbeel. Automatic goal generation for reinforcement learning
agents. arXiv preprint arXiv:1705.06366, 2017.
[12] J. Schmidhuber. Learning to generate sub-goals for action sequences. In Artificial Neural Networks, pages
967–972. Elsevier Science Publishers B.V., North-Holland, 1991.
[13] M. B. Ring. Incremental development of complex behaviors through automatic construction of sensorymotor hierarchies. In Machine Learning: Proceedings of the Eighth International Workshop, pages
343–347. Morgan Kaufmann, 1991.
[14] J. Tenenberg, J. Karlsson, and S. Whitehead. Learning via task decomposition. In From Animals to
Animats 2: Proceedings of the Second International Conference on Simulation of Adaptive Behavior, pages
337–343. MIT Press, 1993.
[15] P. Dayan and G. Hinton. Feudal reinforcement learning. In Advances in Neural Information Processing
Systems (NIPS) 5, pages 271–278. Morgan Kaufmann, 1993.
[16] J. Schmidhuber. On learning how to learn learning strategies. Technical Report FKI-198-94, Fakultät für
Informatik, Technische Universität München, 1994.
[17] G. Weiss. Hierarchical chunking in classifier systems. In Proceedings of the 12th National Conference on
Artificial Intelligence, volume 2, pages 1335–1340. AAAI Press/The MIT Press, 1994.
[18] J. R. Olsson. Inductive functional programming using incremental program transformation. Artificial
Intelligence, 74(1):55–83, 1995.
[19] F. J. Gomez and R. Miikkulainen. Incremental evolution of complex general behavior. Adaptive Behavior,
5:317–342, 1997.
[20] D. Precup, R. S. Sutton, and S. Singh. Multi-time models for temporally abstract planning. In Advances in
Neural Information Processing Systems (NIPS), pages 1050–1056. Morgan Kaufmann, 1998.
[21] T. G. Dietterich. Hierarchical reinforcement learning with the MAXQ value function decomposition. J.
Artif. Intell. Res. (JAIR), 13:227–303, 2000.
[22] R. S. Sutton, D. Precup, and S. Singh. Between MDPs and semi-MDPs: A framework for temporal
abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.
5
[23] I. Menache, S. Mannor, and N. Shimkin. Q-Cut – dynamic discovery of sub-goals in reinforcement learning.
In Proc. ECML’02, pages 295–306, 2002.
[24] M. Ghavamzadeh and S. Mahadevan. Hierarchical policy gradient algorithms. In Proceedings of the
Twentieth Conference on Machine Learning (ICML-2003), pages 226–233, 2003.
[25] A. G. Barto and S. Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete Event
Dynamic Systems, 13(4):341–379, 2003.
[26] B. Bakker and J. Schmidhuber. Hierarchical reinforcement learning based on subgoal discovery and
subpolicy specialization. In Proc. 8th Conference on Intelligent Autonomous Systems IAS-8, pages
438–445, Amsterdam, NL, 2004. IOS Press.
[27] J. Schmidhuber. Optimal ordered problem solver. Machine Learning, 54:211–254, 2004.
[28] S. Whiteson, N. Kohl, R. Miikkulainen, and P. Stone. Evolving keepaway soccer players through task
decomposition. Machine Learning, 59(1):5–30, May 2005.
[29] S. Singh, A. G. Barto, and N. Chentanez. Intrinsically motivated reinforcement learning. In Advances in
Neural Information Processing Systems 17 (NIPS). MIT Press, Cambridge, MA, 2005.
[30] M. Ring, T. Schaul, and J. Schmidhuber. The two-dimensional organization of behavior. In Proceedings of
the First Joint Conference on Development Learning and on Epigenetic Robotics ICDL-EPIROB, Frankfurt,
August 2011.
[31] J. Oh, S. Singh, H. Lee, and P. Kohli. Zero-shot task generalization with multi-task deep reinforcement
learning. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of
Proceedings of Machine Learning Research, pages 2661–2670, International Convention Centre, Sydney,
Australia, 06–11 Aug 2017. PMLR.
[32] A. S. Vezhnevets, S. Osindero, T. Schaul, N. Heess, M. Jaderberg, D. Silver, and K. Kavukcuoglu. FeUdal
networks for hierarchical reinforcement learning. In Proceedings of the 34th International Conference
on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3540–3549,
International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR.
[33] T. D. Kulkarni, K. Narasimhan, A. Saeedi, and J. Tenenbaum. Hierarchical deep reinforcement learning:
Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing
Systems, pages 3675–3683, 2016.
[34] M. Andrychowicz, F. Wolski, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel,
and W. Zaremba. Hindsight experience replay. arXiv preprint arXiv:1707.01495, 2017.
[35] L. Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine
learning, 8(3/4):69–97, 1992.
[36] R. J. Williams. Reinforcement-learning in connectionist networks: A mathematical analysis. Technical
Report 8605, Institute for Cognitive Science, University of California, San Diego, 1986.
[37] R. J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning.
Machine learning, 8(3-4):229–256, 1992.
[38] R. S. Sutton, D. A. McAllester, S. P. Singh, and Y. Mansour. Policy gradient methods for reinforcement
learning with function approximation. In Advances in Neural Information Processing Systems (NIPS) 12,
pages 1057–1063, 1999.
[39] J. Peters and S. Schaal. Reinforcement learning of motor skills with policy gradients. Neural networks, 21
(4):682–697, 2008.
[40] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel. Benchmarking deep reinforcement learning
for continuous control. In Proceedings of International Conference on Machine Learning, pages 1329–1338,
2016.
[41] T. Jie and P. Abbeel. On a connection between importance sampling and the likelihood ratio policy gradient.
In Advances in Neural Information Processing Systems, pages 1000–1008, 2010.
[42] C. M. Bishop. Pattern Recognition and Machine Learning. Information science and statistics. Springer,
2013. ISBN 9788132209065.
[43] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd
International Conference on Learning Representations (ICLR), 2014.
6
| 9cs.NE
|
arXiv:1508.02792v1 [cs.NE] 12 Aug 2015
Possible Mechanisms for Neural Reconfigurability
and their Implications
Thomas M. Breuel
University of Kaiserslautern
Abstract
The paper explores a biologically and evolutionarily plausible neural
architecture that allows a single group of neurons, or an entire cortical
pathway, to be dynamically reconfigured to perform multiple, potentially
very different computations. We observe that reconfigurability can account for the observed stochastic and distributed coding behavior of neurons and provides a parsimonious explanation for timing phenomena in
psychophysical experiments. It also shows that reconfigurable pathways
correspond to classes of statistical classifiers that include decision lists,
decision trees, and hierarchical Bayesian methods. Implications for the
interpretation of neurophysiological and psychophysical results are discussed, and future experiments for testing the reconfigurability hypothesis
are explored.
1
Introduction
A common underlying assumption of much of modern neurophysiology and functional brain imaging is that the brain is divided into areas with different, identifiable functions, and that within each area, neurons perform specific computations.
Consider a pathway like the visual pathway (V1, V2, V4, IT). Commonly,
such a pathway is assume to compute a sequence of representations of the sensory input at increasing levels of abstraction. However, in the visual pathway,
despite extensive attempts to identify the nature of representation in V4 and
IT, not clear representation has been identified.
The reconfigurability hypothesis proposed in [1] states that we should view
such a pathway not as a neural network with a fixed function, but instead as
a reconfigurable network that performs different computational tasks at different times. In effect, the pathway is temporally multiplexed between different
functions.
A reconfigurable pathway can be in one of many different states (configurations). In different configurations, areas may perform different computations
This paper is the long version of talk presented at Snowbwird Learning Workshop, April
2012[1].
1
and generate different representations; in particular, the activities of individual
neurons may be unrelated in different configurations. In a sense, we think of
neural pathways as being similar to digital FPGAs (field programmable gate
arrays).
Reconfiguration of pathways is postulated to be under control of one or more
brain areas outside the pathway. Furthermore, we generally consider pathway
configurations in which the pathway itself operates largely like a feed-forward
network. We will see that combining feed-forward networks with reconfigurability results in additional computational power that allows feed-forward networks
to perform some of the functions usually thought of as requiring recurrent architectures.
As we will see, reconfigurable networks are easy to create with simple, biologically plausible mechanisms. We will argue that they likely confer a significant evolutionary advantage on animals because neurons require large amounts
of energy to maintain and reconfigurability permits an organism to perform
more diverse and complex computations with the same number of neurons. In
addition, reconfigurability provides a simple explanation for a number of puzzling observations in neurophysiology and psychophysics that currently have
no widely accepted, specific computational models, such as stochastic neural
responses and timing behaviors.
A second important issue in understanding the function of neural networks
has been the question of how recurrency is involved in neural computations.
Neural network models (spiking or McCulloch-Pitts-style) are generally divided
into two classes: feed-forward networks and recurrent networks. Feed-forward
networks are widely used and well studied; they compute non-linear functions
that are commonly used as decision functions in classifiers, or function approximators. Recurrent networks have temporal state and dynamics. There is a wide
variety of models, such as the Hopfield model[4], Bayesian message passing algorithms, liquid or echo state networks[6], and Boltzmann machines[3, 10]. In
addition, special-purpose recurrent networks, such as networks for attentional
selection[12] and segmentation have been proposed.
Numerous effects have been observed, both neurophsiologically and psychophysically, that cannot be explained with simple feedforward models, yet
explanation in terms of biologically plausible recurrent networks has been difficult. We will see that reconfigurability not only results in more efficient utilization of neural hardware, but it also extends the computational capabilities
of feed-forward neural networks significantly.
In the rest of the paper, we will first formulate and formalize a simple model
of reconfigurable feed-forward networks, then examine how such models related
to neurophysiology, psychophysics, and statistical learning.
2
Reconfigurable Feed-Forward Models
Although reconfigurable models can be formulated for many different kinds
of networks (including spiking neural networks and recurrent networks), for
2
concreteness, let us consider a simple model primarily based on feed-forward
models for most of the rest of the paper.
2.1
Multiplexed Feed-Forward Models
Consider a feed-forward model with linear combination of input values, followed
(k)
by a non-linearity. Let the activity of unit j at layer k be represented as xj ,
(k)
let the weight matrix at layer be Mi,j , and let the nonlinearity be given by σ(·)
(we assume that this is an element-wise non-linearity, like the sigmoid function,
and/or a normalization function like f (x) = x/||x||). Then a feed-forward model
is given by:
X
(k) (k)
x(k+1) = σ
(1)
Mi,j xj
Reconfigurability postulates an additional set of parameters λr and replaces the
feed-forward equations with:
X X (k,r) (k)
(2)
x(k+1) = σ
Mi,j xj
λr
r
j
(Additional nonlinearities could be introduced inside the outer sum.)
We are primarily considering models in which the pathways are P
in one configuration r of R discrete configurations. That is, λr ∈ {0, 1}R ,
r λr = 1
(more general models of the form λr ∈ R are also possible).
From these equations, it is easy to see that for each r ∈ {1, ..., R}, the
network computes completely the result of applying a different feed-forward
network to the inputs.
Let us call these kinds of feed-forward models a multiplexed feed-forward
model, since the same neural pathway can be used to implement multiple different (and potentially unrelated) computations in sequence. This kind of mechanism allows the same number of neurons to perform many more computations
compared to a non-multiplexed pathway. It is attractive because it is easily
implemented via a biologically plausible mechanism.
2.2
Sequential Control of Configurations
Multiplexed feed-forward models are one component of reconfigurable network
models. The second component we need is a control mechanism that chooses
configurations. In the most general case, a control mechanism might consist of
a recurrent network receiving as inputs sensory inputs, internal state, and outputs from the multiplexed pathway. But, perhaps surprisingly, we can perform
many kinds of useful computations with much more limited control mechanisms.
These kinds of more limited control mechanisms will be useful in formulating
hypotheses about the timing behavior in psychophysical experiments, and they
also provide a plausible evolutionary path by which more complex control mechanisms may have evolved. To describe these control mechanisms, consider a task
3
to be solved, like a sequence of presentations and responses in a psychophysical
vision experiment.
Type I The simplest control mechanism is to choose a fixed configuration at the
beginning of a task and then operate with that configuration throughout
the duration of the task. This is the kind of setting in which we can explain
results such as fast feed-forward object recognition[9]. The configuration
itself is chosen by mechanisms outside the reconfigurable pathway and it
is chosen independently of its outputs.
Type II Configurations may also be chosen using hierarchical control. In that case,
a separate pathway performs analysis of sensory input (say, itself using a
feed-forward architecture), and its output then controls the configuration
of the reconfigurable pathway.
Type III Configurations may be sequentially multiplexed during a task. That is,
the parameters λr cycle through a fixed set of values, corresponding to
different configurations of the pathway. The output of the reconfigurable
pathway switches between the results of different computations. These
may then be integrated by a subsequent feed-forward network (possibly
after some temporal integration).
Type IV Configurations may be sequentially tried out during a task, where each
configuration attempts to complete the task; if a configuration is successful, the task is completed and the sequence of configurations starts again
at the beginning.
Type V Configurations depend on the output of the reconfigurable pathway, but the
only information that is recurrent is the set of parameters λr . The overall
output from the pathway is computed possibly by an additional feedforward computation operating on the output of the configurable pathway,
as in Type III.
Type VI Configurations depend on the output of the reconfigurable pathway, but
in addition, other information is propagated recurrently. For example, the
control mechanism may cause activations to be buffered or stored inside
the pathway.
Except for Type I control, all these mechanisms assume that the successful completion of some recognition requires the use of multiple configurations,
and that the integration of the results from these configurations yields the final
answer to the perceptual problem that is being solved during the task. As a
concrete example, consider a task in which subjects need to classify either outdoors scenes or recognize letters; solving this task would run through a sequence
of configurations specialized for recognizing outdoors scenes and a sequence of
configurations specialized for recognizing letters. Depending on which control
mechanism is used, we would expect different kinds of timing behaviors. Each
configuration would indicate whether it successfully interpreted the input and
4
what class the input belongs to. From a pattern recognition point of view,
the resulting classifier would be analogous to a decision list, a multi-classifier
system, or a decision tree (we will return to this point later).
This division into different types of control is intended to help with discussions about timing behavior; we are not postulating that there are different
kinds of neural structures implementing different kinds of control mechanisms.
Instead, the mechanisms controlling reconfigurability are proposed to be somewhat analogous to motor control in the general case (Types V and VI), but
can perform simple and repetitive types of control (Types I through IV) when
faced with simpler tasks–often the kind of constrained tasks encountered in
psychophysical experiments.
The rest of the paper is mainly concerned with control mechanismsP
I - IV and
assuming simple, discrete reconfigurations in which λr ∈ {0, 1} and
λr = 1.
This should not be taken to mean that there is any reason why more complex
forms of control could not exist. But we will see that these simple and limited
forms of recurrence are sufficient to explain many complex phenomena.
3
Neurophysiological Basis
Merely gaining some computational power through postulating more complex
computations than linear threshold units would by itself be unremarkable. What
makes Equation 2 interesting is that it has a simple, biologically plausible implementation in terms of real neurons.
Recall that most neurons (including most cortical neurons) consist of a (usually) large dendritic tree, having somewhere between 104 and 106 synapses,
a trigger zone, and an axon that eventually branches out and forms further
synapses. The dendritic tree is a large branching membrane that has a topological structure corresponding to a tree structure in mathematics and computer
science. Signals transmitted by synapses can be excitatory or inhibitory and
generally contribute additively to the voltage at the trigger zone of a neuron.
Some synapses can be multiplicative in nature. Such multiplicative synapses
are linked to the topological structure of the tree. That is, they combine multiplicatively with the additive contributions of the synapses within their own
branch.
When we combine additive and multiplicative synapses, potentially arbitrary
boolean expressions can be implemented, with the tree structure of the dendritic
tree corresponding to the nesting of the boolean expression[7]. A criticism of
using the computational capabilities of dendritic trees in this way is that it is
unclear how such structures can be created during ontogeny, or even how such
connectivity might be encoded genetically.
For reconfigurable networks, we take advantage of these computational capabilities of dendritic trees, but without requiring the kind of specific connectivity
needed for the implementation of arbitrary boolean circuits. Instead, we rely
on the existence of a multiplicative synapse type that preferentially forms on
branches near the root of the dendritic tree.
5
Putting these observations together, we then arrive at the following process
for the development of reconfigurable neural networks. First, neurons from the
area controlling the reconfigurable pathway send axons into the pathway. Each
axon makes random multiplicative connections to branches of the dendritic trees
of neurons in the reconfigurable pathway.
For the simple model described by Equation 2, there should be a connection
from each controlling neuron to only a single branch of each neuron in the
target pathway, and multiple controlling neurons should not connect to the
same target branch (more complex forms of control and reconfiguration are
possible, however). Such constraints do not require specificity in the connection
and can be implemented in terms of simple signaling molecules or development
dependent on neural activity during development.
Alternatively, axons from controlling neurons can first establish synapses on
the body of neurons or on a dendritic stump in the target pathways, and this
synapse formation then itself triggers the formation of a larger dendritic subtree
under the control of the controlling neuron.
Through either mechanism, each neuron in the target pathway then ends up
with a collection of dendritic subtrees, each of which is gated by a signal from a
controlling neuron via a multiplicative synapse. Furthermore, the establishment
of this kind of network was based on random connections from controlling to target neurons and requires only simple, local signaling during development. This
distinguishes it from many prior proposals for taking advantage of multiplicative
synapses
A second observation is that, except for the formation of the controlling
synapses, the neurons in the reconfigurable pathway are free to form other
synapses by whatever rules, statistics, and signals that are required for a particular function. Activation of a set of controlling synapses will simply select
a subset of these connections for a computation. If these activation patterns
for the controlling synapses are disjoint, then the reconfigurable pathway will
behave as if it could be switched between multiple, unrelated feed forward networks, as described by Equation 2 (actually, there is nothing in this description
that prohibits the formation of recurrent connections, so recurrent networks can
also be reconfigured using this mechanism).
In addition, signals flowing backwards from the body of the neuron towards
the synapses are also blocked by the kind of multiplicative synapses used in this
construction. This means that the reconfiguration implied by the control signals
does not just establish a multiplexing for feed-forward computations, but also
for learning. That is, the reconfigurable pathway behaves in each configuration
as if it were a separate neural network, both for the purposes of computation
and for the purposes of learning.
4
Evolutionary Rationale
In the previous section, we saw that reconfigurable pathways are easy to implement in terms of neural hardware: the required computations and synapse
6
types are common, and creating the network only requires simple intercellular
signaling and random connections, both commonly observed in the nervous system. But although we have seen (and will explore in more detail below) that
reconfigurable pathways are useful for high-level computations, the question is
how such a complex mechanism might have arisen.
Let us consider a plausible evolutionary history of such a pathway. Neurons
are energetically costly for organisms to maintain. Reducing the number of neurons that an organism needs to maintain is therefore evolutionarily potentially
highly advantageous. The reconfiguration mechanism describe above permits
this by providing a simple mechanism for multiplexing a neural pathway. Such
a circuit allows the same number of neurons to carry out a much larger number
of neural computations, where the additional cost of adding new functions is
limited to the cost of adding the extra membrane needed for the additional dendritic subtree, instead of developing entirely new neurons. The price paid for
this is that multiplexed neural computations need to be carried out sequentially
instead of in parallel.
Initially, control mechanisms may have been limited to Type II and III control described above, which require only a connection from a sensory input or a
simple oscillatory neural circuit to a pathway. In fact, Type II control, in which
a neural pathway is switched depending on sensory input, in its simplest form
is a simple kind of associative learning, in which long term potentiation and/or
long term depression is conditioned on some other sensory input or context.
Once the necessary genetic circuitry for establishing a control/reconfiguration
relationship between two neural areas was established, the next step may have
been increasing the complexity of the control circuit, for example by reusing
neural circuitry for motor control and motor planning. Once complex control of
reconfigurable pathways was available and provided greatly increased learning
and adaptation capabilities, it may then have become replicated and formed the
basis for cortical circuits in higher animals.
Although this evolutionary view is, of course, highly speculative, it predicts
that we should be able to find early forms of reconfigurability already in simple
neural circuits, where they are potentially part of circuits for long term potentiation, long term depression, and classical conditioning. It also predicts that
some of these circuits would be the predecessors of more complex reconfigurable
pathways in higher animals, suggesting that ontogenic signals and neurotransmitters may be related between such simple circuits and corresponding circuits
in higher animals.
5
Interpretation of Existing Experimental Results
The previous sections introduced a biologically plausible mechanism for extending the computational capabilities of feed forward networks (we will return to
this point in a later section). However, by itself, that is not particularly inter-
7
esting, since there are many such possible extensions. What actually motivated
the development of the above architecture was a number of observations in
neurophysiology and psychophysics that were different to explain in terms of
proposed existing mechanisms. This section describes two observations that
motivated the proposal of reconfigurable pathways: the observation of stochastic and distributed representations in neurophysiology and priming behavior in
psychophysics. The intent of giving these examples is not to provide a complete theory or explanation of these phenomena in terms of reconfigurability
(that will require a much more detailed formulation of models and examination
of experimental results than possible here), but merely to illustrate the kinds
of explanations that reconfigurable models can potentially yield for complex
phenomena in neurophysiology and psychophysics.
5.1
Neurophysiology
Neurophysiology has given insights into neural codes in some cases, such as early
visual representations[5] and motor behavior like bird songs[2]. In the case of
early visual representations, neural codes roughly correspond to common image
transformations based on simple statistical principles. In the case of motor
behavior, neural codes represent spatial and temporal activation patterns for
muscles. In both cases, representations are largely determined by the physical
statistical nature of the signals.
However, such easily interpretable representations have not been identified
for higher level concepts. For example, even though object categories and instances appear to have psychological reality, no individual neurons that clearly
correspond to these concepts, so called “grandmother cells”, have been identified.
Another common observation in neurophysiological experiments is that the
response of neurons to stimuli tends to be stochastic; that is, a given stimulus
does not reliably produce an action potential.
Since a simple correspondence between psychological concepts and neural
activity has not been identified, stochastic, distributed representations have been
postulated as representations of psychological concepts such as object class and
identity. In such representations, no neuron responds specifically to a concept or
category, but instead concepts are encoded in the joint activity of many neurons.
For object recognition, this means that instead of observing cells that respond
to specific objects, we expect to observe cells that respond stochastically to
multiple objects. Such cells are referred to as totem pole cells[11].
Intuitively, such representations are attractive because if a system is based on
stochastic and distributed representations, we would expect it to be robust to the
failure of individual neurons. However, stochastic, distributed representations
of the form postulated for neural representations otherwise have not been found
much use for computational purposes. Furthermore, little other support for
their existence has been found in the literature.
An additional problem with stochastic, distributed representations is that
they would seem to require a significant amount of temporal integration in
8
order to decode, for example into a motor response. However, this is difficult to
reconcile with experimental results on fast feed-forward object recognition.
Reconfigurability provides a simpler explanation of these observations.
• The response of individual neurons appears stochastic under some experimental conditions because their responses are different in different configurations. If the configuration input to the pathway were known, the
responses of neurons would be much more predictable relative to that
input.
• The response of neurons to multiple unrelated stimuli (“totem pole cells”)
is explained not as distributed coding, but instead as different functions
of the neurons in different configurations. Within each configurations, the
response of individual neurons is specific.
Let us note that context-dependent changes in the preferred stimuli for neurons has been observed in place cells[8]. Within a single “room” or context,
place cells respond quite specifically to spatial locations. But when the room
or context is changed, the same place cell may respond to a different, unrelated
location. That is, we observe a neuron responding to a high level pattern (in
this case, location within a room), yet that response varies and is context dependent. Reconfigurable pathways in the above sense are a potential mechanisms
for place cells.
Finally, observe that some degree of reconfigurability is also observed physiologically through fMRI (functional magnetic resonance imaging). That is, depending on the task being solved, different brain areas are differentially active.
It has been demonstrated that changes in neural activity cause the physiological
changes observed using fMRI (as opposed to the physiological changes causing
changes in neural activity). That is, in some way, either neural output from a
source area or neural input to a target area is modulated in such a way that the
target area becomes more or less involved in solving a mental task.
The question is now whether we can experimentally distinguish reconfigurable pathways from alternative explanations of the observed neural activity, such as explanations involving Bayesian message passing or stochastic distributed coding. A sufficient experimental result allowing us to distinguish
between the reconfigurability hypothesis and such other explanations would be
the observation of correlations between signals going into axo-dentritic “gating”
synapses near the cell body in an area and the response of the cells.
The reconfigurability hypothesis predicts that if there are multiple such incoming gating signals and we consider the response of a cell only while a specific
gating input is active, the neuron will then behave in a more deterministic and
more specific way; that is, in the context of incoming gating signals, it will
stop behaving like a stochastic totem pole cell, and more like a deterministic
grandmother cell.
A second neurophysiological prediction is that synapses that are suitable
for performing the gating function for implementing a reconfigurable pathway
should respond to stimuli based on context and priming (see below). That is,
9
certain contexts should activate such synapses more frequently, while others
should depress them.
5.2
Priming
In the previous section, we looked at possible explanations for observations in
neurophysiology based on a reconfigurable pathway model. Let us now look at
an example from psychophysics.
Priming is a process observed in many experiments in which presentation
of one stimulus, or context, affects the perception of subsequent stimuli. Most
commonly, priming with a stimulus speeds up recognition of subsequent, related stimuli. Priming also reduces neural activity for subsequent processing,
as measured by neurophysiology, fMRI, and EEG.
Such results are generally interpreted in the framework of a dynamic systems
view of neural networks, typified by Hopfield networks. Within a dynamic
systems framework, neural computations need to converge to a final state and
the time for convergence depends on the state of the network and the starting
point for the computation. However, despite this general view of priming, no
specific neural network models of priming are widely accepted, and priming
effects are difficult to explain with commonly used feedforward neural networks.
Reconfigurable pathways provide a alternative, simple explanation for priming effects. For computation with reconfigurable pathways, the time until a
result is obtained depends on the number of configurations that need to be
explored and potentially the order in which they are explored; when a perceptual task can be solved by utilizing fewer configurations, or a configuration that
solves the task can be identified earlier (depending on which type of control we
assume), the task can be completed more quickly. An explanation for priming
effects in a reconfigurable model is that they change the order and/or choice of
configurations used in the performance of a task.
As a concrete example, consider a visual object recognition task in which
objects are presented sequentially on a display. In a reconfigurable model of
visual object recognition, recognition would be carried out by many different
configurations of the visual pathways, where each configuration recognizes a
category of objects related by invariance properties, degree of shape variation,
relative importance of color vs. shape, etc. Priming with an image or concept
can then be understood as a reordering in which the different configurations are
considered, with configurations related to the priming stimulus being considered
earlier than they otherwise would be. Assuming Type IV control (the simplest
form of the control mechanisms above that is capable of showing variable timing), this would result in the experimentally observed influence of priming on
recognition speed.
This interpretation is also consistent with neurophysiological results showing
reduced and “more focused” activity in the presence of priming. Generally, neural computations at any one time only activate a small fraction of all neurons in
a pathway. If two configurations represent very different kinds of computations,
the probability that a neuron is activated in one and the other configuration
10
would be independent of each other. Therefore, the probability that any specific neuron is activated throughout the performance of a task increases with
the number of configurations used in the task, consistent with what is measured
in priming studies.
It is an interesting question whether we can test and distinguish reconfigurable models from other models purely using psychophysical experiments. Reconfigurability actually makes fairly specific predictions: the amount of time
required to complete a perceptual task grows with the number of configurations
that need to be considered in its solution, and the accuracy depends on the
best configuration that was identified. Experimentally, there are a number of
mechanisms we can potentially use to manipulate and measure the number of
configurations that are used in the solution to a task. For example, if we give
the subject only a fixed amount of time to complete a task, the subject would
ideally choose a statistically optimal subset of configurations for solving that
task. As a special case, for only very short amounts of time for solving a task,
only a single configuration may be available. Based on such manipulations, it
should be possible to determine whether experimentally observed timing and
error behavior is consistent with the predictions of reconfigurability.
It needs to be born in mind, however, that existing explanations of timing
and error behavior in psychophysical experiments tend to be so non-specific
that they are difficult to disprove by any experiment. For example, models
that are based on message passing or dynamical systems views of neural systems generally allow for highly variable response times for perceptual tasks in
psychophysical experiments, but they do not make many specific, falsifiable
predictions.
In other areas of psychophysics, such as visual search, models based on concepts such as “parallel search” and “serial search” are commonly used in the
interpretation of experiments. Such models are, however, merely descriptive
of the timing behavior, not mechanistic; that is, rather than being an alternative hypothesis to reconfigurability, reconfigurability may provide a neural and
mechanistic basis for these timing behaviors.
On the whole, we can be optimistic that psychophysical experiments can
provide evidence for reconfigurable models. At the same time, it is probably
time for models based on dynamical systems, message passing, and stochastic
and distributed representations, to be made concrete to the point where they
can also make specific timing and error predictions for different experimental
conditions.
6
Statistical Learning
Learning and classification in feed-forward networks is a well-researched and
well-understood problem. Computationally, feed-forward networks can approximate functions, implement classifiers, implement boolean circuits, and perform
feature extraction.
Particularly important applications of feed-forward neural networks are their
11
use as classifiers. Commonly used architectures and training techniques for
classifiers are multilayer perceptrons and convolutional networks trained with
backpropagation, radial basis functions trained using least square methods, and
support vector machines. Nearest neighbor classifiers can also be implemented
easily with feed forward networks.
However, many classifiers that are commonly used in pattern recognition,
are less naturally implemented in terms of neural networks. Below, we discuss
this for decision lists, decision trees, multi-classifier systems, and style models,
showing different capabilities of reconfigurable networks.
Such classifiers have in common that they are made up of a potentially large,
training-data dependent number of component classifiers that are invoked in
sequence and whose outputs are integrated into a final classification decision.
Let us call an abstract model of this class of classifiers a sequential compound
classifiers. Such classifiers defined by a finite state machine determining which
classifier to execute next, how to integrate the output of this classifier with
previous classifiers and when to terminate the computation. The transition
matrix is a function of the output of the current classifier. We will study these
kinds of systems and their learning algorithms more formally in a separate paper.
Although, in principle, many of these classifiers can easily be parallelized, by
implementing their component classifiers in parallel and then combining their
outputs within another feed-forward, this is not a natural way of implementing
them, since the number (and sometimes kind) of classifiers is data-dependent
and grows with more training data. Furthermore, many of the component
classifiers are invoked only rarely, making dedicating neural hardware to them
wasteful. Parallel implementations would require mechanisms for allocating new
groups of neurons as more input data becomes available and connecting these
new groups back to existing groups in the right way. Furthermore, the time
required for classification in such classifiers depends on the input data.
Reconfigurable pathways permit a more direct implementation of these kinds
of classifiers: different component classifiers are implemented by different configurations, and the order in which they are invoked and the way their outputs
are combined is handled by the control circuits. Let us look at these kinds of
classifiers in more detail.
For decision lists, a sequence of classifiers needs to be executed. Each such
classifier outputs either the final classification, or a special class indicating that
it could not provide the answer. The control circuit circles through a sequence
of states until the first classifier returns a final classification. Such a control
circuit can be implemented as a simple feed forward network in terms of linear
threshold units.
Let us look at this construction in more detail in the case of decision lists to
illustrate how this works mathematically. For a decision list, we have a sequence
of classifiers fi giving some output y = fi (x), where the fi are implemented
as configurations of the pathway and feedforward networks. As component
classifiers of a decision list classifier, each classifier either outputs a classification,
or an indication that it couldn’t classify the input sample; we assume that this
indicator is a binary output y0 , with y1 . . . yn representing classification output.
12
We augment the output of each configuration with a set of neurons representing
a unary encoding of the configuration number lj = δij , with δij = ⌊i = j⌋. The
control network is itself a feed-forward classifier that, as input, takes y0 and the
lj and produces as output a unary representation λk of the next configuration to
be used. Without the stopping criterion, we would simply have a feed forward
control network that cycles through the configurations (here, (x)R is a shorthand
for x mod R.
λk = lj · δk,(j+1)R
where R is the total number of configurations With the stopping criterion, we
arrive at:
λk = (1 − y0 ) · lj · δk,(j+1)r + y0 δk,0
To guard against noise, we may want to apply a threshold unit instead of using
the linear computation.
In this construction, we have considered the state labels lj an augmentation to the feed forward networks in the reconfigurable pathways. This lets us
consider the entire pathway and its control network to be a single, simple feed
forward network where the only information that flows recurrently is the choice
of configuration r. Anatomically, the state labels lj might, of course, also be
located in the control network, giving the control network the appearance of a
more complex recurrent network.
For decision trees, there are two kinds of classifiers, those that are in the
interior of the tree, and those that are at the leaves. The former return as
output the next classifier to be tried, the latter return a final answer to the
classification problem. The control circuit is similar as for decision lists, but
it needs to keep track of the relationship between classifiers, classifier outputs,
and states.
Multi-classifier systems compute the outputs from multiple individual classifiers and then combine the outputs of the discriminant functions or posteriors
of the classifiers, often by averaging. In the simplest case, the control circuit cycles through a static sequence of component classifiers, although more complex
choices are possible. The control circuit also needs to coordinate the averaging
of the classifier outputs; alternatively, the outputs may simply be averaged using time constants, so that the final output of the pathway is always a weighted
temporal average of the outputs of the past several configurations in the pathway.
Style models are models in pattern recognition in which a group g of related
classification problems {xg,1 , . . . , xg,r } can be solved better by taking advantage
of their relationship. One of the most general models for styles is the hierarchical
Bayesian model, in which we assume that the class conditional densities p(x|c, θ)
are dependent on some parameter θ, and for a group of related samples, θ was
sampled from a prior parameter density p(θ). When the prior parameter density
is a linear combination of delta functions,
X
Λi δ(θ, θi )
p(θ) =
i
13
then the hierarchical Bayesian model becomes a finite mixture model:
X
λi pi (x|c)
p(x|c, θ) = p(x|c, λ) =
i
Often, the Λi are assumed to be in {0, 1} (i.e., the world is in one of a number
of discrete and distinct states for each group of input samples), and hence a
maximum likelihood estimate for the λi is itself in {0, 1}. If we take a Bayesian
approach to classification, the λi take on continuous values. For optimal classification under a zero-one loss function, the classifier should classify using the
posterior probabilities (or a strictly monotonic function thereof) as the discriminant function. The posterior probabilities are easily seen to be expressible in
the form:
X
λi pi (c|x)
p(c|x) =
i
This is similar to the multi-classifier case, but the final output is a weighted
average, not just an average, of the different component classifiers.
The weights themselves are determined by how well each group g of samples
is explained by each model. The optimal decision is also a classification problem,
namely determining whether mg , the model for group g, is given by the set of
class conditional densities pi (x|c). In the absence of class labels or other classrelated information, we can only determining which model fits the data in the
group best by marginalizing and considering pi (x). We then obtain using Bayes
formula and assuming independent samples:
Y
p(mg = i|G) = p(mg = i|xg,1 , . . . , xg,r ) ∝
pi (xg,i )P (mg = i)
The weights λi should be chosen either proportional to p(mg = i|G) or (for
maximum likelihood) as arg maxi p(mg = 1|G). (If extra information is available
prior to the presentation of the training samples, that gives us information
about P (mg ), then we have a situation related to priming in psychophysical
experiments.)
Since neural systems usually face a sequential decision problem, in which
the first elements of a group G need to be classified before the rest of the
batch is available, and in which transitions between different groups may not
be marked explicitly, a fairly simple strategy for estimating the p(mg = i|G) is
to approximate it with a sliding window, i.e. for a given sample xn :
p̂(mg = i|G) ≈
n
Y
pi (xj )
j=n−k
If we take logarithms on both sides, we obtain:
log p̂(mg = i|G) ≈ log
n
X
log pi (xj )
j=n−k
This then gives a simple prescription for style models based on reconfigurable
pathways:
14
• Each configuration of the pathway is a classifier that estimates pi (c|x), as
well as a density estimator that estimates log pi (x) (this can be interpreted
as a confidence the classifier has in classifying the input sample)
• The pathway is cycled through all relevant configurations i (configurations
for which log pi (x) is low during one iteration may simply be skipped on
future iterations)
• The outputs are integrated by temporally integrating the log pi (x) and
averaging the pi (c|x) using as weights the exponential of the integrated
weights.
In this way, we see that reconfigurable pathways can implement a common
form of hierarchical Bayesian classifiers (namely, those represented as mixture
models), even for large numbers of mixture components. Let us note that if
we drop the constraint that neural pathway configurations are totally distinct,
and assume instead that mixed configurations are possible, we can imagine
more complex hierarchical Bayesian models, in which the control of the pathway attempts to maximize p(x) on average by “mixing” different configurations
together.
Overall, the goal of this section has been to show that reconfigurable pathways permit the plausible implementation of a rich set of widely used classifiers
and pattern recognition methods that, due to their sequential nature, used to
be considered not plausible as instances of neural computation.
7
Discussion
The paper has described a biologically plausible mechanism that allows a collection of neurons or an entire neural pathway comprising multiple areas and
connections between them, to be reused for multiple different, potentially unrelated computations. We have seen evolutionary rationales, and some relationships to findings in neurophysiology, psychophysics and machine learning. Of
course, the discussions in this paper are not sufficient to prove the actual use of
reconfigurable pathways in the brain.
Nevertheless, the reconfigurability hypothesis here has been stated in greater
detail and with more experimental predictions than other common theories of
dynamic brain function and distributed representations, such as Bayesian message passing, dynamic systems theories of brain function, and stochastic distributed representations. In particular, the reconfigurability hypothesis makes
specific predictions about the kinds of neural signals and circuits we expect to
find neurophysiologically, as well as predictions about the relationship between
task structure and timing across potentially a wide range of psychophysical
experiments.
The analysis of psychophysical and neurophysiological results above can,
of course, only be considered tentative. The next steps are to look in much
more detail at the degree to which existing results can be explained in terms of
15
reconfigurability, as well as to devise specific experimental tests of the reconfigurability hypothesis.
From a theoretical and machine learning point, the reconfigurability mechanism is significant, not so much in that it contributes new machine learning
methods, but rather in that it provides a plausible neural basis for a large range
of existing machine learning methods (e.g., decision trees, decision lists, hierarchical Bayesian methods), for which previously specific neural implementations
either didn’t exist or even seemed implausible.
No matter whether reconfigurability will ultimately turn out to true, it does
provide a challenge to common interpretations of neurophysiological and psychophysical results. Right now, these experiments are generally interpreted with
the implicit assumption that neurons and brain areas perform specific, relatively
consistent functions, and that properties of, and computations performed by,
say, the visual system remain stable across experiments. The mere possibility
that different experiments may actually be testing completely different neural
computations is a significant factor that needs to be taken into account in future
work.
References
[1] Thomas
Breuel.
A
reconfigurable
classifier
model
of
the
visual
system.
Snowbird
Learning
Workshop,
http://snowbird.djvuzone.org/2012/abstracts/149.pdf, 2012.
[2] Richard HR Hahnloser, Alexay A. Kozhevnikov, and Michale S. Fee. An
ultra-sparse code underliesthe generation of neural sequences in a songbird.
Nature, 419(6902):65–70, 2002.
[3] Geoffrey E. Hinton and Terrance J. Sejnowski. Learning and relearning in
Boltzmann machines. MIT Press, Cambridge, Mass, 1:282–317, 1986.
[4] John J. Hopfield. Neural networks and physical systems with emergent
collective computational abilities. Proceedings of the national academy of
sciences, 79(8):2554–2558, 1982.
[5] David H. Hubel and Torsten N. Wiesel. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal
of physiology, 160(1):106–154, 1962.
[6] Herbert Jaeger. The echo state approach to analysing and training recurrent neural networks-with an erratum note. Bonn, Germany: German
National Research Center for Information Technology GMD Technical Report, 148:34, 2001.
[7] Christof Koch. Biophysics of computation: information processing in single
neurons. Oxford university press, 2004.
16
[8] Edvard I. Moser, Emilio Kropff, and May-Britt Moser. Place cells, grid
cells, and the brain’s spatial representation system. Annu. Rev. Neurosci.,
31:69–89, 2008.
[9] Maximilian Riesenhuber and Tomaso Poggio. Hierarchical models of object
recognition in cortex. Nature neuroscience, 2(11):1019–1025, 1999.
[10] Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton. Restricted
Boltzmann machines for collaborative filtering. In Proceedings of the 24th
international conference on Machine learning, pages 791–798. ACM, 2007.
[11] Simon J. Thorpe. Grandmother Cells and Distributed Representations.
In Visual Population Codes: Toward a Common Multivariate Framework
for Cell Recording and Functional Imaging, pages 23–52. MIT Press Cambridge, 2012.
[12] Dirk Walther, Laurent Itti, Maximilian Riesenhuber, Tomaso Poggio, and
Christof Koch. Attentional selection for object recognitiona gentle way. In
Biologically Motivated Computer Vision, pages 472–479. Springer, 2002.
17
| 9cs.NE
|
arXiv:1604.00959v1 [math.GR] 4 Apr 2016
The semaphore codes attached to a Turing machine via resets and
their various limits
John Rhodes
Department of Mathematics, University of California, Berkeley,
CA 94720, U.S.A.
email: [email protected], [email protected]
Anne Schilling
Department of Mathematics, University of California, Davis,
One Shields Ave., Davis, CA 95616-8633, U.S.A.
email: [email protected]
Pedro V. Silva
Centro de Matemática, Faculdade de Ciências, Universidade do Porto,
R. Campo Alegre 687, 4169-007 Porto, Portugal
email: [email protected]
April 5, 2016
2010 Mathematics Subject Classification: 20M07, 20M30, 54H15, 68Q05, 68Q15, 68Q70
ABSTRACT
We introduce semaphore codes associated to a Turing machine via resets. Semaphore
codes provide an approximation theory for resets. In this paper we generalize the
set-up of our previous paper “Random walks on semaphore codes and delay de
Bruijn semigroups” to the infinite case by taking the profinite limit of k-resets to
obtain (−ω)-resets. We mention how this opens new avenues to attack the P versus
NP problem.
1
Introduction
In our previous paper [9], we developed algebraic foundations centered around the prime decomposition theory for finite semigroups and finite automata (see [5], [6] and [8, Chapter 4]). This analysis
focused on the right zero component action, when the corresponding pseudovariety contains all finite
transformation semigroups (X, S) (or automata) with the property that there exists some k ≥ 1 such
that every product s1 · · · sk (with si ∈ S) is a constant map (or reset) on X. Such finite automata
were called k-reset graphs in [9, Section 3] and their elementary properties were studied, using the
lattice of right congruences on the finite free objects (De Bruijn semigroups), in [9, Sections 5 and 6].
Semaphore codes, which are well known in the literature (see [2]), were proved [9, Sections 4 and
7] to be in bijection with special right congruences and provide a lower approximation to any right
1
congruence with the same hitting time to constant. Thus in many applications right congruences
can be replaced by semaphore codes. Except for Section 4 on semaphore codes, all the material in [9]
up to and including Section 7 is restricted to finite codes and automata. Finally, in [9, Section 8], a
natural random walk on any (finite or infinite) semaphore code was constructed and its stationary
distribution plus hitting time to constant were computed.
In this paper we do the following: first we reveal a main application we have in mind [7] by
introducing the infinite and finite semaphore codes associated to a Turing machine via resets (see
Section 2). Then Sections 3 and 4 take the profinite limits of k-reset graphs yielding (−ω)-reset
graphs.
We consider the pseudovariety D and left infinite words, but by duality we have analogous results
for the pseudovariety K and right infinite words. We need both versions for studying Turing machines.
Generalizing the finite case from [9], we study right congruences and special right congruences in
bijection with infinite semaphore codes and the natural action in Sections 5 and 6 and obtain an
approximation theory as in the finite case of k-resets, but for (−ω)-resets. In the final Section 7, we
make some more remarks on relating Section 2 to Sections 3-6, and the next paper on attacking P
versus NP.
Acknowledgements
We are indebted to Jean-Camille Birget for good advice and patient reading. We thank also Benjamin
Steinberg and Nicolas M. Thiéry for discussions.
The first author thanks the Simons Foundation–Collaboration Grants for Mathematicians for
travel grant #313548. The second author was partially supported by NSF grants OCI–1147247 and
DMS–1500050. The third author was partially supported by CMUP (UID/MAT/00144/2013), which
is funded by FCT (Portugal) with national (MEC) and European structural funds (FEDER), under
the partnership agreement PT2020.
2
Resets and Turing machines
In this section we present a new viewpoint on Turing machines centered in the concept of resets and
their associated semaphore codes.
2.1
Turing machines
For details on Turing machines, the reader is referred to [4].
Let us assume that T = (Q, A, Γ, q0 , F, δ) is a (deterministic) Turing machine, where:
• Q is the (finite) state set;
• A is the (finite) input alphabet;
• Γ is the (finite) tape alphabet (containing A and the blank symbol B);
• q0 ∈ Q is the initial state;
• F ⊆ Q is the set of final states;
• δ : Q × Γ → Q × (Γ \ {B}) × {L, R} is the (partial) transition function.
2
Then we write Ω = Γ ∪ (Γ × Q). To make notation lighter, we shall denote (X, q) ∈ Γ × Q by X q .
To avoid confusion with powers of X, we stipulate that from now on symbols such as q, q ′ , qi will be
reserved to denote states and never integers.
We define two homomorphisms tape : Ω∗ → Γ∗ and heads : Ω∗ → (N, +) by
tape(X) = tape(X q ) = X,
heads(X) = 0,
heads(X q ) = 1
for all X ∈ Γ and q ∈ Q. Now we define the set of all legal words by
Leg(T ) = B ∗ {w ∈ Ω∗ | tape(w) ∈ {1, B}(Γ \ {B})∗ {1, B}, heads(w) ≤ 1}B ∗ .
Note that Leg(T ) is closed under reversal and factors. The illegal words are the elements of the
complement Ω∗ \ Leg(T ), which is the ideal having as generating set all the words of the form
• X1q uX2q
′
• Y BnY ′
• Y B mB q B k Y ′
• Y q BnY ′
• Y ′ B nY q
• B q B nY
• Y BnBq
where X1 , X2 ∈ Γ; Y, Y ′ ∈ Γ \ {B}; u ∈ Γ∗ ; n ≥ 1; m, k ≥ 0.
Note that legal words do not correspond necessarily to the possible content of the tape during a
computation (or a factor of that content), but they contain such words as particular cases.
We define the one-move mapping β : Leg(T ) → Leg(T ) as follows. Given w ∈ Leg(T ), then β(w)
is intended to be obtained from w by performing one single move of T on a tape with content w; if
T admits no such move from w (in particular, if heads(w) = 0), we set β(w) = w. In all cases except
(A) and (B) below, the interpretation of β(w) is clear and |β(w)| = |w|. The following two cases
deserve extra clarification:
(A) if w = X q w′ and δ(q, X) = (. . . , . . . , L);
(B) if w = w′ X q and δ(q, X) = (. . . , . . . , R).
In these cases, we interpret β(w) as β(Bw) (respectively β(wB)), falling into the general case. We
say that legal words of types (A) and (B) are β-singular. Note that |β(w)| = |w| + 1 in the β-singular
case. Note that, by padding the input sequences with sufficiently many B’s before and after, cases
(A) and (B) never occur.
For every n ≥ 0, we denote by β n the n-fold composition if β. We define also a partial mapping
β (n) : Ω∗ × Ω × Ω∗ → Ω
as follows.
3
Let u, v ∈ Ω∗ and X ∈ Ω. If uXv ∈ Leg(T ), then β (n) (u, X, v) is the symbol replacing X at
the designated position in the tape after applying β n times. If uXv ∈
/ Leg(T ), then β (n) (u, X, v) is
undefined.
We say that T is legal-halting if, for every u ∈ Leg(T ), the sequence (β n (u))n is eventually
constant. This implies that the sequence (β (n) (u, X, v))n is also eventually constant for all u, v ∈ Ω∗
and X ∈ Ω such that uXv ∈ Leg(T ). We write
β ω (u) = lim β n (u),
n→∞
β (ω) (u, X, v) = lim β (n) (u, X, v).
n→∞
Note that, from a formal viewpoint, a Turing machine which halts for every input is not necessarily
legal-halting (since not every legal word arises from an input configuration), but it can be made
legal-halting with minimum adaptations.
2.2
Resets
We say that r ∈ Ω∗ is a right reset if
β (n) (t1 rt2 , X, t3 ) = β (n) (t′1 rt2 , X, t3 )
for all t1 , t′1 , t2 , t3 ∈ Ω∗ , X ∈ Ω and n ≥ 0 such that both t1 rt2 Xt3 , t′1 rt2 Xt3 ∈ Leg(T ). Let RRes(T )
denote the set of all right resets of T .
Dually, r ∈ Ω∗ is a left reset if
β (n) (t1 , X, t2 rt3 ) = β (n) (t1 , X, t2 rt′3 )
for all t1 , t2 , t3 , t′3 ∈ Ω∗ , X ∈ Ω and n ≥ 0 such that both t1 Xt2 rt3 , t1 Xt2 rt′3 ∈ Leg(T ). Let LRes(T )
denote the set of all left resets of T .
Lemma 2.1 RRes(T ) and LRes(T ) are ideals of Ω∗ containing all the illegal words.
Proof. We prove the claim for right resets.
If r is illegal then t1 rt2 Xt3 is always illegal and so r ∈ RRes(T ) trivially. In particular, RRes(T )
is nonempty.
Let r ∈ RRes(T ) and let x, y ∈ Ω∗ . Suppose that xry ∈
/ RRes(T ). Then there exist t1 , t′1 , t2 , t3 ∈
∗
′
Ω , X ∈ Ω and n ≥ 0 such that both t1 xryt2 Xt3 and t1 xryt2 Xt3 are legal and
β (n) (t1 (xry)t2 , X, t3 ) 6= β (n) (t′1 (xry)t2 , X, t3 ).
Rewriting this inequality as
β (n) ((t1 x)r(yt2 ), X, t3 ) 6= β (n) ((t′1 x)r(yt2 ), X, t3 ),
we deduce that r ∈
/ RRes(T ), a contradiction. Thus xry ∈ RRes(T ) and so RRes(T ) is an ideal of
Ω∗ .
Lemma 2.2 If u ∈ Ω∗ \ B ∗ , then Bu ∈ RRes(T ) and uB ∈ LRes(T ).
Proof. It is easy to see that Bu is a right reset since the only legal words of the form tBut′ must
arise from t ∈ B ∗ . Similarly, uB is a left reset.
4
Lemma 2.3
(i) β(RRes(T )) ⊆ RRes(T ).
(ii) β(LRes(T )) ⊆ LRes(T ).
Proof. We prove the claim for right resets.
Let r ∈ RRes(T ). We may assume that r is legal, X q occurs in r and δ(X, q) is defined. Let
t1 , t′1 , t2 , t3 ∈ Ω∗ , X ∈ Ω and n ≥ 0 be such that both t1 β(r)t2 Xt3 and t′1 β(r)t2 Xt3 are legal.
Suppose first that r is not β-singular. Then t1 rt2 Xt3 and t′1 rt2 Xt3 are also legal, and
β (n) (t1 β(r)t2 , X, t3 ) = β (n+1) (t1 rt2 , X, t3 ) = β (n+1) (t′1 rt2 , X, t3 ) = β (n) (t′1 β(r)t2 , X, t3 ).
Thus we may assume that r is β-singular. Suppose first that r = X q r ′ and δ(q, X) = (p, Y, L).
Then β(r) = B p Y r ′ . Since t1 B p Y r ′ t2 Xt3 and t′1 B p Y r ′ t2 Xt3 are legal, it is easy to check that
t1 Brt2 Xt3 and t′1 Brt2 Xt3 are legal as well. Hence
β (n) (t1 β(r)t2 , X, t3 ) = β (n+1) (t1 Brt2 , X, t3 ) = β (n+1) (t′1 Brt2 , X, t3 ) = β (n) (t′1 β(r)t2 , X, t3 ).
Finally, suppose that r = r ′ X q and δ(q, X) = (p, Y, R). Then β(r) = r ′ Y B p . Since t1 r ′ Y B p t2 Xt3
and t′1 r ′ Y B p t2 Xt3 are legal, it is easy to check that t1 rBt2 Xt3 and t′1 rBt2 Xt3 are legal as well.
Hence
β (n) (t1 β(r)t2 , X, t3 ) = β (n+1) (t1 rBt2 , X, t3 ) = β (n+1) (t′1 rBt2 , X, t3 ) = β (n) (t′1 β(r)t2 , X, t3 ).
Therefore β(r) is a right reset in any case.
2.3
Semaphore codes and the output function
Given an alphabet X, we define the suffix order on X ∗ by
u ≤s v if v ∈ X ∗ u.
Dually, we define the prefix order ≤p .
We say that S ⊆ X ∗ is a (right) semaphore code if S is a suffix code (i.e. an antichain for the
suffix order) and SX ⊆ X ∗ S. By [9, Proposition 4.3], S is a semaphore code if and only if S is the
set of ≤s -minimal elements in some ideal I E X ∗ , denoted by Iβℓ .
Dually, S ⊆ X ∗ is a left semaphore code if S is a prefix code (i.e. an antichain for the prefix
order) and XS ⊆ SX ∗ . Then S is a left semaphore code if and only if S is the set of ≤p -minimal
elements in some ideal of X ∗ .
We describe now the semaphore code RSC(T ) defined by the right resets. It consists of all
minimal right resets for the suffix order. If 1 ∈
/ RRes(T ) (a trivial case), then RSC(T ) consists of
all right resets Xz with X ∈ Ω and z ∈ Ω∗ such that z ∈
/ RRes(T ). In particular, z must be a legal
word.
Similarly, the left semaphore code LSC(T ) consists of all minimal left resets for the prefix order.
If 1 ∈
/ LRes(T ), then LSC(T ) consists of all left resets zX with z ∈ Ω∗ and X ∈ Ω such that
z∈
/ LRes(T ). In particular, z must be a legal word.
Note also that every right reset contains some s ∈ RSC(T ) as a suffix. Dually, every left reset
contains some s ∈ LSC(T ) as a prefix.
From now on, we assume that T is legal-halting. The output function ϕT is the restriction of the
partial function β (ω) : Ω∗ × Ω × Ω∗ to (RSC(T ) ∪ {1}) × Ω × (LSC(T ) ∪ {1}).
5
Proposition 2.4 Let T be a legal-halting Turing machine. Then β ω is fully determined by the output
function ϕT .
Proof. Clearly, β ω is fully determined by β (ω) . Let u, v ∈ Ω∗ and X ∈ Ω. If uXv is illegal,
then β (ω) (u, X, v) is undefined, hence we may assume that uXv ∈ Leg(T ). It follows that also
BuXvB ∈ Leg(T ) and β (ω) (Bu, X, vB) = β (ω) (u, X, v).
If u ∈ B ∗ , then β (ω) (Bu, X, vB) = β (ω) (1, X, vB). If u ∈
/ B ∗ , then Bu ∈ RRes(T ) by Lemma 2.2
∗
and so Bu = xr for some x ∈ Ω and r ∈ RSC(T ). It follows that β (ω) (Bu, X, vB) = β (ω) (r, X, vB),
so in any case we have
β (ω) (Bu, X, vB) = β (ω) (r, X, vB)
(2.1)
for some r ∈ RSC(T ) ∪ {1}.
Now if v ∈ B ∗ , then β (ω) (r, X, vB) = β (ω) (r, X, 1) = ϕT (r, X, 1), hence we may assume that
v ∈
/ B ∗ . Then vB ∈ LRes(T ) by Lemma 2.2 and so vB = r ′ x′ for some r ′ ∈ LSC(T ) and x ∈
∗
Ω . It follows that β (ω) (r, X, vB) = β (ω) (r, X, r ′ ) = ϕT (r, X, r ′ ), so in view of (2.1) we have that
β (ω) (u, X, v) = β (ω) (Bu, X, vB) is determined by ϕT . Therefore β ω is determined by ϕT .
2.4
Length restrictions
For every ℓ ≥ 0, we define the cofinite ideals
RResℓ (T ) = RRes(T ) ∪ Ωℓ Ω∗ ,
LResℓ (T ) = LRes(T ) ∪ Ωℓ Ω∗ .
Note that
RRes(T ) =
\
RResℓ (T ),
LRes(T ) =
ℓ≥0
\
LResℓ (T ).
(2.2)
ℓ≥0
Since
β(Ωℓ ) ⊆ Ωℓ ∪ Ωℓ+1 ,
it follows from Lemma 2.3 that:
Lemma 2.5
(i) β(RResℓ (T )) ⊆ RResℓ (T ).
(ii) β(LResℓ (T )) ⊆ LResℓ (T ).
The semaphore code RSCℓ (T ) consists of all minimal elements of RResℓ (T ) for the suffix order.
Equivalently,
RSCℓ (T ) = (RSC(T ) ∩ Ω≤ℓ ) ∪ Ω(Ωℓ−1 \ RRes(T )).
(2.3)
Dually, the left semaphore code LSCℓ (T ) consists of all minimal elements of LResℓ (T ) for the prefix
order, or equivalently
LSCℓ (T ) = (LSC(T ) ∩ Ω≤ℓ ) ∪ (Ωℓ−1 \ LRes(T ))Ω.
Therefore RSCℓ (T ) ∪ LSCℓ (T ) ⊆ Ω≤ℓ .
6
(2.4)
2.5
2.5.1
A context-free example
Description
We present now a very elementary example just to illustrate the notation and ideas. Further research
will include much more complicated examples.
Let A = {a, b} and L = {an bn | n ≥ 1}, one of the classical examples of a (deterministic) contextfree language which is not rational. The language L is accepted by the Turing machine T depicted
by
Y |Y R
q5 o
B|ZL
q6
Y |Y R
a|aR
Y |Y R
qO0 `❅
❅❅
a|XR
/ q1
Y |Y R
/ q2
⑦
❅❅
⑦⑦
❅❅
⑦⑦
⑦
❅❅
⑦
X|XR
b|Y L ⑦⑦
❅
⑦
X|XR ❅❅
b|Y L
❅❅
⑦⑦
❅❅ ⑦⑦⑦
~⑦
q4 Z o
q3 Z
a|aL
Y |Y L
a|aL
where q0 is the initial state and q6 the unique final state.
In state q0 we can only read a or Y . In the first case, a is replaced by X and we change to state
q1 . Then we move right across other possible a’s until we reach the first b and replace it by Y to go
to state q3 . If we have done this routine before, we may have to move across older Y ’s – taking us
into state q2 . From state q3 , we intend to move left until we reach X, which means going through
Y ’s and then a’s (if there are some left – state q4 ). So we are back at state q0 and we repeat the
procedure. If we have replaced all the a’s, we are supposed to read Y at state q0 , then move to the
right end of the tape (state q5 ) reading only Y ’s. If we have succeeded on reaching the blank B, then
we accept the input moving to the final state q6 .
It is easy to check that T is legal-halting. Indeed, B can be read at most once, and any long
enough sequence of transitions must necessarily involve replacing a by X or b by Y , which are both
irreversible changes.
We use the notation introduced in Section 2.1 for an arbitrary Turing machine.
2.5.2
Resets
We claim that
S
S
Ω∗ \ RRes(T ) = a∗ {b, Y }∗ ∪ ( i=1,3,4 a∗ aqi a∗ {b, Y }∗ ) ∪ ( i=1,2 a∗ Y ∗ bqi {b, Y }∗ )
S
∪ ( i=1,2,3 a∗ Y ∗ Y qi {b, Y }∗ ).
Let m, n ≥ 0 and u ∈ {b, Y }∗ containing n b’s. Then
β (ω) (aq0 an · am u, b, 1) = Y 6= b = β (ω) (am u, b, 1),
hence am u ∈
/ RRes(T ).
7
(2.5)
Assume now that i ∈ {1, 3, 4}, m, n, k ≥ 0 and u ∈ {b, Y }∗ contains k b’s. Then
β (ω) (Xak+1 · am aqi an u · b, b, 1) = Y 6= b = β (ω) (am aqi an u · b, b, 1),
hence am aqi an u ∈
/ RRes(T ).
Assume next that i ∈ {1, 2}, m, n, k ≥ 0 and u ∈ {b, Y }∗ contains k b’s. Then
β (ω) (Xak+1 · am Y n bqi u, b, 1) = Y 6= b = β (ω) (am Y n bqi u, b, 1),
hence am Y n bqi u ∈
/ RRes(T ).
Finally, assume that i ∈ {1, 2, 3}, m, n, k ≥ 0 and u ∈ {b, Y }∗ contains k b’s. Then
β (ω) (Xak+2 · am Y n Y qi u · b, b, 1) = Y 6= b = β (ω) (am Y n Y qi u · b, b, 1),
hence am Y n Y qi u ∈
/ RRes(T ).
To prove the converse, we must prove that any other word is necessarily a right reset. We make
extensive use from RRes(T ) being an ideal of Ω∗ .
Consider first
r ∈ {B, X, Z} ∪ {B q , X q , Z q | q ∈ Q}.
Suppose that t1 , t′1 , t2 , t3 ∈ Ω∗ , P ∈ Ω are such that both t1 rt2 P t3 , t′1 rt2 P t3 ∈ Leg(T ). Let n ≥ 0. If
β (n) (t1 rt2 , P, t3 ) 6= P , then it is easy to see that t1 ∈ Γ∗ and has no influence in the computation.
Thus β (n) (t1 rt2 , P, t3 ) = β (n) (t′1 rt2 , P, t3 ) and so r ∈ RRes(T ).
Thus we only need to discuss words w ∈ Ω∗ such that tape(w) ∈ {a, b, Y }∗ . We consider next
the word ba. Consider t1 bat2 P t3 ∈ Leg(T ). If t2 P t3 ∈
/ Γ∗ , then t1 is irrelevant to the computation
of β (n) (t1 bat2 , P, t3 ). If t2 P t3 ∈ Γ∗ , then β (n) (t1 bat2 , P, t3 ) = P necessarily. It follows that ba ∈
RRes(T ).
Similarly, Y a, bq a, baq , Y q a, Y aq ∈ RRes(T ) for every q ∈ Q. Hence we have reduced the problem
to words w ∈ Ω∗ such that tape(w) ∈ a∗ {b, Y }∗ .
Since the transition function is undefined for these pairs, we have
{aq2 , aq5 , aq6 , bq0 , bq3 , bq4 , bq5 , bq6 , Y q4 , Y q6 } ⊆ RRes(T ).
Also aq0 ∈ RRes(T ) because T moves to the right and will never get to the left of the new X.
Similarly, Y q0 , Y q5 ∈ RRes(T ). To complete the proof of (2.5), it suffices to show that
bY ∗ bqi ∪ bY ∗ Y qj ⊆ RRes(T )
(2.6)
for i = 1, 2 and j = 1, 2, 3, because any word not containing such a factor has already been established
to be or not to be a right reset.
Indeed, in neither case the head of T can pass to the left of the first b, so (2.6) and therefore
(2.5) hold as claimed.
Similarly, we compute the left resets, in fact we obtain
LRes(T ) = RRes(T ).
8
(2.7)
2.5.3
Semaphore codes
In view of (2.5), it is straightforward to check that
RSC(T ) = (Ω \ {a, aq1 , aq3 , aq4 })a+ {b, Y }∗
∪ (Ω \ {a, b, Y, aq1 , aq3 , aq4 , bq1 , bq2 , Y q1 , Y q2 , Y q3 }){b, Y }∗
S
S
∪ ( i=1,3,4 (Ω \ {a})a∗ aqi a∗ {b, Y }∗ ) ∪ ( i=1,2 (Ω \ {a})a+ Y ∗ bqi {b, Y }∗ )
S
S
∪ ( i=1,2 (Ω \ {a, Y })Y ∗ bqi {b, Y }∗ ) ∪ ( i=1,2,3 (Ω \ {a})a+ Y ∗ Y qi {b, Y }∗ )
S
∪ ( i=1,2,3 (Ω \ {a, Y })Y ∗ Y qi {b, Y }∗ ).
To compute the intersection RSC(T ) ∩ Leg(T ), we replace the 5 last occurrences of Ω by Γ.
Similarly,
LSC(T ) = a∗ (Ω \ {a, b, Y, aq1 , aq3 , aq4 , bq1 , bq2 , Y q1 , Y q2 , Y q3 })
∪ a∗ Y + (Ω \ {b, Y, bq1 , bq2 , Y q1 , Y q2 , Y q3 }) ∪ a∗ Y ∗ b{b, Y }∗ (Ω \ {b, Y })
S
S
∪ ( i=1,3,4 a∗ aqi a∗ (Ω \ {a, b, Y })) ∪ ( i=1,3,4 a∗ aqi a∗ {b, Y }+ (Ω \ {b, Y }))
S
S
∪ ( i=1,2 a∗ Y ∗ bqi {b, Y }∗ (Ω \ {b, Y })) ∪ ( i=1,2,3 a∗ Y ∗ Y qi {b, Y }∗ (Ω \ {b, Y })).
To compute the intersection LSC(T ) ∩ Leg(T ), we replace the 4 last occurrences of Ω by Γ.
2.5.4
Semaphore codes modulo ℓ
In view of (2.3) and (2.4), we can easily compute easily RSCℓ (T ) and LSCℓ (T ) making use of the
computations performed in Sections 2.5.2 and 2.5.3.
3
Free pro-D semigroups
For general background on free pro-D semigroups, see [8, Sections 3.1 and 3.2].
Let A be a finite nonempty alphabet. We denote by A−ω the set of all left infinite words on A,
that is, infinite sequences of the form · · · a3 a2 a1 with ai ∈ A. If u ∈ A+ , we denote the left infinite
word · · · uuu by u−ω .
The free semigroup A+ acts on the right of A−ω by concatenation: given x = · · · a3 a2 a1 ∈ A−ω
and u = a′1 · · · a′n ∈ A+ , we define
xu = · · · a3 a2 a1 a′1 · · · a′n ∈ A−ω .
Given x ∈ A+ ∪ A−ω and y ∈ A−ω , we define also xy = y. Together with concatenation on A+ , this
defines a semigroup structure for A+ ∪ A−ω .
9
The suffix (ultra)metric on A+ ∪ A−ω is defined as follows. Given x, y ∈ A+ ∪ A−ω , let lcs(x, y)
be the longest common suffix of x and y, and define
(
2−|lcs(x,y)| if x 6= y,
d(x, y) =
0
otherwise.
Given x0 ∈ A+ ∪ A−ω and δ > 0, we write
Bδ (x0 ) = {x ∈ A+ ∪ A−ω | d(x, x0 ) < δ}
for the open ball of radius δ around x0 .
If S ∈ D is endowed with the discrete topology and ϕ : A → S is a mapping, then there exists a
unique continuous homomorphism Φ : A+ ∪ A−ω → S such that the diagram
ϕ
A _
✉
✉
✉
✉
/
✉: S
Φ
A+ ∪ A−ω
commutes. This characterizes (A+ ∪ A−ω , d) as the free pro-D semigroup on A. We shall denote it
by ΩA (D). It is well known that ΩA (D) is a complete and compact topological semigroup.
We remark that for a general pseudovariety V, the metric considered for free pro-V semigroups
is the profinite metric, but in the particular case of D we can use this alternative metric that equates
to the normal form.
4
(−ω)-reset graphs
We consider now A-graphs with possibly infinite vertex sets. For general concepts in automata
theory, the reader is referred to [1].
A left infinite path in an A-graph Γ = (Q, E) is an infinite sequence of the form
a
a
a
3
2
1
· · · −→q
3 −→q2 −→q1
such that (qi+1 , ai , qi ) ∈ E and qi ∈ Q for every i ≥ 1. Its label is the left infinite word · · · a3 a2 a1 ∈
A−ω . We write
x
· · · −→q
to denote a left infinite path with label x ending at q.
An A-graph Γ = (Q, E) is:
• deterministic if (p, a, q), (p, a, q ′ ) ∈ E implies q = q ′ ;
• complete if for all p ∈ Q and a ∈ A there exists some edge (p, a, q) ∈ E;
u
• strongly connected if, for all p, q ∈ Q, there exists a path p−→q in Γ for some u ∈ A∗ ;
• (−ω)-deterministic if
x
x
· · · −→q, · · · −→q ′ paths in Γ ⇒ q = q ′
holds for all q, q ′ ∈ Q and x ∈ A−ω ;
10
• (−ω)-complete if every x ∈ A−ω labels some left infinite path in Γ;
• (−ω)-trim if every q ∈ Q occurs in some left infinite path in Γ;
• a (−ω)-reset graph if it is (−ω)-deterministic, (−ω)-complete and (−ω)-trim.
We denote by RG(A) the class of all (−ω)-reset A-graphs.
If Γ = (Q, E) ∈ RG(A), then Q induces a partition
[
A−ω =
A−ω
q ,
q∈Q
x
where A−ω
denotes the set of all x ∈ A−ω labelling some path · · · −→q in Γ. Moreover, A−ω
q
q 6= ∅ for
every q ∈ Q.
Proposition 4.1 Let Γ ∈ RG(A). Then Γ is deterministic and complete.
Proof. Write Γ = (Q, E) and suppose that (p, a, q), (p, a, q ′ ) ∈ E. Since Γ is (−ω)-trim, there exists
x
xa
some left infinite path · · · −→p for some x ∈ A−ω . Hence there exist left infinite paths · · · −→q and
xa ′
· · · −→q , and since Γ is (−ω)-deterministic, we get q = q ′ . Therefore Γ is deterministic.
x
Let p ∈ Q and a ∈ A. Since Γ is (−ω)-trim, there exists some left infinite path · · · −→p for some
xa
x ∈ A−ω . Now xa ∈ A−ω and Γ being (−ω)-complete implies that there exists some path · · · −→q
in Γ, which we may factor as
a
x
· · · −→q ′ −→q.
Since Γ is (−ω)-deterministic, we get q ′ = p, hence (p, a, q) ∈ E and Γ is complete.
We recall now the preorder ≤ introduced in [9, Section 3]. Given A-graphs Γ, Γ′ , we write Γ ≤ Γ′
if there exists a morphism Γ → Γ′ .
Lemma 4.2 Let A be a finite nonempty alphabet and let Γ, Γ′ ∈ RG(A) with Γ ≤ Γ′ ≤ Γ. Then
Γ∼
= Γ′ .
Proof. Let ϕ : Γ → Γ′ and ψ : Γ′ → Γ be morphisms. Write Γ = (Q, E) and Γ′ = (Q′ , E ′ ). It is
easy to see that
−ω
−ω
A−ω
A−ω
q ⊆ Aqϕ ,
q ′ ⊆ Aq ′ ψ
−ω 6= ∅. Since Γ
⊆ A−ω
for all q ∈ Q and q ′ ∈ Q′ . Hence A−ω
q
qϕψ . Since Γ is (−ω)-trim, we have Aq
′
′
is (−ω)-deterministic, we get q = qϕψ. Similarly, q = q ψϕ, hence ϕ and ψ are mutually inverse
bijections and therefore mutually inverse A-graph isomorphisms.
Let [Γ] denote the isomorphism class of Γ. Similarly to [9, Section 3],
[Γ] ≤ [Γ′ ]
if Γ ≤ Γ′
defines a preorder on RG(A)/ ∼
=. Moreover, Lemma 4.2 yields:
Corollary 4.3 Let A be a finite nonempty alphabet. Then ≤ is a partial order on RG(A)/ ∼
=.
11
5
Right congruences on A−ω
Since xy = y for all x ∈ ΩA (D) and y ∈ A−ω , it follows that A−ω is the minimum ideal of ΩA (D).
Following the notation introduced in [9, Section 2.2], we denote by RC(A−ω ) the lattice of right
congruences on A−ω (with respect to the right action of ΩA (D)).
We say that ρ ∈ RC(A−ω ) is closed if ρ is a closed subset of A−ω × A−ω for the product metric
d′ ((x, y), (x′ , y ′ )) = max{d(x, x′ ), d(y, y ′ )},
where d denotes the suffix metric on A−ω . Given x0 , y0 ∈ A+ ∪ A−ω and δ > 0, we write
Bδ ((x0 , y0 )) = {(x, y) ∈ (A+ ∪ A−ω )2 | d′ ((x, y), (x0 , y0 )) < δ}.
By [8, Exercise 3.1.7], this implies that xρ is a closed subset of A−ω for every x ∈ A−ω . The next
example shows that the converse fails.
Example 5.1 Let A = {a, b} and let
w = . . . a4 ba3 ba2 bab.
For all x, y ∈ A−ω , let
xρy
if
(5.1)
x = wu, y = wv with |u| = |v|
or
x = y.
Then ρ ∈ RC(A−ω ) and xρ is closed for every x ∈ A−ω , but ρ is not closed.
Indeed, it is easy to see that, given x ∈ A−ω , there is at most one word u ∈ A∗ such that
x = wu. We call this a w-factorization of x. Hence ρ is transitive and it follows immediately that
ρ ∈ RC(A−ω ). The uniqueness of the w-factorization implies also that xρ is finite (hence closed) for
every x ∈ A−ω . However,
lim (wan , wbn ) = (a−ω , b−ω ) ∈
/ ρ.
n→∞
Since
(wan , wbn )
∈ ρ for every n ≥ 1, then ρ is not closed.
We denote by CRC(A−ω ) (respectively ORC(A−ω )) the set of all closed (respectively open) right
congruences on A−ω .
We consider CRC(A−ω ) (partially) ordered by inclusion. Similarly to [9, Section 5], we can relate
CRC(A−ω ) with RG(A).
Given ρ ∈ RC(A−ω ), the Cayley graph Cay(ρ) is the A-graph Cay(ρ) = (A−ω /ρ, E) defined by
E = {(uρ, a, (ua)ρ) | u ∈ A−ω , a ∈ A}.
Lemma 5.2 Let ρ ∈ RC(A−ω ).
x
(i) For every x ∈ A−ω , there exists a left infinite path · · · −→xρ in Cay(ρ).
(ii) Cay(ρ) is (−ω)-complete and (−ω)-trim.
Proof. (i) Write x = · · · a3 a2 a1 with ai ∈ A. For every n ≥ 1, write xn = · · · an+2 an+1 an . Then
a
a
a
3
2
1
· · · −→x
3 ρ−→x2 ρ−→x1 ρ = xρ
is a left infinite path in Cay(ρ) labeled by x.
(ii) By part (i).
12
Lemma 5.3 Let ρ ∈ CRC(A−ω ). Then:
x
(i) If · · · −→q is a left infinite path in Cay(ρ), then q = xρ.
(ii) Cay(ρ) ∈ RG(A).
Proof. (i) Assume that q = yρ with y ∈ A−ω . Write x = · · · a3 a2 a1 with ai ∈ A. For every n ≥ 1,
un
let un = an · · · a1 . Then there exists some path yn ρ−→yρ
in Cay(ρ) for some yn ∈ A−ω . Hence
yn un ∈ yρ. Since
x = lim un = lim yn un
n→∞
n→∞
and ρ closed implies yρ closed, we get x ∈ yρ, hence xρ = yρ = q.
(ii) By part (i), Cay(ρ) is (−ω)-deterministic. By Lemma 5.2(ii), Cay(ρ) is both (−ω)-complete
and (−ω)-trim, therefore Cay(ρ) ∈ RG(A).
We discuss next open right congruences, relating them in particular with the right congruences
on Ak . Given x ∈ A−ω , let xξk denote the suffix of length k of x. For σ ∈ RC(Ak ), let σ
b be the
relation on A−ω defined by
xb
σ y if (xξk )σ(yξk ).
It is immediate that σ
b ∈ RC(A−ω ).
On the other hand, given ρ ∈ RC(A−ω ) and k ≥ 1, we define a relation ρ(k) on Ak by
uρ(k) v
if (A−ω u × A−ω v) ∩ ρ 6= ∅.
We denote by ρ[k] the transitive closure of ρ(k) .
The next example shows that ρ(k) needs not to be transitive, even in the closed case.
Example 5.4 Let A = {a, b} and let w be given by (5.1). For all x, y ∈ A−ω , let
{x, y} = {wa2 u, wbau} for some u ∈ A∗
or
{x, y} = {wb2 av, wb3 v} for some v ∈ A∗
xρy if
or
x = y.
Then ρ ∈ CRC(A−ω ) but ρ(2) is not transitive.
Indeed, by the uniqueness of the w-factorization remarked in Example 5.1, ρ turns out to be
transitive and therefore a right congruence.
We sketch the proof that ρ is closed. Let (x, y) ∈ (A−ω × A−ω ) \ ρ. Then x 6= y. Write
u = lcs(x, y). We consider several cases:
Case I: {x, y} = {zb2 au, z ′ b3 u}.
Then either z 6= w or z ′ 6= w. We may assume that z 6= w. Let k ≥ 1 be such that w ∈
/ B2−k (z). It
is easy to see that B2−k−3−|u| ((x, y)) ∩ ρ = ∅.
Case II: {x, y} = {za2 u, z ′ bau}.
Then either z 6= w or z ′ 6= w. We may assume that z 6= w. Let k ≥ 1 be such that w ∈
/ B2−k (z). It
is easy to see that B2−k−2−|u| ((x, y)) ∩ ρ = ∅.
Case III: all the remaining cases.
13
It is easy to see that B2−3−|u| ((x, y)) ∩ ρ = ∅.
Therefore ρ is closed.
Now (wa2 , wba) ∈ ρ yields (a2 , ba) ∈ ρ(2) , and (wb2 a, wb3 ) ∈ ρ yields (ba, b2 ) ∈ ρ(2) , However,
(a2 , b2 ) ∈
/ ρ(2) , hence ρ(2) is not transitive.
The following lemma compiles some elementary properties of ρ(k) and ρ[k] . The proof is left to
the reader.
Lemma 5.5 Let A be a finite nonempty alphabet, ρ ∈ RC(A−ω ) and k ≥ 1. Then:
(i) ρ(k) ∈ RC(Ak ) if and only if ρ(k) is transitive;
(ii) ρ[k] ∈ RC(Ak );
\
d
[n] .
(iii) ρ ⊆
ρ
n≥1
We discuss next some alternative characterizations for open right congruences. We recall the
definition of k-reset graph from [9, Section 3].
We say that u ∈ A∗ is a reset word for a deterministic and complete A-graph Γ = (Q, E) if
|Qu| = 1. This is equivalent to say that all paths labeled by u end at the same vertex. Let Res(Γ)
denote the set of all reset words for Γ.
We say that Γ is a k-reset graph if Ak ⊆ Res(Γ). We denote by RGk (A) the class of all strongly
connected deterministic complete k-reset A-graphs.
Proposition 5.6 Let A be a finite nonempty alphabet and ρ ∈ RC(A−ω ). Then the following conditions are equivalent:
(i) ρ is open;
(ii) xρ is an open subset of A−ω for every x ∈ A−ω ;
(iii) ρ = σ
b for some σ ∈ RC(Ak ) and k ≥ 1;
(k) ;
(iv) there exists some k ≥ 1 such that ρ(k) is transitive and ρ = ρd
(v) Cay(ρ) ∈ RGk (A) for some k ≥ 1;
(vi) ρ is closed and has finite index.
Proof. (i) ⇒ (ii). Let x ∈ A−ω . Since (x, x) ∈ ρ, there exists some δ > 0 such that Bδ ((x, x)) ⊆ ρ.
Since Bδ ((x, x)) = Bδ (x) × Bδ (x), we get Bδ (x) ⊆ xρ and so xρ is open.
(ii) ⇒ (vi). Let x, y ∈ A−ω be such that (x, y) ∈
/ ρ. Since xρ and yρ are open, there exists some
δ > 0 such that Bδ (x) ⊆ xρ and Bδ (y) ⊆ yρ. If x′ ∈ xρ and y ′ ∈ yρ, then (x, y) ∈
/ ρ yields (x′ , y ′ ) ∈
/ ρ.
Hence
(Bδ (x) × Bδ (y)) ∩ ρ = ∅,
and so Bδ ((x, y)) ∩ ρ = ∅. Thus the complement of ρ is open and so ρ is closed.
On the other hand, {xρ | x ∈ A−ω } is an open cover of A−ω and so admits a finite subcover since
A−ω is compact. Therefore ρ has finite index.
(vi) ⇒ (v). By Lemma 5.3(ii), we have Cay(ρ) ∈ RG(A). Hence Cay(ρ) is deterministic and
complete by Proposition 4.1.
14
z
Let x, y ∈ A−ω . Since Cay(ρ) is (−ω)-trim, there exists a left infinite path · · · −→yρ in Cay(ρ).
Since ρ has finite index, we may factor this path as
z′
u
v
· · · −→wρ−→wρ−→yρ
with u 6= ε. On the other hand, since Cay(ρ) is complete and ρ has finite index, there exist m ≥ 0
and p ≥ 1 such that there exists a path
um
up
xρ−→x′ ρ−→x′ ρ
in Cay(ρ). It follows that there exist two paths
u−ω
· · · −→wρ,
u−ω
· · · −→x′ ρ
and so wρ = x′ ρ since Cay(ρ) is (−ω)-deterministic. Thus there exists a path
um
v
xρ−→wρ−→yρ
and so Cay(ρ) is strongly connected.
Suppose now that Cay(ρ) ∈
/ RGk (A) for every k ≥ 1. Let P denote the set of pairs of distinct
vertices in Cay(ρ). Then
u
u
k
k
∀k ≥ 1 ∃uk ∈ Ak ∃(p, q) ∈ P ∃ paths · · · −→p,
· · · −→q
in Cay(ρ).
Since P is finite, one of the pairs (p, q) must repeat infinitely often. Hence there exists some (p, q) ∈ P
such that
uk
uk
∀k ≥ 1 ∃uk ∈ A≥k ∃ paths · · · −→p,
· · · −→q
in Cay(ρ).
Since ΩA (D) is compact, we may replace (uk )k by some convergent subsequence. Let x = limk→∞ uk .
Since (|uk |)k is unbounded, we have x ∈ A−ω .
Write p = xp ρ with xp ∈ A−ω . Since Cay(ρ) is (−ω)-trim, there exists some left infinite path
yk uk
yk uk
−ω . By Lemma 5.2(i), there exists a path · · · −−−→(y
· · · −−−→x
p ρ for some yk ∈ A
k uk )ρ in Cay(ρ).
Since Cay(ρ) is (−ω)-deterministic, we get (yk uk )ρ = xp ρ, hence yk uk ∈ xp ρ. Since ρ closed implies
xp ρ closed and
x = lim uk = lim yk uk ,
k→∞
k→∞
x
we get x ∈ xp ρ. By Lemma 5.2(i), there exists a path · · · −→xρ = xp ρ = p in Cay(ρ). Similarly, there
x
exists some path · · · −→q. Since p 6= q, this contradicts Cay(ρ) being (−ω)-deterministic. Therefore
Cay(ρ) ∈ RGk (A) for some k ≥ 1.
(v) ⇒ (iv). Assume that Cay(ρ) ∈ RGk (A) for some k ≥ 1. We show that
xξk = yξk ⇒ xρy
holds for all x, y ∈ A−ω . Indeed, by Lemma 5.2(i), there exists left infinite paths
x
· · · −→xρ,
y
· · · −→yρ
in Cay(ρ). Since xξk = yξk ∈ Ak ⊆ Res(Cay(ρ)), we get xρ = yρ and so (5.2) holds.
15
(5.2)
Suppose now that u, v, w ∈ Ak are such that uρ(k) vρ(k) w. Then there exist some x, y, y ′ , z ∈ A−ω
such that (xu)ρ(yv) and (y ′ v)ρ(zw). Then (yv)ξk = v = (y ′ v)ξk and (5.2) yields (yv)ρ(y ′ v). Thus
(xu)ρ(zw) by transitivity and so uρ(k) w. Therefore ρ(k) is transitive.
d
(k) is well defined and ρ ⊆ ρ
(k) .
Now it follows from Lemma 5.5 that ρd
(k) . Then (xξ , yξ ) ∈ ρ(k) and so there exist x′ , y ′ ∈ A−ω such that
Conversely, let (x, y) ∈ ρd
k
k
′
′
(x (xξk ), y (yξk )) ∈ ρ. Since (x′ (xξk ))ξk = xξk , it follows from (5.2) that (x′ (xξk ))ρx. Similarly,
(k) ⊆ ρ as required.
(y ′ (yξk ))ρy and we get xρy by transitivity. Therefore ρd
(iv) ⇒ (iii). In view of Lemma 5.5(i).
(iii) ⇒ (i). Let (x, y) ∈ ρ = σ
b and let (x′ , y ′ ) ∈ B2−k ((x, y)). Then x′ ξk = xξk and y ′ ξk = yξk .
Hence
xρy ⇒ (xξk )σ(yξk ) ⇒ (x′ ξk )σ(y ′ ξk ) ⇒ x′ ρy ′
and so B2−k ((x, y)) ⊆ ρ. Therefore ρ is open.
The following example shows that closed is required in condition (vi).
Example 5.7 Let A = {a, b} and let ρ be the relation on A−ω defined by xρy if b occurs in both x, y
or in none of them. Then ρ is a right congruence of index 2 on A−ω but it is not closed.
Indeed, it is immediate that ρ is a right congruence of index 2. Since a−ω = limn→∞ b−ω an , ρ is
not closed.
We say that ρ ∈ RC(A−ω ) is profinite if ρ is an intersection of open right congruences. Since
open right congruences are closed by Proposition 5.6, it follows that every profinite right congruence,
being the intersection of closed sets, is itself closed. We denote by PRC(A−ω ) the set of all profinite
right congruences on A−ω .
(k)
Given a graph Γ = (Q, E) and k ≥ 1, we define a relation µΓ on Q by
(k)
pµΓ q
u
u
if there exist paths · · · −→p, · · · −→q in Γ for some u ∈ Ak .
(k)
[k]
[k]
Let µΓ denote the reflexive and transitive closure of µΓ . Then µΓ is an equivalence relation on Q.
Proposition 5.8 Let A be a finite nonempty alphabet and ρ ∈ RC(A−ω ). Then the following conditions are equivalent:
(i) ρ is profinite;
(ii) ρ is an intersection of countably many open congruences;
\
[k] ;
ρc
(iii) ρ =
k≥1
(iv)
\
[k]
µCay(ρ) = id.
k≥1
Proof. (i) ⇒ (iii). Assume that ρ = ∩i∈I τi with τi ∈ ORC(A−ω ) for every i ∈ I.
[k] by Lemma 5.5(iii). To prove the opposite inclusion, we show that
We have ρ ⊆ ∩k≥1 ρc
[k] ⊆ τ .
∀i ∈ I ∃k ≥ 1 ρc
i
16
(5.3)
Indeed, it follows from Proposition 5.6 that there exist some k ≥ 1 and σi ∈ RC(Ak ) such that
τi = σbi . We claim that
(k)
τ i ⊆ σi .
(5.4)
(k)
Assume that (u, v) ∈ τi . Then there exist x, y ∈ A−ω such that (xu, yv) ∈ τi = σbi . Hence
(u, v) = ((xu)ξk , (yv)ξk ) ∈ σi
and (5.4) holds.
(k)
Since ρ ⊆ τi implies ρ(k) ⊆ τi , it follows that ρ(k) ⊆ σi and so ρ[k] ⊆ σi since σi is transitive.
Thus
[k] ⊆ σ
ρc
bi = τi
and (5.3) holds.
Therefore
\
k≥1
[k] ⊆ ∩
ρc
i∈I τi = ρ
as required.
[k] is open by Proposition
(iii) ⇒ (ii). By Lemma 5.5(ii), ρ[k] ∈ RC(Ak ) for every k ≥ 1, hence ρc
5.6 and we are done.
(ii) ⇒ (i). Trivial.
[k]
(k)
(iii) ⇒ (iv). Write µ(k) = µCay(ρ) and µ[k] = µCay(ρ) . By Lemma 5.5(ii), ρ[k] ∈ RC(Ak ) for every
[k] is open (and therefore closed) by Proposition 5.6. Therefore ρ is closed and so
k ≥ 1, hence ρc
Cay(ρ) ∈ RG(A) by Lemma 5.3(ii).
Let x, y ∈ A−ω be such that xρ 6= yρ. Suppose that (xρ, yρ) ∈ µ[k] . Then there exist z0 , . . . , zn ∈
−ω
A
such that z0 = x, zn = y and (zi−1 ρ, zi ρ) ∈ µ(k) for i = 1, . . . , n. For i = 1, . . . , n, there exist
paths
ui
ui
′
zi−1
ρ−→z
zi′′ ρ−→z
i−1 ρ,
iρ
′ , z ′′ ∈ A−ω .
in Cay(ρ) for some ui ∈ Ak and zi−1
i
′
Hence zi−1 ρ = (zi−1 ui )ρ and zi ρ = (zi′′ ui )ρ, yielding
(zi−1 ξk ) ρ(k) ui ρ(k) (zi ξk )
[k] . Since xρ 6= yρ
and so (zi−1 ξk )ρ[k] (zi ξk ). Now (xξk )ρ[k] (yξk ) follows by transitivity, hence (x, y) ∈ ρc
d
implies (x, y) ∈
/ ρ[m] for some m ≥ 1 by condition (iii), it follows that (xρ, yρ) ∈
/ µ[m] and so (iv)
holds.
c
[k] . Conversely, let (x, y) ∈ ∩
[k]
(iv) ⇒ (iii). By Lemma 5.5(iii), we have ρ ⊆ ∩k≥1 ρc
k≥1 ρ . For
each k, we have (xξk , yξk ) ∈ ρ[k] , hence there exist u0 , . . . , un ∈ Ak such that u0 = xξk , un =
yξk and (ui−1 , ui ) ∈ ρ(k) for i = 1, . . . , n. For i = 1, . . . , n, there exist zi−1 , zi′ ∈ A−ω such that
(zi−1 ui−1 , zi′ ui ) ∈ ρ. Write also x = z0′ u0 and y = zn un .
By Lemma 5.2(i), there exist paths
z ′ ui
i
′
· · · −−→(z
i ui )ρ,
zu
i i
· · · −−→(z
i ui )ρ
in Cay(ρ) for i = 0, . . . , n, hence
((zi−1 ui−1 )ρ, (zi ui )ρ) = ((zi′ ui )ρ, (zi ui )ρ) ∈ µ(k) .
17
Thus
((z0 u0 )ρ, (zn un )ρ) ∈ µ[k].
Since ((z0′ u0 )ρ, (z0 u0 )ρ) ∈ µ(k) , we get
(xρ, yρ) = ((z0′ u0 )ρ, (zn un )ρ) ∈ µ[k].
[k] ⊆ ρ as required.
Since k is arbitrary, it follows from condition (iv) that xρ = yρ, hence ∩k≥1 ρc
Every open right congruence on A−ω is trivially profinite and we remarked before that every
profinite right congruence is necessarily closed. Hence
ORC(A−ω ) ⊆ PRC(A−ω ) ⊆ CRC(A−ω ).
We show next that these inclusions are strict if |A| > 1.
For every k ≥ 1, let ρk be the relation on A−ω defined by
xρk y
if xξk = yξk .
It is easy to check that ρk ∈ ORC(A−ω ) for every k ≥ 1. Since ∩k≥1 ρk = id, it follows that the
identity congruence is profinite, while it is clearly not open.
To construct a closed non profinite right congruence is much harder. We do it through the
following example.
Example 5.9 Let A = {a, b}. Given u, v ∈ Ak , write u < v if u = u′ aw and v = v ′ bw for some
(k)
(k)
w ∈ A∗ . Let u1 < · · · < u2k be the elements of Ak , totally ordered by <. Let p1 < p2 < · · · be the
prime natural numbers. For every n ∈ N, let
n
n
n
wn = . . . ap3 bap2 bap1 b.
Let ρ ∈ RC(A−ω ) be generated by the relation
(k)
(k)
R = {(wpi ui , wpi ui+1 ) | k ≥ 1, 1 ≤ i < 2k } ∪ {(b−ω a, a−ω b)}.
k
k
Then ρ is closed but not profinite.
We start by showing that
wpi A∗ ∩ wpi′ A∗ 6= ∅ implies (k = k′ and i = i′ )
k
(5.5)
k′
for all k, k ′ , i, i′ ≥ 1. Indeed, suppose that wpi u = wpi′ v for some u, v ∈ A∗ . By definition of
k
wn , wpi u has only finitely many factors of the form
k
pi
′
pi ′
k′
ba2m b,
pi
and the leftmost must be ba2 k b. Since
′
wpi u = wpi′ v, we get ba2 k b = ba2 k b and so pik = pik′ . Therefore k = k′ and i = i′ , and (5.5) holds.
k
k′
Write
R′ = {(xu, yu) | (x, y) ∈ R ∪ R−1 , u ∈ A∗ }.
Let x ∈ A−ω . We show that
there exists at most one y ∈ A−ω such that (x, y) ∈ R′ .
18
(5.6)
This is obvious if x ∈ b−ω aA∗ ∪ a−ω bA∗ , hence we may assume that x ∈ wpi A∗ for some k ≥ 1 and
k
1 ≤ i < 2k . In view of (5.5), we must have
(k)
(k)
{x, y} = {wpi ui v, wpi ui+1 v}
k
k
(k)
(k)
/ wpi A+ , also ui , ui+1 and v are
for some v ∈ A∗ , and k, i are uniquely determined. Since wpi ∈
k
k
uniquely determined. Thus (5.6) holds.
Suppose that x R′ y R′ z with x 6= y 6= z. Since R′ is symmetric, (5.6) yields x′ = z ′ . It follows
that R′ ∪ id is an equivalence relation, indeed the smallest right congruence containing R. It follows
that
R′ ∪ id = ρ.
Moreover, each ρ-class contains at most two elements.
We prove now that ρ is closed. Let (x, y) ∈ (A−ω × A−ω ) \ ρ. Then x 6= y, hence we may assume
without loss of generality that x = x′ av and y = y ′ bv with v ∈ A∗ . Let
m = max{i ≥ 0 | bi <s x′ , ai <s y ′ }.
Note that the above set is bounded, otherwise x′ = b−ω and y ′ = a−ω , yielding
(x, y) = (b−ω av, a−ω bv) ∈ ρ,
a contradiction. Write x′ = x′′ bm and y ′ = y ′′ am .
For j = 0, . . . , |v|, write v = vj vj′ with |vj | = j. Then am bvj is the successor of bm avj in the
ordering of Am+1+j , hence we may write
(m+1+j)
bm avj = uij
,
(m+1+j)
am bvj = uij +1
for some 1 ≤ ij < 2m+1+j . It follows that
(m+1+j) ′
vj ,
x = x′′ uij
(m+1+j) ′
vj
y = y ′′ uij +1
(5.7)
Let
mj = min{|lcs(x′′ , w
i
j
pm+1+j
)|, |lcs(y ′′ , w
i
j
pm+1+j
Note that mj is a well-defined natural number, otherwise x′′ = w
)|}.
i
j
pm+1+j
(m+1+j) ′
(m+1+j) ′
uij
vj , w ij
u
vj )
ij
pm+1+j
pm+1+j ij +1
(x, y) = (w
= y ′′ and
∈ ρ,
a contradiction.
Let
p = max{m0 , . . . , m|v| } + m + 1 + |v|.
We show that
B2−p ((x, y)) ∩ ρ = ∅.
19
(5.8)
Suppose that (z1 , z2 ) ∈ B2−p ((x, y)) ∩ ρ. Since p > 1 + |v|, we have av <s z1 and bv <s z2 . By
maximality of m, and since p > m + 1 + |v|, we have either abm av <s z1 or bam bv <s z2 . Hence we
must have
(k)
(k)
z1 = wpi ui v ′ , z2 = wpi ui+1 v ′
(5.9)
k
v′ ,
k
|v ′ |
for some
k ≥ 1 and 1 ≤ i <
Clearly,
≤ |v|, hence we must have v ′ = vj′ for j = |v| − |v ′ |.
′′
m
′
m
′
We have x = x b avj vj , hence b avj vj <s z1 . Similarly, y = y ′′ am bvj vj′ yields am bvj vj′ <s z2 .
Suppose that k < m + 1 + j. Since |lcs(x, z1 )| > m + 1 + |v|, it follows from (5.9) that wpi ends
k
with a b. Similarly, |lcs(y, z2 )| > m + 1 + |v| implies that wpi ends with an a, a contradiction. Hence
k
k ≥ m + 1 + j.
Suppose now that k > m + 1 + j. By maximality of m, we must have one of the following cases:
2k .
(k)
• abm avj ≤s ui
(k)
• bm avj ≤s ui
(k)
and am bvj ≤s ui+1 ;
(k)
and bam bvj ≤s ui+1 .
(k)
(k)
Any of these cases contradicts ui+1 being the successor of ui
m + 1 + j and we may write
z1 = wpi
m+1+j
(m+1+j) ′
vj ,
ui
z2 = wpi
m+1+j
for the ordering of Ak , hence k =
(m+1+j) ′
vj .
ui+1
′
′
Since d(z1 , x) < 2−mj −m−1−|v | = 2−mj −m−1−j−|vj | , it follows from (5.7) that i = ij and
|lcs(x′′ , w
i
)| > mj .
|lcs(y ′′ , w
i
)| > mj ,
j
pm+1+j
Similarly,
j
pm+1+j
contradicting the definition of mj .
Thus (5.8) holds and so (A−ω × A−ω ) \ ρ is open. Therefore ρ is closed.
(k)
(k)
(k) (k)
Let k ≥ 1. Since (wpi ui , wpi ui+1 ) ∈ ρ, we have (ui , ui+1 ) ∈ ρ(k) for every 1 ≤ i < 2k . Since
k
k
(k)
(k)
[k] . Since k is arbitrary,
u1 = ak and u2k = bk , it follows that (ak , bk ) ∈ ρ[k] and so (a−ω , b−ω ) ∈ ρc
we get
\
[k] .
ρc
(a−ω , b−ω ) ∈
k≥1
However,
(a−ω , b−ω ) ∈
/ R′ ∪ id = ρ,
[k] and so ρ is not profinite by Proposition 5.8.
hence ρ 6= ∩k≥1 ρc
6
Special right congruences on A−ω
To avoid trivial cases, we assume throughout this section that A is a finite alphabet containing at
least two elements.
Given P ⊆ A∗ , we define a relation τP on A−ω by:
xτP y if x = y or x, y ∈ A−ω u for some u ∈ P.
Lemma 6.1 Let P ⊆ A∗ . Then τP is an equivalence relation on A−ω .
20
Proof. It is immediate that τP is reflexive and symmetric. For transitivity, we may assume that
x, y, z ∈ A−ω are distinct and x τP y τP z. Then there exist u, v ∈ P such that u <s x, y and v <s y, z.
Since u and v are both suffixes of y, one of them is a suffix of the other. Hence either u <s x, z or
v <s x, z. Therefore τP is transitive.
If we consider left ideals, being a right congruence turns out to be a special case:
Proposition 6.2 Let L Eℓ A∗ . Then the following conditions are equivalent:
(i) τL ∈ RC(A−ω );
(ii) τL ∈ PRC(A−ω );
(iii) L E A∗ ;
(iv) (Lβℓ )A ⊆ A∗ (Lβℓ ).
(v) Lβℓ is a semaphore code.
Proof. (i) ⇒ (iv). We may assume that |A| > 1. Let u ∈ L and a ∈ A. Take b ∈ A \ {a}. Then
(a−ω u, b−ω u) ∈ τL and by (i) we get (a−ω ua, b−ω ua) ∈ τL . It follows that ua has some suffix in L,
hence LA ⊆ A∗ L = L and so
(Lβℓ )A ⊆ LA ⊆ L = A∗ (Lβℓ ).
(iv) ⇒ (iii). We have
LA = A∗ (Lβℓ )A ⊆ A∗ (Lβℓ ) = L.
It follows that LA∗ ⊆ L. Since L Eℓ A∗ , we get L E A∗ .
(iii) ⇒ (ii). By Lemma 6.1, τL is an equivalence relation. Let x, y ∈ A−ω be such that xτL y. We
may assume that there exists some u ∈ L such that u <s x, y. Since L E A∗ , we have ua ∈ L and
ua <s xa, ya yields (xa, ya) ∈ τL . Therefore τL ∈ RC(A−ω ).
Let (x, y) ∈ (A−ω × A−ω ) \ τL . Then x 6= y. Let u = lcs(x, y) and let m = |u| + 1. We claim that
[m]
(xξm )τL
= {xξm }.
(m)
(6.1)
Indeed, suppose that (xξm , v) ∈ τL and v 6= xξm . Then there exist z, z ′ ∈ A−ω such that
(z(xξm ), z ′ v) ∈ τL . Since v 6= xξm , then z(xξm ) and z ′ v must have a common suffix w ∈ L,
(m)
and |w| < m. But then w ≤s u, yielding u ∈ L and xτL y, a contradiction. Thus (xξm , v) ∈ τL
implies v = xξm , and so (6.1) holds.
d
[m]
[m]
Suppose that (x, y) ∈ τL . Then (xξm , yξm ) ∈ τL , hence xξm = yξm by (6.1), contradicting
d
d
d
[m]
[m]
[m]
m > |u|. Thus (x, y) ∈
/ τL and so ∩k≥1 τL ⊆ τL . Hence τL = ∩k≥1 τL by Lemma 5.5(iii), and so
τL is profinite by Proposition 5.8.
(ii) ⇒ (i). Trivial.
(iv) ⇔ (v). By Lemma [9, Lemma 4.1], since Lβℓ is always a suffix code.
21
We say that ρ ∈ RC(A−ω ) is a special right congruence on A−ω if ρ = τI for some I E A∗ . In
view of Proposition 6.2, this is equivalent to say that ρ = τS for some semaphore code S on A. We
denote by SRC(A−ω ) the set of all special right congruences on A−ω .
The next result characterizes the open special right congruences. Recall that a suffix code S ⊂ A∗
is said to be maximal if S ∪ {u} fails to be a suffix code for every u ∈ A∗ \ S.
Proposition 6.3 Let I E A∗ . Then the following conditions are equivalent:
(i) τI ∈ ORC(A−ω );
(ii) Iβℓ is a finite maximal suffix code;
(iii) A∗ \ I is finite.
Proof. (i) ⇒ (ii). Let u, v ∈ Iβℓ be distinct. Then
((A−ω u) × (A−ω v)) ∩ τI = ∅.
Since τI has finite index by Proposition 5.6, it follows that the suffix code Iβℓ is finite.
Suppose now that Iβℓ ∪ {u} is a suffix code for some u ∈ A∗ \ (Iβℓ ). It is easy to see that no
two elements of A−ω u are τI equivalent, a contradiction since τI has finite index. Therefore Iβℓ is a
maximal suffix code.
(ii) ⇒ (iii). Let m denote the maximum length of the words in Iβℓ . Suppose that v ∈ A∗ \ I
has length > m. It is straightforward to check that Iβℓ ∪ {v} is a suffix code, contradicting the
maximality of Iβℓ . Thus A∗ \ I ⊆ A≤m and is therefore finite.
(iii) ⇒ (i). We have τI ∈ RC(A−ω ) by Proposition 6.2.
Let m ≥ 1 be such that A∗ \ I ⊆ A≤m . Then
xξm+1 = yξm+1 ⇒ xτI y
holds for all x, y ∈ A−ω and so τI has finite index. Since τI is profinite (and therefore closed) by
Proposition 6.2, it follows from Proposition 5.6 that τI is open.
The proof of [9, Lemma 7.4] can be adapted to show that inclusion among left ideals determines
inclusion for the equivalence relations τL :
Lemma 6.4 Let |A| > 1 and L, L′ Eℓ A∗ . Then
τ L ⊆ τ L′ ⇔ L ⊆ L ′ .
Note that Lemma 6.4 does not hold for |A| = 1, since |A−ω | = 1.
Similarly, we adapt [9, Proposition 7.6]:
Proposition 6.5 Let |A| > 1. Then:
(i) τI∩J = τI ∩ τJ and τI∪J = τI ∪ τJ for all I, J E A∗ ;
(ii) SRC(A−ω ) is a full sublattice of RC(A−ω );
(iii) the mapping
I(A) → SRC(A−ω )
I 7→ τI
is a lattice isomorphism.
22
Given ρ ∈ RC(A−ω ) and C ∈ A−ω /ρ, we say that C is nonsingular if |C| > 1. If C is nonsingular,
we denote by lcs(C) the longest common suffix of all words in C. We define
• Λρ = {lcs(C) | C ∈ A−ω /ρ is nonsingular},
• Λ′ρ = {lcs(x, y) | (x, y) ∈ ρ, x 6= y}.
Lemma 6.6 Let ρ ∈ RC(A−ω ). Then:
(i) A∗ Λρ = A∗ Λ′ρ ;
(ii) Λ′ρ Er A∗ ;
(iii) A∗ Λ′ρ E A∗ .
Proof. (i) Let C ∈ A−ω /ρ be nonsingular and let w = lcs(C). By maximality of w there exist
a, b ∈ A distinct and x, y ∈ A−ω such that xaw, ybw ∈ C. Thus w = lcs(xaw, ybw) and so
Λρ ⊆ Λ′ρ .
(6.2)
Therefore A∗ Λρ ⊆ A∗ Λ′ρ .
Conversely, let (x, y) ∈ ρ with x 6= y. Then xρ is nonsingular and lcs(xρ) is a suffix of lcs(x, y),
hence Λ′ρ ⊆ A∗ Λρ and so A∗ Λρ = A∗ Λ′ρ .
(ii) Let u ∈ Λ′ρ and a ∈ A. Then u = lcs(x, y) for some (x, y) ∈ ρ with x 6= y. Then (xa, ya) ∈ ρ.
Since lcs(xa, ya) = ua, we get ua ∈ Λ′ρ . Therefore Λ′ρ Er A∗ .
(iii) Clearly, A∗ Λ′ρ Eℓ A∗ . Now we use part (ii).
Given ρ ∈ RC(A−ω ), we write
Res(ρ) = Res(Cay(ρ)).
We refer to the elements of Res(ρ) as the resets of ρ.
Lemma 6.7 Let ρ ∈ RC(A−ω ). Then:
(i) Res(ρ) E A∗ ;
(ii) if ρ is closed, then
Res(ρ) = {w ∈ A∗ | (xw, yw) ∈ ρ for all x, y ∈ A−ω }.
Proof. (i) Immediate.
(ii) Let w ∈ Res(ρ) and x, y ∈ A−ω . By Lemma 5.2(i), there exist paths
yw
xw
· · · −→(xw)ρ,
· · · −→(yw)ρ
in Cay(ρ). Now w ∈ Res(ρ) yields (xw)ρ = (yw)ρ.
w
w
Now let w ∈ A∗ \ Res(ρ). Then there exist paths p−→q and p′ −→q ′ in Cay(ρ) with q 6= q ′ . Since
Cay(ρ) is (−ω)-trim by Lemma 5.2(ii), there exist left infinite paths
in Cay(ρ), hence paths
y
· · · −→p,
x
· · · −→p′
xw
· · · −→q ′ .
· · · −→q,
yw
Since ρ is closed, it follows from Lemma 5.3(i) that (xw)ρ = q 6= q ′ = (yw)ρ and we are done.
23
Adapting the proof of [9, Proposition 7.9], we obtain:
Proposition 6.8 Let |A| > 1, ρ ∈ RC(A−ω ) and I E A∗ . Then:
(i) ρ ⊆ τI ⇔ Λρ ⊆ I ⇔ Λ′ρ ⊆ I;
(ii) if ρ is closed, then τI ⊆ ρ ⇔ I ⊆ Res(ρ).
Given R ⊆ A−ω × A−ω , we denote by R♯ the right congruence on A−ω generated by R, i.e. the
intersection of all right congruences on A−ω containing R.
Given ρ ∈ RC(A−ω ), we denote by NS(ρ) the set of all nonsingular ρ-classes.
We can now prove several equivalent characterizations of special right congruences.
Proposition 6.9 Let |A| > 1 and ρ ∈ RC(A−ω ). Then the following conditions are equivalent:
(i) ρ ∈ SRC(A−ω );
(ii) lcs : NS(ρ) → A∗ is injective, Λρ is a suffix code and
∀x ∈ A−ω ∀w ∈ Λρ (xw)ρ ∈ NS(ρ);
(6.3)
(iii) ρ = τA∗ Λρ ;
(iv) ρ = τA∗ Λ′ρ ;
(v) ρ = τL♯ for some L Eℓ A∗ .
Proof. (i) ⇒ (ii). By a straightforward adaptation of the proof of (i) ⇒ (ii) in [9, Proposition 7.10],
we check that lcs : NS(ρ) → A∗ is injective and Λρ is a suffix code.
Now let x ∈ A−ω and w ∈ Λρ . Then w = lcs(yρ) for some yρ ∈ NS(ρ). By (6.2), we may
write w = lcs(y ′ , y ′′ ) for some y ′ , y ′′ ∈ yρ distinct. Since ρ = τI , it follows that w ∈ I, hence
(xw, y), (xw, y ′ ) ∈ τI = ρ. Since y ′ 6= y ′′ , it follows that either xw 6= y ′ or xw 6= y ′′ , so in any case
(xw)ρ ∈ NS(ρ) as required.
(ii) ⇒ (iii). Write I = A∗ Λρ . If (x, y) ∈ ρ and x 6= y, then lcs(xρ) ∈ Λρ ⊆ I is a suffix of both x
and y, hence (x, y) ∈ τI . Thus ρ ⊆ τI .
Conversely, let (x, y) ∈ τI . We may assume that x 6= y, hence there exists some w ∈ Λρ such that
w <s x, y. Hence (6.3) yields xρ, yρ ∈ NS(ρ).
Suppose that lcs(xρ) 6= w. Then lcs(xρ) <s w or w <s lcs(xρ), contradicting Λρ being a suffix
code. Hence lcs(xρ) = w. Similarly, lcs(yρ) = w. Since lcs : NS(ρ) → A∗ is injective, we get xρ = yρ.
Thus ρ = τI .
(iii) ⇔ (iv). By Lemma 6.6(i).
(iii) ⇒ (v). Write L = A∗ Λρ . By (iii), we have τL♯ = ρ♯ = ρ. Since L E A∗ by Lemma 6.6, (iv)
holds.
(v) ⇒ (i). Let I = LA∗ E A∗ . Since L ⊆ I, it follows from Lemma 6.4 that τL ⊆ τI , hence
ρ = τL♯ ⊆ τI♯ = τI
by Proposition 6.2.
Conversely, let (x, y) ∈ τI . We may assume that x 6= y. Then there exist factorizations x = x′ w
and y = y ′ w with w ∈ I. Write w = zw′ with z ∈ L. Then (x′ z, y ′ z) ∈ τL and so
(x, y) = (x′ w, z ′ w) = (x′ zw′ , y ′ zw′ ) ∈ τL♯ = ρ.
Thus τI ⊆ ρ as required.
24
Proposition 6.10 Let |A| > 1 and ρ ∈ CRC(A−ω ). Then the following conditions are equivalent:
(i) ρ ∈ SRC(A−ω );
(ii) lcs : NS(ρ) → A∗ is injective, Λρ is a suffix code and
∀x ∈ A−ω ∀w ∈ Λρ (xw)ρ ∈ NS(ρ);
(iii) ρ = τA∗ Λρ ;
(iv) ρ = τA∗ Λ′ρ ;
(v) ρ = τL♯ for some L Eℓ A∗ ;
(vi) ρ = τRes(ρ) ;
(vii) Λρ ⊆ Res(ρ);
(viii) Λ′ρ ⊆ Res(ρ);
(ix) whenever
aw
p−→q,
bw
p′ −→q,
w
p′′ −→r
(6.4)
are paths in Cay(ρ) with a, b ∈ A distinct, then q = r.
Proof. (i) ⇔ (ii) ⇔ (iii) ⇔ (iv) ⇔ (v). By Proposition 6.9.
(i) ⇒ (vi). If ρ = τI for some I E A∗ , then I ⊆ Res(ρ) by Proposition 6.8(ii). Since Res(ρ) E A∗
by Lemma 6.7(i), then Proposition 6.8(ii) also yields
τRes(ρ) ⊆ ρ = τI ,
hence Res(ρ) ⊆ I by Lemma 6.4. Therefore I = Res(ρ).
(vi) ⇒ (vii) ⇔ (viii). By Lemma 6.7(i), Res(ρ) E A∗ . Now we apply Proposition 6.8(i).
(viii) ⇒ (i). We have A∗ Λ′ρ , Res(ρ)EA∗ by Lemmas 6.6(iii) and 6.7(i). It follows from Proposition
6.8 that
τRes(ρ) ⊆ ρ ⊆ τA∗ Λ′ρ .
Since Λ′ρ ⊆ Res(ρ) yields A∗ Λ′ρ ⊆ Res(ρ) and therefore τA∗ Λ′ρ ⊆ τRes(ρ) by Lemma 6.4, we get
ρ = τRes(ρ) ∈ SRC(A−ω ).
(viii) ⇒ (ix). Consider the paths in (6.4). By Lemma 5.2(ii), there exist left infinite paths
x′
x
· · · −→p′
· · · −→p,
in Cay(ρ), hence (xaw, x′ bw) ∈ ρ by Lemma 5.3(i) and so
w = lcs(xaw, x′ bw) ∈ Λ′ρ ⊆ Res(ρ).
Thus q = r as required.
(ix) ⇒ (viii). Let w ∈ Λ′ρ . Then w = lcs(x, y) for some (x, y) ∈ ρ such that x 6= y. We may write
x = x′ aw and y = y ′ bw with a, b ∈ A distinct. By Lemma 5.2(i), there exist in Cay(ρ) paths of the
form
y′
bw
x′
aw
· · · −→p−→xρ, · · · −→p′ −→yρ = xρ.
Now (ix) implies that w ∈ Res(ρ).
25
We can now prove that not all open right congruences are special, even for |A| = 2:
Example 6.11 Let A = {a, b} and let σ be the equivalence relation on A3 defined by the following
partition:
{a3 , aba, ba2 } ∪ {bab, a2 b} ∪ {ab2 } ∪ {b2 a} ∪ {b3 }.
Then σ
b ∈ ORC(A3 ) \ SRC(A3 ).
Indeed, it is routine to check that σ ∈ RC(A3 ), hence ρ = σ
b ∈ ORC(A−ω ) by Proposition 5.6.
−ω
−ω
2
Since lcs(a ρ) = a and lcs((b a)ρ) = b a, then Λρ is not a suffix code and so ρ ∈
/ SRC(A−ω ) by
Proposition 6.9.
Let ρ ∈ RC(A−ω ) and let
ρ = ∨{τ ∈ SRC(A−ω ) | τ ⊆ ρ},
ρ = ∧{τ ∈ SRC(A−ω ) | τ ⊇ ρ}.
By Proposition 6.5(ii), we have ρ, ρ ∈ SRC(A−ω ).
Proposition 6.12 Let |A| > 1 and ρ ∈ CRC(A−ω ). Then:
(i) ρ = τRes(ρ) ;
(ii) ρ = τA∗ Λρ = τA∗ Λ′ρ .
Proof. (i) By Lemma 6.7(i), we have Res(ρ) E A∗ . Now the claim follows from Proposition 6.8(ii).
(ii) Similarly, we have A∗ Λρ = A∗ Λ′ρ E A∗ by Lemma 6.6(iii), and the claim follows from Proposition 6.8(i).
The straightforward adaptation of [9, Example 7.14] shows that the pair (ρ, ρ) does not univocally
determine ρ ∈ RC(A−ω ), even in the open case:
Example 6.13 Let A = {a, b} and let σ, σ ′ be the equivalence relations on A3 defined by the following
partitions:
{a3 , aba, ba2 } ∪ {bab, a2 b} ∪ {ab2 } ∪ {b2 a} ∪ {b3 },
{a3 , b2 a, ba2 } ∪ {bab, a2 b} ∪ {ab2 } ∪ {aba} ∪ {b3 }.
Let ρ = σ
b and ρ′ = σb′ . Then ρ, ρ′ ∈ ORC(A−ω ), ρ = ρ′ and ρ = ρ′ .
This same example shows also that ρ does not necessarily equal or cover ρ in SRC(A−ω ). Indeed,
in this case we have
Res(ρ) = A∗ A3 ∪ {a2 , ab} ⊂ I ⊂ A+ \ {b, b2 } = A∗ Λρ
for I = A∗ A3 ∪ {a2 , ab, ba} E A∗ . By Lemma 6.4, we get
ρ ⊂ τI ⊂ ρ.
26
7
Conclusion and future work
We enter now into random walks on infinite semigroups. The most sophisticated approach is described
in [3]. We use profinite limits (see [8]) as an alternative approach, as developed in Sections 3-6.
Indeed, if I1 , I2 , . . . is a sequence of ideals in A∗ with I = ∩k≥1 Ik , let Jk βℓ the semaphore
code determined by the ideal Jk = I1 ∩ . . . ∩ Ik . Whenever k ≥ m, we may define a mapping
ϕkm : Jk βℓ → Jm βℓ by setting uϕkm to be the unique suffix of u in Jm βℓ . It is routine to check that:
• ϕkm is onto;
• ϕkm preserves the action of A∗ on the right;
• (ϕkm ) constitutes a projective system of surjective morphisms with respect to this action;
• Iβℓ is the projective limit of this system.
In view of (2.2), each Turing machine T provides an instance of this setting when Ik = RResk (T )
and I = RRes(T ). Moreover, each ideal RResk (T ) is cofinite and τRRes(T ) is a profinite congruence
on A−ω , indeed the intersection of the open congruences τRResk (T ) .
Using the left-right duals of Sections 3-6, we have similar results for LRes(T ) and the sequence
(LResk (T )).
In a subsequent paper, we intend to characterize polynomial time Turing machines in this framework, including the natural semaphore codes action and the action of β (n) and β (∞) . The approach
will constitute a variation of [10]: we will need to consider certain metrics that will give the same
topology as in Sections 3-6, but conditions involving the metrics will take us from the realm of
topology into that of geometry.
References
[1] J. Berstel, Transductions and context-free languages, Teubner, Stuttgart, 1979. 10
[2] J. Berstel, D. Perrin and C. Reutenauer, Codes and automata, Encyclopedia of Mathematics
and its Applications 129, Cambridge University Press, Cambridge, 2010. 1
[3] G. Hognas and A. Mukherjea, Probability measures on semigroups, Springer Series in Probability
and its Applications, Springer, 2011. 27
[4] J. E. Hopcroft and J. D. Ullman, Introduction to automata theory, languages, and computation,
Addison-Wesley, 1979. 2
[5] K. Krohn and J. Rhodes, Algebraic theory of machines. I. Prime decomposition theorem for
finite semigroups and machines, Trans. Amer. Math. Soc. 116 (1965) 450–464. 1
[6] J. Rhodes, Applications of automata theory and algebra. Via the mathematical theory of complexity to biology, physics, psychology, philosophy, and games, With an editorial preface by
Chrystopher L. Nehaniv and a foreword by Morris W. Hirsch. World Scientific Publishing Co.
Pte. Ltd., Hackensack, NJ, 2010. xviii+274 pp. 1
27
[7] J. Rhodes and P. V. Silva, Turing machines and bimachines, Theoret. Comput. Sci. 400 (2008),
no. 1-3, 182–224. 2
[8] J. Rhodes and B. Steinberg, The q-theory of finite semigroups, Springer Monographs in Mathematics, Springer, 2009. 1, 9, 12, 27
[9] J.Rhodes, A. Schilling and P. V. Silva, Random walks on semaphore codes and delay de Bruijn
semigroups, preprint arXiv:150903383. 1, 2, 5, 11, 12, 14, 21, 22, 24, 26
[10] J. Rhodes and P. Weil, Algebraic and topological theory of languages, RAIRO Theoret. Informatics Appl. 29 (1995), 1–44. 27
28
| 4math.GR
|
BASMAJIAN–TYPE INEQUALITIES
FOR MAXIMAL REPRESENTATIONS
arXiv:1611.00286v2 [math.GT] 14 Dec 2017
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Abstract. For suitable metrics on the locally symmetric space associated to a maximal
representation, we prove inequalities between the length of the boundary and the lengths
of orthogeodesics that generalize the classical Basmajian’s identity from Teichmüller theory.
Any equality characterizes diagonal embeddings.
1. Introduction
In the last few decades, many authors have proved elegant identities over moduli spaces
of hyperbolic surfaces. One of the first identities was proven by Basmajian in [Bas93]. Even
though his work applies more generally to hyperbolic n-manifolds for any n ≥ 2, his most
celebrated result is the following:
Basmajian’s identity ([Bas93]). Let Σ be a hyperbolic surface with nonempty geodesic
2
boundary ∂Σ, and let OH
Σ denote the set of unoriented orthogeodesics of Σ (i.e. the set of
geodesics with endpoints on the boundary and orthogonal to it). Then
`(∂Σ) = 4
X
log coth
2
`(α)
.
2
α∈OH
Σ
Another identity involving orthogeodesics is due to Bridgeman and Kahn [BK10], who
show that the volume of a hyperbolic n-manifold with geodesic boundary can be computed
in terms of the length of orthogeodesics. Vlamis and Yarmola generalized this result to a
larger class of manifolds in [VY17b].
Other well-known identities are McShane’s identity [McS98] and its generalization by
Mirzakhani [Mir07], Luo–Tan’s identity [LT14] and Bridgeman’s identity [Bri11]. Note that,
besides the intrinsic interest of these results, some also have important applications: for instance, Mirzakhani used in [Mir07] the generalization of McShane’s identity to give a recursive
formula for the Weil-Petersson volume of moduli spaces of surfaces with boundary.
Our main goal is to prove inequalities which generalize Basmajian’s identity to the context
of maximal representations and such that equality is attained exactly when the maximal
representation is a diagonal embedding of a hyperbolization. A tool to prove these inequalities
is a result of independent interest: an identity involving cross-ratios (in the sense of Labourie),
which should be thought of as the higher rank analogue of the fact that the limit set of the
fundamental group of a surface with boundary has measure zero.
Date: March 7, 2018.
1
2
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
1.1. Maximal representations as higher Teichmüller theory. Maximal representations
form a class of discrete and injective representations of fundamental groups of surfaces1 into
the symplectic group. They consist of the representations maximizing the Toledo invariant,
a notion of volume defined using bounded cohomology. A fundamental result of Burger,
Iozzi and Wienhard [BIW10, Theorem 8], characterizes them as the representations into
Sp(2n, R) admitting an equivariant “well-behaved” boundary map φ : ∂π1 (Σ) → L(R2n )
from the boundary of the fundamental group of the surface to the Lagrangians of R2n (see
[BIW10] or Section 3 for the precise definitions).
These representations are a generalization of Teichmüller space: it was proven by Goldman
in [Gol80] that Teichmüller space can be interpreted as the parameter space of representations
of the fundamental group of a surface into PSL(2, R) = PSp(2, R) maximizing the Euler
class, of which the Toledo invariant is a higher rank generalization. In particular, maximal
representations into PSp(2, R) are holonomies of hyperbolizations and hence correspond to
the Teichmüller space. The study of common patterns of maximal representations and other
special representations, most notably Hitchin representations and positive representations, is
often referred to as higher Teichmüller theory.
A recent trend in higher Teichmüller theory is to see which results of hyperbolic geometry
can be generalized in the context of representations of surface groups. For instance, a classical
and extremely useful result about hyperbolic surfaces is the collar lemma. First proven by
Keen [Kee74], it states the existence, for a simple closed geodesic γ, of a neighborhood which
is an embedded cylinder of width depending only on `(γ), diverging when `(γ) shrinks to
zero. In particular, one can deduce a lower bound in terms of `(γ) to the length of any simple
closed geodesic intersecting γ. This corollary has been generalized to the Hitchin component
by Lee and Zhang [LZ17] and to maximal representation by Burger and the second author
[BP17].
Also identities in the higher Teichmüller setting have attracted attention in recent years:
Labourie and McShane generalized McShane-Mirzakhani’s identities to arbitrary cross-ratios
and in particular to Hitchin representations in [LM09], while in [VY17a] Vlamis and Yarmola
obtained a generalization of Basmajian’s identity to the Hitchin component.
1.2. Main results. A maximal representation ρ : π1 (Σ) → Sp(2n, R) induces an action
of Γ = π1 (Σ) on the symmetric space X associated to the symplectic group. The action
is properly discontinuous and therefore the locally symmetric space ρ(Γ)\X is a smooth
manifold. As we will describe in Section 2.2, we consider three Sp(2n, R)-invariant distances
on X : the determinant Finsler distance dF , the Riemannian distance dR and a Weyl chamber
+
valued distance dā . The latter has values in a Weyl chamber ā+ , a specific subset of Rn that
+
parametrizes the orbits of Sp(2n, R) on the tangent bundle TX . We denote by `R , `F and `ā
+
the length of paths in X computed using dR , dF and dā respectively. If γ is an element of
the fundamental group, its length (with respect to each metric) will be its translation length:
the infimum over all points X ∈ X of the distance between X and its translate ρ(γ) · X.
The first goal of the paper is to find the suitable generalization of classical orthogeodesics
to the context of maximal representations. In the classical setup, an orthogeodesic between
1We will only consider oriented surfaces with negative Euler characteristic.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
3
two boundary components of a hyperbolic surface lifts in H2 to a geodesic segment orthogonal
to two lifts of the boundary components in the hyperbolic plane. It is pointed out in [BP17]
that a good generalization of geodesics in the setting of maximal representations is given by
the so-called R-tubes, parallel sets of specific singular geodesics (see Section 2.3). In fact, to
each γ ∈ Γ we can naturally associate an R-tube Yγ , which is a ρ(γ)-invariant subspace of
X . In Section 4.2, we observe that given any pair of primitive peripheral elements γ and δ
in Γ, there exists a unique R-tube Yα (sometimes simply denoted by α) which is orthogonal
to both Yγ and Yδ ; moreover, this tube meets each subspace in a point. We call Yα an
orthotube, and its length is defined to be the distance between the intersection points with Yγ
and Yδ . We denote by OΣ the collection of orthotubes and by OΣ (γ) the subset of orthotubes
orthogonal to a fixed Yγ . If α ∈ OΣ (γ), we denote by δα the peripheral element associated
to α and different from γ.
The length of orthotubes has also geometric significance: in Section 3.3 we construct what
we call the holomorphic double of a maximal representation (with suitable hypotheses on the
peripheral elements). The double is a specific representation of the fundamental group of the
double of the surface that restrict to the given representation. Then we can show that the
Finsler length of an orthotube is half the length of the corresponding curve in the double
(Proposition 4.4).
Given v ∈ ā+ , we denote by vn the smallest coordinate of v. Our geometric generalizations
of Basmajian’s identity to the setting of maximal representations are the following:
Theorem A. For any maximal representation ρ : Γ → Sp(2n, R) with the property that the
image of peripheral elements are Shilov hyperbolic, we have
(A1)
2n
X
α∈OΣ
log coth
+
X
`ā (α)n
`F (α)
≥ `F (∂Σ) ≥ 2n
log coth
2
n
α∈OΣ
and
(A2)
+
√ X
√ X
`ā (α)n
`R (α)
4 n
log coth
≥ `R (∂Σ) ≥ 4 n
log coth √
2
2 n
α∈OΣ
α∈OΣ
with equalities if and only if ρ is, up to a character in a compact group, the composition of a
holonomy representation of a hyperbolization into SL(2, R) with the diagonal representation
of SL(2, R) into Sp(2n, R).
Note that adding the condition on the images of peripheral elements is equivalent to requiring that the representation is Anosov (see Section 3). Interestingly, for both metrics
the difference between the middle term and the right term can be arbitrarily large (Proposition 7.7), namely there are sequences of maximal representations in which the length of the
boundary components stay bounded away from zero, but the R-tubes associated to any two
peripheral elements are arbitrarily far apart. This is discussed in Section 7.3.
Remark 1.1. If we drop the hypothesis on the Shilov hyperbolic image, the right-hand side
inequalities in (A1) and (A2) still hold, as long as there is at least one peripheral element
whose image is Shilov hyperbolic. Moreover, the characterization of any equality still holds
under these milder assumptions.
4
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Remark 1.2. A natural question is how this work generalizes to maximal representations ρ
Z
in other Hermitian Lie groups G. It is well known that then the Zariski closure H = ρ(Γ)
of the image of ρ is a Hermitian Lie group of tube type [BIW10, Theorem 5]. Under the mild
assumption that H has no factor isogenous to SL(2, R) or E6(−14) , there exists a (virtual)
tight holomorphic embedding
ι : H → Sp(2n, R)
(for some explicit n depending on H only [BIW09, Example 8.7]) that is isometric, up to
suitably rescaling the metric on the irreducible factors of H. Furthermore the Shilov boundary
of H is naturally a subspace of L(R2n ) by [BIW09, Theorem 7].
The composition ρ0 = ι ◦ ρ : Γ → Sp(2n, R) is then a maximal representation and Theorem
A holds for ρ0 . Note that, if H is irreducible, the Riemannian translation distance for ρ0 is
an explicit multiple of the Riemannian translation distance for the action ρ of Γ on the symmetric space XH associated to H. Since there is a ι-equivariant totally geodesic holomorphic
inclusion XH → X , it is possible to verify that, for every orthotube Yα that we consider,
the endpoints of Yα belong to the Shilov boundary of H. This follows from Lemma 4.2 and
the observation that the totally geodesic holomorphic and tight image of a polydisk in XH
is a partially diagonal subset of a polydisk of X (the rank of XH is in general smaller than
the rank of X ). In turn this implies that intersection Yα ∩ XH is non-empty, consists of the
parallel set of a maximally singular geodesic in XH (an H-tube) and is orthogonal to the
H-tubes in XH associated to the two peripheral elements corresponding to α.
It is furthermore possible to verify that this does not generalize to the Fuchsian locus
within Hitchin maximal representations: this is not a contradiction since the totally geodesic
equivariant map is, in that case, not holomorphic. This also partially justifies why, for
representations in such Fuchsian locus, equality in Theorem A does not hold.
As a corollary of our main result we are able to deduce an interesting geometric property
of the locally symmetric space ρ(Γ)\X : not only the image hρ(γ)i\Yγ in ρ(Γ)\X of Yγ is an
embedded manifold, but it also admits an embedded tubular neighborhood whose width can
be explicitly computed as a function of the translation length of Γ:
Corollary 1.3. If γ ∈ Γ corresponds to a simple closed curve or boundary component of Σ
and ρ : Γ → Sp(2n, R) is an Anosov maximal representation, then hρ(γ)i\Yγ ⊂ ρ(Γ)\X has
an embedded tubular neighbourhood C(γ) of width
R
√
` (γ)
√
w(γ) := n arccoth exp
2 n
with respect to the Riemannian metric. Moreover, if δ corresponds to a simple closed curve
or boundary component that is disjoint from γ, then the neighborhoods C(γ) and C(δ) are
disjoint.
The main difference of our approach with respect to Vlamis and Yarmola’s one is that with
our inequalities we want to relate intrinsic geometric quantities of the locally symmetric space
associated to a representation. Instead, following Labourie–McShane, they mainly work with
algebraic versions of the identities and use cross-ratios to translate Basmajian’s result in the
language of representations. As a partial step towards the proof of Theorem A we prove the
analogue of Vlamis and Yarmola’s result in our context:
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
5
Theorem B. Let ρ : Γ → Sp(2n, R) be a maximal representation with the property that
the image of each peripheral element is Shilov hyperbolic. Then for every peripheral element
γ ∈ Γ we have
X
log B(γ − , δα+ , γ + , δα− ),
`B (γ) =
α∈OΣ (γ)
where B is the R-valued cross-ratio constructed by Labourie in [Lab08].
Our proof of this result is very different from Vlamis and Yarmola’s proof: they build on
the fact that the image of the boundary map associated to a Hitchin representation is C 1 ,
which is in general far from being true in the case of maximal representations. Instead, we
adapt a more geometric proof which could also be used to get estimates on the Hausdorff
dimension of the image of the boundary map.
We conclude this introduction by mentioning that Xu [Xu16] used orthogeodesics to study
the metric completion of the pressure metric (see [BCLS15]) on the Teichmüller space of
surfaces with boundary. He proved that, in that case, the pressure metric is not a constant
multiple of the Weil-Petersson metric. We hope that our study will have application in the
study of the pressure metric on the space of maximal representations as well.
1.3. Plan of the paper. In Section 2 we discuss properties of the geometry of the symmetric
space associated to Sp(2n, R) and of the synthetic geometry of R-tubes. We recall the
results of [BP17] and expand them when needed. In particular we relate, in Section 2.4
the Finsler length and the translation on the Euclidean factor of an R-tube (Lemma 2.17),
we define causal paths in R-tubes (these will arise naturally while dealing with maximal
representations), and give an explicit bound on the length of the projection of a causal path
on the symmetric space for SL(n, R) in terms of its length in the Euclidean factor (Lemma
2.20). Section 3 contains the necessary preliminaries about maximal representations and the
construction of the holomorphic double of a representation (Proposition 3.8). In Section
4 we define our generalization of orthogeodesic in the context of maximal representations
and prove the relation between the length of an orthotube and the length of the associated
element in the double of the representation (Proposition 4.4). Section 5 describes the idea of
the proof of Basmajian’s identity and the difficulties that arise when generalizing it to higher
rank. Section 6 is devoted to the proof of Theorem B, generalizing the classical strategy
described in Section 6.1. In the final section we prove the geometric inequalities announced
in the introduction and the characterization of diagonal representations (Theorem A).
Acknowledgements. The second author would like to thank Ursula Hamenstädt for asking
if the equality in the Collar lemma of [BP17] characterizes the Fuchsian locus – that question gave the initial motivation for this work. We are grateful to Brian Bowditch for useful
conversations and in particular for suggesting a proof of Proposition 6.1 that turned out to
be amenable to generalization to maximal representations, to the referees for many detailed
suggestions, and to Jean-Louis Clerc for asking about possible generalizations to other symmetric domains.
We acknowledge support of Swiss National Science Foundation grants number P2FRP2 161723
(Federica Fanoni) and P2EZP2 159117 (Maria Beatrice Pozzetti).
6
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
2. The symmetric space associated to Sp(2n, R)
Recall that the symplectic group Sp(2n, R) is the subgroup of SL(2n, R) preserving the
symplectic form ω(·, ·) represented, with respect to the standard basis, by the matrix
0 Idn
Jn = −Id
.
n 0
The symmetric space X associated to the symplectic group Sp(2n, R) is often referred to
as Siegel space. In this paper we will be concerned with locally symmetric spaces arising as
the quotient of X by the image ρ(Γ) < Sp(2n, R) of a maximal representation. We will be
interested in two models for X : the upper-half space and the image of the Borel embedding.
The upper-half space is the generalization of the upper-half plane model of the hyperbolic
plane, given by a specific set of symmetric matrices:
X = {X + iY |X ∈ Sym(n, R), Y ∈ Sym+ (n, R)},
where Sym(n, R) denotes the set of n-dimensional symmetric matrices with coefficients in R
and Sym+ (n, R) is the subset of Sym(n, R) given by positive definite matrices. The group
Sp(2n, R) acts by fractional linear transformations in this model:
A B
· Z = (AZ + B)(CZ + D)−1 .
C D
The image of the Borel embedding is
X = {l ∈ L(C2n ) | iω(·, σ(·))C |l×l is positive definite}.
Here L(C2n ) is the set of Lagrangians, the maximal isotropic subspaces of C2n for the complexification of the symplectic form ω(·, ·)C , and σ : C2n → C2n denotes the complex conjugation.
We will consider the affine chart ι : Sym(n, C) → L(C2n ) that associates to a symmetric
matrix Z the linear subspace of C2n spanned by the columns of the matrix IdZn . It is easy
to verify that ι is well defined and induces an Sp(2n, R)-equivariant identification ι : X → X
(cfr. [BP17, Section 2.2] for more detail). The complex conjugation σ : C2n → C2n induces
a map on L(C2n ) that will still be denoted by σ with a slight abuse of notation. It is easy to
verify that σ corresponds, via ι, to the complex conjugation on Sym(n, C).
A maximal polydisk in X is the image of a totally geodesic and holomorphic embedding of
the Cartesian product of n copies of the Poincaré disk into X . Maximal polydisks exist and
they are all conjugate under the action of Sp(2n, R) (see [Wol72, p. 280]). Polydisks arise as
complexifications of maximal flats. Thus each pair of points (x, y) is contained in a maximal
polydisk, that is unique if the direction determined by (x, y) is regular.
2.1. Lagrangians and boundaries. The set of real Lagrangians L(R2n ) naturally arises
as the unique closed Sp(2n, R) orbit in the boundary of X in its Borel embedding and for
this reason L(R2n ) is the Shilov boundary of the bounded domain realization of X (see
[Wie04] for details). The restriction of the affine chart ι to the subspace Sym(n, R) provides
a parametrization of the set of real Lagrangians that are transverse (as linear subspaces)
to he1 , . . . , en i. We will denote he1 , . . . , en i by l∞ , since it is at infinity in the affine chart
we chose. Whenever this won’t generate confusion we will omit the map ι and identify
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
7
symmetric matrices (with real or complex coefficients) with Lagrangian subspaces (of R2n
and C2n respectively).
Since X has higher rank for n > 1, the visual boundary ∂∞ X is not homogeneous [Ebe96],
and the relation between the visual boundary and the closure of X in the Borel embedding is
in general fairly complicated. However there is a closed orbit in ∂∞ X which is naturally identified with the Lagrangians [Loo77, Theorem 9.11]. In particular this allows us to associate
to a point l ∈ L(R2n ) a class of asymptotic directions.
Denote by L(R2n )(k) the set of k-tuples of pairwise transverse Lagrangians. It is well
known and easy to prove that the group Sp(2n, R) acts transitively L(R2n )(2) . Moreover, it
has (n + 1) orbits in L(R2n )(3) , indexed by the Maslov cocycle [LV80, Section 1.5]. The value
of the Maslov cocycle is maximal on the orbit of
(he1 , . . . , en i, hen+1 , . . . , e2n i, he1 + en+1 , . . . , en + e2n i) = (l∞ , 0, Id).
Definition 2.1. A triple of pairwise transverse Lagrangians is called maximal if it is in the
Sp(2n, R)-orbit of (l∞ , 0, Id).
Maximal triples should be regarded as a generalization of positively oriented triples in the
circle. As we will shortly see, maximal triples play a fundamental role in the definition and
study of maximal representations. We will need a concrete criterion to check when triples
of Lagrangians are maximal. The following is well known, and a proof can be found, for
example, in [BP17, Lemma 2.10].
Lemma 2.2. The following hold:
(1) any cyclic permutation of a maximal triple is maximal;
(2) the triple (l∞ , X, Y ) is maximal if and only if Y − X is positive definite;
(3) if Z − X is positive definite, the triple (X, Y, Z) is maximal if and only if Z − Y and
Y − X is positive definite.
More generally, we can define maximal m-tuples:
Definition 2.3. An m-tuple (l1 , . . . , lm ) ∈ L(R2n )(m) is maximal if for every i < j < k the
triple (li , lj , lk ) is maximal.
We will often consider maximal 4-tuples and it will be useful to consider special representatives in an Sp(2n, R)-orbit of a maximal 4-tuple. Two representatives are described in
[BP17, Prop. 2.11] and another is given in the following Lemma:
Lemma 2.4. If (l1 , l2 , l3 , l4 ) is a maximal 4-tuple, there exists Λ = diag(λ1 , . . . λn ) with
λi ∈ (0, 1) and an element g ∈ Sp(2n, R) such that
g(l1 , l2 , l3 , l4 ) = (−Id, −Λ, Λ, Id).
Proof. By [BP17, Prop. 2.11], we can assume (l1 , l2 , l3 , l4 ) = (−Id, 0, D, l∞ ) for a diagonal
matrix D = diag(d1 , . . . , dn ) with d1 ≥ · · · ≥ dn > 0. Note
that for every di > 0, there
αi βi
exists λi = λi (di ) and a matrix Ai = Ai (di ) =
such that Ai sends the 4-tuple
γi δi
(−1, 0, di , ∞) ∈ (∂ H2 )4 to (−1, −λi , λi , 1). Consider the element g ∈ Sp(2n, R) given by
diag(α1 , . . . , αn ) diag(β1 , . . . , βn )
g=
.
diag(γ1 , . . . , γn ) diag(δ1 , . . . , δn )
8
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Then it is straightforward to check that
g(−Id, 0, D, l∞ ) = (−Id, −Λ, Λ, Id),
where Λ = diag(λ1 , . . . , λn ).
2.2. Sp(2n, R)-invariant distances. We are interested in three Sp(2n, R)-invariant distances on the symmetric space X : the vectorial distance, the Riemannian distance and the
(determinant) Finsler distance.
Fix a point p in a maximal flat F and a Weyl chamber ā+ ⊂ Tp F . This is a fundamental
domain for the action of Sp(2n, R) on the tangent bundle T X . In our case we have
ā+ = {(x1 , . . . , xn ) ∈ Rn |x1 ≥ . . . ≥ xn ≥ 0}.
A vector in the model Weyl chamber is regular if all the inequalities are strict, which is
equivalent to being contained in a unique flat. We say that v is singular pointing in the
direction of a Lagrangian if x1 = . . . = xn > 0. Indeed, if a vector v is singular pointing
in the direction of a Lagrangian and γ is the geodesic determined by exponentiating v, the
endpoints of γ in the visual boundary ∂∞ X are two Lagrangians.
In order to define the projection onto the model Weyl chamber, we need to recall from
[BP17] the definition of an endomorphism-valued cross-ratio. If two real or complex Lakl
grangians l1 and l2 are transverse (denoted by l1 t l2 ), we denote by pl1 2 : R2n → l1 (resp.
kl
by pl1 2 : C2n → l1 ) the projection to l1 parallel to l2 .
Definition 2.5. For Lagrangians l1 , . . . , l4 ∈ L(C2n ) such that l1 t l2 and l3 t l4 , the
cross-ratio R(l1 , l2 , l3 , l4 ) is the endomorphism of l1 given by
kl
kl
R(l1 , l2 , l3 , l4 ) = pl1 2 ◦ pl4 3
l1
.
We will use multiple times the explicit expression for the cross-ratio R on the affine chart
ι(Sym(n, C)) of L(C2n ) (cfr. [BP17, Lemma 4.2]):
(2.1)
R(X1 , X2 , X3 , X4 ) = (X1 − X2 )−1 (X4 − X2 )(X4 − X3 )−1 (X1 − X3 ).
X1
Here R is expressed with respect to the basis of X1 given by the columns of the matrix Id
.
n
Since the symmetric space X is a complete, negatively curved Riemannian manifold, any
pair of points (a, b) ∈ X 2 is joined by a unique geodesic γ and therefore corresponds to a
unique vector v ∈ Ta X . In [Sie43], Siegel proved that the cross-ratio we just introduced
can be used to describe the projection of a pair of points in X onto the Weyl chamber, the
fundamental domain for the Sp(2n, R)-action on T X :
Theorem 2.6 ([Sie43]). The projection onto the Weyl chamber is given by
X 2 → ā+
(X, Z) 7→ (log(λ1 ), . . . , log(λn ))
where λi =
√
1+ ri
√
1− ri
and 1 > r1 ≥ . . . ≥ rn ≥ 0 are the eigenvalues of R(X, Z, Z, X).
We are now ready to define the distances on X .
+
Definition 2.7. The vectorial distance dā is the projection onto the Weyl chamber ā+ .
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
9
The vectorial distance is not R-valued; however there is a natural partial order on the
Weyl chamber that allows us to talk about triangular inequality. This, together with many
+
interesting properties of dā , is proven by Parreau in [Par]. Since in the symmetric space associated to Sp(2n, R) the opposition involution is trivial, the vectorial distance is symmetric:
+
+
for each pair x, y ∈ X it holds dā (x, y) = dā (y, x).
As the Weyl chamber is a fundamental domain for the Sp(2n, R)-action on X 2 , any
Sp(2n, R)-invariant distance can be obtained composing the vectorial distance with a suitable
function. We will use two R-valued distances:
+
Definition 2.8. The Riemannian distance dR is the composition of dā with the function
ā+ → R
qX
(x1 , . . . , xn ) 7→
x2i .
Notice that the Riemannian distance is the distance induced by the unique Sp(2n, R)invariant metric on X with minimal holomorphic sectional curvature equal to −1. The normalization is chosen so that the polydisks are isometrically and holomorphically embedded.
+
Definition 2.9. The (determinant) Finsler metric dF is the composition of dā
function
with the
ā+ → R
1X
xi .
(x1 , . . . , xn ) 7→
2
An easy application of Cauchy-Schwarz’s inequality shows:
Lemma 2.10. For every a, b ∈ X
dR (a, b) ≤ 2dF (a, b) ≤
√
ndR (a, b).
As we will see, the computation of Finsler and vectorial distance often reduces to computations about eigenvalues of positive definite symmetric matrices. We will denote by h·, ·i the
standard Euclidean inner product on R2n . One of the tools we will use many times is the
(Courant–Fischer–Weyl) min–max principle, which we recall here:
Lemma 2.11. Let A be a symmetric matrix of eigenvalues a1 ≥ · · · ≥ an . Then
hAv, vi
hAv, vi
ai =
min
max
| 0 6= v ∈ U = max min
| 0 6= v ∈ U .
kvk2
kvk2
dim(U )=i
dim(U )=n−i+1
In particular
a1 = max
v6=0
hAv, vi
hAv, vi
and an = min
.
2
v6=0 kvk2
kvk
The following lemmas will be handy:
Lemma 2.12. Let A, B be positive definite symmetric matrices. Then
max evA max evB ≥ max ev(AB) ≥ max evA min evB.
10
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Proof. Notice that since B is symmetric and positive definite it admits a positive square root.
Moreover, since the eigenvalues of a matrix are invariant under conjugation, the eigenvalues of
AB coincide with the eigenvalues of B 1/2 AB 1/2 . The result then follows using the min–max
principle:
hB 1/2 AB 1/2 v, vi
kvk2
v∈Rn \0
max ev(B 1/2 AB 1/2 ) = max
hB 1/2 AB 1/2 v, vi hB 1/2 v, B 1/2 vi
kvk2
v∈Rn \0
kB 1/2 vk2
iAw, wh
iBv, vh
≥ max
min
w6=0 kwk2
v6=0 kvk2
= max ev(A) min ev(B).
= max
Lemma 2.13. Let A, B be positive definite symmetric matrices. Then the difference A − B
is positive definite if and only if all eigenvalues of AB −1 are bigger than one.
Proof. This follows from the observation that
B −1/2 AB −1/2 = B −1/2 (A − B)B −1/2 + Id
and the fact that a symmetric matrix M is positive definite if and only if t N M N is positive
definite for any invertible matrix N .
Using Lemma 2.13, we get:
Lemma 2.14. Let A and B be positive definite symmetric matrices such that the difference
B − A is positive definite. Let µ1 ≥ · · · ≥ µn be the eigenvalues of A−1 B. Then
+
dā (iA, iB) = (log µ1 , . . . , log µn )
and
1
log det(A−1 B).
2
Proof. Note first that by Lemma 2.13 the eigenvalues of A−1 B are all bigger than one. We
then have
+
dā (iA, iB) = (log λ1 , . . . , log λn )
dF (iA, iB) =
√
1+ r
where λi = 1−√rii and the ri are the eigenvalues of R(iA, −iB, iB, −iA). Using (2.1), one
can compute that
R(iA, −iB, iB, −iA) = (Id + A−1 B)−2 (A−1 B − Id)2 ,
so ri =
(µi −1)2
.
(µi +1)2
But since all the µi are all bigger than one, we deduce that λi = µi . Hence
+
dā (iA, iB) = (log µ1 , . . . , log µn )
and
dF (iA, iB) =
n
n
j=1
j=1
Y
1X
1
1
log µj = log
µj = log det(A−1 B).
2
2
2
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
11
2.3. R-tubes. In this section we recall the definition of R-tubes, subspaces of X that play
the role of geodesics in H2 , and we will prove some results that we will use in the remainder
of the paper. Let {a, b} be an unordered pair of transverse real Lagrangians.
Definition 2.15. The R-tube Ya,b associated to {a, b} is the set
Ya,b = {l ∈ X | R(a, l, σ(l), b) = −Id}.
We will refer to the real Lagrangians a, b as the endpoints of Ya,b .
It can be proven (see [BP17, Section 4.2]) that Ya,b is a totally geodesic subspace of X of
the same real rank as X and that it is the parallel set of the Riemannian singular geodesics,
whose endpoints in the visual boundary of X are the Lagrangians a and b. The stabilizer of
Ya,b is StabSp(2n,R) ({a, b}) and is a Z/2Z-extension of GL(n, R) that acts transitively on the
R-tube.
Up to the symplectic group action we can reduce to a model R-tube, the one with endpoints
0 and l∞ . In the upper-half space model this standard tube consists of matrices of the form
Y0,l∞ = {iY | Y ∈ Sym+ (n, R)}
and GLn (R) acts on it as G · iX = i(GX(tG)).
It was exploited in [BP17] (cfr. also [BILW05]) that the incidence structure of R-tubes
in the Siegel space forms a synthetic geometry that shares many common features with the
hyperbolic geometry in H2 . In particular, the following result shows how the intersection
pattern of R-tubes reflects the intersection pattern of geodesics in the hyperbolic plane.
Proposition 2.16. If (l1 , l2 , l3 , l4 ) is maximal, the intersection Yl1 ,l3 ∩ Yl2 ,l4 consists of a
single point and Yl1 ,l2 ∩ Yl3 ,l4 is empty.
Proof. The first result is proven in [BP17, Lemma 4.7]. With the same techiques we can
prove that the second intersection is empty. We reproduce the argument for completeness.
Up to the action of the symplectic group, we can assume that
(l1 , l2 , l3 , l4 ) = (−Id, Λ, 0, l∞ ),
where Λ = diag(λ1 , . . . , λn ) and λi ∈ (−1, 0) (cfr. [BP17, Proposition 2.11]). Now a point y
belongs to Yl1 ,l2 ∩ Yl3 ,l4 = Y−Id,Λ ∩ Y0,`∞ if and only if y = iY , for some Y ∈ Sym(n, R), and
iY ∈ Y−Id,Λ . By definition iY ∈ Y−Id,Λ if and only if R(−Id, iY, −iY, Λ) = −Id. Using (2.1)
we have
R(−Id, iY, −iY, Λ) = −Id ⇔ (−Id − iY )−1 (Λ − iY )(Λ + iY )−1 (−Id + iY ) = −Id
⇔ (Λ − iY )(Λ + iY )−1 = (Id + iY )(−Id + iY )−1
⇔ (Λ − iY )(Λ + iY )−1 = (−Id + iY )−1 (Id + iY )
⇔ (−Id + iY )(Λ − iY ) = (Id + iY )(Λ + iY )
⇔ Y 2 = Λ.
But as Λ is negative definite, there is no solution to this equation.
12
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
2.4. The product structure of a tube and causal maps. We now turn to a more
precise description of the geometry of a single R-tube. Recall that the standard model for
the symmetric space associated to GL(n, R) is
XGL(n,R) = Sym+ (n, R)
and GL(n, R) acts on XGLn (R) by G · X = GX t G. We endow XGL(n,R) with the Riemannian
distance given by
v
u n
uX
dGL (X, Y ) = t (log λi )2 ,
i=1
where λi are the eigenvalues of XY −1 . This is twice the normalization for the Riemannian
distance chosen in [Ben97, Par12], but it is better suited to our purposes because with this
choice the natural identification XGL(n,R) ∼
= Y0,l∞ is an isometry (where Y0,l∞ is equipped
with the Riemannian metric).
Recall that a model for the symmetric space associated to SL(n, R) is
XSL(n,R) = {X ∈ Sym+ (n, R)| det(X) = 1}
and its model Weyl chamber is
(
ā+
SL(n,R) =
(x1 , . . . , xn ) ∈ Rn |x1 ≥ . . . ≥ xn ,
n
X
)
xi = 0 .
i=1
We associate to the pair (X, Y ) the vector (log λ1 , . . . , log λn ) where λ1 ≥ . . . ≥ λn are the
eigenvalues of XY −1 . We normalize the Riemannian metric on XSL(n,R) so that
v
u n
uX
dSL (X, Y ) = t (log λi )2 .
i=1
Coherently with above, this is twice the standard normalization.
The group GL(n, R) is reductive and its symmetric space, endowed with the Riemannian
metric, splits as the direct product XGL(n,R) = R × XSL(n,R) ; R is the Euclidean factor of
the reducible symmetric space XGL(n,R) . Explicitly:
Lemma 2.17. The map
F = π R × π SL : XGL(n,R) −→ R × XSL(n,R)
log det X
1
√
,
X
X 7−→
n
(det X)1/n
is an isometry.
Proof. Clearly the map is a bijection, so we just need to show that it preserves distances. For
any X, Y ∈ X we have
|π R (X) − π R (Y )| =
log det X − log det Y
log det(XY −1 )
√
√
=
.
n
n
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
13
Moreover
v
u n
uX
SL
SL
dSL (π (X), π (Y )) = t (log µi )2
i=1
where µ1 ≥ · · · ≥ µn are the eigenvalues of
−1
1
1
1
X
Y
=
XY −1 .
(det X)1/n
(det Y )1/n
(det(XY −1 ))1/n
So
µi =
1
λi
(det(XY −1 ))1/n
where the λi are the eigenvalues of XY −1 . Set d := det(XY −1 ). We have:
d(F (X), F (Y ))2 = |π R (X) − π R (Y )|2 + dSL (π SL (X), π SL (Y ))2
2
n
X
1
1
2
= (log d) +
log λi − log d
n
n
i=1
=
n
X
n
(log λi )2 +
i=1
X
2
log d (log d −
log λi ) = dGL (X, Y )2
n
i=1
|
{z
}
=0
√
√
Remark 2.18. On the model flat π R is the scalar product with the unit vector (1/ n, . . . , 1/ n)
and the tangent space to SL(n, R) is its orthogonal.
We now define causal maps and show some of the properties we will need.
Definition 2.19. Let K be a subset of R; a map f : K → XGL(n,R) is causal if for each pair
x < y ∈ K, f (y) − f (x) is positive definite.
The next lemma summarizes a property of causal paths that will be crucial for our future
analysis: for any causal path the length of its projection on the Euclidean factor gives an
upper bound on the length of the projection to XSL(n,R) :
Lemma 2.20. Let A, B ∈ XGL(n,R) . If B − A is positive definite, then
π R (B) > π R (A)
and
√
n − 1(π R (B) − π R (A)) > dSL (π SL (B), π SL (A)).
Proof. It follows from Lemma 2.13 that if B − A is positive definite then all eigenvalues of
BA−1 are bigger than 1. In particular det B > det A, which gives the first inequality.
In order to verify the second statement we can assume that A = Id: indeed GL(n, R) acts
transitively on XGL(n,R) by isometries, and since the splitting of Lemma 2.17 is isometric,
both terms in the second inequality are invariant by the GL(n, R)-action. Denoting by
14
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
b1 ≥ . . . ≥ bn > 1 the eigenvalues of B we have π R (B) =
√1
n
Pn
i=1 log bi ,
π R (A) = 0, and the
SL
SL
vector in ā+
SL(n,R) associated to the pair (π (B), π (A)) is
π R (B)
π R (B)
.
log b1 − √ , . . . , log bn − √
n
n
Then
2
n
X
π R (B)
dSL (π (B), π (A)) =
log bi − √
n
SL
SL
2
i=1
=
n
X
(log bi )2 − π R (B)2 ≤ (n − 1)π R (B)2 ,
i=1
where the last inequality follows from the fact that, for positive numbers xi , it holds
!2
n
n
X
X
2
xi ≤
xi .
i=1
i=1
In rank two we have the following stronger result:
Corollary 2.21. If n = 2, the matrix B − A is positive definite if and only if
π R (B) − π R (A) > dSL (π SL (B), π SL (A)).
Proof. Given Lemma 2.20, we just need to show that if the condition on the projections holds,
B − A is positive definite, i.e. (by Lemma 2.13) that the eigenvalues of BA−1 are strictly
bigger than one. As in the proof of Lemma 2.20, we can reduce ourselves to the case where
A = Id, so we need to show that the eigenvalues b1 ≥ b2 of B are bigger than one. But in
this case we have
1
π R (B) − π R (S) = √ (log b1 + log b2 )
2
and
s
2
2
1
1
SL
SL
dSL (π (B), π (A)) =
log b1 − (log b1 + log b2 ) + log b2 − (log b1 + log b2 )
2
2
1
= √ (log b1 − log b2 )
2
so the hypothesis implies log b2 > 0, i.e. b2 > 1, and hence b1 > 1 as well.
Remark 2.22. By restating Lemma 2.14 using the notation of this section, we get that
given a pair of points iX, iY ∈ Y0,l∞ such that X − Y is positive definite, the Finsler distance
dF (X, Y ) is, up to an explicit factor, equal to the difference of the projections of the two
points on the Euclidean factor:
√
2dF (iX, iY ) = n(π R (X) − π R (Y )).
This observation will play a crucial role in our proof of Theorem A.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
15
2.5. Orthogonality and orthogonal projection. In this section we define the concept of
orthogonality, following [BP17].
Definition 2.23. We say that two R-tubes Ya,b and Yc,d are orthogonal (and we write
Ya,b ⊥ Yc,d ) if they are orthogonal as submanifolds of the symmetric space endowed with the
Riemannian metric.
Concretely, Ya,b and Yc,d are orthogonal if they meet in a point x and there their tangent
spaces are orthogonal as subspaces of Tx X (with respect to the Riemannian metric). It was
observed in [BP17, Section 4.3] that the orthogonality relation can be expressed as a property
of the cross-ratio of the boundary points: if (a, c, b, d) is maximal, the R-tubes Ya,b and Yc,d
are orthogonal if and only if R(a, c, b, d) = 2Id.
Sp
Let us now fix a tube Ya,b . In [BP17, Section 4.3] the authors construct an involution σa,b
of Sp(2n, R), which is induced by the complex conjugation that fixes the real form Va,b of
C2n given by
Va,b = hv + iw|v ∈ a, w ∈ bi.
We denote by σa,b ∈ GL(2n, R) the matrix corresponding to such linear map. Note that σa,b
is not a symplectic matrix, but for any two tubes Ya,b and Yc,d the product σa,b σc,d belongs
to Sp(2n, R).
Sp
X of X whose fixed point set
The involution σa,b
induces an anti-holomorphic map σa,b
consists precisely of Ya,b . It was verified in [BP17, Cor. 4.7] that the R-tubes orthogonal to
Ya,b foliate the symmetric space X . Moreover σa,b induces also an involution
L
: ((a, b)) → ((b, a))
σa,b
where ((a, b)) := {c ∈ L | (a, c, b) is maximal}. For each Lagrangian c such that (a, c, b) is
L (c) is the unique Lagrangian d with the property that Y
maximal, σa,b
a,b ⊥ Yc,d . Using these
observations it is possible to define [BP17, Cor. 4.7] the orthogonal projection
pa,b : X ∪ ((a, b)) → Ya,b .
L and the restriction of
It will be useful to have concrete expression for the reflection σa,b
orthogonal projection pa,b to ((a, b)), for (a, b) = (0, l∞ ). Recall that we identify Sym(n, R)
with the set of Lagrangians in L(R2n ) that are transverse to l∞ via the restriction
of the
Idn 0
2n
affine chart ι : Sym(n, C) → L(C ). From the definitions, it follows that σ0,l∞ = 0 −Idn .
Further, one can also prove that:
Lemma 2.24. For any A ∈ Sym(n, R), the R-tubes YA,−A and Y0,l∞ are orthogonal and
L (A) = −A and p
their unique intersection point is iA. In particular σ0,l
0,l∞ (A) = iA.
∞
We will need the fact that the vectorial distance of the projection of two Lagrangians x, y
to an R-tube Ya,b can be computed in term of the eigenvalues of the cross-ratio of the four
Lagrangians.
Lemma 2.25. If (a, x, y, b) ∈ L(R2n )(4) is a maximal 4-tuple and pa,b is the orthogonal
+
projection onto Ya,b , the distance dā (pa,b (x), pa,b (y)) is (log µ1 , . . . , log µn ), where the µi are
the eigenvalues of R(a, x, y, b).
Proof. Up to the action of the symplectic group, we can reduce to the case (a, b) = (0, l∞ ).
In this case, the result follows from explicit computations.
16
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
3. Maximal representations
As mentioned in the introduction, maximal representations are the representations that
maximize the Toledo invariant, an invariant defined with the aid of bounded cohomology.
It follows from a deep result of Burger, Iozzi and Wienhard that these representations can
be equivalently characterized as representations admitting a “well-behaved” boundary map.
Precisely, let Γ be the fundamental group of an oriented surface Σ with negative Euler
characteristic and boundary ∂Σ (which could be empty). Fix a finite area hyperbolization of
Σ inducing an action of Γ on S1 = ∂H2 .
Definition 3.1. A representation ρ : Γ → Sp(2n, R) admits a maximal framing if there exists
a ρ-equivariant map φ : S1 → L(R2n ) which is monotone (i.e. the image of any positively
oriented triple in the circle is a maximal triple) and right continuous.
Burger, Iozzi and Wienhard (see [BIW10, Theorem 8]) proved that a representation admits
a maximal framing if and only if it is maximal. Since we will not directly need bounded
cohomology in the rest of the paper we refer the interested reader to [BIW10] for a definition
of the Toledo invariant. The following structural result about maximal representations allows
us to associate, to each maximal representation, a locally symmetric space whose fundamental
group is isomorphic to the fundamental group of Σ. In the paper we will be interested in the
geometry of such locally symmetric space.
Theorem 3.2 ([BIW10, Theorem 5]). Maximal representations are injective and have discrete image.
Given a maximal representation ρ : Γ → Sp(2n, R), the image ρ(γ) ∈ Sp(2n,R) of each
non-peripheral element γ ∈ Γ is Shilov hyperbolic: it is conjugate to A0 t A0−1 for a matrix A in GL(n, R) with all eigenvalues with absolute value greater than one (see [Str15]).
−
Equivalently, ρ(γ) fixes two transverse Lagrangians Λ+
γ and Λγ on which it acts expandingly
(resp. contractingly). Furthermore if φ : S1 → L(R2n ) is the equivariant boundary map, the
±
Lagrangians Λ±
γ are the images φ(γ ).
Assumption: from now on, whenever a surface Σ has boundary, we will restrict to maximal representations ρ such that the image of each peripheral element is Shilov hyperbolic.
It is possible to show that this is equivalent to the requirement that the representation
ρ : Γ → Sp(2n, R) is Anosov in the sense of [GW12], and therefore we will denote maximal
representations satisfying our assumption Anosov maximal representation. Moreover, we fix
an orientation on the boundary components such that the surface lies to the right of the
boundary. This corresponds to a choice between a peripheral element and its inverse. We
will always choose primitive peripheral elements according to the orientation of the boundary
components.
The synthetic geometry whose lines are R-tubes is particularly adapted to the study of
maximal representations: if ρ is a maximal representation, the boundary map φ allows us to
select specific R-tubes associated to the elements of the fundamental group. More precisely,
for each element γ ∈ Γ, ρ(γ) stabilizes an R-tube, denoted Yγ , whose endpoints are the
2n
attractive and repulsive fixed points Λ±
γ of ρ(γ) in L(R ). It follows from Proposition 2.16
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
17
e For ease
that such tubes have the same intersection pattern as the axes of the elements in Σ.
of notation, if a tube is associated to an element γ ∈ Γ, we will denote the projection defined
Sp
in section 2.5 by pγ and the associated involution by σγSp (or σρ(γ)
, if we want to underline
the representation we are considering).
For points x, y ∈ S1 , we denote by ((x, y)) the subset of S1 given by points z such that
(x, z, y) is a positive oriented triple. The following observation will be useful:
Remark 3.3. The transitivity of the Sp(2n, R)-action on the space of tubes and the identification of Y0,l∞ with XGL(n,R) , allows us to generalize Definition 2.19 and define the notion
of a causal path f : K → Y for any tube Y. With this definition, if φ : S1 → L(R2n ) is the
boundary map of a maximal representation ρ, then for each x < y the image of the map
((x, y)) → Yφ(x),φ(y)
t 7→ pφ(x),φ(y) (φ(t))
is a causal path.
3.1. Some special maximal representations. A special type of maximal representations
we will be interested in are the ones obtained using the diagonal embedding ∆ of n identical
copies of SL(2, R) into Sp(2n, R). The centralizer ZSp(2n,R) (∆(SL(2, R))) is the subgroup,
isomorphic to O(n), consisting of matrices of the form A0 A0 , for matrices A ∈ O(n).
Let ρ0 : Γ → PSL(2, R) be the holonomy representation of a hyperbolic structure on Σ.
The representation ρ0 can be lifted to a maximal representation ρ : Γ → SL(2, R). For
any ρ obtained this way and any character χ : Γ → ZSp(2n,R) (∆(SL(2, R))), the product
∆ ◦ ρ × χ is a maximal representation (see [BIW10] for details) whose image is contained
in ∆(SL2 (R)) × ZSp(2n,R) (∆(SL(2, R))) < Sp(2n, R). Observe that such a representation
preserves the image of the diagonal inclusion of the Poincaré disk in the standard polydisk.
Definition 3.4. We say that a representation ρ : Γ → Sp(2n, R) is a diagonal embedding of
a hyperbolization if there exists a lift ρ0 : Γ → SL(2, R) of the holonomy of a hyperbolization
and a character χ : Γ → ZSp(2n,R) (∆(SL(2, R))) such that ρ is conjugated to ∆ ◦ ρ0 × χ.
Another generalization of Teichmüller space that has attracted a lot of attention is the
Hitchin component. If iN : SL(2, R) → SL(N, R) denotes the unique irreducible representation, the Hitchin component HitN (Γ) is the component of the character variety
.
Hom(Γ, SL(N, R))
SL(N, R)
containing iN ◦ ρ, where ρ the holonomy of a hyperbolization. Representations in the Hitchin
component are called Hitchin representations. When N = 2n is even, the representation i2n
factors through Sp(2n, R). It turns out that all the Hitchin representations whose image is
contained in Sp(2n, R) are also maximal representations [BILW05, Example 3.10]. We call
such representations Hitchin maximal representations.
+
3.2. Distances and maximal representations. Given γ ∈ Γ, we define the length `ā (ρ(γ))
(resp. `R (ρ(γ)), `F (ρ(γ))) of ρ(γ) with respect to the vectorial (resp. Riemannian, Finsler)
metric to be the translation length of ρ(γ) acting on X computed with the corresponding
metric.
18
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Let A be a matrix representing the action of ρ(γ) on its attractive Lagrangian Λ+
γ and
suppose |a1 | ≥ · · · ≥ |an | > 1 are the absolute values of the eigenvalues of A. It is well
known (see [Ben97] and [Par12]) that the vectorial length of ρ(γ) is explicitely related to the
eigenvalues of A:
+
`ā (ρ(γ)) = (2 log |a1 |, . . . , 2 log |an |).
From this it is easy to deduce that
`F (ρ(γ)) = log det A
and
v
u n
uX
R
` (ρ(γ)) = 2t (log |ai |)2 .
i=1
Note that the Finsler metric assigns to a peripheral element a length which is closely related
to its action on the attractive Lagrangians. Moreover many points in Yγ realize the Finsler
translation length of the element ρ(γ):
Lemma 3.5. Let g ∈ Sp(2n, R) be Shilov hyperbolic with associated tube Y. For any x ∈ Y
such that (x, gx) is a causal segment we have
`F (g) = dF (gx, x).
In particular, if ρ is a maximal representation and φ is the associated boundary map, for any
x ∈ S1 ,
`F (ρ(γ)) = dF (pγ (φ(x)), pγ (φ(γx))).
Proof. Using the transitivity of the Sp(2n, R)-action we can assume that Λ+
g = l∞ and
A 0
=
0,
so
that
g
has
expression
Λ−
,
and,
up
to
conjugating
A,
we
can
assume that
g
0 tA−1
t
t
x = iId. In this case we get gx = iA A. The causality condition tells us that A A is positive
definite, and hence (by Lemma 2.14) dF (iId, iA tA) = log det A = `F (g).
Another important advantage of the Finsler metric is that it is additive on causal curves:
Lemma 3.6. Suppose x, y, z ∈ ((γ − , γ + )) ⊂ S 1 are positively oriented. Then
dF (pγ (φ(x)), pγ (φ(z))) = dF (pγ (φ(x)), pγ (φ(y))) + dF (pγ (φ(y)), pγ (φ(z))).
Proof. Up to the action of the symplectic group, we can assume φ(γ − ) = 0, φ(γ + ) = l∞ ,
φ(x) = Id, φ(y) = A and φ(z) = B. By monotonicity, we know that A − Id and B − A are
positive definite. By Lemma 2.24, the R-tubes passing through Id, A and B and orthogonal
to Y0,l∞ are Y−Id,Id , Y−A,A and Y−B,B respectively, so pγ (φ(x)) = iId, pγ (φ(y)) = iA and
pγ (φ(z)) = iB.
By Lemma 2.14
1
dF (pγ (φ(y)), pγ (φ(z))) = log det(A−1 B)
2
1
dF (pγ (φ(x)), pγ (φ(y))) = log det A
2
1
dF (pγ (φ(x)), pγ (φ(z))) = log det B,
2
which implies the desired equality.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
19
The Finsler metric is also closely related to a cross-ratio in the sense of Labourie. Precisely,
in [Lab07] Labourie introduced a notion of R-valued cross-ratio on the boundary at infinity
of the fundamental group of a surface. In [Lab08] he showed that
B : ∂Γ4∗ → R
(x, y, z, t) 7→ det R(φ(x), φ(t), φ(y), φ(z))
is a cross-ratio in the sense of [Lab07], where
∂Γ4∗ = {(x, y, z, t) ∈ ∂Γ4 | x 6= t and y 6= z}.
Moreover, given a cross-ratio B, the period of a non-trivial element γ ∈ Γ is defined as
`B (γ) := log | B(γ − , γ · y, γ + , y)|.
It is easy to check that
1
`F (γ) = `B (γ).
2
Labourie uses this cross-ratio in [Lab08] to show that maximal representations are well displacing (cfr. also [HS12] for a functorial extension to general Hermitian Lie groups). Recently
Martone and Zhang used this language to prove systolic inequalities for the Finsler distance2
[MZ16].
As opposed to the Finsler metric, the Riemannian metric is not additive on causal paths.
However we have the following (cfr. [BP17, Lemma 9.3]):
Lemma 3.7. Let x0 , . . . , xk ∈ Yγ be on a causal curve. Then
k
X
dR (xi−1 , xi ) ≤
√ R
nd (x1 , xk ).
i=1
3.3. Doubles. Let Σ be a surface with nonempty boundary. The purpose of this section is to
construct, for each Anosov maximal representation ρ : π1 (Σ, v) → Sp(2n, R), what we call the
holomorphic double of ρ: a specific maximal representation Dρ : π1 (DΣ, v) → Sp(2n, R) of
the fundamental group of the double of the surface Σ that restricts to the given representation
ρ.
Denote by c0 , . . . , cm the boundary components, and fix a basepoint v on c0 . Recall that
the double DΣ of the surface Σ is the surface obtained gluing two copies Σ, Σ of the surface
Σ along its boundary components (where Σ is Σ endowed with the opposite orientation). We
denote by j0 : Σ → DΣ and j1 : Σ → DΣ the natural inclusions and by j : DΣ → DΣ the
involution fixing the boundary components pointwise and with the property that j ◦ j0 = j1 .
With a slight abuse of notation we will also denote by j (resp. ji ) the maps induced at the
level of fundamental groups. Given two paths α and β such that α ends at the starting point
of β, we denote by α ∗ β their concatenation.
We denote by ci also loops parametrizing the respective boundary components and leaving
the surface to the right. Recall that for any γ ∈ Γ we denote by σρ(γ) the element of GL(2n, R)
inducing the involution σγSp (see Sections 2.5 and 3). We set σρ := σρ(c0 ) .
2Their result holds for the larger class of positively ratioed representations.
20
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Each arc α starting at v and ending at a boundary component ci determines the peripheral
element γα = α ∗ ci ∗ α−1 , which in turn gives a matrix σρ(γα ) ∈ GL(2n, R).
Our holomorphic double will be obtained amalgamating the representation ρ on j0 (π1 (Σ, v))
and the σρ ρσρ on j1 (π1 (Σ, v)):
Proposition 3.8. Let Σ be a surface with nonempty boundary. Let ρ : π1 (Σ, v) → Sp(2n, R)
be a maximal representation and assume that the image of every peripheral element is Shilov
hyperbolic. Then ρ is the restriction of a unique maximal representation Dρ : π1 (DΣ, v) →
Sp(2n, R) such that:
(1) for all elements γ in π1 (DΣ, v), Dρ(j(γ)) = σρ Dρ(γ)σρ ;
(2) for every arc α joining v to a boundary component of Σ we have
Dρ(α ∗ j(α)−1 ) = σρ(γα ) σρ .
The double of a Hitchin representation was defined by Labourie and McShane in [LM09,
Section 9]. The crucial difference of our setting is that in general maximal representations
do not have diagonalizable image, as opposed to Hitchin representations. For this reason
we cannot deduce our result from [LM09] and we need to choose a different involution of
Sp(2n, R). As a result, if ρ is a Hitchin maximal representation, the holomorphic double we
define here is different from the Hitchin double defined in [LM09]: if i2n is the irreducible
representation of SL(2, R) into Sp(2n, R) and h : Γ → SL(2, R) is the holonomy of a hyperbolization, Di2n (h) is different from i2n (Dh). On the other hand, we have ∆(Dh) = D∆(h),
and this motivated the choice of the definition of our double as a holomorphic double.
Proof of Proposition 3.8. To simplify the notation, we will drop the reference to the representation ρ both in σρ and σρ(γα ) , which we will simply denote by σ and σγα .
For all i between 1 and m, fix an arc αi joining v to the boundary component ci . If xi
denotes the concatenation αi ∗ j(αi )−1 , and γi ∈ π1 (Σ, v) is the class of the concatenation
γi = αi ∗ ci ∗ αi−1 , a presentation for the group DΓ is given by
π1 (DΣ, v) = j0 (π1 (Σ, v)), j1 (π1 (Σ, v)), x1 , . . . , xm j0 (c0 )j1 (c0 )−1 , j0 (γi )−1 xi j1 (γi )x−1
.
i
In particular, there exists at most one representation Dρ : DΓ → Sp(2n, R) satisfying the
hypotheses of Proposition 3.8, because the requirements of Proposition 3.8 uniquely determine
the image of Dρ on the generators. Indeed, we need to set for all γ ∈ π1 (Σ, v)
(3.1)
Dρ(j0 (γ)) = ρ(γ)
(3.2)
Dρ(j1 (γ)) = σρ(γ)σ
and for all i ∈ {1, . . . , m}
(3.3)
Dρ(xi ) = σγi σ.
To show that Dρ exists, we just need to show that (3.3), (3.1) and (3.2) determine a well
defined maximal representation. To prove that it is well defined we need to check that:
• property (1) holds for the xi ;
• the relations are mapped to the identity by Dρ;
• for any α ∗ j(α) different from xi , (2) holds.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
21
For the first point, we have
Dρ(j(xi )) = Dρ(j(αi ) ∗ αi−1 ) = Dρ((αi ∗ j(αi )−1 )−1 )
(3.3)
= [Dρ(αi ∗ j(αi )−1 )]−1 = (σγi σ)−1
(3.3)
= σσγi σσ = σDρ(xi )σ.
For the first relation, since σ and ρ(c0 ) commute, we have:
Dρ(j0 (c0 )j1 (c0 )−1 ) = (Dρ(j0 (c0 )) Dρ(j(j0 (c0 )−1 )
(3.1),(3.2)
=
ρ(c0 )σρ(c0 )−1 σ = Id.
For the other relations, we use the fact that σγi commutes with ρ(γi ). This implies:
Dρ(j0 (γi )−1 xi j1 (γi )x−1
i )
(3.1),(3.2),(3.3)
=
(ρ(γi−1 ))(σγi σ)(σρ(γi )σ)(σσγi ) = Id.
Let now α be any other arc with endpoint in the component ci . Denote by δα the concatenation δα = α ∗ αi−1 , so that α = δα ∗ αi . We get:
Dρ(α ∗ j(α)−1 ) = Dρ(δα (αi ∗ j(αi )−1 )j(δα )−1 )
(3.1),(3.3)
=
ρ(δα )σγi σσρ(δα )−1 σ
and since one can verify that
ρ(δα )σγi ρ(δα )−1 = σγα
we deduce that
Dρ(α ∗ j(α)−1 ) = σγα σ,
as required.
The fact that Dρ is a maximal representation follows from the additivity formula for
the Toledo invariant (see [BIW10, Theorem 1(3)], and also [Str15]) and the fact that the
restriction of Dρ to j1 (π1 (Σ, v)) is maximal being conjugate via an anti-holomorphic isometry
to a maximal representation of a surface with the opposite orientation.
For how we defined it, the double of a representation depends on the choice of a boundary
component and of a base point on it. The goal of the next proposition is to show that we
can forget about this choice, as up to conjugation in the fundamental group we get the same
double.
Proposition 3.9. Let v, w ∈ ∂Σ and β an arc from v to w. Let η : π1 (Σ, w) → Sp(2n, R)
be an Anosov maximal representation. Then
Dη(δ) = Dρ(β ∗ δ ∗ β −1 ),
where ρ : π1 (Σ, v) → Sp(2n, R) is the representation given by
ρ(γ) = η(β −1 ∗ γ ∗ β).
22
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Proof. Denote by c be boundary component containing v and by d the one containing w.
Note first that since η is maximal, ρ is maximal as well, so we can define its double as
in Proposition 3.8. To show the equality, we will prove that θ(γ) := Dρ(β ∗ γ ∗ β −1 ) is a
maximal representation which extends η and satisfies the two conditions of Proposition 3.8.
By uniqueness, this will imply that θ = Dη.
Clearly θ is maximal, since Dρ is. Suppose δ ∈ π1 (Σ, w). Then β ∗ δ ∗ β −1 ∈ π1 (Σ, v), so
θ(j0 (δ)) = Dρ(j0 (β ∗ δ ∗ β −1 )) = ρ(β ∗ δ ∗ β −1 ) = η(δ),
i.e. θ extends η.
We now prove property (1). We have:
θ(j(δ)) = Dρ(β ∗ j(δ) ∗ β −1 ) = Dρ(β ∗ j(β)−1 )Dρ(j(β ∗ δ ∗ β −1 ))Dρ(j(β) ∗ β −1 )
= σρ(γβ ) σρ σρ θ(δ)σρ σρ σρ(γβ )
and since ρ(γβ ) = ρ(β ∗ d ∗ β −1 ) = η(d), we have σρ(γβ ) = ση , which shows (1).
Finally, we show that property (2) holds. Let α be an arc joining w with a boundary
component e. Then
θ(α ∗ j(α)−1 ) = Dρ(β ∗ α ∗ j(α)−1 ∗ β −1 )
= Dρ((β ∗ α) ∗ j(β ∗ α)−1 )Dρ(j(β) ∗ β −1 )
= σρ(γβ∗α ) σρ σρ σρ(γβ ) .
Since we have
ρ(γβ∗α ) = ρ(β ∗ α ∗ e ∗ α−1 ∗ β −1 ) = η(γα )
and we know from before that σρ(γβ ) = ση , we get θ(α ∗ j(α)−1 ) = ση(γα ) ση .
If a surface has nonempty boundary, the boundary map associated to a maximal representation of its fundamental group is only right-continuous in general, while for closed surfaces
the boundary map has especially good properties:
Theorem 3.10 ([BILW05, Corollary 6.3]). Let ρ : π1 (Σ) → Sp(2n, R) be a maximal representation, where Σ is a closed surface. Then there is a ρ-equivariant continuous injective
map φ : S1 → L(R2n ) with rectifiable image.
We will associate to a maximal representation ρ of a surface with boundary the boundary
map of Dρ, so that we have a continuous map in this case as well. More precisely, we denote
by DΓ the fundamental group of the double of Σ and fix once and for all an action h of DΓ
on H2 , inducing an action of DΓ on S1 . Let φ : S1 → L(R2n ) be the continuous boundary
map associated to Dρ. We will denote by Λ(Γ) ⊂ S1 the limit set of Γ on ∂H2 . If we fix any
finite-area hyperbolization of Σ with holonomy h0 and denote by t : S1 → Λ(Γ) the (h0 , h|Γ )equivariant map associating to any parabolic point p for h0 the attractive fixed point of h(p),
then the composition φ ◦ t is the equivariant boundary map of Definition 3.13.
3Cfr. also [BT17] for a different explicit construction of the boundary map.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
23
4. Defining orthotubes
The purpose of the section is to define a notion of orthogeodesic in our setting, which will
allow us to give a geometric interpretation of our main result as inequalities relating intrinsic
geometric quantities in the locally symmetric space associated to a maximal representation.
We will define orthotubes and their length, and we will prove that the length of an orthotube
is half the length of the corresponding curve in the double of the surface.
4.1. The classical definition. Let Σ be a hyperbolic surface with nonempty geodesic
boundary and fundamental group Γ. Fix a boundary component c and consider the set
of oriented orthogeodesics starting from c, that is, the set of oriented geodesic segments with
first endpoint on c, second endpoint in ∂Σ and orthogonal to the boundary at both endpoints.
If we fix a fundamental domain I in a lift c̃ of c in Σ̃ = H2 , any oriented orthogeodesic starting
from c can be lifted uniquely to a geodesic segment in H2 orthogonal to I and to the lift d˜ of
a boundary component d (c and d can be the same, but c̃ and d˜ are different).
On the one hand, this implies that the set of orthogeodesics is in bijection with the set
of arcs starting at c and ending at a boundary component modulo homotopy relative to the
boundary of Σ. On the other, the set of oriented orthogeodesics starting from c is also in
bijection with the set
.
e d˜ 6= c̃}
{d˜ ⊆ ∂ Σ,
StabΓ (c̃) .
Now, each lift of a boundary component corresponds to a unique primitive peripheral element
of Γ with the correct orientation4. So if we fix the peripheral element γ corresponding to c,
the set of oriented orthogeodesics starting from c is also in bijection with
.
2
{δ
=
6
γ
peripheral}
OH
(γ)
:=
Σ
hγi ,
where hγi acts by conjugation.
2
Similarly, if OH
Σ denotes the set of all unoriented orthogeodesic in Σ, we can see that it is
in bijection with
.
2
{{γ, δ} | γ 6= δ peripheral}
OH
Σ :=
Γ.
4.2. Orthotubes in X . Interpreting orthogeodesics as a pairs of peripheral elements gives
a natural way to define orthotubes as orthogonal R-tubes in X :
Definition 4.1. An orthotube corresponding to the pair α = (γ, δ) is a tube Yα in X orthogonal to Yγ and Yδ .
It is easy to show that for every pair of peripheral elements there exists a unique orthotube:
Lemma 4.2. For each pair of peripheral elements, an orthotube exists and is unique. The
assignment of orthotubes is Γ-equivariant.
Proof. Consider γ and δ; by [BP17, Proposition 2.11] we can assume that φ(γ + ) = −Id,
φ(γ − ) = Λ, φ(δ + ) = 0 and φ(δ − ) = l∞ , for some diagonal matrix Λ with eigenvalues between
−1 and 0. If Y is orthogonal to Yδ , it must then be of the form Y−A,A for some positive
4From now on peripheral element will mean primitive peripheral element with compatible orientation.
24
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
definite symmetric matrix A. We show that there exists a unique matrix A such that Y is
orthogonal to Yγ as well, i.e. such that
R(−Id, −A, Λ, A) = 2Id.
By explicit computations using (2.1), this is equivalent to A2 = −Λ, which has indeed one
and only one positive definite solution.
We define the length of an orthotube Yα to be the distance, with respect to the vectorial,
Riemannian or Finsler metric, of the unique intersection points with the tubes associated
+
with the peripheral elements determining α. We denote such distances by `ā (α), `R (α) and
`F (α) respectively. Observe that any Riemannian geodesic segment in X with endpoints in
Yγ and Yδ and orthogonal to both tubes is necessarily contained in Yα and has length `R (α).
In particular in the locally symmetric space Γ\X the Riemannian length of an orthotube is
the length of a local length minimizer between the projections of two peripheral tubes.
In analogy with the hyperbolic case, we define OΣ (γ) to be the set of all orthotubes
associated to pairs (γ, δ), for every peripheral element δ 6= γ, up to the action of hγi by
conjugation, and OΣ to be the union of all OΣ (γ) up to the action of Γ by conjugation and
of Z/2Z by switching endpoints. Note that any two orthotubes in the same class have the
same length, so we can talk about the length of an element of OΣ (γ) or OΣ .
Remark 4.3. Each element α ∈ OΣ appears twice in the union of all OΣ (γ). So for any
positive real-valued function f , if Σ is a surface with m boundary components represented
by γ1 , . . . , γn , we have:
2
X
α∈OΣ
+
f (`ā (α)) =
m
X
X
+
f (`ā (α)).
i=1 α∈OΣ (γi )
4.3. Orthotubes and doubles. In the case of hyperbolic surfaces, any orthogeodesic doubles to a closed geodesic in the double of the surface, whose length is twice the length of
the orthogeodesic. The purpose of the section is to show that the same holds in our setting.
Formally if α = (δ1 , δ2 ) is an orthotube of Σ, it corresponds to a homotopy class of paths,
also denoted by α with a slight abuse of notation, between two boundary components of Σ,
e between the
and hence to an element Dα ∈ π1 (DΣ, v). Explicitly, if β denotes a path in Σ
preferred lift of v and the axis of δ1 , and if β is the projection to Σ of β, Dα is the class, in
π1 (Σ, v) corresponding to β ∗ α ∗ j(α)−1 ∗ β −1 . We have the following:
Proposition 4.4. For each orthotube α we have
(1) Yα = YDα ;
(2) 2`F (α) = `F (Dα).
Proof. We show the result for orthotubes α starting from the boundary component c0 . The
general result will follow from Proposition 3.9.
Since Dα = α ∗ j(α)−1 , we know that Dρ(Dα) = σγα σρ . Note that by [BP17, Lemma
4.15], if Ya,b ⊥ Yc,d , then setwise σa,b Yc,d = Yc,d , and by [BP17, Lemma 4.11] σa,b Ya,b = Ya,b .
As a consequence,
Dρ(Dα)Yα = Yα .
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
25
−
Suppose the endpoints of Yα are Lagrangians a, b, where (Λ+
γα , a, Λγα , b) is maximal. We want
to show that b (resp. a) is the repulsive (resp. attractive) Lagrangian of Dρ(Dα). Note first
that since Dρ(Dα) fixes the tube Yα , it either fixes or exchanges a and b. But
Dρ(Dα)(a) = σγα σρ (a) = σγα (b),
where the second equality holds by [BP17, Lemma 4.15]. But since b is not an endpoint of
Yγα , it is not fixed by σγα , which implies that Dρ(Dα)(a) = a and Dρ(Dα)(b) = b.
−
−
−1 −
Moreover, since Dρ(Dα)−1 Λ−
γα = σρ Λγα , we get that (a, Λγα , Dρ(Dα) Λγα , b) is maximal.
We claim that this implies that b and a are the repulsive and attractive Lagrangians of
Dρ(Dα). This follows from the following observation:
Remark 4.5. Let g ∈ Sp(2n, R) be Shilov hyperbolic fixing two Lagrangians a and b. If
−
there exists a point x ∈ ((a, b)) with (a, x, g −1 x, b) maximal, then a = Λ+
g and b = Λg .
To prove the remark, observe that upto the symplectic group action we can assume that
(a, x, b) = (l∞ , Id, 0) and g = A0 t A0−1 , and we need to verify that l∞ is the attractive
Lagrangian for g, namely that the eigenvalues of A are precisely the eigenvalues of g that
have absolute value bigger than 1. But the hypothesis that (a, x, g −1 x, b) is maximal implies
that (a, gx, x, b) is maximal and so
1 < min ev(At A) ≤ | min ev(A)|2 ,
as requested.
So we get Yα = YDα .
Λ−
γα
Λ+
c0
a
Λ+
γα
Yα
σρ Λ−
γα
Λ−
c0
b
σρ Λ+
γα
Figure 1. The R-tubes appearing in the proof
We now want to show the statement about the Finsler lengths. Denoting the projection
onto Yα by pα , we have
−
`F (α) = dF (pα (Λ+
γα ), pα (Λc0 ))
=
1
−
F
−
+
dF (pα (Λ+
γα ), pα (Λc0 )) + d (pα (σρ (Λc0 )), pα (σρ (Λγα ))
2
| {z }
Λ−
c0
1
+
= dF (pα (Λ+
γα ), pα (σρ (Λγα ))
2
26
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
where the last equality follows from the additivity of the Finsler metric on causal curves
+
(Lemma 3.6). But Dρ(Dα)−1 (Λ+
γα ) = σρ Λγα , thus by Lemma 3.5
+
F
dF (pα (Λ+
γα ), pα (σρ (Λγα )) = ` (Dα),
which implies that `F (α) = 12 `F (Dα).
5. Strategy of proof
The idea of the proof of Basmajian’s identity is the following. Fix a lift c̃ of a boundary
component c to H2 , with endpoints x and y. Pick z ∈ c̃ and let γ be the peripheral element
with axis c̃. We can write (z, γz) as
[
˜
(z, γz) = pc̃ ( ((x, y)) ∩ Λ(Γ)) ∪
pc̃ (d)
∩ (z, γz)
˜ Σ
e
d⊂∂
with endpoints
in ((x,y))
where pc̃ is the orthogonal projections onto c̃ and Λ(Γ) is the limit set of Γ. For every other
component d˜ of ∂ Σ̃, we can compute the length of the projection onto c̃ in terms of the length
of the corresponding orthogeodesic, using hyperbolic trigonometry. Since the limit set has
measure zero (see Proposition 6.1), we can deduce that its projection onto c̃ has measure zero
as well. The length of (z, γz) is the length of c, and hence we obtain Basmajian’s identity.
d˜3
y
c̃
d˜2
γz
d˜1
z
x
Figure 2. Some projections on the lift c̃ (orthogonals in red)
Our proof of (A1), Basmajian-type inequalities for the Finsler metric, follows the same
strategy of the classical proof, using additivity of the Finsler metric along causal paths
(Lemma 3.6) and the important fact that the Finlser translation length of an element is
attained at any point along a causal curve (Lemma 3.5). With this at hand, given any peripheral element γ, we look at the R-tubes corresponding to the other peripheral elements δ
and we compute the vectorial distance between the projection of the two endpoints of a tube
Yδ onto Yγ in terms of the vectorial length of the orthogeodesic between them (Lemma 7.1),
which in turn (Lemma 7.3) gives an inequality for the Finsler metric. The generalization
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
27
(Theorem 6.4) of the fact that the limit set has measure zero is proven in Section 6, following
the strategy explained in Section 6.1.
The characterization of diagonal embeddings as representations attaining the equalities in
(A1) follows from the observation that having equalities is equivalent (Lemma 7.1) to the fact
+
−
+
that the cross-ratios of the form R(Λ−
γ , Λδ , Λδ , Λγ ) are multiples of the identity. By Lemma
7.4, this implies that the image of the boundary map is in the boundary of a diagonal disk,
which allows us to deduce the characterization we want.
The Riemannian case is a priori harder, since neither Lemma 3.5 nor Lemma 3.6 hold true
for the Riemannian distance. However we rely on the observation that the translation length
of a peripheral element γ is at most the translation length of γ on the Euclidean factor of
Yγ , and the latter quantity is, up to a constant, its Finsler translation length (Remark 2.22).
This allows us to deduce (A2) as a consequence of the result about the Finsler metric.
6. An identity between cross-ratios
Basmajian’s proof of his celebrated identity builds on the following well known fact:
Proposition 6.1. Let Γ be a Fuchsian group corresponding to the holonomy of a hyperbolic
surface Σ with nonempty geodesic boundary. Then the Lebesque measure of the limit set
Λ = Λ(Γ) is zero.
The goal of this section is to show that the analogous result holds for Anosov maximal
representation, by adapting to the higher rank setting a proof which has been kindly suggested
to us by Brian Bowditch (a similar argument can be found in [Tuk84]).
We will use the following result, giving a sufficient condition for a subset of an interval to
have Lebesgue measure zero.
Lemma 6.2. Let [0, `] be an interval in R, X be a union of open subintervals {Iα }α of [0, `]
and Y = [0, `] \ X. Suppose there exists λ > 0 such that for every interval I ⊂ [0, `] there
exists a finite union of closed intervals J ⊂ I such that
(1) J ⊂ X ∩ I, and
(2) µLeb (J) ≥ λ µLeb (I).
Then µLeb (Y ) = 0.
Proof. We construct by induction a sequence of sets Yk ⊃ Y which is a union of intervals and
such that µLeb (Yk ) ≤ (1 − λ)k `.
For k = 0, consider I = [0, `]. By hypothesis, there exists a finite union of subintervals
S
J0 ⊂ X such that µLeb (J0 ) ≥ λ`. Set Y0 := X \ J0 . Suppose we have constructed Yk = β Iβk
with µLeb (Yk ) ≤ (1 − λ)k `. Then for each Iβk we know that there exists a finite union of
S
subintervals Jβk satisfying (1) and (2); we set Yk+1 := Yk \ β Jβk . Then Yk+1 is still a union
of intervals and
µLeb (Yk+1 ) ≤ (1 − λ) µLeb (Yk ) ≤ (1 − λ)k+1 `.
28
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
6.1. The classical proof. In this section we sketch the proof of Proposition 6.1.
We will show that for any compact interval T ⊂ ∂ H2 ' S1 , µLeb (T ∩ Λ) = 0. Note that
the measure µLeb on T is the one induced by the measure on S1 . Fix an identification of ∂ H2
with R ∪ {∞} so that T = [0, L]. Since (T, µLeb ) is C-biLipschitz to T endowed with the
distance | · | induced by R, it is enough to prove our claim using | · |.
We want to apply Lemma 6.2 and we claim that it is enough to verify the assumptions
only for subintervals I = [a, b] ⊂ T whose extrema a, b belong to Λ(Γ). Indeed, consider
a0 = inf{x ∈ [a, b]|x ∈ Λ} and b0 = sup{x ∈ [a, b]|x ∈ Λ}. We have I = [a, a0 ) ∪ [a0 , b0 ] ∪ (b0 , b]
and the first and the last subintervals are contained in I \ Λ.
We will use the following observation:
Remark 6.3. If Σ is a complete hyperbolic structure on a compact surface with non-empty
boundary, there is a constant r = r(Σ) such that for every p ∈ Σ, the distance of p from the
boundary is at most r.
Consider the unique point p in the geodesic between a and b with imaginary part (b − a)/2.
e because a and b are in the limit set. By Remark 6.3, there is a point q ∈ ∂ Σ
e
Note that p ∈ Σ
at distance at most r from p, which implies that Im q ≥ e−r (b − a)/2. Moreover q belongs to
a lift of a boundary component of Σ with endpoints c < d in I, so (c, d) ⊂ I \ Λ, and we have
(d − c) ≥ 2 Im q ≥ e−r (b − a).
So we can set λ = e−r and apply Lemma 6.2 to deduce that |Λ ∩ T | = 0.
Dr (p)
p
q
a c
d
b
Figure 3. Finding the subinterval
6.2. Maximal representations. Our objective is to prove the following result:
Theorem 6.4. Let Σ be a surface with nonempty boundary and let ρ : π1 (Σ) → Sp(2n, R)
be an Anosov maximal representation. For every peripheral element γ we have:
X
`F (γ) =
dF (pγ (φ(δα+ )), pγ (φ(δα− )))
α∈OΣ (γ)
where φ is the boundary map associated to ρ and δα is the peripheral element associated to α
and different from γ.
Remark 6.5. Observe that the inequality
X
`F (γ) ≥
dF (pγ (φ(δα+ )), pγ (φ(δα− )))
α∈OΣ (γ)
immediately follows from the fact that the Finsler translation distance of γ is attained at the
projection pγ (φ(x)) of any point φ(x) in the image of the boundary map (Lemma 3.5) together
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
29
with the additivity of the Finsler distance along causal paths (Lemma 3.6). Obviously the
same inequality holds also if some peripheral element is not Shilov hyperbolic, as long as ρ(γ)
is, restricting the sum to the orthogeodesic corresponding to pairs of peripheral elements
whose image is Shilov hyperbolic.
Remark 6.6. Note that we can rewrite Theorem 6.4 as the Basmajian identity for cross-ratios
associated to maximal representations, stated in the introduction as Theorem B. Indeed, it
follows from Lemma 2.25 that
dF (pγ (φ(δα+ )), pγ (φ(δα− ))) =
1
log B(γ − , δα− , γ + , δα+ )
2
and we have already noticed in Section 3.2 that for any element γ ∈ Γ
1
`F (γ) = `B (γ).
2
To prove Theorem 6.4, we will use the same strategy explained in Section 6.1. The similarities will be evident as we will consider the upper-half space model for the symmetric space
associated to Sp(2, R); the maximum eigenvalue of the imaginary part of a point in X will
play the role of the imaginary part of a point in H2 .
Consider [[x, γx]] = {x, γx} ∪ ((x, γx)) ⊂ ∂ H2 and denote by ` the Finsler translation
distance of γ. We can assume that φ(γ + ) = l∞ , φ(γ − ) = 0, φ(x) = Id, up to conjugating with
an element in Sp(2n, R). We define a monotone map
θ : [[x, γx]] → [0, `]
y 7→ dF (pγ (φ(y)), pγ (φ(x))) =
1
log det(φ(y)φ(x)−1 )
2
where the equality follows from Lemma 2.14 and Remark 3.3 (see also the proof of Lemma
3.6). For any orthotube α so that δα+ , δα− ∈ ((x, γ · x)), let Iα be
Iα := (θ(δα+ ), θ(δα− )) ⊂ [0, `],
S
and X = α Iα .
Proving Theorem 6.4 is then equivalent to showing that µLeb (X) = `, which in turn is the
same as proving that µLeb ([0, `] \ X) = 0. We want to use Lemma 6.2; as in the classical
case, we know that we can reduce ourselves to consider subintervals of [0, `] with endpoints
in θ(Λ(Γ)). So it is enough to prove the following:
Proposition 6.7. There exists a constant λ > 0 such that for every a, b ∈ θ(Λ(Γ)) there is
an interval Iα ⊂ [a, b] ∩ X with µLeb (Iα ) ≥ λ(b − a).
To be able to use Remark 6.3 in this setting, we first need a ρ-equivariant map from the
unit tangent bundle of the hyperbolic plane into X . Recall that we fixed a cocompact action
h of DΓ on H2 . We parametrize the unit tangent bundle by positively oriented triples of
points (a, b, c) on the boundary of the hyperbolic plane: a point (p, v) ∈ T 1 H2 determines an
oriented geodesic l and a and c denote its start and end points. Moreover, b is the point at
infinity of the geodesic ray starting from p, orthogonal to l and to the right of l.
30
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
c
(p, v)
b
a
Figure 4. The parametrization of T 1 H2
We use the boundary map φ to define a Dρ-equivariant map
F : T 1 H2 → X
(a, b, c) 7→ pYφ(a),φ(c) (φ(b)).
Proposition 6.8. The map F is a quasi-isometric embedding with respect to the Riemannian
and Finsler metric on X .
Proof. Since the Riemannian and Finsler metrics on X are quasi-isometric, it is enough to
prove the result for the Riemannian metric. It is shown in [BILW05, Corollary 6.2] that the
restriction of F to a DΓ-orbit is a quasi-isometric embedding. The result then follows since
DΓ acts cocompactly on H2 and F is continuous (since φ is).
In the proof of Proposition 6.7 it will be useful to be able to relate the maximum eigenvalue
of B − A to the logarithm of det BA−1 , for A, B ∈ φ(∂Γ).
Lemma 6.9. Assume (Id, A, B, mId) is maximal, for some m > 1. Then there exists k1 , k2
depending only on m and n such that
k2 log det(BA−1 ) > max ev(B − A) > k1 log det(BA−1 ).
Proof. Since (Id, A, B, mId) is maximal, Lemma 2.2 and Lemma 2.13 imply that all eigenvalues of A and B are between 1 and m. Moreover it follows from Lemma 2.13 that all
eigenvalues of BA−1 are bigger than 1, and since B − B 1/2 A−1 B 1/2 B is positive definite
they are also smaller than m (compare with Lemma 2.12). Moreover B − A is conjugate to
(A−1/2 BA−1/2 − Id)A, so by Lemma 2.12 (and since the eigenvalues of A are bigger than 1)
we have
max ev(BA−1 ) − 1 ≤ max ev(B − A) < max ev(A) max ev(BA−1 ) − 1 .
Because log x ≤ x − 1 for any positive x,
1
max ev(BA−1 ) − 1 ≥ log max ev(BA−1 ) ≥ log det(BA−1 )
n
where in the second inequality we used the fact that (max ev(BA−1 ))n ≥ det(BA−1 ).
x
Furthermore, log
x−1 is monotone decreasing for x > 1, so
max ev(BA−1 ) − 1 ≤
m−1
m−1
log max ev(BA−1 ) ≤
log det(BA−1 ).
log(m)
log(m)
Combining these inequalities and using the fact that the eigenvalues of A are smaller than m
we get the result.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
31
In the hyperbolic setting, given two points x, y ∈ Λ, we know that there exists a point
p ∈ Se ⊂ H2 with Im p = y−x
2 and belonging to the geodesic of endpoints x and y. In the case of
maximal representations, given x, y ∈ Λ, with φ(x) = A and φ(y) = B, we cannot guarantee
that there is a point with imaginary part 12 (B − A) belonging to the path pA,B (φ(z)), for
z ∈ ((x, y)). Our next goal is to show that there exists z ∈ ((x, y)) so that the maximum
eigenvalue of the imaginary part of pA,B (φ(z)) is large enough with respect to the maximum
eigenvalue of B − A. We begin by computing the imaginary part of the projection onto a
tube with endpoints (−C 2 , C 2 ):
Lemma 6.10. Assume that (−C 2 , T, C 2 ) is a maximal triple. Then
Im p−C 2 ,C 2 (T ) = C(Id − M 2 )(Id + M 2 )−1 C,
where M = C −1 T C −1 .
Proof. Let us choose γ ∈ Sp(2n, R) such that γ(0, l∞ ) = (−C 2 , C 2 ). For example
!
√1 C
√1 C
−
2
2
γ=
.
√1 C −1 √1 C −1
2
2
The preimage Z := γ −1 · T ∈ ((0, l∞ )) is
Z = (C −1 T + C)(−C −1 T + C) = (C −1 T C −1 + Id)(−C −1 T C −1 + Id)−1 .
Hence
p−C 2 ,C 2 (T ) = γ(pY0,∞ (γ −1 · T )) = γ · iZ =
= (iCZ − C)(iC −1 Z + C −1 )−1 = C(iZ − Id)(iZ + Id)−1 C =
= C(iZ − Id)(iZ − Id)(iZ − Id)−1 (iZ + Id)−1 C
so
Im p−C 2 ,C 2 (T ) = 2CZ(Z 2 + Id)−1 C,
which becomes, after some computations,
Im p−C 2 ,C 2 (T ) = C(Id − M 2 )(Id + M 2 )−1 C
where M = C −1 T C −1 as required.
The following Proposition is the hardest step in the proof of Theorem 6.4, and requires a
careful analysis of causal paths.
Proposition 6.11. There is an uniform constant k3 depending on the representation ρ such
that, for any ((x, y)) ⊂ ((γ − , γ + )), there exists z ∈ ((x, y)) with
max ev(Impφ(x),φ(y) (φ(z))) > k3 max ev(φ(y) − φ(x)).
Note that while the constant k3 is independent on γ, for the statement to make sense we
need to know what φ(x) and φ(y) are in ((0, l∞ )) = ((φ(γ − ), φ(γ + ))), so that we can identify
them with positive definite symmetric matrices.
32
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
K
Proof. Up to the action of an element H ∈ Sp(2n, R) of the form H = Id
0 Id (whose action
doesn’t change the imaginary part) we can assume φ(x) = −C 2 and φ(y) = C 2 , for some
positive definite matrix C. So
2 max ev(C 2 ) = max ev(φ(y) − φ(x)).
As in Lemma 6.10 we will denote
M (t) := C −1 φ(t)C −1
for t ∈ ((x, y)). M (t) forms a continuous monotone curve with M (x) = −Id and M (y) = Id.
Without loss of generality, we can assume that e1 is an eigenvector of C for its maximum
eigenvalue. Then we get
max ev Impφ(x),φ(y) (φ(t)) = max hC(Id − M (t)2 )(Id + M (t)2 )−1 Cv, vi
kvk=1
≥ hC(Id − M (t)2 )(Id + M (t)2 )−1 Ce1 , e1 i
= max ev(C 2 )[(Id − M (t)2 )(Id + M (t)2 )−1 ]11 .
So it is enough to show that there exists an uniform k3 > 0 such that for each such path
there exists t ∈ ((x, y)) with [(Id − M (t)2 )(Id + M (t)2 )−1 ]11 > 2k3 .
Set X(t) = M (t)2 . Notice that, for every t ∈ (x, y), the matrix X(t) is a positive semidefinite matrix, and all its eigenvalues and diagonal coefficients are smaller than one: indeed, by
monotonicity, all eigenvalues of M (t) are in absolute value smaller than one.
Since any matrix in Sym(n − 1, R) is orthogonally congruent
to
a diagonal matrix, for
1 0
every t there exist a orthogonal matrix D(t) of the form 0 D1 (t) with D1 (t) ∈ O(n − 1)
such that D(t)XD(t)−1 has the form
a(t) b2 (t) . . . bn (t)
b2 (t) c2 (t) . . .
0
(6.1)
..
.
..
.
.
0
bn (t)
0
. . . cn (t)
Here the only non-zero values of D(t)XD(t)−1 are the numbers a(t), bi (t), ci (t), and ci (t) are
the eigenvalues of the (n − 1)-dimensional lower right block of X(t). Observe that a(t) =
(X(t)e1 , e1 ) and there exist vectors vi (t) ∈ Rn of norm one such that ci (t) = (X(t)vi , vi ). In
particular, since, for all t, X(t) is a positive semidefinite matrix whose eigenvalues are strictly
smaller than one, we deduce 0 ≤ ci (t) < 1 and 0 ≤ a(t) < 1.
We get from Equation 6.1
!−1
n
2
X
det(Id
+
X(t)|
)
b
(t)
he2 ,...,en i
i
x(t) := [(Id + X(t))−1 ]11 =
= 1 + a(t) −
,
det(Id + X(t))
1 + ci (t)
i=2
where the third equality follows by expanding the determinant of Id + X(t) with respect
to the first row. Here we denote by (Id + X(t)|he2 ,...,en i the (n − 1)-dimensional lower right
block of X(t), the choice of the notation is motivated by the fact that we understand the
symmetric matrix Id+X(t) as representing a bilinear form. In this notation we have det(Id+
X(t)|he2 ,...,en i ) = c2 (t) . . . cn (t)
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
Now, assume by contradiction that for all t, x(t) is smaller than
1 > a(t) = x(t)−1 − 1 +
n
X
bi (t)2
≥
1 + ci (t)
i=2
1
2
1
2
+ ρ with ρ <
33
1
.
64n3
Then
1
− 1 > 1 − 4ρ.
+ρ
Since 0 ≤ ci (t) < 1:
kb(t)k2 =
n
X
bi (t)2 ≤ 2
i=2
n
X
bi (t)2
≤
1 + ci (t)
i=2
≤ 2(1 + a(t) − x(t)−1 ) < 2 2 −
1
2
1
+ρ
!
< 8ρ.
Let us now write the matrix M (t) in block form as
m(t) t d(t)
M (t) =
d(t) N (t)
where N (t) is a matrix in Sym(n − 1, R). We have
m(t)2 + kd(t)k2
m(t)t d(t) + t d(t)N (t)
X(t) =
.
N (t)d(t) + m(t)d(t)
d(t)t d(t) + N (t)2
The rest of the proof is devoted to deducing a contradiction from the fact that N (t) is
a monotone path in Sym(n − 1, R), −m(t) is a strictly decreasing function, but N (t) has
almost eigenvectors for the value m(t) (since kb(t)k = kN (t)d(t) + m(t)d(t)k is very small).
In order to make this idea precise, we focus on the subinterval J of ((x, y)) on which m(t)2
is smaller than 1/2 − 4ρ (which is nonempty because m(t) varies continuously between −1
and 1). Then
1
1 − 4ρ < a(t) = m(t)2 + kd(t)k2 < − 4ρ + kd(t)k2 ,
2
which implies that kd(t)k2 > 21 .
Fix an orthonormal basis v2 (t), . . . , vn (t) of eigenvectors for N (t) (these exist since N (t)
is symmetric), where vi (t) is an eigenvector corresponding to the eigenvalue ni (t). Observe
that since M (t) forms a monotone path, also N (t) forms a monotone path in Sym(n, R), and
in particular ni (t) are monotone functions with ni (x) = −1, ni (y) = 1. We can write
d(t) =
n
X
αi (t)vi (t).
i=2
Then
N (t)d(t) + m(t)d(t) =
n
X
αi (t)(ni (t) + m(t))vi (t)
i=2
which implies, taking the norm squared, that:
n
X
αi (t)2 (ni (t) + m(t))2 < 8ρ
i=2
and hence for each i
|αi (t)(ni (t) + m(t))| <
p
8ρ.
34
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
Pn
αi (t)2 = kd(t)k2 > 21 , for every t there is an i such that αi (t) >
n i=2
o
i.e. the sets Ji = t ∈ J|αi (t) > √12n form an open cover of J. But in each Ji
√
|ni (t) + m(t)| < 4 nρ
Note that since
√1 ,
2n
that is, ni (t) and −m(t) are very close. Since ni (t) is strictly increasing and −m(t) is strictly
decreasing, the Ji are connected.
Consider the subinterval I of J where m(t)2 ≤ 41 . We can write it as
I = [x0 , x1 ] ∪ [x1 , x2 ] ∪ · · · ∪ [xk−1 , xk ]
where k ≤ n and for each j there exists an i such that [xj−1 , xj ] ⊂ Ji . We know that
m(xk ) − m(x0 ) = 1; at the same time
m(xk ) − m(x0 ) =
k
X
m(xj ) − m(xj−1 ).
j=1
Since
m(xj ) − m(xj−1 ) = m(xj ) + ci (xj ) − ci (xj ) − m(xj−1 ) + ci (xj−1 ) − ci (xj−1 ) =
√
√
≤ 8 nρ + (ci (xj−1 ) − ci (xj )) < 8 nρ
|
{z
}
<0
√
we have 1 = m(xk ) − m(x0 ) ≤ 8n nρ which gives a contradiction because ρ is smaller than
1
.
64n3
The following Lemma should be understood as an analogue, in the Siegel space, of the fact
that the closest point to i in the horoball {x + iy0 |x ∈ R} ⊂ H2 is iy0 .
Lemma 6.12. For every Z ∈ Sym(n, R) and Y, W ∈ Sym+ (n, R) with Y − W positive
definite, we have:
max ev(Y )
dR (iY, Z + iW ) ≥ log
.
max ev(W )
Proof. We know that
ā+
d (iY, Z + iW )j = log
√ !
rj
√
1 − rj
1+
where 1 ≥ r1 ≥ . . . ≥ rn ≥ 0 are the eigenvalues of the cross-ratio R(iY, Z −iW, Z +iW, −iY ).
Since the function
1+x
x 7→ log
1−x
+
R
ā
is monotone increasing
√ on
(0, 1) and d (iY, Z + iW ) ≥ d (iY, Z + iW )1 , we want to estimate
1+ r
the quantity log 1−√r1 .
1
We set S = Y −1/2 ZY −1/2 and T = Y −1/2 W Y −1/2 . By Lemma 2.13, T is positive definite
and all its eigenvalues are smaller than 1. One can show by explicit computations that
R(iY, Z − iW, Z + iW, −iY ) is conjugate to
Id − 4[S 2 + (T + Id)2 + i(ST − T S)]−1 T = Id − X −1
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
35
where X = 14 T −1/2 [S 2 + (T + Id)2 + i(ST − T S)]T −1/2 . Now r1 is 1 − x−1
1 , where x1 is the
maximum eigenvalue of X.
Note that X is Hermitian, so its maximum eigenvalue is at least the maximum eigenvalue
of its real part: indeed, because Im(X) is antisymmetric, if λ is the maximum eigenvalue of
Re(X) and v is an eigenvector for λ of norm one, we have
x1 ≥ ((Re(X) + i Im(X))v, v) = λ + i (Im(X)v, v) = λ.
As a consequence,
x1 ≥ max ev(Re(X)) ≥ max ev
1 −1
T (T + Id)2
4
=
1 (m + 1)2
,
4
m
where m is the minimum eigenvalue of T . Thus
r1 = 1 −
which implies that
√
1
(m − 1)2
≥
x1
(m + 1)2
r1 ≥
1−m
,
m+1
because m < 1. Using Lemma 2.12, we get
√
1
max ev(Y )
1 + r1
√ ≥
= max ev(T −1 ) ≥ max ev(Y ) min ev(W −1 ) =
.
min ev(T )
max ev(W )
1 − r1
Next we prove that if Y ∈ YC,D has imaginary part with large maximum eigenvalue, the
segment with endpoints θ(C) and θ(D) in [0, `] is long.
Lemma 6.13. If Y ∈ YC,D , then
1
2
max ev(D − C) ≥ max ev(ImY ).
Proof. Up to translating horizontally (as at the beginning of the proof of Proposition 6.11),
we can assume C = −E 2 and D = E 2 , for some matrix E. We know from Lemma 6.10 that
the imaginary part of a point Y ∈ Y−E 2 ,E 2 can be written as E(Id − M 2 )(Id + M 2 )−1 E for
some positive semidefinite matrix M whose eigenvalues are smaller than 1. By Lemma 2.12
we have
max ev(Im Y ) ≤ (max ev(E))2 max ev(Id − M 2 )(Id + M 2 )−1 ≤ max ev(E 2 ).
We can now prove the key step for Theorem 6.4.
Proof of Proposition 6.7. Given a = θ(s) and b = θ(t), with s, t ∈ Λ(Γ), we want to find
a boundary component δα with d − c := dF (pγ (φ(δα− )), pγ (φ(δα+ ))) ≥ λ(b − a), for some
λ independent on a and b. Let z ∈ ((s, t)) given by Proposition
6.11. By Remark 6.3 and
1
e at distance at most log k4
Proposition 6.8, we know that there exist k4 > 1 and w ∈ F T ∂ Σ
from pφ(s),φ(t) (φ(z)). Here F : T 1 H2 → X is the Dρ-equivariant quasi-isometric embedding
studied in Proposition 6.8. Denote by s0 < t0 the points in ((s, t)) such that w ∈ Yφ(s0 ),φ(t0 )
and let c = θ(s0 ) and d = θ(t0 ).
Note that (φ(x), φ(s0 ), φ(t0 ), φ(γ · x)) is maximal, and hence so is (φ(x), φ(s0 ), φ(t0 ), mId)
where m = max ev(φ(γ · x)). But since the translation length of γ is ` and we assumed that
36
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
φ(x) = Id (which implies that all eigenvalues of φ(γ · x) are bigger than one), we can deduce
that m < 2e` and hence
(φ(x), φ(s0 ), φ(t0 ), 2e` Id)
is maximal.
We have [c, d] ⊂ [a, b] ∩ X and
1
log det(φ(t0 )φ(s0 )−1 )
2
(i) 1
max ev(φ(t0 ) − φ(s0 ))
>
2k2
(ii) 1
max ev(Im(w))
≥
k2
(iii) 1
max ev(Im pφ(s),φ(t) (φ(z)))
≥
k2 k4
(iv) k3
>
max ev(φ(t) − φ(s))
k2 k4
(v) 2k1 k3
>
(b − a)
k2 k4
where we apply Lemma 6.9 in (i) and (v), Lemma 6.13 in (ii), Lemma 6.12 in (iii) and
Proposition 6.11 in (iv).
d−c=
∂ H2
t
t0
z
s
φ
pφ(s),φ(t) (φ(z))
s0
w
φ(s)
θ
φ(s0 )
φ(t0 ) φ(t)
0 ac d b `
Figure 5. A schematic picture of the situation described in the proof
7. Geometric Basmajian-type inequalities
This section is dedicated to the proof of Theorem A and to showing that the gap between
the middle and right-hand side term in the inequalities (A1) and (A2) can be arbitrarily
large.
We will need a preliminary lemma concerning the vectorial length of orthotubes.
Lemma 7.1. Given two peripheral elements γ, δ, we have
+
`ā (O(γ, δ))n−i+1
d (pγ (φ(δ )), pγ (φ(δ )))i = 2 log coth
.
2
Proof. Let Ya,b be the R-tube orthogonal to both Yγ and Yδ . Since (φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + ))
is a maximal 4-tuple, by Lemma 2.4 we can assume (up to the action of Sp(2n, R)) that
ā+
−
+
(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) = (−Id, −Λ, Λ, Id),
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
where Λ = diag
1
1
λn , . . . , λ1
37
, with λ1 ≥ . . . ≥ λn > 1.
l∞
O(γ, δ)
−Id
−Λ
0
Λ
Id
Figure 6. A schematic picture of two R-tubes and their orthogeodesic
By Lemma 2.24, Y0,l∞ is the R-tube orthogonal to Yγ and Yδ and it intersects them in iId
and iΛ. As a consequence, by Lemma 2.14
+
`ā (O(γ, δ)) = (log λ1 , . . . , log λn ).
+
Moreover, Lemma 2.25 implies that the distance dā (pγ (φ(δ − )), pγ (φ(δ + ))) is the vector
given by the logarithm of the eigenvalues of R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )). Now using (2.1)
we obtain
R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) = (Id − Λ)−2 (Λ + Id)2 ,
so
ā+
d (pγ (φ(δ − )), pγ (φ(δ + ))) =
log
λn + 1
λn − 1
2
, . . . , log
λ1 + 1
λ1 − 1
2 !
.
7.1. The Finsler metric. We show the Basmajian-type inequality for the Finsler metric
and for a single boundary component:
Theorem 7.2. Let Σ be a compact surface with boundary and γ ∈ Γ = π1 (Σ) a peripheral
element. Given an Anosov maximal representation ρ : Γ → Sp(2n, R), we have
n
X
α∈OΣ (γ)
log coth
+
X
`ā (α)n
≥ `F (ρ(γ)) ≥ n
2
α∈OΣ (γ)
log coth
`F (α)
n
with equality if and only if ρ is the diagonal embedding of a hyperbolization.
This result, together with Remark 4.3, implies the Finsler metric part of Theorem A.
The first step is to relate `F (O(γ, δ)) to the distance between the orthogonal projections
of the extremal points of Yδ on Yγ . For the reader’s convenience, since we will use multiple
properties of log coth(x), Figure 7 shows the graph of this function.
Lemma 7.3. Given two peripheral elements γ, δ, we have
+
`ā (O(γ, δ))n
`F (O(γ, δ))
≥ dF (pγ (φ(δ − )), pγ (φ(δ + ))) ≥ n log coth
,
2
n
with equalities if and only if all the eigenvalues of R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) are equal.
n log coth
38
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
2.5
2.0
1.5
1.0
0.5
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Figure 7. The graph of log coth(x) (drawn using Mathematica)
Proof. We have
n
X
+
`ā (O(γ, δ))i
2
i=1
Pn ā+
` (O(γ, δ))i
≥ n log coth i=1
2n
F
` (O(γ, δ)))
= n log coth
n
dF (pγ (φ(δ − )), pγ (φ(δ + ))) =
log coth
where the first equality is given by Lemma 7.1 and the inequality follows from the convexity
of log coth(x). On the other hand by the monotonicity of log coth(x)
F
−
+
d (pγ (φ(δ )), pγ (φ(δ ))) =
n
X
i=1
+
+
`ā (O(γ, δ))i
`ā (O(γ, δ))n
log coth
≤ n log coth
.
2
2
Since log coth is strictly convex and strictly monotone, there are equalities if and only if
+
+
`ā (O(γ, δ))i = `ā (O(γ, δ))j for all i, j. Because log coth(x) is strictly decreasing, this is also
equivalent to the fact that, for each pair i, j,
+
+
dā (pγ (φ(δ − )), pγ (φ(δ + ))i = dā (pγ (φ(δ − )), pγ (φ(δ + ))j ,
namely that all eigenvalues of R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) are equal (Lemma 2.25).
In order to prove that the equalities characterize diagonal representations we need the
following lemma:
Lemma 7.4. Assume that an equality in Theorem 7.2 holds. Then π SL ◦ pγ ◦ φ is constant
on Λ(Γ) \ {γ − , γ + }.
Proof. Denote by Φγ,SL the map π SL ◦ pγ ◦ φ and by Φγ,R the map π R ◦ pγ ◦ φ. We will prove
that for each x, y ∈ Λ(Γ) \ {γ − , γ + } we have Φγ,SL (x) = Φγ,SL (y).
Observe that since, by assumption, an equality holds in Theorem 7.2, we have necessarily
that the corresponding equality in Lemma 7.3 is satisfied. If in particular, δ ∈ Γ is such
that (x, δ + , δ − , y) is positively oriented, as a consequence of Lemma 7.3 we have that all
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
39
eigenvalues of R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) are equal. Since R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + ))
is conjugated to a symmetric matrix, this means
R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) = λId
for some λ. Up to the action of the symplectic group, we can assume
(φ(γ − ), pγ φ(δ + ), pγ φ(δ − ), φ(γ + )) = (0, iId, iA, l∞ )
for some A ∈ Sym+ (n, R). Under these assumptions, by [BP17, Lemma 4.3],
R(φ(γ − ), φ(δ + ), φ(δ − ), φ(γ + )) = A.
Thus A = λId, which implies that Φγ,SL (δ − ) = Φγ,SL (δ + ) = Id.
In particular
dSL (Φγ,SL (x), Φγ,SL (y)) ≤ dSL (Φγ,SL (x), Φγ,SL (δ + )) + dSL (Φγ,SL (δ − ), Φγ,SL (y))
√
≤ n − 1 Φγ,R (δ + ) − Φγ,R (x) + Φγ,R (y) − Φγ,R (δ − ) ,
where the last inequality is a consequence of Lemma 2.20. By repeating the argument, we
obtain the same inequality for any finite sum of δ. Taking the limit we get
X
√
Φγ,R (δ − ) − Φγ,R (δ + )
dSL (Φγ,SL (x), Φγ,SL (y)) ≤ n − 1 Φγ,R (y) − Φγ,R (x) −
δ:((δ + ,δ − ))⊂((x,y))
Using Remark 2.22 and Theorem 6.4, we deduce
2
Φγ,R (y) − Φγ,R (x) = √ dF (pγ φ(x), pγ φ(y))
n
X
2
dF (pγ φ(δ − ), pγ φ(δ + ))
=√
n
δ:((δ + ,δ − ))⊂((x,y))
X
=
Φγ,R (δ − ) − Φγ,R (δ + ) ,
δ:((δ + ,δ − ))⊂((x,y))
so dSL (Φγ,SL (x), Φγ,SL (y)) = 0, which means that Φγ,SL (x) = Φγ,SL (y).
Proof of Theorem 7.2. Theorem 6.4 and Proposition 7.3 imply that both inequalities in Theorem 7.2 hold.
If ρ is the diagonal embedding of a hyperbolization, then the inequalities are identities by
the classical Basmajian’s identity. Let us then prove the reverse implication. Observe that
the boundary of a diagonal disk is given by
∂Diag = {l∞ } ∪ {λId | λ ∈ R} ⊂ L(R2n ).
So as a consequence of Lemma 7.4 we get that, up to conjugating the representation in
Sp(2n, R), the image φ(Λ(Γ)) is contained in ∂Diag. This implies that ρ(Γ) preserves ∂Diag
and hence
ρ(Γ) ⊂ StabSp(2n,R) (∂Diag).
It can be directly checked that
aX bX
a b
StabSp(2n,R) (∂Diag) =
∈ SL(2, R), X ∈ O(n) ' SL(2, R) × O(n),
cX dX
c d
40
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
which finishes the proof.
7.2. The Riemannian metric. We begin the section deducing from Theorem 6.4 an inequality for the Riemannian distances.
Proposition 7.5. Let ρ : π1 (Σ) → Sp(2n, R) be an Anosov maximal representation, for Σ
with nonempty boundary. For every peripheral element γ we have:
X
dR (pγ (φ(δα− )), pγ (φ(δα+ ))).
`R (γ) ≤
α∈OΣ (γ)
Proof. Consider the peripheral elements δ such that ((δ + , δ − )) ⊂ ((x, γx)) for some fixed
x ∈ Λ(Γ). As they are countably many, we can index them as {δi }i∈N . For every k ∈ N,
consider {δ1 , . . . , δk }. Up to reordering them, assume that δ1+ , δ1− , . . . , δk+ , δk− are positively
oriented. Then `R (γ) ≤ Ak + Bk , where
Ak :=
k
X
dR (pγ (φ(δi+ )), pγ (φ(δi− )))
i=1
and
R
Bk := d
(x, pγ (φ(δ1+ )))
+
k−1
X
+
dR (pγ (φ(δi− )), pγ (φ(δi+1
))) + dR (pγ (φ(δk− )), γx).
i=1
By definition, as k → ∞
X
Ak →
dR (pγ (φ(δ − )), pγ (φ(δ + ))) =
δ:(δ + ,δ − )⊂(x,γx)
X
dR (pγ (φ(δα− )), pγ (φ(δα+ ))).
α∈OΣ (γ)
So it is enough to show that Bk → 0 as k → ∞. For every k, by Lemma 2.10 we have
Bk ≤ 2BkF
where
BkF = dF (x, pγ (φ(δ1+ ))) +
k−1
X
−
dF (pγ (φ(δi+ )), pγ (φ(δi+1
))) + dF (pγ (φ(δk− )), γx).
i=1
By Theorem 6.4 BkF tends to zero as k tends to infinity.
As in the Finsler case, the inequalities (A2) follow from a computation for each boundary
component (Theorem 7.6) and Remark 4.3.
Theorem 7.6. For any Anosov maximal representation ρ : Γ → Sp(2n, R) and any peripheral element γ ∈ Γ we have
√
2 n
X
+
√ X
`ā (α)n
`R (α)
log coth
≥ `R (γ) ≥ 2 n
log coth √
2
2 n
α∈OΣ (γ)
α∈OΣ (γ)
with equalities if and only if ρ is the diagonal embedding of a hyperbolization.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
41
Proof. We have
`R (γ) ≥
√ X
√ X
2`F (γ)
`F (α)
`R (α)
√
≥2 n
log coth
log coth √ ,
≥2 n
n
n
2 n
α∈OΣ (γ)
α∈OΣ (γ)
where the first and last inequalities follow from Lemma 2.10 and the middle one is given
by Theorem 7.6. Note that the first inequality has also a geometric interpretation: since
ρ(γ) preserves the tube Yγ , the Riemannian translation length of ρ(γ), `R (γ) is at most the
translation length in the tube, that, in turn, is at most the translation length in the Euclidean
factor R.
For the other inequality, we have:
+
X
X
√
`ā (α)n
R
−
+
R
d (pγ (φ(δα )), pγ (φ(δα ))) ≤ 2 n
` (γ) ≤
log coth
2
α∈OΣ (γ)
α∈OΣ (γ)
where the first inequality is Proposition 7.5 and the second follows from Lemma 7.1 and the
monotonicity of log coth(x).
If ρ is the diagonal embedding of a hyperbolization, then the equalities holds as a consequence of the classical Basmajian equality. On the other hand assume that some equality
holds. Then the corresponding equality holds in Theorem 7.2 and hence ρ is a diagonal
embedding.
7.3. Extremal cases. In this section we show that there are sequences of maximal representations in which the length of the boundary components stay bounded away from zero,
but the R-tubes associated to any two peripheral elements are arbitrarily far apart.
Proposition 7.7. For any n ≥ 2, L > 0 and η > 0, there is a maximal representation
ρ : Γ0,n+1 → Sp(2n, R) and a peripheral element γ such that
√
nL
and `R (γ) = nL
`F (γ) =
2
while
X
X
√
`F (α)
`R (α)
n
log coth
< η and 2 n
log coth √ < η.
n
2 n
α∈OΣ (γ)
α∈OΣ (γ)
The following lemma from hyperbolic geometry will be useful to prove Proposition 7.7:
Lemma 7.8. For every L > 0, ε > 0, and n ∈ N, there exists a hyperbolic structure on a
sphere with n + 1 boundary components such that one boundary component γ0 has length L
and there is an orthogeodesic α between γ0 and another boundary component satisfying
2 log coth
`(α)
= L − ε.
2
Proof. Explicitly, we want α of length
e(L−ε)/2 + 1
.
e(L−ε)/2 − 1
It is enough to construct a pair of pants with one boundary component of length L and the
orthogonal between this boundary and another one of length g(L, ε). Equivalently, we can
construct a right-angled hexagon with one side of length L2 and an adjacent one of length
g(L, ε) = log
42
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
y
x
L
2
g(L, ε)
Figure 8. The hexagon we want to construct
g(L, ε). By properties of hyperbolic hexagons (see [Bus10, Section 2.4]), it is enough to find
positive x and y satisfying
L
L
cosh y = sinh sinh x cosh g(L, ε) − cosh cosh x.
2
2
If we find such x and y, there is a right-angled hyperbolic hexagon with non-consecutive
sides of length L2 , x and y (and the side between the ones of length L2 and x has length
g(L, ε)). But since cosh is bijective between (0, ∞) and (1, ∞), it is enough to show that
there exists x > 0 such that
L
L
sinh sinh x cosh g(L, ε) − cosh cosh x > 1. (?)
2
2
Now there is some positive η such that
cosh g(L, ε) =
eL−ε + 1
eL + 1
cosh(L/2)
=
+η =
+ η.
L−ε
L
e
−1
e −1
sinh(L/2)
Using this, inequality (?) becomes
L
L
> (cosh x − sinh x) cosh + 1.
2
2
Let x0 be such that cosh x − sinh x < 1; then it is enough to choose x > x0 satisfying
η sinh x sinh
sinh x >
1 + cosh L2
δ sinh L2
.
Proof of Proposition 7.7. Choose 0 < ε = nη2 . We write the fundamental group Γ0,n+1 of an
(n + 1)-punctured sphere as
D
E
Y
Γ0,n+1 = γ0 , . . . , γn
γi = 1 .
Consider n hyperbolizations ρi : Γ0,n+1 → SL(2, R) so that ρi (γ0 ) has length L and the
orthogeodesic αi between the axis of γ0 and the axis of γi satisfies the hypothesis of Lemma
7.8. Consider the representation ρ = diag(ρ1 , . . . , ρn ). Then
√
nL
and `R (γ0 ) = nL.
`F (γ0 ) =
2
The classical Basmajian’s identity tells us that
X
`ρ (α)
2
log cosh i
= `ρi (γ0 ) = L.
2
α
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
` (α )
Using the fact that the left-hand side term is bigger than 2 log cosh ρi 2 j + 2 log cosh
for any j 6= i and the assumption on `ρi (αi ) we can deduce that for any j 6= i
43
`ρi (αi )
2
`ρi (αj ) > `ρi (αi ) = `ρj (αj )
so
min `ρi (αj ) 6= `ρj (αj ).
i
Look now at each term log coth
have
log coth
`F (α)
n
: by convexity and monotonicity of log coth(x), we
n
n
i=1
i=1
X `ρ (α)
`ρ (α)
`F (α)
1X
i
= log coth
≤
log coth i
n
2n
n
2
`ρ (α)
`ρ (α)
≤ max log coth i
= log coth min i
.
i
i
2
2
If α = αj , we know thus that
log coth
`ρ (α)
`F (αj ) X
≤
log coth i
n
2
i6=j
while if α 6= αj for all j we just consider that
n
log coth
`ρ (α)
`F (α) X
≤
.
log coth i
n
2
i=1
Using these inequalities and reordering the terms we get
n
X
X
X
`
(α)
`F (α)
ρ
.
≤n
n
log coth
log coth i
n
2
i=1
α∈OΣ (γ0 )
α6=αi
On the other hand, the classical Basmajian’s identity tells us that
X
2 log coth
α6=αi
`ρi (α)
`ρ (αi )
≤ `(γ0 ) − 2 log coth i
= ε.
2
2
and this implies that
√
2 n
X
α∈OΣ (γ0 )
log coth
`R (α)
√ ≤ 2n
2 n
X
α∈OΣ (γ0 )
log coth
`F (α)
≤ n2 ε < η,
n
where the first inequality follows from Lemma 2.10.
7.4. Proof of Corollary 1.3. We conclude the paper with a proof of the geometric consequence of Theorem A mentioned in the introduction:
Proof of Corollary 1.3. We can assume that γ is a boundary component, otherwise we cut
the surface along the corresponding curve and consider the restriction of the representation
44
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
to the fundamental groups of the surface(s) obtained. Note that Theorem 7.6 implies that,
for any orthotube α ∈ OΣ (γ),
R
√
` (γ)
R
√
=: 2w(γ).
` (α) > 2 n arccoth exp
2 n
Suppose, by contradiction, that the neighborhood C(γ) = hρ(γ)i\Nw(γ) (Yγ ) is not embedded.
This implies that there is an element δ ∈ Γ \ hγi such that C(γ) ∩ δC(γ) 6= ∅. Looking at the
universal cover X , this means that
dR (Yγ , Yδγδ−1 ) < 2w(γ)
and hence that there is an orthotube of length less than 2w(γ), a contradiction.
Similarly, consider another element δ corresponding to a simple closed curve or boundary
component disjoint from the curve represented by γ. If by contradiction the two neighborhoods intersect, we get that
dR (Yγ , Yδ ) < w(γ) + w(δ) ≤ 2 max w(γ), w(δ)
so we have again an orthogeodesic (seen as element of Oγ (Σ) or Oδ (Σ), depending on whether
w(γ) or w(δ) is maximum) which is too short, a contradiction.
References
[Bas93]
Ara Basmajian. The orthogonal spectrum of a hyperbolic manifold. Amer. J. Math., 115(5):1139–
1159, 1993.
[BCLS15] Martin Bridgeman, Richard Canary, François Labourie, and Andres Sambarino. The pressure metric
for Anosov representations. Geom. Funct. Anal., 25(4):1089–1179, 2015.
[Ben97] Y. Benoist. Propriétés asymptotiques des groupes linéaires. Geom. Funct. Anal., 7(1):1–47, 1997.
[BILW05] Marc Burger, Alessandra Iozzi, François Labourie, and Anna Wienhard. Maximal representations of
surface groups: symplectic Anosov structures. Pure Appl. Math. Q., 1(3, Special Issue: In memory
of Armand Borel. Part 2):543–590, 2005.
[BIW09] Marc Burger, Alessandra Iozzi, and Anna Wienhard. Tight homomorphisms and Hermitian symmetric spaces. Geom. Funct. Anal., 19(3):678–721, 2009.
[BIW10] Marc Burger, Alessandra Iozzi, and Anna Wienhard. Surface group representations with maximal
Toledo invariant. Ann. of Math. (2), 172(1):517–566, 2010.
[BK10]
Martin Bridgeman and Jeremy Kahn. Hyperbolic volume of manifolds with geodesic boundary and
orthospectra. Geom. Funct. Anal., 20(5):1210–1230, 2010.
[BP17]
Marc Burger and Maria Beatrice Pozzetti. Maximal representations, non-Archimedean Siegel
spaces, and buildings. Geom. Topol., 21(6):3539–3599, 2017.
[Bri11]
Martin Bridgeman. Orthospectra of geodesic laminations and dilogarithm identities on moduli
space. Geom. Topol., 15(2):707–733, 2011.
[BT17]
Jean-Philippe Burelle and Nicolaus Treib. Schottky groups and maximal representations. Geometriae Dedicata, Oct 2017.
[Bus10] Peter Buser. Geometry and spectra of compact Riemann surfaces. Modern Birkhäuser Classics.
Birkhäuser Boston, Inc., Boston, MA, 2010. Reprint of the 1992 edition.
[Ebe96] Patrick B. Eberlein. Geometry of nonpositively curved manifolds. Chicago Lectures in Mathematics.
University of Chicago Press, Chicago, IL, 1996.
[Gol80]
William Mark Goldman. Discontinuous groups and the Euler class. ProQuest LLC, Ann Arbor, MI,
1980. Thesis (Ph.D.)–University of California, Berkeley.
[GW12] O. Guichard and A. Wienhard. Anosov representations: domains of discontinuity and applications.
Invent. Math., 190(2):357–438, 2012.
BASMAJIAN–TYPE INEQUALITIES FOR MAXIMAL REPRESENTATIONS
[HS12]
[Kee74]
[Lab07]
[Lab08]
[LM09]
[Loo77]
[LT14]
[LV80]
[LZ17]
[McS98]
[Mir07]
[MZ16]
[Par]
[Par12]
[Sie43]
[Str15]
[Tuk84]
[VY17a]
[VY17b]
[Wie04]
[Wol72]
[Xu16]
45
Tobias Hartnick and Tobias Strubel. Cross ratios, translation lengths and maximal representations.
Geom. Dedicata, 161:285–322, 2012.
Linda Keen. Collars on Riemann surfaces. In Discontinuous groups and Riemann surfaces (Proc.
Conf., Univ. Maryland, College Park, Md., 1973), pages 263–268. Ann. of Math. Studies, No. 79.
Princeton Univ. Press, Princeton, N.J., 1974.
François Labourie. Cross ratios, surface groups, PSL(n, R) and diffeomorphisms of the circle. Publ.
Math. Inst. Hautes Études Sci., (106):139–213, 2007.
François Labourie. Cross ratios, Anosov representations and the energy functional on Teichmüller
space. Ann. Sci. Éc. Norm. Supér. (4), 41(3):437–469, 2008.
François Labourie and Gregory McShane. Cross ratios and identities for higher TeichmüllerThurston theory. Duke Math. J., 149(2):279–345, 2009.
O. Loos. Bounded symmetric domains and Jordan pairs. Mathematical Lectures. Department of
Mathematics, University of California at Irvine, 1977.
Feng Luo and Ser Peow Tan. A dilogarithm identity on moduli spaces of curves. J. Differential
Geom., 97(2):255–274, 2014.
G. Lion and M. Vergne. The Weil representation, Maslov index and theta series, volume 6 of
Progress in Mathematics. Birkhäuser, Boston, Mass., 1980.
Gye-Seon Lee and Tengren Zhang. Collar lemma for Hitchin representations. Geom. Topol.,
21(4):2243–2280, 2017.
Greg McShane. Simple geodesics and a series constant over Teichmuller space. Invent. Math.,
132(3):607–632, 1998.
Maryam Mirzakhani. Simple geodesics and Weil-Petersson volumes of moduli spaces of bordered
Riemann surfaces. Invent. Math., 167(1):179–222, 2007.
G. Martone and T. Zhang. Positively ratioed representations. ArXiv e-prints, September 2016.
A. Parreau. La distance vectorielle dans les immeubles affines et les espaces symmetriques. In
preparation.
Anne Parreau. Compactification d’espaces de représentations de groupes de type fini. Math. Z.,
272(1-2):51–86, 2012.
Carl Ludwig Siegel. Symplectic geometry. Amer. J. Math., 65:1–86, 1943.
Tobias Strubel. Fenchel-Nielsen coordinates for maximal representations. Geom. Dedicata, 176:45–
86, 2015.
Pekka Tukia. The hausdorff dimension of the limit set of a geometrically finite kleinian group. Acta
Mathematica, 152(1):127–140, 1984.
Nicholas G. Vlamis and Andrew Yarmola. Basmajian’s identity in higher teichmllerthurston theory.
Journal of Topology, 10(3):744–764, 2017.
Nicholas G. Vlamis and Andrew Yarmola. The bridgeman–kahn identity for hyperbolic manifolds
with cusped boundary. Geometriae Dedicata, Jul 2017.
Anna Katharina Wienhard. Bounded cohomology and geometry. Bonner Mathematische Schriften
[Bonn Mathematical Publications], 368. Universität Bonn, Mathematisches Institut, Bonn, 2004.
Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn, Bonn, 2004.
Joseph A. Wolf. Fine structure of Hermitian symmetric spaces. In Symmetric spaces (Short Courses,
Washington Univ., St. Louis, Mo., 1969–1970), pages 271–357. Pure and App. Math., Vol. 8.
Dekker, New York, 1972.
B. Xu. Incompleteness of pressure metric on Teichmüller space of a bordered surface. ArXiv e-prints,
August 2016.
46
FEDERICA FANONI AND MARIA BEATRICE POZZETTI
(Federica Fanoni) Mathematical Institute, University of Heidelberg, Heidelberg, Germany
E-mail address: [email protected]
(Maria Beatrice Pozzetti) Mathematical Institute, University of Heidelberg, Heidelberg, Germany
E-mail address: [email protected]
| 4math.GR
|
arXiv:1308.3604v4 [math.GR] 21 Dec 2016
An approximation principle for congruence
subgroups
Tobias Finis
Erez Lapid
December 22, 2016
Abstract
The motivating question of this paper is roughly the following: given a flat group
scheme G over Zp , p prime, with semisimple generic fiber GQp , how far are open
subgroups of G(Zp ) from subgroups of the form X(Zp )Kp (pn ), where X is a subgroup scheme of G and Kp (pn ) is the principal congruence subgroup Ker(G(Zp ) →
G(Z/pn Z))? More precisely, we will show that for GQp simply connected there exist constants J ≥ 1 and ε > 0, depending only on G, such that any open subgroup
of G(Zp ) of level pn admits an open subgroup of index ≤ J which is contained in
X(Zp )Kp (p⌈εn⌉ ) for some proper, connected algebraic subgroup X of G defined
over Qp . Moreover, if G is defined over Z, then ε and J can be taken independently
of p.
We also give a correspondence between natural classes of Zp -Lie subalgebras of
gZp and of closed subgroups of G(Zp ) that can be regarded as a variant over Zp of
Nori’s results on the structure of finite subgroups of GL(N0 , Fp ) for large p [Nor87].
As an application we give a bound
Q for the volume of the intersection of a conjugacy class in the group G(Ẑ) = p G(Zp ), for G defined over Z, with an arbitrary
open subgroup. In a companion paper, we apply this result to the limit multiplicity
problem for arbitrary congruence subgroups of the arithmetic lattice G(Z).
Keywords. Lattices in Lie groups, uniform pro-p groups, Lie algebras
Contents
1 Introduction
2
Tobias Finis: Universität Leipzig, Mathematisches Institut, Postfach 100920, D-04009 Leipzig, Germany; e-mail: [email protected]
Erez Lapid: Department of Mathematics, The Weizmann Institute of Science, Rehovot 7610001, Israel;
e-mail: [email protected]
Mathematics Subject Classification (2010): Primary 20G25; Secondary 22E40, 22E60
1
2
Tobias Finis, Erez Lapid
2 The pro-p approximation theorem
4
3 The approximation theorem for open subgroups of G(Zp )
20
4 Closed subgroups of G(Zp ) and closed subalgebras of gZp
29
5 Intersections of conjugacy classes and open compact subgroups
42
A Bounds for the number of solutions of polynomial congruences
61
1 Introduction
In this paper we study closed subgroups of the profinite groups G(Zp ), where G is a
semisimple algebraic group defined over Q with a fixed Z-model, and p is a prime. The
first main topic, which is treated in §§2 and 3, is the comparison between arbitrary open
subgroups H of G(Zp ) and special open subgroups of the form (X(Qp ) ∩ G(Zp ))Kp (pn ),
where X is a proper, connected algebraic subgroup of G defined over Qp , and Kp (pn ) =
Ker(G(Zp ) → G(Z/pn Z)) is the principal congruence subgroup of G(Zp ) of level pn .
These special open subgroups are ubiquitous in the literature, especially in the case where
X is a parabolic subgroup of G,
To motivate our theorem, recall that commensurability classes of closed subgroups of
G(Zp ) are in one-to-one correspondence with Qp -Lie subalgebras of gQp = LieQp G (cf.
[DdSMS99, Theorem 9.14]), and that by Chevalley’s theorem [Bor91, Corollary 7.9] any
proper, Qp -Lie subalgebra of gQp is contained in the Lie algebra of a proper, connected
algebraic subgroup X of G defined over Qp . These statements do not say anything nontrivial about open subgroups. Recall that the level of an open subgroup H of G(Zp ) is
defined as min{l = pn ≥ 1 : Kp (l) ⊂ H}. Our first main result (Theorem 3.1) says
that if G is simply connected then there exist constants J ≥ 1 and ε > 0, depending
only on G, such that any open subgroup of G(Zp ) of level pn admits an open subgroup of
index ≤ J which is contained in (X(Qp ) ∩ G(Zp ))Kp (p⌈εn⌉ ) for some proper, connected
algebraic subgroup X of G defined over Qp .
Simple considerations show that it is not possible to take ε = 1 here, and that in fact
necessarily ε ≤ 12 for every G (cf. Remark 2.3 and Lemma 2.20). The question of the
optimal value of ε for a given G (or of the optimal asymptotic value as n → ∞) remains
unresolved except in the case of G = SL(2) (cf. Lemma 2.21).
We also have a variant of Theorem 3.1 (Theorem 2.2) for subgroups of the pro-p
groups Kp (p′ ), where p′ = p if p is odd and p′ = 4 if p = 2. Here G does not need
to be simply connected, and we do not need to pass to a finite index subgroup. In fact,
we first prove Theorem 2.2 in §2 and then modify our arguments to deal with arbitrary
Approximation principle
3
open subgroups in §3. Our proofs are based on the general correspondence between (a
large class of) subgroups of Kp (p′ ) and Zp -Lie subalgebras of p′ gZp . This correspondence
reduces Theorem 2.2 to a Zp -Lie algebra analog (Theorem 2.12). The latter is deduced
from a general lifting result of M. Greenberg [Gre74], which in the case at hand allows
one (under natural conditions) to lift Lie subalgebras modulo pν to Lie subalgebras in
characteristic zero. To infer Theorem 3.1 from Theorem 2.2 we use in addition the results
of Nori [Nor87] on the structure of finite subgroups of G(Fp ) for large p.
The discussion above spurs the next result (§4) which is a natural correspondence
(valid for all p large enough with respect to G) between the set of Zp -Lie subalgebras of
gZp that are generated by residually nilpotent elements and the set of closed subgroups
of G(Zp ) that are generated by residually unipotent elements. Note that the analogous
correspondence over Fp is a consequence of Nori’s results, which not surprisingly, play
an important role in the proof of the Zp -analog. In addition, our theorem generalizes (and
relies on) the previously known correspondence between closed pro-p subgroups and Lie
subalgebras consisting of residually nilpotent elements, which is given by the application
of the logarithm and exponential maps [Ila95, Klo05, GSK09]. Apart from these results,
our proof of the correspondence uses Steinberg’s algebraicity theorem [Ste68, Theorem
13.3] and Jantzen’s semisimplicity theorem [Jan97] for characteristic p representations
of the groups X(Fp ), where X is semisimple and simply connected over Fp , as well as
the Curtis–Steinberg–Tits presentations of these groups [Ste62, Cur65, Ste81, Tit81]. In
a somewhat different direction, results on closed subgroups of G(Zp ) connected to Nori’s
theorems have been obtained by Larsen [Lar10].
As an application of the approximation theorems, we give in §5 a bound for the volQ
ume of the intersection of a conjugacy class in the profinite group G(Ẑ) = p G(Zp )
with an arbitrary open subgroup. (The results of §4 are not logically necessary for either
Theorem 3.1 or this application.) The main technical result is Theorem 5.3 and the bound
in question is given in Corollary 5.8 (which is stated for a slightly more general class of
groups G). In Corollary 5.9, we give a variant of this result which deals with arithmetic
lattices in semisimple Lie groups. We note that essentially the same result has been obtained independently in [ABB+ , §5] by a different method. A rough form of the resulting
bound can be stated as follows. For an arbitrary group Γ, a finite index subgroup ∆, and
an element γ ∈ Γ set
c∆ (γ) = |{δ ∈ Γ/∆ : δ −1 γδ ∈ ∆}|,
which is also the number of fixed points of γ in the permutation representation of Γ on
the finite set Γ/∆. Let G be semisimple and simply connected and assume that for no
Q-simple factor H of G the group H(R) is compact. Let K = G(Ẑ) ⊂ G(Afin ) and let
Γ = G(Z) = G(Q) ∩ K, which is an arithmetic lattice in the connected semisimple Lie
group G(R). For any open subgroup K ⊂ K let ∆ = G(Q) ∩ K be the associated finite
Tobias Finis, Erez Lapid
4
index subgroup of Γ. By definition, these subgroups are the congruence subgroups of Γ.
Then there exists a constant ε > 0, depending only on Γ, such that for all congruence
subgroups ∆ of Γ and all γ ∈ Γ that are not contained in any proper normal subgroup of
G (which we may assume to be defined over Q) we have
c∆ (γ)
≪Γ,γ [Γ : ∆]−ε .
[Γ : ∆]
Moreover, we can also control the dependence of the implied constant on γ very explicitly.
In the companion paper [FL15] we will use this result to extend the (partly conditional)
results of [FLM15] about limit multiplicities for subgroups of the lattice Γ from the case
of principal congruence subgroups to arbitrary congruence subgroups.
The appendix contains some elementary bounds on the number of solutions of polynomial congruences that are needed in §5. These results are probably well known but we
were unable to retrieve them in the literature in a form convenient for us.
It would be of interest to study the case of groups defined over function fields over
finite fields and over the corresponding local fields (cf. [LS94, BS97, ANS03]). We do
not say anything about this here.
We thank the Centre Interfacultaire Bernoulli, Lausanne, and the Max Planck Institute
for Mathematics, Bonn, where a part of this paper was worked out. We thank Yiftach
Barnea, Vladimir Berkovich, Tsachik Gelander, Ehud Hrushovski, Michael Larsen, Alex
Lubotzky and Aner Shalev for useful discussions and interest in the subject matter of this
paper.
2 The pro-p approximation theorem
2.1 The general setup
Throughout §§2 – 4 (except for §2.4, where we consider a more general situation) let
G = GQ be a connected semisimple algebraic group defined over Q, together with a
fixed embedding ρ0 : G → GL(N0 ) defined over Q. (Actually, only the image of ρ0
inside GL(N0 ) is relevant for our purposes.) The choice of ρ0 endows G with the structure of a flat group scheme over Z, which we denote by GZ . Concretely, it is given as
Spec Z[GL(N0 )]/(I ∩ Z[GL(N0 )]), where R[GL(N0 )] = R[y, xij , i, j = 1, . . . , N0 ]/(1 −
y det x) for any ring R and I is the ideal of Q[GL(N0 )] defining the image of ρ0 . If
R is a ring which is flat over Z, i.e., if the total quotient ring Q(R) is a Q-algebra,
then G(R) = {g ∈ G(Q(R)) : ρ0 (g) ∈ GL(N0 , R)}. For any commutative ring R
let GR = GZ × Spec R be the base extension to R of the group scheme GZ . If R is a field
of characteristic zero, then GR is a connected semisimple group over R. For almost all
primes p, the group scheme GZp is smooth over Zp and GFp is a connected semisimple
group defined over Fp [Tit79, §3.8.1, §3.9.1].
Approximation principle
5
Note that we can always choose ρ0 such that G is smooth over Z, and therefore smooth
over Zp for all p. (Indeed, by [ibid., §3.4.1], for each p there exists a smooth group scheme
over Zp with generic fiber GQp , and we can glue these as in [Gro96] to obtain a smooth
Z-model for G.) However, only under strong restrictions on G it is possible to have G
smooth over Z and GFp connected semisimple for all p [Gro96].
For each prime p we set
Kp := G(Zp ) = {g ∈ G(Qp ) : ρ0 (g) ∈ GL(N0 , Zp )},
a compact open subgroup of G(Qp ). The principal congruence subgroups of Kp are defined by
Kp (pn ) := Ker (G(Zp ) → G(Z/pn Z)) = {g ∈ Kp : ρ0 (g) ≡ 1
(mod pn )},
n ≥ 1.
They are clearly pro-p normal subgroups of Kp , and form a neighborhood base of the
identity element. For n = 0 we set Kp (1) := Kp .
Definition 2.1. Let H be an open subgroup of Kp . The level of H is
min{q = pn : Kp (q) ⊂ H}.
The main result of this section is the following theorem on open subgroups of the prop groups Kp (p′ ) where p′ = p if p is odd and p′ = 4 if p = 2. As usual, we denote by ⌊x⌋
(resp., ⌈x⌉) the largest integer ≤ x (resp., the smallest integer ≥ x).
Theorem 2.2. There exists a constant ε > 0, depending only on G and ρ0 , such that for
any prime p and any open subgroup H ⊂ Kp (p′ ) of level pn there exists a proper, connected algebraic subgroup X of G defined over Qp such that H ⊂ (X(Qp )∩Kp (p′ ))Kp (p⌈εn⌉ ).
Alternatively, we can reformulate the theorem as a statement on the projective system
consisting of the finite groups G(Z/pn Z) and the canonical reduction homomorphisms
πn,m : G(Z/pn Z) → G(Z/pm Z), n ≥ m. Note that Kp = lim G(Z/pn Z). We write
←−
π∞,n for the reduction homomorphism Kp → G(Z/pn Z). A subgroup H̄ of G(Z/pn Z)
is called essential if it is not the pullback via πn,n−1 of a subgroup of G(Z/pn−1 Z), i.e. if
H̄ does not contain the kernel of πn,n−1 . Then an equivalent form of Theorem 2.2 is the
following
Theorem 2.2′ . There exists a constant ε > 0, depending only on G and ρ0 , such that for
any prime p, an integer n ≥ 1 and an essential subgroup H̄ of π∞,n (Kp (p′ )) ⊂ G(Z/pn Z)
there exists a proper, connected algebraic subgroup X of G defined over Qp such that
πn,⌈εn⌉ (H̄) ⊂ π∞,⌈εn⌉ (X(Qp ) ∩ Kp (p′ )).
Tobias Finis, Erez Lapid
6
Note here that if GZp is smooth (which is the case for almost all p) then the p-group
π∞,n (Kp (p′ )) coincides with the kernel of the homomorphism πn,ǫp : G(Z/pn Z) →
G(Z/p′ Z) where ǫp = 1 if p is odd and ǫ2 = 2.
Remark 2.3. In the statement of Theorem 2.2 there is no harm in assuming that n ≥ n0
for some fixed positive integer n0 depending only on G. In fact, for any integer n ≥ 1
let ν(n) be the largest integer ν such that for any p and any open subgroup H ⊂ Kp (p′ )
of level pn there exists a proper, connected algebraic subgroup X of G defined over Qp
such that H ⊂ (X(Qp ) ∩ Kp (p′ ))Kp (pν ). Then the content of Theorem 2.2 is that εG :=
lim inf n→∞ ν(n)
> 0. It is natural to ask what further can be said about εG . It is easy to
n
see that εG ≤ 21 for any G (see Lemma 2.20 below). Thus, the introduction of the factor
Kp (p⌈εn⌉ ) instead of Kp (pn ) is essential for Theorem 2.2 to hold. For G = SL(2) we
have εG = 21 (see Lemma 2.21 below). With our current knowledge we cannot rule out the
possibility that εG = 12 for all G. However, our method of proof falls short of even proving
that εG is bounded away from 0 independently of G.
Theorem 2.2 will be proved in §2.3 below.
2.2 Uniform pro-p groups
A key ingredient in the proof of Theorem 2.2 is a linearization argument. In order to carry
it out, we need to recall some further structural properties of the congruence subgroups
Kp (pn ). We first recall some standard terminology and basic results from [DdSMS99]
which have their root in the fundamental work of Lazard [Laz65].
For any group H and any integer n we write H {n} = {xn : x ∈ H} and denote by H n
the group generated by H {n} . We will use the standard notation [x, y] = xyx−1 y −1 for
the commutator. Also, for any subgroups H1 , H2 of H we write [H1 , H2] for the subgroup
generated by the commutators [h1 , h2 ], h1 ∈ H1 , h2 ∈ H2 .
In the following, all pro-p groups that we consider will be implicitly assumed to be
topologically finitely generated.
Definition 2.4. Let L be a pro-p group.
′
1. L is called powerful if [L, L] ⊂ Lp .
2. L is called uniformly powerful (or simply uniform) if it is powerful and torsion-free
(cf. [ibid., Theorem 4.5]).
3. The rank of L is the smallest integer n such that every closed (or alternatively by
[ibid., Proposition 3.11], every open) subgroup of L can be topologically generated
by n elements.3
3
We say that L is of infinite rank if no such integer exists. In this paper, we will only work with pro-p
groups of finite rank.
Approximation principle
7
4. A subgroup S of L is called isolated if the only elements x ∈ L with xp ∈ S are the
elements of S.
Similarly, a free Zp -Lie algebra M of finite rank is called powerful if [M, M] ⊂ p′ M and
a subalgebra S of M is called isolated if x ∈ M and px ∈ S implies x ∈ S.
The prime examples of uniform pro-p groups [ibid., Theorem 5.2] are the groups
Γ(N0 , pn ) = {g ∈ GL(N0 , Zp ) : g ≡ 1
(mod pn )},
n ≥ ǫp .
We list the following properties of powerful and uniform pro-p groups. (The unreferenced parts are clear.)
Lemma 2.5.
1. ([ibid., Theorem 3.6 and Proposition 1.16]) If L is a powerful pro-p
i
i
group then Lp = L{p } , i ≥ 0. These groups form a neighborhood base of the
identity.
2. ([ibid., Theorem 3.8]) The rank of a uniform pro-p group L is the minimum number
of generators of L, which is also the dimension of the Fp -vector space L/Lp .
3. Let L be a free Zp -module of finite rank. Then the map U 7→ U ∩ L defines a
bijection between the set of vector subspaces of the Qp -vector space L ⊗ Qp and
the set of isolated Zp -submodules of L. Moreover, if L is a Zp -Lie algebra then this
map induces a bijection between the Qp -Lie subalgebras of L ⊗ Qp and the isolated
closed Lie subalgebras of L.
4. An isolated subgroup of a uniform group is uniform. Similarly, an isolated subalgebra of a powerful Lie algebra is powerful.
5. The only isolated open subgroup of a pro-p group L is the group L itself.
6. ([ibid., Theorem 3.10]) For any positive integer d there exists a non-negative integer
N = N(d) (in fact, we may take N(d) = d(⌈log2 d⌉ + ǫp − 1)) such that any open
subgroup H of a uniform pro-p group L of rank d contains a uniform characteristic
N
open subgroup V of index ≤ pN (d) (which implies in particular that H p ⊂ V ).
For any commutative ring R let gl(N0 , R) be the Lie algebra of N0 × N0 matrices
over R with the usual Lie bracket. Let g be the Lie algebra of G over Q, regarded as
a subalgebra of gl(N0 , Q), and let gZ = g ∩ gl(N0 , Z) be its Z-form, which is a free
Z-module of rank d = dim G. Let gR = gZ ⊗ R for any ring R.
P
P∞
n
Let exp and log be the power series exp x = ∞
n=0 x /n! and log x = −
n=1 (1 −
n
x) /n, whenever convergent.
If F is a field of characteristic zero and ξ ∈ gl(N0 , F ), then for any regular function
f on GL(N0 ) over F we can form the formal power series f (exp(tξ)) ∈ F [[t]]. The
following lemma is probably well known. For convenience we provide a proof.
Tobias Finis, Erez Lapid
8
Lemma 2.6. Let X be an algebraic subgroup of GL(N0 ) defined over a field F of characteristic zero and IF (X) the ideal of all regular functions on GL(N0 ) over F vanishing on X. Let ξ ∈ gl(N0 , F ). Then ξ ∈ LieF X if and only if the formal power series
f (exp(tξ)) ∈ F [[t]] vanishes for all f ∈ IF (X).
In particular, if F = Qp and ξ ∈ gl(N0 , Qp ) is such that the power series exp(tξ)
converges for all t ∈ Zp , then ξ ∈ LieQp X is equivalent to exp ξ ∈ X(Qp ).
Proof. Using the notation of [Bor91, §3.7], we have the differentiation formula
d
g(exp(tξ)) = (g ∗ ξ)(exp(tξ))
dt
P
for any regular function g on GL(N0 ) over F . To see this, write g(x · y) = i ui (x)vi (y)
with regular functions ui and vi as in [loc. cit.], which gives the relation g(exp((t1 +
P
t2 )ξ)) = i ui (exp(t1 ξ))vi (exp(t2 ξ)) in F [[t1 , t2 ]]. Taking the derivative with respect to
P
t2 at t2 = 0 yields dtd g(exp(tξ)) t=t1 = i ui (exp(t1 ξ))(ξvi ) = (g ∗ ξ)(exp(t1 ξ)) (by
formula (2) in [loc. cit.]), as claimed.
It follows that the coefficient of ti in the power series of f (exp(tξ)) is i!1 fi (1) where
f0 = f and fi = fi−1 ∗ ξ, i > 0. By [ibid., Proposition 3.8 (ii)], for f ∈ IF (X) and
ξ ∈ LieF X the functions fi vanish on X for all i, whence the ‘only if’ part of the lemma.
For the ‘if’ part, note that the vanishing of f (exp(tξ)) for all f ∈ IF (X) implies the
vanishing of the linear terms of these series, i.e., of ξf = (f ∗ ξ)(1), for all such f . By
[ibid., Proposition 3.8 (i)], this gives ξ ∈ LieF X.
The last assertion follows since a convergent power series on Zp which vanishes on
the rational integers is identically zero.
If A and B are sets and f : A → B and g : B → A are functions, we write
f
⇀
A−
↽
−B
g
to mean that f and g are mutually inverse bijections, i.e., g ◦ f = 1A and f ◦ g = 1B .
The following proposition summarizes the correspondence between subgroups of Γ(N0 , p′ )
and Lie subalgebras of p′ gl(N0 , Zp ). It is discussed in [DdSMS99].
Proposition 2.7.
1. ([ibid., Proposition 6.22 and Corollary 6.25]) The power series
exp (resp. log) converge on the Zp -Lie algebra p′ gl(N0 , Zp ) (resp., the uniform
group Γ(N0 , p′ )) and define mutually inverse bijections
exp
′
−
⇀
p′ gl(N0 , Zp ) −
↽
−
− Γ(N0 , p ).
log
2. The maps exp and log induce mutually inverse bijections
exp
′
n
−
⇀
p′ gl(N0 , Zp )/pn gl(N0 , Zp ) −
↽
−
− Γ(N0 , p )/Γ(N0 , p ),
log
n ≥ ǫp .
Approximation principle
9
3. ([ibid., Corollary 7.14 and Theorem 9.10]) Applying the maps exp and log to subsets gives rise to bijections
{powerful closed Lie subalgebras of p′ gl(N0 , Zp )}
exp
−−
⇀
↽
−
−
log
{uniform closed subgroups of Γ(N0 , p′ )}.
4. ([ibid., Theorem 4.17 and scholium to Theorem 9.10]) Under these bijections the
rank of a uniform subgroup H of Γ(N0 , p′ ) is dimQp log H ⊗ Qp and
{isolated closed Lie subalgebras of p′ gl(N0 , Zp )}
exp
−−
⇀
↽
−
−
log
{isolated closed subgroups of Γ(N0 , p′ )}.
5. For any algebraic subgroup X of GL(N0 ) defined over Qp we have
exp(LieQp X ∩ pn gl(N0 , Zp )) = X(Qp ) ∩ Γ(N0 , pn ),
n ≥ ǫp .
In particular, for all n ≥ ǫp we have exp(pn gZp ) = Kp (pn ) and Kp (pn ) is a uniform
group.
Remark 2.8. In §3 below we will consider the exponential and logarithm maps on larger
domains, provided that p is sufficiently large with respect to N0 .
Proof. To prove part 2, we need to show that
exp(p′ x + pn y) ≡ exp(p′ x)
(mod pn gl(N0 , Zp ))
for any x, y ∈ gl(N0 , Zp ) and n ≥ ǫp . Expanding the power series as an infinite linear
combination of products of x and y (which in general do not commute), it is enough to
show that
−vp (k!) + ǫp (k − i) + ni ≥ n
for any 0 < i ≤ k. Equivalently,
vp (k!) ≤ n(i − 1) + ǫp (k − i),
which holds since vp (k!) ≤ (k − 1)/(p − 1) = (i − 1)/(p − 1) + (k − i)/(p − 1) and
1/(p − 1) ≤ n, ǫp . Similarly, the weaker inequality
−vp (k) + ǫp (k − i) + ni ≥ n
Tobias Finis, Erez Lapid
10
shows that
log((1 + p′ x)(1 + pn y)) ≡ log(1 + p′ x)
(mod pn gl(N0 , Zp ))
for any x, y ∈ gl(N0 , Zp ) and n ≥ ǫp . Part 2 follows. (Alternatively, this part also follows from [ibid., Theorem 5.2 and Corollary 7.14] together with the first part of [ibid.,
Corollary 4.15].)
Given part 1, part 5 follows from Lemma 2.6 applied to the elements of pn gl(N0 , Zp ).
Remark 2.9. In fact, one can intrinsically endow any uniform pro-p group with the structure of a Lie algebra over Zp which is free of finite rank as a Zp -module [ibid., §4.5] and
from which we can recover the group structure. We will not recall this construction here,
since for our purposes it is advantageous to use the realization of the Lie algebra inside
a matrix space via ρ0 (cf. [ibid., §7.2]).
We can immediately deduce analogous results for subgroups of Kp (p′ ).
Proposition 2.10.
1. The maps exp and log induce mutually inverse bijections
exp
′
n
−
⇀
p′ gZp /pn gZp −
↽
−
− Kp (p )/Kp (p ),
log
n ≥ ǫp .
2. The application of exp and log to subsets gives rise to bijections
exp
′
−
⇀
{powerful closed Lie subalgebras of p′ gZp } −
↽
−
− {uniform closed subgroups of Kp (p )}
log
and
exp
′
−
⇀
{isolated closed Lie subalgebras of p′ gZp } −
↽
−
− {isolated closed subgroups of Kp (p )}.
log
3. For any algebraic subgroup X of G defined over Qp we have
exp(LieQp X ∩ pn gZp ) = X(Qp ) ∩ Kp (pn ),
n ≥ ǫp .
Remark 2.11. In general, a closed subgroup of Kp (p′ ) is not necessarily uniform. Nevertheless, by a result of Ilani [Ila95] (combined with [DdSMS99, Corollary 7.14]) for
p ≥ dim G the application of exp and log to subsets gives rise to bijections
exp
′
−−
⇀
{closed Lie subalgebras of p′ gZp } ↽
−
− {closed subgroups of Kp (p )}.
log
If moreover p > N0 + 1, then this bijection can be extended to all closed pro-p subgroups
of Kp ([Klo05, GSK09]). See §4, where we also give a further generalization that is valid
for all p large enough with respect to G.
Approximation principle
11
Note that Ilani’s theorem implies (together with Proposition 2.7, part 2) that for p ≥
dim G we have [Kp (p) : H] ≥ pn−1 for all open subgroups H of Kp (p) of level pn . If G
is simply connected and p is sufficiently large, then by [LS03, Window 9, Lemma 5]
for any proper open subgroup H of Kp the image H̄ of H in G(Fp ) ≃ Kp /Kp (p)
is a proper subgroup of G(Fp ). (2.1)
Since G(Fp ) is generated by elements of p-power order, we have [G(Fp ) : H̄] ≥ p, and
therefore if H is of level pn , n ≥ 1, then [Kp : H] = [G(Fp ) : H̄][Kp (p) : H ∩ Kp (p)] ≥
pn , which is a well-known result of Lubotzky [Lub95, Lemma 1.6]. By Lemma 2.5, part 6,
and Proposition 2.10, for every p we have at least [Kp : H] ≥ cpn for all open subgroups
H of Kp of level pn , where c > 0 depends only on G.
2.3 Proof of Theorem 2.2
The essential step in the proof of Theorem 2.2 is to establish the following Lie algebra
analog. If Lp is a free Zp -module of finite rank, we say that an open submodule M ⊂ Lp
has level pn if n is the minimal integer such that pn Lp ⊂ M.
Theorem 2.12. Let L be a Lie algebra over Z which is free of finite rank as a Z-module.
Then there exist an integer D > 0 and a constant 0 < ε ≤ 1 with the following property.
For any prime p and an open Lie subalgebra M ⊂ Lp = L ⊗ Zp of level pn , where
n > vp (D), there exists a proper, isolated closed subalgebra I of Lp such that M ⊂
I + p⌈εn⌉ Lp .
Before proving Theorem 2.12 we show how it implies Theorem 2.2.
Proof of Theorem 2.2. As in Remark 2.3, we are free to assume throughout that n ≥ n0 ,
where n0 depends only on G (and will be determined below).
Let H ⊂ Kp (p′ ) be an arbitrary subgroup of level pn and V ⊂ H a uniform subgroup
as in part 6 of Lemma 2.5. Clearly, V has level at least pn . Since V is uniform, log V is by
part 2 of Proposition 2.10 a Lie subalgebra of p′ gZp of level at least pn (inside gZp ). We
now apply Theorem 2.12 to L = gZ and M = log V , and obtain under the assumption
n > vp (D) the existence of a proper, isolated closed subalgebra I of gZp with
log V ⊂ I + p⌈εn⌉ gZp .
N
Since V ⊃ H {p } , where N = N(dim G) depends only on G, we get
pN log H = log H {p
N}
⊂ I + p⌈εn⌉ gZp ,
and hence, if ⌈εn⌉ ≥ N + ǫp , the set log H is contained in
(Qp I + p⌈εn⌉−N gZp ) ∩ p′ gZp = p′ I + p⌈εn⌉−N gZp ,
Tobias Finis, Erez Lapid
12
where the last equality holds since I is isolated in gZp . It follows from part 1 of Proposition
2.10 that
H = exp log H ⊂ exp(p′ I)Kp (p⌈εn⌉−N ).
(2.2)
Using the notation of [Bor91, §7.1], let X = A(Qp I) be the smallest algebraic subgroup of G defined over Qp such that LieQp X ⊃ Qp I. Note that X is connected and that
its Lie algebra LieQp X = a(Qp I) is the smallest algebraic subalgebra of gQp containing
the proper subalgebra Qp I. Since gQp is semisimple,4 it follows from [ibid., Corollary
7.9] that LieQp X is still a proper subalgebra of gQp , and we may therefore replace I by
LieQp X ∩ gZp . Combining part 3 of Proposition 2.10 with (2.2), we obtain
H ⊂ (X(Qp ) ∩ Kp (p′ ))Kp (p⌈εn⌉−N ).
Since we are free to replace ε by a smaller positive constant, we conclude Theorem 2.2.
Remark 2.13. If p ≥ dim G then using Ilani’s theorem (see Remark 2.11) we can avoid
using V in the argument above and keep the same value of ε as in Theorem 2.12.
It remains to prove Theorem 2.12. We first need an abstract lifting result for isolated
subalgebras.
Lemma 2.14. Let L be a Lie algebra over Z which is free of rank d as a Z-module. Then
there exist an integer D1 > 0 and a constant 0 < ε1 ≤ 1 with the following property. Let
p be a prime and r = 1, . . . , d − 1. Assume that x1 , . . . , xr are elements of Lp = L ⊗ Zp
which are linearly independent modulo p and satisfy
[xi , xj ] ∈
r
X
k=1
Zp xk + pν Lp ,
1 ≤ i, j ≤ r,
with ν > vp (D1 ). Then there exist y1 , . . . , yr ∈ Lp such that yi ≡ xi (mod p⌈ε1 ν⌉ ) and
I = Zp y1 + · · · + Zp yr is an isolated closed subalgebra of Lp of rank r.
Remark 2.15. An analogous result in the archimedean case was proved in [EMV09]
using the Lojasiewicz inequality.
The proof is an application of the following general lifting theorem of M. Greenberg
[Gre74].
Theorem 2.16 (Greenberg). Let f1 , . . . , fm be polynomials in N variables over Z. Then
there exist positive integers C and D with the following property. Suppose that x1 , . . . , xN ∈
Zp and ν > vp (D) are such that fi (x1 , . . . , xN ) ≡ 0 (mod pν ) for all i. Then there exist
ν−vp (D)
y1 , . . . , yN ∈ Zp such that yi ≡ xi (mod p⌈ C ⌉ ) and fi (y1 , . . . , yN ) = 0 for all i.
4
This is the only place in the proof where we use the semisimplicity of G.
Approximation principle
13
Remark 2.17. If the affine variety over Q defined by f1 , . . . , fm is smooth, then the proof
in [Gre74] shows that we can in fact take C = 1 (cf. [Gre66], where the analogous result
over Zp is credited to Néron). The point of Theorem 2.16 is that it holds without any
restriction on f1 , . . . , fm .
Proof of Lemma 2.14. By fixing a Z-basis for L, we can identify L with Zd . The space of
k-tuples (x1 , . . . , xk ) of vectors in L ⊗ Q can then be identified with an affine space Xk
of dimension kd. For any subset S ⊂ {1, . . . , d} of cardinality k let DS (y1 , . . . , yk ) be the
corresponding k × k-minor of the k × d-matrix formed by the vectors y1 , . . . , yk ∈ L.
In the statement of the lemma we can clearly fix 1 ≤ r < d. Let k = r +1 and consider
the polynomials fij,S (x1 , . . . , xr ) := DS (x1 , . . . , xr , [xi , xj ]) on Xr for all 1 ≤ i, j ≤ r
and subsets S ⊂ {1, . . . , d} of cardinality k. These polynomials have coefficients in Z
and they define a closed subvariety Vr of Xr .
Whenever x1 , . . . , xr ∈ Lp are linearly independent, and (x1 , . . . , xr ) ∈ Vr (Qp ), the
associated free Zp -submodule M = Zp x1 ⊕ · · · ⊕ Zp xr of Lp (of rank r) is a closed Lie
subalgebra. If furthermore the reductions modulo p of the vectors x1 , . . . , xr are already
linearly independent, then M = Zp x1 ⊕ · · · ⊕ Zp xr is an isolated closed Lie subalgebra
of Lp .
Let now x1 , . . . , xr ∈ Lp be as in the statement of Lemma 2.14. Then fij,S (x1 , . . . , xr ) ≡
0 (mod pν ) for all i, j and S. Applying Theorem 2.16 to the polynomials fij,S , if ν >
vp (D) there exists (y1 , . . . , yr ) ∈ Vr (Qp ) with yi ≡ xi (mod pµ ), where
ν − vp (D)
ν
µ=
≥
.
C
C + vp (D)
In particular, the vectors y1 , . . . , yr are congruent modulo p to the vectors x1 , . . . , xr ,
which implies that they are linearly independent modulo p. Therefore I = Zp y1 ⊕ · · · ⊕
Zp yr is an isolated closed Lie subalgebra of Lp . We obtain the assertion with ε1 = (C +
maxp vp (D))−1.
Proof of Theorem 2.12. Let M be an open Lie subalgebra of Lp of level pn . In particular,
M is a Zp -submodule of finite index. By the elementary divisor theorem, there exist a
Zp -basis x1 , . . . , xd of Lp and integers 0 ≤ α1 ≤ · · · ≤ αd such that (pαi xi )1≤i≤d is a
basis of M. The fact that M is of level pn means that αd = n.
Fix an arbitrary constant 0 < c < 12 . Let 0 < r ≤ d − 1 be the maximal index such
that αr < cαr+1 , if such an index exists; otherwise set r = 0. Note that
αr+1 ≥ cαr+2 ≥ · · · ≥ cd−r−1 αd = cd−r−1 n.
P
Write [xi , xj ] = k cijk xk with cijk ∈ Zp . Since M is a Lie subalgebra of Lp , we have
pαi +αj cijk ∈ pαk Zp . For i, j ≤ r and k > r we obtain cijk ∈ pαk −αi −αj Zp ⊂ pαr+1 −2αr Zp .
Here,
αr+1 − 2αr > (1 − 2c)αr+1 ≥ (1 − 2c)cd−r−1 n.
Tobias Finis, Erez Lapid
14
Summing up, we obtain the existence of 0 ≤ r ≤ d − 1 such that [xi , xj ] ∈
pν L for 1 ≤ i, j ≤ r, where
ν = (1 − 2c) cd−r−1 n .
Pr
k=1 Zp xk
+
(2.3)
We may apply Lemma 2.14 to the elements x1 , . . . , xr and obtain an isolated Lie subalgebra I = Zp y1 ⊕ · · · ⊕ Zp yr of Lp with xi ≡ yi (mod p⌈ε1 ν⌉ ) whenever ν > vp (D1 ).
The last condition is evidently satisfied when n > vp (D), where D = D1e with e =
⌈(1 − 2c)−1 c1−d ⌉. Under this condition we get
M ⊂ Zp x1 ⊕ · · · ⊕ Zp xr + pαr+1 L ⊂ I + p⌈εn⌉ L
for ε = ε1 (1 − 2c)cd−1 , since αr+1 ≥ cd−r−1 n ≥ εn.
We remark that Theorem 2.12 admits a local counterpart:
Theorem 2.18. Let p be a prime number and let L be a Lie algebra over Zp which is
free of finite rank as a Zp -module. Then there exist an integer n0 ≥ 0 and a constant
0 < ε ≤ 1 with the following property. For any open Lie subalgebra M ⊂ L of level
pn , where n > n0 , there exists a proper, isolated closed subalgebra I of L such that
M ⊂ I + p⌈εn⌉ L.
The proof is the same except that we use the natural variant of Theorem 2.16 where
the ring Z is replaced by Zp (which is contained in [Gre74], but also already in [Gre66,
Theorem 1]). For our purposes, the additional uniformity in p of ε and n0 in Theorem 2.12
is important.
Similarly, we have the following analog of Theorem 2.2 for arbitrary uniform pro-p
groups.
Theorem 2.19. Let L be a uniform pro-p group. Then there exist an integer n0 ≥ 0 and
a constant 0 < ε ≤ 1 such that for any open subgroup H of L with n = min{m ≥ 0 :
m
⌈εn⌉
Lp ⊂ H} > n0 there exists a proper, isolated closed subgroup I of L with H ⊂ ILp .
Proof. Let H be a subgroup of L and n as above. We use the Lie algebra structure on L
(and in particular the additive structure) defined in [DdSMS99, §4]. In particular, xm =
mx for any x ∈ L and m ∈ Z [ibid., Lemma 4.14]. Let V ⊂ H be as in part 6 of Lemma
2.5. Since V is uniform, it is a Lie subalgebra of L of level ≥ pn [ibid., Theorem 9.10].
Applying Theorem 2.18, there exists a proper, isolated closed subalgebra I of L such that
V ⊂ I + p⌈εn⌉ L.
Since V ⊃ H {p
N}
= pN H we get
pN H ⊂ I + p⌈εn⌉ L,
Approximation principle
15
and hence, if ⌈εn⌉ ≥ N,
H ⊂ I + p⌈εn⌉−N L,
since I is isolated in L. For any integer m ≥ 0 the identity map induces a bijection
m
between the quotient group L/Lp and the quotient L/pm L with respect to the additive
structure on L [ibid., Corollary 4.15]. We obtain
H ⊂ ILp
⌈εn⌉−N
,
from which the theorem follows. (Recall that by [ibid., Scholium to Theorem 9.10], I is
an isolated subgroup of L.)
We end up the discussion of Theorem 2.2 with a couple of comments about the function ν(n) and the number εG of Remark 2.3 above.
Lemma 2.20. We have ν(n) ≤ ⌈ n2 ⌉ and hence εG ≤ 12 .
Proof. For n ≥ 2ǫp − 1 the logarithm map induces an isomorphism
n
n
n
Kp (p⌈ 2 ⌉ )/Kp (pn ) ≃ p⌈ 2 ⌉ gZp /pn gZp ≃ gZp /p⌊ 2 ⌋ gZp
2
(mod pn )
of abelian groups. (To see this, note that we have log x ≡ x − 1 − (x−1)
2
n
and therefore log(xy) ≡ log x + log y (mod pn ) for x, y ∈ Γ(N0 , p⌈ 2 ⌉ ). The bijectivity
follows from part 1 of Proposition 2.10.)
n
Therefore, the map H 7→ p−⌈ 2 ⌉ log H gives rise to a bijection between the open subn
n
groups of Kp (p⌈ 2 ⌉ ) ⊂ Kp (p′ ) of level pn and the essential subgroups of gZp /p⌊ 2 ⌋ gZp ,
n
n
namely those which do not contain p⌊ 2 ⌋−1 gZp /p⌊ 2 ⌋ gZp .
Assume on the contrary that ν = ν(n) > ⌈ n2 ⌉. Then it follows from this discusn
sion and Proposition 2.10 that every essential subgroup H̄ of gZp /p⌊ 2 ⌋ gZp is contained
n
n
in (LieQp X ∩ gZp ) + pν−⌈ 2 ⌉ gZp /p⌊ 2 ⌋ gZp for some proper, connected algebraic subgroup
n
X of G. Equivalently, the projection of H̄ to gZp /pν−⌈ 2 ⌉ gZp is contained in the reduction
n
n
modulo pν−⌈ 2 ⌉ of LieQp X ∩ gZp , which is a proper Lie subalgebra of gZp /pν−⌈ 2 ⌉ gZp .
Considering the subgroups H̄ obtained as kernels of arbitrary surjective group homomorn
n
phisms gZp /p⌊ 2 ⌋ gZp → Z/p⌊ 2 ⌋ Z, one sees that at least for almost all p this is impossible.
We conclude that ν(n) ≤ ⌈ n2 ⌉.
In general, the proof of Theorem 2.12 does not give a good value for εG , since we
used Greenberg’s theorem to lift a Lie algebra modulo pν to characteristic zero. We note
however that for L = sl2 (Z) and p odd one gets the optimal value (and that even for p = 2
we get an almost optimal result).
Tobias Finis, Erez Lapid
16
Lemma 2.21. Consider the Lie algebra L = sl2 (Z). Then for p odd, Theorem 2.12 is true
for all n > 0 with ε = 12 , which is optimal. Similarly, Theorem 2.2 holds for G = SL(2)
over Z and all odd primes p with ε = 21 .
For p = 2, Theorem 2.12 holds for all n ≥ 3 with ⌈εn⌉ replaced by ⌈ n2 ⌉ − 1. We have
ν(n) ≥ ⌈ n2 ⌉ − 10 for G = SL(2) over Z, and in particular εG = 12 .5
Proof. Consider an open Lie subalgebra M = pα1 Zp x1 ⊕ pα2 Zp x2 ⊕ pn Zp x3 of Lp , where
x1 , x2 , x3 form a Zp -basis of Lp and 0 ≤ α1 ≤ α2 ≤ n. Assume first that p is odd. If
α2 ≥ n2 we set I = Zp x1 and obtain M ⊂ I + pα2 Lp . Otherwise, the rank two isolated
submodule J = Zp x1 ⊕ Zp x2 of Lp maps to a Lie subalgebra of L/pn−α1 −α2 L, where
n − α1 − α2 > 0. We claim that we can lift J to a rank two isolated subalgebra I ⊂ Lp
with J + pn−α1 −α2 Lp = I + pn−α1 −α2 Lp . Granted this claim, we get M ⊂ I + pn−α2 Lp .
n
Thus, in both cases we have M ⊂ I + p⌈ 2 ⌉ Lp . The statement for SL(2) follows then from
Remark 2.13.
For p = 2 we modify the argument slightly and distinguish the cases α2 ≥ n2 − 1
and α2 < n2 − 1. In the second case J maps to a Lie subalgebra of L/2m L with m =
n − α1 − α2 ≥ n − 2α2 > 2, and we claim that there exists a rank two isolated subalgebra
n
I ⊂ L2 with J + 2m−2 L2 = I + 2m−2 L2 . We obtain M ⊂ I + 2⌈ 2 ⌉−1 L2 . The statement
for SL(2) follows from the proof of Theorem 2.2, noting that we can take N = N(3) = 9
in part 6 of Lemma 2.5.
It remains to establish our claim on the lifting of subalgebras (for arbitrary p). We may
c2
parametrize rank two isolated submodules J of Lp by elements c = ( cc13 −c
) ∈ sl2 (Qp )
1
with min(vp (2c1 ), vp (c2 ), vp (c3 )) = 0, considered up to multiplication by scalars in Z×
p,
by setting J = J(c) = {x ∈ Lp : tr(cx) = 0}. It is easy to check that the submodule
J(c) is a Lie subalgebra of Lp if and only if 2 tr(c2 ) = (2c1 )2 + 4c2 c3 = 0, and that for
m ≥ 1 the module J(c) maps to a Lie subalgebra of Lp /pm Lp if and only if 2 tr(c2 ) =
(2c1 )2 + 4c2 c3 ≡ 0 (mod pm ). For p > 2 the quadric defined by this equation is smooth
over Zp and points modulo pm lift to Zp -points. For p = 2 it is easy to see that for
m ≥ 3, points modulo 2m are congruent modulo 2m−2 to Z2 -points. Therefore, whenever
J(c) maps to a Lie subalgebra modulo pm there exists a rank two isolated Lie subalgebra
I ⊂ Lp with J(c) + pm Lp = I + pm Lp for p odd (resp., J(c) + 2m−2 L2 = I + 2m−2 L2
for p = 2).
As one sees from the proof, it is not difficult to obtain an explicit parametrization
of open Lie subalgebras of sl2 (Zp ) and of open subgroups of the principal congruence
subgroups of SL(2, Zp ), at least for odd p. For more details and the explicit computation
of the corresponding counting functions see [Ila99, dS00, dST02, KV09].
5
Using [CP84, Proposition 2.9, Theorem 2.10] instead of Lemma 2.5, one may improve this bound to
ν(n) ≥ ⌈ n−3
2 ⌉. We omit the details.
Approximation principle
17
2.4 A finiteness property for maximal connected algebraic subgroups
For a linear algebraic group G defined over a field F and a field extension E ⊃ F , we
denote by MSGRE (G) the set of maximal (proper) connected algebraic subgroups of
GE defined over E. Note that in the statement of Theorem 2.2 we can clearly take X to
be in MSGRQp (G). As a supplement to this theorem, we now consider a certain simple
finiteness property of the family or algebraic groups MSGRQp (G), where p ranges over
all prime numbers. This property is crucial for the application of our results in §5 below.
We first quote a basic finiteness property of affine varieties [Sei74, §65]. Denote by
n
AR = Spec R[x1 , . . . , xn ] the n-dimensional affine space over a ring R.
Lemma 2.22 (Seidenberg). Let n and d be positive integers. Then there exists a constant
C = C(n, d) with the following property. Let F be an algebraically closed field and let
V be a closed subvariety of AnF whose defining ideal I(V ) is generated in degrees ≤ d.
Then the number of irreducible components of the variety V is at most C and the defining
ideal of any irreducible component W is generated in degrees ≤ C. Moreover, whenever
√
I is an arbitrary ideal of F [X1 , . . . , Xn ] generated in degrees ≤ d, the radical I, which
is the defining ideal of the zero set of I, is generated in degrees ≤ C.
Remark 2.23. Let F be an arbitrary field and V be a closed subvariety of AnF̄ defined
over F with defining ideal I(V ) ⊂ F̄ [X1 , . . . , Xn ]. Then the following are equivalent:
• The ideal IF (V ) = I(V )∩F [X1 , . . . , Xn ] ⊂ F [X1 , . . . , Xn ] is generated in degrees
≤ d.
• I(V ) is generated in degrees ≤ d.
This is simply because I(V ) is generated as an F̄ -vector space by IF (V ).
Definition 2.24. Let Y = Spec A be an affine scheme over a ring R. An admissible family
X is a family of pairs (E, X), where E is a field together with a ring homomorphism
R → E and X ⊂ YE = Y ×Spec R Spec E is a closed subscheme. We say that X is (BG)
(boundedly generated) if there exists a finitely generated R-submodule M ⊂ A such that
for all (E, X) ∈ X the defining ideal IE (X) of the subscheme X is generated by elements
of M ⊗R E.
Remark 2.25. If Y is of finite type over R, then we can reformulate this definition as
follows. Fix a closed embedding of Y into AnR (equivalently, a choice of generators
a1 , . . . , an of A as an R-algebra). We then have corresponding closed embeddings of
YE into AnE for any E, and can define the notion of the degree of a regular function on
YE . An admissible family X is (BG) if and only if there exists an integer N such that for all
(E, X) ∈ X the defining ideal IE (X) of the subscheme X is generated by regular functions defined over E of degree at most N. Note that this property is actually independent
of the choice of embedding.
18
Tobias Finis, Erez Lapid
We will mostly use this definition in the situation where we are given a linear algebraic
group G defined over a field F and a family of algebraic subgroups defined over field
extensions E ⊃ F . We will also consider the case of a flat affine group scheme over Z
and of the fields E = Fp as in §2.1.
Lemma 2.26. Let G be an affine group scheme of finite type over a ring R and let X be
an admissible family such that for all (E, X) ∈ X the base extension GE is geometrically
reduced (i.e., a linear algebraic group defined over E) and X is an algebraic subgroup
of GE defined over E. Fix a closed embedding of G into AnR for some n.
Then X is (BG) if and only if there exist an integer N and for every (E, X) ∈ X
an E-rational representation (ρ, V ) of GE , where V is an E-vector space, and a line
D ⊂ V , such that the matrix coefficients v ∨ (ρ(g)v), v ∈ V , v ∨ ∈ V ∨ , have degree
≤ N and X = {g ∈ GE : ρ(g)D = D}. Moreover, if this is the case, then we may
require in addition that ρ is injective, that dim V ≤ N and that the associated orbit map
G → P(V ), g 7→ ρ(g)D, is separable, i.e., that LieE X = {v ∈ gE : dρ(v)D ⊂ D}
[Bor91, Proposition 6.7].
Proof. The ‘only if’ part (with the additional injectivity and separability conditions and
the boundedness of dim V ) follows from the proof of [Bor91, Theorem 5.1]. The main
point is the existence of an integer N ′ such that for all (E, X) ∈ X there exists a right
GE -invariant E-subspace U of the space of regular functions on GE of degree ≤ N ′ ,
which generates the ideal of X. This claim follows from the proof of [ibid., Proposition
1.9]. The representation ρ is then given as the direct sum of a suitable exterior power of
U and some injective representation ρ0 . If we obtain the representation ρ0 by the construction of [ibid., Proposition 1.10] from the coordinate functions provided by the fixed
closed embedding of G into AnR , then the degrees of the matrix coefficients of ρ0 and its
dimension are bounded in terms of G. Therefore the degrees of the matrix coefficients of
ρ and its dimension are bounded in terms of G and N ′ .
To see the ‘if’ part, note that the defining ideal I(X) of X over Ē is the radical of
the ideal generated by the matrix coefficients v ∨ (ρ(g)v), v ∈ D, v ∨ ∈ D ⊥ , and hence
by Lemma 2.22 I(X) is generated in degrees bounded in terms of G and N only. On the
other hand, I(X) is by assumption generated by regular functions defined over E. In view
of Remark 2.23, this shows that IE (X) is generated in degrees bounded in terms of G and
N.
For the following lemma, recall that the connected component of the identity X ◦ of an
algebraic group X defined over E is again defined over E [Bor91, Proposition 1.2].
Lemma 2.27. Let G and X be as in Lemma 2.26. Suppose that the family X is (BG).
Then the family of its identity components {(E, X ◦ ) : (E, X) ∈ X } is also (BG) and the
sizes of the component groups X/X ◦ , (E, X) ∈ X , are bounded.
Approximation principle
19
Proof. This follows immediately from Lemma 2.22.
Lemma 2.28. Let G be an affine group scheme of finite type over a ring R and E a
collection of perfect fields E together with homomorphisms R → E for which the schemes
GE are geometrically reduced. Then the family of all pairs (E, NGE (U)), where E ∈ E
and U ranges over the subspaces of the E-vector space gE = LieR G ⊗ E, is (BG). The
family of all pairs (E, NGE (U)◦ ) is also (BG).
Note here that NGE (U) is defined over E because E is perfect.
Proof. Let d = dim U and consider the d-th exterior power Vd = ∧d Ad of the adjoint
representation of GE . Then the group NGE (U) is the stabilizer of the line ∧d U ⊂ Vd ⊗ E
(cf. [Bor91, §5.1, Lemma]). The degrees of the matrix coefficients of the representations
Vd are clearly bounded in terms of G only. Hence we can use Lemma 2.26. The last
assertion follows immediately from Lemma 2.27.
Corollary 2.29. Let G be a semisimple algebraic group defined over a field F of characteristic zero and E a collection of field extensions of F . Then the family
MAX E = {(E, X) : E ∈ E, X ∈ MSGRE (G)}
is (BG).
By Remark 2.31 below, the analogous statement for G over R = Z (as in §2.1) and the
family {(Fp , X) : X ∈ MSGRFp (G)} is not true in general.
Proof. Let E ∈ E and X ∈ MSGRE (G). Consider the normalizer N = NG (x) of the
Lie algebra x of X. Since we are in characteristic zero, we have N = NG (X) ⊃ X.
If N = G, then X is a normal subgroup of G, which cannot be maximal, since G is
semisimple. Therefore necessarily N ◦ = X. It remains to apply Lemma 2.28.
In fact, in the setting of Corollary 2.29 one can show a stronger finiteness property
for the family MAX E . (This will not be used in the sequel.) Let us say that a family X
satisfies (FC) if there exists a finite set S of subgroups of GF̄ such that for any (E, X) ∈ X
the group XĒ is G(Ē)-conjugate to YĒ for some Y ∈ S. (By Remark 2.23, this clearly
implies that X is (BG).) We claim that the following families satisfy (FC):
1. {(E, X) : E ∈ E, X contains a maximal torus of G defined over E},
2. {(E, X) : E ∈ E, X is a semisimple connected subgroup of G defined over E},
3. MAX E .
Tobias Finis, Erez Lapid
20
The first case follows from [Bor91, Corollary 11.3 and Proposition 13.20] (for any field
E). The second case follows from Richardson’s rigidity theorem [Ric67], or alternatively
from Dynkin’s description of the connected semisimple subgroups over an algebraically
closed field, which is based on the classification of semisimple groups [Dyn52b, Dyn52a].
The third case follows from the first two together with the fact (valid for any field E) that
for X ∈ MSGRE (G) there are three possibilities:
1. X is not reductive,
2. X is the centralizer of a torus defined over E,
3. X is semisimple.
(To see this, note that for X reductive but not semisimple, Z(X)◦ is a non-trivial torus
defined over E [Bor91, Proposition 11.21, Theorem 18.2 (ii)], and that X ⊂ CG (Z(X)◦ ),
which is defined over E [ibid., Corollary 9.2] and connected [ibid., Corollary 11.12].
Therefore X = C(Z(X)◦ ) and we are in the second case.) In the first case X is a maximal
parabolic subgroup of GE defined over E (See [Mor56]; in fact, by [BT71, Corollaire 3.3]
this is true for any perfect field E.)
Remark 2.30. Corollary 2.29 does not hold in general if G is not semisimple. For example, if G a torus of dimension ≥ 2 which is split over E then MSGRE (G) is the set of all
codimension one subtori, and it is easy to see that this set is not (BG).
Remark 2.31. Corollary 2.29 does not hold in the positive characteristic case. Indeed,
even if F is algebraically closed the set MSGRF (G) may not be (BG) (see [LT04, Lemma
2.4] for a description of MSGRF (G) in this case). On the other hand, if G is a simple
algebraic group over (an algebraically closed field) F then the finiteness of the number of
G(F )-conjugacy classes in MSGRF (G) (or in fact, the finiteness of the number of G(F )conjugacy classes of maximal subgroups of G of positive dimension) has been established
in [LS04]. However, even in that case the set of all connected semisimple subgroups is
usually not (BG) (cf. [LT04, esp. Corollary 4.5]).
Remark 2.32. In the situation where F = Q and E is the family of the p-adic fields
E = Qp , finiteness of Galois cohomology [PR94, Theorem 6.14] implies that each of the
sets MSGRQp (G) consists of finitely many classes under conjugation by G(Qp ) (instead
of G(Qp )).
3 The approximation theorem for open subgroups of G(Zp)
3.1 Statement of the result
Let G and ρ0 be as in §2.1 above. We will now prove an extension of Theorem 2.2 to
arbitrary open subgroups of Kp in the case where G is simply connected.
Approximation principle
21
Theorem 3.1. Suppose that G is simply connected. Then there exist constants J ≥ 1
and ε > 0, depending only on G and ρ0 , with the property that for every open subgroup
H ⊂ Kp of level pn there exists a proper, connected algebraic subgroup X ⊂ G defined
over Qp such that [H : H ∩ (X(Qp ) ∩ Kp )Kp (p⌈εn⌉ )] ≤ J.
Remark 3.2.
1. Note that Theorem 3.1 implies the existence of an open normal sub′
group H of H of index ≤ J! which is contained in the group (X(Qp )∩Kp )Kp (p⌈εn⌉ ).
2. It is clear that just like Theorem 2.2, we can rephrase Theorem 3.1 in terms of
essential subgroups of the finite groups G(Z/pn Z). Note here that when GZp is
smooth over Zp (which is the case for almost all p), the reduction maps π∞,n :
Kp → G(Z/pn Z) are surjective for all n.
3. As will be explained in §3.3 below, the case n = 1 (or n bounded) follows rather
directly from results of Nori on subgroups of G(Fp ) [Nor87]. (A variant of these
results has been obtained independently by Gabber [Kat88, Ch. 12].)
3.2 Consequences of Nori’s Theorem
We first prove a corollary of an algebraization theorem of Nori for subgroups of G(Fp )
[Nor87]. While this corollary is probably known (cf. [HP95, Proposition 8.1] for a less
explicit result), we include the deduction from Nori’s result (and Jordan’s classic theorem on finite subgroups of GL(N, F ) of order prime to the characteristic of F ), since we
will use the essential part of the argument also in §3.3 below. One may in fact consider
the following as an introduction to the proof of Theorem 3.1. Also, for our application
in §5 below, Proposition 3.3 is in fact sufficient. Note that a more general algebraization
theorem for finite subgroups of GL(N, F ), F an arbitrary field, has been obtained by
Larsen–Pink [LP11]. Of course, one can also deduce such theorems (with optimal constants) from the classification of finite simple groups [Col07, Col08]. However, the proofs
of Nori and Larsen–Pink are independent of the classification.
We write n(G) for a constant (which may change from one occurrence to the other)
depending only on G and ρ0 .
Note first that for p > n(G) the group scheme GZp is smooth over Zp and GFp is
semisimple and simply connected. In particular, for those p the reduction map Kp →
G(Fp ) is surjective.
Proposition 3.3. Assume that G is simply connected. Then there exists a (BG) family X
such that for any prime p > n(G) and any proper subgroup H of G(Fp ) there exists
(Fp , X) ∈ X such that X is a proper, connected algebraic subgroup of GFp defined over
Fp and [H : H ∩ X(Fp )] ≤ n(G).
Tobias Finis, Erez Lapid
22
Remark 3.4. As above, we obtain from Proposition 3.3 the existence of a normal subgroup H ′ of H of index ≤ n(G) that is contained in the group X(Fp ).
To prove this proposition, we need to recall the results of Nori and Jordan. We will use
the following variant of Jordan’s theorem to deal with subgroups of order prime to p.
Proposition 3.5. For any integer n there exists a constant J = J(n) with the following
property. Let G be a reductive group defined over a field F with a faithful representation
of degree n. Then for all finite subgroups H of G(F ) with char F 6 ||H| there exists a
maximal torus T of G defined over F such that [H : H ∩ T (F )] ≤ J.
Proof. Jordan’s theorem implies the existence of an abelian subgroup of H of index at
most J ′ (n). On the other hand, by [BS53, BM55, SS70] we can embed any supersolvable
subgroup of G(F ) into NG (T )(F ) for some maximal torus T of G defined over F . It
remains to note that the index [NG (T ) : T ] is bounded in terms of the F̄ -rank of G, i.e.,
in terms of n.
We now state Nori’s main result, which describes (for p > n(G)) the subgroups of
G(Fp ) generated by their elements of order p in terms of certain connected algebraic
subgroups of G defined over Fp , or equivalently by certain Lie subalgebras of gFp . It
is based on the following construction. Recall the truncated logarithm and exponential
functions
p−1 i
p−1
X
X
y
(1 − x)i
, exp(p) y =
,
log(p) x = −
i
i!
i=0
i=1
which are defined over any ring in which the primes < p are invertible. Fix n and
let F be a field of characteristic p ≥ n. Denote by gl(n, F )nilp (resp., GL(n, F )unip )
the set of nilpotent (resp., unipotent) elements of gl(n, F ) (resp., GL(n, F )). Note that
gl(n, F )nilp = {x ∈ gl(n, F ) : xp = 0} and GL(n, F )unip = {x ∈ GL(n, F ) : xp = 1}
since p ≥ n. Then
exp(p)
−
⇀
gl(n, F )nilp −
↽
−−
−
− GL(n, F )unip .
log(p)
We say that an algebraic subgroup of GL(n) over F is exponentially generated (exp. gen.),
if it is generated by one-dimensional unipotent subgroups of the form t 7→ exp(p) (ty),
where y ∈ gl(n, F )nilp . By [Bor91, Proposition 2.2] such a group is automatically connected. For any subset S ⊂ gl(n, F ) write Exp S for the group generated by t 7→
exp(p) (ty), y ∈ Snilp = S ∩ gl(n, F )nilp . We call a Lie subalgebra of gl(n, F ) nilpotently
generated (nilp. gen.), if it is generated as a vector space by nilpotent elements.
We now quote the main part of [Nor87, Theorem A].
Theorem 3.6 (Nori). Assume that p > n(G). Then
L7→Exp L
{nilp. gen. Lie subalgebras of gFp } −
↽−
−−
−−
−−
−−
−⇀
− {exp. gen. subgroups of GFp }.
LieFp X←[X
Approximation principle
23
Moreover, if X is an exp. gen. subgroup of G over Fp then LieFp X is spanned by the
elements log(p) x, where x ∈ X(Fp )unip = X(Fp ) ∩ GL(n, Fp )unip .
Following Nori, for any subgroup H of G(Fp ) we define the algebraic subgroup H̃ =
Exp {log(p) x : x ∈ Hunip } of GFp . Also, we denote by H + the subgroup of H generated
by Hunip , i.e., by the p-Sylow subgroups of H. Thus, H + is a characteristic subgroup of
H of index prime to p. The following result is [Nor87, Theorem B].
Theorem 3.7 (Nori). For p > n(G) and any subgroup H of G(Fp ) we have H̃(Fp )+ =
H + and LieFp H̃ is spanned by log(p) x, x ∈ Hunip .
Proof of Proposition 3.3. We first remark that using Lemma 2.27 we may drop the connectedness statement in the conclusion of the proposition, since we may pass to the family
of identity components. We will take X to be the family
{(Fp , NG (I)) : I is a subspace of LieFp G with NG (I) 6= G}.
By Lemma 2.28, X is (BG).
Let H be a proper subgroup of G(Fp ) and let H̃ be as before. Note that H̃ is a proper
(connected) algebraic subgroup of GFp since otherwise we would have H + = H̃(Fp )+ =
G(Fp )+ = G(Fp ), because GFp is simply connected. Clearly, H is contained in the group
of Fp -points of NG (H̃), which coincides with NG (Lie H̃) by Theorem 3.6. If H̃ is not
normal in G then NG (Lie H̃) = NG (H̃) ∈ X and we are done. Otherwise, as in the proof
of [Nor87, Theorem C], using the Frattini argument we can write H = H1 H + where
H1 is a subgroup of H of order prime to p. (Namely, we choose a p-Sylow subgroup P
of H and take H1 to be a complement of P in NH (P ).) By Proposition 3.5, there exist a
subgroup A of H1 of bounded index and a maximal torus T of G defined over Fp such that
A ⊂ T (Fp ). Then we take X = NG (Lie(T H̃)). For p > n(G) the G-invariant subspaces
of gFp are precisely the Lie algebras of the connected normal subgroups of GFp . Therefore,
Lie(T H̃) ⊂ gFp is not G-invariant, and X ∈ X .
3.3 Proof of Theorem 3.1
We now prove Theorem 3.1. We first need to modify our previous intermediate results
Theorem 2.12 and Lemma 2.14 by incorporating the action of a finite abelian subgroup
of Kp of order prime to p. The following is the analog of Lemma 2.14 in this setting.
Lemma 3.8. Let G be as above and s a positive integer. Then there exist an integer
D1 > 0 and a constant 0 < ε1 ≤ 1, depending only on G, ρ0 , and s, with the following
property. Suppose that we are given x1 , . . . , xr ∈ gZp , 1 ≤ r < d, a1 , . . . , as ∈ Kp and
ν > vp (D1 ) such that
Tobias Finis, Erez Lapid
24
1. x1 , . . . , xr are linearly independent modulo p,
2. [xi , xj ] ∈
Pr
3. Ad(ai )xj ∈
k=1
Pr
Zp xk + pν gZp , 1 ≤ i, j ≤ r,
k=1
Zp xk + pν gZp , 1 ≤ i ≤ s, 1 ≤ j ≤ r.
Then there exist yi ∈ xi + p⌈ε1 ν⌉ gZp , i = 1, . . . , r, and bj ∈ aj Kp (p⌈ε1 ν⌉ ), j = 1, . . . , s,
such that
1. I = Zp y1 + · · · + Zp yr is an isolated closed subalgebra of gZp of rank r,
2. Ad(bj )I = I, j = 1, . . . , s.
Proof. Modify the proof of Lemma 2.14 by considering instead of Xr the affine variety
Xr × Gs , with the Z-structure induced from the Z-structure of the affine variety G, and its
subvariety Vr,s ⊂ Vr × Gs defined by the polynomials fij,S (x1 , . . . , xr ) introduced above
together with the regular functions DS (x1 , . . . , xr , Ad(yi )xj ), i = 1, . . . , s, j = 1, . . . , r,
of the variables (x1 , . . . , xr ) ∈ Xr and y1 , . . . , ys ∈ G. An application of Theorem 2.16
yields the assertion.
The following is the analog of Theorem 2.12.
Lemma 3.9. Let G be as above and s a positive integer. Then there exist an integer D > 0
and a constant 0 < ε ≤ 1 with the following property. For any open Lie subalgebra M ⊂
gZp of level pn , where n > vp (D), and elements a1 , . . . , as ∈ Kp with Ad(ai )M = M,
i = 1, . . . , s, there exist a proper, isolated closed subalgebra I of gZp such that M ⊂
I + p⌈εn⌉ gZp , and elements bj ∈ aj Kp (p⌈εn⌉ ), j = 1, . . . , s such that Ad(bj )I = I.
Proof. This is proved exactly like Theorem 2.12 except that we use Lemma 3.8 instead
of Lemma 2.14. We only need to check that the extra condition of Lemma 3.8 is satisfied
for the elements x1 , . . . , xr constructed in the proof above. Write
Ad(ai )xj =
d
X
dijk xk ,
k=1
dijk ∈ Zp ,
for i = 1, . . . , s and j = 1, . . . , d. Then pαj Ad(ai )xj ∈ M and therefore dijk ∈ pαk −αj Zp .
For 1 ≤ i ≤ s and 1 ≤ j ≤ r < k ≤ d we have
dijk ∈ pαk −αj Zp ⊂ pαr+1 −αr Zp ⊂ p⌈(1−c)αr+1 ⌉ Zp ⊂ p⌈(1−c)c
where ν is as in (2.3). Therefore, Ad(ai )xj ∈
as required.
Pr
k=1 Zp xk
d−r−1 n⌉
Zp ⊂ pν Zp ,
+ pν gZp , 1 ≤ i ≤ s, 1 ≤ j ≤ r,
Approximation principle
25
We also need to know that for p large we can extend the domain of definition of the
exponential and logarithm functions. Let
n
GL(N0 , Zp )resunip = {x ∈ GL(N0 , Zp ) : xp ≡ 1
pn
= {x ∈ GL(N0 , Zp ) : x
(mod p) for some n}
→ 1,
n → ∞}
be the open set of residually unipotent elements of GL(N0 , Zp ). Thus, for p ≥ N0 we
have
GL(N0 , Zp )resunip = {x ∈ GL(N0 , Zp ) : xp ≡ 1 (mod p)}.
Similarly, let
gl(N0 , Zp )resnilp = {y ∈ gl(N0 , Zp ) : y n ≡ 0
(mod p) for some n}
n
= {y ∈ gl(N0 , Zp ) : y → 0,
n → ∞}
be the set of residually nilpotent elements of gl(N0 , Zp ), so that for p ≥ N0 we have
gl(N0 , Zp )resnilp = {y ∈ gl(N0 , Zp ) : y p ≡ 0
(mod p)}.
Lemma 3.10. The power series exp (resp., log) converges in the domain gl(N0 , Zp )resnilp
(resp., GL(N0 , Zp )resunip ) provided that p > N0 + 1. Assume that p > 2N0 . Then
1. We have
exp
−−
⇀
gl(N0 , Zp )resnilp ↽
−
− GL(N0 , Zp )resunip .
log
2. The diagrams
exp
gl(N0 , Zp )resnilp −−−→ GL(N0 , Zp )resunip
y
y
gl(N0 , Fp )nilp
(3.1)
exp(p)
−−−→ GL(N0 , Fp )unip
log
GL(N0 , Zp )resunip −−−→ gl(N0 , Zp )resnilp
y
y
GL(N0 , Fp )unip
(3.2)
log(p)
−−−→ gl(N0 , Fp )nilp
commute, where the vertical arrows denote reduction modulo p.
3. For any n ≥ 1 we get induced maps
exp
n
−
⇀
gl(N0 , Zp )resnilp /pn gl(N0 , Zp ) −
↽
−
− GL(N0 , Zp )resunip /Γ(N0 , p ).
log
(3.3)
Tobias Finis, Erez Lapid
26
P
k
Proof. We will prove the first part assuming only p > N0 +1. Consider exp x = ∞
n=0 x /k!.
If x ∈ gl(N0 , Zp )resnilp then xN0 ∈ pgl(N0 , Zp ). It follows that kxk k ≤ p−⌊k/N0 ⌋ for all k.
Since vp (k!) ≤ (k − 1)/(p − 1) for k > 0, we get
kxk /k!k ≤ p−⌊k/N0 ⌋+⌊(k−1)/(p−1)⌋ → 0 as k → ∞,
provided that p−1 > N0 , which gives the convergence of exp x. Moreover, under the same
restriction on p we have kxk /k!k ≤ 1 for all k, which shows that the matrices xk = xk /k!
and their sum exp x have integral entries. Since the elements xk , k > 0 are residually
nilpotent and pairwise commute, we conclude that exp x ∈ GL(N0 , Zp )resunip . A similar
reasoning applies to log. That the maps exp : gl(N0 , Zp )resnilp → GL(N0 , Zp )resunip and
log : GL(N0 , Zp )resunip → gl(N0 , Zp )resnilp so obtained are mutually inverse follows from
[Bou98, §II.8.4, Proposition 4] (applied to the algebra of all square matrices of size N0
with entries in Cp equipped with the operator norm).6
If moreover p ≥ 2N0 then we have ⌊k/N0 ⌋ ≥ 1 + ⌊(k − 1)/(p − 1)⌋ for any k ≥ p.
Thus, if p > 2N0 then the diagram (3.1) commutes, and similarly we obtain (3.2).
The existence of the maps in (3.3) is proved like part 2 of Proposition 2.7. We need to
show that
exp(x + pn y) ≡ exp x (mod pn gl(N0 , Zp ))
(3.4)
for any x ∈ gl(N0 , Zp )resnilp , y ∈ gl(N0 , Zp ) and n ≥ 1. Expanding the power series as
an infinite linear combination of products of x and y (which do not commute in general),
we may first observe that we only need to consider terms of total degree
1≤k<
N0−1
n
,
− (p − 1)−1
(3.5)
since for all larger k we have kxk /k!k, k(x + pn y)k /k!k ≤ p−⌊k/N0 ⌋+⌊(k−1)/(p−1)⌋ ≤ p−n .
We will show that all summands in the range (3.5) involving y are actually ≡ 0 (mod pn ).
If y occurs at least twice, then the corresponding summand has norm ≤ p−2n+vp (k!) =
p−n p−n+vp (k!) < p−n , since vp (k!) < k/(p − 1) < n for k as in (3.5) and p > 2N0 . It
remains to consider the terms xi yxk−1−i /k! for 0 ≤ i ≤ k − 1. We can bound the norm
of this term by p−n−⌊i/N0 ⌋−⌊(k−1−i)/N0 ⌋+vp (k!) . For k < p this is clearly ≤ p−n . For k ≥ p
we have
i
k−1
k−1−i
k−1
p−1
+
− vp (k!) >
−2−
≥
− 3 ≥ −1,
N0
N0
N0
p−1
N0
and therefore −n − ⌊i/N0 ⌋ − ⌊(k − 1 − i)/N0 ⌋ + vp (k!) ≤ −n, which establishes the
congruence (3.4). The analogous congruence for log can be proven similarly.
6
In fact, [ibid.] also gives the convergence of exp and log on gl(N0 , Zp )resnilp and GL(N0 , Zp )resunip ,
respectively.
Approximation principle
27
Using these extensions of log and exp, we have the following consequence.
Lemma 3.11. Assume p > 2N0 . Let X ⊂ G be an algebraic subgroup defined over
Qp , and h ∈ Kp with hp ∈ (X(Qp ) ∩ Kp (p))Kp (pm ) for some m ≥ 1. Then h ∈
(X(Qp ) ∩ Kp )Kp (pm−1 ).
Proof. The lemma is trivially true for m = 1, so we can assume that m > 1. Note that
hp ∈ Kp (p), so that h ∈ GL(N0 , Zp )resunip . By the first part of Lemma 3.10 we can write
h = exp y where y = log h ∈ gl(N0 , Zp )resnilp . Then py = log(hp ) ∈ log((X(Qp ) ∩
Kp (p))Kp (pm )) = p(LieQp X ∩ gZp ) + pm gZp by parts 1 and 3 of Proposition 2.10,
and therefore y ∈ (LieQp X ∩ gZp ) + pm−1 gZp . Let y ′ ∈ LieQp X ∩ gZp be such that
y − y ′ ∈ pm−1 gZp . Clearly, y ′ ∈ gl(N0 , Zp )resnilp . By the last part of Lemma 3.10 we have
exp y ∈ (exp y ′)Kp (pm−1 ). On the other hand, by Lemma 2.6 we have exp(y ′ ) ∈ X(Qp ).
The lemma follows.
In the following, for any closed subgroup H ⊂ Kp we will denote by H + the open
normal subgroup of H generated by the set Hresunip .
Proof of Theorem 3.1. First note that we may assume without loss of generality that p >
n(G) since otherwise we can apply Theorem 2.2 to H ′ = H ∩Kp (p′ ) as long as J ≥ [Kp :
Kp (p′ )]. In particular, we may assume in the following that p ≥ max(2N0 + 1, dim G).
Let H ⊂ Kp be an arbitrary open subgroup. The quotient H/H + is finite of order
prime to p. Repeating the argument of the proof of [Nor87, Theorem C] in the setting of
profinite groups, we see that H = H1 H + for a finite subgroup H1 of H of order prime to
p. Applying Proposition 3.5 to H1 , there exists a maximal torus T of G defined over Qp
such that [H : AH + ] ≤ J where A = T (Qp ) ∩ H1 and J is a constant depending only on
G. Clearly, A can be generated by elements a1 , . . . , as , where s is bounded in terms of G
only, namely by the Q-rank of G.
Let ε be as in Lemma 3.9. We will show below that there exists a proper, connected
algebraic subgroup Y of G defined over Qp such that
H + ⊂ (Y (Qp ) ∩ Kp )Kp (p⌈εn/2⌉ ) and NG (Y )(Qp ) ∩ ai Kp (p⌈εn/2⌉ ) 6= ∅, i = 1, . . . , s.
(3.6)
Let us see how to obtain the theorem from this assertion. If Y is normal in G, we simply
take X = Y T . Otherwise, let X = NG (Y )◦ ⊃ Y (which holds since Y is connected) so
that H + ⊂ (X(Qp ) ∩ Kp )Kp (p⌈εn/2⌉ ). The index of X in NG (Y ) is bounded in terms
of G only by Lemma 2.22. Therefore, choosing bi ∈ NG (Y )(Qp ) ∩ ai Kp (p⌈εn/2⌉ ), i =
1, . . . , s, there exists an exponent e, depending only on G, such that bei ∈ X(Qp ) ∩ Kp ,
i = 1, . . . , s. The projection of AH + to the factor group Kp /Kp (p⌈εn/2⌉ ) is generated
by the image of the group H + and the images of the elements bi . We conclude that the
28
Tobias Finis, Erez Lapid
open subgroup Ae H + of H, which has index at most J ′ = es J, is contained in the group
(X(Qp ) ∩ Kp )Kp (p⌈εn/2⌉ ).
It remains to prove (3.6). Assume first that n ≥ 2ε . Consider the open subgroup H ∩
Kp (p) of Kp (p), which is normalized by the elements a1 , . . . , as . Since we assume that
p ≥ dim G, the subset M = log(H ∩ Kp (p)) ⊂ pgZp is by Remark 2.11 a Lie subalgebra,
which is clearly of level pn in gZp and stable under Ad(a1 ), . . . , Ad(as ). (Alternatively,
we may use part 6 of Lemma 2.5 and imitate the proof of Theorem 2.2 above, noting that
the uniform subgroup V is characteristic in H ∩ Kp (p).) From Lemma 3.9 we obtain a
proper, isolated Lie subalgebra I ⊂ gZp such that
M ⊂ (I + p⌈εn⌉ gZp ) ∩ pgZp = pI + p⌈εn⌉ gZp ,
and elements bi ∈ ai Kp (p⌈εn⌉ ), i = 1, . . . , s, such that Ad(bi )I = I. We infer that
H ∩ Kp (p) ⊂ (exp pI)Kp (p⌈εn⌉ ). As in the proof of Theorem 2.2, consider the algebraic
subgroup Y = A(Qp I) of G, which is a proper, connected algebraic subgroup of G
defined over Qp . Clearly, b1 , . . . , bs ∈ NG (Y )(Qp ). We can now invoke Lemma 3.11 to
lift the relation H ∩ Kp (p) ⊂ exp(pI)Kp (p⌈εn⌉ ) ⊂ (Y (Qp ) ∩ Kp (p))Kp (p⌈εn⌉ ) to H + and
obtain H + ⊂ (Y (Qp ) ∩ Kp )Kp (p⌈εn⌉−1 ). Thus, (3.6) holds by the assumption on n.
Consider now the case where 1 ≤ n < 2ε . Let H̄ be the image of H 6= Kp in
Kp /Kp (p) ≃ G(Fp ). Since G is simply connected, H̄ is a proper subgroup of G(Fp )
(for p > n(G)) by (2.1) and Nori’s algebraic envelope Ȳ of H̄ is a proper subgroup of
GFp (cf. the proof of Proposition 3.3). It follows from Theorems 3.6 and 3.7 that LieFp Ȳ
is spanned by log(p) x, x ∈ H̄unip , and is a proper Lie subalgebra of gFp . Let M be the
inverse image of LieFp Ȳ under gZp → gZp /pgZp , so that M is a proper Zp -Lie subalgebra
of gZp . Since Hresunip is the inverse image of H̄unip under the reduction map H → H̄, the
P
commutativity of (3.2) gives M = h∈Hresunip Zp log h + pgZp .
Since p > n(G), we can use Lemma 3.9 to obtain the inclusion M ⊂ I + pgZp for a
proper, isolated subalgebra I ⊂ gZp and elements bi ∈ ai Kp (p), i = 1, . . . , s, such that
Ad(bi )I = I. As above, we can assume that I = LieQp Y ∩ gZp for a proper, connected
algebraic subgroup Y of G defined over Qp and retain the property that I is stable under
the operators Ad(bi ). Since H + is generated by elements h with hp ∈ Kp (p), and for such
an h we have hp ∈ exp(pM) ⊂ exp(pI)Kp (p2 ) = (Y (Qp ) ∩ Kp (p))Kp (p2 ), we obtain
from Lemma 3.11 that H + ⊂ (Y (Qp ) ∩ Kp )Kp (p), so that (3.6) holds in this case as well.
(Alternatively, Nori’s proof shows that Ȳ extends to a smooth group scheme over Zp , and
in particular lifts to characteristic zero.)
If we are willing to use Theorem 4.3, the main result of §4, then we may simplify the
proof above (see Remark 4.4 below).
Approximation principle
29
4 Closed subgroups of G(Zp) and closed subalgebras of
gZp
4.1 Statement of the correspondence
We continue to use the conventions of §2.1. In this section we will establish (for p large
with respect to G) a correspondence between subgroups of G(Zp ) which are residually
exponentially generated and Zp -Lie subalgebras of gZp which are residually nilpotently
generated. On the one hand, this bijection extends previously known results on pro-p
subgroups of G(Zp ), and on the other hand it is compatible, via reduction modulo p, with
Nori’s correspondence between subgroups of G(Fp ) and subalgebras of gFp .
We first recall that Nori’s two theorems (Theorems 3.6 and 3.7) establish for p > n(G)
bijective correspondences between three different collections of objects: subgroups H̄ of
G(Fp ) that are generated by their elements of order p (i.e. for which H̄ + = H̄), nilpotently
generated subalgebras of gFp , and exponentially generated algebraic subgroups H of GFp .
We will focus on the bijection between the first two collections, which we can describe
as follows: to a subgroup H̄ of G(Fp ) we associate L̄(H̄), the Fp -span of log(p) H̄unip
(which is a Lie algebra), and to h̄ the subgroup Ḡ(h̄) of G(Fp ) generated by exp(p) h̄nilp.
Regarding the third collection, to an algebraic group H ⊂ GFp correspond its Lie algebra
LieFp H and the finite group H(Fp )+ . The opposite bijections were described in §3 above.
Next, we recall the following correspondence due to Klopsch [Klo05] (see also [GSK09])
generalizing Ilani’s correspondence.
Theorem 4.1 (Lazard–Ilani–Klopsch). Assume that p ≥ max(N0 + 2, dim G). Then the
exponential and logarithm maps (applied to subsets of gZp ,resnilp and G(Zp )resunip , respectively) induce mutually inverse bijections between the set of all closed pro-p subgroups
H ⊂ G(Zp ) and the set of all closed Lie subalgebras h ⊂ gZp consisting of residually
nilpotent elements.
Remark 4.2. Just like in Ilani’s work, the correspondence in [Klo05] is stated in terms
of the intrinsic Lie algebra structure on a saturable pro-p group in the sense of Lazard.
However, by [Laz65, Proposition IV.3.2.3] the intrinsic Lie algebra is isomorphic to the
one defined using the logarithm map (cf. [DdSMS99, §7.2]).
Throughout we assume that p ≥ max(N0 + 2, dim G). In fact, we also assume that
p > 2N0 , so that the exponential and logarithm bijections satisfy the properties of Lemma
3.10.
The basic construction underlying our extension of this theorem to more general subgroups of G(Zp ) is a generalization of Nori’s correspondence. Let H be a subgroup of
G(Zp ). We associate to H the Zp -submodule L(H) ⊂ gZp generated by the set log Hresunip .
Tobias Finis, Erez Lapid
30
We will show in Proposition 4.7 below that L(H) is in fact always a Zp -Lie algebra provided that p > dim G + 1. Conversely, to any Lie subalgebra h ⊂ gZp we associate the
closed subgroup G(h) of G(Zp ) topologically generated by the set exp hresnilp.
By Theorem 4.1, for pro-p groups H we have L(H) = log H and for subalgebras h
contained in gZp ,resnilp we have G(h) = exp h.
We extend Theorem 4.1 to a correspondence between the following classes of subgroups and subalgebras. On one side, we consider the set SG rug (G(Zp )) of closed subgroups H ⊂ G(Zp ) which are topologically generated by their residually unipotent elements. Clearly, H ∈ SG rug (G(Zp )) precisely when H + = H, where H + is as in §3 above.
It is equivalent to demand that H̄ ⊂ G(Fp ), the reduction modulo p of H, is generated by
its elements of order p.
On the other side, we consider the set SArng (gZp ) of all closed Lie subalgebras of gZp
which are spanned over Zp by their residually nilpotent elements. Clearly, h ∈ SArng (gZp )
if and only if the reduction h̄ is a nilpotently generated subalgebra of gFp .
Theorem 4.3. Suppose that p is large with respect to G. Then
L
−
⇀
SG rug (G(Zp )) ↽
− SArng (gZp ).
G
These bijections preserve the level of open subgroups and subalgebras, extend those of
Theorem 4.1, and are compatible with those of Nori under reduction modulo p, that is if
H ∈ SG rug (G(Zp )) and h = L(H) then the image H̄ of H in G(Fp ) corresponds under
Nori to the image h̄ of h in gFp , i.e., h̄ = L̄(H̄).
The theorem will be proved below. Clearly, the special case of subgroups and subalgebras of level p reduces to Nori’s correspondence.
Remark 4.4. Using Theorem 4.3 we can streamline the proof of Theorem 3.1. Namely,
to prove (3.6) (with ε instead of ε/2) we apply Lemma 3.9 to M = L(H + ), which is
of level pn in gZp and stable under Ad(a1 ), . . . , Ad(as ). We obtain a proper, isolated
Lie subalgebra I ⊂ gZp such that M ⊂ I + p⌈εn⌉ gZp and elements bi ∈ ai Kp (p⌈εn⌉ ),
i = 1, . . . , s, such that Ad(bi )I = I. We infer that H + ⊂ G(I)Kp (p⌈εn⌉ ). Taking the
algebraic subgroup Y = A(Qp I) of G, which is a proper, connected algebraic subgroup
of G defined over Qp , we see that b1 , . . . , bs ∈ NG (Y )(Qp ) and H + ⊂ Y (Qp )Kp (p⌈εn⌉ ),
as required. We opted to include the original proof of Theorem 3.1 since it is simpler and
more self-contained.
4.2 The Lie algebra associated to a subgroup
We start with some simple considerations (in the spirit of [Nor87, §1]) and establish that
L(H) is indeed a Lie subalgebra of gZp (for p > dim G + 1).
Approximation principle
31
Lemma 4.5. Let V be a finitely generated free Zp -module, U a submodule of V and
s ∈ EndZp (V )resnilp . Assume that sU ⊂ U. Then s U ∈ EndZp (U)resnilp . In particular,
srk U (U) ⊂ pU.
Proof. We know that sn → 0 as n → ∞. Clearly, this implies that sn
n → ∞, or s U ∈ EndZp (U)resnilp .
U
= (s U )n → 0 as
Corollary 4.6. Let V be a free Zp -module of rank r < p − 1, U a submodule of V and
s ∈ EndZp (V )resnilp . Then U is s-invariant if and only if it is exp(s)-invariant.
Proof. Suppose that U is s-invariant. Then it follows from Lemma 4.5 that for all n ≥ 1
we have (n!)−1 (s U )n ∈ pkn EndZp (U) where kn = ⌊n/ rk U⌋ − ⌊(n − 1)/(p − 1)⌋ ≥ 0.
Since kn → ∞ as n → ∞ and U is closed, we conclude that U is invariant under exp s.
The converse is proved in a similar way using log.
Proposition 4.7. Assume that p > dim G + 1 and let H be a subgroup of G(Zp ). Then the
Zp -submodule h = L(H) of gZp generated by the set log Hresunip is a Zp -Lie subalgebra.
For p > n(G), the image of h in gFp is the Lie algebra LieFp H̃ of Nori’s algebraic
envelope H̃ ⊂ GFp of the image H̄ of H in G(Fp ).
We remark that under our standing assumption p > 2N0 the image of h in gFp is always
a Lie algebra (to see this, combine [Nor87, Lemma 1.6] with (3.2)).
Proof. Clearly, h is Ad(h)-invariant for any h ∈ H, and in particular for any h ∈ Hresunip .
We have Ad(h) = exp(ad(log h)) for any h ∈ Hresunip . By Corollary 4.6 we conclude
that h is ad(log h)-invariant for any h ∈ Hresunip . The first claim follows.
The second assertion follows from the commutativity of (3.2) and from Theorems 3.6
and 3.7 as in the proof of Theorem 3.1.
Remark 4.8. At this stage we can already prove that if H is an open subgroup of G(Zp )
of level pn , then h = L(H) has level pn or pn−1 . (Eventually we will prove that h has level
pn .) Indeed, log(H ∩ Kp (p)) is a Lie subalgebra of gZp of level pn by Remark 2.11, and
hp ∈ H ∩ Kp (p) for all h ∈ Hresunip . Therefore log(H ∩ Kp (p)) ⊂ h ⊂ p−1 log(H ∩
Kp (p)), which shows that h has level pn or pn−1 . Moreover, it is also clear at this point
that if H has level p, then h has level p.
The remaining parts of Theorem 4.3 are easily deduced from the following two statements which will be proved below.
For any H ∈ SG rug (G(Zp )) we have L(H) ∩ pgZp = log(H ∩ Kp (p)).
For any h ∈ SArng (gZp ) we have G(h) ∩ Kp (p) = exp(h ∩ pgZp ).
(4.1a)
(4.1b)
32
Tobias Finis, Erez Lapid
Indeed, given H ∈ SG rug (G(Zp )) let h = L(H) and H ′ = G(h). It is clear that H ′ ⊃ H.
By Nori’s theorems, the reductions modulo p of H ′ and H coincide. Moreover, by (4.1a)
and (4.1b) we have H ′ ∩ Kp (p) = exp(h ∩ pgZp ) = H ∩ Kp (p). Therefore, H ′ = H. The
other direction can be shown in a similar way.
The following lemma provides a complement to Theorem 4.1 that will be useful below.
Lemma 4.9. Let p > dim G + 1. Let P = exp p be a pro-p subgroup of G(Zp ) and h ∈
G(Zp )resunip a residually unipotent element normalizing P . Then we have [log h, p] ⊂ p,
hP ⊂ G(Zp )resunip and log(P h) = log(hP ) = log h + p.
Similarly, if p is a closed Lie subalgebra of gZp contained in gZp ,resnilp, and a residually
nilpotent element u ∈ gZp ,resnilp satisfies [u, p] ⊂ p, we have u + p ⊂ gZp ,resnilp and
exp(u + p) = exp(u)P = P exp(u), where P = exp p.
Proof. Since h normalizes P , the group Q generated by h and P is a pro-p group. Therefore hP ⊂ Q ⊂ G(Zp )resunip . Furthermore, hP h−1 = P implies that Ad(h)p = p, and
by Corollary 4.6 we obtain [log h, p] ⊂ p.
Analogously, if we have p contained in gZp ,resnilp, and u ∈ gZp ,resnilp satisfies [u, p] ⊂ p,
then the Lie algebra generated by ū and p̄ inside gFp clearly consists of nilpotent elements,
and therefore u + p ⊂ gZp ,resnilp.
P
Let p ⊂ gZp ,resnilp and u ∈ gZp ,resnilp with [u, p] ⊂ p. Denote by Φ(x, y) = ∞
n=1 Φn (x, y)
the Hausdorff series [Bou98, §II.6.4], where Φn is a homogeneous Lie polynomial of degree n, and Φ1 (x, y) = x+y. Since Zp u+p is a Zp -Lie algebra contained in gl(N0 , Zp )resnilp,
the Hausdorff series converges on (Zp u + p)2 and we have Φ(x, y) = log(exp x exp y) for
all x, y ∈ Zp u + p ([Bou98, §II.8.3,II.8.4]).
It remains to show that Φ(u, p) ⊂ u + p and Φ(−u, u + p) ⊂ p. as these conditions
are equivalent to exp(u)P ⊂ exp(u + p) and exp(u + p) ⊂ exp(u)P , respectively. (The
relation log(P h) = log(hP ) = log h + p will then follow by taking log.) For this, it
suffices to show that Φn (u, p), Φn (−u, u + p) ⊂ p for all n ≥ 2. By [Klo05, Lemma
4.2] applied to K = p, H = Zp u + p and j = 1, and taking into account that p >
dim G + 1, the iterated commutator of an element of p and n − 1 elements of Zp u + p lies
n−1
in p⌊ p−1 ⌋ p. On the other hand, by a standard estimate for the denominator of Φn [Bou98,
n−1
§II.8.1, Proposition 1], we know that p⌊ p−1 ⌋ Φn (x, y) is a Zp -linear combination of iterated
commutators of degree n in x and y. This implies that indeed Φn (u, p), Φn (−u, u+p) ⊂ p,
which finishes the proof.
4.3 Application of Nori’s theorem
The proof of (4.1a) uses crucially Nori’s correspondences recalled above. We first summarize some easy complements to Nori’s theorems in the following lemma.
Approximation principle
33
Lemma 4.10. Let H̄ be a subgroup of G(Fp ) with H̄ + = H̄, h̄ = L̄(H̄) ⊂ gFp the
associated Lie algebra, and H̃ the associated exponentially generated algebraic subgroup
of GFp . Let P̄ be the maximal normal p-subgroup of H̄ and P̃ the unipotent radical of H̃
(which is also the solvable radical of H̃). Finally, let S sc be the simply connected covering
group of the semisimple group H̃/P̃ . Then
1. The following four conditions are equivalent: H̄ is a p-group; H̃ is a unipotent
algebraic group; h̄ is contained in gFp ,nilp; h̄ is a nilpotent Lie algebra.
2. H̄1 is a normal subgroup of H̄2 if and only if H̃1 is a normal subgroup of H̃2 if and
only if h̄1 is an ideal of h̄2 .
3. P̄ corresponds to P̃ and to the nilradical of h (which equals the radical of h̄ as well
as the largest ideal of h̄ contained in gFp ,nilp ).
4. The covering map κ : S sc → H̃/P̃ induces a surjection S sc (Fp ) → H̄/P̄ (cf.
[Nor87, Remark 3.6]). The adjoint action of H̄ on h̄ induces an action of H̄/P̄ on
h̄/p̄ which is compatible under the above map with the adjoint action of S sc on its
Lie algebra.
5. A Sylow p-subgroup of H̄ corresponds to a maximal unipotent subgroup of H̃ (i.e.,
to the unipotent radical of a Borel subgroup) and to a maximal nilpotent subalgebra
of h̄ contained in gFp ,nilp , or equivalently, to the nilradical of a maximal solvable
subalgebra of h̄.
Proof. We only need to observe that if h is a nilpotently generated Lie subalgebra of
gl(N0 , Fp ) then h is nilpotent if and only if h is solvable if and only if h is contained in
gl(N0 , Fp )nilp. This follows from Engel’s and Lie’s theorems (the latter is valid over the
algebraic closure for p > N0 ).
Let now H ∈ SG rug (G(Zp )) and write h = L(H). Let H̄ be the image of H in G(Fp ).
Clearly, H acts on h by Ad. Let P be the largest normal pro-p subgroup of H and P̄ the
image of P in G(Fp ). Write S := H/P . It is clear that P contains H ∩ Kp (p). Thus,
under reduction modulo p we have the isomorphism S ≃ H̄/P̄ . Also, p = log P is a Lie
subalgebra of h by Theorem 4.1, and by Corollary 4.6 in fact an ideal. It remains to show
that h ∩ pgZp ⊂ p, since in this case
h ∩ pgZp = p ∩ pgZp = log(P ∩ Kp (p)) = log(H ∩ Kp (p)).
Let s = h/p. Our task is to show that the canonical surjective map π : s → s̄ = h̄/p̄
is an isomorphism. Note first that the adjoint action of H on h preserves p and therefore induces an action of H on s. Moreover, for h ∈ Hresunip and p ∈ P we have
Tobias Finis, Erez Lapid
34
Table 1: Nori’s correspondence
subgroups H of G(Fp )
with H + = H
exponentially generated algebraic subgroups of GFp
nilpotently generated Lie
subalgebras of gFp
p-groups
unipotent groups
Lie subalgebras contained
in gFp ,nilp
normal subgroups
normal subgroups
ideals
unipotent (or solvable) radical
nilradical (or radical)
a maximal unipotent subgroup
the nilradical of a maximal
solvable subalgebra
maximal
subgroup
normal
a Sylow p-subgroup
p-
Ad(p)(log h) = log php−1 = log[p, h]h ∈ log(P h) = log h+p by Lemma 4.9. Therefore,
P acts trivially on s, and the H-action on s descends to an action of S. Under the map
π : s → s̄, this action is compatible with the canonical action of H̄/P̄ ≃ S on s̄ = h̄/p̄.
To study the injectivity of π, we apply Nori’s correspondence (Theorem 3.7) to the
groups H̄ and P̄ and associate to them their algebraic envelopes H̃ and P̃ inside GFp . By
Lemma 4.10, the group P̃ is the unipotent radical of H̃, and the quotient group H̃/P̃ is
therefore a semisimple algebraic group defined over Fp . Let S sc be the simply connected
covering group of the group H̃/P̃ . The quotient S ≃ H̄/P̄ is isomorphic to the quotient
of S sc (Fp ) by a central subgroup K. We can therefore regard s and s̄ as representation
spaces of S sc (Fp ). The action on s̄ is given by the adjoint representation of S sc on its Lie
algebra LieFp S sc .
Let N ⊃ P be a Sylow pro-p subgroup of H and n = log N ⊂ h the associated Lie
algebra. By Lemma 4.10, the image of n in h̄ is the Lie algebra of a maximal unipotent
subgroup of H̃. Passing to the quotient by p̄, we obtain that the image of n in s̄ is the
Lie algebra of a maximal unipotent subgroup of S sc . Consider the kernel ker π = (p +
pgZp ) ∩ h/p ⊂ s of the map π. If an element log n + p, where n ∈ N, lies in ker π,
then log n ∈ p + pgZp , which by Lemma 4.9 implies that log(nm) ∈ pgZp for a suitable
m ∈ P . Consequently, nm ∈ H ∩ Kp (p) ⊂ P and therefore n ∈ P . This means that the
restriction of π to the subspace n/p ⊂ s is injective. Finally, by the very definition of h,
the space n/p spans s under the action of S sc (Fp ).
The assertion (4.1a) immediately follows from the following result on characteristic p
representations of the group S sc (Fp ). Note that dim s ≤ dim G in the case at hand.
Proposition 4.11. Let S sc be a simply connected semisimple algebraic group defined
over Fp , s̄ = LieFp S sc with the adjoint representation and π : s → s̄ a surjection of
Fp -representations of S sc (Fp ). Assume that p ≥ 2 dim s and that there exists a subspace
n ⊂ s such that
Approximation principle
35
1. The space s is spanned as a representation of S sc (Fp ) by n.
2. The image n̄ = π(n) ⊂ s̄ of n under π is the Lie algebra of a maximal unipotent
subgroup of S sc .
3. The induced map π|n : n → n̄ is an isomorphism.
Then the map π is an isomorphism.
4.4 Proof of Proposition 4.11
To finish the proof of (4.1a), it remains to prove Proposition 4.11. For this we need to
recall the representation theory of S sc (Fp ) in characteristic p.
Q
Write S sc = i ResFqi /Fp Si , where each qi is a power of p and Si is an absolutely
Q
simple, simply connected algebraic group defined over Fqi . Set S = i Si , a semisimple,
Q
simply connected group defined over Fp , and let σ = i Frobqi , a Steinberg endomorphism of S [Ste68]. We have then S sc (Fp ) = S σ . The σ-stable Borel subgroups of S
are in one-to-one correspondence with the Borel subgroups of S sc defined over Fp . We
take the σ-stable Borel subgroup B of S corresponding to the normalizer of the subspace
n̄ ⊂ s̄. We also fix a σ-stable maximal torus T of S contained in B, and let U ⊂ B be the
unipotent radical of B, a maximal unipotent subgroup of S stable under σ.
Let Φ be the root system of S with respect to T , Φ+ ⊂ Φ the system of positive roots
associated to B, and ∆ the set of simple roots. For any root α ∈ Φ we set q(α) = qi if α
Q
belongs to the simple factor Si in the factorization S = i Si above. By [Ste68, §11.2,
11.6], there exists a permutation ρ of Φ, preserving Φ+ and ∆, such that σ ∗ ρα = q(α)α
for all α ∈ Φ, where σ ∗ denotes the action of σ on the weight lattice X = X ∗ (T ). Let
X + ⊂ X the set of dominant weights and let
Xσ+ = {λ ∈ X : 0 ≤ hλ, α∨i < q(α),
α ∈ ∆} ⊂ X + .
For each λ ∈ X + let L(λ) be the irreducible representation of S of highest weight
λ with coefficients in Fp . By Steinberg’s algebraicity theorem [Ste68, Theorem 13.3] the
irreducible representations of S σ over the field Fp are precisely the restrictions to S σ of
the S-representations L(λ) for λ ∈ Xσ+ .
For a T -representation V and a character χ ∈ X we write V (χ) for the χ-eigenspace
in V , and similarly for T σ -representations and Fp -valued characters of T σ .
σ
Lemma 4.12. For any λ ∈ Xσ+ we have L(λ)U = L(λ)U = L(λ)(λ) , and this space
is one-dimensional. Moreover, any non-trivial U σ -invariant subspace of L(λ) contains
L(λ)U .
Tobias Finis, Erez Lapid
36
Proof. The first equality is contained in [Cur70, Theorem 4.3], and the second equality is
standard (cf. [Bor70, Theorem 5.3]). The second part follows from the first one and the
well-known fact that any representation of a p-group over a field of characteristic p admits
a non-trivial vector fixed under the action [Ser77, §8.3, Proposition 26].
Lemma 4.13. Let V be a representation of S σ over Fp with dim V <
µ ∈ X with V (µ) 6= 0 we have V (µ) = V (µ|T σ ) .
p+1
.
2
Then for every
′
Proof. We need to show that for µ, µ′ ∈ X with V (µ) , V (µ ) 6= 0 the identity µ|T σ = µ′ |T σ
implies that µ = µ′ .
For this, we first claim that every µ ∈ X with V (µ) 6= 0 satisfies
|hµ, α∨ i| ≤
q(α) − 1
for all α ∈ ∆.
2
Since V 7→ V (µ) is an exact functor, we may assume without loss of generality that V is
irreducible, say V = L(λ). We may also assume that S is simple. Then q(α) = q = pr
Pr−1 i
for some r ≥ 1 and we may write λ = i=0
p λi with 0 ≤ hλi , α∨i < p for all α ∈ ∆.
By Steinberg’s tensor product theorem [Ste63], the S-representation L(λ) is isomorphic
to the tensor product of the representations L(pi λi ) for i = 0, . . . , r − 1, and L(pi λi ) is
Qr−1
dim L(λi )
isomorphic to the i-th Frobenius twist of L(λi ). Therefore dim L(λ) = i=0
and in particular dim L(λi ) < (p + 1)/2 for all i. By [Jan97, Lemma 1.2], this implies
that hλi , α∨ i < (p − 1)/2 for all α ∈ Φ. Hence, |hµi , α∨ i| < (p − 1)/2 for any weight µi
Pr−1 i
i
of L(λi ). Suppose that µ ∈ X with L(λ)(µ) ≃ ⊗L(pi λi )(p µi ) 6= 0. Then µ = i=0
p µi
(µi )
with L(λi )
6= 0 for all i and therefore
∨
|hµ, α i| ≤
r−1
X
i=0
r−1
p−1X i q−1
p |hµi , α i| <
p =
,
2 i=0
2
∨
i
which establishes the claim.
For each α ∈ Φ let d(α) be the number of elements in the ρ-orbit of α. The kernel of
the restriction map µ 7→ µ|T σ is then given by
d(α)−1
{µ ∈ X : q(α)
d(α)
X
−1
µ, σ j α∨ for all α ∈ ∆}.
j=0
For all µ ∈ X with V (µ) 6= 0 we have here
d(α)−1
|
X
j=0
d(α)−1
j
∨
µ, σ α | = |
X
j=0
d(α)−1
1 X
1
q(α) µ, ρ α | <
q(α)j (q(α)−1) = (q(α)d(α) −1).
2 j=0
2
j
∨
j
Thus if µ, µ′ ∈ X are weights of V with µ|T σ = µ′ |T σ , then necessarily
d(α)−1
d(α)−1
X
j=0
j
µ, σ α
∨
=
X
j=0
µ′ , σ j α ∨ ,
Approximation principle
37
which necessitates that hµ, σ j α∨i = hµ′ , σ j α∨i for all j. Thus, hµ, α∨i = hµ′ , α∨i for all
α ∈ ∆, so that µ = µ′ as required.
We will now use a rather deep result of Jantzen [Jan97] (which crucially relies on
earlier work by Cline–Parshall–Scott–van den Kallen [CPSvdK77]). Namely, every Fp representation of S σ of dimension less than p − 1 is completely reducible, and consequently isomorphic to a direct sum of representations L(λ) with λ ∈ Xσ+ . (See [McN99]
for an extension of this result. Note that by [AJL83] there exists an Fp -representation of
SL(2, Fp ) of dimension p − 1 that is not completely reducible.)
Lemma 4.14. Let V be a representation of S σ over Fp with dim V <
a Bσ -invariant subspace that spans V as a S σ -representation.
p+1
.
2
Let U ⊂ V be
1. Let φ : V → L(λ), λ ∈ Xσ+ , be a non-trivial S σ -map. Then φ(U (λ|T σ ) ) =
L(λ)(λ|T σ ) .
2. The multiplicity of L(λ), λ ∈ Xσ+ , in V is at most dim U (λ|T σ ) . In particular, the
restriction to T σ of the highest weight of any irreducible constituent of V appears
in the T σ -space U.
Proof. Consider the first assertion. Since φ(U) spans L(λ) as an S σ -space, we have
φ(U) 6= 0. Combining Lemmas 4.12 and 4.13, and noting that φ(U) is U σ -invariant,
σ
we have φ(U)U = L(λ)(λ) = L(λ)(λ|T σ ) , which is a one-dimensional space. Thus,
φ(U)(λ|T σ ) = L(λ)(λ|T σ ) . Since U is Bσ -invariant, we also have φ(U (λ|T σ ) ) = φ(U)(λ|T σ ) .
The first part follows.
For the second assertion assume on the contrary that the multiplicity n of L(λ) in V
is strictly greater than dim U (λ|T σ ) . By the semisimplicity of V there exists a surjection
φ : V → W = L(λ)n of S σ -modules. Then φ(U) spans W as an S σ -space while φ(U)(λ)
is a proper subspace of the n-dimensional space W (λ) . Now the restriction map from
HomS σ (W, L(λ)) to HomFp (W (λ) , L(λ)(λ) ) is an isomorphism. Therefore there exists a
surjective S σ -homomorphism ψ : W → L(λ) with ψ(φ(U)(λ) ) = 0. But then the composition ψ ◦ φ : V → L(λ) contradicts the first assertion, which finishes the proof.
For the proof of Proposition 4.11 we have to study the adjoint representation of S σ =
S sc (Fp ) on the Lie algebra LieFp (S sc ) ⊗ Fp . It decomposes as the direct sum of the representations L(pi α̃), where α̃ ranges over the highest roots of the irreducible components
of Φ and i is such that pi |q(α̃).
We will prove a lemma on the Bσ -representations L(pi α̃). For this we first need the
following easy lemma on root systems, a variant of a standard result on sums of roots
[Bou02, §VI.1.6, Proposition 19].
Tobias Finis, Erez Lapid
38
Lemma 4.15. Let Ψ be a root system, Ψ+ a system of positive roots for Ψ, and ∆ the
P
associated set of simple roots. Let β ∈ Ψ+ and write β = α∈∆ nα α with non-negative
integers nα . Then the coefficient nα of a simple root α ∈ ∆ is positive if and only if there
exists a sequence β1 , . . . , βk ∈ Ψ+ of positive roots starting with β1 = α and ending with
βk = β such that βi+1 − βi ∈ Ψ+ for all i = 1, . . . , k − 1.
Proof. Clearly, the existence of a sequence β1 , . . . , βk as above implies that nα > 0. It
P
remains to show the reverse implication. For this we proceed by induction on γ∈∆ nγ .
There has to exist a simple root γ ∈ ∆ with hβ, γ ∨i > 0, which implies that either β = γ
or β − γ ∈ Ψ+ . In the first case, β = γ = α and there is nothing to prove. If in the
second case we have γ = α, then we may simply take k = 2. Otherwise, we may apply
the induction hypothesis to β − γ and obtain a sequence β1 , . . . , βk−1 ∈ Ψ+ with β1 = α,
βk−1 = β − γ and βi+1 − βi ∈ Ψ+ . Setting βk = β yields the assertion.
For any irreducible root system Ψ we denote by Ψns the set of all roots which are not
contained in the Z-span of the short simple roots. (By convention, if Ψ is simply laced
then all roots are long, so that Ψns = Ψ.) If Ψ+ is a system of positive roots, we write
P
+
+
+
Ψ+
ns = Ψns ∩ Ψ . A root β =
α∈∆ nα α ∈ Ψ is contained in Ψns if and only if nα > 0
for at least one long simple root α ∈ ∆ ∩ Ψlong . Note that for any non-zero dominant
weight the coefficients with respect to the basis ∆ are all positive, since this holds for the
fundamental weights [Hum78, §III.13, exercise 8]. It follows that the highest short root
α̃short of Ψ+ belongs to Ψ+
ns .
Lemma 4.16. Assume that p ≥ 2 dim S sc . Let α̃ be the highest root of an irreducible
component Ψ of Φ and pi a divisor of q(α̃). Then in the S σ -representation L(pi α̃), the
Bσ -span D of the subspace
X
i
L(pi α̃)(p α) ⊂ L(pi α̃)
α∈∆∩Ψlong
i
is the sum of the weight spaces L(pi α̃)(p β) for β ∈ Ψ+
ns .
Proof. By Frobenius twist, we reduce to the case i = 0. Note that the S-representation
L(α̃) is nothing else than the adjoint representation of the simple factor of S containing
α̃. For each β ∈ Ψ the corresponding weight space L(α̃)(β) = L(α̃)(β|T σ ) (where we use
Lemma 4.13 to identify T - and T σ -eigenspaces) is one-dimensional.
Since it is stable under the action of T σ , the Bσ -span of any sum of weight spaces is
again a sum of weight spaces. Furthermore, it is clear that the weights appearing in D are
(β)
roots in Ψ+
, β ∈ Ψ+
ns . It remains to show that each weight space L(α̃)
ns is contained in
D.
For this consider the following claim: for any β ∈ Ψ+ , the Bσ -span of L(α̃)(β) contains
′
L(α̃)(β ) for all β ′ ∈ Ψ+ for which β ′ − β is a positive root. Granted this claim, we may
Approximation principle
39
argue as follows: by Lemma 4.15, for every β ∈ Ψ+
ns there exists a sequence β1 , . . . , βk ∈
+
Ψ , where β1 = α is a long simple root, βk = β and βi+1 − βi ∈ Ψ+ for all i =
1, . . . , k − 1. By our claim, for any i = 1, . . . , k − 1, the Bσ -span of L(α̃)(βi ) contains
L(α̃)(βi+1 ) . This obviously implies the assertion.
It remains to prove the claim above. For each β ∈ Ψ let uβ ∈ L(α̃)(β) be a non-trivial
element of the one-dimensional weight space. The action of a root unipotent xγ (ξ) ∈ S
on uβ is given by
xγ (ξ)uβ = uβ + C1,βγ ξuβ+γ +
X
Ci,βγ ξ i uβ+iγ
(4.2)
i≥2
with Ci,βγ ∈ Fp and C1,βγ 6= 0, since p > 3 and therefore p does not divide any structure
constant of a semisimple Lie algebra.
Let now β, β ′ ∈ Ψ+ with γ = β ′ −β ∈ Ψ+ and let d(γ) be the cardinality of the ρ-orbit
of γ. It is enough to show that the U σ -span of uβ contains an element with non-vanishing
projection to the β ′ -weight space, since we can then use the T σ -action to get the desired
inclusion.
Note that by [Ste68, 11.2] the σ-action on the root unipotents xα (ξ) is given by
×
σxα (ξ) = xρα (c(α)ξ q(α) ) for suitable c(α) ∈ Fp . For d(γ) = 1, application of a suitable
xγ (ξ) ∈ U σ with ξ 6= 0 to uβ provides an element in the U σ -span of uβ with non-vanishing
′
projection to L(α̃)(β ) , as asserted. In the general case, we can always find a product
Q
x = ni=1 xγi (ξi ) ∈ U σ with ξ1 6= 0 and positive roots γ1 , . . . , γn ∈ Ψ+ , such that γ1 = γ,
P
and ni=1 νi γi = γ for integers νi ≥ 0 if and only if ν1 = 1, ν2 = · · · = νn = 0. By
′
repeated application of (4.2) we obtain that xuβ has non-vanishing projection to L(α̃)(β ) ,
which finishes the proof.
Proof of Proposition 4.11. By tensoring the spaces s, s̄, n and n̄ with Fp , we may pass
to representations of S σ and Bσ over Fp , which for the remainder of this proof we will
denote by the same letters. (Note that n is Bσ -invariant since n̄ is.) The same applies to
the surjection π : s → s̄. Obviously, to establish the proposition it is equivalent to show
that π is an isomorphism in the new setting.
As noted above, s̄ ≃ LieFp (S sc ) ⊗ Fp decomposes as the direct sum of the representations L(pi α̃), where α̃ ranges over the highest roots of the irreducible components of Φ
and pi |q(α̃). The weights of T σ on n ≃ n̄ are given by pi α, where α ∈ Φ+ and pi divides
q(α), and they all have multiplicity one.
Applying Lemma 4.14 to U = n ⊂ V = s, we obtain that the possible highest
weights of the irreducible constituents of s are pi α̃, pi |q(α̃), and pi α̃short , pi |q(α̃short ),
where α̃ ranges over the highest roots of the irreducible components of Φ and α̃short
over the highest short roots of the components that are not simply laced. Moreover, all
irreducible constituents of s appear with multiplicity one.
Tobias Finis, Erez Lapid
40
It only remains to show that no representation L(pi α̃short ) can be a quotient of s. Denote by l the (unique) lift of the representation s̄ to a subspace of s. Note that for any
i
long root β the corresponding weight space s(p β) is necessarily contained in l since the
non-zero weights in L(α̃short ) are the short roots.
Let α̃short be the highest short root of a component Ψ of Φ that is not simply laced. For
i
any long simple root α of Ψ the weight space n(p α) is contained in l. By Lemma 4.16,
i
i
i
i
the weight space n(p α̃short ) ≃ n̄(p α̃short ) lies in the Bσ -span of the spaces n(p α) ≃ n̄(p α) ,
i
α ∈ ∆ ∩ Ψlong . Therefore n(p α̃short ) is contained in l. By the first part of Lemma 4.14, it
follows that the representation s/l does not admit L(pi α̃short ) as a quotient.
This shows that s = l, and that π is an isomorphism, as asserted.
4.5 From subalgebras to subgroups
For the proof of the remaining identity (4.1b), we need the following easy consequence
of the standard presentations of the groups S sc (Fp ), where S sc is a simply connected
semisimple group defined over Fp .
Lemma 4.17. Let S sc be a simply connected semisimple group defined over Fp . Let Γ =
Γ(S sc ) be the group defined by generators γu , u ∈ S sc (Fp )unip , and relations
1. (restricted multiplication) γu1 γu2 = γu1 u2 for any unipotent subgroup Ũ of S sc
defined over Fp and u1 , u2 ∈ Ũ(Fp ),
for any u1 , u2 ∈ S sc (Fp )unip .
2. (conjugation) γu1 γu2 γu−1
= γu1 u2 u−1
1
1
Then the map γu 7→ u, u ∈ S sc (Fp )unip , extends to an isomorphism of groups s : Γ →
S sc (Fp ).
Proof. It is clear that the map γu 7→ u extends to a group homomorphism s, which is
surjective since S sc (Fp ) is generated by its unipotent elements. It remains to show that
s is injective. For this, we may use any presentation (X, R) of S sc (Fp ) with generators
X ⊂ S sc (Fp )unip and relations R. We only need to verify that the elements γu , u ∈ X
generate Γ and satisfy the relations R in Γ, i.e. that R is a consequence of the relations
above.
We may assume that S sc is simple over Fp . For S sc of Fp -rank one the assertion follows
(at least for p > 7, when the Schur multiplier of S sc (Fp ) is trivial) from [Cur65, Theorem 1.2]. Alternatively, it follows from [Ste62, Théorème 3.3] and [Ste81, Theorem 5.1]
without restriction on p. Namely, we consider a maximal unipotent subgroup U ⊂ S sc
and the opposite unipotent subgroup Ū. By the Bruhat decomposition, every unipotent
element of S sc (Fp ) is either contained in U(Fp ) or of the form uūu−1 for u ∈ U(Fp ) and
ū ∈ Ū (Fp ), and therefore Γ is generated by the γu ’s for u ∈ U(Fp ) ∪ Ū(Fp ). Furthermore,
the relations of [Cur65, Theorem 1.2] follow from the conjugation relations in Γ.
Approximation principle
41
In rank > 1 the assertion follows from the Curtis–Steinberg–Tits presentation [ibid.,
Theorem 1.4] (at least for p ≥ 5, when the Schur multiplier of S sc (Fp ) is trivial, cf.
also [Ste62, Ste81, Tit81]). Fix a maximal split torus T of S sc and a Borel subgroup B
containing T , and let U be the unipotent radical of B, which is a maximal unipotent
subgroup of S sc . Then every element of U(Fp ) is a product of root unipotents u with
respect to the roots of T on U. The Bruhat decomposition and the above relations imply
that Γ is generated by the elements γu , where u ranges over the elements of the unipotent
root subgroups with respect to T . Moreover, the Curtis–Steinberg–Tits relations certainly
follow from the relations between the γu ’s. This finishes the proof of the lemma.
We now show (4.1b). Let h ∈ SArng (gZp ) and let p be the largest ideal of h contained
in hresnilp. Let H = G(h) and P = exp p, and let as usual H̄ and P̄ be the images of
these groups in G(Fp ). It is clear that P is a normal pro-p subgroup of H. Note that p is
the inverse image under reduction modulo p of the largest ideal p̄ of h̄ contained in the set
h̄nilp . In particular, p ⊃ h ∩ pgZp and therefore P ⊃ exp(h ∩ pgZp ).
It remains to prove that H ∩ Kp (p) ⊂ P , since in this case
H ∩ Kp (p) = P ∩ Kp (p) = exp(log(P ∩ Kp (p))) = exp(p ∩ pgZp ) = exp(h ∩ pgZp ).
Let H̃ ⊃ P̃ be the algebraic subgroups of GFp associated by Nori’s correspondence to
the nilpotently generated Lie subalgebras h̄ ⊃ p̄ of gFp . The corresponding subgroups of
G(Fp ) are H̄ and P̄ , respectively. By Lemma 4.10, P̃ is the unipotent radical of H̃. Let
S sc be the simply connected covering group of the semisimple group H̃/P̃ and let K be
the kernel of the surjective covering map κ : S sc (Fp ) → H̃(Fp )+ /P̃ (Fp ) = H̄/P̄ . Note
that κ restricts to a bijection between S sc (Fp )unip and the image of H̄unip in H̄/P̄ . We may
assume that p does not divide the size of the center of S sc , and in particular the size of K.
We need to show that the canonical surjective homomorphism r : H/P → H̄/P̄ is an
isomorphism of groups. Using Lemma 4.17, we will show this by constructing a surjective
homomorphism φ : Γ = Γ(S sc ) → H/P with r ◦ φ = κ ◦ s. Note that in any case the
kernel of r equals (H ∩ P Kp (p))/P , and that it is therefore a p-group.
For any residually nilpotent subalgebra u ⊂ h containing p, the group U = exp u is
a pro-p subgroup of H containing P and r restricts to an isomorphism r|U/P : U/P →
Ū /P̄ , where Ū ⊂ H̄ is the image of U in G(Fp ). Given any p-group Ū ⊂ H̄, we can lift
the associated Lie algebra ū = L̄(Ū ) = log(p) Ū ⊂ h̄ = L̄(H̄) to its inverse image u under
the reduction map h → h̄ and obtain a pro-p group U ⊂ H containing P which projects
onto Ū under reduction modulo p and for which r|U/P : U/P → Ū /P̄ is an isomorphism.
As a consequence, for any unipotent subgroup Ñ of S sc defined over Fp , we may
consider the preimage Ũ of κ(Ñ ) in the group H̃ and lift the p-group Ũ (Fp ) ⊂ H̄ to a
pro-p subgroup U ⊂ H such that r maps U/P isomorphically onto Ñ(Fp ) ≃ Ũ(Fp )/P̄ ⊂
H̄/P̄ .
42
Tobias Finis, Erez Lapid
In particular, we may lift any unipotent element u of S sc (Fp ) to an element f (u) ∈
H/P with r(f (u)) = κ(u), namely to f (u) := (exp v)P for some v ∈ hresnilp such that
exp v maps under reduction modulo p to a representative of κ(u) ∈ H̄/P̄ in H̄. This
lift does not depend on the choice of v: for any v1 , v2 ∈ hresnilp it follows from Lemma
4.9 that exp v1 and exp v2 map to the same element of H̄/P̄ if and only if v1 − v2 ∈
p + h ∩ pgZp = p, which is in turn equivalent to exp v1 ∈ (exp v2 )P . Therefore, we obtain
a well-defined map f : S sc (Fp )unip → H/P with r ◦ f = κ|S sc (Fp )unip . Moreover, for any
unipotent Fp -subgroup Ñ ⊂ S sc the restriction of f to Ñ(Fp ) is the inverse of the group
isomorphism r|U/P : U/P → Ñ (Fp ) constructed above, and in particular it is a group
homomorphism.
We now claim that in the setting of Lemma 4.17 the definition φ(γu ) := f (u), u ∈
S sc (Fp )unip , extends to a group homomorphism φ : Γ → H/P , i.e., that it respects the
relations defining Γ.
Clearly,
r(φ(γu )) = r(f (u)) = κ(u) for all u ∈ S sc (Fp )unip .
(4.3)
For a unipotent subgroup Ñ ⊂ S sc defined over Fp we know that the restriction f |Ñ(Fp )
is a group homomorphism, which means that φ(γu1 u2 ) = f (u1 u2 ) = f (u1 )f (u2) =
φ(γu1 )φ(γu2 ) for any u1 , u2 ∈ Ñ (Fp ).
Let us check the conjugation relations. Suppose that Lie algebra elements vi ∈ hresnilp
reduce modulo p to log(p) u′i , where u′i ∈ H̄unip , i = 1, 2, and that u′i maps to κ(ui ) ∈ H̄/P̄
for some ui ∈ S sc (Fp )unip . Then φ(γui ) = f (ui ) = (exp vi )P . Now Ad(exp v1 )v2 ∈
hresnilp maps under reduction modulo p to Ad(u′1 ) log(p) (u′2 ) = log(p) (u′1u′2 (u′1 )−1 ), and
u′1 u′2 (u′1 )−1 maps to κ(u1 u2u−1
1 ) in H̄/P̄ . Hence,
) = exp(Ad(exp v1 )v2 )P = exp v1 exp v2 exp(−v1 )P = φ(γu1 )φ(γu2 )φ(γu1 )−1 ,
φ(γu1 u2 u−1
1
which finishes the proof of the claim.
It is clear from the definition of H that the homomorphism φ is surjective. Moreover,
it is clear from (4.3) that r ◦ φ = κ ◦ s. Since s is an isomorphism by Lemma 4.17, we
conclude that the order of Ker r divides the order of K. Since Ker r is also a p-group, we
conclude that r is an isomorphism.
This finishes the proof of (4.1b).
5 Intersections of conjugacy classes and open compact
subgroups
5.1 The global bound
As an application of the approximation theorem, we prove in this section an estimate for
the volume of the intersection of a conjugacy class in the group of Ẑ-points of a reductive
Approximation principle
43
group defined over Q with an arbitrary open subgroup. This is a key technical result in
our approach to the limit multiplicity problem for arbitrary congruence subgroups of an
arithmetic group, which is the subject matter of [FL15]. It is convenient to formulate the
result in a slightly more general way, namely in terms of the commutator map. Theorem
5.3 below gives the most general formulation, and Corollary 5.8 the main application.
Corollary 5.9 is a variant of this result in the language of lattices in semisimple Lie groups.
At the end of this section we also prove some auxiliary results that will be applied to the
limit multiplicity problem.
We will temporarily consider more general groups G than before. (However, in Theorem 5.3 and its proof G will be assumed to be semisimple and simply connected.) For the
following definitions let G be a (possibly non-connected) reductive group defined over
Q such that its adjoint group Gad as well as its derived group Gder are connected. (The
connectedness of Gder is not essential, and we assume it only for simplicity.) We fix an
embedding ρ0 : G → GL(N0 ). We denote the canonical action of Gad on G by ad and
the adjoint representation of Gad on g = LieQ G by Ad. We have the commutator map
[·, ·] :
Gad × Gad → Gder ,
which is a morphism of algebraic varieties over Q.
Q
For every prime p we set Kp = ρ−1
0 (GL(N0 , Zp )) as before, and let K =
p Kp , an
open compact subgroup of G(Afin ). Let
K(N) = {g ∈ K : ρ0 (g) ≡ 1 (mod N)} =
Y
p
Kp (pvp (N ) ),
N ≥ 1,
be the principal congruence subgroups of K. They are normal open subgroups of K and
form a neighborhood base of the identity element. For an open subgroup K ⊂ K (or, more
Q
generally, an open compact subgroup K ⊂ G(Afin )) let its level lev(K) = p levp (K) be
the smallest integer N for which K(N) ⊂ K. We note that a change of the representation
ρ0 changes the level of a given open compact subgroup K only by a factor that is bounded
from above and below.
Sometimes we will need a generalization of this notion. If H ⊂ G(Afin ) is an arbitrary
Q
closed subgroup, then we let lev(K; H) = p levp (K; H) be the smallest integer N for
which K(N) ∩ H ⊂ K. In particular, this applies to the group H = G(Afin )+ , the image
of Gsc (Afin ) under the natural homomorphism psc : Gsc → G, where Gsc is the simply
connected covering group of Gder .
ad
ad −1
ad
We also fix an embedding ρad
→ GL(N0ad ) and let Kad
0 : G
p = (ρ0 ) (GL(N0 , Zp )) ⊂
Q
ad
ad n
ad
Gad (Qp ) and Kad = p Kad
p ⊂ G (Afin ). Let Kp (p ) ⊂ Kp be the principal congruad
ence subgroups with respect to ρad
0 . We normalize the Haar measures on the groups Kp
and on their product Kad to have volume one.
Tobias Finis, Erez Lapid
44
Definition 5.1. Let K̃ be a compact subgroup of Gad (Afin ) with its normalized Haar
measure. For an open subgroup K ⊂ G(Afin ) define
φK,K̃(x) = vol {k ∈ K̃ : [k, x] ∈ K} ,
x ∈ Gad (Afin ).
For K̃ = Kad we simply write φK (x) = φK,Kad (x). Analogously, we write
φKp ,K̃p (xp ) = vol {k ∈ K̃p : [k, xp ] ∈ Kp } ,
xp ∈ Gad (Qp ),
for open subgroups Kp ⊂ G(Qp ) and compact subgroups K̃p ⊂ Gad (Qp ), and set
(x).
φKp (xp ) = φKp ,Kad
p
We will estimate the function φK (x) (indeed φK,K̃(x) for certain subgroups K̃ ⊂ Kad )
for x ∈ Kad and arbitrary open subgroups K ⊂ K in terms of the level of K. To state
the dependence of the bound on x in a convenient manner, we introduce the following
notation. Fix a Z-lattice Λ in g such that Λ ⊗ Ẑ is Kad -stable.
Definition 5.2. For xp ∈ Gad (Qp ) set
n
λG
p (xp ) = max{n ∈ Z∪{∞} : (Ad(xp )−1) Prh (Λ⊗Zp ) ⊂ p (Λ⊗Zp ) for some h 6= 0 },
where h ranges over the non-trivial semisimple ideals of the Lie algebra g ⊗ Qp , and Prh
denotes the corresponding projection g ⊗ Qp → h ⊂ g ⊗ Qp .
See Remark 5.23 below for an alternative, somewhat more concrete expression for
ad
this function. Note that λG
p (xp ) ≥ 0 for xp ∈ Kp . A change of the lattice Λ will change
λG
p (xp ) at most by a constant cp , where cp = 0 for almost all p. Similarly, for any compact
set Ω ⊂ Gad (Afin ) there exist constants Ap , with Ap = 0 for almost all p, such that
−1
G
ad
|λG
p (y xy) − λp (x)| ≤ Ap for all primes p, x ∈ G (Afin ), y ∈ Ω.
Q
Theorem 5.3. Let G be semisimple and simply connected. Let K̃ = p K̃p , where for
ad
each p the group K̃p is an open subgroup of Kad
p and the indices [Kp : K̃p ] are bounded.
Then there exist constants ε, δ > 0, depending only on G, ρ0 , ρad
0 and K̃, such that for all
Q np
open subgroups K of K of level N = lev(K) = p p and all x ∈ Kad we have
φK,K̃(x) ≪ β G (N, x, δ)−ε
with
β G (N, x, δ) =
Y
pn p .
p|N, λG
p (x)<δnp
Here, the implied constant depends on G, ρ0 , ρad
0 , K̃ and Λ.
Remark 5.4. Let
ad
Kad
: λp (x) < ∞ for all p and λp (x) = 0 for almost all p},
reg := {x ∈ K
(5.1)
Approximation principle
45
which is a dense subset of Kad . For x ∈ Kad
reg we obtain from Theorem 5.3 the estimate
Y
ε/δ
φK,K̃ (x) ≪K̃
pλp (x)
lev(K)−ε .
p
ad
ad
By Lemma 5.25 below, x ∈ Kad
that is not
reg whenever x is an element of G (Q) ∩ K
ad
contained in any proper normal subgroup of G (which we may of course assume to be
defined over Q).
Remark 5.5. In fact, the proof of Theorem 5.3 will yield the sharper estimate
Y
C Y
′
G
min(1, pε (c+λp (x)−εnp ) ),
φK,K̃(x) ≤
min(1, )
p
G
p|N, λp (x)=0
(5.2)
p|N
where ε > 0 is the constant provided by the approximation theorem (Theorem 2.2) for G
and ρ0 , and ε′ , C > 0 and c are additional constants depending only on G, ρ0 , ρad
0 and K̃.
Remark 5.6. For square-free level N and x ∈ Kad
reg we obtain from (5.2) the estimate
φK,K̃(x) ≪x,K̃,ε N −ε for any ε < 1. Up to the determination of the constant C, the
estimate (5.2) is best possible even for groups of arbitrarily large rank. (It might not be
best possible for particular groups G.) Indeed, if G = SL(n) over Q, Gad = PGL(n),
n−1
¯
Kad
(Fp )
p = PGL(n, Zp ) and Kp ⊂ Kp = SL(n, Zp ) is the stabilizer of a point ξ ∈ P
under the natural action, then
r
1 X ni
(p − 1) ,
φKp (xp ) = n
p − 1 i=1
if xp ∈ GL(n, Zp ) stabilizes ξ¯ and the dimensions of the Fp -eigenspaces of the image
x̄p of xp in GL(n, Fp ) are n1 , . . . , nr . In particular, for n1 = n − 1 and n2 = 1 we get
φKp (xp ) ≥ p1 . Note that for Λ = sl(n, Z), we have λp (xp ) = 0 if x̄p is not a scalar matrix.
Remark 5.7. For an arbitrary N, it is not possible to take ε arbitrarily close to 1 in
(5.1). For example, consider for G = SL(2) the congruence subgroups Kp = Γ0 (pn ) ⊂
Kp = SL(2, Zp ), n ≥ 1. Let xp = ( a0 db ) ∈ SL(2, Zp ) and r = λp (xp ) = min(vp (d −
a), vp (b)). Then φKp (xp ) is just the number of fixed points of xp on P1 (Z/pn Z) divided by
′
|P1 (Z/pn Z)| = pn (1 + p1 ). Assume that r < n. Write xp = a + pr y with y = 00 db′ and b′ ,
d′ ∈ Zp not both divisible by p. Then φKp (xp ) is also the number of eigenvectors of y in
P1 (Z/pn−r Z) divided by pn−r (1 + p1 ). After possibly conjugating by an upper triangular
matrix in SL(2, Zp ), we may assume that b′ = 1. Then the number of eigenvectors in question is just the number of solutions to the quadratic congruence ξ 2 −d′ ξ ≡ 0 (mod pn−r ),
where vp (d′ ) = vp (d − a) − r. Direct computation yields the result
−1
2 1 + 1
,
p−(n−vp (d−a)) , vp (d − a) < n+r
p
2
−1
φKp (xp ) =
(5.3)
1+ 1
n+r
⌉
−⌈ n−r
2
,
v
(d
−
a)
≥
.
p
p
p
2
This implies that ε ≤
1
2
for G = SL(2).
Tobias Finis, Erez Lapid
46
The following corollary concerning conjugacy classes in G(Ẑ) for an arbitrary reductive group G has applications to the limit multiplicity problem. Essentially the same
result has been obtained independently in [ABB+ , §5] without using the approximation
theorem. We note that in the case of G = SL(2) or the group of norm one elements of a
quaternion algebra over Q, and for particular choices of x, very explicit estimates of this
type have been already obtained in [CP84].
Corollary 5.8. Let G be a (possibly non-connected) reductive group defined over Q whose
derived group Gder and adjoint group Gad are connected. Let ρ0 : G → GL(N0 ) be a
faithful Q-rational representation and K = ρ−1
0 (GL(N0 , Ẑ)) ⊂ G(Afin ). Then there exist
constants ε, δ > 0 such that for all open compact subgroups K0 ⊂ G(Afin ) we have
vol {k ∈ K : kxk −1 ∈ K} ≪ρ0 , K0 β G (lev(K; G(Afin )+ ), x, δ)−ε
for all open subgroups K ⊂ K0 and all x ∈ G(Afin ). (Here we pull back λp to G via the
projection G → Gad .)
The simple proof will be given in §5.4 below.
For the convenience of the reader, and to facilitate comparison with [ABB+ ], we also
give a variant of this result concerning lattices in semisimple Lie groups. For an arbitrary
group Γ, a finite index subgroup ∆, and an element γ ∈ Γ set
c∆ (γ) = |{δ ∈ Γ/∆ : δ −1 γδ ∈ ∆}|,
which is also the number of fixed points of γ in the permutation representation of Γ on
the finite set Γ/∆.
Corollary 5.9. Let G be a semisimple and simply connected group defined over Q such
that for no Q-simple factor H of G the group H(R) is compact, and let K ⊂ G(Afin )
be as above. Let Γ = G(Q) ∩ K, which is a lattice in the Lie group G(R). For any
open subgroup K ⊂ K let ∆ = G(Q) ∩ K be the associated finite index subgroup of Γ.
Then there exist constants ε, δ > 0, depending only on G and ρ0 , such that for all open
subgroups K and all γ ∈ Γ we have
c∆ (γ)
≪ β G (lev(K), γ, δ)−ε .
[Γ : ∆]
In particular, if γ is not contained in any proper normal subgroup of G (which we may
assume to be defined over Q) then we have
Y
C
c∆ (γ)
≪
pλp (γ) [Γ : ∆]−ε ,
[Γ : ∆]
p
where C > 0 depends only on G.
Approximation principle
47
Proof. The assumptions on G imply (and in fact, are equivalent to) that G has the strong
approximation property, i.e., that G(Q) is dense in G(Afin ) [PR94, Theorem 7.12]. From
this we get that [Γ : ∆]−1 c∆ (γ) = vol ({k ∈ K : kγk −1 ∈ K}), and we may apply
Corollary 5.8 to deduce the first inequality. For the second, we invoke Remark 5.4 to
Q
get the estimate ≪ ( p pλp (γ) )C lev(K)−ε . However, we have [Γ : ∆] = [K : K] ≤
vol(K(lev(K)))−1 ≪ lev(K)dim G , and we may therefore replace lev(K) by [Γ : ∆] (at
the price of replacing ε by ε/ dim G).
The groups ∆ = G(Q) ∩ K ⊂ Γ are called the congruence subgroups of Γ. Under
the assumptions of the corollary, they are in one-to-one correspondence with the open
subgroups K of K.
5.2 Reduction to two statements on open subgroups of Kp
We will now derive Theorem 5.3 (and the refined estimate (5.2)) from the following two
statements concerning open subgroups of the groups Kp for arbitrary p. As in the theorem,
G will be assumed to be semisimple and simply connected. (The assumption that G is
simply connected is relevant only for Proposition 5.10 below.) The first statement bounds
φKp (xp ) for all proper subgroups Kp of Kp , while the second one concerns subgroups of
level pn for large n. The propositions will be proved in §5.3 below.
Proposition 5.10. For any prime p, any proper subgroup Kp of Kp , and any xp ∈ Kad
p
−1
p
.
with λp (xp ) = 0, we have φKp (xp ) ≪ρ0 ,ρad
0
Proposition 5.11. There exist constants ε, ε′ > 0 and c, depending only on G, ρ0 and ρad
0 ,
such that for any prime p, any subgroup Kp of Kp of level pn and any xp ∈ Kad
p , we have
ε′ (c+λp (xp )−εn)
.
φKp (xp ) ≤ p
As in Remark 5.5, the constant ε in Proposition 5.11 is the constant provided by Theorem 2.2 applied to G and ρ0 .
In fact Proposition 5.10 is essentially already known. More precisely, at least in the
case where xp lies in the image of Kp in Gad (Qp ), it is shown in [LS91] that for any
G one can take the implied constant (say C) to be 2 for almost all p.7 This estimate is
optimal for G = SL(2), as one sees from (5.3) for n = 1 and r = vp (d − a) = 0. If
one excludes the case where G has a factor of type A1 , then one may lower the value
of C further. (By Remark 5.6 above, we need to have C ≥ 1 even if we omit finitely
many possibilities for G.) However, the proof in [loc. cit.] is based on a detailed analysis of a great number of particular cases, and uses explicit information on the maximal subgroups of the finite simple groups of Lie type, while our proof, which does not
7
Lemma 5.13 below shows that the framework of [loc. cit.] is applicable for almost all p.
Tobias Finis, Erez Lapid
48
give the optimal value of C, uses only Nori’s theorem. For more refined recent bounds
see [LS99, Bur07b, Bur07a, Bur07c, Bur07d] (concerning classical groups) and [LLS02]
(concerning exceptional groups), as well as the references cited therein.
We will now show how to deduce Theorem 5.3 from Propositions 5.10 and 5.11. Arguments of this type can be found already in [CP84] and [LS03, §6.1]. We first state a
simple property of the functions φK,K̃ which follows by straightforward calculation.
Lemma 5.12. Let L ⊂ K be open subgroups of K. Then
X
φK,K̃(x) =
φad(kη )−1 (L),K̃ (x),
η∈L\K: [K̃,x]∩Lη6=∅
x ∈ Kad ,
where kη ∈ K̃ is an arbitrary element with [kη , x] ∈ Lη.
In the following lemma we collect some standard facts on the behavior of the groups
G and Gad modulo p for almost all primes p.
Lemma 5.13. Given G, ρ0 , Gad , ρad
0 , for almost all primes p we have:
1. The group schemes G and Gad are smooth over Zp and thus Kp /Kp (p) ≃ G(Fp )
ad
ad
ad
and Kad
p /Kp (p) ≃ G (Fp ). Moreover, the group schemes GFp and GFp are semisimple algebraic groups over Fp .
2. The maps ad : Gad × G → G and [·, ·] : Gad × Gad → G map Kad
p × Kp and
ad
ad
Kp × Kp to Kp , and moreover descend modulo p to corresponding maps of the
groups of Fp -points.
3. Let hi , i = 1, . . . , rp , be the minimal non-zero ideals of g ⊗ Qp . Then the Z-lattice
L rp
Λ ⊂ g used to define λp (x) satisfies Λ⊗Zp = i=1
(Λ⊗Zp )∩hi . The corresponding
ad
ad
as
factorizations G = G1 × · · · × Grp and Gad = Gad
1 × · · · × Grp of G and G
products of Qp -simple algebraic groups extend to factorizations of group schemes
that are smooth over Zp . In particular, we have corresponding factorizations of G
and Gad over Fp (cf. [LS03, pp. 392–393]).
4. All proper normal subgroups of the groups Gi (Fp ), i = 1, . . . , rp , are central
[PR94, Proposition 7.5], and the center of G(Fp ) is the set of Fp -points of the
center of G, i.e., the kernel of the adjoint representation of G(Fp ).
5. For any m ≥ 1 the m-th iterated Frattini subgroup of Kp is equal to Kp (pm ). In
particular, the Frattini subgroup Φ(Kp ) is Kp (p) [LS03, Window 9, Lemma 5 and
Corollary 6].
We also need a standard estimate for the number of Fp -points of a linear algebraic
group [Nor87, Lemma 3.5].
Approximation principle
49
Lemma 5.14. For almost all p, depending on Gad and ρad
0 , we have
(p − 1)dim G ≤ |Gad (Fp )| = vol Kad
p (p)
−1
≤ (p + 1)dim G .
For convenience, we also isolate a key technical consequence of Proposition 5.11 as a
separate lemma.
Lemma 5.15. Let ε and ε′ be as in Proposition 5.11. Let K̃p ⊂ Kad
p be open subgroups
and B a positive integer with [Kad
p : K̃p ] ≤ B for all p. There exists a constant c,
depending only on G, ρ0 , and B, such that for any prime p, for all subgroups Kp ⊂ Kp (p)
of level pn and for all xp ∈ Kad
p , we have the estimate
′
φad(kp )Kp ,K̃p (xp ) ≤ min(1, pε (c+λp (xp )−εn) )φad(kp )K(p),K̃p (xp ).
Proof. By the second and fifth parts of Lemma 5.13, for all but finitely many primes p the
m
operators ad(kp ), kp ∈ Kad
p , act on Kp , and each of the groups Kp (p ) is a characteristic
subgroup of Kp . We therefore have ad(kp )Kp (pm ) = Kp (pm ) as well as lev(ad(kp )Kp ) =
lev(Kp ) = pn for almost all p. Treating the remaining finitely many primes p one at a time,
an easy compactness argument shows that lev(ad(kp )Kp ) ≥ pn−c1 for all p, with a constant c1 depending only on G and ρ0 . Again by compactness, for each single p the values
φad(kp )(Kp (p)),K̃p (xp ) for xp , kp ∈ Kad
p are bounded away from zero. Moreover, by Lemma
5.14, for almost all p we have φad(kp )Kp (p),K̃p (xp ) = φKp (p),K̃p (xp ) ≥ vol(Kad
p (p))) ≥
−c3
−c2
for all p with
p , with a suitable constant c2 , and therefore φad(kp )Kp (p),K̃p (xp ) ≥ p
a suitable c3 . Since φad(kp )Kp ,K̃p (xp ) ≤ Bφad(kp )Kp (xp ), the lemma follows now from
Proposition 5.11.
Proof of Theorem 5.3. Let K̃ ⊂ Kad be as in Theorem 5.3. Let K ⊂ K be an arbitrary
Q
Q
open subgroup and write N = lev(K) = p pnp . Set N1 = p|N p and consider the
groups K̃ = KK(N1 ) and L = K ∩ K(N1 ). Clearly, L is a normal subgroup of K of
level N and
L\K ≃ K(N1 )\K̃.
(5.4)
Q
Q
Note that we can factor L = p|N Lp p6|N Kp , where Lp ⊂ Kp (p) is a pro-p group.
We can now apply Lemma 5.12 to L ⊂ K to obtain
X
φad(kη )−1 (L),K̃ (x).
(5.5)
φK,K̃ (x) =
η∈L\K: [K̃,x]∩Lη6=∅
By (5.4), we may choose the same representatives η ∈ K and kη ∈ K̃ also for the pair of
groups K(N1 ) ⊂ K̃. We obtain the corresponding equation
X
φK̃,K̃ (x) =
φad(kη )−1 (K(N1 )),K̃ (x).
(5.6)
η
Tobias Finis, Erez Lapid
50
Consider φad(kη )−1 L,K̃ (x) =
Q
p|N
φad(kη,p )−1 Lp ,K̃p (xp ). Applying Lemma 5.15 yields
φad(kη )−1 L,K̃ (x) ≤ φad(kη )−1 K(N1 ),K̃ (x)
Y
′
min(1, pε (c+λp (xp )−εnp ) ).
p|N
Inserting this into (5.5) and using (5.6), we get
φK,K̃(x) ≤ φK̃,K̃ (x)
Y
′
min(1, pε (c+λp (xp )−εnp ) ).
(5.7)
p|N
It remains to estimate φK̃,K̃ (x). There exists a constant C ≥ 1, depending only on G and
ρ0 , such that for all primes p|N with p ≥ C the image of K in the factor group Kp (p)\Kp
is a proper subgroup of this group [LS03, p. 116]. Therefore,
K̃ ⊂
Y
K̃p ,
p
where for the primes p|N with p ≥ C the group K̃p is a proper subgroup of Kp and K̃p =
Kp for all other primes p. For a suitable value of C we can therefore apply Proposition
5.10 to estimate
Y
C
φK̃,K̃ (x) ≤
min(1, ).
p
p|N :λp (xp )=0
Combining this with (5.7) yields immediately (5.2).
It is now a routine matter to derive (5.1) from (5.2). For this assume without loss of
generality that c ≥ 0 and let 0 < δ < (c + 1)−1 ε. Observe first that we can estimate
Y
min(1,
p|N :λp (xp )=0
where we set C2 =
C
)≤
p
1
Y
p|N :p≥C 2, λp (xp )=0
p− 2 ≤ C 2
Y
1
p− 2 ,
p|N :λp (xp )=0
1
Q
p<C 2
φK̃,K̃ (x) ≤ C2
p 2 . Therefore, we obtain
Y
p|N :λp (xp )=0
1
Y
p− 2
′
min(1, pε (c+λp (xp )−εnp ) ).
p|N :λp (x)<δnp
Consider now any prime p|N for which λp (x) < δnp . In the case np ≤ δ −1 , the inequality
λp (x) < δnp implies that λp (x) = 0 and the first product contains therefore the factor
δ
1
p− 2 ≤ p− 2 np . In case np > δ −1 , the second product contains the factor
′
′
pε (c+λp (xp )−εnp ) ≤ pε ((c+1)δ−ε)np = p−ε
with ε′′ = ε′ (ε − (c + 1)δ) > 0. This clearly implies (5.1).
′′ n
p
Approximation principle
51
5.3 Proof of the local bounds
We now prove Propositions 5.10 and 5.11. For this we use the general estimates of Appendix A for the number of solutions of polynomial congruences. The first proposition is
easily deduced from Proposition 3.3.
Proof of Proposition 5.10. We can assume that p is sufficiently large (depending on G).
In particular, we can assume that we are in the situation of Lemma 5.13.
Since under this assumption on p we have Φ(Kp ) = Kp (p), the projection H of the
proper subgroup Kp of Kp to G(Fp ) is a proper subgroup. Replacing Kp by the preimage
of H and taking into account the first two parts of Lemma 5.13, we are reduced to proving
that
|{k ∈ Gad (Fp ) : [k, x̄] ∈ H}|
C
≤
(5.8)
ad
|G (Fp )|
p
ad
for a suitable constant C, where x̄ denotes the image of xp ∈ Kad
p in G (Fp ).
Recall that by the algebraization result of Proposition 3.3, there exist an integer N,
depending only on G, and a proper, connected algebraic subgroup X of GFp , defined over
Fp , such that [H : H ∩X(Fp )] ≤ N. Moreover, the ideal defining the subvariety X of GFp
is generated by regular functions of degree ≤ N (with respect to the affine embedding of
GFp provided by ρ0 ). Given this, it follows easily from Lemma 5.12 that it is enough to
establish the estimate (5.8) for the group X(Fp ) instead of H.
The condition λp (xp ) = 0 implies that the group
¯
C(x̄)
:= [Gad (Fp ), x̄] ⊂ G(Fp )
¯
is the full group G(Fp ). Indeed, by the identity [yx, z] = (ad(y)[x, z])[y, z] the group C(x̄)
is a normal subgroup of G(Fp ) for any x̄. It is also clearly the product of its projections
C¯i (x̄) to the factors Gi (Fp ), i = 1, . . . , rp . Moreover, by our assumption on xp the projections of x̄ ∈ Gad (Fp ) to the factors Gad
i (Fp ) are all non-trivial. Therefore the normal
¯
subgroups Ci (x̄) ⊂ Gi (Fp ) are all non-central. By the fourth part of Lemma 5.13 they
¯ = G(Fp ).
therefore have to be the full factor groups Gi (Fp ), and we obtain that C(x̄)
Thus we have [Gad (Fp ), x̄] 6⊂ X(Fp ). Therefore, in any generating set of the defining
ideal of X, there exists an element f (a regular function on GFp ) such that g = f ([·, x̄])
ad
(a regular function on Gad
Fp ) does not vanish on G (Fp ). Since the degree of f can be
bounded by N, the degree of the function g is clearly also bounded in terms of G and N.
Since we have
|{k ∈ Gad (Fp ) : [k, x̄] ∈ X(Fp )}| ≤ |{k ∈ Gad (Fp ) : g(k) = 0}|,
and moreover
|Gad (Fp )| ≥ cpdim G
Tobias Finis, Erez Lapid
52
for a suitable constant c > 0 by Lemma 5.14, it remains to invoke Lemma A.1 to establish
(5.8) for X(Fp ) and to finish the proof.
We now turn to the proof of Proposition 5.11, which is based on the combination of
Theorem 2.2 with Lemma A.2.
We first clarify the group-theoretic meaning of λp (x) following Larsen and Lubotzky.
Definition 5.16. For xp ∈ Kad
p let
Cp (xp ) := [Kad
p (p), xp ] ⊂ G(Qp ).
Note that the group Cp (xp ) is invariant under ad(Kad
p (p)). (For almost all p it is by
Lemma 5.13 therefore a normal subgroup of Kp (p).)
Lemma 5.17. There exists a constant n0 , depending only on G, such that for all xp ∈ Kad
p
with λp (xp ) < ∞ we have
Cp (xp ) ⊃ Kp (pλp (xp )+n0 ).
Proof. The assertion is essentially proven in [LL04, p. 453–454], where the case of Qp simple groups G is treated. Since GQp can be factored as a product of Qp -simple groups,
we can easily reduce to this case. Although it is not explicitly stated there, the proof in
[ibid.] also shows that n0 may be chosen independently of p.
In the proof of Proposition 5.11 it turns out to be convenient to consider a variant of
the principal congruence subgroups which is provided by the following definition.
Definition 5.18. Let
ρ : G → GL(N)
be a Qp -rational representation of G with ρ(Kp ) ⊂ GL(N, Zp ), and assume that no nonN
zero vector in QN
/ pZN
p is fixed by ρ(G). For any primitive v ∈ Zp (i.e., v ∈
p ) and any
m ≥ 1 set
Kp (ρ, v; m) = {g ∈ Kp : ∃λ ∈ Z∗p : ρ(g)v ≡ λv
(mod pm )}.
Clearly, the groups Kp (ρ, v; m) are open subgroups of Kp . We can estimate their levels
as follows.
Lemma 5.19. There exist constants m0 and m1 , depending only on G and N, such that
pm+m1 ≥ lev(Kp (ρ, v; m)) ≥ pm−m0
(5.9)
for all ρ as in Definition 5.18, all primitive v ∈ ZN
p and all m ≥ 1. Moreover, for all p
outside a finite set of primes depending only on G and N, we have lev(Kp (ρ, v; m)) = pm .
Approximation principle
53
Proof. We first observe that there exists ep ≥ 1 such that
ρ(Kp (pep )) ⊂ Γ(N, pǫp ),
where Γ(N, pm ), m ≥ 1, are the principal congruence subgroups of GL(N, Zp ). Indeed,
the reduction modulo p of the representation ρ induces a homomorphism Kp (pǫp ) →
GL(N, Fp ), whose image is contained in a p-Sylow subgroup of GL(N, Fp ). Recall that
Kp (pn ) = Kp (pǫp ){p
n−ǫp
}
(5.10)
for all n ≥ ǫp . Hence we can take ep = ǫp + ⌈logp N⌉ if p is odd, while for p = 2 we
consider the reduction modulo 4 and set e2 = 2 + ν, where 2ν is the exponent of the
2-Sylow subgroup of the finite group GL(N, Z/4Z).
Using (5.10) again, for any m ≥ ǫp we have
ρ(Kp (pm+ep −ǫp )) = ρ(Kp (pep ){p
m−ǫp
}
) ⊂ Γ(N, pǫp ){p
m−ǫp
}
= Γ(N, pm ).
Consequently, Kp (pm+ep −ǫp ) ⊂ Kp (ρ, v; m) for all v. This shows that lev(Kp (ρ, v; m)) ≤
pm+m1,p for m1,p = ep − ǫp .
Moreover, ρ(Kp (pep )) is a uniform subgroup of the uniform pro-p group Γ(N, pǫp ),
and by Proposition 2.7 we have therefore ρ(Kp (pep )) = exp(Lρ,ep ) for a powerful Lie
subalgebra Lρ,ep ⊂ pǫp gl(N, Zp ). We claim that there exists fp ≥ ǫp with the property:
/ Zp v + pfp +1 ZN
For all primitive v ∈ ZN
p .
p there exists l ∈ Lρ,ep with lv ∈
(5.11)
For assume the contrary. Because the set of primitive elements of ZN
p is compact, there
N
would then exist a primitive v ∈ Zp with Lρ,ep v ⊂ Zp v, and therefore ρ(Kp (pep )) would
ep
stabilize the line Qp v ⊂ QN
p . Since Kp (p ) is Zariski dense in G [PR94, Lemma 3.2],
this would imply that the representation ρ stabilizes a line in QN
p , and therefore contains
the trivial representation (since G is semisimple), contrary to our assumption. This proves
the claim.
Assume now that n ≥ ep and m ≥ ǫp are such that
Kp (pn ) ⊂ Kp (ρ, v; m).
Since ρ(Kp (pn )) = exp(pn−ep Lρ,ep ), we can rewrite this relation as
exp(pn−ep Lρ,ep )v ⊂ Zp v + pm ZN
p .
Let l ∈ Lρ,ep be arbitrary and let ξ = exp(pn−ep l) ∈ Γ(N, pǫp ). Since ξv ∈ Zp v + pm ZN
p
′
m
′
ǫp
′
we can write ξ ∈ ξ Γ(N, p ) for some ξ ∈ Γ(N, p ) with ξ v ∈ Zp v. Therefore, by part
2 of Proposition 2.7 we have pn−ep l = log ξ ∈ log ξ ′ + pm gl(N, Zp ) and (log ξ ′)v ∈ Zp v,
m−n+ep N
which together implies that pn−ep lv ∈ Zp v+pm ZN
Zp ,
p , and therefore lv ∈ Zp v+p
54
Tobias Finis, Erez Lapid
/ Zp v+
since v is primitive. On the other hand, by (5.11) there exists l ∈ Lρ,ep such that lv ∈
fp +1 N
p
Zp . We infer that n ≥ m + ep − fp . We conclude that lev(Kp (ρ, v; m)) ≥ pm−m0,p
for all m ≥ ǫp with m0,p = fp − ep .
It remains to see that for almost all p we can choose ep = fp = 1. Consider the set
Kp,resunip of residually unipotent elements of Kp . Note that the proof of [LS03, Window
9, Lemma 5] shows that for almost all p, the group Kp (p) is topologically generated by the
elements up , u ∈ Kp,resunip. Clearly, if u ∈ Kp,resunip then its image ρ(u) ∈ GL(N, Zp )
is also residually unipotent, and for p ≥ N this implies that in fact ρ(u)p ≡ 1 (mod p).
Therefore, after excluding a finite set of primes that depends only on G and N, we have
ρ(Kp (p)) ⊂ Γ(N, p), so that we can take ep = 1.
Moreover, for all such p the reduction modulo p of the representation ρ gives a representation ρ̄ : G(Fp ) → GL(N, Fp ). On the other hand, the representation ρ decomposes
over Q̄p into irreducibles, which are parametrized by their highest weights. For almost all
p, the root coordinates of these weights are small with respect to p, which implies that
the reduction modulo p of each irreducible constituent of ρ remains irreducible [Jan87,
Corollary II.5.6]. By the Brauer–Nesbitt theorem, the semisimplification of the representation ρ̄ (considered over Fp ) is given by the direct sum of the reductions modulo p of
the irreducible constituents of ρ. Since by assumption the trivial representation is not a
constituent of ρ, we conclude that if we exclude a finite set of primes p that depends
only on G and N, then the characteristic p representation ρ̄ does not contain the trivial
representation.
N
Therefore, for any 0 6= v̄ ∈ FN
p the line Fp v̄ ⊂ Fp is not stabilized by the operators ρ(u), u ∈ Kp,resunip, and hence by the operators log(p) ρ(u), u ∈ Kp,resunip [Nor87,
Lemma 1.4]. By the commutativity of (3.2), we conclude that the line Fp v̄ is not stabilized by the reduction mod p of the logarithms log ρ(u), u ∈ Kp,resunip. However, the Lie
algebra Lρ,1 contains the elements p log ρ(u), u ∈ Kp,resunip. This means that we may take
fp = 1 above and conclude that lev(Kp (ρ, v; m)) = pm . This proves the lemma.
Next, we reformulate Lemma A.2 in the case at hand. Recall that ρad
0 fixes an affine
ad
embedding of the affine variety G , which allows us to speak of the degree of regular
function on Gad (and similarly for G).
Corollary 5.20. For any integer d > 0 there exists a constant ε(d) > 0, depending only
ad
on Gad and ρad
0 , such that for any m, n ∈ Z, and a regular function f on G , defined
m
over Qp , of degree ≤ d such that f (Kad
p (p)) 6⊂ p Zp , we have
n
−ε(d)(n−m+1)
vol {g ∈ Kad
.
p (p) : f (g) ≡ 0 (mod p )} ≪d,Gad p
Proof. We take V = Gad in Lemma A.2 together with the affine embedding provided by
ad
ad
ρad
0 . Let µp be the normalized Haar measure on G (Qp ) such that vol(Kp ) = 1 and µ̃p
Approximation principle
55
the measure on Gad (Qp ) induced by the fixed affine embedding of Gad (cf. [Ser81, §3.3],
[Oes82, §3]). Since Gad is a smooth variety over Qp , for any p we clearly have
cp µ̃p |Kp (p) ≤ µp |Kad
≤ c′p µ̃p |Kad
p (p)
p (p)
for suitable c′p ≥ cp > 0. Moreover, for almost all p the measure µ̃p |Kad
is a conp (p)
ad
− dim G
stant multiple of the measure µp , and since it gives Kp (p) total volume p
, while
ad
−1
ad
µp (Kp (p)) = |G (Fp )| for almost all p, we have
µp |Kad
=
p (p)
pdim G
µ̃p | ad
|Gad (Fp )| Kp (p)
for almost all p. By Lemma 5.14, the normalizing factor satisfies
− dim G
− dim G
1
1
pdim G
≤ 1−
1+
≤ ad
,
p
|G (Fp )|
p
and is therefore bounded in both directions in terms of dim G only. The lemma follows
therefore from Lemma A.2.
Remark 5.21. Assume that G is split over Q. Then we can realize Kad
p (p) as an explicit
compact subset of an affine space of dimension dim G and by Lemma A.9 it is therefore
possible to take any ε(d) < d1 in Corollary 5.20. In general, the affine variety Gad does
not need to be rational over Q (or even over Qp ).
We can now prove a variant of Proposition 5.11 for the groups Kp = Kp (ρ, v; m).
Lemma 5.22. There exist constants ε′ > 0 and c′ ≥ 0, depending only on G and N, such
that for any prime p, any Qp -rational representation ρ : G → GL(N) with ρ(Kp ) ⊂
ad
GL(N, Zp ) without fixed vectors, any primitive v ∈ ZN
p , m ≥ 1, and any xp ∈ Kp we
have
′ ′
vol {k ∈ Kad (p) : [k, xp ] ∈ Kp (ρ, v; m)} ≤ pε (c +λp (xp )−m) .
PN
Proof. Let h·, ·i be the standard bilinear form on QN
p . Write v =
i=1 vi ei and consider
the regular functions
fij,v (g) = hρ([g, xp ])v, vj ei − vi ej i ,
i 6= j,
ad
on Gad . It follows directly from Definition 5.18 that for any primitive v ∈ ZN
p , k ∈ Kp
and m ≥ 1 we have
fij,v (k) ∈ pm Zp ∀i 6= j
if and only if
[k, xp ] ∈ Kp (ρ, v; m),
and therefore (see Definition 5.16)
m
fij,v (Kad
p (p)) ⊂ p Zp ∀i 6= j
if and only if
Cp (xp ) ⊂ Kp (ρ, v; m).
Tobias Finis, Erez Lapid
56
Combining Lemma 5.17 with Lemma 5.19, the inclusion Cp (xp ) ⊂ Kp (ρ, v; m) implies
the inequality m ≤ λp (xp ) + n0 + m0 . Therefore, for any v there exist indices i 6= j with
λp (xp )+n0 +m0 +1
fij,v (Kad
Zp .
p (p)) 6⊂ p
Over the algebraic closure of Q, there are only finitely many isomorphism classes
of representations of G of dimension N. Since ρ is necessarily Q̄p -isomorphic to such
a representation, we can bound the degrees of the functions fij,v in terms of G and N.
Therefore, we can apply Corollary 5.20 to estimate
vol {k ∈ Kad
(p)
:
[k,
x
]
∈
K
(ρ,
v;
m)}
≤
p
p
p
m
(p)
:
f
(k)
∈
p
Z
}
≪G p−ε(m−λp (xp )−n0 −m0 ) ,
min vol {k ∈ Kad
ij,v
p
p
i6=j
for suitable ε as required.
We are now ready to finish the proof of Proposition 5.11 (and therefore of Theorem
5.3).
Proof of Proposition 5.11. Let Kp be an open subgroup of Kp of level pn . Without loss
of generality we may assume that n ≥ 2. Let xp ∈ Kad
p .
First note that by the continuity of the map [·, ·], there exist integers lp ≥ 1 with the
lp
ad
ǫp
property that [Kad
p (p ), Kp ] ⊂ Kp (p ). Moreover, since by Lemma 5.13, for almost all
p the map [·, ·] descends to a map Gad (Fp ) × Gad (Fp ) → G(Fp ), we may take lp = 1 for
almost all p. We may then write
φKp (xp ) =
X
ξ
lp
−1
ǫp
vol {k ∈ Kad
p (p ) : [k, ξxp ξ ] ∈ Kp ∩ Kp (p )} ,
(5.12)
lp
ad
where the summation is over those classes in Kad
p (p )\Kp for which there exists a representative ξ with [ξ, xp ] ∈ Kp . By Lemma 5.14, the total number of summands in (5.12)
′′
is clearly bounded by pc for a constant c′′ ≥ 0 depending only on G.
From Theorem 2.2, applied to the group Kp ∩ Kp (pǫp ) of level pn , we obtain the
existence of a proper, connected algebraic subgroup X of G, defined over Qp , such that
Kp ∩ Kp (pǫp ) ⊂ (X(Qp ) ∩ Kp )Kp (p⌈εn⌉ ), where ε depends only on G. Of course, here
we may assume X to be maximal among proper, connected subgroups defined over Qp ,
i.e., X ∈ MSGRQp (G) in the notation of §2.4. We can find a Qp -rational representation
ρ : G → GL(N), not containing the trivial representation, such that X is the stabilizer of
a line Qp v ⊂ QN
p . Moreover, by Corollary 2.29 (taken together with Lemma 2.26) we can
bound here N in terms of G only. By conjugating ρ and adjusting v, we can also assume
N
that ρ(Kp ) ⊂ GL(N, Zp ) and v ∈ ZN
p \ pZp . In this situation we have from Lemma 5.19
that lev(Kp (ρ, v; m)) ≤ pm+m1 with m1 depending only on G and ρ0 .
Approximation principle
57
Clearly, we have now X(Qp ) ∩ Kp ⊂ Kp (ρ, v; m) for any m ≥ 1, and in addition
Kp (p⌈εn⌉ ) ⊂ Kp (ρ, v; m) for
m := ⌈εn⌉ − m1 ,
if n >
m1
,
ε
as we may assume. Therefore,
Kp ∩ Kp (pǫp ) ⊂ (X(Qp ) ∩ Kp )Kp (p⌈εn⌉ ) ⊂ Kp (ρ, v; m).
We can now apply Lemma 5.22 to the individual summands in (5.12) to obtain
φKp (xp ) ≤ pc
′′ +ε′ (c′ +λ (x )−m)
p p
.
The proposition follows.
5.4 Some supplementary results
We turn to the proof of Corollary 5.8. Until the end of this section we assume that G and K
are as in Corollary 5.8, i.e., that G is a (possibly non-connected) reductive group defined
over Q whose derived group Gder and adjoint group Gad are connected, and that K =
ρ−1
0 (GL(N0 , Ẑ)) ⊂ G(Afin ) for a faithful Q-rational representation ρ0 : G → GL(N0 ).
Proof of Corollary 5.8. Writing
vol {k ∈ K : kxk −1 ∈ K} ≤
X
y∈K∩K0 \K
vol {k ∈ K0 : kyxy −1k −1 ∈ K} ,
we may first reduce to the case where K0 = K and x ∈ K.
Let Gsc be the simply connected covering group of the derived group Gder of G and let
psc : Gsc → G be the associated canonical homomorphism. Let ρsc : Gsc → GL(N sc ) be
a faithful Q-rational representation such that (psc )−1 (K) ⊂ Ksc = (ρsc )−1 (GL(N sc , Ẑ)).
We can assume that there exists k0 ∈ K with k0 xk0−1 ∈ K, for otherwise the bound is
trivial. Then
vol {k ∈ K : kxk −1 ∈ K} = vol {k ∈ K : [k, k0 xk0−1 ] ∈ K} .
To estimate the right-hand side, we can apply Theorem 5.3 to the group Gsc , the open
subgroup (psc )−1 (K) of Ksc and the image of k0 xk0−1 in Gad (Afin ). Here we let K̃p be the
ad
image of Kp in Gad (Qp ) and take any representation ρad
0 such that K̃p ⊂ Kp for all p.
In the final estimate we may replace the level of (psc )−1 (K) ⊂ Ksc with respect to ρsc by
lev(K; G(Afin )+ ), since the quotient of these two quantities is bounded from above and
below in terms of G, ρ0 and ρsc
0 .
We now give an alternative description of the functions λp (x), as well as three lemmas
on their behavior, which are useful in the application to limit multiplicities.
58
Tobias Finis, Erez Lapid
Q
ad
ad
Remark 5.23. We can write Gad = ri=1 Gad
i with Gi = ResFi /Q Gi for finite extensions F1 , . . . , Fr of Q and absolutely simple adjoint Fi -groups Giad [BT65, §6.21 (ii)]. The
Lr
ad
gi
individual factors Gad
i are then the Q-simple factors of G . The Lie algebra g =
Qr i=1
acquires naturally the structure of a module over the semisimple algebra A = i=1 Fi .
Q
Let oA be the ring of integers of A. Furthermore, we can take Kad = v Kad
v , where
Q
ad
v ranges over the prime ideals of oA , which implies the factorizations Kp = v|p Kad
v
for all primes p. Assume finally that Λ is an oA -lattice in g. For v|p let ̟v be a prime
element of the local field Av and ev the corresponding ramification index. For n ≥ 0 set
n
ad
n
Kad
v (̟v ) = {xv ∈ Kv : Ad(xv )l ≡ l (mod ̟v Λ) ∀l ∈ Λ}. Then for xp = (xv )v|p ∈
Kad
p we can compute λp (xp ) as the largest integer n ≥ 0 for which there exists a place v
ev n
above p with xv ∈ Kad
v (̟v ) (and λp (xp ) = ∞ if xv = 1 for some v|p).
Definition 5.24. We say that an element γ ∈ G(Q) is non-degenerate if the smallest
normal subgroup of G containing γ (which is necessarily defined over Q) contains the
derived group Gder .
Lemma 5.25. Let γ ∈ G(Q) be non-degenerate. Then λp (γ) < ∞ for all p and λp (γ) = 0
for almost all p (depending on γ).
Proof. If λp (γ) = ∞ for some p, then there exists a semisimple ideal 0 6= h of g⊗Qp such
that Ad(γ)|h = 1, i.e., γ lies in the kernel of the corresponding projection π : G → H of
reductive algebraic groups defined over Qp given by the action of G on h. This contradicts
the assumption on γ.
Consider the description of λp given in Remark 5.23. To have λp (γ) > 0 means that
Ad(γ)l ≡ l (mod ̟v Λ) for some place v of A above p. If this is the case for infinitely
many p, then we may conclude that the linear map Ad(γ) acts as the identity on a Qsimple summand gi for some i, which again contradicts the assumption on γ.
Lemma 5.26. Let P be a parabolic subgroup of G defined over Q, U its unipotent radical,
and M a Levi subgroup of P defined over Q. Let µ ∈ M(Q) be non-degenerate in G. Then
for every p we have λp (µu) ≪µ 1 for all u ∈ U(Afin ) ∩ K. Moreover, for almost all p
(depending on µ) we have λp (µu) = 0 for all u ∈ U(Afin ) ∩ K.
Proof. Assume on the contrary that λp (µu) is unbounded for some p. Since U(Afin ) ∩ K
is compact and p−λp (µ·) is a continuous function, we can then find u ∈ U(Afin ) ∩ K with
λp (µu) = ∞, or Ad(µu)|h = 1 for some semisimple ideal 0 6= h of g ⊗ Qp . Consider
the corresponding projection π : G → H of reductive algebraic groups defined over Qp .
We have π(µu) = 1, and therefore π(µ) = π(u) = 1, since π(µ) is semisimple and π(u)
unipotent. This means that µ is contained in the kernel of π, a normal subgroup of G not
containing Gder , which contradicts the assumption on µ.
Approximation principle
59
To show the second assertion, we may assume that p is such that we are in the situation
of Lemma 5.13 (applied to Gad ) and that µ ∈ Kp . If λp (µu) > 0 for some u ∈ U(Afin ) ∩
K, then we can apply the previous argument to the reduction modulo p of Ad(µu) ∈ Kad
p
and conclude that we have in fact λp (µ) > 0. By Lemma 5.25, this is only possible for
finitely many p under our assumption on µ.
We now consider λp on the unipotent radical U of a parabolic subgroup P of G that
is defined over Q. Recalling the description of Remark 5.23, if u is the Lie algebra of
Lr
U, then we can write u =
i=1 ui , where each ui is an Fi -vector space. Moreover, the
spaces ui are non-trivial for all i = 1, . . . , r if and only if the smallest normal subgroup
of G containing U is the full derived group Gder .
Lemma 5.27. Let P be a parabolic subgroup of G defined over Q and U its unipotent
radical, and assume that the smallest normal subgroup of G containing U is the derived
L
group Gder . Let u be the Lie algebra of U and write u = ri=1 ui , where ui is a non-trivial
Fi -vector space.
Let Lu ⊂ u be an oA -lattice and set
λ̃p (exp x) := max{k ≥ 0 : ∃v|p : x ∈ ̟vev k Lu ⊗ Zp },
x ∈ Lu ⊗ Ẑ,
where v ranges over the places of A lying over p, ̟v is a prime element of the associated
local field Av and ev the corresponding ramification index.
Then we have the following:
1. For every prime p the difference |λp (exp x) − λ̃p (exp x)| is bounded on Lu ⊗ Ẑ.
2. For almost all p we have λp (exp x) = λ̃p (exp x), x ∈ Lu ⊗ Ẑ.
3. For a suitable choice of Lu we have log(U(Afin ) ∩ K) ⊂ Lu ⊗ Ẑ.
Proof. The third assertion is clear, since log(U(Afin ) ∩ K) is a compact subset of u ⊗ Afin ,
and is therefore contained in a set of the form Lu ⊗ Ẑ.
To show the first and second assertions, it suffices to consider the Q-simple factors of
ad
G one at a time. So, we may assume that Gad is Q-simple and therefore of the form
ResF/Q G ad for an absolutely simple adjoint group G ad defined over a finite extension F of
Q. Consider the restriction to u of the adjoint representation ad : g → EndF (g) of F -Lie
algebras. Clearly ad(Lu ) is a finitely generated oF -submodule of EndF (g) and therefore
there exists a positive integer N such that ad(Lu )Λ ⊂ N −1 Λ. Let r be a positive integer
such that ad(u)i = 0 for all u ∈ u and i > r.
If we now have x ∈ ̟vk Lu ⊗ Zp for some k ≥ 0, then we obtain ad(x)Λ ⊗ Zp ⊂
N −1 ̟vk Λ⊗Zp , and therefore Ad(exp x) = exp ad(x) satisfies Ad(exp x)l−l ∈ (r!)−1 N −r ̟vk Λ⊗
Tobias Finis, Erez Lapid
60
Zp for all l ∈ Λ. We conclude that λp (exp x) ≥ λ̃p (exp x) − vp (r!N r ). Using the logarithm map on ad(u), which is again given by polynomials, we obtain also the opposite
inequality λ̃p (exp x) ≥ λp (exp x) − vp (M) for a suitable non-zero integer M. Taken
together, these inequalities amount to the first two assertions of the lemma.
Finally, we show how to bound the unipotent orbital integrals which appear naturally
in the limit multiplicity problem.
Corollary 5.28. There exists a constant ε > 0, depending only on G and ρ0 , with the
following property. Let K0 be an open compact subgroup of G(Afin ). Let P be a parabolic
subgroup of G defined over Q and U its unipotent radical, and let H ⊂ Gder be the
smallest normal subgroup of G containing U. Then
Z
Z
1K (k −1 uk) dk du ≪P, K0 lev(K, H(Afin )+ )−ε
U (Afin )
K
for all open subgroups K ⊂ K0 .
Proof. Estimating the integral on the left-hand side by
Z
Z
[K0 : K ∩ K0 ]
1K∩K (k −1 uk) dk du,
U (Afin )
K
we may first reduce to the case where K is contained in K.
We may assume that the Haar measure on U(Afin ) is the product of the measures
on U(Qp ) that assign the open compact subgroups U(Qp ) ∩ Kp measure one. Let N =
Q
lev(K, H(Afin )+ ) = p pnp . We apply Corollary 5.8 to the group H to estimate
Z
Y
1K (k −1 uk) dk ≪ β(N, u, δ)−ε =
βp (pnp , up , δ)−ε
K
p
for all u ∈ U(Afin ) ∩ K, where
(
pn p ,
βp (pnp , up , δ) =
1,
We obtain
Z
Z
U (Afin )
−1
K
1K (k uk) dk du ≪
λH
p (up ) < δnp ,
otherwise.
Z
β(N, u, δ)−εdu
U (Afin )∩K
YZ
=
p|N
βp (pnp , up , δ)−ε dup .
U (Qp )∩Kp
To estimate the integral over U(Qp ) ∩ Kp , we write
Z
βp (pnp , up , δ)−ε dup = µp (⌈δnp ⌉) + p−εnp (1 − µp (⌈δnp ⌉)),
U (Qp )∩Kp
(5.13)
Approximation principle
61
where
µp (n) = vol{up ∈ U(Qp ) ∩ Kp : λH
p (up ) ≥ n},
n ≥ 0.
Therefore it only remains to estimate µp (n). It follows from Lemma 5.27 that we have
′
the bound µp (n) ≪G,U,ρ0 p−n . This yields a bound for (5.13) of the form Cp p−ε np , where
ε′ , Cp > 0 and Cp = 1 for almost all p, which shows the assertion.
A Bounds for the number of solutions of polynomial congruences
In this appendix we give some simple general bounds for the number of solutions of
polynomial congruences that are essential for the argument in §5 of the body of the paper.
The emphasis lies on general non-trivial bounds which are uniform in all parameters and
are obtained by elementary methods, and not on the quality of the bounds.
Let V = Spec BZ be an affine scheme that is flat and of finite type over Spec Z. Fix
once and for all generators y1 , . . . , yr of the affine coordinate ring BZ of V , or equivalently
a closed embedding of V into an affine space Ar over Z. Henceforth, all implied constants
will depend on V and on this embedding. For any commutative ring R let BR = BZ ⊗R be
the base change of the ring BZ to R, VR = Spec BR the base change of V , and V (R) the
set of R-points of V . We assume throughout that the generic fiber VQ of V is an absolutely
irreducible variety, i.e., that the ring BQ̄ is an integral domain. Let s be the dimension of
the variety VQ .
For any commutative ring R and any integer d ≥ 0 we let BR,≤d be the R-submodule
of BR generated by the monomials of degree ≤ d in the generators y1 , . . . , yr . We define
the degree deg f of an element f ∈ BR as the smallest integer d ≥ 0 such that f ∈ BR,≤d .
The main results of this appendix are the following two lemmas. The first lemma
concerns polynomial congruences modulo p.
Lemma A.1. For all primes p and all non-zero f ∈ BFp we have
#{x ∈ V (Fp ) : f (x) = 0} ≪V,yi deg f · pdim VQ −1 .
For the next lemma, which treats congruences modulo pn , note that the given affine
embedding V ֒→ Ar induces for every p a measure on the smooth part of the p-adic
analytic set V (Qp ) ⊂ Qrp (cf. [Ser81, §3.3], [Oes82, §3]).
Lemma A.2. Assume that the generic fiber VQ of V is a smooth variety containing the
origin of Ar . Then for each d > 0 there exists a constant ε(d) > 0, depending only on V ,
such that for any f ∈ BQp of degree ≤ d and any m, n ∈ Z with f (V (Zp )) 6⊂ pm Zp we
have
vol {x ∈ V (Zp ) ∩ pZrp : f (x) ≡ 0 (mod pn )} ≪V,yi ,d p−ε(d)(n−m+1) .
Tobias Finis, Erez Lapid
62
Remark A.3. For any Zariski closed subset W ⊂ ArFp defined by polynomials of degree ≤
n, the cardinality of W (Fp ) can be bounded by C(r, n)pdim W (cf. [CvdDM92, Proposition
3.3], which is based on the Lang-Weil estimates). This implies the statement of Lemma A.1
with deg f replaced by an unspecified function of deg f , which would be sufficient for our
purposes. However, the proof given below is much simpler than the proof of the Lang-Weil
estimates. A proof of the lemma can also be obtained by combining [Oes82, Proposition
1] (applied to the set Ω = Frp ) with basic intersection theory in the affine space ArFp .
Remark A.4. In Lemma A.2, the smoothness of the variety VQ is not essential, but it
simplifies the argument considerably. In the main part of the paper we are interested in
the case of linear algebraic groups, which are smooth varieties.
Remark A.5. For fixed p and f , the Poincaré series associated to the sequence
vn = vol {x ∈ V (Zp ) ∩ pZrp : f (x) ≡ 0 (mod pn )} ,
n ≥ 0,
is given by a rational function [Igu89, Theorem 1]. (In fact, we may allow general varieties VQp and more general functions f in this statement.) We will not make use of this
fact.
Remark A.6. An estimate of the form ≪V,f p−n for the volume of the subset
W (f, n) = {x ∈ V (Zp ) : ∃y ∈ V (Zp ) : f (y) = 0, y ≡ x (mod pn )}
of V (Zp ) (and for the volume of its intersection with pZrp ) can be obtained from [Oes82,
Théorème 1]. For f ∈ BZp we obviously have the inclusion W (f, n) ⊂ {x ∈ V (Zp ) :
f (x) ≡ 0 (mod pn )}, but in general no equality.
The proofs of Lemmas A.1 and A.2 are based on the Noether normalization lemma
(e.g., [Nag62, Theorem 14.4]), which gives the existence of algebraically independent
elements x1 , . . . , xs ∈ BZ and of a positive integer D such that BZ [1/D] is integral
over AZ [1/D], where AZ = Z[x1 , . . . , xs ] is the polynomial ring over Z generated by
x1 , . . . , xs inside BZ . This means that each generator yj , j = 1, . . . , r, satisfies a monic
polynomial equation over AZ [1/D]. In fact, one may take x1 , . . . , xs to be elements of the
Z-module generated by y1 , . . . , yr inside BZ (cf. [ibid., §14, Exercise]). Let
ξ = (x1 , . . . , xs ) :
V → As
be the associated finite morphism, and for any commutative ring R write ξR : VR → AsR
for its base change to R. If we are given a field extension K of Q and a smooth point v of
V (K), then we may arrange that the morphism ξK is smooth at v. (To see this, note that if
we write x1 , . . . , xs as linear combinations of the generators y1 , . . . , yr with coefficients
Approximation principle
63
cij ∈ Z, then the conclusion of the normalization lemma holds for all (cij ) in an open
dense subset of the space of all matrices. The smoothness condition also defines an open
dense subset. Since the intersection of both subsets is still open and dense, it contains a
point with coordinates in Z.) In the following we write AR = AZ ⊗ R = R[x1 , . . . , xs ] ⊂
BR for a commutative ring R.
Denote by Q(R) the quotient field of an integral domain R. Recall a simple fact: if A
is an integrally closed integral domain and L a finite extension of K = Q(A), then the
integral closure B of A in L is precisely the set of all elements x ∈ L whose minimal
polynomial over K has coefficients in A. In particular, we have then NL/K (x) ∈ A and
NL/K (x) ∈ xA[x] for all x ∈ B.
We apply this to the field extension Q(BZp )/Q(AZp ) for arbitrary p and to the corresponding norm map
N = NQ(BZp )/Q(AZp ) : Q(BZp ) → Q(AZp ).
We record the following facts.
Lemma A.7. Let ÃZp be the integral closure of the polynomial ring AZp inside the ring
BQp . We have then:
1. N(BQp ) ⊂ AQp .
2. Qp ÃZp = BQp .
3. N(ÃZp ) ⊂ AZp .
4. N(f ) ∈ f ÃZp
for all f ∈ ÃZp .
5. deg N(f ) ≪ deg f for all f ∈ BQp .
Proof. The first assertion follows immediately from the fact that BQp is integral over AQp .
The second to fourth assertions are a consequence of the remarks preceding the lemma.
For the fifth assertion, recall that each yj , j = 1, . . . , r, satisfies a monic polynomial
equation over AZ [1/D], of degree nj , say. Let zk , k = 1, . . . , l be an enumeration of
the monomials y1α1 . . . yrαr where 0 ≤ αi < ni , i = 1, . . . , r. Then the zk ’s generate the
AQ -module BQ , and therefore also the AQp -module BQp for any p. Moreover, an arbitrary
monomial of degree n in y1 , . . . , yr can be expressed as a linear combination of the zk ’s
with coefficients in AQ of degree ≪ n. In addition, if f ∈ BQp is expressed as a linear
P
combination fk zk with fk ∈ AQp , then N(f ) is given by a fixed polynomial in the fk ’s
with coefficients in AQ . This clearly shows the assertion.
We now list several additional properties that hold (after fixing x1 , . . . , xs ) for almost
all p. For all p prime to the positive integer D introduced above we have ÃZp ⊃ BZp
Tobias Finis, Erez Lapid
64
and the ring BFp is integral over AFp = Fp [x1 , . . . , xs ]. In addition, for almost all p the
scheme VFp is an absolutely irreducible variety over Fp , and in particular the ring BFp is
an integral domain. For all such p we can consider the characteristic p norm map
N̄ = NQ(BFp )/Q(AFp ) : Q(BFp )× → Q(AFp )× .
Lemma A.8. We have for almost all p:
1. N(BZp ) ⊂ AZp .
2. ÃZp = BZp .
3. N̄(BFp ) ⊂ AFp .
4. N̄(f ) ∈ f BFp for all f ∈ BFp .
5. The diagram
N =NQ(BZ
)/Q(AZ )
p
N̄ =NQ(BF
)/Q(AF )
p
p
BZp −−−−−−−−−−−→
y
p
AZp
y
BFp −−−−−−−−−−−→ AFp
commutes, where the vertical lines are the reduction maps modulo p.
Proof. The first, third and fourth assertions are an immediate consequence of the remarks
preceding Lemmas A.7 and A.8. To prove the last assertion, fix a basis b1 , . . . , bt of Q(BZ )
over Q(AZ ) consisting of elements of BZ . For any f ∈ BZ let M(f ) be the matrix (with
entries in the field Q(AZ ) = Q(x1 , . . . , xs )) representing multiplication by f in the basis
b1 , . . . , bt . Then N(f ) = det M(f ). Let S be the finite set of all primes p that divide
the denominator of an entry of M(yj ) for some j = 1, . . . , r. Since BZ is generated by
the elements y1 , . . . , yr as a Z-algebra, it follows that only primes in S appear in the
denominator of M(f ) for any f ∈ BZ . For almost all p, the images b̄1 , . . . , b̄t of b1 , . . . , bt
/ S we may
in BFp form a basis of Q(BFp ) over Q(AFp ), and under the assumption p ∈
reduce M(f ) modulo p and obtain the matrix representing multiplication by f¯ in the basis
b̄1 , . . . , b̄t . Taking determinants we obtain the fifth assertion.
As for the second assertion, it only remains to show that ÃZp ⊂ BZp for almost all p.
/ pBZp , with p−1 f ∈ ÃZp . By Lemma
Assuming the contrary, there exists f ∈ BZp , f ∈
A.7, we have then N(p−1 f ) = p−t N(f ) ∈ AZp , while on the other hand N(f ) = N̄(f¯) 6=
0 by the fifth assertion of the present lemma, which is a contradiction. This proves the
second assertion.
We can now prove the first of the two lemmas stated at the beginning.
Approximation principle
65
Proof of Lemma A.1. Our proof is based on the following elementary estimate (see [Sch76,
Ch. IV, Lemma 3A]) for the cardinality of the zero set in Fsp of a non-zero polynomial
g ∈ Fp [x1 , . . . , xs ]:
#{x ∈ Fsp : g(x) = 0} ≤ (deg g)ps−1 .
(A.1)
In the situation of Lemma A.1 it is clearly enough to prove the assertion for almost all
p (depending on V ). We can therefore assume the existence of algebraically independent
elements x1 , . . . , xs ∈ BZ such that the assertions of Lemma A.8 hold. Let ξ¯ = ξFp :
VFp → AsFp be the reduction modulo p of the morphism ξ. Note that since the generators
yj satisfy some fixed monic equations with coefficients in AZ [1/D], the cardinality of the
fibers of ξ¯ is bounded by a constant that is independent of p (namely by the product of the
degrees of these equations).
By Lemma A.8, for any 0 6= f ∈ BFp we have g = N̄(f ) ∈ AFp = Fp [x1 , . . . , xs ],
¯
g 6= 0, and g ∈ f BFp . Therefore, the zero set of f is contained in the zero set of g ◦ ξ.
Invoking (A.1), this implies that
¯
#{v ∈ V (Fp ) : f (v) = 0} ≤ #{v ∈ V (Fp ) : g(ξ(v))
= 0}
≪ #{x ∈ Fsp : g(x) = 0}
≤ (deg g)ps−1.
Combining the last assertions of Lemmas A.7 and A.8, we have the degree bound deg g ≪
deg f and obtain the assertion.
We now turn to the proof of Lemma A.2. Again, we first consider the case of an affine
space and replace (A.1) by the following estimate.
Lemma A.9. Let f ∈ Zp [X1 , . . . , Xs ] be a polynomial of degree ≤ d and suppose that f
is not congruent to zero modulo p. Then we have
n
s
n
s n+s−1
p− d
vol {x ∈ Zp : f (x) ≡ 0 (mod p )} ≤ d
(A.2)
s−1
for any n ≥ 0. In other words, the number of solutions to f (x1 , . . . , xs ) = 0 in (Z/pn Z)s
n(s− 1 )
d .
is bounded by ds n+s−1
p
s−1
Remark A.10. This result is essentially [Art88, Lemma 7.1]. For convenience we include
a proof. Asymptotic estimates for the left-hand side of (A.2) as n → ∞ have been obtained by Igusa [Igu75, Igu77] and Loeser [Loe86]. Here, we are just interested in a
non-trivial bound that depends in a very transparent way on the polynomial f .
Proof. We may assume that n > 0. We prove the statement by induction on the number
of variables s. In the case s = 1 we can factorize f over a suitable algebraic extension F
Tobias Finis, Erez Lapid
66
of Qp as
d
Y
f (x) =
(αi x + βi ),
i=1
where max(|αi |, |βi |) = 1 for all i = 1, . . . , d and |·| denotes the extension of the standard
absolute value of Qp to F . If x ∈ Zp and |f (x)| ≤ p−n then |αi x + βi | ≤ p−n/d for some
i, where necessarily |αi | = 1. Thus
n
vol ({x ∈ Zp : f (x) ≡ 0 (mod pn )}) ≤ dp− d ,
as claimed.
P
For the induction step we write f = di=0 fi (x1 , . . . , xs−1 )xis where at least one of the
polynomials fi , say fi0 , is non-zero modulo p. Fix x1 , . . . , xs−1 . Let
j = max{l ≥ 0 : fi (x1 , . . . , xs−1 ) ≡ 0 (mod pl ) for all i}.
Applying the one-variable case to p−j f (x1 , . . . , xs−1 , ·) we get
vol({xs ∈ Zp : f (x1 , . . . , xs ) ≡ 0
(mod pn )}) ≤ min(dp−
n−j
d
, 1).
Thus, the left-hand side of (A.2) is bounded by
n
X
≤d
j=0
n
X
j=0
dp−
p−
n−j
d
n−j
d
vol {x ∈ Zps−1 : fi (x1 , . . . , xs−1 ) ≡ 0 (mod pj ) for all i}
vol {x ∈ Zps−1 : fi0 (x1 , . . . , xs−1 ) ≡ 0 (mod pj )} .
It remains to apply the induction hypothesis and the binomial identity
n+s−1
.
s−1
Pn
j+s−2
j=0
s−2
=
Proof of Lemma A.2. Note first that it is enough to establish the following two claims.
1. For every p and any v ∈ V (Zp ) there exists an integer N(p, v) ≥ 0 such that for all
d ≥ 1 we have
vol {u ∈ V (Zp ) ∩ (v + pN (p,v) Zrp ) : f (u) ≡ 0
(mod pn )} ≪p,d,v p−ε(p,d,v)(n−m+1)
for any f ∈ BQp of degree ≤ d and m, n ∈ Z with f (V (Zp )) 6⊂ pm Zp , where
ε(p, d, v) > 0 depends on p, d and v.
2. The assertion of the lemma is true for almost all p (depending only on V ).
Approximation principle
67
Namely, granted the second claim, it only remains to consider the assertion of the
lemma for finitely many primes p, which we may treat one at a time. For each such p we
invoke the first claim and cover the compact set V (Zp ) ∩ pZrp by finitely many sets of the
form V (Zp ) ∩ (v + pN (p,v) Zrp ). It is then enough to sum the resulting estimates to obtain
the assertion.
It remains to prove the two claims. We start with the first one. Let p be arbitrary and
v ∈ V (Zp ). As noted above, we can choose x1 , . . . , xs (depending on p and v) such that
the associated morphism ξQp : VQp → AsQp is smooth at v. This implies the existence of
an integer N(p, v) ≥ 0 such that ξQp induces an injection of V (Zp ) ∩ (v + pN (p,v) Zrp ) into
Zsp that multiplies volumes by a positive constant C(p, v).
For any p and f ∈ BQp , f 6= 0, set
νp (f ) := max{i ∈ Z : p−i f ∈ ÃZp }.
Clearly, vp (f (x)) ≥ νp (f ) for all x ∈ V (Zp ). In the situation of the first claim we have
therefore νp (f ) ≤ m − 1. We are reduced to prove the estimate
vol {u ∈ V (Zp ) ∩ (v + pN (p,v) Zrp ) : f (u) ≡ 0 (mod pn )} ≪d p−ε(p,d,v)n (A.3)
for all f ∈ BQp of degree ≤ d with νp (f ) = 0, or equivalently for all f ∈ ÃZp \ pÃZp
with deg f ≤ d.
Let Ã≤d = ÃZp ∩ BZp ,≤d , which is a free Zp -module of finite rank. Since N(f ) 6= 0
for all f 6= 0, and the set ̲d \ p̲d is compact, we see that
n(p, d) := max{vp (N(f )) : f ∈ Ã≤d \ pÃ≤d } < ∞.
(Here vp is the usual p-valuation on AZp .)
It follows from Lemma A.7, part 4, that the set of all u ∈ V (Zp ) for which f (u) ∈
n
p Zp is contained in the set of all u with F (ξ(u)) ∈ pn Zp , where F = N(f ) ∈ AZp . This
implies that
vol {u ∈ V (Zp ) ∩ (v + pN (p,v) Zrp ) : f (u) ∈ pn Zp }
≤ vol {u ∈ V (Zp ) ∩ (v + pN (p,v) Zrp ) : F (ξ(u)) ∈ pn Zp }
≤C(p, v)−1 vol {x ∈ Zsp : F (x) ∈ pn Zp } .
Invoking Lemma A.9 and using that vp (F ) ≤ n(p, d), we obtain the desired estimate
(A.3). This finishes the proof of the first claim.
We now consider the second claim. By a suitable choice of x1 , . . . , xs we can arrange
that the morphism ξQ : VQ → AsQ is smooth at the origin. For almost all p the map ξQp
induces then a volume preserving injection of V (Zp ) ∩ pZrp into Zsp . Moreover, by Lemma
A.8 for almost all p we have ÃZp = BZp and furthermore f ∈ ÃZp \ pÃZp = BZp \ pBZp
implies that N(f ) = N̄(f¯) 6= 0, or vp (N(f )) = 0. This means that n(p, d) = 0 for all d.
Tobias Finis, Erez Lapid
68
For all such p we can replace in the previous argument V (Zp ) ∩ (v + pN (p,v) Zrp ) by
V (Zp ) ∩ pZrp , and C(p, v) by 1, and since we have n(p, d) = 0, we obtain
vol {u ∈ V (Zp ) ∩ pZrp : f (u) ≡ 0
(mod pn )} ≪d p−ε(d)n
for all f ∈ BQp of degree ≤ d with νp (f ) = 0, which establishes the second claim and
finishes the proof.
Acknowledgments. T.F. partially supported by DFG Heisenberg grant # FI 1795/1-1. E.L. partially supported
by grant #711733 from the Minerva Foundation.
References
[ABB+ ]
Miklos Abert, Nicolas Bergeron, Ian Biringer, Tsachik Gelander, Nikolay
Nikolov, Jean Raimbault, and Iddo Samet, On the growth of L2 -invariants
for sequences of lattices in Lie groups, arXiv:1210.2961.
[AJL83]
Henning Haahr Andersen, Jens Jørgensen, and Peter Landrock, The projective indecomposable modules of SL(2, pn ), Proc. London Math. Soc. (3) 46
(1983), no. 1, 38–52. MR 684821 (84f:20044)
[ANS03]
Miklós Abért, Nikolay Nikolov, and Balázs Szegedy, Congruence subgroup
growth of arithmetic groups in positive characteristic, Duke Math. J. 117
(2003), no. 2, 367–383. MR 1971298 (2004a:20057)
[Art88]
James Arthur, The local behaviour of weighted orbital integrals, Duke Math.
J. 56 (1988), no. 2, 223–293. MR 932848 (89h:22036)
[BM55]
A. Borel and G. D. Mostow, On semi-simple automorphisms of Lie algebras,
Ann. of Math. (2) 61 (1955), 389–405. MR MR0068531 (16,897d)
[Bor70]
Armand Borel, Properties and linear representations of Chevalley groups,
Seminar on Algebraic Groups and Related Finite Groups (The Institute for
Advanced Study, Princeton, N.J., 1968/69), Lecture Notes in Mathematics,
Vol. 131, Springer, Berlin, 1970, pp. 1–55. MR MR0258838 (41 #3484)
[Bor91]
, Linear algebraic groups, second ed., Graduate Texts in Mathematics, vol. 126, Springer-Verlag, New York, 1991. MR MR1102012
(92d:20001)
[Bou98]
Nicolas Bourbaki, Lie groups and Lie algebras. Chapters 1–3, Elements
of Mathematics (Berlin), Springer-Verlag, Berlin, 1998, Translated from
the French, Reprint of the 1989 English translation. MR MR1728312
(2001g:17006)
Approximation principle
69
[Bou02]
, Lie groups and Lie algebras. Chapters 4–6, Elements of Mathematics (Berlin), Springer-Verlag, Berlin, 2002, Translated from the 1968 French
original by Andrew Pressley. MR MR1890629 (2003a:17001)
[BS53]
A. Borel and J.-P. Serre, Sur certains sous-groupes des groupes de Lie
compacts, Comment. Math. Helv. 27 (1953), 128–139. MR MR0054612
(14,948d)
[BS97]
Yiftach Barnea and Aner Shalev, Hausdorff dimension, pro-p groups, and
Kac-Moody algebras, Trans. Amer. Math. Soc. 349 (1997), no. 12, 5073–
5091. MR 1422889 (98b:20041)
[BT65]
Armand Borel and Jacques Tits, Groupes réductifs, Inst. Hautes Études Sci.
Publ. Math. (1965), no. 27, 55–150. MR MR0207712 (34 #7527)
[BT71]
A. Borel and J. Tits, Éléments unipotents et sous-groupes paraboliques de
groupes réductifs. I, Invent. Math. 12 (1971), 95–104. MR MR0294349 (45
#3419)
[Bur07a]
Timothy C. Burness, Fixed point ratios in actions in finite classical groups.
II, J. Algebra 309 (2007), no. 1, 80–138. MR 2301234 (2008a:20003)
[Bur07b]
, Fixed point ratios in actions of finite classical groups. I, J. Algebra
309 (2007), no. 1, 69–79. MR 2301233 (2008a:20002)
[Bur07c]
, Fixed point ratios in actions of finite classical groups. III, J. Algebra
314 (2007), no. 2, 693–748. MR 2344583 (2008j:20148a)
[Bur07d]
, Fixed point ratios in actions of finite classical groups. IV, J. Algebra
314 (2007), no. 2, 749–788. MR 2344584 (2008j:20148b)
[Col07]
Michael J. Collins, On Jordan’s theorem for complex linear groups, J. Group
Theory 10 (2007), no. 4, 411–423. MR 2334748 (2008g:20106)
[Col08]
, Modular analogues of Jordan’s theorem for finite linear groups, J.
Reine Angew. Math. 624 (2008), 143–171. MR 2456628 (2009j:20071)
[CP84]
David A. Cox and Walter R. Parry, Genera of congruence subgroups in Qquaternion algebras, J. Reine Angew. Math. 351 (1984), 66–112. MR 749678
(85i:11029)
[CPSvdK77] E. Cline, B. Parshall, L. Scott, and Wilberd van der Kallen, Rational and
generic cohomology, Invent. Math. 39 (1977), no. 2, 143–163. MR 0439856
(55 #12737)
[Cur65]
Charles W. Curtis, Central extensions of groups of Lie type, J. Reine Angew.
Math. 220 (1965), 174–185. MR 0188299 (32 #5738)
70
[Cur70]
Tobias Finis, Erez Lapid
C. W. Curtis, Modular representations of finite groups with split (B, N)pairs, Seminar on Algebraic Groups and Related Finite Groups (The Institute for Advanced Study, Princeton, N.J., 1968/69), Springer, Berlin, 1970,
pp. 57–95. MR 0262383 (41 #6991)
[CvdDM92] Zoé Chatzidakis, Lou van den Dries, and Angus Macintyre, Definable sets
over finite fields, J. Reine Angew. Math. 427 (1992), 107–135. MR 1162433
(94c:03049)
[DdSMS99] J. D. Dixon, M. P. F. du Sautoy, A. Mann, and D. Segal, Analytic pro-p
groups, second ed., Cambridge Studies in Advanced Mathematics, vol. 61,
Cambridge University Press, Cambridge, 1999. MR 1720368 (2000m:20039)
[dS00]
Marcus du Sautoy, The zeta function of sl2 (Z), Forum Math. 12 (2000), no. 2,
197–221. MR 1740889 (2001c:11097a)
[dST02]
Marcus du Sautoy and Gareth Taylor, The zeta function of sl2 and resolution
of singularities, Math. Proc. Cambridge Philos. Soc. 132 (2002), no. 1, 57–73.
MR 1866324 (2002j:11100)
[Dyn52a] E. B. Dynkin, Maximal subgroups of the classical groups, Trudy Moskov.
Mat. Obšč. 1 (1952), 39–166. MR 0049903 (14,244d)
[Dyn52b]
, Semisimple subalgebras of semisimple Lie algebras, Mat. Sbornik
N.S. 30(72) (1952), 349–462 (3 plates). MR 0047629 (13,904c)
[EMV09] M. Einsiedler, G. Margulis, and A. Venkatesh, Effective equidistribution for
closed orbits of semisimple groups on homogeneous spaces, Invent. Math.
177 (2009), no. 1, 137–212. MR 2507639 (2011a:22030)
[FL15]
Tobias Finis and Erez Lapid, An approximation principle for congruence subgroups II: application to the limit multiplicity problem, 2015,
arXiv:1504.04795.
[FLM15]
Tobias Finis, Erez Lapid, and Werner Müller, Limit multiplicities for principal
congruence subgroups of GL(n) and SL(n), J. Inst. Math. Jussieu 14 (2015),
no. 3, 589–638. MR 3352530
[Gre66]
Marvin J. Greenberg, Rational points in Henselian discrete valuation rings,
Inst. Hautes Études Sci. Publ. Math. (1966), no. 31, 59–64. MR 0207700 (34
#7515)
[Gre74]
, Strictly local solutions of Diophantine equations, Pacific J. Math. 51
(1974), 143–153. MR 0351994 (50 #4482)
[Gro96]
Benedict H. Gross, Groups over Z, Invent. Math. 124 (1996), no. 1-3, 263–
279. MR 1369418 (96m:20075)
Approximation principle
71
[GSK09]
Jon González-Sánchez and Benjamin Klopsch, Analytic pro-p groups of
small dimensions, J. Group Theory 12 (2009), no. 5, 711–734. MR 2554763
(2010m:20036)
[HP95]
E. Hrushovski and A. Pillay, Definable subgroups of algebraic groups
over finite fields, J. Reine Angew. Math. 462 (1995), 69–91. MR 1329903
(97f:20059)
[Hum78]
James E. Humphreys, Introduction to Lie algebras and representation theory,
Graduate Texts in Mathematics, vol. 9, Springer-Verlag, New York, 1978,
Second printing, revised. MR 499562 (81b:17007)
[Igu75]
Jun-ichi Igusa, Complex powers and asymptotic expansions. II. Asymptotic
expansions, J. Reine Angew. Math. 278/279 (1975), 307–321. MR 0404215
(53 #8018)
[Igu77]
Jun-Ichi Igusa, Some observations on higher degree characters, Amer. J.
Math. 99 (1977), no. 2, 393–417. MR 0441933 (56 #324)
[Igu89]
Jun-ichi Igusa, Universal p-adic zeta functions and their functional equations,
Amer. J. Math. 111 (1989), no. 5, 671–716. MR 1020825 (91e:11142)
[Ila95]
Ishai Ilani, Analytic pro-p groups and their Lie algebras, J. Algebra 176
(1995), no. 1, 34–58. MR 1345293 (96h:20059)
[Ila99]
, Zeta functions related to the group SL2 (Zp ), Israel J. Math. 109
(1999), 157–172. MR 1679595 (2000b:11107)
[Jan87]
Jens Carsten Jantzen, Representations of algebraic groups, Pure and Applied
Mathematics, vol. 131, Academic Press Inc., Boston, MA, 1987. MR 899071
(89c:20001)
[Jan97]
, Low-dimensional representations of reductive groups are semisimple, Algebraic groups and Lie groups, Austral. Math. Soc. Lect. Ser.,
vol. 9, Cambridge Univ. Press, Cambridge, 1997, pp. 255–266. MR 1635685
(99g:20079)
[Kat88]
Nicholas M. Katz, Gauss sums, Kloosterman sums, and monodromy groups,
Annals of Mathematics Studies, vol. 116, Princeton University Press, Princeton, NJ, 1988. MR 955052 (91a:11028)
[Klo05]
Benjamin Klopsch, On the Lie theory of p-adic analytic groups, Math. Z. 249
(2005), no. 4, 713–730. MR 2126210 (2005k:22012)
[KV09]
Benjamin Klopsch and Christopher Voll, Zeta functions of three-dimensional
p-adic Lie algebras, Math. Z. 263 (2009), no. 1, 195–210. MR 2529493
(2011a:11209)
72
Tobias Finis, Erez Lapid
[Lar10]
Michael Larsen, Exponential generation and largeness for compact p-adic Lie
groups, Algebra Number Theory 4 (2010), no. 8, 1029–1038. MR 2832632
[Laz65]
Michel Lazard, Groupes analytiques p-adiques, Inst. Hautes Études Sci. Publ.
Math. (1965), no. 26, 389–603. MR 0209286 (35 #188)
[LL04]
Michael Larsen and Alexander Lubotzky, Normal subgroup growth of linear
groups: the (G2 , F4 , E8 )-theorem, Algebraic groups and arithmetic, Tata Inst.
Fund. Res., Mumbai, 2004, pp. 441–468. MR 2094120
[LLS02]
Ross Lawther, Martin W. Liebeck, and Gary M. Seitz, Fixed point ratios in
actions of finite exceptional groups of Lie type, Pacific J. Math. 205 (2002),
no. 2, 393–464. MR 1922740 (2003g:20004)
[Loe86]
F. Loeser, Une estimation asymptotique du nombre de solutions approchées
d’une équation p-adique, Invent. Math. 85 (1986), no. 1, 31–38. MR
MR842046 (87j:11136)
[LP11]
Michael J. Larsen and Richard Pink, Finite subgroups of algebraic groups, J.
Amer. Math. Soc. 24 (2011), no. 4, 1105–1158. MR 2813339
[LS91]
Martin W. Liebeck and Jan Saxl, Minimal degrees of primitive permutation
groups, with an application to monodromy groups of covers of Riemann surfaces, Proc. London Math. Soc. (3) 63 (1991), no. 2, 266–314. MR 1114511
(92f:20003)
[LS94]
Alexander Lubotzky and Aner Shalev, On some Λ-analytic pro-p groups, Israel J. Math. 85 (1994), no. 1-3, 307–337. MR 1264349 (95f:20047)
[LS99]
Martin W. Liebeck and Aner Shalev, Simple groups, permutation groups, and
probability, J. Amer. Math. Soc. 12 (1999), no. 2, 497–520. MR 1639620
(99h:20004)
[LS03]
Alexander Lubotzky and Dan Segal, Subgroup growth, Progress in Mathematics, vol. 212, Birkhäuser Verlag, Basel, 2003. MR 1978431 (2004k:20055)
[LS04]
Martin W. Liebeck and Gary M. Seitz, The maximal subgroups of positive dimension in exceptional algebraic groups, Mem. Amer. Math. Soc. 169 (2004),
no. 802, vi+227. MR 2044850 (2005b:20082)
[LT04]
Martin W. Liebeck and Donna M. Testerman, Irreducible subgroups of
algebraic groups, Q. J. Math. 55 (2004), no. 1, 47–55. MR 2043006
(2005b:20087)
[Lub95]
Alexander Lubotzky, Subgroup growth and congruence subgroups, Invent.
Math. 119 (1995), no. 2, 267–295. MR 1312501 (95m:20054)
Approximation principle
73
[McN99]
George J. McNinch, Semisimple modules for finite groups of Lie type, J. London Math. Soc. (2) 60 (1999), no. 3, 771–792. MR 1753813 (2001k:20096)
[Mor56]
V. V. Morozov, Proof of the theorem of regularity, Uspehi Mat. Nauk (N.S.)
11 (1956), no. 5(71), 191–194. MR 0088488 (19,527c)
[Nag62]
Masayoshi Nagata, Local rings, Interscience Tracts in Pure and Applied
Mathematics, No. 13, Interscience Publishers a division of John Wiley & Sons
New York-London, 1962. MR 0155856 (27 #5790)
[Nor87]
Madhav V. Nori, On subgroups of GLn (Fp ), Invent. Math. 88 (1987), no. 2,
257–275. MR 880952 (88d:20068)
[Oes82]
Joseph Oesterlé, Réduction modulo pn des sous-ensembles analytiques fermés
de ZN
p , Invent. Math. 66 (1982), no. 2, 325–341. MR 656627 (83j:12014)
[PR94]
Vladimir Platonov and Andrei Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics, vol. 139, Academic Press Inc., Boston,
MA, 1994, Translated from the 1991 Russian original by Rachel Rowen. MR
1278263 (95b:11039)
[Ric67]
R. W. Richardson, Jr., A rigidity theorem for subalgebras of Lie and associative algebras, Illinois J. Math. 11 (1967), 92–110. MR 0206170 (34 #5992)
[Sch76]
Wolfgang M. Schmidt, Equations over finite fields. An elementary approach,
Lecture Notes in Mathematics, Vol. 536, Springer-Verlag, Berlin, 1976. MR
0429733 (55 #2744)
[Sei74]
A. Seidenberg, Constructions in algebra, Trans. Amer. Math. Soc. 197
(1974), 273–313. MR 0349648 (50 #2141)
[Ser77]
Jean-Pierre Serre, Linear representations of finite groups, Springer-Verlag,
New York, 1977, Translated from the second French edition by Leonard L.
Scott, Graduate Texts in Mathematics, Vol. 42. MR 0450380 (56 #8675)
[Ser81]
, Quelques applications du théorème de densité de Chebotarev,
Inst. Hautes Études Sci. Publ. Math. (1981), no. 54, 323–401. MR 644559
(83k:12011)
[SS70]
T. A. Springer and R. Steinberg, Conjugacy classes, Seminar on Algebraic
Groups and Related Finite Groups (The Institute for Advanced Study, Princeton, N.J., 1968/69), Lecture Notes in Mathematics, Vol. 131, Springer, Berlin,
1970, pp. 167–266. MR 0268192 (42 #3091)
[Ste62]
Robert Steinberg, Générateurs, relations et revêtements de groupes
algébriques, Colloq. Théorie des Groupes Algébriques (Bruxelles, 1962),
74
Tobias Finis, Erez Lapid
Librairie Universitaire, Louvain, 1962, pp. 113–127. MR MR0153677 (27
#3638)
[Ste63]
, Representations of algebraic groups, Nagoya Math. J. 22 (1963),
33–56. MR MR0155937 (27 #5870)
[Ste68]
, Endomorphisms of linear algebraic groups, Memoirs of the American Mathematical Society, No. 80, American Mathematical Society, Providence, R.I., 1968. MR MR0230728 (37 #6288)
[Ste81]
, Generators, relations and coverings of algebraic groups. II, J. Algebra 71 (1981), no. 2, 527–543. MR MR630615 (83d:14025)
[Tit79]
J. Tits, Reductive groups over local fields, Automorphic forms, representations and L-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 1, Proc. Sympos. Pure Math., XXXIII, Amer. Math.
Soc., Providence, R.I., 1979, pp. 29–69. MR 546588 (80h:20064)
[Tit81]
Jacques Tits, Définition par générateurs et relations de groups avec BNpaires, C. R. Acad. Sci. Paris Sér. I Math. 293 (1981), no. 6, 317–322. MR
637803 (82k:20069)
| 4math.GR
|
Global parameter identification of stochastic
reaction networks from single trajectories
arXiv:1111.4785v1 [q-bio.MN] 21 Nov 2011
Christian L. Müller† , Rajesh Ramaswamy† , and Ivo F. Sbalzarini
Abstract We consider the problem of inferring the unknown parameters of a
stochastic biochemical network model from a single measured time-course of the
concentration of some of the involved species. Such measurements are available,
e.g., from live-cell fluorescence microscopy in image-based systems biology. In
addition, fluctuation time-courses from, e.g., fluorescence correlation spectroscopy
provide additional information about the system dynamics that can be used to more
robustly infer parameters than when considering only mean concentrations. Estimating model parameters from a single experimental trajectory enables single-cell
measurements and quantification of cell–cell variability. We propose a novel combination of an adaptive Monte Carlo sampler, called Gaussian Adaptation, and efficient exact stochastic simulation algorithms that allows parameter identification
from single stochastic trajectories. We benchmark the proposed method on a linear
and a non-linear reaction network at steady state and during transient phases. In addition, we demonstrate that the present method also provides an ellipsoidal volume
estimate of the viable part of parameter space and is able to estimate the physical
volume of the compartment in which the observed reactions take place.
1 Introduction
Systems biology implies a holistic research paradigm, complementing the reductionist approach to biological organization [16, 15]. This frequently has the goal of
mechanistically understanding the function of biological entities and processes in
interaction with the other entities and processes they are linked to or communicate
with. A formalism to express these links and connections is provided by network
Christian L. Müller, Rajesh Ramaswamy, and Ivo F. Sbalzarini:
Institute of Theoretical Computer Science and Swiss Institute of Bioinformatics, ETH Zurich, CH–
8092 Zurich, Switzerland, e-mail: [email protected], [email protected], [email protected].
† These authors contributed equally to this work.
1
2
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
models of biological processes [4, 1]. Using concepts from graph theory [26] and
dynamic systems theory [44], the organization, dynamics, and plasticity of these
networks can then be studied.
Systems biology models of molecular reaction networks contain a number of
parameters. These are the rate constants of the involved reactions and, if spatiotemporal processes are considered, the transport rates, e.g. diffusion constants, of the
chemical species. In order for the models to be predictive, these parameters need to
be inferred. The process of inferring them from experimental data is called parameter identification. If in addition also the network structure is to be inferred from
data, the problem is called systems identification. Here, we consider the problem of
identifying the parameters of a biochemical reaction network from a single, noisy
measurement of the concentration time-course of some of the involved species.
While this time series can be long, ensemble replicas are not possible, either because the measurements are destructive or one is interested in variations between
different specimens or cells. This is particularly important in molecular systems biology, where cell–cell variations are of interest or large numbers of experimental
replica are otherwise not feasible [45].
This problem is particularly challenging and traditional genomic and proteomic
techniques do not provide single-cell resolution. Moreover, in individual cells the
molecules and chemical reactions can only be observed indirectly. Frequently, fluorescence microscopy is used to observe biochemical processes in single cells. Fluorescently tagging some of the species in the network of interest allows measuring the
spatiotemporal evolution of their concentrations from video microscopy and fluorescence photometry. In addition, fluorescence correlation spectroscopy (FCS) allows
measuring fluctuation time-courses of molecule numbers [23].
Using only a single trajectory of the mean concentrations would hardly allow
identification of network parameters. There could be several combinations of network parameters that lead to the same mean dynamics. A stochastic network model,
however, additionally provides information about the fluctuations of the molecular
abundances. The hope is that there is then only a small region of parameter space that
produces the correct behavior of the mean and the correct spectrum of fluctuations
[31]. Experimentally, fluctuation spectra can be measured at single-cell resolution
using FCS.
The stochastic behavior of biochemical reaction networks can be due to low copy
numbers of the reacting molecules [39, 10]. In addition, biochemical networks may
exhibit stochasticity due to extrinsic noise. This can persist even at the continuum
scale, leading to continuous–stochastic models. Extrinsic noise can, e.g., arise from
environmental variations or variations in how the reactants are delivered into the system. Also measurement uncertainties can be accounted for in the model as extrinsic
noise, modeling our inability to precisely quantify the experimental observables.
We model stochastic chemical kinetics using the chemical master equation
(CME). Using a CME forward model in biological parameter identification amounts
to tracking the evolution of a probability distribution, rather than just a single
function. This prohibits predicting the state of the system and only allows statements about the probability for the system to be in a certain state, hence requiring
Global stochastic parameter identification
3
sampling-based parameter identification methods. In the stochastic–discrete context, a number of different approaches have been suggested. Boys et al. proposed
a fully Bayesian approach for parameter estimation using an explicit likelihood for
data/model comparison and a Markov Chain Monte Carlo (MCMC) scheme for
sampling [5]. Zechner et al. developed a recursive Bayesian estimation technique
[46, 45] to cope with cell–cell variability in experimental ensembles. Toni and coworkers used an approximate Bayesian computation (ABC) ansatz, as introduced
by Marjoram and co-workers [25], that does not require an explicit likelihood [43].
Instead, sampling is done in a sequential Monte Carlo (or particle filter) framework.
Reinker et al. used a hidden Markov model where the hidden states are the actual
molecule abundances and state transitions model chemical reactions [40]. Inspired
by Prediction Error Methods [24], Cinquemani et al. identified the parameters of a
hybrid deterministic–stochastic model of gene expression from multiple experimental time courses [7]. Randomized optimization algorithms have been used, e.g., by
Koutroumpas et al. who applied a Genetic Algorithm to a hybrid deterministic–
stochastic network model [21]. More recently, Poovathingal and Gunawan used
another global optimization heuristic, the Differential Evolution algorithm [32]. A
variational approach for stochastic two-state systems is proposed by Stock and coworkers based on Maximum Caliber [41], an extension of Jaynes’ Maximum Entropy principle [14] to non-equilibrium systems. If estimates are to be made based
on a single trajectory, the stochasticity of the measurements and of the model leads
to very noisy similarity measures, requiring optimization and sampling schemes that
are robust against noise in the data.
Here, we propose a novel combination of exact stochastic simulations for a
CME forward model and an adaptive Monte Carlo sampling technique, called
Gaussian Adaptation, to address the single-trajectory parameter estimation problem
for monostable stochastic biochemical reaction networks. Evaluations of the CME
model are done using exact partial-propensity stochastic simulation algorithms [35].
Parameter optimization uses Gaussian Adaptation. The method iteratively samples
model parameters from a multivariate normal distribution and evaluates a suitable
objective function that measures the distance between the dynamics of the forward
model output and the experimental measurements. In addition to estimates of the
kinetic parameters in the network, the present method also provides an ellipsoidal
volume estimate of the viable part of parameter space and is able to estimate the
physical volume of the intra-cellular compartment in which the reactions take place.
We assume that quantitative experimental time series of either a transient or the
steady state of the concentrations of some of the molecular species in the network
are available. This can, for example, be obtained from single-cell fluorescence microscopy by translating fluorescence intensities to estimated chemical concentrations. Accurate methods that account for the microscope’s point-spread function
and the camera noise model are available to this end [12, 13, 6]. Additionally, FCS
spectra can be analyzed in order to quantify molecule populations, their intrinsic
fluctuations, and lifetimes [23, 34, 39]. The present approach requires only a single
stochastic trajectory from each cell. Since the forward model is stochastic and only a
single experimental trajectory is used, the objective function needs to robustly mea-
4
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
sure closeness between the experimental and the simulated trajectories. We review
previously considered measures and present a new distance function in Sec. 5. First,
however, we set out the formal stochastic framework and problem description below.
We then describe Gaussian Adaptation and its applicability to the current estimation
task. The evaluation of the forward model is outlined in Sec. 4. We consider a cyclic
linear chain as well as a non-linear colloidal aggregation model as benchmark test
cases in Sec. 6 and conclude in Sec. 7.
2 Background and problem statement
We consider a network model of a biochemical system given by M coupled chemical
reactions
N
kj
N
∑ νi,−j Si −−−−−→ ∑ νi,+j Si
i=1
∀ j = 1, . . . , M
(1)
i=1
between N species, where ν − = [νi,−j ] and ν + = [νi,+j ] are the stoichiometry matrices
of the reactants and products, respectively, and Si is the ith species in the reaction
network. Let ni be the population (molecular copy number) of species Si . The reactions occur in a physical volume Ω and the macroscopic reaction rate of reaction j
is k j . This defines a dynamic system with state n(t) = [ni (t)] and M + 1 parameters
θ = [k1 , . . . , kM , Ω ].
The state of such a system can be interpreted as a realization of a random variable
n(t) that changes over time t. All one can know about the system is the probability
for it to be in a certain state at a certain time t j given the system’s state history, hence
P(n(t j ) | n(t j−1 ), . . . , n(t1 ), n(t0 )) dNn
= Prob{n(t j ) ∈ [n(t j ), n(t j ) + dn) | n(ti ), i = 0, . . . , j − 1 } .
(2)
A frequently made model assumption, substantiated by physical reasoning, is
that the probability of the current state depends solely on the previous state, i.e.,
P(n(t j ) | n(t j−1 ), . . . , n(t1 ), n(t0 )) = P(n(t j ) | n(t j−1 )) .
(3)
The system is then modeled as a first-order Markov chain where the state n evolves
as
n(t + ∆t) = n(t) + Ξ (∆t; n,t)
(4)
This is the equation of motion of the system. If n is real-valued, it defines a
continuous–stochastic model in the form of a continuous-state Markov chain. Discrete n, as is the case in chemical kinetics, amount to discrete–stochastic models expressed as discrete-state Markov chains. The Markov propagator Ξ is itself a random variable, distributed with probability distribution Π (ξ | ∆t; n,t) =
P(n + ξ ,t + ∆t | n,t) for the state change ξ . For continuous-state Markov chains,
Π is a continuous probability density function (PDF), for discrete-state Markov
Global stochastic parameter identification
5
chains a discrete probability distribution. If Π (ξ ) = δ (ξ − ξ 0 ), with δ the Dirac
delta distribution, then the system’s state evolution becomes deterministic with predictable discrete or continuous increments ξ 0 . Deterministic models can hence be
interpreted as a limit case of stochastic models [22].
In chemical kinetics, the probability distribution Π of the Markov propagator
is a linear combination of Poisson distributions with weights given by the reaction
stoichiometry. This leads to the equation of motion for the population n given by
ψ1
n(t + ∆t) = n(t) + (ν + − ν − ) . . . ,
(5)
ψM
where ψi ∼ P(ai (n(t))∆t) is a random variable from the Poisson distribution with
rate λ = ai (n(t))∆t. The second term on the right-hand side of Eq. 5 follows a
probability distribution Π (ξ | ∆t; n,t) whose explicit form is analytically intractable
in the general case. The rates a j , j = 1, . . . M, are called the reaction propensities
and are defined as
N
kj
ni
aj = ∏
.
(6)
−
ν
1+
ν−
∑N
i, j
i0 =1 i0 , j
i=1
Ω
They depend on the macroscopic reaction rates and the reaction volume and can be
interpreted as the probability rates of the respective reactions. Advancing Eq. 5 with
a ∆t such that more than one reaction event happens per time step yields an approximate simulation of the biochemical network as done in approximate stochastic
simulation algorithms [9, 3].
An alternative approach consists in considering the evolution of the state probability distribution P(n,t | n0 ,t0 ) of the Markov chain described by Eq. 5, hence:
!
M
N ν − −ν +
∂P
i, j
i, j
(7)
= ∑ ∏ Ei E j − 1 a j (n(t))P(n,t)
∂t
j=1 i=1
with the step operator Eip f (n) = f (n + pî ) for any function f where î is the Ndimensional unit vector along the ith dimension. This equation is called the chemical
master equation (CME). Directly solving it for P is analytically intractable, but
trajectories of the Markov chain governed by the unknown state probability P can
be sampled using exact stochastic simulation algorithms (SSA) [8]. Exact SSAs
are exact in the sense that they sample Markov chain realizations from the exact
solution P of the CME, without ever explicitly computing this solution. Since SSAs
are Monte Carlo algorithms, however, a sampling error remains.
Assuming that the population n increases with the volume Ω , n can be approximated as a continuous random variable in the limit of large volumes, and Eq. 5
becomes
η1
n(t + ∆t) = n(t) + (ν + − ν − ) . . . ,
(8)
ηM
6
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
where ηi ∼ N (ai (n(t))∆t, ai (n(t))∆t) are normally distributed random variables.
The second term on the right-hand side of this equation is a random variable that is
distributed according to the corresponding Markov propagator Π (ξ | ∆t; n,t), which
is a Gaussian. Equation 8 is called the chemical Langevin equation with Π given by
1
T Σ −1 (ξ −µ)
Π (ξ | ∆t; n,t) = (2π)−N/2 |Σ |−1/2 e− 2 (ξ −µ)
,
(9)
where
a1 (n(t))
µ = ∆t (ν + − ν − ) . . . and Σ = ∆t (ν + − ν − ) diag (n(t))(ν + − ν − )T .
aM (n(t))
The corresponding equation for the evolution of the state PDF is the non-linear
Fokker-Planck equation, given by
∂P
1
T
=∇
D ∇ − F P(n,t) ,
(10)
∂t
2
where
∂
∂
,...,
∇ =
,
∂ n1
∂ nN
T
1
∆t→0 ∆t
Fi = lim
Z +∞
−∞
dξi ξi Π (ξ | ∆t; n,t) ,
(11)
(12)
and
1
Di j = lim
∆t→0 ∆t
Z +∞ Z +∞
−∞
−∞
dξi dξ j ξi ξ j Π (ξ | ∆t; n,t) .
(13)
At much larger Ω , when the population n is on the order of Avogadro’s number,
Eq. 8 can be further approximated as
φ1 (n(t))∆t
+
−
,
...
n(t + ∆t) = n(t) + (ν − ν )
(14)
φM (n(t))∆t
1−∑N
ν−
ν−
i, j
i0 =1 i0, j
where φ j (n) = k j Ω
∏Ni=1 ni (νi,−j !)−1 . Note that the second term on the
right-hand side of this equation is a random variable whose probability distribution
is the Dirac delta
φ1 (n(t))∆t
.
...
Π (ξ | ∆t; n,t) = δ ξ − (ν + − ν − )
(15)
φM (n(t))∆t
Global stochastic parameter identification
7
Equation 14 hence is a deterministic equation of motion. In the limit ∆t → 0 this
equation can be written as the ordinary differential equation
φ1 (x(t))
dx
= (ν + − ν − ) . . .
(16)
dt
φM (x(t))
for the concentration x = nΩ −1 . This is the classical reaction rate equation for the
system in Eq. 1.
By choosing the appropriate probability distribution Π of the Markov propagator,
one can model reaction networks in different regimes: small population n (small Ω )
using SSA over Eq. 7, intermediate population (intermediate Ω ) using Eq. 8, and
large population (large Ω ) using Eq. 16. The complete model definition therefore is
M (θ ) = {ν − , ν + , Π }.
The problem considered here can then be formalized as follows: Given a forward
model M (θ ) and a single noisy trajectory of the population of the chemical species
n̂(t0 + (q − 1)∆texp ) at K discrete time points t = t0 + (q − 1)∆texp , q = 1, . . . , K, we
wish to infer θ = [k1 , . . . , kM , Ω ]. The time between two consecutive measurements
∆texp and the number of measurements K are given by the experimental technique
used. As a forward model we use the full CME as given in Eq. 7 and sample trajectories from it using the partial-propensity formulation of Gillespie’s exact SSA as
described in Sec. 4.
3 Gaussian Adaptation for global parameter optimization,
approximate Bayesian computation, and volume estimation
Gaussian Adaptation (GaA), introduced in the late 1960’s by Gregor Kjellström
[17, 19], is a Monte Carlo technique that has originally been developed to solve
design-centering and optimization problems in analog electric circuit design. Design centering solves the problem of determining the nominal values (resistances,
capacitances, etc.) of the components of a circuit such that the circuit output is
within specified design bounds and is maximally robust against random variations
in the circuit components with respect to a suitable criterion or objective function.
This problem is a superset of general optimization, where one is interested in finding a parameter vector that minimizes (or maximizes) the objective function without
any additional robustness criterion. GaA has been specifically designed for scenarios where the objective function f (θ ) is only available in a black-box (or oracle)
model that is defined on a real-valued domain A ⊆ Rn and returns real-valued output R. The black-box model assumes that gradients or higher-order derivatives of
the objective function may not exist or may not be available, hence including the
class of discontinuous and noisy functions. The specific black-box function used
here is presented in Sec. 5.
8
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
The principle idea behind GaA is the following: Starting from a user-defined
point in parameter space, GaA explores the space by iteratively sampling single
parameter vectors from a multivariate Gaussian distribution N (m, Σ ) whose mean
m ∈ Rn and covariance matrix Σ ∈ Rn×n are dynamically adapted based on the information from previously accepted samples. The acceptance criterion depends on the
specific mode of operation, i.e., whether GaA is used as an optimizer or as a sampler [28, 27]. Adaptation is performed such as to maximize the entropy of the search
distribution under the constraint that acceptable search points are found with a predefined, fixed hitting (success) probability p < 1 [19]. Using the
qdefinition of the
(2πe)n det(Σ )
entropy of a multivariate Gaussian distribution H (N ) = log
shows that this is equivalent to maximizing the determinant of the covariance matrix Σ . GaA thus follows Jaynes’ Maximum Entropy principle [14].
GaA starts by setting the mean m(0) of the multivariate Gaussian to an initial
acceptable point θ (0) and the Cholesky factor Q(0) of the covariance matrix to the
identity matrix I. At each iteration g > 0, the covariance Σ (g) is decomposed as:
T
T
Σ (g) = r · Q(g) r · Q(g) = r2 Q(g) Q(g) , where r is the scalar step size
that controls the scale of the search. The matrix Q(g) is the normalized square root
of Σ (g) , found by eigen- or Cholesky decomposition of Σ (g) . The candidate parameter vector in iteration g + 1 is sampled from a multivariate Gaussian according to
θ (g+1) = m(g) + r(g) Q(g) η (g) , where η (g) ∼ N (0, I). The parameter vector is then
evaluated by the objective function f (θ (g+1) ).
Only if the parameter vector is accepted, the following adaptation rules are applied: The step size r is increased as r(g+1) = fe · r(g) , where fe > 1 is termed the
expansion factor. The mean of the proposal distribution is updated as
1
1 (g+1)
(g+1)
m
m(g) +
= 1−
θ
.
(17)
Nm
Nm
Nm is a weighting factor that
rate of the method. The successful
controls the learning
(g+1)
(g+1)
(g)
= θ
−m
is used to perform a rank-one update of
search direction d
the covariance matrix: Σ (g+1) = 1 − N1C Σ (g) + N1C d (g+1) d (g+1) T . NC weights the
influence of the accepted parameter vector on the covariance matrix. In order to
decouple the volume of the covariance (controlled by r(g+1) ) from its orientation,
Q(g+1) is normalized such that det(Q(g+1) ) = 1.
In case θ (g+1) is not accepted at the current iteration, only the step size is adapted
as r(g+1) = fc · r(g) , where fc < 1 is the contraction factor.
The behavior of GaA is controlled by several strategy parameters. Kjellström
analyzed the information-theoretic optimality of the acceptance probability p for
GaA in general regions [19]. He concluded that the efficiency E of the process and
p are related as E ∝ −p log p, leading to an optimal p = 1e ≈ 0.3679, where e is
Euler’s number. A proof is provided in [18]. Maintaining this optimal hitting prob-
Global stochastic parameter identification
9
ability corresponds to leaving the volume of the distribution, measured by det(Σ ),
constant under stationary conditions. Since det(Σ ) = r2n det(Q QT ), the expansion
and contraction factors fe and fc expand or contract the volume by a factor of fe2n
and fc2n , respectively. After S accepted and F rejected samples, a necessary condiS
, and
tion for constant volume thus is: ∏Si=1 ( fe )2n ∏Fi=1 ( fc )2n = 1. Using p = S+F
introducing a small β > 0, the choice fe = 1 + β (1 − p) and fc = 1 − β p satisfies
the constant-volume condition to first order. The scalar rate β is coupled to NC . NC
influences the update of Σ ∈ Rn×n , which contains n2 entries. Hence, NC should be
related to n2 . We suggested using NC = (n + 1)2 / log(n + 1) as a standard value, and
coupling β = N1C [29]. A similar reasoning is also applied to Nm . Since Nm influences the update of m ∈ Rn , it is reasonable to set Nm ∝ n. We propose Nm = en as
a standard value.
Depending on the specific acceptance rule used, GaA can be turned into a global
optimizer [29], an adaptive MCMC sampler [28, 27], or a volume estimation method
[30], as described next.
3.1 GaA for global black-box optimization
In a minimization scenario, GaA uses an adaptive-threshold acceptance mechanism.
(0)
Given an initial scalar cutoff threshold cT , we accept a parameter vector θ (g+1) at
(g)
iteration
g + 1 if f (θ (g+1) ) < cT . Upon acceptance, the threshold cT is lowered as
(g+1)
(g)
cT
= 1 − N1T cT + N1T f (θ (g+1) ), where NT controls the weighting between the
old threshold and the objective-function value of the accepted sample. This sampledependent threshold update renders the algorithm invariant to linear transformations
of the objective function. The standard strategy parameter value is NT = en [28].
We refer to [28] for further information about convergence criteria and constraint
handling techniques in GaA.
3.2 GaA for approximate Bayesian computation and viable volume
estimation
Replacing the threshold acceptance-criterion by a probabilistic Metropolis criterion,
and setting Nm = 1, turns GaA into an adaptive MCMC sampler with global adaptive scaling [2]. We termed this method Metropolis-GaA [28, 27]. Its strength is that
GaA can automatically adapt to the covariance of the target probability distribution
while maintaining the fixed hitting probability. For standard MCMC, this cannot be
achieved without fine-tuning the proposal using multiple MCMC runs. We hypothesize that GaA might also be an effective tool for approximate Bayesian computation
(ABC) [43]. In essence, the ABC ansatz is MCMC without an explicit likelihood
10
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
function [25]. The likelihood is replaced by a distance function — which plays the
same role as our objective function — that measures closeness between a parameterized model simulation and empirical data D, or summary statistics thereof. When a
uniform prior over the parameters and a symmetric proposal are assumed, a parameter vector in ABC is unconditionally accepted if its corresponding distance function
value f (θ (g+1) ) < cT [25]. The threshold cT is a problem-dependent constant that
is fixed prior to the actual computation. Marjoram and co-workers have shown that
samples obtained in this manner are approximately drawn from the posterior parameter distribution given the data D. While Pritchard et al. used a simple rejection
sampler [33], Marjoram and co-workers proposed a standard MCMC scheme [25].
Toni and co-workers used sequential MC for sample generation [43]. To the best of
our knowledge, however, the present work presents the first application of an adaptive MCMC scheme for ABC in biochemical network parameter inference. Finally,
we emphasize that when GaA’s mean, covariance matrix, and hitting probability p
stabilize during ABC, they provide direct access to an ellipsoidal estimation of the
volume of the viable parameter space as defined by the threshold cT [30]. Hafner
and co-workers have shown how to use such viable volume estimates for model
discrimination [11].
4 Evaluation of the forward model
In each iteration of the GaA algorithm, the forward model of the network needs
to be evaluated for the proposed parameter vector θ . This requires an efficient and
exact SSA for the chemical kinetics of the reaction network, used to generate trajectories n(t) from M (θ ). Since GaA could well propose parameter vectors that lead
to low copy numbers for some species, it is important that the SSA be exact since
approximate algorithms are not appropriate at low copy number.
In its original formulation, Gillespie’s SSA has a computational cost that is
linearly proportional to the total number M of reactions in the network. If many
model evaluations are required, as in the present application, this computational cost
quickly becomes prohibitive. While more efficient formulations of SSA have been
developed for weakly coupled reaction networks, their computational cost remains
proportional to M for strongly coupled reaction networks [35]. A reaction network is
weakly coupled if the number of reactions that are influenced by any other reaction
is bounded by a constant. If a network contains at least one reaction whose firing
influences the propensities of a fixed proportion (in the worst case all) of the other
reactions, then the network is strongly coupled [35]. Scale-free networks as seem to
be characteristic for systems biology models [1, 42] are by definition strongly coupled. This is due to the existence of hubs that have a higher connection probability
than other nodes. These hubs frequently correspond to chemical reactions that produce or consume species that also participate in the majority of the other reactions,
such as water, ATP, or CO2 in metabolic networks.
Global stochastic parameter identification
11
We use partial-propensity methods [35, 36] to simulate trajectories according to
the solution of the chemical master Eq. 7 of the forward model. Partial-propensity
methods are exact SSAs whose computational cost scales at most linearly with the
number N of species in the network [35]. For large networks, this number is usually much smaller than the number of reactions. Depending on the network model
at hand, different partial-propensity methods are available for its efficient simulation. Strongly coupled networks where the rate constants span only a limited spectrum of values are best simulated with the partial-propensity direct method (PDM)
[35]. Multi-scale networks where the rate constants span many orders of magnitude are most efficiently simulated using the sorting partial-propensity direct method
(SPDM) [35]. Weakly coupled reaction networks can be simulated at constant computational cost using the partial-propensity SSA with composition-rejection sampling (PSSA-CR) [37]. Lastly, reaction networks that include time delays can be
exactly simulated using the delay partial-propensity direct method (dPDM) [38].
Different combinations of the algorithmic modules of partial-propensity methods
can be used to constitute all members of this family of SSAs [36]. We refer to the
original publications for algorithmic details, benchmarks of the computational cost,
and a proof of exactness of partial-propensity methods.
5 Objective function
In the context of parameter identification of stochastic biochemical networks, a
number of distance or objective functions have previously been suggested. Reinker
et al. proposed an approximate maximum-likelihood measure under the assumption
that only a small number of reactions fire between two experimental measurement
points, and a likelihood based on singular value decomposition that works when
many reactions occur per time interval [40]. Koutroumpas et al. compared objective functions based on least squares, normalized cross-correlations, and conditional
probabilities using a Genetic Algorithm [21]. Koeppl and co-workers proposed the
Kantorovich distance to compare experimental and model-based probability distributions [20]. Alternative distance measures include the Earth Mover’s distance or
the Kolomogorov-Smirnov distance [32]. These distance measures, however, can
only be used when many experimental trajectories are available. In order to measure the distance between a single experimental trajectory n̂(t) and a single model
output n(t), we propose a novel cost function f (θ ) = f (M (θ ), n̂) that reasonably
captures the kinetics of a monostable system. We define a compound objective function f (θ ) = f1 (θ ) + f2 (θ ) with
4
f1 (θ ) = ∑ γi ,
i=1
where
N
z
x
|ACFl (n̂i ) − ACFl (ni )|
∑l=0
,
zx
ACFl (n̂i )
∑l=0
i=1
f2 (θ ) = ∑
(18)
12
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
N
γi =
s
∑
j=1
µi (n j ) − µi (n̂ j )
µi (n̂ j )
2
(19)
with the central moments given by
(
if i = 1
∑Kp=1 n j (t0 + (p − 1)∆texp )
i
µi (n j ) =
(| ∑Kq=1 n j (t0 + (q − 1)∆texp ) − µ1 (n j ) |)1/i otherwise
(20)
and the time-autocorrelation function (ACF) at lag l given by
ACFl (ni ) =
ni (t0 )ni (t0 + l ∆texp ) − (µ1 (ni ))2
.
µ2 (ni )
The variable zx is the lag at which the experimental ACF crosses 0 for the first time.
The function f1 (θ ) measures the difference between the first four moments of n and
n̂. This function alone would, however, not be enough to capture the kinetics since
it lacks information about correlations in time. This is taken into account by f2 (θ ),
measuring the difference in the lifetimes of all chemical species. These lifetimes are
systematically modulated by the volume Ω [39], hence enabling volumetric measurements of intra-cellular reaction compartments along with the identification of
the rate constants.
The present objective function allows inclusion of experimental readouts from
image-based systems biology. The moment-matching part is a typical readout from
fluorescence photometry, whereas the autocorrelation of the fluctuations can directly
be measured using, e.g., FCS.
a
b
c
100
300
300
250
250
80
n̂1 200
n̂1
200
n̂1
60
n̂2
n̂2
150
n̂3
40
n̂2
150
100
20
0
0
100
50
20
40
60
time [s]
80
100
0
0
50
20
40
60
time [s]
80
100
0
0
20
40
60
80
100
time [s]
Fig. 1 In silico data for all test cases. a. Time evolution of the populations of three species in the
cyclic chain model at steady state (starting at t0 = 2000). b. Time evolution of the populations of
two species in the aggregation model at steady state (starting at t0 = 5000). c. Same as b, but during
the transient phase (starting at t0 = 0).
Global stochastic parameter identification
13
6 Results
We estimate the unknown parameters θ for two reaction networks: a weakly coupled
cyclic chain and a strongly coupled non-linear colloidal aggregation network. For
the cyclic chain we estimate θ at steady state. For the aggregation model we estimate
θ both at steady state and in the transient phase. Every kinetic parameter is allowed
to vary in the interval [10−3 , 103 ] and the reaction volume Ω in [1, 500]. Each GaA
run starts from a point selected uniformly at random in logarithmic parameter space.
6.1 Weakly coupled reaction network: cyclic chain
The cyclic chain network is given by:
k
i
Si −
→
Si+1
kN
Si −→ S1
i = 1, . . . , N − 1 ,
i=N.
(21)
In this linear network, the number of reactions M is equal to the number of species
N. The maximum degree of coupling of this reaction network is 2, irrespective of the
size of the system (length of the chain), rendering it weakly coupled [35]. We hence
use PSSA-CR to evaluate the forward model with a computational complexity in
O(1) [37]. In the present test case, we limit ourselves to 3 species and 3 reactions,
i.e., N = M = 3. The parameter vector for this case is given by θ = [k1 , k2 , k3 ], since
the kinetics of linear reactions is independent of the volume Ω [39].
We simulate steady-state “experimental” data n̂ using PSSA-CR with ground
truth k1 = 2, k2 = 1.5, k3 = 3.2 (see Fig. 1a). We set the initial population of the
species to n1 (t = 0) = 50, n2 (t = 0) = 50, and n3 (t = 0) = 50 and sample a single
CME trajectory at equi-spaced time points with ∆texp = 0.1 between t = t0 and
t = t0 + (K − 1)∆texp with t0 = 2000 and K = 1001 for each of the 3 species S1 , S2 ,
and S3 . For the generated data we find zx = 7.
We generate trajectories from the forward model for every parameter vector θ
proposed by GaA using PSSA-CR between t = 0 and t = (K − 1)∆texp = 100, starting from the initial population ni (t = 0) = n̂i (t = t0 ).
Before turning to the actual parameter identification, we illustrate the topography
of the objective function landscape for the present example. We fix k3 = 3.2 to its
optimal value and perform a two-dimensional grid sampling for k1 and k2 over the
full search domain. We use 40 logarithmically spaced sample points per parameter,
resulting in 402 parameter combinations. For each combination we evaluate the objective function. The resulting landscapes of f1 (θ ), f2 (θ ), and f (θ ) are depicted in
Fig. 2a. Figure 2b shows refined versions around the global optimum. We see that
the moment-matching term f1 (θ ) is largely responsible for the global single-funnel
topology of the landscape. The autocorrelation term f2 (θ ) sharpens the objective
function near the global optimum and renders it locally more isotropic.
14
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
We perform both global optimization and ABC runs using GaA. In each of the 15
independent optimization runs, the number of function evaluations (FES) is limited
to MAX FES= 1000M = 3000. We set the initial step size to r(0) = 1 and perform all
searches in logarithmic scale of the parameters. Independent restarts from uniformly
random points are performed when the step size r drops below 10−4 [29]. For each
of the 15 independent runs, the 30 parameter vectors with the smallest objective
function value are collected and displayed in the box plot shown in the left panel of
Fig. 3a. All 450 collected parameter vectors have objective function values smaller
than 1.6. The resulting data suggest that the present method is able to accurately
determine the correct scale of the kinetic parameters from a single experimental
trajectory, although an overestimation of the rates is apparent.
We use the obtained optimization results for subsequent ABC runs. We conduct
15 independent ABC runs using cT = 2. The starting points for the ABC runs are
selected uniformly at random from the 450 collected parameter vectors in order to
ensure stable initialization. For each run, we again set MAX FES= 1000M = 3000.
The initial step size r(0) is set to 0.1, and the parameters are again explored in logarithmic scale. For all runs we observe rapid convergence of the empirical hitting
a
3
3
3
2
8
ï1
6
4
ï2
1
2.5
0
ï1
2
ï2
1.5
log10 k2
1
log10 k2
log10 k2
0
16
3
12
10
18
2
14
1
20
3.5
16
2
14
12
0
10
ï1
8
6
ï2
4
2
ï3
ï3
ï2
ï1
0
log10 k1
1
2
ï3
ï3
3
ï2
ï1
0
log10 k1
1
2
ï3
ï3
3
ï2
ï1
0
log10 k1
1
2
3
2
b
1
1
1
4.5
1.9
0.8
2.5
0.4
2
1.5
0.2
0
0
0.2
0.4
0.6
log10 k1
0.8
1
1.7
0.6
log10 k2
log10 k2
3
5.5
5
1.8
3.5
0.6
6
0.8
1.6
1.5
0.4
4.5
0.6
log10 k2
4
0.8
2
4
3.5
0.4
1.4
1.3
0.2
1
1.2
0.5
1.1
0
0
0.2
0.4
0.6
log10 k1
0.8
1
3
0.2
2.5
2
0
0
0.2
0.4
0.6
0.8
1
log10 k1
Fig. 2 a. Global objective function landscape for the cyclic chain over the complete search domain
for optimal k3 = 3.2. The three panels from left to right show f1 (θ ), f2 (θ ), and f (θ ), respectively.
b. A refined view of the global objective function landscape near the global optimum. The three
panels from left to right show f1 (θ ), f2 (θ ), and f (θ ), respectively. The white dots mark the true,
optimal parameters.
Global stochastic parameter identification
15
probability pemp to the optimal p = 1e (see Sec. 3). We collect the ABC samples
along with the means and covariances of GaA as soon as |pemp − p| < 0.05. As an
example we show the histograms of the posterior samples for a randomly selected
run in Fig. 3b. The means of the posterior distributions are again larger than the true
kinetic parameters. Using GaA’s means, covariance matrices, and the corresponding hitting probabilities that generated the posterior samples, we can construct an
ellipsoidal volume estimation [30]. This is done by multiplying each eigenvalue of
the average of the collected covariance matrices with c pemp = inv χn2 (pemp ), the ndimensional inverse Chi-square distribution evaluated at the empirical hitting probability. The product of these scaled eigenvalues and the volume of the n-dimensional
n
unit sphere, |S(n)| = Γ (πn 2+1) , then yields the ellipsoid volume with respect to a uni2
form distribution (see [30] for details). The resulting ellipsoid contains the optimal
kinetic parameter vector and is depicted in the right panel of Fig. 3a. It has a volume
of 0.045 in log-parameter space. This constitutes only 0.0208% of the initial search
space volume, indicating that GaA significantly narrows down the viable parameter
space around the true optimal parameters.
a
0.7
0.6
log10 k3
log10 ki
0.5
0.4
0.3
0.2
0.1
b
log10 k2
1
2
i
3
log10 k1
90
90
90
80
80
80
70
70
70
60
60
60
50
50
50
40
40
40
30
30
30
20
20
20
10
10
0
0.1
0.2
0.3
0.4
log10 k1
0.5
0.6
0
0
10
0.1
0.2
0.3
log10 k2
0.4
0.5
0
0.3
0.4
0.5
0.6
0.7
0.8
log10 k3
Fig. 3 a. Left panel: Box plot of the 30 best parameter vectors from each of the 15 independent
optimization runs. The blue dots mark the true parameter values. Right panel: Ellipsoidal volume
estimation of the parameter space below an objective-function threshold cT = 2 from a single ABC
run. b. Empirical posterior distributions of the kinetic parameters from the same single ABC run
with cT = 2. The red lines indicate the true parameters.
16
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
6.2 Strongly coupled reaction network: colloidal aggregation
The colloidal aggregation network is given by:
kon
1
→
0/ −−
S1
ki j
Si + S j −→ Si+ j
k̄i j
Si+ j −→ Si + S j
kioff
Si −−→
0/
i + j = 1, . . . , N
i + j = 1, . . . , N
i = 1, . . . , N .
(22)
j 2k
For this network of N species, the number of reactions is M = N2 + N + 1. The
maximum degree of coupling of this reaction network is proportional to N, rendering the network strongly coupled [35]. We hence use SPDM to evaluate the
forward model with a computational complexity in O(N) [35]. We use SPDM instead of PDM since the search path of GaA is unpredictable and could well generate parameters that lead to multi-scale networks. For this test case, we limit ourselves to two species, i.e., N = 2 and M = 5. The parameter vector for this case is
θ = [k11 , k̄11 , k1on , k1off , k2off , Ω ].
We perform GaA global optimization runs following the same protocol as for the
cyclic chain network with MAX FES = 1000(M + 1) = 6000.
6.2.1 At steady state
We simulate “experimental” data n̂ using SPDM with ground truth k11 = 0.1, k̄11 =
1.0, k1on = 2.1, k1off = 0.01, k2off = 0.1, and Ω = 15 (see Fig. 1b). We set the initial
population of the species to n1 (t = 0) = 0, n2 (t = 0) = 0, and n3 (t = 0) = 0 and
sample K = 1001 equi-spaced data points between t = t0 and t = t0 + (K − 1)∆texp
with t0 = 5000 and ∆texp = 0.1.
We generate trajectories from the forward model for every parameter vector θ
proposed by GaA using SPDM between t = 0 and t = (K − 1)∆texp = 100, stating
from the initial population ni (t = 0) = n̂i (t = t0 ).
The optimization results are summarized in the left panel of Fig. 4a. For each
of 15 independent runs, the 30 lowest-objective parameters are collected and shown
in the box plot. We observe that the true parameters corresponding to θ2 = k̄11 ,
θ3 = k1on , θ4 = k1off , and θ5 = k2off are between the 25th and 75th percentiles of the
identified parameters. Both the first parameter and the reaction volume are, on average, overestimated. Upon rescaling the kinetic rate constants with the estimated
volume, we find θ norm = [θ1 /θ6 , θ2 , θ3 θ6 , θ4 , θ5 ], which are the specific probability
rates of the reactions. The identified values are shown in the right panel of Fig. 4a.
The median of the identified θ3norm coincides with the true specific probability rate.
Likewise, θ1norm is closer to the 25th percentile of the parameter distribution. This
Global stochastic parameter identification
17
suggests a better estimation performance of GaA in the space of specific probability
rates, at the expense of not obtaining an estimate of the reaction volume.
6.2.2 In the transient phase
We simulate “experimental” data in the transient phase of the network dynamics
using the same parameters as above between t = t0 and t = (K − 1)∆texp with t0 =
0, ∆texp = 0.1, and K = 1001 (see Fig. 1c). We evaluate the forward model with
ni (t = 0) = n̂i (t = t0 ) to obtain trajectories between t = 0 and t = (K − 1)∆texp for
every proposed parameter vector θ .
The optimization results for the transient case are summarized in Fig. 4b. We
observe that the true parameters corresponding to θ3 = k1on , θ5 = k2off , and θ6 = Ω
are between the 25th and 75th percentiles of the identified parameters. The remaining
parameters are, on average, overestimated. In the space of rescaled parameters θ norm
we do not observe a significant improvement of the estimation.
3
3
2
2
log10 θinorm
a
log10 θi
1
0
1
0
ï1
ï1
ï2
ï2
ï3
ï3
1
2
3
4
5
6
1
2
3
i
4
5
1
2
3
i
4
5
i
3
3
2
2
log10 θinorm
b
log10 θi
1
0
1
0
ï1
ï1
ï2
ï2
ï3
ï3
1
2
3
4
i
5
6
Fig. 4 a. Left panel: Box plot of the 30 best parameter vectors from each of the 15 independent
optimization runs for the steady-state data set. Right panel: Box plots of the normalized parameters
(see main text for details). b. Left panel: Box plot of the 30 best parameter vectors from each of the
15 independent optimization runs for the transient data set. Right panel: Box plot of the normalized
parameters (see main text for details). The blue dots indicate the true parameter values.
18
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
7 Conclusions and Discussion
We have considered parameter estimation of stochastic biochemical networks from
single experimental trajectories. Parameter identification from single time series is
desirable in image-based systems biology, where per-cell estimates of the fluorescence evolution and its fluctuations are available. This enables quantifying cell–cell
variability on the level of network parameters. The histogram of the parameters
identified for different cells provides a biologically meaningful way of assessing
phenotypic variability beyond simple differences in the fluorescence levels.
We have proposed a novel combination of a flexible Monte Carlo method, the
Gaussian Adaptation (GaA) algorithm, and efficient exact stochastic simulation algorithms, the partial-propensity methods. The presented method can be used for
global parameter optimization, approximate Bayesian inference under a uniform
prior, and volume estimation of the viable parameter space. We have introduced an
objective function that measures closeness between a single experimental trajectory
and a single trajectory generated by the forward model. The objective function comprises a moment-matching and a time-autocorrelation part. This allows including
experimental readouts from, e.g., fluorescence photometry and fluorescence correlation spectroscopy.
We have applied the method to estimate the parameters of two monostable reaction networks from a single simulated temporal trajectory each, both at steady
state and during transient phases. We considered the linear cyclic chain network and
a non-linear colloidal aggregation network. For the linear model we were able to
robustly identify a small region of parameter space containing the true kinetic parameters. In the non-linear aggregation model, we could identify several parameter
vectors that fit the simulated experimental data well. There are two possible reasons
for this reduced parameter identifiability: either GaA cannot find the globally optimal region of parameter space due to high ruggedness and noise in the objective
function, or the non-linearity of the aggregation network modulates the kinetics in
a non-trivial way [39, 10]. Both cases are not accounted for in the current objective
function, thus leading to reduced performance for non-linear reaction networks.
We also used GaA as an adaptive MCMC method for approximate Bayesian inference of the posterior parameter distributions in the linear chain network. This enabled estimating the volume of the viable parameter space below a given objectivefunction value threshold. We found these volume estimates to be stable across independent runs. We thus believe that GaA might be a useful tool for exploring the
parameter spaces of stochastic systems.
Future work will include (i) alternative objective functions that include temporal cross-correlations between species and the derivative of the autocorrelation; (ii)
longer experimental trajectories; (iii) multi-stable and oscillatory systems; and (iv)
alternative global optimization schemes. Moreover, the applicability of the present
method to large-scale, non-linear biochemical networks and real-world experimental data will be tested in future work.
Global stochastic parameter identification
19
Acknowledgements RR was financed by a grant from the Swiss SystemsX.ch initiative (grant
WingX), evaluated by the Swiss National Science Foundation. This project was also supported
with a grant from the Swiss SystemsX.ch initiative, grant LipidX-2008/011, to IFS.
References
1. Albert, R.: Scale-free networks in cell biology. J. Cell Sci. 118(21), 4947–4957 (2005)
2. Andrieu, C., Thoms, J.: A tutorial on adaptive MCMC. Statistics and Computing 18(4), 343–
373 (2008)
3. Auger, A., Chatelain, P., Koumoutsakos, P.: R-leaping: Accelerating the stochastic simulation
algorithm by reaction leaps. J. Chem. Phys. 125, 084,103 (2006)
4. Barabási, A.L., Oltvai, Z.N.: Network biology: understanding the cell’s functional organization. Nat. Rev. Genet. 5(2), 101–113 (2004)
5. Boys, R.J., Wilkinson, D.J., Kirkwood, T.B.L.: Bayesian inference for a discretely observed
stochastic kinetic model. Statistics and Computing 18(2), 125–135 (2008)
6. Cardinale, J., Rauch, A., Barral, Y., Székely, G., Sbalzarini, I.F.: Bayesian image analysis with
on-line confidence estimates and its application to microtubule tracking. In: Proc. IEEE Intl.
Symposium Biomedical Imaging (ISBI), pp. 1091–1094. IEEE, Boston, USA (2009)
7. Cinquemani, E., Milias-Argeitis, A., Summers, S., Lygeros, J.: Stochastic dynamics of genetic
networks: modelling and parameter identification. Bioinformatics 24(23), 2748–2754 (2008)
8. Gillespie, D.T.: A rigorous derivation of the chemical master equation. Physica A 188, 404–
425 (1992)
9. Gillespie, D.T.: Approximate accelerated stochastic simulation of chemically reacting systems. J. Chem. Phys. 115(4), 1716–1733 (2001)
10. Grima, R.: Noise-induced breakdown of the Michaelis-Menten equation in steady-state conditions. Phys. Rev. Lett. 102(21), 218103 (2009). DOI 10.1103/PhysRevLett.102.218103. URL
http://link.aps.org/abstract/PRL/v102/e218103
11. Hafner, M., Koeppl, H., Hasler, M., Wagner, A.: ‘Glocal’ robustness analysis and model discrimination for circadian oscillators. PLoS Comput. Biol. 5(10), e1000,534 (2009)
12. Helmuth, J.A., Burckhardt, C.J., Greber, U.F., Sbalzarini, I.F.: Shape reconstruction of subcellular structures from live cell fluorescence microscopy images. J. Struct. Biol. 167, 1–10
(2009)
13. Helmuth, J.A., Sbalzarini, I.F.: Deconvolving active contours for fluorescence microscopy images. In: Proc. Intl. Symp. Visual Computing (ISVC), Lecture Notes in Computer Science,
vol. 5875, pp. 544–553. Springer, Las Vegas, USA (2009)
14. Jaynes, E.T.: Information theory and statistical mechanics. Phys. Rev. 106(4), 620–630 (1957).
DOI 10.1103/PhysRev.106.620
15. Kitano, H.: Computational systems biology. Nature 420(6912), 206–210 (2002)
16. Kitano, H.: Systems biology: A brief overview. Science 295(5560), 1662–1664 (2002)
17. Kjellström, G.: Network optimization by random variation of component values. Ericsson
Technics 25(3), 133–151 (1969)
18. Kjellström, G.: On the efficiency of Gaussian Adaptation. J. Optim. Theory Appl. 71(3),
589–597 (1991)
19. Kjellström, G., Taxen, L.: Stochastic optimization in system design. IEEE Trans. Circ. and
Syst. 28(7), 702–715 (1981)
20. Koeppl, H., Setti, G., Pelet, S., Mangia, M., Petrov, T., Peter, M.: Probability metrics to calibrate stochastic chemical kinetics. In: Proc. IEEE Intl. Symp. Circuits and Systems, pp.
541–544. Paris, France (2010)
21. Koutroumpas, K., Cinquemani, E., Kouretas, P., Lygeros, J.: Parameter identification for
stochastic hybrid systems using randomized optimization: A case study on subtilin production by Bacillus subtilis. Nonlinear Anal.: Hybrid Syst. 2(3), 786–802 (2008)
20
C. L. Müller, R. Ramaswamy, and I. F. Sbalzarini
22. Kurtz, T.G.: Relationship between stochastic and deterministic models for chemical reactions.
J. Chem. Phys. 57(7), 2976–2978 (1972)
23. Lakowicz, J.R.: Principles of Fluorescence Spectroscopy. Springer US (2006). DOI
10.1007/978-0-387-46312-4
24. Ljung, L.: Prediction error estimation methods. Circuits, systems, and signal processing 21(1),
11–21 (2002)
25. Marjoram, P., Molitor, J., Plagnol, V., Tavare, S.: Markov chain Monte Carlo without likelihoods. Proc. Natl. Acad. Sci. USA 100(26), 15,324–15,328 (2003)
26. Mason, O., Verwoerd, M.: Graph theory and networks in biology. Systems Biology, IET 1(2),
89–119 (2007)
27. Müller, C.L.: Exploring the common concepts of adaptive MCMC and Covariance Matrix Adaptation schemes. In: A. Auger, J.L. Shapiro, D. Whitley, C. Witt (eds.) Theory of Evolutionary Algorithms, no. 10361 in Dagstuhl Seminar Proceedings. Schloss
Dagstuhl – Leibniz-Zentrum für Informatik, Germany, Dagstuhl, Germany (2010). URL
http://drops.dagstuhl.de/opus/volltexte/2010/2813
28. Müller, C.L., Sbalzarini, I.F.: Gaussian Adaptation as a unifying framework for continuous
black-box optimization and adaptive Monte Carlo sampling. In: Proc. IEEE Congress on
Evolutionary Computation (CEC), pp. 2594–2601. Barcelona, Spain (2010)
29. Müller, C.L., Sbalzarini, I.F.: Gaussian Adaptation revisited — an entropic view on covariance matrix adaptation. In: Proc. EvoStar, Lect. Notes Comput. Sci., vol. 6024, pp. 432–441.
Springer, Istanbul, Turkey (2010)
30. Müller, C.L., Sbalzarini, I.F.: Gaussian Adaptation for robust design centering. In: Proc. EuroGen, Intl. Conf. Evolutionary and Deterministic Methods for Design, Optimization and Control. Capua, Italy (2011)
31. Munsky, B., Trinh, B., Khammash, M.: Listening to the noise: random fluctuations reveal gene
network parameters. Mol. Sys. Biol. 5(1), 318 (2009)
32. Poovathingal, S.K., Gunawan, R.: Global parameter estimation methods for stochastic biochemical systems. BMC Bioinformatics 11(1), 414 (2010)
33. Pritchard, J.K., Seielstad, M.T., Perez-Lezaun, A., Feldman, M.W.: Population growth of human Y chromosomes: A study of Y chromosome microsatellites. Molecular Biology and
Evolution 16(12), 1791–1798 (1999)
34. Qian, H., Elson, E.L.: Fluorescence correlation spectroscopy with high-order and dual-color
correlation to probe nonequilibrium steady states. Proc. Natl. Acad. Sci. USA 101(9), 2828–
2833 (2004)
35. Ramaswamy, R., González-Segredo, N., Sbalzarini, I.F.: A new class of highly efficient exact
stochastic simulation algorithms for chemical reaction networks. J. Chem. Phys. 130(24),
244,104 (2009)
36. Ramaswamy, R., Sbalzarini, I.F.: Fast exact stochastic simulation algorithms using partial
propensities. In: Proc. ICNAAM, Numerical Analysis and Applied Mathematics, International Conference, pp. 1338–1341. AIP (2010)
37. Ramaswamy, R., Sbalzarini, I.F.: A partial-propensity variant of the composition-rejection
stochastic simulation algorithm for chemical reaction networks. J. Chem. Phys. 132(4),
044,102 (2010)
38. Ramaswamy, R., Sbalzarini, I.F.: A partial-propensity formulation of the stochastic simulation
algorithm for chemical reaction networks with delays. J. Chem. Phys. 134, 014,106 (2011)
39. Ramaswamy, R., Sbalzarini, I.F., González-Segredo, N.: Noise-induced modulation of the relaxation kinetics around a non-equilibrium steady state of non-linear chemical reaction networks. PLoS ONE 6(1), e16,045 (2011)
40. Reinker, S., Altman, R.M., Timmer, J.: Parameter estimation in stochastic biochemical reactions. IEE Proc.-Syst. Biol 153(4), 168 (2006)
41. Stock, G., Ghosh, K., Dill, K.A.: Maximum Caliber: A variational approach applied to twostate dynamics. J. Chem. Phys. 128(19), 194,102 (2008)
42. Strogatz, S.H.: Exploring complex networks. Nature 410, 268–276 (2001)
Global stochastic parameter identification
21
43. Toni, T., Welch, D., Strelkowa, N., Ipsen, A., Stumpf, M.P.H.: Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems. J. R. Soc.
Interface 6(31), 187–202 (2009)
44. Wolkenhauer, O.: Systems biology: The reincarnation of systems theory applied in biology?
Briefings in Bioinformatics 2(3), 258 (2001)
45. Zechner, C., Ganguly, A., Pelet, S., Peter, M., Koeppl, H.: Accounting for extrinsic variability in the estimation of stochastic rate constants. Intl. J. Robust Nonlinear Control (2011,
submitted)
46. Zechner, C., Pelet, S., Peter, M., Koeppl, H.: Recursive Bayesian estimation of stochastic rate
constants from heterogeneous cell populations. In: Proc. IEEE CDC, Conference on Decision
and Control (2011, submitted)
| 5cs.CE
|
Population stability: regulating size in the presence of an adversary
arXiv:1803.02540v1 [cs.DC] 7 Mar 2018
Shafi Goldwasser∗
Rafail Ostrovsky†
Alessandra Scafuro‡
Adam Sealfon
§
Abstract
We introduce a new coordination problem in distributed computing that we call the population stability problem. A system of agents each with limited memory and communication, as
well as the ability to replicate and self-destruct, is subjected to attacks by a worst-case adversary
that can at a bounded rate (1) delete agents chosen arbitrarily and (2) insert additional agents
with arbitrary initial state into the system. The goal is perpetually to maintain a population
whose size is within a constant factor of the target size N . The problem is inspired by the ability of complex biological systems composed of a multitude of memory-limited individual cells
to maintain a stable population size in an adverse environment. Such biological mechanisms
allow organisms to heal after trauma or to recover from excessive cell proliferation caused by
inflammation, disease, or normal development.
We present a population stability protocol in a communication model that is a synchronous
variant of the population model of Angluin et al. In each round, pairs of agents selected at
random meet and exchange messages, where at least a constant fraction of agents is matched
in each round. Our protocol uses three-bit messages and ω(log2 N ) states per agent. We
emphasize that our protocol can handle an adversary that can both insert and delete agents, a
setting in which existing approximate counting techniques do not seem to apply. The protocol
relies on a novel coloring strategy in which the population size is encoded in the variance of the
distribution of colors. Individual agents can locally obtain a weak estimate of the population
size by sampling from the distribution, and make individual decisions that robustly maintain a
stable global population size.
∗
MIT. Department of Computer Science. [email protected]. Supported by NSF MACS CNS-1413920,
DARPA IBM W911NF-15-C-0236, and a Simons Investigator award agreement dated 6-5-12.
†
University of California, Los Angeles. Department of Computer Science and Mathematics. [email protected].
Research supported in part by NSF grant 1619348, DARPA, US-Israel BSF grant 2012366, OKAWA Foundation
Research Award, IBM Faculty Research Award, Xerox Faculty Research Award, B. John Garrick Foundation Award,
Teradata Research Award, and Lockheed-Martin Corporation Research Award. The views expressed are those of the
authors and do not reflect position of the Department of Defense or the U.S. Government.
‡
North Carolina State University. Department of Computer Science. [email protected].
§
MIT. Department of Computer Science. [email protected]. Supported by a DOE CSGF fellowship, NSF
MACS CNS-1413920, DARPA IBM W911NF-15-C-0236, and a Simons Investigator award agreement dated 6-5-12.
1
Introduction
A single fertilized mouse egg and human egg develop into organisms with vastly different numbers
of cells. How do cellular mechanisms regulate the number of cells in complex biological systems?
In order to function in adverse environments, organisms must maintain stability and be able to
recover from unplanned circumstances. For example, a lizard that loses its tail can grow a new one,
and internal organs require mechanisms to recover from cell loss caused by injury or from execessive
cell proliferation due to development or disease. But how do individual cells know how to respond
in order to reestablish the desired population size?
Regulation of population size may be achieved through a combination of internal programs
running within each cell and intercellular communication. One approach could be for individual cells
to count the population using a distributed protocol. But an interesting question is how to control
the population size if each cell lacks sufficient memory to count. Understanding the mechanisms
for regulating population size in an adversarial environment, in light of memory constraints of
individual agents, is a natural computational question.
In this work, we study the problem of robustly maintaining a stable population size from the
perspective of distributed computing. We introduce a new coordination question that we call the
population stability problem. Consider a population of agents with the ability to replicate and
self-destruct. How can such distributed systems detect and recover from adversarial deletions and
insertions of agents so as to maintain the desired population size?
Our focus is on systems that consist of huge numbers of agents, where each agent individually
has very limited memory and connective capability and can directly communicate with only a few
other agents in the system. We model communication using a synchronous variant of the population
model of Angluin et al. [AAD+ 06, AAE07]. In each round, a constant fraction of agents is matched
at random and can exchange messages, where the matched agents are chosen independently in each
round. Population size must be maintained in this setting in the presence of an adversary that
observes the entire state of the system and can continually delete or insert agents. We describe the
model in more detail below.
The population stability problem augments a growing body of work that uses the language and
ideas of distributed computing to model biological systems consisting of a collection of resourceconstrained components that collectively accomplish complex tasks. Naturally, we do not claim
direct relevance of our results to biological systems due to potential modeling differences. Regardless, the population stability problem makes sense in any system consisting of individual components
with the ability to reproduce.
1.1
Contributions
Our main contributions are as follows.
A New Problem in Distributed Computing: The Population Stability Problem. We
introduce a new problem in distributed computing. A population of N memory-constrained agents
(i.e. processors with the ability to reproduce and self-destruct) is subjected to adversarial attacks.
Whereas many attacks can be envisioned, we consider a worst-case adversary that can delete or
insert agents at a bounded rate. The goal is to maintain a stable population size within a small
multiplicative factor of the original size N . This problem appears fundamentally different from the
1
classical problems of distributed computing, such as consensus, leader election, majority, common
coin flipping, or computing general functions of the joint state of the parties.
Models for Communication and the Adversary. The communication model we consider is
a synchronous variant of the population model of [AAD+ 06, AAE07]. That model was designed
to represent sensor networks consisting of very limited mobile agents with no control over their
own movement and whose goal is to compute some function of their inputs or evaluate a property
of the system. Whereas [AAD+ 06] assumed that pairs of agents can communicate via pairwise
interactions as scheduled by a uniformly random matching process, we assume in addition that
agents are synchronized and interact with one another in rounds. Within each round, at least a
γ fraction of agents participates in pairwise interactions, again scheduled uniformly at random.
The agents additionally have the ability to self-destruct and to reproduce by producing a second
identical copy of themselves.
It is clear that we cannot allow the adversary to delete most or all of the agents in a single
round, since maintaining a stable population size in the presence of such an adversary is impossible.
Consequently, we give the adversary a budget of K alterations to perform in each round, where an
alteration consists of removing, inserting or modifying the memory of a single agent. We allow the
adversary to observe the memory contents of every agent before determining its alterations for a
round. Both γ and K are parameters of the model. The model is described in detail in Section 2.
Protocol for Population Stability. We present a protocol with three-bit messages requiring
polylog(N ) states (i.e. Θ(log log N ) bits of memory) per agent that tolerates K = N 1/4−ǫ worstcase insertions or deletions in each round, for any constant ǫ > 0. Formally, our main theorem is
the following.
Theorem 1. Let α, γ, ǫ be positive constants, where γ is a lower bound on the fraction of agents that
is matched in each round. Then there exists a population stability protocol with three-bit messages
and polylog(N ) states per agent and guaranteeing that if the adversary inserts or deletes at most
K = O(N 1/4−ǫ ) agents in each round, then with all but negligible probability the population will
remain between (1 − α)N and (1 + α)N for any polynomial number of rounds.
New Techniques. The main idea employed in our construction is to color the agents with
the values 0 and 1 in such a way that information about the population size is encoded in the
distribution of colors. That is, given a set of agents with assigned colors in {0, 1}, consider the
distribution specified by choosing an agent at random and observing its color. We are able to
assign colors to agents in such a way that approximate population size is encoded in the variance
of this distribution. Subsequently, each individual agent locally computes a very weak estimator of
whether the variance is too large or too small, and makes a local decision of whether to reproduce
or self-destruct. Although each individual agent’s estimate is noisy, we show that in the aggregate,
the local decisions are globally able to maintain a stable population size even in the presence of a
powerful adversary.
We discuss the model and results further in Section 1.2. We provide a more in-depth overview
of our techniques in Section 1.3 below. The model is described in further detail in Section 2. In
Section 3 we provide a full description of the protocol, and in Section 4 we present the analysis.
2
1.2
Discussion and extensions
Challenges. The first obstacle we must confront in designing a protocol for population stability
is the memory constraint. Each agent does not have sufficient memory to store a unique identifier
or to count to the population target N . Yet collectively, they must have a good approximation of
N and must individually decide whether to replicate if the population is too low or to self-destruct
if the population is too high. Making their task even more challenging, these memory-constrained
agents must correct deviations in the size of the population and make their decisions while the
adversary is acting.
A second major challenge is that the adversary is very powerful. Although the number of agents
inserted or deleted in each round is bounded, the adversary can observe the complete state of the
system1 before deciding on its actions, and may insert agents of arbitrary initial state. These
capabilities present difficulties for many standard techniques and abstractions used in distributed
computing. For instance, many protocols are based on leader election, where a single leader processor is chosen to direct or facilitate the task at hand. However, since our adversary is able to observe
the state of every agent, the adversary can simply wait for a leader to be chosen and then delete
it. Furthermore, even without observing the internal memory of agents, the adversary could insert
many additional agents that are all identical in state to the leader. Indeed, since agents can be in
one of only polylog(N ) distinct states and the adversary is able to insert K = N 1/4−ǫ ∈ poly(N )
agents per round, the adversary can even insert many copies of every possible agent type in each
round. Consequently any approach that relies on the existence of agents of unique or extremely
special state, such as leader election, seems doomed to failure. This appears to render ineffectual
a large part of the distributed computing toolbox.
Adversarial insertions. Recall that we allow the adversary to insert new agents with arbitrary
initial state. Starting from that internal state, we assume that the inserted agents execute the
same protocol as honest agents. We could instead consider an even stronger adversary that inserts
agents running arbitrary malicious protocols specifying their subsequent behavior. However, the
population stability problem as described above is clearly impossible in the face of this stronger
adversary, since our model does not include the ability to destroy agents that do not cooperate.
A malicious agent can simply ignore all interactions with other agents and replicate itself at every opportunity. Such malicious agents would quickly replicate themseves out of control, rapidly
exceeding the population target.
One may consider a different model that allows agents not only to self-destruct but also to
remove other agents it encounters. In such a setting, our protocol can be extended to achieve
population stability even if the adversary is allowed to insert agents that execute arbitrary malicious
programs, as long as there is a bound on how frequently malicious agents can replicate and an agent
is able to detect when it encounters an agent whose program is different from its own. However,
that setting is not the focus of this work.
Correction and detection. The objective in the population stability problem is to maintain a
population size that is close to a target value N , and to correct the population if it deviates too far
from this target. A closely related problem is that of simply detecting whether the population has
deviated too far from the target or whether it has exceeded some threshold, objectives that seem
1
That is, the adversary has the ability to read the memory contents of every agent.
3
very similar to the problem of approximate counting [Mor78]. It would be natural to try to solve
our problem by first detecting whether the population is too large or too small and then correcting
appropriately. With a weaker adversary that can only delete agents but not insert additional ones
(and additionally is oblivious to the internal states and coin flips of agents), this approach can be
made to work using approximate counting techniques. In the adversarial model considered here,
constructing approximate counters and detecting changes in population size are interesting open
questions.
Synchrony. In this work we study a synchronous model, where all agents communicate and
perform updates in rounds. As has commonly been the case across distributed computing, it is
natural to study a new problem first in a synchronous setting to distill key ideas and techniques
before adding additional complications in an asynchronous setting.
We note, however, that synchronization is far from an unreasonable assumption in biological
systems. Indeed, many multicellular systems do achieve synchrony either through regular external stimuli such as sunlight or through chemical control mechanisms. For instance, heart cells
maintained in culture were able to achieve a high degree of synchronization of their rhythmic contractions [JMPT87]. Neuronal cells too exhibit highly synchronized behavior. Even when grown in
culture, specialized neuronal cells show the capacity to synchronize the release of particular hormones at regular time intervals [DLECW92]. Bacterial populations have also been shown to have
the capability of producing coordinated oscillations [MO09, DMPTH10].
Nonetheless, an extremely natural and interesting question is how to solve the population stability problem in a setting without synchrony or with only partial synchrony. For instance, one
could consider a setting where agents have clocks that have bounded drift relative to one another.
Related to this is the typical random scheduler setting of population protocols [AAE07], in which a
single pair of agents at a time is chosen to interact and update state. By a concentration argument,
this process allows agents to maintain clocks that do not drift too quickly relative to one another.
While the construction in this paper requires synchrony, there are some known techniques in the
population protocol literature for maintaining approximate synchronization in a non-synchronous
setting; see, for instance, the recent work of [AAG18]. A natural extension is to show whether our
techniques can be combined with synchronizers to achieve population stability in such settings.
Alternate communication models. Another very interesting question is to explore the population stability problem under a different communication model. In this paper, pairs of agents that
communicate are chosen independently at random in each round. Alternatively, one could consider
settings in which the neighbors of an agent are consistent over time, perhaps reflecting underlying
geometric constraints.
One natural approach is to use a fixed sparse communication graph, for instance an expander.
However, modeling problems arise in determining how connectivity changes upon agent replication,
insertion, or deletion. In various settings along these lines, it is straightforward for the adversary to
disconnect the communication graph and consequently to violate population stability. An alternate
approach could be to associate agents with points in Rd , and to allow each agent to communicate
with a small number of the nearest other agents.
Population stability in the high-memory setting. We note that in the absence of memory
constraints, there is a trivial protocol both for approximate counting and for the population stability
4
problem if the adversary can only delete and not insert new agents. Each agent simply flips N
coins to generate a unique identifier id ∈ {0, 1}N . For an interval of O(poly log(N )) rounds each
agent broadcasts the set of identifiers it has received so far. With high probability, all identifiers are
unique and agents receive the identifiers of every agent that was alive throughout the interval, so
each agent learns a close approximation of the population size and can make a decision of whether
to self-destruct, replicate, or neither. However, this protocol relies heavily on agents having very
large memory, and does not yield an approach to solve the problem in the low-memory setting.
A note about success probability. Theorem 1 states that our protocol maintains a stable
population for any polynomial rounds with all but negligible probability. Recall that a function is
negligible if it tends to zero faster than any polynomial. That is, a function ν(x) is negligible if for
any c ∈ N there exists an x0 such that |ν(x)| < 1/xc for all sufficiently large x ≥ x0 .
Throughout this paper, we use the phrases “with high probability” and “with overwhelming
probability” to mean with probability 1 − ν(N ) for some negligible function ν.
1.3
1.3.1
Technical overview
Preliminary attempts
We first describe two preliminary attempts at protocols for the population stability problem which,
while unsound, will provide useful intuition toward the design of our actual protocol.
Attempt 1: non-interactive leader election As a first attempt in the low-memory setting,
consider the following approach, which is based on an idea from [AAE+ 17]. Each agent flips a biased
coin where Pr[c = 1] = N1 , where outcome 1 means that the agent is a leader. For O(polylog(N ))
rounds, each agent sends any agent it encounters the bit 0 if its coin was zero and it has not
received the message 1, and the bit 1 if its coin was 1 or it has received a 1 from another agent.
In the absence of an adversary, this allows every agent to learn whether a 1 was obtained in any
of the initial coin flips. The probability of this event differs noticeably depending on whether the
population is too small or too large. After repeating to amplify the signal, with high probability
the agents can detect if the population is too small or too large and can replicate or self-destruct
accordingly. A protocol of this form can be shown to work in the presence of an adversary that
can only delete and not insert agents, and is additionally oblivious to the coin flips made by the
agents. However, in the adversary model considered here with insertion as well as deletion and full
knowledge of the states of agents, the protocol will fail. The adversary can either insert an agent
with coin value c = 1 in each phase, or else identify the agent or agents with coin value 1 and
selectively remove these agents. Consequently the adversary can cause the population to grow or
shrink arbitrarily.
This attempt highlights a fundamental difficulty in designing protocols in our adversarial model.
The protocol relied on a non-interactive strategy related to leader election, where the presence or
absence of a leader could be used to infer the approximate size of the population. However, as
we have discussed above, the use of a special state with one or only a few agents of that state
(in this case agents with coin value c = 1) provides the adversary with an easy avenue of attack,
namely the deletion of agents of that state or the insertion of many additional agents of that state.
Consequently, constructions of this flavor seem to have little promise in this adversarial setting.
5
Attempt 2: independent coloring As a next attempt, consider the simple protocol in which
each agent flips a fair coin, receiving at random a color c ∈ {0, 1}. For each agent, compare the colors
of the next two agents encountered. If the colors are equal, then split, and otherwise self-destruct.
Observe that if an agent encounters the same agent twice, the colors must be the same, while if
an agent encounters two different agents, the colors are independently random. Consequently if
1
), which is slightly
the population currently has size m, then the probability of splitting is 12 + Θ( m
1
1
larger than the probability 2 − Θ( m ) of self-destructing, and so this protocol would cause the
population to increase slightly over time. To compensate this, modify the protocol to split only
with probability 1 − Θ( N1 ) if the colors are equal while still splitting with probability 1 if the colors
are unequal.2
Now, the population will stay the same size in expectation if its current size m is equal to
the target N , will decrease in expectation if m > N , and will increase in expectation if m < N .
Qualitatively, this is exactly the behavior that we want. However, tending in expectation to correct
itself is insufficient for maintaining a stable population. In fact, despite a very weak bias to correct
drifts in the population, the signal is overwhelmed by the noise, and the size of the population under
this protocol will behave very much like a random walk. Even in the absence of an adversary, this
protocol will cause the population to drift extremely far from its initial size.
In some sense the protocol we have just outlined behaves even worse than the empty protocol, in
that it fails to maintain a stable population when there is no adversary at all. However, the protocol
does have one intriguing feature. It entirely lacks any “special” agent types for the adversary to
exploit. Consequently, if we could design a more sophisticated protocol along these lines that could
maintain a stable population in the absence of an adversary, we might hope that it could do the
same even in the presence of an adversary.
1.3.2
Overview of our protocol
We now describe our actual protocol. At a very high level, the idea behind our protocol is as follows.
Through some coloring process which we will discuss below, agents are colored with the colors {0, 1}.
After the agents are colored, agents run a step called the evaluation phase in which agents make
the decision of whether to reproduce or self-destruct. The coloring process and evaluation phase
are then repeated indefinitely. We will refer to each iteration of the coloring process followed by
the evaluation phase as an epoch.
During the evaluation phase, each agent that is matched with another agent in this round
compares its own color with the color of its neighbor (i.e. the agent to which it is matched). If
the two agents have the same color, then the agent will replicate itself with some probability psplit .
If the two agents have different colors, then the agent will self-destruct. Note that if the coloring
process consisted of every agent tossing its own coin, then this would be essentially the same as
Attempt 2 above. We will instead employ a more structured coloring process that results in agent
colors that are not generated independently at random.
The coloring process will consist of two phases, a √
(noninteractive) leader selection phase and a
recruitment phase. In the leader selection phase, Θ(1/ N ) of the agents will become “leaders”3 and
2
Another perspective on why the protocol without this step cannot maintain a stable population is that the
protocol run by each agent has no dependence on the population target N . Consequently, if agents are added or
removed (either by adversary, or even as a result of random drift) then the protocol should behave as if those agents
were there to begin with and will not correct this deviation in the population.
√
3
That is, each agent will become a leader with probability Θ(1/ N ).
6
each
√ leader will choose a random color in {0, 1}. In the recruitment phase, each leader will identify
N uncolored agents and color each
√ of them with its own color. We note that the leader will not
directly encounter each of these N agents, but will directly color some agents which in turn will
color other agents. To begin with, each leader activates the first inactive agent that it encounters,
√
sharing its color with the new agent. Each agent is subsequently responsible for recruiting N /2
inactive agents in the same manner, forming a recruitment tree of depth 21 log N . By delegating
the coloring in this manner, the recruitment process can be√performed in only O(log2 N ) rounds.4
For each leader, at the end of the recruitment phase, N agents will
√ have obtained a color
based on the original coin toss of that leader. We will refer to these N agents as the cluster
associated with the leader, and we will sometimes describe agents as belonging to the same cluster
or different clusters. At the end of the recruitment phase and the beginning of the evaluation phase,
a constant fraction of the population will have been colored having been recruited into the clusters
of the various leaders, roughly half with color 0 and the other half with color 1.
Consider a particular agent in the evaluation phase. If the agent meets another agent from the
same cluster, then they necessarily have the same color. If the agent meets an agent from a different
cluster, then their colors are independently
random. Consequently, if the current population size is
√
1
N
m, then with probability 2 + Θ( m ) the two agents will have the same color, and with probability
√
1
2
− Θ( mN ) the two agents will have different colors. We can choose the splitting probability
√
psplit = 1 − Θ(1/ N ) so that the expected change in population is zero for m = N . For m ≫ N the
expected change in population will be negative, and for m ≪ N the expected change in population
will be positive. Moreover, unlike in Attempt 2 above, which behaved similarly to a random walk,
here the effect is strong enough to maintain the population in a small interval around the target
value N , with all but negligible probability.
Moreover, the adversary can do little to influence
√ the result of the protocol. For a population
size m = Θ(N ), the number of leaders selected is N , and so the standard deviation of the number
of leaders with each color is roughly N 1/4 . Consequently, an adversary that can insert or delete
o(N 1/4 ) agents can do little to influence the distribution of colors. Even if the adversary selectively
inserts or deletes agents that are leaders and have a particular color, the effect of the adversary on
the distribution of colors is dominated by the random deviation of the sampling process. Unlike in
Attempt 1 above, there are many leaders, and so the adversary is unable to delete enough of them
or to insert enough additional leaders to overwhelm the protocol. As we will show, the adversary
cannot cause substantial deviations in the size of the population.
Note that one strategy the adversary may attempt is inserting agents that do not know the
correct round number within the epoch. In the protocol described so far, there is no mechanism
for detecting and correcting this, and so over many rounds, adversarial insertions may lead to a
population of agents attempting to execute different portions of the protocol. In order to address
this, agents can exchange which round they are in, and self-destruct upon encountering an agent
that is in a different round of the epoch. This results in the self-destruction of any agent with
the wrong round number as soon as it encounters an agent with the correct round number. A
corresponding number of correct agents are also destroyed, but we will show that the number of
correct agents removed in this manner is sufficiently small.
As discussed briefly in Section 1.1, we can think of this protocol as encoding the population size
in the variance of a distribution and then sampling from this distribution to obtain a weak estimate
4
In order to achieve constant message size, our full protocol will be slightly different and will use additional rounds
for the recruitment process.
7
of the variance. Since the variance of the fraction of successes in many independent Bernoulli trials
decreases as the number of trials increases, if the number of leaders is larger, then the fraction
of colored agents with color 0 will be more closely concentrated around 1/2. On the other hand,
if the number of leaders is smaller, then we expect the fraction of colored agents with color 0 to
be farther from 1/2. Since the expected number of leaders is proportional to the current size of
the population, an approximation to the population size is encoded in the fraction of agents of
each color. Consider the distribution obtained by selecting an agent at random and reading its
color. Comparing the colors of two agents serves as a very weak estimate of the variance of this
distribution, while aggregating the results of many agents’ individual choices of whether to replicate
or self-destruct serves to amplify the accuracy of this estimate.
Achieving constant-size messages. The protocol described above involves messages of size
Θ(log log N ) bits, essentially as large as the entire memory of an agent. We now outline how
to modify the protocol to use constant-size messages. The only large portions of the messages
described so far consist of the current round in the epoch and the depth in the recruitment tree,
each of which can be encoded in Θ(log log N ) bits.
The current round in the epoch is sent to prevent the adversary from confusing the protocol by
inserting agents with the wrong round number. However, rather than sending the exact round, we
will instead send the single bit specifying whether or not the agent is currently in the evaluation
round. If an agent is entering the evaluation round and its neighbor is not, then both agents will
self-destruct. We can show that this suffices to maintain the invariant that a large majority of the
agents are in the same round of the epoch.
The depth of the recruitment tree is sent to allow each leader to induce the recruitment of the
correct number of agents. Note that we cannot simply recruit for 21 log N rounds, since recruiting
agents may encounter other agents that are already colored and cannot recruit them. However, we
can slow down the recruitment process to allow the depth in the recruitment tree to be determined
as a function of the round number. To do this, we recruit for Θ(log3 N ) rounds, divided into 21 log N
subphases of log2 N rounds each.5 In a single subphase, a recruiting agent will recruit only the
first inactive agent it sees even if it encounters many inactive agents in the subphase. This allows
agents to determine their depth in the recruitment tree based on the round in which they were
recruited. Since a subphase consists of ω(log N ) rounds, we will show that an inactive agent will
be encountered with high probability.
This yields a population control protocol with constant-size messages, since messages consist
of four binary values, namely an agent’s color, whether or not it is active, whether or not it is
recruiting, and whether or not it is currently in the evaluation round. In the analysis we will see
how to achieve the same result with only three-bit messages.
1.4
Related work
Population Protocols. The population protocol model was introduced by Angluin et al. [AAD+ 04,
AAD+ 06]. In this model a collection of agents, which are modeled by finite state machines, move
around unpredictably and have pairwise interactions. The original definition considers a worst-case
environment/scheduler, while later formulations [AAE07] consider the case where each interaction
occurs between a pair of agents chosen uniformly at random. In a population protocol, agents start
5
We actually only require that subphases are ω(log N ) rounds, so it is sufficient to recruit for ω(log2 N ) rounds.
8
with an initial configuration, and the goal is to jointly compute a function of this input. Previous
works have tried to identity the class of functions that can be computed in such a model [AAER07],
and the tradeoffs between the resources need to do so (e.g. [AAE+ 17]). In these works, the agents
are always active throughout the execution of the protocol.
Another line of work expands the population model to the case in which agents can crash or
undergo transient failures that corrupt their states. Delporte-Gallet et al. [DFGR06] consider a
setting in which agents must compute a function of their inputs in presence of such failures. They
construct a compiler that takes as input a protocol that works in the failure-free model, and outputs
a protocol that works in the presense of failures as long as modifying a small number of inputs
does not change the function output. Angluin et al. [AAFJ08] incorporated the notion of selfstabilization into the population protocol model, giving self-stabilizing protocols for some classical
problems such as leader election and token passing. They focus on the goal of stably maintaining
some property such as having a unique leader or a legal coloring of the communication graph.
Unlike these works, in our work agents have the ability to reproduce and self-destruct, and the
goal of maintaining a consistent population size must be carried out in the presence of an adversary
with the corresponding capability to insert and delete agents.
Approximate Counting. The problem of maintaining the population size of a collection of
memory-constrained agents is related to the problem of counting N items when the available memory of the agents is less than log N bits. Approximate counters were introduced by Morris [Mor78] as
technique to accurately approximate a value N using only Θ(log log N ) bits of memory. Techniques
for approximate counting in the population model were developed in [ABBS16, AAE+ 17].
A sequence of works by Di Luna et al. [LBBC14b, LBBC14a] consider the problem of estimating
the size of a network of agents that communicate according to a dynamic connection graph, in
presence of an adversary that can add and remove edges in the graph.
Cellular Automata. Cellular automata were proposed by von Neumann [vN51] as a model to
reason about artificial self-reproduction. A cellular automaton consists of a regular grid of cells,
each assuming one of a finite number of states. Over time, the states of cells change according to
some fixed rule (e.g. a mathematical function) that determines the new state of each cell in terms
of the current state of the cell and the states of the cells in its neighborhood. Conway’s Game of
Life [Gar70] is a cellular automaton that works with a simpler set of rules than von Neumann’s rules,
and was shown to be Turing-complete by Berlekamp, Conway and Guy [BCG04]. Cook [Coo04]
proved that rule 110 (a binary, one-dimensional cellular automata) is Turing-complete.
Our setting is crucially different from the cellular automaton setting, since agents in our model
do not simply change state but can be deleted from the system during the computation. Another
difference between our setting and the cellular automaton setting is that we consider an adversarial
model whereas in cellular automata cells deterministically change state. In some sense, in Conway’s
game, death “plays by the rules” while in our game death is sudden and unpredictable.
Dynamic Environments. A recent work of Goldreich and Ron [GR17] considers environments
that evolve according to a fixed local rule. They define an environment as a collection of small
components of a large system which interact in a local level, and change state according to a fixed
rule. As an example, they focus on the model of a two-dimensional cellular automata. They ask
how many queries a global observer must make about local components in order to test whether the
evolution of the environment obeys a fixed known rule or to predict the state of the system at a given
time and location. Although their work seems very different than ours, it bears some intellectual
similarity in seeking information about a global property of the system from local information.
9
However, whereas in [GR17] a global observer who can query a limited number of individual cells
asks “does the global system obey a specific evolution rule,” in our case individual agents need to
decide locally what they should do to maintain the overall global property of population size.
Self-Stabilization. Also related to our question is the self-stabilization problem introduced by
Dijkstra [Dij74]. Given a system that starts in an arbitrary state, the goal of a stabilization algorithm is to eventually converge to the correct state. In this setting, however, deletion of system
components is not considered. Super-stabilization [DH97] is the problem of achieving self stabilization in dynamic networks, that is, network where nodes are dynamically added and removed. While
this setting is closer to ours, super-stabilization algorithms make additional assumptions about the
system, such as that each node in the system is uniquely identified.
Distributed Algorithms Explaining Ant Colony Behaviors. A single ant has very limited
communication and processing power, yet collectively a colony of ants can perform complex tasks
such as consensus decision-making, leader election, and navigation. In [CDLN14] Cornejo et al. give
a mathematical model for the problem of task allocation in ant colonies, and propose a very efficient
protocol for satisfying this task. One of the main goals of their paper is to provide a formal model
enabling the comparison of the various task allocation algorithms proposed in the biology literature.
Similarly, in [GMRL15] Ghaffari et al. use techniques from distributed computing theory in order
to gain insight into the ant colony house-hunting problem, where a set of agents need to identify
potential nests, evaluate the quality of candidates, and reach consensus in a distributed manner.
2
The Population Stability Problem
As discussed above, the population stability problem is concerned with a system of agents with
bounded memory and the ability to reproduce and self-destruct. The system is subjected to adversarial attacks that delete or insert processors. The objective is to maintain a stable population
size despite these adversarial attacks. In this section we give a formal description of the problem.
Parameters. The population stability problem is parameterized by the initial number of agents
N , the number of distinct memory states M each agent can be in, the number of alterations K the
adversary is allow to make in each round (where an alteration consists of removing or inserting an
processor), a value α specifying how tightly concentrated the population size must remain around
the target value N , and a value γ specifying a lower bound on the fraction of processors that are
matched with other processors in each round. Below we describe each of these components of the
problem.
We note that one could consider separate parameters for the number of adversarial deletions
and insertions, allowing the adversary to make a different number of each. In this paper we will
consider both to be bounded by a single parameter K.
Agents. We consider agents with bounded memory. Each agent can be in one of M possible
states, so we can think of agents as having log M bits of memory. Agents can communicate by
message-passing as specified by the connectivity structure of the system, which will be discussed
further below. In our setting we will have M ≪ N , so each individual agent has insufficient memory
to count the total population, to posses a unique ID, or to address messages to a particular recipient.
Each agent has the ability to flip unbiased coins, to split into two identical agents, or to self destruct.
10
That is, in any round an agent may choose to split into two daughter agents which both inherit
specified state from the parent agent, or it can decide to delete itself from the system.
Connectivity. As discussed above, we consider a synchronous version of the population model
of Angluin et al. [AAD+ 04, AAE07], where we assume the existence of a global clock. We assume
that the pairs of agents that are able to communicate in each round are selected by choosing a
random matching of at least a γ fraction of surviving agents. We think of the parameter γ as a
constant (e.g. γ = 1/4). That is, each agent is matched with at most one other agent (that we call
its neighbor ) in each round, and there is no consistency from round to round, since connectivity
in different rounds is determined by sampling independently random matchings. The schedule of
these matchings is unknown to the adversary in advance.
Adversary. We consider a worst-case, computationally unbounded adversary that can arbitrarily
choose which agents to delete in each round and can insert agents with arbitrary state in each round.
The adversary also can observe the entire history of agent interactions, including the memory
contents of every agent. While the initial state of inserted agents is determined by the adversary,
the newly inserted agents are assumed to follow the protocol (that is, the agents introduced by the
adversary do not behave maliciously).
Objective. The goal in the population stability problem is to maintain a number of agents within
a small interval [(1−α)N, (1+α)N ] around the initial population size N . That is, initially the system
consists of N agents. In each round some agents may be removed or inserted by the adversary, and
some agents may decide to replicate or to self-destruct. Let Ni denote the number of agents in the
system after the ith round. The adversary wins in round i if Ni ∈
/ [(1 − α)N, (1 + α)N ] at the end
of the round. We say that protocol Π is a population stability protocol if for any polynomial p and
any adversary, the probability that the adversary wins in at most p(N ) rounds is negligible in N .
3
The protocol
We now provide a formal specification of the main protocol (Algorithm 1). Agents continually
run the protocol throughout their lifetime. We will think of time as partitioned into epochs of
T = 21 log3 N rounds. Each epoch consists of three phases, the leader selection phase, the recruitment phase, and the evaluation phase. The recruitment phase consists of 12 log N subphases each
consisting of roughly Tinner = log2 N rounds.6 We will elaborate on each of these phases below.
Recall that agents have the ability to toss coins, to send and receive a message upon encountering
another agent, to reproduce by splitting into two identical copies of itself, and to self-destruct.
These capabilities are notated by the following functions. We denote flipping an unbiased coin by
$
x ← {0, 1}. Agent splitting and death are implemented in commands Split() and Die(). Finally,
the command Z := Communicate(X) sends message X to the neighboring agent in the present
round, if any, simultaneously receiving in response message Z.7 If the agent is unmatched in this
round and has no neighboring agent, then the return value is assumed to be ⊥. In the protocol
6
More generally, we want T = Tinner · 21 log N for any Tinner = ω(log N ). The first and last subphase will each be
shorter by one round to account for the leader selection and evaluation phases.
7
Recall that an agent’s neighbor is the other agent the agent is randomly matched with in this round and that
matchings in each round are independent and uniformly random.
11
below, messages will consist of four boolean values (inEvalPhase, active, color, recruiting). Upon
receiving message Z, the value of each of these variables can be accessed by writing Z.inEvalPhase,
Z.active, and so forth. If Z = ⊥ then we will follow the convention that each of these components
will also have value ⊥.
The main variables that describe the state of an agent are round ∈ [0, T − 1] and four boolean
variables active ∈ {0, 1} color ∈ {0, 1}, recruiting ∈ {0, 1}, and inEvalPhase ∈ {0, 1}. The variable Nbr ∈ {0, 1}4 stores the most recent message received, consisting of four boolean values
(Nbr.inEvalPhase, Nbr.active, Nbr.color, Nbr.recruiting). An additional variable, to recruit ∈ [0, 21 log N ],
is not necessary for the protocol itself but is used in the analysis. We emphasize that these variables are local to a single agent, so that different agents have independent copies of each of these
variables. Initially, at the onset of the system, for each agent we will have that all variables are set
to zero.
The variable round keeps track of which round it is within the epoch. The variable is incremented
modulo T after each round, ensuring that agents begin and end each phase and each epoch at the
same time.
The variables active and color specify whether an agent has been activated and colored, as
well as the color of the agent. In the first round of each epoch, some of the agents will designate
$
themselves as leaders and will become active, choosing at random a color color ← {0, 1}, while the
rest of the agents remain inactive. During the recruitment phase, additional agents will become
active and will receive colors in {0, 1}, as inherited from a leader. The value of variable color is
only relevant for active agents.
The variable recruiting specifies whether or not an active agent is trying to recruit in the present
subphase. Each active agent should recruit only one additional agent in a single subphase, so this
variable specifies whether or not the agent is still looking for an inactive agent to recruit in the
subphase.
The variable inEvalPhase specifies whether the agent is currently in the evaluation phase, which
is true exactly when round = T − 1.
Finally, the variable to recruit specifies the number of additional followers an active agent is
tasked with recruiting directly, which is the logarithm of the total number of agents that should be
activated as a result of the given agent. When a new leader first becomes active,
it sets to recruit =
√
1
recruit
to
= N agents. Each time a
2 log N , indicating that it is tasked with recruiting a total of 2
new agent is recruited, the value of to recruit is decremented and shared with the newly recruited
agent, indicated that each of the two is responsible for recruiting only half of the total. For instance,
after the first time a leader recruits another agent, both agents will have to recruit = 21 log N − 1,
√
and so each of the two agents subsequently are
√ responsible for recruiting only N /2 agents. In this
way a leader can induce the recruitment of N agents in roughly a logarithmic number of rounds.
Although the variable is not used by the algorithm itself, we will refer to it in the analysis.
We first present the main procedure run by each agent in every round. Each agent first exchanges
messages with its neighboring agent in this round, if any. The agent then performs a consistency
check on its on state and the state of its neighbor. Then, depending on the value of variable round
modulo T , the program calls the appropriate subroutine for the corresponding phase.
12
Algorithm 1 Main protocol
1: procedure MainProtocolStep
2:
ExchangeMessages()
3:
CheckRoundConsistency()
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
if round = 0 then
DetermineIfLeader()
round := round + 1
else if round < T − 1 then
RecruitmentPhase()
round := round + 1
else
EvaluationPhase()
round := 0
end if
end procedure
⊲ Handle initialization of leaders
⊲ Perform recruitment phase
⊲ Final round of phase. Perform split or death.
We now describe the subroutine that exchanges messages with the neighboring agent, if any. An
agent simply computes the indicator value of whether it is in the evaluation phase, and sends this
information along with its activation state, color, and recruiting status. It receives a corresponding
message from its neighbor, or the value ⊥ if it has no neighbor in this round.
Algorithm 2 Subroutine to send and receive messages with neighboring agent
1: procedure ExchangeMessages
2:
if round = T − 1 then
3:
inEvalPhase := 1
4:
else
5:
inEvalPhase := 0
6:
end if
7:
MyStatus := (inEvalPhase, active, color, recruiting)
8:
Nbr := Communicate(MyStatus)
9: end procedure
We now describe the leader selection subroutine, which comprises the first round of each epoch.
This is an entirely
√ non-interactive process in which each agent becomes a leader with some fixed
probability 1/(8 N ) by tossing its own coins, entirely independently of each other agent. With
overwhelming probability,
√ if the total number of agents is Θ(N ), the number of leaders chosen in
this phase will be Θ( √
N ). Each newly activated leader chooses a random color in {0, 1} and is
tasked with recruiting N agents and assigning them this color.
13
Algorithm 3 Leader selection phase subroutine
1: procedure DetermineIfLeader √
2:
active := TossBiasedCoin(log(8 N )).
√
⊲ active = 1 with probability 1/(8 N ), and otherwise active = 0.
3:
if active = 1 then
$
4:
color ← {0, 1}
5:
recruiting := 1
6:
to recruit := 12 log(N )
7:
end if
8: end procedure
The leader selection phase, as well as the evaluation
phase below, requires the ability to flip
√
biased coins with bias 1/poly(N ), in particular 1/Θ( N ), where bias refers to the probability of the
coin having value 1 (so that a fair coin has bias 1/2). Recall that we assume only the ability to toss
unbiased coins. We now give a simple procedure to obtain the desired bias using only O(log log(N ))
bits of memory, assuming that log N is an even integer. More generally, we show how to obtain
bias 2−a for any integer a using 1 + ⌈log a⌉ bits of memory. We note that it is sufficient to toss a
coins and report 1 if they all landed heads and 1 otherwise. This requires counting to a, which can
be done with log a memory.
Algorithm 4 Subroutine to toss a biased coin that equals 1 with probability 2−a and equals 0
otherwise
1: procedure TossBiasedCoin(a)
⊲ Flip a biased coin with bias Pr[c = 1] = 2−a
2:
c := 1
3:
for i = 1 to a do
$
4:
b ← {0, 1}
5:
if b = 0 then
6:
c := 0
7:
end if
8:
end for
9:
return c
10: end procedure
We now describe the recruitment phase, which is the second phase executed by each agent in
every epoch and is the main source of interaction in the protocol. The phase lasts for T = Θ(log3 N )
rounds, consisting of 21 log N subphases each of length Tinner = Θ(log2 N ) rounds. As discussed
√
above, during this phase each leader is tasked with finding N inactive agents (i.e. with active = 0)
and coloring each of√them with the color of the leader. We note again that the leader will not directly
meet each of these N inactive agents, but rather that this is done by propagation, where the leader
will activate some agents and each of these will activate additional agents. In each subphase each
active agent will attempt to recruit a single nonactive
√ agent, which will then start to recruit in
1
the following subphase. Since there are 2 log N = log N subphases, if each attempt to recruit is
√
successful, then a single leader will result in the activation of a total of N agents.
14
Algorithm 5 Recruitment phase subroutine
1: procedure RecruitmentPhase
2:
if Nbr.active = 0 and recruiting = 1 then
3:
recruiting := 0
4:
to recruit := to recruit − 1
5:
else if active = 0 and Nbr.recruiting = 1 then
6:
active := 1
7:
color := Nbr.color
8:
recruiting := 0
9:
to recruit := 12 log N − ⌈(round + 1)/Tinner ⌉
10:
end if
11:
if round ≡ −1 (mod Tinner ) then
12:
recruiting := 1.
13:
end if
14: end procedure
⊲ Other agent has been activated
⊲ This agent must be activated
⊲ Prepare for next round
The final phase of the algorithm is the evaluation phase, which occurs on the last round of each
epoch. In this phase, each matched active agent compares its color to that of its neighbor and
makes a decision of whether to replicate itself or to self-destruct.
Algorithm 6 Subroutine to perform splitting and self-destruction at the end of each phase
1: procedure EvaluationPhase
2:
if active = 1 and Nbr.active = 1 then
√
3:
if Nbr.color = color then
⊲ Colors same: split with probability 1 − 16/ N
√
4:
c := TossBiasedCoin(log( N /16))
5:
if c = 0 then
6:
Split()
7:
end if
8:
else
⊲ Colors different: self-destruct with probability 1
9:
Die()
10:
end if
11:
end if
12:
active := 0
13:
color := 0
14:
recruiting := 0
15: end procedure
Finally, we give the subroutine invoked at the very beginning of each round, that performs a
consistency check on the round values of the agent and its neighbor. In the absence of adversarial
insertions this subroutine is unnecessary, since agents will always have the correct round value in
the epoch. However, it is necessary if the adversary is allowed to insert agents with an incorrect
round value. If left to increase unchecked, the presence of many agents with different round values
would interfere with the operation of the protocol. We will prevent this from happening by causing
agents to self-destruct as soon as they encounter an agent with a different round value. However,
implementing this exactly would require Θ(log log N )-bit messages, since agents would need to
15
exchange their round numbers. To avoid this, we will instead have agents exchange only an indicator
variable for whether or not they are in the evaluation phase. An agent will self-destruct if it is in
the evaluation phase and its neighbor is not, or if its neighbor is in the evaluation phase and it
is not. This process deletes a small number of agents with the correct round number along with
agents with the incorrect round number. We will show in the analysis below that this will ensure
that there are few agents with the incorrect round number and that only a small number of agents
with the correct round number will self-destruct as a result of this procedure.
Algorithm 7 Subroutine to determine if agent knows the correct round
1: procedure CheckRoundConsistency
2:
if Nbr 6= ⊥ and inEvalPhase 6= Nbr.inEvalPhase then
3:
Die()
4:
end if
5: end procedure
4
Analysis
In this section we prove the main theorem.
Theorem 2. Let α, γ, ǫ be positive constants, where γ ≤ 1 is a lower bound on the fraction of
agents that is matched in each round. Then Algorithm 1 is a population stability protocol using
ω(log2 N ) states per agent and three-bit messages8 guaranteeing that if the adversary inserts and
deletes at most K = O(N 1/4−ǫ ) agents in each round, then with all but negligible probability the
population will remain between (1 − α)N and (1 + α)N for any polynomial number of rounds.
For ease of presentation, the version of the protocol described above uses four-bit messages.
However, we can reduce the message size to three bits, as follows. If the agent is in the evaluation
phase (i.e. inEvalPhase = 1) then the message must contain the values active and color, but need
not contain recruiting. If inEvalPhase = 0, then the message must contain the value color but not
active if recruiting = 1 and the value active but not color if recruiting = 0. Consequently, the desired
information can be encoded in only three bits.
For the memory requirements, log T bits are needed to store the variable round ∈ {0, 1}, and
the other variables stored by each agent consist of eight boolean values. The invocations of the
subroutine TossBiasedCoin() require log log N bits of local memory. However, the subroutine is
only invoked in two rounds of each epoch, the leader selection round and the evaluation round.
Consequently, using additional indicator bits to specify whether an agent is in each of these those
two rounds, the memory used to store the variable round can be used as the helper memory for the
subroutine, and so additional memory is not necessary. For T = 12 log3 N , the total number of states
is therefore Θ(log3 N ). However, it suffices to have T = Tinner · log N for any Tinner = ω(log N ), and
so we can reduce the number of states to ω(log2 N ).
Roadmap to proof We must show that the population size will remain close to the target value
N . We will do this by means of two key steps.
8
A straightforward implementation of the protocol described above would use Θ(log4 N ) states and four-bit messages. We describe below how to achieve the improved bounds stated here.
16
The first step is to show that in any single epoch the population size will be relatively stable.
That is, we show that the population in the middle or end of an epoch will not be too much larger
or smaller than the population at the beginning of the epoch. We achieve this by showing that
irrespective of the adversary’s actions, with overwhelming probability the number of agents of each
color in the evaluation phase will be concentrated around one-sixteenth of the total number of
agents. This step is formalized in Lemmas 6 and 7 in Section 4.2.
The second step is to show that the population size will tend to correct itself if it has deviated
too far from the target value N . More precisely, we show that if the population is far from N , then
in expectation the population at the end of the phase will be substantially closer to N than the
population at the start of the phase. This step highlights a key tension of the proof, namely the
difficulty analyzing a system that contains both random components in the matching schedule and
worst-case components in the adversary’s insertions and deletions. Indeed, it is not even clear a
priori what it means to discuss expectation in a system with a worst-case adversary. This step is
formalized in Lemma 8 in Section 4.3.
Putting these two steps together will enable us to conclude the proof of the theorem. Consider
the first epoch in which the population size lies outside the interval [(1 − α2 )N, (1 + α2 )N ]. Since the
population size does not deviate too much in a single epoch, the population at the start of the next
epoch will be close to (1 ± α2 )N . But for each epoch in which the population remains outside the
interval [(1 − α2 )N, (1 + α2 )N ], in expectation the change in population will be in the direction of
the target value N . Considering the next N 0.01 epochs, a Chernoff-Hoeffding bound then implies
that with overwhelming probability the population will return to the interval [(1 − α2 )N, (1 + α2 )N ].
Since the population does not change much in each epoch, it follows further that the population will
remain inside the interval [(1 − α)N, (1 + α)N ] during these N 0.01 epochs. We will conclude that
with all but negligible probability, the population will remain in the interval [(1 − α)N, (1 + α)N ]
for any polynomial number of rounds.
We now outline the remainder of the section. In Section 4.1 we prove some preliminary lemmas
about the protocol. These lemmas provide us with invariants that we will need in order to prove
the two key steps above. In particular, we show that nearly every agent knows the correct round
in the epoch, that at least half of the agents are inactive (i.e. have active = 0) at any point in the
execution of the protocol, and that
√ any leader selected in the first round of the epoch will succeed
in recruiting a full cluster of size N unless either the adversary deletes some agent in that cluster
or an agent with the wrong round number interferes with the recruitment.
In Section 4.2 we prove the first of the key steps, showing that with high probability the
population size does not deviate too much in a single epoch. In Section 4.3 we prove the second
of the key steps, showing that if the population has drifted too far from the target value, then in
expectation the population will correct itself. Finally, in Section 4.4 we conclude the proof of the
main theorem.
4.1
Bookkeeping lemmas
We will first prove several bookkeeping lemmas to guarantee that with overwhelming probability,
certain invariants continue to hold throughout the execution of the protocol. We will subsequently
use these invariants to prove stronger statements that will enable us to conclude the correctness
the protocol.
Recall that during the protocol, agents keep track of the current round within the epoch, that is,
the round number modulo T = log3 N . However, the adversary is empowered to insert new agents
17
into the system with arbitrary initial state, and in particular may insert agents with the incorrect
round number. Our first lemma provides a bound on the number of agents with the incorrect round
number. Recall that γ = Θ(1) is a lower bound on the fraction of parties in each round that are
matched. We can assume that α ≤ 1/2, since the desired statement is stronger for smaller α.
Lemma 3. For any t > 0, suppose that the population size remains above Nmin = N/2 for the first
t rounds of the protocol. Then there exists some negligible function ν such that conditioning on this
event, with probability 1 − t · ν(N ), all but (1 + γ −1 )N 1/4 of the agents will have the same value for
variable round in each of these t rounds.
Proof. We prove this by induction on the round number. Initially all agents have round = 0.
Assume for induction that at round r all but γ −1 N 1/4 of the agents have variable round = 0. We
will show that with high probability, the same statement holds at the start of round r + T . The
adversary may add an additional K agents in each of these T rounds, for a total of KT ≤ N 1/4 /8
agents. Note that each agent with the wrong round value may split at most once in this epoch
of T rounds, when it reaches its evaluation phase. If there are v = O(N 1/4 ) agents which differ
from the majority value for variable round, then the probability of such an agent being matched
with another such agent in its evaluation phase is at most γ · Nvmin = γ · O(N −3/4 ). It follows
that with high probability, the number of agents with the wrong round value that split during this
epoch is at most N 0.01 /8 < N 1/4 /8. Consequently at any point in this epoch of T rounds, the
number of agents with the wrong round value will be at most γ −1 N 1/4 + N 1/4 /4. Each agent with
the wrong round value at the start of the majority evaluation phase (i.e. round r + T − 1) has
probability at least γ(Nmin − (γ −1 + 1/4)N 1/4 )/(Nmin ) ≥ γ − 3N −3/4 of being matched with an
agent starting the evaluation phase, so with all-but-negligible probability, at most (1/γ − 1/2)N 1/4
of these agents will not be matched with an agent with different round value and will survive the
round. It follows that at the start of round r +T , at most γ −1 N 1/4 agents have round value different
from 0. Consequently, by induction we have that with overwhelming probability, the number of
agents with variable round 6= 0 in any round r ≡ 0 (mod T ) is at most γ −1 N 1/4 . Since we showed
above that the number of agents with the wrong round number that can be added during the epoch
is small with overwhelming probability, the lemma follows.
Lemma 4. With high probability, if the population is in the interval [(1 − α)N, (1 + α)N ] at the
start of an epoch, at any point in the epoch, at most 1/2 of the agents have active = 1.
Proof. Let m ∈ [(1 − α)N, (1 + α)N ] be the population at the start
With all but
√ of the epoch.
0.01 ) by a Chernoffo(N
negligible probability the number of leaders chosen will be m/(8 N ) ±
√
Hoeffding bound. Each leader may induce the activation of at most N total agents. During
the epoch, the adversary
may insert an additional T · K ≤ N 1/4 agents, which may each induce
√
the activation of N total agents. Consequently at any point in the epoch, the number of active
agents will be at most m/8 + N 3/4 . Prior to the evaluation step, the adversary can have killed
at most T · K ≤ N 1/4 agents. By the previous lemma, at most γ −1 N 1/4 agents at the start of
the epoch can have the wrong value for round, and at most T · K additional such agents can
be introduced during the protocol, so at most O(γ −1 N 1/4 ) = O(N 1/4 ) agents can be killed in
the CheckRoundConsistency() procedure. Consequently the population throughout the epoch
until the evaluation phase will be at least m − O(N 1/4 ). The conclusion follows.
18
Lemma 5. Suppose the population is in the interval [(1 − α)N, (1 + α)N ] at the start of an epoch.
Then with high probability, in the last round of the epoch, every active agent entering the evaluation
phase that was not inserted by the adversary during this epoch will have to recruit = 0.
Proof. By Lemma 4, at most half of the agents are active in each round, so the probability in each
round of encountering an inactive agent at each round is at least γ/2 = Θ(1). With overwhelming
probability, in any sequence of ω(log N ) steps an agent will encounter an inactive agent and will be
able to recruit it. Applying a union bound, we have that in each cycle of Tinner steps, each of the
O(N ) active agents attempting to recruit will be successful in finding an inactive agent to recruit.
Consequently each agent will be able to recruit the desired number of additional agents, and the
lemma follows.
4.2
Bounded deviation
In this section we show that with high probability, the population size does not change by too much
in any single epoch.
Lemma 6. Let m be the population at the start of an epoch. With high probability, the number of
agents with each color b ∈ {0, 1} at the start of the evaluation phase will be m/16 ± O(N 3/4−ǫ ).
Proof. Let m be the population at the start of the epoch. With all but negligible √
probability, the
number of leaders selected with color 0 at the beginning of the epoch will be m/(16 N ) ± o(N 0.01 ),
and similarly for the number
√ of leaders selected with color 1. In the absence of adversarial deletions,
each leader will recruit N followers with the same coin value, inducing the presence of m/16 ±
o(N 0.51 ) agents of each color by the final round of the epoch.
The adversary may insert or delete K · T = Õ(N 1/4−ǫ
√ ) agents over the course of the epoch.
Each inserted agent can induce the activation
√ of at most N additional agents, and similarly each
removed agent could have activated up to N additional agents. Additionally, by Lemma 3, at most
O(N 1/4 ) agents will be removed in procedure CheckRoundConsistency() upon encountering an
agent with a different value for variable round. Overall the actions of the adversary can affect the
number of agents of each color by at most O(N 3/4−ǫ ). It follows that despite adversarial action,
the number of agents of each color at the start of the evaluation phase will be m/16 ± O(N 3/4−ǫ )
with all but negligible probability.
Lemma 7. With all but negligible probability, if the population is in the√interval [(1−α)N, (1+α)N ]
at the start of an epoch, the population will have deviated by at most Õ( N ) by the end of the epoch.
Proof. Let m ∈ [(1 − α)N, (1 + α)N ] be the population at the start of the epoch. By the previous
lemma, at the start of the evaluation phase the number of agents with color 0 will be m0 =
m/16 ± O(N 3/4−ǫ ) with all but negligible probability, and likewise the number of agents with
color 1 will be m1 = m/16 ± O(N 3/4−ǫ ). We condition on these events. Since the adversary can
insert at most K agents in each of the T = log3 N rounds for a total of Õ(N 1/4−ǫ ), and no other
new agents with the correct value of variable round can be produced until the evaluation phase,
Lemma 3 implies that the total number of agents at the start of the evaluation phase is at most
m+O(γ −1 N 1/4 ). Similarly, since the adversary can have directly removed at most K ·T = K log3 N
agents, and at most 2((γ −1 + 1)N 1/4 + K log3 N ) agents with the correct value of round may have
been removed as a result of procedure CheckRoundConsistency() after encountering an agent
19
with a different value of round, it follows that for γ = Θ(1), the total number of agents at the start
of the evaluation phase is m ± O(N 1/4 ).
Consequently the communication graph for the evaluation phase is a random matching of size
q = Ω(γN ) = Ω(N ). Sample such a matching by first choosing a set of q left vertices and a set of q
right vertices, and then associating corresponding vertices on the left and right. Let q0 = qm0 /m′
and q1 = qm0 /m′√
. With high probability the number of left (respectively, right) vertices with color
0 will be q0 ± Õ( N ), and similarly for vertices of color 1.
It follows that with all but negligible probability, the number√of left-vertices of color b that split
after being matched with a right-vertex of color b is qb2 /q ± Õ( N ) for each b ∈ {0, 1}. Similarly,
the number of left-vertices of √
color b that self-destruct after being matched with a right-vertex
of color 1 − b is qb q1−b /q ± Õ( N ). Consequently will all but negligible probability, noting that
q0 − q1 = O(N 3/4 ), we have that the change in population during the evaluation phase is
as desired.
4.3
√
√
2
· q02 + q12 − 2q0 q1 ± Õ( N ) = Õ( N )
q
Correcting population drift
In this section we show that if the population has drifted too far from the target value N , in
expectation it will tend to correct itself.
Lemma 8. If the population is in the interval [(1 − α)N, (1 − α2 )N ] at the start of an epoch, then
√
for any adversarial strategy, in expectation the population will increase by Ω( N ) by the end of the
epoch. If the population is in the interval [(1 + α2 )N, (1 + α)N ] at the start of an epoch, then in
√
expectation the population will decrease by at least Ω( N ) by the end of the epoch.
Proof. The behavior of the system during the evaluation phase depends on the distribution of coin
values of active agents in this phase. We would like to argue that each pair of clusters has its
colors assigned by independent, fair coin flips. This is clearly true in the absence of an adversary.
However, in our setting, adversarial instertions and deletions can bias the joint distribution of the
colors of a pair of agents in different clusters.9 Nonetheless, we will argue that the adversary’s
influence is limited, and that for most pairs of agents in different clusters, we can think of the joint
distribution of colors as unbiased and uniform. We will label clusters as honest or adversarial,
where the colors of a pair of honest clusters can be regarded as independently sampled, and no
assumption is made on the colors of the adversarial clusters.
Consider any fixed adversarial strategy. Note that the adversary can insert or delete a total of
no more than 2 · T · K = Õ(N 1/4−ǫ ) agents during the epoch, and consequently can influence no
more than this many clusters. Any strategy of the adversary during this epoch can be emulated
by deferring any deletions of colored agents (with the correct round value) to the beginning of the
evaluation phase, deleting instead an inactive
agent. Recall that by Lemma 5, each cluster not
√
affected by the adversary will consist of N agents at the beginning of the evaluation phase. At
9
For instance, in the first round of the epoch, the adversary can instert additional leaders all with color 0, or can
delete several leaders that have color 1. This difficulty arises because we allow the adversary to observe the internal
memory of all agents, including the results of coin tosses.
20
the beginning of the evaluation phase, we allow this new√adversary not only to delete the specified
agent, but also to set active = 0 for any subset of the N − 1 other agents in the cluster. Note
that any attack that could be accomplished by the original adversary can still be carried out by
this new adversary that defers all of its deletions of colored agents to the evaluation phase but is
subsequently allowed to modify up to Õ(N 3/4−ǫ ) agents in Õ(N 1/4−ǫ ) different clusters.
Since we now defer deletions of colored agents to the beginning of the evaluation
√ phase, each
cluster induced by a leader selected in the first round of the epoch will have the full N members,
and consequently these clusters are indistinguishable except for their color. Consequently as long as
the adversary can modify agents in Õ(N 1/4−ǫ ) clusters of each color, which specific cluster of each
color is irrelevant. Consider an arbitrary indexing of the agents before the first round of the epoch,
1
and consider the first Õ(N 1/4− 2 ǫ ) leaders chosen in this round. With all but negligible probability,
this set will contain at least Õ(N 1/4−ǫ ) agents of each color color ∈ {0, 1}, and so the strategy of
the adversary can be carried out by manipulating only the clusters of agents in this set. Let the
clusters induced by these agents be the adversarial clusters along with the clusters induced by any
agents inserted by the adversary, and let the remaining clusters be the honest clusters.
Now we have reduced
√ to a setting in which we have achieved the desired property that each
honest cluster has size N , and the coin flips of any two honest clusters are independent. However,
we are now dealing with a modified adversary that can affect a larger overall number of agents. Let
m′ be the number of agents at the start of the evaluation phase. By Lemma 6, it follows that with
1
all but negligible probability, the number of agents in honest clusters is mh = m′ /8 ± O(N 3/4− 2 ǫ ),
1
and the number of agents in clusters influenced by the adversary is ma = O(N 3/4− 2 ǫ ).
Pick a random pair of vertices at the start of the evaluation phase. Then with probability 1/64−
O(N 3/4 /m′ ) = Ω(1) both agents will belong to honest clusters, with probability o(N 3/4 /m′ ) one
agent will belong to an honest cluster and the other to an adversarial cluster, and with probability
o(N 3/2 /m′2 ) both will belong to adversarial clusters.
A pair of vertices belonging to honest clusters will have the same color if they belong to the
same cluster, and independently random colors if they belong to different√clusters. Consequently
the probability that such a pair of vertices will have the same color is 21 + 2mNh − O( m1h ). It follows
that the expected change in population resulting from the matching of a pair of vertices belonging
to honest clusters is
√
√ !
√ !
1
16
16
2 N
N
N
O(1)
· 1− √
−O
−√ −
.
=
1+
− 1−
mh
mh
mh
mh
mh
N
N
√
Recalling that α ≤ 1/2 is a fixed constant, for m < (1 − α2 )N this quantity is Θ(1/ N ), and
√
for m > (1 + α2 )N this quantity is negative, with magnitude Θ(1/ N ).
The honest clusters consist of nearly the same number of agents of each color (m/16 ± O(N 3/4 )
of each). It follows that the expected change in population resulting from the matching of a vertex
in an honest cluster and a vertex in an adversarial cluster has magnitude O(N −1/4 ). Making no
assumption about the distribution of colors in the adversarial clusters, the matching of two vertices
in an adversarial cluster can have a change in population of O(1).
Consider a random pair of vertices at the start of the evaluation phase. For m < (1 − α2 )N ,
the expected change in population resulting from matching this pair of vertices is Ω(N −1/2 ) −
o(N −1/4 N −1/4 ) − o(N −1/2 · 1) = Ω(N −1/2 ). Since the number of matched pairs of vertices in the
evaluation phase is γm′ = Θ(N ), it follows from linearity of expectation that the expected change
21
√
in population during the evaluation phase is Ω( N ). Similarly, for m > (1 + α2 )N we have that the
√
expected change in population during the evaluation phase is −Ω( N ). The adversary can delete
or insert only K · T = o(N 1/4 ) agents during the epoch, and Lemma 3 implies that O(N 1/4 ) agents
will self-destruct during procedure CheckRoundConsistency() during the epoch, so the other
terms dominate, and the conclusion follows.
4.4
Putting everything together
We now show that if the population size leaves the interval [(1 − α2 )N, (1 + α2 )N ] during an epoch,
with high probability it will return to this interval during one of the next few epochs. With this
final lemma, we then conclude the proof of the theorem.
Lemma 9. Consider an epoch in which the population has drifted outside the interval [(1− α2 )N, (1+
α
2 )N ]. With all but negligible probability, the population will once again be in the interval [(1 −
α
α
0.01 epochs.
2 )N, (1 + 2 )N ] at the start of one of the next t = N
Proof. Assume to the contrary, and let epoch 0 denote the first epoch after the population has
exceeded the interval [(1 − α2 )N, (1 + α2 )N ]. For concreteness, suppose the population has dropped
below (1 − α2 )N . By Lemma 7, with all but negligible probability the population is still above
√
(1 − α2 )N − Õ( N ). For i ∈ {1, . . . , t}, let Xi be the random variable denoting the difference in
population between the start of epoch i and the start of epoch i − 1, and let X = (X1 + · · · + Xt )/t.
7, √
with all but negligible probability, each random variable Xi is bounded in the range
By Lemma
√
[−Õ( N ), Õ( N )]. Since by assumption the population is below (1 − α2 )N at the start of each
√
epoch, Lemma 8 implies that E[Xi ] = Ω( N ). It follows by a Chernoff-Hoeffding
bound that with
√
all but negligible probability, for any constant c, X − E[X] ≤ c N , so with all but negligible
√
probability we have that X = Ω N and X1 + · · · + Xt = Ω(N 0.51 ). Consequently the population at
the start of epoch t exceeds (1 − α2 )N . The argument is identical when the population has exceeded
(1 + α2 )N .
We now conclude the proof of Theorem 2.
Proof of Theorem 2. Consider any polynomial f , and suppose that for some adversarial strategy
the population deviates from the interval [(1 − α)N, (1 + α)N ] in f (N ) rounds with non-negligible
probability. It follows that for some pair of epochs i < j ∈ [1 + ⌊f (N )/T ⌋], with non-negligible
probability the population deviates from interval [(1 − α)N, (1 + α)N ] for the first time in epoch j
after deviating from the interval [(1 − α2 )N, (1 + α2 )N ] for the last time in epoch i. We condition on
this event. Lemma 7 implies that until√epoch j, with all but neligible probability the population will
deviate in each epoch by at most Õ( N ). But then Lemma 9 implies that with all but negligible
probability the population will return to interval [(1 − α2 )N, (1 + α2 )N ] within N 0.01 epochs, which
is a contradiction. Consequently the population will remain between (1 − α)N and (1 + α)N with
high probability for any polynomial number of rounds, as desired.
References
[AAD+ 04]
Dana Angluin, James Aspnes, Zoë Diamadi, Michael J. Fischer, and René Peralta.
Computation in networks of passively mobile finite-state sensors. In Proceedings of
the Twenty-Third Annual ACM Symposium on Principles of Distributed Computing,
22
PODC 2004, St. John’s, Newfoundland, Canada, July 25-28, 2004, pages 290–299,
2004.
[AAD+ 06]
Dana Angluin, James Aspnes, Zoë Diamadi, Michael J. Fischer, and René Peralta.
Computation in networks of passively mobile finite-state sensors. Distributed Computing, 18(4):235–253, 2006.
[AAE07]
Dana Angluin, James Aspnes, and David Eisenstat. A simple population protocol
for fast robust approximate majority. In Distributed Computing, 21st International
Symposium, DISC 2007, Lemesos, Cyprus, September 24-26, 2007, Proceedings, pages
20–32, 2007.
[AAE+ 17]
Dan Alistarh, James Aspnes, David Eisenstat, Rati Gelashvili, and Ronald L. Rivest.
Time-space trade-offs in population protocols. In Proceedings of the Twenty-Eighth
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona,
Spain, Hotel Porta Fira, January 16-19, pages 2560–2579, 2017.
[AAER07]
Dana Angluin, James Aspnes, David Eisenstat, and Eric Ruppert. The computational
power of population protocols. Distributed Computing, 20(4):279–304, 2007.
[AAFJ08]
Dana Angluin, James Aspnes, Michael J. Fischer, and Hong Jiang. Self-stabilizing
population protocols. TAAS, 3(4):13:1–13:28, 2008.
[AAG18]
Dan Alistarh, James Aspnes, and Rati Gelashvili. Space-optimal majority in population protocols. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium
on Discrete Algorithms, pages 2221–2239. SIAM, 2018.
[ABBS16]
James Aspnes, Joffroy Beauquier, Janna Burman, and Devan Sohier. Time and space
optimal counting in population protocols. In 20th International Conference on Principles of Distributed Systems, OPODIS 2016, December 13-16, 2016, Madrid, Spain,
pages 13:1–13:17, 2016.
[BCG04]
E. R Berlekamp, John Horton Conway, and R. K. Guy. Winning ways for your
mathematical plays. In A K Peters 1st Ed., 2001-2004.
[CDLN14]
Alejandro Cornejo, Anna R. Dornhaus, Nancy A. Lynch, and Radhika Nagpal. Task
allocation in ant colonies. In Distributed Computing - 28th International Symposium,
DISC 2014, Austin, TX, USA, October 12-15, 2014. Proceedings, pages 46–60, 2014.
[Coo04]
Matthew Cook. Universality in elementary cellular automata, 2004.
[DFGR06]
Carole Delporte-Gallet, Hugues Fauconnier, Rachid Guerraoui, and Eric Ruppert.
When birds die: Making population protocols fault-tolerant. In Distributed Computing in Sensor Systems, Second IEEE International Conference, DCOSS 2006, San
Francisco, CA, USA, June 18-20, 2006, Proceedings, pages 51–66, 2006.
[DH97]
Shlomi Dolev and Ted Herman. Superstabilizing protocols for dynamic distributed
systems. Chicago J. Theor. Comput. Sci., 1997.
[Dij74]
Edsger W. Dijkstra. Self-stabilizing systems in spite of distributed control. Commun.
ACM, pages 643–644, 1974.
23
[DLECW92] G Martinez De La Escalera, AL Choi, and Richard I Weiner. Generation and synchronization of gonadotropin-releasing hormone (gnrh) pulses: intrinsic properties of
the gt1-1 gnrh neuronal cell line. Proceedings of the National Academy of Sciences,
89(5):1852–1855, 1992.
[DMPTH10] Tal Danino, Octavio Mondragón-Palomino, Lev Tsimring, and Jeff Hasty. A synchronized quorum of genetic clocks. Nature, 463(7279):326, 2010.
[Gar70]
Martin Gardner. The fantastic combinations of john conway’s new solitaire game
“life”, 1970.
[GMRL15]
Mohsen Ghaffari, Cameron Musco, Tsvetomira Radeva, and Nancy A. Lynch. Distributed house-hunting in ant colonies. In Proceedings of the 2015 ACM Symposium
on Principles of Distributed Computing, PODC 2015, Donostia-San Sebastián, Spain,
July 21 - 23, 2015, pages 57–66, 2015.
[GR17]
Oded Goldreich and Dana Ron. On learning and testing dynamic environments.
Journal of the ACM (JACM), 64(3):21, 2017.
[JMPT87]
HJ Jongsma, M Masson-Pevet, and L Tsjernina. The development of beat-rate
synchronization of rat myocyte pairs in cell culture. Basic research in cardiology,
82(5):454–464, 1987.
[LBBC14a]
Giuseppe Antonio Di Luna, Roberto Baldoni, Silvia Bonomi, and Ioannis Chatzigiannakis. Conscious and unconscious counting on anonymous dynamic networks.
In Distributed Computing and Networking - 15th International Conference, ICDCN
2014, Coimbatore, India, January 4-7, 2014. Proceedings, pages 257–271, 2014.
[LBBC14b]
Giuseppe Antonio Di Luna, Roberto Baldoni, Silvia Bonomi, and Ioannis Chatzigiannakis. Counting in anonymous dynamic networks under worst-case adversary.
In IEEE 34th International Conference on Distributed Computing Systems, ICDCS
2014, Madrid, Spain, June 30 - July 3, 2014, pages 338–347, 2014.
[MO09]
Joseph S Markson and Erin K O’Shea. The molecular clockwork of a protein-based
circadian oscillator. FEBS letters, 583(24):3938–3947, 2009.
[Mor78]
Robert Morris. Counting large numbers of events in small registers. In ommunications
of the ACM, pages 840–842, 1978.
[vN51]
John von Neumann. The general and logical theory of automata, 1951.
24
| 8cs.DS
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.